Computer Vision - ECCV 2008

Lecture Notes in Computer Science 5305
Commenced Publication in 1973

Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen
Editorial Board
David Hutchison
Lancaster University, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Alfred Kobsa
University of California, Irvine, CA, USA
Friedemann Mattern
ETH Zurich, Switzerland
John C. Mitchell
Stanford University, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
Oscar Nierstrasz
University of Bern, Switzerland
C. Pandu Rangan
Indian Institute of Technology, Madras, India
Bernhard Steffen
University of Dortmund, Germany
Madhu Sudan
Massachusetts Institute of Technology, MA, USA
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max-Planck Institute of Computer Science, Saarbruecken, Germany
David Forsyth Philip Torr
Andrew Zisserman (Eds.)
Computer Vision –
ECCV 2008
10th European Conference on Computer Vision

Marseille, France, October 12-18, 2008
Proceedings, Part IV
13
Volume Editors
David Forsyth
University of Illinois at Urbana-Champaign, Computer Science Department
3310 Siebel Hall, Urbana, IL 61801, USA
E-mail: daf@cs.uiuc.edu
Philip Torr
Oxford Brookes University, Department of Computing
Wheatley, Oxford OX33 1HX, UK
E-mail: philiptorr@brookes.ac.uk
Andrew Zisserman
University of Oxford, Department of Engineering Science
Parks Road, Oxford OX1 3PJ, UK
E-mail: az@robots.ox.ac.uk
Library of Congress Control Number: 2008936989
CR Subject Classification (1998): I.4, I.2.10, I.5.4, I.5, I.7.5
LNCS Sublibrary: SL 6 – Image Processing, Computer Vision, Pattern Recognition,

and Graphics
ISSN 0302-9743
ISBN-10 3-540-88692-3 Springer Berlin Heidelberg New York
ISBN-13 978-3-540-88692-1 Springer Berlin Heidelberg New York
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,
in its current version, and permission for use must always be obtained from Springer. Violations are liable
to prosecution under the German Copyright Law.
Springer is a part of Springer Science+Business Media
springer.com
© Springer-Verlag Berlin Heidelberg 2008
Printed in Germany
Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India
Printed on acid-free paper SPIN: 12553631 06/3180 543210
Preface
Welcome to the 2008 European Conference on Computer Vision. These proceed-

ings are the result of a great deal of hard work by many people. To produce
them, a total of 871 papers were reviewed. Forty were selected for oral presen-
tation and 203 were selected for poster presentation, yielding acceptance rates
of 4.6% for oral, 23.3% for poster, and 27.9% in total.
We applied three principles. First, since we had a strong group of Area Chairs,
the final decisions to accept or reject a paper rested with the Area Chair, who
would be informed by reviews and could act only in consensus with another Area
Chair. Second, we felt that authors were entitled to a summary that explained
how the Area Chair reached a decision for a paper. Third, we were very careful
to avoid conflicts of interest.
Each paper was assigned to an Area Chair by the Program Chairs, and each
Area Chair received a pool of about 25 papers. The Area Chairs then identified
and ranked appropriate reviewers for each paper in their pool, and a constrained
optimization allocated three reviewers to each paper. We are very proud that
every paper received at least three reviews.
At this point, authors were able to respond to reviews. The Area Chairs then
needed to reach a decision. We used a series of procedures to ensure careful
review and to avoid conflicts of interest. Program Chairs did not submit papers.
The Area Chairs were divided into three groups so that no Area Chair in the
group was in conflict with any paper assigned to any Area Chair in the group.
Each Area Chair had a “buddy” in their group. Before the Area Chairs met,
they read papers and reviews, contacted reviewers to get reactions to submissions
and occasionally asked for improved or additional reviews, and prepared a rough
summary statement for each of the papers in their pool.
At the Area Chair meeting, groups met separately so that Area Chairs could
reach a consensus with their buddies, and make initial oral/poster decisions.
We met jointly so that we could review the rough program, and made final
oral/poster decisions in groups. In the separate meetings, there were no conflicts.
In the joint meeting, any Area Chairs with conflicts left the room when relevant
papers were discussed. Decisions were published on the last day of the Area
Chair meeting.
There are three more somber topics to report. First, the Program Chairs
had to deal with several double submissions. Referees or Area Chairs identified
potential double submissions, we checked to see if these papers met the criteria
published in the call for papers, and if they did, we rejected the papers and did
not make reviews available. Second, two submissions to ECCV 2008 contained
open plagiarism of published works. We will pass details of these attempts to
journal editors and conference chairs to make further plagiarism by the respon-
sible parties more difficult. Third, by analysis of server logs we discovered that
VI Preface
there had been a successful attempt to download all submissions shortly after
the deadline. We warned all authors that this had happened to ward off dangers
to intellectual property rights, and to minimize the chances that an attempt at
plagiarism would be successful. We were able to identify the responsible party,
discussed this matter with their institutional management, and believe we re-
solved the issue as well as we could have. Still, it is important to be aware that
no security or software system is completely safe, and papers can leak from
conference submission.
We felt the review process worked well, and recommend it to the community.
The process would not have worked without the efforts of many people. We thank
Lyndsey Pickup, who managed the software system, author queries, Area Chair
queries and general correspondence (most people associated with the conference
will have exchanged e-mails with her at some point). We thank Simon Baker,
Ramin Zabih and especially Jiřı́ Matas for their wise advice on how to organize
and run these meetings; the process we have described is largely their model from
CVPR 2007. We thank Jiřı́ Matas and Dan Večerka, for extensive help with,
and support of, the software system. We thank C. J. Taylor for the 3-from-5
optimization code. We thank the reviewers for their hard work. We thank the
Area Chairs for their very hard work, and for the time and attention each gave
to reading papers, reviews and summaries, and writing summaries.
We thank the Organization Chairs Peter Sturm and Edmond Boyer, and
the General Chair, Jean Ponce, for their help and support and their sharing of
the load. Finally, we thank Nathalie Abiola, Nasser Bacha, Jacques Beigbeder,
Jerome Bertsch, Joëlle Isnard and Ludovic Ricardou of ENS for administra-
tive support during the Area Chair meeting, and Danièle Herzog and Laetitia
Libralato of INRIA Rhône-Alpes for administrative support after the meeting.
August 2008 Andrew Zisserman

David Forsyth
Philip Torr
Organization
Conference Chair
Jean Ponce Ecole Normale Supérieure, France
Honorary Chair
Jan Koenderink EEMCS, Delft University of Technology,
The Netherlands
Program Chairs
David Forsyth University of Illinois, USA
Philip Torr Oxford Brookes University, UK
Andrew Zisserman University of Oxford, UK
Organization Chairs
Edmond Boyer LJK/UJF/INRIA Grenoble–Rhône-Alpes, France
Peter Sturm INRIA Grenoble–Rhône-Alpes, France
Specialized Chairs
Frédéric Jurie Workshops Université de Caen, France
Frédéric Devernay Demos INRIA Grenoble–Rhône-Alpes,
France
Edmond Boyer Video Proc. LJK/UJF/INRIA
Grenoble–Rhône-Alpes, France
James Crowley Video Proc. INPG, France
Nikos Paragios Tutorials Ecole Centrale, France
Emmanuel Prados Tutorials INRIA Grenoble–Rhône-Alpes,
France
Christophe Garcia Industrial Liaison France Telecom Research, France
Théo Papadopoulo Industrial Liaison INRIA Sophia, France
Jiřı́ Matas Conference Software CTU Prague, Czech Republic
Dan Večerka Conference Software CTU Prague, Czech Republic
Program Chair Support

Lyndsey Pickup University of Oxford, UK
VIII Organization
Administration
Danile Herzog INRIA Grenoble–Rhône-Alpes, France
Laetitia Libralato INRIA Grenoble–Rhône-Alpes, France
Conference Website
Elisabeth Beaujard INRIA Grenoble–Rhône-Alpes, France
Amaël Delaunoy INRIA Grenoble–Rhône-Alpes, France
Mauricio Diaz INRIA Grenoble–Rhône-Alpes, France
Benjamin Petit INRIA Grenoble–Rhône-Alpes, France
Printed Materials
Ingrid Mattioni INRIA Grenoble–Rhône-Alpes, France
Vanessa Peregrin INRIA Grenoble–Rhône-Alpes, France
Isabelle Rey INRIA Grenoble–Rhône-Alpes, France
Area Chairs
Horst Bischof Graz University of Technology, Austria
Michael Black Brown University, USA
Andrew Blake Microsoft Research Cambridge, UK
Stefan Carlsson NADA/KTH, Sweden
Tim Cootes University of Manchester, UK
Alyosha Efros CMU, USA
Jan-Olof Eklund KTH, Sweden
Mark Everingham University of Leeds, UK
Pedro Felzenszwalb University of Chicago, USA
Richard Hartley Australian National University, Australia
Martial Hebert CMU, USA
Aaron Hertzmann University of Toronto, Canada
Dan Huttenlocher Cornell University, USA
Michael Isard Microsoft Research Silicon Valley, USA
Aleš Leonardis University of Ljubljana, Slovenia
David Lowe University of British Columbia, Canada
Jiřı́ Matas CTU Prague, Czech Republic
Joe Mundy Brown University, USA
David Nistér Microsoft Live Labs/Microsoft Research, USA
Tomáš Pajdla CTU Prague, Czech Republic
Patrick Pérez IRISA/INRIA Rennes, France
Marc Pollefeys ETH Zürich, Switzerland
Ian Reid University of Oxford, UK
Cordelia Schmid INRIA Grenoble–Rhône-Alpes, France
Bernt Schiele Darmstadt University of Technology, Germany
Christoph Schnörr University of Mannheim, Germany
Steve Seitz University of Washington, USA
Organization IX
Richard Szeliski Microsoft Research, USA

Antonio Torralba MIT, USA
Bill Triggs CNRS/Laboratoire Jean Kuntzmann, France
Tinne Tuytelaars Katholieke Universiteit Leuven, Belgium
Luc Van Gool Katholieke Universiteit Leuven, Belgium
Yair Weiss The Hebrew University of Jerusalem, Israel
Chris Williams University of Edinburgh, UK
Ramin Zabih Cornell University, USA
Conference Board
Horst Bischof Graz University of Technology, Austria
Hans Burkhardt University of Freiburg, Germany
Bernard Buxton University College London, UK
Roberto Cipolla University of Cambridge,UK
Jan-Olof Eklundh Royal Institute of Technology, Sweden
Olivier Faugeras INRIA, Sophia Antipolis, France
Anders Heyden Lund University, Sweden
Aleš Leonardis University of Ljubljana, Slovenia
Bernd Neumann University of Hamburg, Germany
Mads Nielsen IT University of Copenhagen, Denmark
Tomáš Pajdla CTU Prague, Czech Republic
Giulio Sandini University of Genoa, Italy
David Vernon Trinity College, Ireland
Program Committee
Sameer Agarwal Tamara Berg Thomas Brox
Aseem Agarwala James Bergen Andrés Bruhn
Jörgen Ahlberg Marcelo Bertalmio Antoni Buades
Narendra Ahuja Bir Bhanu Joachim Buhmann
Yiannis Aloimonos Stan Bileschi Hans Burkhardt
Tal Arbel Stan Birchfield Andrew Calway
Kalle Åström Volker Blanz Rodrigo Carceroni
Peter Auer Aaron Bobick Gustavo Carneiro
Jonas August Endre Boros M. Carreira-Perpinan
Shai Avidan Terrance Boult Tat-Jen Cham
Simon Baker Richard Bowden Rama Chellappa
Kobus Barnard Edmond Boyer German Cheung
Adrien Bartoli Yuri Boykov Ondřej Chum
Benedicte Bascle Gary Bradski James Clark
Csaba Beleznai Chris Bregler Isaac Cohen
Peter Belhumeur Thomas Breuel Laurent Cohen
Serge Belongie Gabriel Brostow Michael Cohen
Moshe Ben-Ezra Matthew Brown Robert Collins
Alexander Berg Michael Brown Dorin Comaniciu
X Organization
James Coughlan Christopher Geyer Esther Koller-Meier

David Crandall Michael Goesele Vladimir Kolmogorov
Daniel Cremers Dan Goldman Nikos Komodakis
Antonio Criminisi Shaogang Gong Kurt Konolige
David Cristinacce Leo Grady Jana Košecká
Gabriela Csurka Kristen Grauman Zuzana Kukelova
Navneet Dalal Eric Grimson Sanjiv Kumar
Kristin Dana Fred Hamprecht Kyros Kutulakos
Kostas Daniilidis Edwin Hancock Ivan Laptev
Larry Davis Allen Hanson Longin Jan Latecki
Andrew Davison James Hays Svetlana Lazebnik
Nando de Freitas Carlos Hernández Erik Learned-Miller
Daniel DeMenthon Anders Heyden Yann Lecun
David Demirdjian Adrian Hilton Bastian Leibe
Joachim Denzler David Hogg Vincent Lepetit
Michel Dhome Derek Hoiem Thomas Leung
Sven Dickinson Alex Holub Anat Levin
Gianfranco Doretto Anthony Hoogs Fei-Fei Li
Gyuri Dorko Daniel Huber Hongdong Li
Pinar Duygulu Sahin Alexander Ihler Stephen Lin
Charles Dyer Michal Irani Jim Little
James Elder Hiroshi Ishikawa Ce Liu
Irfan Essa David Jacobs Yanxi Liu
Andras Ferencz Bernd Jähne Brian Lovell
Rob Fergus Hervé Jégou Simon Lucey
Vittorio Ferrari Ian Jermyn John Maccormick
Sanja Fidler Nebojsa Jojic Petros Maragos
Mario Figueiredo Michael Jones Aleix Martinez
Graham Finlayson Frédéric Jurie Iain Matthews
Robert Fisher Timor Kadir Wojciech Matusik
François Fleuret Fredrik Kahl Bruce Maxwell
Wolfgang Förstner Amit Kale Stephen Maybank
Charless Fowlkes Kenichi Kanatani Stephen McKenna
Jan-Michael Frahm Sing Bing Kang Peter Meer
Friedrich Fraundorfer Robert Kaucic Etienne Mémin
Bill Freeman Qifa Ke Dimitris Metaxas
Brendan Frey Renaud Keriven Branislav Mičušı́k
Andrea Frome Charles Kervrann Krystian Mikolajczyk
Pascal Fua Ron Kikinis Anurag Mittal
Yasutaka Furukawa Benjamin Kimia Theo Moons
Daniel Gatica-Perez Ron Kimmel Greg Mori
Dariu Gavrila Josef Kittler Pawan Mudigonda
James Gee Hedvig Kjellström David Murray
Guido Gerig Leif Kobbelt Srinivasa Narasimhan
Theo Gevers Pushmeet Kohli Randal Nelson
Organization XI
Ram Nevatia Radim Šára John Tsotsos

Jean-Marc Odobez Eric Saund Peter Tu
Björn Ommer Silvio Savarese Matthew Turk
Nikos Paragios Daniel Scharstein Oncel Tuzel
Vladimir Pavlovic Yoav Schechner Carole Twining
Shmuel Peleg Kondrad Schindler Ranjith Unnikrishnan
Marcello Pelillo Stan Sclaroff Raquel Urtasun
Pietro Perona Mubarak Shah Joost Van de Weijer
Maria Petrou Gregory Shakhnarovich Manik Varma
Vladimir Petrovic Eli Shechtman Nuno Vasconcelos
Jonathon Phillips Jianbo Shi Olga Veksler
Matti Pietikäinen Kaleem Siddiqi Jakob Verbeek
Axel Pinz Leonid Sigal Luminita Vese
Robert Pless Sudipta Sinha Thomas Vetter
Tom Pock Josef Sivic René Vidal
Fatih Porikli Cristian Sminchişescu George Vogiatzis
Simon Prince Anuj Srivastava Daphna Weinshall
Long Quan Drew Steedly Michael Werman
Ravi Ramamoorthi Gideon Stein Tomáš Werner
Deva Ramanan Björn Stenger Richard Wildes
Anand Rangarajan Christoph Strecha Lior Wolf
Ramesh Raskar Erik Sudderth Ying Wu
Xiaofeng Ren Josephine Sullivan Eric Xing
Jens Rittscher David Suter Yaser Yacoob
Rómer Rosales Tomáš Svoboda Ruigang Yang
Bodo Rosenhahn Hai Tao Stella Yu
Peter Roth Marshall Tappen Lihi Zelnik-Manor
Stefan Roth Demetri Terzopoulos Richard Zemel
Volker Roth Carlo Tomasi Li Zhang
Carsten Rother Fernando Torre S. Zhou
Fred Rothganger Lorenzo Torresani Song-Chun Zhu
Daniel Rueckert Emanuele Trucco Todd Zickler
Dimitris Samaras David Tschumperlé Lawrence Zitnick
Additional Reviewers
Lourdes Agapito Ross Beveridge Yixin Chen
Daniel Alexander V. Bhagavatula Dmitry Chetverikov
Elli Angelopoulou Edwin Bonilla Sharat Chikkerur
Alexandru Balan Aeron Buchanan Albert Chung
Adrian Barbu Michael Burl Nicholas Costen
Nick Barnes Tiberio Caetano Gabriela Oana Cula
João Barreto Octavia Camps Goksel Dedeoglu
Marian Bartlett Sharat Chandran Hervé Delingette
Herbert Bay François Chaumette Michael Donoser
XII Organization
Mark Drew Mike Langer Michael Ross

Zoran Duric Georg Langs Szymon Rusinkiewicz
Wolfgang Einhauser Neil Lawrence Bryan Russell
Aly Farag Sang Lee Sudeep Sarkar
Beat Fasel Boudewijn Lelieveldt Yoichi Sato
Raanan Fattal Marc Levoy Ashutosh Saxena
Paolo Favaro Michael Lindenbaum Florian Schroff
Rogerio Feris Chengjun Liu Stephen Se
Cornelia Fermüller Qingshan Liu Nicu Sebe
James Ferryman Manolis Lourakis Hans-Peter Seidel
David Forsyth Ameesh Makadia Steve Seitz
Jean-Sébastien Franco Ezio Malis Thomas Serre
Mario Fritz R. Manmatha Alexander Shekhovtsov
Andrea Fusiello David Martin Ilan Shimshoni
Meirav Galun Daniel Martinec Michal Sofka
Bogdan Georgescu Yasuyuki Matsushita Jan Solem
A. Georghiades Helmut Mayer Gerald Sommer
Georgy Gimel’farb Christopher Mei Jian Sun
Roland Goecke Paulo Mendonça Rahul Swaminathan
Toon Goedeme Majid Mirmehdi Hugues Talbot
Jacob Goldberger Philippos Mordohai Chi-Keung Tang
Luis Goncalves Pierre Moreels Xiaoou Tang
Venu Govindaraju P.J. Narayanan C.J. Taylor
Helmut Grabner Nassir Navab Jean-Philippe Thiran
Michael Grabner Jan Neumann David Tolliver
Hayit Greenspan Juan Carlos Niebles Yanghai Tsin
Etienne Grossmann Ko Nishino Zhuowen Tu
Richard Harvey Thomas O’Donnell Vaibhav Vaish
Sam Hasinoff Takayuki Okatani Anton van den Hengel
Horst Haussecker Kenji Okuma Bram Van Ginneken
Jesse Hoey Margarita Osadchy Dirk Vandermeulen
Slobodan Ilic Mustafa Ozuysal Alessandro Verri
Omar Javed Sharath Pankanti Hongcheng Wang
Qiang Ji Sylvain Paris Jue Wang
Jiaya Jia James Philbin Yizhou Wang
Hailin Jin Jean-Philippe Pons Gregory Welch
Ioannis Kakadiaris Emmanuel Prados Ming-Hsuan Yang
Joni-K. Kämäräinen Zhen Qian Caspi Yaron
George Kamberov Ariadna Quattoni Jieping Ye
Yan Ke Ali Rahimi Alper Yilmaz
Andreas Klaus Ashish Raj Christopher Zach
Georg Klein Visvanathan Ramesh Hongyuan Zha
Reinhard Koch Christopher Rasmussen Cha Zhang
Mathias Kolsch Tammy Riklin-Raviv Jerry Zhu
Andreas Koschan Charles Rosenberg Lilla Zollei
Christoph Lampert Arun Ross
Organization XIII
Sponsoring Institutions
Table of Contents – Part IV
Segmentation
Image Segmentation in the Presence of Shadows and Highlights . . . . . . . . 1
Eduard Vazquez, Joost van de Weijer, and Ramon Baldrich
Image Segmentation by Branch-and-Mincut . . . . . . . . . . . . . . . . . . . . . . . . . 15

Victor Lempitsky, Andrew Blake, and Carsten Rother
What Is a Good Image Segment? A Unified Approach to Segment

Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Shai Bagon, Oren Boiman, and Michal Irani
Computational Photography
Light-Efficient Photography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Samuel W. Hasinoff and Kiriakos N. Kutulakos
Flexible Depth of Field Photography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

Hajime Nagahara, Sujit Kuthirummal, Changyin Zhou, and
Shree K. Nayar
Priors for Large Photo Collections and What They Reveal about
Cameras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Sujit Kuthirummal, Aseem Agarwala, Dan B Goldman, and
Shree K. Nayar
Understanding Camera Trade-Offs through a Bayesian Analysis of

Light Field Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Anat Levin, William T. Freeman, and Frédo Durand
Poster Session IV
CenSurE: Center Surround Extremas for Realtime Feature Detection
and Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Motilal Agrawal, Kurt Konolige, and Morten Rufus Blas
Searching the World’s Herbaria: A System for Visual Identification of

Plant Species . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
Peter N. Belhumeur, Daozheng Chen, Steven Feiner,
David W. Jacobs, W. John Kress, Haibin Ling, Ida Lopez,
Ravi Ramamoorthi, Sameer Sheorey, Sean White, and Ling Zhang
XVI Table of Contents – Part IV
A Column-Pivoting Based Strategy for Monomial Ordering in

Numerical Gröbner Basis Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Martin Byröd, Klas Josephson, and Kalle Åström
Co-recognition of Image Pairs by Data-Driven Monte Carlo Image
Exploration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Minsu Cho, Young Min Shin, and Kyoung Mu Lee
Movie/Script: Alignment and Parsing of Video and Text
Transcription . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Timothee Cour, Chris Jordan, Eleni Miltsakaki, and Ben Taskar
Using 3D Line Segments for Robust and Efficient Change Detection
from Multiple Noisy Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
Ibrahim Eden and David B. Cooper
Action Recognition with a Bio–inspired Feedforward Motion Processing
Model: The Richness of Center-Surround Interactions . . . . . . . . . . . . . . . . . 186
Maria-Jose Escobar and Pierre Kornprobst
Linking Pose and Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
Andrea Fossati and Pascal Fua
Automated Delineation of Dendritic Networks in Noisy Image Stacks . . . 214
Germán González, François Fleuret, and Pascal Fua
Calibration from Statistical Properties of the Visual World . . . . . . . . . . . . 228
Etienne Grossmann, José António Gaspar, and Francesco Orabona
Regular Texture Analysis as Statistical Model Selection . . . . . . . . . . . . . . . 242
Junwei Han, Stephen J. McKenna, and Ruixuan Wang
Higher Dimensional Affine Registration and Vision Applications . . . . . . . 256
Yu-Tseh Chi, S.M. Nejhum Shahed, Jeffrey Ho, and
Ming-Hsuan Yang
Semantic Concept Classification by Joint Semi-supervised Learning of
Feature Subspaces and Support Vector Machines . . . . . . . . . . . . . . . . . . . . . 270
Wei Jiang, Shih-Fu Chang, Tony Jebara, and Alexander C. Loui
Learning from Real Images to Model Lighting Variations for Face
Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
Xiaoyue Jiang, Yuk On Kong, Jianguo Huang, Rongchun Zhao, and
Yanning Zhang
Toward Global Minimum through Combined Local Minima . . . . . . . . . . . . 298
Ho Yub Jung, Kyoung Mu Lee, and Sang Uk Lee
Differential Spatial Resection - Pose Estimation Using a Single Local
Image Feature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
Kevin Köser and Reinhard Koch
Table of Contents – Part IV XVII
Riemannian Anisotropic Diffusion for Tensor Valued Images . . . . . . . . . . . 326

Kai Krajsek, Marion I. Menzel, Michael Zwanger, and Hanno Scharr
FaceTracer: A Search Engine for Large Collections of Images with

Faces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
Neeraj Kumar, Peter Belhumeur, and Shree Nayar
What Does the Sky Tell Us about the Camera? . . . . . . . . . . . . . . . . . . . . . . 354

Jean-François Lalonde, Srinivasa G. Narasimhan, and
Alexei A. Efros
Three Dimensional Curvilinear Structure Detection Using Optimally

Oriented Flux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
Max W.K. Law and Albert C.S. Chung
Scene Segmentation for Behaviour Correlation . . . . . . . . . . . . . . . . . . . . . . . 383

Jian Li, Shaogang Gong, and Tao Xiang
Robust Visual Tracking Based on an Effective Appearance Model . . . . . . 396

Xi Li, Weiming Hu, Zhongfei Zhang, and Xiaoqin Zhang
Key Object Driven Multi-category Object Recognition, Localization

and Tracking Using Spatio-temporal Context . . . . . . . . . . . . . . . . . . . . . . . . 409
Yuan Li and Ram Nevatia
A Pose-Invariant Descriptor for Human Detection and Segmentation . . . 423

Zhe Lin and Larry S. Davis
Texture-Consistent Shadow Removal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437

Feng Liu and Michael Gleicher
Scene Discovery by Matrix Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451

Nicolas Loeff and Ali Farhadi
Simultaneous Detection and Registration for Ileo-Cecal Valve Detection

in 3D CT Colonography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
Le Lu, Adrian Barbu, Matthias Wolf, Jianming Liang, Luca Bogoni,
Marcos Salganicoff, and Dorin Comaniciu
Constructing Category Hierarchies for Visual Recognition . . . . . . . . . . . . . 479

Marcin Marszalek and Cordelia Schmid
Sample Sufficiency and PCA Dimension for Statistical Shape Models . . . 492
Lin Mei, Michael Figl, Ara Darzi, Daniel Rueckert, and
Philip Edwards
Locating Facial Features with an Extended Active Shape Model . . . . . . . 504

Stephen Milborrow and Fred Nicolls
XVIII Table of Contents – Part IV
Dynamic Integration of Generalized Cues for Person Tracking . . . . . . . . . 514

Kai Nickel and Rainer Stiefelhagen
Extracting Moving People from Internet Videos . . . . . . . . . . . . . . . . . . . . . . 527

Juan Carlos Niebles, Bohyung Han, Andras Ferencz, and Li Fei-Fei
Multiple Instance Boost Using Graph Embedding Based Decision Stump

for Pedestrian Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
Junbiao Pang, Qingming Huang, and Shuqiang Jiang
Object Detection from Large-Scale 3D Datasets Using Bottom-Up and

Top-Down Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
Alexander Patterson IV, Philippos Mordohai, and Kostas Daniilidis
Making Background Subtraction Robust to Sudden Illumination

Changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
Julien Pilet, Christoph Strecha, and Pascal Fua
Closed-Form Solution to Non-rigid 3D Surface Registration . . . . . . . . . . . . 581

Mathieu Salzmann, Francesc Moreno-Noguer, Vincent Lepetit, and
Pascal Fua
Implementing Decision Trees and Forests on a GPU . . . . . . . . . . . . . . . . . . 595

Toby Sharp
General Imaging Geometry for Central Catadioptric Cameras . . . . . . . . . . 609

Peter Sturm and João P. Barreto
Estimating Radiometric Response Functions from Image Noise

Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623
Jun Takamatsu, Yasuyuki Matsushita, and Katsushi Ikeuchi
Solving Image Registration Problems Using Interior Point Methods . . . . . 638

Camillo Jose Taylor and Arvind Bhusnurmath
3D Face Model Fitting for Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 652

Frank B. ter Haar and Remco C. Veltkamp
A Multi-scale Vector Spline Method for Estimating the Fluids Motion

on Satellite Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 665
Till Isambert, Jean-Paul Berroir, and Isabelle Herlin
Continuous Energy Minimization Via Repeated Binary Fusion . . . . . . . . . 677

Werner Trobin, Thomas Pock, Daniel Cremers, and Horst Bischof
Unified Crowd Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691

Peter Tu, Thomas Sebastian, Gianfranco Doretto, Nils Krahnstoever,
Jens Rittscher, and Ting Yu
Table of Contents – Part IV XIX
Quick Shift and Kernel Methods for Mode Seeking . . . . . . . . . . . . . . . . . . . 705

Andrea Vedaldi and Stefano Soatto
A Fast Algorithm for Creating a Compact and Discriminative Visual
Codebook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 719
Lei Wang, Luping Zhou, and Chunhua Shen
A Dynamic Conditional Random Field Model for Joint Labeling of
Object and Scene Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733
Christian Wojek and Bernt Schiele
Local Regularization for Multiclass Classification Facing Significant
Intraclass Variations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 748
Lior Wolf and Yoni Donner
Saliency Based Opportunistic Search for Object Part Extraction and
Labeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 760
Yang Wu, Qihui Zhu, Jianbo Shi, and Nanning Zheng
Stereo Matching: An Outlier Confidence Approach . . . . . . . . . . . . . . . . . . . 775
Li Xu and Jiaya Jia
Improving Shape Retrieval by Learning Graph Transduction . . . . . . . . . . . 788
Xingwei Yang, Xiang Bai, Longin Jan Latecki, and Zhuowen Tu
Cat Head Detection - How to Effectively Exploit Shape and Texture
Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 802
Weiwei Zhang, Jian Sun, and Xiaoou Tang
Motion Context: A New Representation for Human Action
Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 817
Ziming Zhang, Yiqun Hu, Syin Chan, and Liang-Tien Chia
Active Reconstruction
Temporal Dithering of Illumination for Fast Active Vision . . . . . . . . . . . . . 830
Srinivasa G. Narasimhan, Sanjeev J. Koppal, and
Shuntaro Yamazaki
Compressive Structured Light for Recovering Inhomogeneous
Participating Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845
Jinwei Gu, Shree Nayar, Eitan Grinspun, Peter Belhumeur, and
Ravi Ramamoorthi
Passive Reflectometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 859
Fabiano Romeiro, Yuriy Vasilyev, and Todd Zickler
Fusion of Feature- and Area-Based Information for Urban Buildings
Modeling from Aerial Imagery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 873
Lukas Zebedin, Joachim Bauer, Konrad Karner, and Horst Bischof
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 887

Image Segmentation in the Presence of Shadows
and Highlights
Eduard Vazquez, Joost van de Weijer, and Ramon Baldrich
Computer Vision Center \ Dpt. Ciencies de la Computacio. Edifici O Universitat

Autonoma de Barcelona. 08193 Cerdanyola del Valles, Barcelona, Spain
Abstract. The segmentation method proposed in this paper is based

on the observation that a single physical reflectance can have many dif-
ferent image values. We call the set of all these values a dominant colour.
These variations are caused by shadows, shading and highlights and due
to varying object geometry. The main idea is that dominant colours trace
connected ridges in the chromatic histogram. To capture them, we pro-
pose a new Ridge based Distribution Analysis (RAD) to find the set of
ridges representative of the dominant colour. First, a multilocal crease-
ness technique followed by a ridge extraction algorithm is proposed. Af-
terwards, a flooding procedure is performed to find the dominant colours
in the histogram. Qualitative results illustrate the ability of our method
to obtain excellent results in the presence of shadow and highlight edges.
Quantitative results obtained on the Berkeley data set show that our
method outperforms state-of-the-art segmentation methods at low com-
putational cost.
1 Introduction
Image segmentation is a computer vision process consisting in the partition of an

image into a set of non-overlapped regions. A robust and efficient segmentation
is required as a preprocessing step in several computer vision tasks such as
object recognition or tracking. On real images the varying shapes of the objects
provoke several effects related with the illumination such as shadows, shading
and highlights. These effects, are one of the main difficulties that have to be
solved to obtain a correct segmentation.
There exist several different methods covering a broad spectrum of points
of view. The work presented by Skarbek and Koschan [1], draws the basis of
the current classifications of segmentation methods. Some other comprehensive
surveys of colour segmentation techniques are presented in [2] and [3], where a
similar schema is followed. From these works segmentation methods are divided
in four main categories: feature-based, image-based, physics-based and hybrid
approaches. Feature-based approaches are focused on the photometric informa-
tion of an image represented on its histogram [4],[5]. Image-based approaches
exploit the spatial information of the colour in an image, named spatial coher-
ence [6]. Physics-based methods use physics and psychophysics information to
D. Forsyth, P. Torr, and A. Zisserman (Eds.): ECCV 2008, Part IV, LNCS 5305, pp. 1–14, 2008.

c Springer-Verlag Berlin Heidelberg 2008
2 E. Vazquez, J. van de Weijer, and R. Baldrich
perform the segmentation. Finally, hybrid techniques combine methods of the

previous categories.
This paper introduces a method that exploits exclusively the photometric in-
formation of an image on its histogram. Therefore, it belongs to the category
of feature-based segmentation methods. This category can be further split in
three main categories, i.e., histogram thresholding, clustering and fuzzy cluster-
ing. Histogram thresholding techniques assume that there exist a threshold value
that isolates all pixels representative of an object in a scene. This basic concept
is exploited in several ways as explained in [7]. Clustering techniques, also named
hard clustering, perform a partition of the feature space under different criteria
such a distance measure as k-means or ISODATA, probabilistic/statistical ap-
proaches, such as Mean Shift [8], or the spectral analysis of the data [9], based
on the Karhunen-Loeve transformation. Fuzzy clustering includes methods such
as fuzzy k-means, Gath-Geva clustering, or mixture models [10], [11] which are
a way to look for areas of high density. The most related technique with the
one introduced in this paper, is the Mean shift, which will be commented and
compared with our method in section 4. Each technique has its own advantages
and drawbacks. There is a difficulty shared between all these methods, i.e., their
behaviour in the presence of shadows, shading, and highlights. Furthermore, the
work presented by Martin et al. in [12], points out the existence of strong edges
related with this physical effects in an image that are not considered in a hu-
man segmentation. These edges are detected for both image and feature based
methods.
Our approach. Our approach to colour image segmentation is based on the
insight that the distributions formed by a single-colored object have a physically
determined shape in colour histogram-space. We model an image as being gener-
ated by a set of dominant colours (DC), where each dominant colour is described
by a distribution in histogram-space. Each DC is related to a semantic object in
the image. For example, in Figure 1 we distinguish between four different DC’s,
namely: red for the pepper, green and brown for the branch and black for the
background.
A DC generates many image values due to geometrical and photometric varia-
tions. Our main aim is to find a good representation of the topologies which DC’s
are likely to form in histogram space. For this purpose, consider the distribution
of a single DC as described by the dichromatic reflection model [13]:
f (x) = mb (x) cb + mi (x) ci (1)
in which f = {R, G, B}, cb is the body reflectance, ci the surface reflectance,

mb and mi are geometry dependent scalars representing the magnitude of body
and surface reflectance. Bold notation is used to indicate vectors. For one DC
we expect both cb and ci to be almost constant, whereas mb (x) and mi (x) are
expected to vary significantly.
The two parts of the dichromatic reflectance model are clearly visible in the
histogram of Figure 1b. Firstly, due to the shading variations the distribution
of the red pepper traces an elongated shape in histogram-space. Secondly, the
Image Segmentation in the Presence of Shadows and Highlights 3
surface reflectance forms a branch which points in the direction of the reflected
illuminant. In conclusion, the distribution of a single DC forms a ridge-like struc-
ture in histogram space.
a) b) c) d)
Fig. 1. (a) An image from [14] and (b) its histogram. The effects of shading and
highlights are clearly visible in the red colours of the histogram. (c) Segmented images
using RAD. (d) Ridges found with RAD. Note that the three branches of the red
pepper are correctly connected in a single ridge.
To illustrate the difficulty of extracting the distributions of DCs consider

Figure 2c, which contains a patch of the horse image. The 2D Red-Green his-
togram of the patch is depicted in Figure 2d to see the number of occurrences of
each chromatic combination. This is done for explanation purposes. In this 2D
histogram it can be clearly seen that the density of the geometric term mb (x)
varies significantly, and the distribution is broken in two parts. However, we have
an important clue that the two distributions belong to the same DC: the orien-
tation of the two distribution is similar, which means they have a similar cb . We
exploit this feature in the ridge extraction algorithm by connecting neighboring
distributions with similar orientation.
a) b) c) d)
Fig. 2. (a) An image and (b) its 3D RGB histogram. (c) A patch of a) and its RGB
histogram. (d) 2D histogram of c) to illustrate the discontinuities appearing on a DC.
In literature several methods have explicitly used the dichromatic model to

obtain image segmentation, e.g. [15]. A drawback of such methods is however
that for many images Eq. 1 does only approximately model the data. This can
be caused by many reasons, such as non-linear acquisition systems, clipped high-
lights, and image compression. In this article we use Eq. 1 only to conclude that
objects described by this equation will trace connected ridges in histogram space.
This makes the method more robust to deviations from the dichromatic model.
This paper is organized as follows: in section 2 RAD is presented as a feature
space analysis method. Afterwards, in section 3 RAD is introduced as a segmen-
tation technique. The results obtained and a comparison with Mean Shift and
various other state-of-the-art methods on the Berkeley dataset is presented in
section 4. Finally, conclusions of the current work are given in section 5.
2 A Ridge Based Distribution Analysis Method (RAD)

In this section we present a fast algorithm to extract DCs from histogram space.
Here we propose a method to find dominant structures (DS) for a d-dimensional
feature space. In the context of this paper the dominant colours are the dominant
structures of the 3D chromatic histogram. The proposed method is divided in two
main steps. First, we propose a method to extract ridges as a representative of a
DS. Afterwards a flooding process is performed to find the DSs from its ridges.
2.1 First Step: Ridge Extraction

To extract a DS descriptor we need to find those points containing the most
meaningful information of a DS, i.e., the ridge of a DS. We propose to apply
a multilocal creaseness algorithm to find the best ridge point candidates. This
operator avoids to split up ridges due to irregularities on the distribution, mainly
caused by the discrete nature of the data. Afterwards, we apply a ridge extraction
algorithm to find the descriptor.
Multilocal Creaseness: Finding Candidates and Enhancing Connectiv-

ity. In order to deal with this commonly heavily jagged DS (see Figure 2d) , we
propose to apply the MLSEC-ST operator introduced by Lopez et al. in [16] to
enhance ridge points. This method is used due to its good performance compared
with other ridge detection methods [16] on irregular and noisy landscapes.
The Structure Tensor (ST) computes the dominant gradient orientation in a
neighbourhood of size proportional to σd . Basically, this calculus enhances those
situations where either a big attraction or repulsion exists in the gradient direc-
tion vectors. Thus, it assigns the higher values when a ridge or valley occurs.
Given a distribution Ω(x), (the histogram in the current context), and a sym-
metric neighbourhood of size σi centered at point x, namely, N(x, σi ) the ST
field S is defined as:
S(x, σ) = N(x, σi ) ∗ (∇Ω(x, σd ) · ∇Ωt (x, σd )) (2)
where σ = {σi , σd }, and the calculus of the gradient vector field ∇Ω(x, σd ) has
been done with a Gaussian Kernel with standard deviation σd .
If w(x, σ) is the eigenvector corresponding to the largest eigenvalue of S(x, σ),
then, the dominant gradient orientation w(x, σ) in a neighbourhood of size pro-
portional to σi centered at x is:
w(x, σ) = sign(wt (x, σ) · ∇t Ω(x, σd ))w(x, σ) (3)
The creaseness measure of Ω(x) for a given point x, named k(x, σ), is computed
with the divergence between the dominant gradient orientation and the normal
vectors, namely nk , on the r -connected neighbourhood of size proportional to
σi . That is:
d t
r
k(x, σ) = −Div(w(x, σ)) = − w (k, σ) · nk (4)
r
k=1
where d is the dimension of Ω(x). The creaseness representation of Ω(x) will be

referred hereafter as Ω σ .
As an example, Figure 3a shows the opponent colour 2D histogram of 3g.
Its creaseness values are showed in 3b. There are three enhanced areas which
corresponds with the three dominant colours of the original image. They appear
as three mountains in 3b, clearly separated by two valleys. Note that higher
creaseness values have a larger probability to become a ridge point.
a) b) c) d)
e) f) g) h)
Fig. 3. A graphical example of the whole process. (a) Opponent Red-Green and Blue-
Yellow histogram Ω(x) of g). (b) Creaseness representation of a). (c) Ridges found in
b). (d)Ridges fitted on original distribution. (e) Top-view of d). (f)Dominant structures
of a). (g) Original image. (h)Segmented image.
Ridge Detection. In the previous section we have detected a set of candidate

ridge points. In this section we discard superfluous points. As a result only those
points necessary to maintain the connectivity of a DS remain. These points form
the ridges of Ω σ .
We classify ridge points in three categories. First, Transitional Ridge Points
(TRP): when there is a local maximum in a single direction. Second, Saddle
Points (SP): when there is a local maximum in one direction and a local minimum
in another one. Third, Local Maximum Points (LMP). Formally, let Ω(x, y) be
a continuous 2D surface and ∇Ω(x, y) be the gradient vector of the function
Ω(x, y). We define ω1 and ω2 as the unit eigenvectors of the Hessian matrix and
λ1 and λ2 its corresponding eigenvalues with | λ1 |≤| λ2 |. Then, for the 2D case:
LM P (Ω(x, y)) = {(x, y)|(∇Ω(x, y) = 0), λ1 < 0, λ2 < 0} (5)
T RP (f (x, y)) = {(x, y)|∇Ω(x, y) = 0, λ1 < 0, ∇Ω(x, y) · ω1 = 0,

∇Ω(x, y) = 0, λ2 < 0, ∇Ω(x, y) · ω2 = 0, (6)
∇Ω(x, y) = 0, λ1 < 0, λ2 = 0}
SP (f (x, y)) = {(x, y)|∇Ω(x, y) = 0, λ1 · λ2 < 0} (7)

This definition can be extended for an arbitrary dimension using the combinato-
rial of the eigenvalues. Hereafter we will refer
to these
three categories as ridge
points (RP). Thus, RP (Ω(x, y)) = LM P T RP SP . A further classification
of ridges and its singularities can be found in [17] and [18].
A common way to detect RP is to find zero-crossing in the gradient of a
landscape for a given gradient direction. Thus, we need to compute all gradient
directions and detect changes following the schema proposed in [18]. In our
case, we propose a way to extract a ridge without the need to calculate the
gradient values for all points in the landscape. We begin on a local maxima
of the landscape and follow the ridge by adding the higher neighbours of the
current point, if there is a zero-crossing on it, until it reaches a flat region. This
method can be easily applied to an arbitrary dimension. A further explanation
can be found in [19].
Figure 3c depicts the RP found on Ω σ with black dots. Figures 3d,e show
a 3D view and a 2D projection view respectively of how these RPs fit in the
original distribution as a representative of the three DS. Finally, from the set of
RPs of a distribution we can perform the calculus of each DS. A second example
is shown in Figure 1. The complicated colour distribution of the pepper, caused
by shading and highlight effects, is correctly connected in a single ridge.
2.2 Second Step: DS Calculus from Its RPs

In this final step we find the DS belonging to each ridge found. From topological
point of view, it implies finding the portion of landscape represented by each
ridge. These portions of landscape are named catchments basins. Vincent and
Soille [20] define a catchment basin associated with a local minimum M as the set
of pixels p of Ω σ such that a water drop falling at p flows down along the relief,
following a certain descending path called the downstream of p, and eventually
reaches M. In our case, M are the set of RPs found and then, DSs are found
using the algorithm proposed in [20] applied on the inverse Ω σ distribution. The
proposed algorithm, is not based on the gradient vectors of a landscape [21] but
on the idea of immersion which is more stable and reduces over-segmentation.
Basically, the flooding process begins on the local minima and, iteratively, the
landscape sinks on the water. Those points where the water coming from different
local minima join, compose the watershed lines. To avoid potential problems with
irregularities [16], we force the flooding process to begin at the same time in all
DS descriptors, on the smoothed Ω(x) distribution with a Gaussian kernel of
standard deviation σd (already computed on the ST calculus). Then, we define
RAD as the operator returning the set of DS of Ω σ using RPs as marks:
RAD(Ω(x)) = W (Ω σ , RP (Ω σ )) (8)
Following this procedure, Figure 3f depicts the 2D projection of the DSs found
on 3a.
3 Colour Image Segmentation Using RAD

Once RAD has been applied we need to assign a representative colour to each
DS found. Thus, let DSn = {x1 , ..., xr } be the nth DS of Ω(x), and Ω(xi ) the
function returning the number of occurrences of xi in Ω. Then, the dominant
colour of DSn , namely, DC(DSn ) will be the mass center of Ω(DSn ):
r
i=1 xi · Ω(xi )
DC(DSn ) = r (9)
i=1 Ω(xi )
The segmented image will have as many colours as the number of DSs found.
Figure, 3h shows the segmentation obtained with RAD from 3g. This segmenta-
tion has been performed in the opponent colour histogram. Although RAD can
be applied to any chromatic representation of an image such as CIE, RGB, Ohta
spaces or 2-dimensional ones such as Opponent or normalized RGB.
4 Results and Performance Evaluation

In the experiments we qualitatively and quantitatively evaluate the proposed
segmentation method. Firstly, RAD is compared with Mean Shift (MS) [8], [22].
MS has been chosen because it is widely used, has a public available version, the
EDISON one [23] and it has demonstrated its good performance [24]. Addition-
ally, Mean Shift is a feature space analysis technique, as well as RAD, and yields
a segmentation in a rather reasonable time, in opposition to other set of methods
such as the Graph-Based approaches [25] , (with the exception of the efficient
graph-based segmentation method introduced in [26]). Secondly, our method is
compared on the Berkeley data set against a set of state-of-the-art segmentation
methods.
The MS method [22], consists of finding the modes of the underlying proba-
bility function of a distribution. The method finds the Mean Shift vectors in the
histogram of an image that point to the direction of higher density. All values
of the histogram attracted by one mode compound the basis of attraction of
it. In a second step, the modes which are near of a given threshold are joined
in one unique mode. Finally, the basis of attraction of these modes will com-
pose a dominant colour of the image. Mean Shift has two basic parameters to
adapt the segmentation to an specific problem, namely, hs , which controls a
smoothing process, and hr related with the size of the kernel used to determine
the modes and its basis of attraction. To test the method, we have selected the
set parameters (hs , hr ) = {(7, 3), (7, 15), (7, 19), (7, 23), (13, 7)(13, 19), (17, 23)}
given in [24] and [5]. The average times for this set of parameters, expressed in
seconds, are 3.17, 4.15, 3.99, 4.07, 9.72, 9.69, 13.96 respectively. Nevertheless,
these parameters do not cover the complete spectrum of possibilities of the MS.
Here we want to compare RAD and MS from a soft oversegmentation to a soft
undersegmentation. Hence, in order to reach an undersegmentation with MS, we
add the following parameter settings (hs , hr ) = {(20, 25), (25, 30), (30, 35)}. For
these settings, the average times are 18.05, 24.95 and 33.09 respectively.
The parameters used for RAD based segmentation are (σd ,σi )={ (0.8,0.05),
(0.8,0.5), (0.8,1), (0.8,1.5), (1.5,0.05), (1.5,0.5), (1.5,1.5), (2.5,0.05), (2.5,0.5),
(2.5,1.5) }. These parameters vary from a soft oversegmentation to an underseg-
mentation, and have been selected experimentally. The average times for RAD
are 6.04, 5.99, 6.11, 6.36, 6.11, 5.75, 6.44, 5.86, 5.74 and 6.35. These average
times, point out the fact that RAD is not dependent of the parameters used.
In conclusion, whereas the execution time of Mean Shift increases significantly
with increasing spatial scale, the execution time of RAD remains constant from
an oversegmentation to an undersegmentation.
The experiments has been performed on the publicly available Berkeley image
segmentation dataset and benchmark [12]. We use the Global Constancy Error
(GCE) as an error measure. This measure was also proposed in [12] and takes
care of the refinement between different segmentations. For a given pixel pi ,
consider the segments (sets of connected pixels), S1 from the benchmark and S2
Fig. 4. Examples of segmentation. Original image. Columns from 2 to 5: segmentation

for RAD on RGB with (σd ,σi )={(0.8,0.05),(1.5,0.05),(2.5,0.05),(2.5,1.5)}. Last column:
human segmentation.
from the segmented image that contain this pixel. If one segment is a proper
subset of the other, then pi lies in an area of refinement and the error measure
should be zero. If there is no subset relationship, then S1 and S2 overlap in an
inconsistent manner and the error is higher than zero, (up to one in the worst
possible case). MS segmentation has been done on the CIE Luv space since
this is the space used in [24] and [5]. RAD based segmentation has been done
on the RGB colour space for two reasons. First, the Berkeley image dataset
does not have calibrated images and, consequently, we can not assure a good
transformation from sRGB to CIE Luv. Second, because the size of L, u and
v, is not the same and the method will require six parameters, instead of two,
that is, −
σ→ −
→ −
→
L , σu and σv . Nonetheless, for the sake of clarity, we also present some
results of RAD on CIE Luv to directly compare results with MS.
Figure 4 depicts a set of examples for RAD on RGB. From left to right: original
image, RAD for (σd ,σi )={ (0.8,0.05) , (1.5,0.05) , (2.5,0.05) , (2.5,1.5) } and hu-
man segmentation. Figure 5 shows some results for the mean shift segmentation,
corresponding to (hs , hr ) = {(7, 15), (13, 19), (17, 23), (20, 25), (25, 30), (30, 35)}.
These results point out the main advantage of RAD in favor of MS, namely, the
capability of RAD to capture the DS of a histogram, whereas MS is ignorant
to the physical processes underlying the structure of the DSs as Abd-Almageed
and S. Davis explain in [10]. Graphically, the set of images depicted in the first
row of Figure 5, shows this behavior in a practical case. In the last column, MS
joins rocks with the mountain, and the mountain with the sky, but is not able
to find one unique structure for a rock or for the mountain, whereas RAD, as
shown in Figure 4, is able to do.
A danger of RAD is that for some parameter settings it is prone to underseg-
menting. Consequently it finds only one dominant colour for the whole image.
This happens in some cases for (σd ,σi )={(2.5,1),(2.5,1.5)}, as Figure 6 illus-
trates. In the first example, the aircraft has a bluish colour similar to the sky,
as well as the fish and its environment in the second example.
Additional examples related to the presence of physical effects, such as shad-
ows, shading and highlights are shown in Figure 7. The good performance of
RAD in these conditions can be clearly observed for the skin of the people, the
elephants and buffalos, as well as for the clothes of the people.
Fig. 5. MS segmentation examples for different parameters. Columns from 1 to 5:

(hs , hr ) = {(7, 15), (13, 19), (17, 23), (20, 25), (25, 30)}.
Fig. 6. Examples of undersegmentation. Original image. Columns from 2 to 5: segmen-

tation for RAD with (σd ,σi )={(0.8,0.05),(1.5,0.05),(2.5,0.05),(2.5,1.5)}.
Fig. 7. Examples of segmentation in presence of shadows and highlights
RAD RGB MS Luv RAD RGB MS

20 20 0.8 0.8
std along max and min GCE
std along max and min GCE
18 18 0.7 0.7
16 16
Percent of images
Percent of images
0.6 0.6
14 14
0.5 0.5
12 12
10 10 0.4 0.4
8 8 0.3 0.3
6 6
0.2 0.2
4 4
0.1 0.1
2 2
0 0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
mean GCE mean GCE set of parameters set of parameters
a) b) c) d)
RAD(RGB)−MS (mean GCE index)
1 0.3
RAD(Luv)−MS (mean GCE index)
0.3
0.9
0.2 0.2 0.2
RAD(Luv)−RAD(RGB)
0.8
GCE index (mean)
0.7 0.1 0.1 0.1
0.6
0 0 0
0.5
−0.1 −0.1 −0.1
0.4
0.3 −0.2 −0.2 −0.2

0.2
−0.3 −0.3 −0.3
0.1
0 −0.4 −0.4 −0.4

0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Image Rank Image Rank Image Rank Image Rank
e) f) g) h)
Fig. 8. (a,b)Mean GCE values for each set of parameters. (c,d) Standard deviation of
GCE along maximum and minimum values for each set of parameters. (e)Mean GCE
values for each image sorted form lower to higher. (f)Values higher than zero: images
where MS performs better RAD. (g,h)The same as f) but for MS and RAD Luv and
for RAD RGB versus RAD Luv.
Table 1. Global Constancy Error for several state-of the-art methods: seed [27], fow
[28], MS, and nCuts [29]. Values taken from [27] and [5].
human RAD seed fow MS nCuts

GCE 0.080 0.2048 0.209 0.214 0.2598 0.336
The histogram of the mean GCE values versus the percentage of images for
each GCE value are shown in Figures 8a,b for RAD on RGB and MS respectively.
As more bars are accumulated on the left, the better is the method. Figures 8c,d
show the standard deviation along the maximum and the minimum GCE val-
ues (red lines) for each of the 10 sets of parameters for RAD on RGB and MS.
Note that the behaviour of both methods in this sense is almost the same. A
low and similar standard deviation along all parameters means that the method
has a stable behaviour. Figure, 8e depicts the mean GCE index for each image
ordered by increasing index for MS (green), RAD on RGB (black) and RAD on
Luv (red). This plot shows, not only the good performance of RAD, but that
RAD has a similar behavior on RGB and CIE Luv spaces, even with the afore-
mentioned potential problems on Luv. Figure 8f plots the GCE index differences
for each image between RAD on RGB and MS. Values lower than zero indicate
the number of images where RAD performs better than MS. The same but for
RAD on Luv versus MS, and RAD on RGB versus RAD on Luv is depicted on
Figure 8g,h.
Additionally, table 1 shows GCE values for several state-of-the-art methods.
These values are taken from [27] and [5]. These experiments have been performed
using the train set of 200 images. For both RAD and MS we present the results
obtained with the best parameter settings. For our method the best results
were obtained with (σd ,σi )={(2.5,0.05)}. The mean number of dominant colours
found using RAD had been 5, but it is not directly translated in 5 segments on
segmented images. Often, some segments of few pixels appear due to chromaticity
of surfaces as can bee seen in figure 3h. CGE evaluation favors oversegmentation
[12]. Hence, to make feasible a comparison with other methods using GCE, we
have performed the segmentation without considering segments of an area lower
than 2% of the image area. In this case, the mean number of segments for the 200
test images is 6.98 (7 segments). The number of segments for the other methods
varies from 5 to 12.
As can be seen our method obtains the best results. Furthermore, it should be
noted that the method is substantially faster than the seed and the nCuts [29]
method. In addition, the results obtained with the MS need an additional step.
Namely, a final combination step, which requires a new threshold value, is used
to fuse adjacent segments in the segmented image if their chromatic difference
is lower than the threshold (without pre- an postprocessing MS obtains a score
of 0.2972). For our RAD method we do not apply any pre- or postprocessing
steps.
5 Conclusions
This paper introduces a new feature space segmentation method that extracts the
Ridges formed by a dominant colour in an image histogram. This method is robust
against discontinuities appearing in image histograms due to compression and ac-
quisition conditions. Furthermore, those strong discontinuities, related with the
physical illumination effects are correctly treated due to the topological treatment
of the histogram. As a consequence, the presented method yields better results
than Mean shift on a widely used image dataset and error measure. Additionally,
even with neither preprocessing nor postprocessing steps, RAD has a better per-
formance than the state-of-the-art methods. It points out that the chromatic in-
formation is an important cue on human segmentation. Additionally, the elapsed
time for RAD is not affected by its parameters. Due to that it becomes a faster
method than Mean Shift and the other state-of-the-art methods.
The next step is to add spatial coherence to help the method in those areas
which are not well-represented by a dominant colour. Furthermore, improvement
is expected by looking for dominant colours only in interesting regions in the
image instead of in the whole image at once.
Acknowledgements
This work has been partially supported by projects TIN2004-02970, TIN2007-
64577 and Consolider-Ingenio 2010 CSD2007-00018 of Spanish MEC (Ministery
of Science) and the Ramon y Cajal Program.
References
1. Skarbek, W., Koschan, A.: Colour image segmentation — a survey. Technical re-
port, Institute for Technical Informatics, Technical University of Berlin (October
1994)
2. Cheng, H., Jiang, X., Sun, Y., Wang, J.: Color image segmentation:advances and
prospects. Pattern Recognition 34(6), 2259–2281 (2001)
3. Lucchese, L., Mitra, S.: Color image segmentation: A state-of-the-art survey. INSA-
A: Proceedings of the Indian National Science Academy, 207–221 (2001)
4. Agarwal, S., Madasu, S., Hanmandlu, M., Vasikarla, S.: A comparison of some
clustering techniques via color segmentation. In: ITCC 2005: Proceedings of the In-
ternational Conference on Information Technology: Coding and Computing (ITCC
2005), vol. II, pp. 147–153. IEEE Computer Society Press, Washington (2005)
5. Yang, Y., Wright, J., Sastry, S., Ma, Y.: Unsupervised segmentation of natural
images via lossy data compression (2007)
6. Freixenet, J., Munoz, X., Raba, D., Mart, J., Cuf, X.: Yet another survey on image
segmentation: Region and boundary information integration. In: Heyden, A., Sparr,
G., Nielsen, M., Johansen, P. (eds.) ECCV 2002. LNCS, vol. 2352, pp. 408–422.
Springer, Heidelberg (2002)
7. Sezgin, M., Sankur, B.: Survey over image thresholding techniques and quantitative
performance evaluation. J. Electron. Imaging 13(1), 146–165 (2004)
8. Fukunaga, K., Hostetler, L.D.: The estimation of the gradient of a density func-
tion, with applications in pattern recognition. IEEE Transactions on Information
Theory 121(1), 32–40 (1975)
9. Verma, D., Meila, M.: A comparison of spectral clustering algorithms. technical
report uw-cse-03-05-01, university of washington
10. Abd-Almageed, W., Davis, L.: Density Estimation Using Mixtures of Mixtures of
Gaussians. In: 9th European Conference on Computer Vision (2006)
11. Bilmes, J.: A Gentle Tutorial of the EM Algorithm and its Application to Param-
eter Estimation for Gaussian Mixture and Hidden Markov Models. International
Computer Science Institute 4 (1998)
12. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A Database of Human Segmented
Natural Images and its Application to Evaluating Segmentation Algorithms and
Measuring Ecological Statistics. In: Proc. Eighth Int’l Conf. Computer Vision,
vol. 2, pp. 416–423 (2001)
13. Shafer, S.A.: Using color to seperate reflection components. COLOR research and
application 10(4), 210–218 (1985)
14. Geusebroek, J.M., Burghouts, G.J., Smeulders, A.W.M.: The Amsterdam library
of object images. Int. J. Comput. Vision 61(1), 103–112 (2005)
15. Klinker, G., Shafer, S.: A physical approach to color image understanding. Int.
Journal of Computer Vision 4, 7–38 (1990)
16. López, A.M., Lumbreras, F., Serrat, J., Villanueva, J.J.: Evaluation of methods for
ridge and valley detection. IEEE Trans. Pattern Anal. Mach. Intell. 21(4), 327–335
(1999)
17. Wang, L., Pavlidis, T.: Direct gray-scale extraction of features for character recog-
nition. IEEE Trans. Pattern Anal. Mach. Intell. 15(10), 1053–1067 (1993)
18. Bishnu, A., Bhowmick, P., Dey, S., Bhattacharya, B.B., Kundu, M.K., Murthy,
C.A., Acharya, T.: Combinatorial classification of pixels for ridge extraction in a
gray-scale fingerprint image. In: ICVGIP (2002)
19. Vazquez, E., Baldrich, R., Vazquez, J., Vanrell, M.: Topological histogram reduc-
tion towards colour segmentation. In: Martı́, J., Benedı́, J.M., Mendonça, A.M.,
Serrat, J. (eds.) IbPRIA 2007. LNCS, vol. 4477, pp. 55–62. Springer, Heidelberg
(2007)
20. Vincent, L., Soille, P.: Watersheds in digital spaces: an efficient algorithm based
on immersion simulations. IEEE Transactions on Pattern Analysis and Machine
Intelligence 13(6), 583–598 (1991)
21. Gauch, J.M., Pizer, S.M.: Multiresolution analysis of ridges and valleys in grey-
scale images. IEEE Trans. Pattern Anal. Mach. Intell. 15(6), 635–646 (1993)
22. Comaniciu, D., Meer, P.: Mean shift: A robust approach toward feature space
analysis. IEEE Trans. Pattern Anal. Mach. Intell. 24(5), 603–619 (2002)
23. Christoudias, C., Georgescu, B., Meer, P.: Synergism in low level vision. Interna-
tional Conference on Pattern Recognition 4, 150–155 (2002)
24. Pantofaru, C., Hebert, M.: A comparison of image segmentation algorithms. Tech-
nical Report CMU-RI-TR-05-40, Robotics Institute, Carnegie Mellon University,
Pittsburgh, PA (September 2005)
25. Ge, F., Wang, S., Liu, T.: New benchmark for image segmentation evaluation.
Journal of Electronic Imaging 16, 033011 (2007)
26. Felzenszwalb, P., Huttenlocher, D.: Efficient graph-based image segmentation. Intl.
Journal of Computer Vision 59(2) (2004)
27. Micusık, B., Hanbury, A.: Automatic image segmentation by positioning a seed. In:
Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3952. Springer,
Heidelberg (2006)
28. Fowlkes, C., Martin, D., Malik, J.: Learning affinity functions for image segmen-
tation: combining patch-based and gradient-based approaches. In: Proceedings of
IEEE Computer Society Conference on Computer Vision and Pattern Recognition,
2003, vol. 2 (2003)
29. Shi, J., Malik, J.: Normalized cuts and image segmentation. IEEE Transactions on
Pattern Analysis and Machine Intelligence 22(8), 888–905 (2000)
Image Segmentation by Branch-and-Mincut
Victor Lempitsky, Andrew Blake, and Carsten Rother
Microsoft Research Cambridge
Abstract. Efficient global optimization techniques such as graph cut

exist for energies corresponding to binary image segmentation from low-
level cues. However, introducing a high-level prior such as a shape prior or
a color-distribution prior into the segmentation process typically results
in an energy that is much harder to optimize. The main contribution
of the paper is a new global optimization framework for a wide class
of such energies. The framework is built upon two powerful techniques:
graph cut and branch-and-bound. These techniques are unified through
the derivation of lower bounds on the energies. Being computable via
graph cut, these bounds are used to prune branches within a branch-
and-bound search.
We demonstrate that the new framework can compute globally opti-
mal segmentations for a variety of segmentation scenarios in a reasonable
time on a modern CPU. These scenarios include unsupervised segmen-
tation of an object undergoing 3D pose change, category-specific shape
segmentation, and the segmentation under intensity/color priors defined
by Chan-Vese and GrabCut functionals.
1 Introduction
Binary image segmentation is often posed as a graph partition problem. This is
because efficient graph algorithms such as mincut permit fast global optimiza-
tion of the functionals measuring the quality of the segmentation. As a result,
difficult image segmentation problems can be solved efficiently, robustly, and
independently of initialization. Yet, while graphs can represent energies based
on localized low-level cues, they are much less suitable for representing non-local
cues and priors describing the foreground or the background segment as a whole.
Consider, for example, the situation when the shape of the foreground segment
is known a priori to be similar to a particular template (segmentation with shape
priors). Graph methods can incorporate such a prior for a single pre-defined and
pre-located shape template[13,20]. However, once the pose of the template is al-
lowed to change, the relative position of each graph edge with respect to the tem-
plate becomes unknown, and the non-local property of shape similarity becomes
hard to express with local edge weights. Another example would be the segmen-
tation with non-local color priors, when the color of the foreground and/or back-
ground is known a priori to be described by some parametric distribution (e.g.
a mixture of the Gaussians as in the case of GrabCut [25]). If the parameters of
these distributions are allowed to change, such a non-local prior depending on the
segment as a whole becomes very hard to express with the local edge weights.

16 V. Lempitsky, A. Blake, and C. Rother
An easy way to circumvent the aforementioned difficulties is to alternate the

graph partitioning with the reestimation of non-local parameters (such as the
template pose or the color distribution). A number of approaches [6,16,25,15]
follow this path. Despite the use of the global graph cut optimization inside
the loop, local search over the prior parameters turns these approaches into
local optimization techniques akin to variational segmentation [7,9,24,27]. As a
result, these approaches may get stuck in local optima, which in many cases
correspond to poor solutions.
The goal of this paper is to introduce a new framework for computing glob-
ally optimal segmentations under non-local priors. Such priors are expressed by
replacing fixed-value edge weights with edge weights depending on non-local pa-
rameters. The global minimum of the resulting energy that depends on both the
graph partition and the non-local parameters is then found using the branch-
and-bound tree search. Within the branch-and-bound, lower bounds over tree
branches are efficiently evaluated by computing minimal cuts on a graph (hence
the name Branch-and-Mincut).
The main advantage of the proposed framework is that the globally opti-
mal segmentation can be obtained for a broad family of functionals depending
on non-local parameters. Although the worst case complexity of our method is
large (essentially, the same as the exhaustive search over the space of non-local
parameters), we demonstrate that our framework can obtain globally optimal
image segmentation in a matter of seconds on a modern CPU. Test scenarios in-
clude globally optimal segmentation with shape priors where the template shape
is allowed to deform and to appear in various poses as well as image segmenta-
tion by the optimization of the Chan-Vese [7] and the GrabCut [25] functionals.
In all cases, bringing in high-level non-local knowledge allows to solve difficult
segmentation problems, where local cues (considered by most current global op-
timization approaches) were highly ambiguous.
2 Related Work
Our framework employs the fact that a submodular quadratic function of boolean
variables can be efficiently minimized via minimum cut computation in the as-
sociated graph [2,11,18]. This idea has been successfully applied to binary image
segmentation [3] and quickly gained popularity. As discussed above, the approach
[3] still has significant limitations, as the high-level knowledge such as shape or
color priors are hard to express with fixed local edge weights. These limitations
are overcome in our framework, which allows the edge weights to vary.
In the restricted case, when unary energy potentials are allowed to vary and
depend on a single scalar non-local parameter monotonically, efficient algorithms
known as parametric maxflow have been suggested (see e.g. [19]). Our framework
is however much more general then these methods (at a price of having higher
worst-case complexity), as we allow both unary and pairwise energy terms to
depend non-monotonically on a single or multiple non-local parameters. Such
generality gives our framework flexibility in incorporating various high-level pri-
ors while retaining the globality of the optimization.
Image Segmentation by Branch-and-Mincut 17
Image segmentation with non-local shape and color priors has attracted a
lot of interest in the last years. As discussed above, most approaches use either
local continuous optimization [27,7,24,9] or iterated minimization alternating
graph cut and search over non-local parameter space [25,6,16]. Unfortunately,
both groups of methods are prone to getting stuck in poor local minima. Global-
optimization algorithms have also been suggested [12,26] . In particular, simulta-
neous work [10] presented a framework that also utilizes branch-and-bound ideas
(paired with continuous optimization in their case). While all these global opti-
mization methods are based on elegant ideas, the variety of shapes, invariances,
and cues that each of them can handle is limited compared to our method.
Finally, our framework may be related to branch-and-bound search methods
in computer vision (e.g. [1,21]). In particular, it should be noted that the way our
framework handles shape priors is related to previous approaches like [14] that
used tree search over shape hierarchies. However, neither of those approaches
accomplish pixel-wise image segmentation.
3 Optimization Framework
In this section, we discuss our global energy optimization framework for obtain-
ing image segmentations under non-local priors1 . In the next sections, we detail
how it can be used for the segmentation with non-local shape priors (Section 4)
and non-local intensity/color priors (Section 5).
3.1 Energy Formulation
Firstly, we introduce notation and give the general form of the energy that can
be optimized in our framework. Below, we consider the pixel-wise segmentation
of the image. We denote the pixel set as V and use letters p and q to denote
individual pixels. We also denote the set of edges connecting adjacent pixels
as E and refer to individual edges as to the pairs of pixels (e.g. p, q). In our
experiments, the set of edges consisted of all 8-connected pixel pairs in the raster.
The segmentation of the image is given by its 0−1 labeling x ∈ 2V , where
individual pixel labels xp take the values 1 for the pixels classified as the fore-
ground and 0 for the pixels classified as the background. Finally, we denote the
non-local parameter as ω and allow it to vary over a discrete, possibly very large,
set Ω. The general form of the energy function that can be handled within our
framework is then given by:

E(x, ω) = C(ω)+ F p (ω)·xp + B p (ω)·(1−xp )+ P pq (ω)·|xp −xq | . (1)
p∈V p∈V p,q∈E
Here, C(ω) is a constant potential, which does not depend directly on the seg-
mentation x; F p (ω) and B p (ω) are the unary potentials defining the cost for
1
The C++ code for this framework is available at the webpage of the first author.
assigning the pixel p to the foreground and to the background respectively;

P pq (ω) is the pairwise potential defining the cost of assigning adjacent pixels
p and q to different segments. In our experiments, the pairwise potentials were
taken non-negative to ensure the tractability of E(x, ω) as the function of x for
graph cut optimization [18].
All potentials in our framework depend on the non-local parameter ω ∈ Ω. In
general, we assume that Ω is a discrete set, which may be large (e.g. millions of
elements) and should have some structure (although, it need not be linearly or
partially ordered). For the segmentation with shape priors, Ω will correspond to
the product space of various poses and deformations of the template, while for
the segmentation with color priors Ω will correspond to the set of parametric
color distributions.
3.2 Lower Bound

Our approach optimizes the energy (1) exactly, finding its global minimum using
branch-and-bound tree search [8], which utilizes the lower bound on (1) derived
as follows:
⎡

min E(x, ω) = min min ⎣C(ω) + F p (ω)·xp + B p (ω)·(1 − xp )+
x∈2V ,ω∈Ω x∈2V ω∈Ω
p∈V p∈V
⎤ ⎡

P pq (ω)·|xp − xq |⎦ ≥ min ⎣min C(ω) + min F p (ω)·xp +
x∈2V ω∈Ω ω∈Ω
p,q∈E p∈V
⎤

min B p (ω)·(1 − xp ) + min P pq (ω)·|xp − xq |⎦ =
ω∈Ω ω∈Ω
p∈V p,q∈E
⎡ ⎤

min ⎣CΩ + FΩp ·xp + p
BΩ ·(1 − xp ) + PΩpq ·|xp − xq |⎦ = L(Ω) . (2)
x∈2V
p∈V p∈V p,q∈E
Here, CΩ , FΩp , BΩ
p
, PΩpq denote the minima of C(ω), F p (ω), B p (ω), P pq (ω) over
ω ∈ Ω referred below as aggregated potentials. L(Ω) denotes the derived lower
bound for E(x, ω) over 2V ⊗ Ω. The inequality in (2) is essentially the Jensen
inequality for the minimum operation.
The proposed lower bound possesses three properties crucial to the Branch-
and-Mincut framework:
Monotonicity. For the nested domains of non-local parameters Ω1 ⊂ Ω2 the

inequality L(Ω1 ) ≥ L(Ω2 ) holds (the proof is given in [23]).
Computability. The key property of the derived

lower bound is the ease of its evaluation. Indeed,
this bound equals the minimum of a submodular
quadratic pseudo-boolean function. Such function
can be realized on a network graph such that each
configuration of the binary variables is in one-to-one
correspondence with an st-cut of the graph having
the weight equal to the value of the function (plus The fragment of the
a constant CΩ ) [2,11,18]. The minimal st-cut corre- network graph realiz-
sponding to the minimum of L(Ω) then can be coming L(Ω) (edge weights
puted in a low-polynomial of |V| time e.g. with the shown in boxes). (see
popular algorithm [5]. e.g.[18] for details)
Tightness. For a singleton Ω the bound is tight: L({ω}) = minx∈2V E(x, ω). In
such case, the minimal st-cut also yields the segmentation x optimal for this ω
(xp = 0 iff the respective vertex belongs to the s-component of the cut).
Note, that the fact that the lower bound (2) may be evaluated via st-mincut
gives rise to a whole family of looser, but cheaper, lower bounds. Indeed, the
minimal cut on a network graph is often found by pushing flows until the flow
becomes maximal (and equal to the weight of the mincut) [5]. Thus, the se-
quence of intermediate flows provides a sequence of the increasing lower bounds
on (1) converging to the bound (2) (flow bounds). If some upper bound on
the minimum value is imposed, the process may be terminated earlier without
computing the full maxflow/mincut. This happens when the new flow bound
exceeds the given upper bound. In this case it may be concluded that the value
of the global minimum is greater than the imposed upper bound.
3.3 Branch-and-Bound Optimization
Finding the global minimum of (1) is, in general, a very difficult problem. Indeed,
since the potentials can depend arbitrarily on the non-local parameter spanning
arbitrary discrete set Ω, in the worst-case any optimization has to search ex-
haustively over Ω. In practice, however, any segmentation problem has some
specifically-structured space Ω. This structure can be efficiently exploited by
the branch-and-bound search detailed below.
We assume that the discrete domain Ω can be hierarchically clustered and
the binary tree of its subregions TΩ = {Ω = Ω0 , Ω1 , . . . ΩN } can be constructed
(binarity of the tree is not essential). Each non-leaf node corresponding to the
subregion Ωk then has two children corresponding to the subregions Ωch1(k) and
Ωch2(k) such that Ωch1(k) ⊂ Ωk , Ωch2(k) ⊂ Ωk . Here, ch1(·) and ch2(·) map the
index of the node to the indices of its children. Also, leaf nodes of the tree are
in one-to-one correspondence with singleton subsets Ωl = {ωt }.
Given such tree, the global minimum of (1) can be efficiently found using
the best-first branch-and-bound search [8]. This algorithm propagates a front of
nodes in the top-down direction (Fig. 1). During the search, the front contains a
set of tree nodes, such that each top-down path from the root to a leaf contains
Fig. 1. Best-first branch-and-bound optimization on the tree of nested regions finds

the globally-optimal ω by the top-down propagation of the active front (see text for
details). At the moment when the lowest lower bound of the front is observed at leaf
node, the process terminates with the global minimum found without traversing the
whole tree.
exactly one active vertex. In the beginning, the front contains the tree root
Ω0 . At each step the active node with the smallest lower bound (2) is removed
from the active front, while two of its children are added to the active front (by
monotonicity property they have higher or equal lower bounds). Thus, an active
front moves towards the leaves making local steps that increase the lowest lower
bound of all active nodes. Note, that at each moment, this lowest lower bound
of the front constitutes a lower bound on the global optimum of (1) over the
whole domain.
At some moment of time, the active node with the smallest lower bound
turns out to be a leaf {ω }. Let x be the optimal segmentation for ω (found via
minimum st-cut). Then, E(x , ω ) = L(ω ) (tightness property) is by assumption
the lowest bound of the front and hence a lower bound on the global optimum
over the whole domain. Consequently, (x , ω ) is a global minimum of (1) and
the search terminates without traversing the whole tree. In our experiments,
the number of the traversed nodes was typically very small (two-three orders
of magnitude smaller then the size of the full tree). Therefore, the algorithm
performed global optimization much faster than exhaustive search over Ω.
In order to further accelerate the search, we exploit the coherency between
the mincut problems solved at different nodes. Indeed, the maximum flow as
well as auxiliary structures such as shortest path trees computed for one graph
may be “reused” in order to accelerate the computation of the minimal st-cut
on another similar graph [3,17]. For some applications, this trick may give an
order of magnitude speed-up for the evaluation of lower bounds.
In addition to the best-first branch-and-bound search we also tried the depth-
first branch-and-bound [8]. When problem-specific heuristics are available that
give good initial solutions, this variant may lead to moderate (up to a factor
of 2) time savings. Interestingly, the depth-first variant of the search, which
maintains upper bounds on the global optimum, may benefit significantly from
the use of flow bounds discussed above. Nevertheless, we stick with the best-first
branch-and-bound for the final experiments due to its generality (no need for
initialization heuristics).
In the rest of the paper we detail how the general framework developed above
may be used within different segmentation scenarios.
4 Segmentation with Shape Priors

4.1 Constructing Shape Prior
We start with the segmentation with shape priors. The success of such segmen-
tation crucially depends on the way shape prior is defined. Earlier works have
often defined this prior as a Gaussian distribution of some geometrical shape
statistics (e.g. control point positions or level set functions) [27,24]. In reality,
however, pose variance and deformations specific to the object of interest lead
to highly non-Gaussian, multi-modal prior distributions. For better modeling of
prior distributions, [9] suggested the use of non-parametric kernel densities. Our
approach to shape modeling is similar in spirit, as it also uses exemplar-based
prior. Arguably, it is more direct, since it involves the distances between the
binary segmentations themselves, rather than their level set functions.
We assume that the prior is defined by the set of exemplar binary segmenta-
tions {yω |ω ∈ Ω}, where Ω is a discrete set indexing the exemplar segmentations.
Then the following term introduces a joint prior over the segmentation and the
non-local parameter into the segmentation process:

Eprior (x, ω) = ρ(x, yω ) = (1 − ypω )·xp + ypω ·(1 − xp ) , (3)
p∈V p∈V
where ρ denotes the Hamming distance between segmentations. This term clearly
has the form (1) and therefore its combinations with other terms of this form can
be optimized within our framework. Being optimized over the domain 2V ⊗Ω, this
term would encourage the segmentation x to be close in the Hamming distance
to some of the exemplar shapes (note that some other shape distances can be
used in a similar way).
The full segmentation energy then may be defined by adding a standard
contrast-sensitive edge term [3]:
||Kp −Kq ||
e− σ
Eshape (x, ω) = Eprior (x, ω) + λ ·|xp − xq | , (4)
|p − q|
p,q∈E
where ||Kp − Kq || denote the SAD (L1) distance between RGB colors of the
pixels p and q in the image (λ and σ were fixed throughout the experiments
described in this section), |p − q| denotes
√ the distance between the centers of the
pixels p and q (being either 1 or 2 for the 8-connected grid). The functional
(4) thus incorporates the shape prior with edge-contrast cues.
In practice, the set Ωshape could be huge, e.g. tens of millions exemplars,
which poses a problem for hierarchical clustering as well as pre-computing and
storing aggregated potentials. Fortunately, for many scenarios all these tasks
can be accomplished with reasonable amount of time and memory provided that
Exemplars yω Non-local shape prior+Edge cues Intensity+Edge cues
Fig. 2. Using the shape prior constructed from the set of exemplars (left column) our
approach can accomplish segmentation of an object undergoing general 3D pose changes
within two differently illuminated sequences (two middle columns). Note the varying
topology of the segmentations. For comparison, we give the results of a standard graph
cut segmentation (right column): even with parameters tuned specifically to the test
images, separation is entirely inaccurate.
translational invariance is exploited, so that subsets of shapes yω , which differ

by translation are handled in a similar way. We give the details in [23].
Note the three properties of our approach to segmentation with shape priors.
Firstly, since any shapes can be included in Ωshape , general 3D pose transfor-
mations and deformations may be handled. Secondly, the segmentations may
have general varying topology not restricted to segments with single-connected
boundaries. Thirdly, our framework is general enough to introduce other terms
in the segmentation process (e.g. regional terms used in a standard graph cut
segmentation [3]). These properties of our approach are demonstrated within the
following experiments.
4.2 Experiments
Single object+3D pose changes. In our first experiment, we constructed
a shape prior for a single object (a coffee cup) undergoing 3D pose changes.
We obtained a set of outlines using “blue-screening”. We then normalized these
outlines (by centering at the origin, resizing to a unit scale and orienting the
principle axes with the coordinate axes). After that we clustered the normalized
outlines using k-means. A representative of each cluster was then taken into the
exemplar set. After that we added scale variations, in-plane rotations, and trans-
lations. As a result, we got a set {yω |ω ∈ Ωshape } containing about 30,000,000
exemplar shapes.
The results of the global optimization of the functional (4) for the frames from
the two sequences containing clutter and camouflage are shown in Fig. 2. On
Fig. 3. Results of the global optimization of (5) on some of the 170 UIUC car images
including 1 of the 2 cases where localization failed (bottom left). In the case of the
bottom right image, the global minimum of (4) (yellow) and the result of our feature-
based car detector (blue) gave erroneous localization, while the global minimum of
their combination (5) (red) represented an accurate segmentation.
average, we observed that segmenting 312x272 image took about 30 seconds of

an Intel-2.40 GHz CPU and less than 1 Gb of RAM. The proportion of the nodes
of the tree traversed by the active front was on average about 1 : 5000. Thus,
branch-and-bound tree search used in our framework improved very considerably
over exhaustive search, which would have to traverse all leaves (1 : 2 of the tree).
As a baseline algorithm, we considered the segmentation with a “standard”
graph cut functional,
replacing non-local shape prior term with a local intensity-
based term p∈V (I − Ip )·xp , adjusting the constant I for each frame so that
it gives the best results. However, since the intensity distributions of the cup
and the backgrounds overlapped significantly, the segmentations were grossly
erroneous (Fig. 2 – right column).
Object class+translation invariance. In the second experiment, we per-

formed the segmentation with shape priors on UIUC car dataset (the version
without scale variations), containing 170 images with cars in uncontrolled en-
vironment (city streets). The shape prior was built by manually segmenting 60
translation-normalized training images and then adding translations (dependent
on the extent of each test image). While the test image sizes varyied from 110x75
to 360x176, the size of Ωshape varied from 18,666 to 2,132,865. We computed the
globally optimal segmentations under the constructed prior using the energy (4).
Using the bounding boxes of the cars provided with the dataset, we found
that in 6.5% of the images the global minima corresponded to clutter rather
than cars. To provide a baseline for localization accuracy based on edge cues and
a set of shape templates, we considered Chamfer matching (as e.g. in [14]). For
the comparison we used the same set of templates, which were matched against
truncated Canny-based chamfer distance (with optimally tuned truncation and
Canny sensitivity parameters). In this way, the optimal localization failed (i.e.
corresponded to clutter rather than a car) in 12.4% of the images.
Clearly, segmenting images using (4) takes into account the shape prior and
edge-contrast cues, but ignores the appearance typical for the object category
under consideration. At the same time, there exists a large number of algorithms
working with image appearance cues and performing object detection based
on these cues (see e.g. [22] and references therein). Typically, such algorithms
produce the likelihood of the object presence either as a a function of a bounding
box or even in the form of per-pixel “soft segmentation” masks. Both types of
the outputs can be added into the functional (1) either via constant potential
C(Ω) or via unary potentials. In this way, such appearance-based detectors can
be integrated with shape prior and edge-contrast cues.
As an example of such integration, we devised a simple detector similar in
spirit to [22]. The detector looked for the appearance features typical for cars
(wheels) using normalized cross-correlation. Each pixel in the image then “voted”
for the location of the car center depending on the strength of the response to
the detector and the relative position of the wheels with respect to the car center
observed on the training dataset. We then added an additional term Cvote (ω) in
our energy (1) that for each ω equaled minus the accumulated strength of the
votes for the center of yω :
||Kp −Kq ||
e− σ
Eshape&detect(x, ω) = Cvote (ω) + Eprior (x, ω) + λ ·|xp − xq | , (5)
|p − q|
p,q∈E
Adding the appearance-based term improved the robustness of the segmen-

tation, as the global optima of (5) corresponded to clutter only in 1.2% of the
images. The global minima found for some of the images are shown in Fig. 3.
Note, that for our simple detector on its own the most probable bounding box
corresponded to clutter on as much as 14.7% of the images.
In terms of the performance, on average, for the functional (5) the segmenta-
tion took 1.8 seconds and the proportion of the tree traversed by the active front
was 1 : 441. For the functional (4), the segmentation took 6.6 seconds and the
proportion of the tree traversed by the active front was 1 : 131. This difference
in performance is natural to branch-and-bound methods: the more difficult and
ambiguous is the optimization problem, the larger is the portion of the tree that
has to be investigated.
5 Segmentation with Color/Intensity Priors

Our framework can also be used to impose non-local priors on the intensity or
color distributions of the foreground and background segments, as the examples
below demonstrate.
5.1 Segmenting Grayscale Images: Chan-Vese Functional

In [7] Chan and Vese have proposed the following popular functional for the
variational image segmentation problem:

2
2
E(S, c , c ) = μ
f b
dl + ν dp + λ1 I(p) − c f
dp + λ2 I(p) − cb dp ,
∂S S S S̄
(6)
where S denotes the foreground segment, and I(p) is a grayscale image. The
first two terms measure the length of the boundary and the area, the third and
the forth terms are the integrals over the fore- and background of the difference
between image intensity and the two intensity values cf and cb , which correspond
to the average intensities of the respective regions. Traditionally, this functional
is optimized using level set framework converging to one of its local minima.
Below, we show that the discretized version of this functional can be optimized
globally within our framework. Indeed, the discrete version of (6) can be written
as (using notation as before):
μ
E(x, (cf , cb )) = ·|xp − xq |+
|p − q|
p,q∈E

2 (7)
ν + λ1 (I(p) − cf )2 ·xp + λ2 I(p) − cb ·(1 − xp ) .
p∈V p∈V
Here, the first term approximates the first term of (6) (the accuracy of the
approximation depends on the size of the pixel neighborhood [4]), and the last
two terms express the last three terms of (6) in a discrete setting.
The functional (7) clearly has the form (1) with non-local
parameter ω = {cf , cb }. Discretizing intensities cf and cb into
255 levels and building a quad-tree over their joint domain,
we can apply our framework to find the global minima of
(6). An example of a global minimum of (7) is shown to the
right (this 183x162 image was segmented in 3 seconds, the
proportion of the tree traversed was 1:115). More examples
are given in [23].
5.2 Segmenting Color Images: GrabCut functional
In [25], the GrabCut framework for the interactive color image segmentation
based on Gaussian mixtures was proposed. In GrabCut, the segmentation is
driven by the following energy:

EGrabCut (x, (GM f , GM b )) = − log(P(Kp | GM f ))·xp +
p∈V
||K −Kq ||2 (8)
λ1 + λ2 ·e− p β
+ − log(P(Kp | GM ))·(1 − xp )+
b
·|xp − xq | .
|p − q|
p∈V p,q∈E
Here, GM f and GM b are Gaussian mixtures in RGB color space and the first
two terms of the energy measure how well these mixtures explain colors Kp of
pixels attributed to fore- and background respectively. The third term is the con-
trast sensitive edge term, ensuring that the segmentation boundary is compact
and tends to stick to color region boundaries in the image. In addition to this
energy, the user provides supervision in the form of a bounding rectangle and
brush strokes, specifying which parts of the image should be attributed to the
foreground and to the background.
The original method [25] minimizes the energy within EM-style process, al-
ternating between (i) the minimization of (8) over x given GM f and GM b and
(ii) refitting the mixtures GM f and GM b given x. Despite the use of the global
graph cut optimization within the segmentation update step, the whole process
yields only a local minimum of (8). In [25], the segmentation is initialized to the
provided bounding box and then typically shrinks to one of the local minima.
The energy (8) has the form (1) and therefore can be optimized within Branch-
and-Mincut framework, provided that the space of non-local parameters (which
in this case is the joint space of the Gaussian mixtures for the foreground and
for the background) is discretized and the tree of the subregions is built. In this
scenario, however, the dense discretization of the non-local parameter space is
infeasible (if the mixtures contain n Gaussians then the space is described by
20n − 2 continuous parameters). It is possible, nevertheless, to choose a much
smaller discrete subset Ω that is still likely to contain a good approximation to
the globally-optimal mixtures.
To construct such Ω, we fit a mixture of M = 8 Gaussians G1 , G2 , ...GM with
the support areas a1 , a2 , ...aM to the whole image. The support area ai here
counts the number of pixels p such as ∀j P(Kp |Gi ) ≥ P(Kp |Gj ). We assume
that the components are ordered such that the support areas decrease (ai >
ai+1 ). Then, the Gaussian mixtures we consider are defined by the binary vector
β = {β1 , β2 . . . βM } ∈ {0, 1}M specifying
which Gaussians
should be included
into the mixture: P(K| GM (β)) = i βi ai P(K|Gi ) / i βi ai .
The overall set Ω is then defined as {0, 1}2M , where odd bits correspond to
the foreground mixture vector β f and even bits correspond to the background
mixture vector β b . Vectors with all even bits and/or all odd bits equal to zero
do not correspond to meaningful mixtures and are therefore assigned an infinite
cost. The hierarchy tree is naturally defined by the bit-ordering (the first bit
corresponding to subdivision into the first two branches etc.).
Depending on the image and the value of M , the solutions found by Branch-
and-Mincut framework may have larger or smaller energy (8) than the solutions
found by the original EM-style method [25]. This is because Branch-and-Mincut
here finds the global optimum over the subset of the domain of (8) while [25]
searches locally but within the continuous domain. However, for all 15 images in
our experiments, improving Branch-and-Mincut solutions with a few EM-style
iterations [25] gave lower energy than the original solution of [25]. In most cases,
these additional iterations simply refit the Gaussians properly and change very
few pixels near boundary (see Fig. 4).
In terms of performance, for M = 8 the segmentation takes on average a few
dozen seconds (10s and 40s for the images in Fig. 4) for 300x225 image. The
proportion of the tree traversed by an active front is one to several hundred
(1:963 and 1:283 for the images in Fig. 4).
Image+input GrabCut[25](−618) Branch&Mincut(−624) Combined(−628)
Image+input GrabCut[25](−593) Branch&Mincut(−584) Combined(−607)
Fig. 4. Being initialized with the user-provided bounding rectangle (shown in green
in the first column) as suggested in [25], EM-style process [25] converges to a local
minimum (the second column). Branch-and-Mincut result (the third column) escapes
that local minimum and after EM-style improvement lead to the solution with much
smaller energy and better segmentation accuracy (the forth column). Energy values
are shown in brackets.
This experiment suggests the usefulness of Branch-and-Mincut framework as

a mean of obtaining good initial point for local methods, when the domain space
is too large for an exact branch-and-bound search.
6 Conclusion
The Branch-and-Mincut framework presented in this paper finds global optima
of a wide class of energies dependent on the image segmentation mask and non-
local parameters. The joint use of branch-and-bound and graph cut allows effi-
cient traversal of the solution space. The developed framework is useful within a
variety of image segmentation scenarios, including segmentation with non-local
shape priors and non-local color/intensity priors.
Future work includes the extension of Branch-and-Mincut to other problems,
such as simultaneous stitching and registration of images, as well as deriving
analogous branch-and-bound frameworks for combinatorial methods other than
binary graph cut, such as minimum ratio cycles and multilabel MRF inference.
Acknowledgements
We would like to acknowledge discussions and feedback from Vladimir Kol-
mogorov and Pushmeet Kohli. Vladimir has also kindly made several modifi-
cations of his code of [5] that allowed to reuse network flows more efficiently.
References
1. Agarwal, S., Chandaker, M., Kahl, F., Kriegman, D., Belongie, S.: Practical Global
Optimization for Multiview Geometry. In: Leonardis, A., Bischof, H., Pinz, A.
(eds.) ECCV 2006. LNCS, vol. 3951. Springer, Heidelberg (2006)
2. Boros, E., Hammer, P.: Pseudo-boolean optimization. Discrete Applied Mathemat-
ics 123(1-3) (2002)
3. Boykov, Y., Jolly, M.-P.: Interactive Graph Cuts for Optimal Boundary and Region
Segmentation of Objects in N-D Images. In: ICCV 2001 (2001)
4. Boykov, Y., Kolmogorov, V.: Computing Geodesics and Minimal Surfaces via
Graph Cuts. In: ICCV 2003 (2003)
5. Boykov, Y., Kolmogorov, V.: An Experimental Comparison of Min-Cut/Max-Flow
Algorithms for Energy Minimization in Vision. PAMI 26(9) (2004)
6. Bray, M., Kohli, P., Torr, P.: PoseCut: Simultaneous Segmentation and 3D Pose
Estimation of Humans Using Dynamic Graph-Cuts. In: Leonardis, A., Bischof, H.,
Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3952. Springer, Heidelberg (2006)
7. Chan, T., Vese, L.: Active contours without edges. Trans. Image Process 10(2)
(2001)
8. Clausen, J.: Branch and Bound Algorithms - Principles and Examples. Parallel
Computing in Optimization (1997)
9. Cremers, D., Osher, S., Soatto, S.: Kernel Density Estimation and Intrinsic Align-
ment for Shape Priors in Level Set Segmentation. IJCV 69(3) (2006)
10. Cremers, D., Schmidt, F., Barthel, F.: Shape Priors in Variational Image Segmen-
tation: Convexity, Lipschitz Continuity and Globally Optimal Solutions. In: CVPR
2008 (2008)
11. Greig, D., Porteous, B., Seheult, A.: Exact maximum a posteriori estimation for
binary images. Journal of the Royal Statistical Society 51(2) (1989)
12. Felzenszwalb, P.: Representation and Detection of Deformable Shapes. PAMI 27(2)
(2005)
13. Freedman, D., Zhang, T.: Interactive Graph Cut Based Segmentation with Shape
Priors. In: CVPR 2005 (2005)
14. Gavrila, D., Philomin, V.: Real-Time Object Detection for ”Smart” Vehicles. In:
ICCV 1999 (1999)
15. Huang, R., Pavlovic, V., Metaxas, D.: A graphical model framework for coupling
MRFs and deformable models. In: CVPR 2004 (2004)
16. Kim, J., Zabih, R.: A Segmentation Algorithm for Contrast-Enhanced Images.
ICCV 2003 (2003)
17. Kohli, P., Torr, P.: Effciently Solving Dynamic Markov Random Fields Using Graph
Cuts. In: ICCV 2005 (2005)
18. Kolmogorov, V., Zabih, R.: What Energy Functions Can Be Minimized via Graph
Cuts. In: Heyden, A., Sparr, G., Nielsen, M., Johansen, P. (eds.) ECCV 2002.
LNCS, vol. 2352. Springer, Heidelberg (2002)
19. Kolmogorov, V., Boykov, Y., Rother, C.: Applications of Parametric Maxflow in
Computer Vision. In: ICCV 2007 (2007)
20. Pawan Kumar, M., Torr, P., Zisserman, A.: OBJ CUT. In: CVPR 2005 (2005)
21. Lampert, C., Blaschko, M., Hofman, T.: Beyond Sliding Windows: Object Local-
ization by Efficient Subwindow Search. In: CVPR 2008 (2008)
22. Leibe, B., Leonardis, A., Schiele, B.: Robust Object Detection with Interleaved
Categorization and Segmentation. IJCV 77(3) (2008)
23. Lempitsky, V., Blake, A., Rother, C.: Image Segmentation by Branch-and-Mincut.
Microsoft Technical Report MSR-TR-2008-100 (July 2008)
24. Leventon, M., Grimson, E., Faugeras, O.: Statistical Shape Influence in Geodesic
Active Contours. In: CVPR 2000 (2000)
25. Rother, C., Kolmogorov, V., Blake, A.: ”GrabCut”: interactive foreground extrac-
tion using iterated graph cuts. ACM Trans. Graph. 23(3) (2004)
26. Schoenemann, T., Cremers, D.: Globally Optimal Image Segmentation with an
Elastic Shape Prior. In: ICCV 2007 (2007)
27. Wang, Y., Staib, L.: Boundary Finding with Correspondence Using Statistical
Shape Models. In: CVPR 1998 (1998)
What Is a Good Image Segment?
A Unified Approach to Segment Extraction
Shai Bagon, Oren Boiman, and Michal Irani
Weizmann Institute of Science, Rehovot, Israel
Abstract. There is a huge diversity of definitions of “visually meaning-

ful” image segments, ranging from simple uniformly colored segments,
textured segments, through symmetric patterns, and up to complex se-
mantically meaningful objects. This diversity has led to a wide range of
different approaches for image segmentation. In this paper we present
a single unified framework for addressing this problem – “Segmentation
by Composition”. We define a good image segment as one which can be
easily composed using its own pieces, but is difficult to compose using
pieces from other parts of the image. This non-parametric approach cap-
tures a large diversity of segment types, yet requires no pre-definition or
modelling of segment types, nor prior training. Based on this definition,
we develop a segment extraction algorithm – i.e., given a single point-of-
interest, provide the “best” image segment containing that point. This
induces a figure-ground image segmentation, which applies to a range
of different segmentation tasks: single image segmentation, simultaneous
co-segmentation of several images, and class-based segmentations.
1 Introduction
One of the most fundamental vision tasks is image segmentation; the attempt to
group image pixels into visually meaningful segments. However, the notion of a
“visually meaningful” image segment is quite complex. There is a huge diversity
in possible definitions of what is a good image segment, as illustrated in Fig. 1.
In the simplest case, a uniform colored region may be a good image segment
(e.g., the flower in Fig. 1.a). In other cases, a good segment might be a textured
region (Fig. 1.b, 1.c) or semantically meaningful layers composed of disconnected
regions (Fig. 1.c) and all the way to complex objects (Fig. 1.e, 1.f).
The diversity in segment types has led to a wide range of approaches for
image segmentation: Algorithms for extracting uniformly colored regions (e.g.,
[1,2]), algorithms for extracting textured regions (e.g., [3,4]), algorithm for ex-
tracting regions with a distinct empirical color distribution (e.g., [5,6,7]). Some
algorithms employ symmetry cues for image segmentation (e.g., [8]), while others
use high-level semantic cues provided by object classes (i.e., class-based segmen-
tation, see [9,10,11]). Some algorithms are unsupervised (e.g., [2]), while others
require user interaction (e.g., [7]). There are also variants in the segmentation

Author names are ordered alphabetically due to equal contribution.

What Is a Good Image Segment? 31
(a) (b) (c) (d) (e) (f)
Fig. 1. What is a good image segment? Examples of visually meaningful image

segments. These vary from uniformly colored segments (a) through textured segments
(b)-(c), symmetric segments (d), to semantically meaningful segments (e)-(f ). These
results were provided by our single unified framework.
Fig. 3. Notations:

Seg = S, S, ∂S denotes a
Fig. 2. Segmentation by composition: A good figure-ground segmentation.
segment S (e.g., the butterfly or the dome) can be eas- S is the foreground segment,
ily composed of other regions in the segment. Regions S (its compliment) is the
R1 , R2 are composed from other corresponding regions background, and ∂S is the
in S (using transformations T1 , T2 respectively). boundary of the segment.
tasks, ranging from segmentation of a single input image, through simultaneous

segmentation of a pair of images (“Cosegmentation” [12]) or multiple images.
The large diversity of image segment types has increased the urge to devise a
unified segmentation approach. Tu et al. [13] provided such a unified probabilis-
tic framework, which enables to “plug-in” a wide variety of parametric models
capturing different segment types. While their framework elegantly unifies these
parametric models, it is restricted to a predefined set of segment types, and
each specific object/segment type (e.g., faces, text, texture etc.) requires its own
explicit parametric model. Moreover, adding a new parametric model to this
framework requires a significant and careful algorithm re-design.
In this paper we propose a single unified approach to define and extract visu-
ally meaningful image segments, without any explicit modelling. Our approach
defines a “good image segment” as one which is “easy to compose” (like a puzzle)
using its own parts, yet it is difficult to compose it from other parts of the image
(see Fig. 2). We formulate our “Segmentation-by-Composition” approach, using
32 S. Bagon, O. Boiman, and M. Irani
a unified non-parametric score for segment quality. Our unified score captures
a wide range of segment types: uniformly colored segments, through textured
segments, and even complex objects. We further present a simple interactive
segment extraction algorithm, which optimizes our score – i.e., given a single
point marked by the user, the algorithm extracts the “best” image segment con-
taining that point. This in turn induces a figure-ground segmentation of the
image. We provide results demonstrating the applicability of our score and al-
gorithm to a diversity of segment types and segmentation tasks. The rest of
this paper is organized as follows: In Sec. 2 we explain the basic concept behind
our “Segmentation-by-Composition” approach for evaluating the visual quality
of image segments. Sec. 3 provides the theoretical formulation of our unified
segment quality score. We continue to describe our figure-ground segmentation
algorithm in Sec. 4. Experimental results are provided in Sec. 5.
2 Basic Concept – “Segmentation By Composition”
Examining the image segments of Fig. 1, we note that good segments of signifi-
cantly different types share a common property: Given any point within a good
image segment, it is easy to compose (“describe”) its surrounding region using
other chunks of the same segment (like a ‘jigsaw puzzle’), whereas it is difficult to
compose it using chunks from the remaining parts of the image. This is trivially
true for uniformly colored and textured segments (Fig. 1.a, 1.b, 1.c), since each
portion of the segment (e.g., the dome) can be easily synthesized using other
portions of the same segment (the dome), but difficult to compose using chunks
from the remaining parts of the image (the sky). The same property carries to
more complex structured segments, such as the compound puffins segment in
Fig. 1.f. The surrounding region of each point in the puffin segment is easy to
“describe” using portions of other puffins. The existence of several puffins in the
image provides ‘visual evidence’ that the co-occurrence of different parts (orange
beak, black neck, white body, etc.) is not coincidental, and all belong to a single
compound segment. Similarly, one half of a complex symmetric object (e.g., the
butterfly of Fig. 1.d, the man of Fig. 1.e) can be easily composed using its other
half, providing visual evidence that these parts go together. Moreover, the sim-
pler the segment composition (i.e., the larger the puzzle pieces), the higher the
evidence that all these parts form together a single segment. Thus, the entire
man of Fig. 1.e forms a better single segment than his pants or shirt alone.
The ease of describing (composing) an image in terms of pieces of another
image was defined by [14], and used there in the context of image similarity.
The pieces used for composition are structured image regions (as opposed to un-
structured ‘bags’/distributions of pointwise features/descriptors, e.g., as in [5,7]).
Those structured regions, of arbitrary shape and size, can undergo a global geo-
metric transformation (e.g., translation, rotation, scaling) with additional small
local non-rigid deformations. We employ the composition framework of [14] for
the purpose of image segmentation. We define a “good image segment” S as
one that is easy to compose (non-trivially) using its own pieces, while difficult to
compose from the remaining parts of the image S = I \S. An “easy” composition
consists of a few large image regions, whereas a “difficult” composition consists
of many small fragments. A segment composition induces a description of the
segment, with a corresponding “description length”. The easier the composition,
the shorter the description length. The ease of composing S from its own pieces
is formulated in Sec. 3 in terms of the description length DL (S|S). This is con-
trasted with the ease of composing
S from pieces of the remaining image parts
S, which is captured by DL S|S . This gives rise to a “segment quality score”
Score (S), which is measured by the difference between these two description
lengths: Score (S) = DL S|S − DL (S|S).
Our definition of a “good image segment” will maximize this difference in de-
scription lengths. Any deviation from the optimal segment S will reduce this differ-
ence, and accordingly decrease Score (S). For example, the entire dome in Fig. 1.b
is an optimal image segment S; it is easy to describe non-trivially in terms of its own
pieces (see Fig. 2), and difficult to describe in terms of the background sky. If, how-
ever, we were to define the segment S to be only a smaller part of the dome, then the
background S would contain the sky alongwiththe parts of the dome excluded from
S. Consequently, this would decrease DL S|S and therefore Score (S) would de-
crease. It can be similarly shown that Score (S) would decrease if we were to define
S which is larger than the dome and contains also parts of the sky. Note that unlike
previous simplistic formulations of segment description length (e.g., entropy of sim-
ple color distributions [5]), our composition-based description length can capture
also complex structured segments.
A good figure-ground segmentation Seg = S, S, ∂S (see Fig. 3) partitions
the image into a foreground segment S and a background segment S, where at
least one of these two segments (and hopefully both) is a ‘good image segment’
according to the definition above. Moreover, we expect the segment boundary ∂S
of a good figure-ground segmentation to coincide with meaningful image edges.
Boiman and Irani [14] further employed the composition framework for coarse
grouping of repeating patterns. Our work builds on top of [14], providing a gen-
eral segment quality score and a corresponding image segmentation algorithm,
which applies to a large diversity of segment types, and can be applied for vari-
ous segmentation tasks. Although general, our unified segmentation framework
does not require any pre-definition or modelling of segment types (in contrast to
the unified framework of [13]).
3 Theoretical Formulation
The notion of ‘description by composition’ was introduced by Boiman and Irani

in [14], in the context of image similarity. They provided a similarity measure
between a query image Q and a reference image Ref , according to how easy it
is to compose Q from pieces of Ref . Intuitively speaking, the larger those pieces
are, the greater the similarity. Our paper builds on top of the basic compositional
formulations of [14]. To make our paper self-contained, we briefly review those
basic formulations.
The composition approach is formulated as a generative process by which

the query image Q is generated as a composition of arbitrarily shaped pieces
(regions) taken from the reference image Ref . Each such region from Ref can
undergo a geometric transformation (e.g., shift, scale, rotation, reflection) before
being “copied” to Q in the composition process. The likelihood of an arbitrarily
shaped region R ⊂ Q given a reference image Ref is therefore:

p (R|Ref ) = p (R|T, Ref ) p (T |Ref ) (1)
T
where T is a geometric transformation from Ref to the location of R in Q.

p (R|T, Ref ) is determined by the degree of similarity of R to a region in Ref
which is transformed by T to the location of R. This probability is marginal-
ized over all possible transformations T using a prior over the transformations
p (T |Ref ), resulting in the ‘frequency’ of region R in Ref . Given a partition of Q
into regions R1 , . . . , Rk (assumed i.i.d. given the partition), the likelihood that
a query image Q is composed from Ref using this partition is defined by [14]:

k
p (Q|Ref ) = p (Ri |Ref ) (2)
i=1
Because there are many possible partitions of Q into regions, the righthand side
of (2) is marginalized over all possible partitions in [14].
p (Q|Ref ) /p (Q|H0 ) is the likelihood-ratio between the ‘ease’ of generating
Q from Ref vs. the ease of generating Q using a “random process” H0 (e.g.,
a default image distribution). Noting that the optimal (Shannon) description
length of a random variable x is DL (x) ≡ − log p (x) [15], Boiman and Irani [14]
defined their compositional similarity score as: log (p (Q|Ref )/p (Q|H0 )) =
DL (Q|H0 ) − DL (Q|Ref ) i.e., the “savings” in the number of bits obtained
by describing Q as composed from regions in Ref vs. the ‘default’ number of
bits required to describe Q using H0 . The larger the regions Ri composing Q
the higher the savings in description length. High savings in description length
provide high statistical evidence for the similarity of Q to Ref .
In order to avoid the computationally-intractable marginalization over all pos-
sible query partitions, the following approximation was derived in [14]:

DL (Q|H0 ) − DL (Q|Ref ) ≈ PES (i|Ref ) (3)
i∈Q
where PES (i|Ref ) is a pointwise measure (a Point-Evidence-Score) of a pixel i:

1 p (R|Ref )
PES (i|Ref ) = max log (4)
R⊂Q,i∈R |R| p (R|H0 )
Intuitively, given a region R, |R|
1
log (p (R|Ref ) /p (R|H0 )) is the average savings
per pixel in the region R. Thus, PES (i|Ref ) is the maximum possible savings per
pixel for any region R containing the point i. We refer to the region which obtains
this maximal value PES (i|Ref ) as a ‘maximal region’ around i. The approximate
computation of (4) can be done efficiently (see [14] for more details).
3.1 The Segment Quality Score

A good segment S should be easy to compose from its own pieces using a
non-trivial composition, yet difficult to compose from the rest of the image S
(e.g., Fig. 2). Thus,
we expect that for good segments, the description length
DL S|Ref = S should be much larger than DL (S|Ref = S). Accordingly, we

define Score (S) = DL S|Ref = S − DL (S|Ref = S). We use (2) to com-
pute p (S|Ref ) (the segment S taking the role of the query Q), in order to
define the likelihood and the description length of the segment S, once w.r.t.
to itself (Ref = S), and once w.r.t. to the rest of the image (Ref = S).
We note that DL S|Ref = S = − log p S|Ref = S , and DL (S|Ref = S) =
− log p (S|Ref = S). In order to avoid the trivial (identity) composition when
composing S from its own pieces, we exclude transformations T from (1) that
are close to the identity transformation (e.g., when T is a pure shift, it should be
of at least 15 pixels.) Using the approximation of (3), we can rewrite Score (S):

Score (S) = DL S|Ref = S − DL (S|Ref = S) (5)

= (DL (S|H0 ) − DL (S|Ref = S)) − DL (S|H0 ) − DL S|Ref = S

≈ PES (i|S) − PES i|S = PES (i|S) − PES i|S (6)
i∈S i∈S i∈S

Thus, Score (S) accumulates for every pixel i ∈ S the term PES (i|S)−PES i|S ,
which compares the ‘preference’ (the pointwise evidence) of the pixel i to belong
to the segment S, relative to its ‘preference’ to belong to S.
3.2 The Segmentation Quality Score

A good figure-ground segmentation is such that at least one of its two segments,
S or S, is ‘a good image segment’ (possibly both), and with a good segmenta-
tion boundary ∂S (e.g., coincides with strong image edges, is smooth, etc.) We
therefore define a figure-ground
segmentation quality score as: Score (Seg) =
Score (S) + Score S + Score (∂S), where Score (∂S) denotes the quality of the
segmentation boundary ∂S. Using (6), Score (Seg) can be rewritten as:

Score (Seg) = Score (S) + Score S + Score (∂S) (7)

= PES (i|S) − PES i|S + PES i|S − PES (i|S) + Score (∂S)
i∈S i∈S
The quality of the segmentation boundary, Score (∂S), is defined as follows:

Let P r (Edgei,j ) be the probability of an edge between every two neighbor-
ing pixels i, j (e.g., computed using [16]).
We define the likelihood of a seg-
mentation boundary ∂S as: p (∂S) = i∈S, j∈S,(i,j)∈N P r (Edgei,j ) , where
N is the set of neighboring pixels. We define the score of the boundary ∂S
byits ‘description length’, i.e.: Score (∂S) = DL (∂S) = − log p (∂S) =
− i∈S, j∈S log P r (Edge i,j ). Fig. 4 shows quantitatively that Score (Seg)
peaks at proper segment boundaries, and decreases when ∂S deviates from it.
(a) (b)
Fig. 4. Score (Seg) as a function of deviations in boundary position ∂S:

(a) shows the segmentation score as a function of the boundary position. It obtains
a maximum value at the edge between the two textures. (b) The segmentation score as
a function of the deviation from the recovered segment boundary for various segment
types (deviations were generated by shrinking and expanding the segment boundary).
The above formulation can be easily extended to a quality score of a gen-

m segmentation of an image into m segments,
eral S1 , . . . , Sm : Score (Seg) =
i=1 Score (Si ) + Score (∂S) , s.t. ∂S = mi=1 ∂Si .
3.3 An Information-Theoretic Interpretation
We next show that our segment quality score, Score (S), has an interesting
information-theoretic interpretation, which reduces in special sub-cases to com-
monly used information-theoretic measures. Let us first examine the simple case
where the composition of a segment S is restricted to degenerate one-pixel sized
regions Ri . In this case, p (Ri |Ref = S) in (1) reduces to the frequency of the
color of the pixel Ri inside S (given by the color histogram of S). Using (2) with
one-pixel sized regions Ri , the description length DL (S|Ref = S) reduces to:

DL (S|Ref = S) = − log p (S|Ref = S) = − log p (Ri |Ref = S)
i∈S

=− log p (Ri |Ref = S) = |S| · Ĥ (S)
i∈S
where Ĥ (S) is the empirical entropy1 of the regions {Ri } composing S, which
is
the color entropy of S in case of one-pixel sized Ri . Similarly, DL S|Ref = S =

− i∈S log p Ri |Ref = S = |S| · Ĥ S, S , where Ĥ S, S is the empirical
cross-entropy of regions Ri ⊂ S in S (which reduces to the color cross-entropy

1
The empirical entropy of the sample x1 , .., xn is Ĥ (x) = − n1 i log p (xi ) which
approaches the statistical entropy H (x) as n → ∞.
in case of one-pixel sized Ri ). Using these observations, Score (S) of (5) reduces
to the empirical KL divergence between the region distributions of S and S:

Score (S) = DL S|S − DL (S|S) = |S| · Ĥ S, S − Ĥ (S) = |S| · KL S, S
In the case of single-pixel-sized regions Ri , this reduces to the KL divergence

between the color distributions of S and S.
A similar derivation can be applied to the general case of composing S from
arbitrarily shaped regions Ri . In that case, p (Ri |Ref ) of (1) is the frequency
of regions Ri ⊂ S in Ref = S or in Ref = S (estimated non-parametrically
using region composition). This gives rise to an interpretation of the description
length DL (S|Ref ) as a Shannon entropy measure, and our segment quality score
Score (S) of (5) can be interpreted as a KL divergence between the statistical
distributions of regions (of arbitrary shape and size) in S and in S.
Note that in the degenerate case when the regions Ri ⊂ S are one-pixel sized,
our framework reduces to a formulation closely related to that of GrabCut [7]
(i.e., figure-ground segmentation into segments of distinct color distributions).
However, our general formulation employs regions of arbitrary shapes and sizes,
giving rise to figure-ground segmentation with distinct region distributions. This
is essential when S and S share similar color distributions (first order statistics),
and vary only in their structural patterns (i.e., higher order statistics). Such an
example can be found in Fig. 5 which compares our results to that of GrabCut.
3.4 The Geometric Transformations T
The family of geometric transformations T applied to regions R in the compo-

sition process (Eq. 1) determines the degree of complexity of segments that can
be handled by our approach. For instance, if we restrict T to pure translations,
then a segment S may be composed by shuffling and combining pieces from
Ref . Introducing scaling/rotation/affine transformations enables more complex
compositions (e.g., compose a small object from a large one, etc.) Further in-
cluding reflection transformations enables composing one half of a symmetric
object/pattern from its other half. Note that different regions Ri ⊂ S are ‘gen-
erated’ from Ref using different transformations Ti . Combining several types
of transformations can give rise to composition of very complex objects S from
their own sub-regions (e.g., partially symmetric object as in Fig. 10.b).
4 Figure-Ground Segmentation Algorithm
In this section we outline our figure-ground segmentation algorithm, which op-

timizes Score (Seg) of (7). The goal of figure-ground segmentation is to extract
an object of interest (the “foreground”) from the remaining parts of the image
(the “background”). In general, when the image contains multiple objects, a user
input is required to specify the “foreground” object of interest.
Input Our results Results of GrabCut [7]

image Init+ recovered S Init bounding box recovered S
Fig. 5. Our result vs. GrabCut [7]. GrabCut fails to segment the butterfly (fore-
ground) due to the similar colors of the flowers in the background. Using composition
with arbitrarily shaped regions, our algorithm accurately segments the butterfly. We
used the GrabCut implementation of www.cs.cmu.edu/∼ mohitg/segmentation.htm
Different figure-ground segmentation algorithms require different amounts

of user-input to specify the foreground object, whether in the form of fore-
ground/background scribbles (e.g., [6]), or a bounding-box containing the fore-
ground object (e.g., [7]). In contrast, our figure-ground segmentation algorithm
requires a minimal amount of user input – a single user-marked point on the
foreground segment/object of interest. Our algorithm proceeds to extract the
“best” possible image segment containing that point. In other words, the algo-
rithm recovers a figure-ground segmentation Seg = S, S, ∂S s.t. S contains
the user-marked point, and Seg maximizes the segmentation score of (7). Fig. 6
shows how different user-selected points-of-interest extract different objects of
interest from the image (inducing different figure-ground segmentations Seg).
A figure-ground segmentation can be described by assigning a label li to every
pixel i in the image, where li = 1 ∀i ∈ S, and li = −1 ∀i ∈ S. We can rewrite
Score (Seg) of (7) in terms of these labels:
1
Score (Seg) = li · PES (i|S) − PES i|S + |li − lj |·log P r (Edgei,j )
2
i∈I (i,j)∈N
(8)
where N is the set of all pairs of neighboring pixels. Maximizing (8) is equiv-
alent to an energy minimization formulation which
can be optimized using a
MinCut algorithm [17], where PES (i|S) − PES i|S form the data term, and
log P r (Edgei,j ) is the “smoothness” term. However, the data term has a com-
plicated
dependency
on the segmentation into S, S, via the terms PES (i|S) and
PES i|S . This prevents straightforward application of MinCut. To overcome
this problem, we employ EM-like iterations, i.e., alternating between estimating
the data term and maximizing Score (Seg) using MinCut (see Sec. 4.1).
In our current implementation the “smoothness” term P r (Edgei,j ), is com-
puted based on the edge probabilities of [16], which incorporates texture, lumi-
nance and color cues. The computation of PES (i|Ref ) for every pixel i (where
Input image Extracted S (red) around user selected points (green)
Fig. 6. Different input points result in different foreground segments
Ref is either S or S) involves finding a ‘maximal region’ R surrounding i which

has similar regions elsewhere in Ref , i.e., a region R that maximizes (4). An
image region R (of any shape or size) is represented by a dense and structured
‘ensemble of patch descriptors’ using a star-graph model. When searching for a
similar region, we search for a similar ensemble of patches (similar both in their
patch descriptors, as well as in their relative geometric positions), up to a global
transformation T (Sec. 3.4) and small local non-rigid deformations (see [18]).
We find these ‘maximal regions’ R using the efficient region-growing algorithm
of [18,14]: Starting with a small surrounding region around a pixel i, we search
for similar such small regions in Ref . These few matched regions form seeds for
the region growing algorithm. The initial region around i with its matching seed
regions are simultaneously grown (in a greedy fashion) to find maximal matching
regions (to maximize PES (i|Ref )). For more details see [18,14].
4.1 Iterative Optimization

Initialization. The input to our segment extraction algorithm is an image and a
single user-marked point of interest q. We use the region composition procedure
to generate maximal regions for points in the vicinity of q. We keep only the
maximal regions that contain q and have high evidence (i.e., PES) scores. The
(a) (b) (c) (d) (e) (f)
Input Image Input point Score=418 Score=622 Score=767 S
Fig. 7. Progress of the iterative process: Sequence of intermediate segmentations

of the iterative process. (a) The input image. (b) The user marked point-of-interest.
(c) Initialization of S. (d) S after 22 iterations. (e) Final segment S after 48 iterations.
(f ) The resulting figure-ground segments, S and S. The iterations converged accurately
to the requested segment after 48 iterations.
Input image pair Our cosegmentation Cosegmentation of [12]
Fig. 8. Cosegmentation of image pair: Comparing our result to that of [12]
union of these regions, along with their corresponding reference regions, is used
as a crude initialization, S0 , of the segment S (see Fig. 7.c for an example).
Iterations. Our optimization algorithm employs EM-like iterations:

In each it-
eration we first fix the current segmentation Seg = S, S, ∂S and compute the

data term by re-estimating PES (i|S) and PES i|S . Then, we fix the data term
and maximize Score (Seg) using MinCut [17] on (8). This process is iterated
until convergence (i.e., when Score (Seg) ceases to improve). The iterative pro-
cess is quite robust – even a crude initialization suffices for proper convergence.
For computational efficiency, in each iteration t we recompute PES (i|Ref ) and
relabel pixels only for pixels i within a narrow working band around the current
boundary ∂St . The segment boundary recovered in the next iteration, ∂St+1 ,
is restricted to pass inside that working band. The size of the working band is
∼ 10% of the image width, which restricts the computational complexity, yet
enables significant updates of the segment boundary in each iteration.
Input image 4 class images Init point+recovered Seg
Fig. 9. Class-based Segmentation: Segmenting a complex horse image (left) using

4 unsegmented example images of horses
During the iterative process, similar regions may have conflicting labels. Due
to the EM-like iterations, such regions may simultaneously flip their labels, and
fail to converge (since each such region provides “evidence” for the other to flip
its label). Therefore, in each iteration, we perform two types of steps successively:
(i) an “expansion” step, in which only background pixels in St are allowed to flip
their label to foreground pixels. (ii) a “shrinking” step, in which only foreground
pixels in St are allowed to flip their label to background pixels. Fig. 7 shows a
few steps in the iterative process, from initialization to convergence.
4.2 Integrating Several Descriptor Types

The composition process computes similarity of image regions, using local descrip-
tors densely computed within the regions. To allow for flexibility, our framework
integrates several descriptor-types, each handles a different aspect of similarity be-
tween image points (e.g., color, texture). Thus, several descriptor types can collab-
orate to describe a complex segment (e.g., in a “multi-person” segment, the color
descriptor is dominant in the face regions, while the shape descriptor may be more
dominant in other parts of the body). Although descriptor types are very differ-
ent , the ‘savings’ in description length obtained by each descriptor type are all in
the same units (i.e., bits). Therefore, we can integrate different descriptor-types
by simply adding their savings. A descriptor type that is useful for describing a
region will increase the savings in description length, while non-useful descriptor
types will save nothing. We used the following descriptor types: (1) SIFT (2)
Color: based on a color histogram (3) Texture: based on a texton histogram (4)
Shape: An extension of Shape Context descriptor of Belongie et al. (5) The Self
Similarity descriptor of Shechtman and Irani.
5 Results
We applied our segment extraction algorithm to a variety of segment types and
segmentation tasks, using images from several segmentation databases [19,20,7].
In each case, a single point-of-interest was marked (a green cross in the fig-
ures). The algorithm extracted the “best” image segment containing that point
User selected point (green) with the recovered

Input
figure-ground segmentation
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
(j)
Fig. 10. Examples of figure-ground segmentations

(highlighted in red). Higher resolution images and many more results can be
found in www.wisdom.weizmann.ac.il/∼vision/GoodSegment.html.
Single-Image Segmentation. Fig. 10 demonstrates the capability of our ap-
proach to handle a variety of different segments types: uniformly colored seg-
ments (Fig. 10.f), complex textured segments (Fig. 10.h), complex symmetric
objects (e.g., the butterfly in Fig. 5, the Man in Fig. 1.e). More complex objects
can also be segmented (e.g., a non-symmetric person Fig. 10.b, or the puffins
Fig. 10.g), resulting from combinations of different types of transformations Ti
for different regions Ri within the segment, and different types of descriptors.
We further evaluated our algorithm on the benchmark database of [19], which
consists of 100 images depicting a single object in front of a background, with
ground-truth human segmentation. The total F-measure score of our algorithm
was 0.87±0.01 (F = 2·Recall·P recision
Recall+P recision ), which is state-of-the-art on this database.
“Cosegmentation”. We applied our segmentation algorithm with no modifica-

tions to a simultaneous co-segmentation of an image pair – the algorithm input
is simply the concatenated image pair. The common object in the images is
extracted as a single compound segment (Fig. 8, shows a comparison to [12]).
Class-Based Segmentation. Our algorithm can perform class-based segmen-

tation given unsegmented example images of an object class. In this case, we
append the example images to the reference Ref = S of the foreground seg-
ment S. Thus the object segment can be composed using other parts in the
segment as well as from parts in the example images. This process requires no
pre-segmentation and no prior learning stage. Fig. 9 shows an example of ex-
tracting a complex horse segment using 4 unsegmented examples of horse images.
References
1. Comaniciu, D., Meer, P.: Mean shift: a robust approach toward feature space anal-
ysis. PAMI (2002)
2. Shi, J., Malik, J.: Normalized cuts and image segmentation. PAMI (2000)
3. Malik, J., Belongie, S., Shi, J., Leung, T.K.: Textons, contours and regions: Cue
integration in image segmentation. In: ICCV (1999)
4. Galun, M., Sharon, E., Basri, R., Brandt, A.: Texture segmentation by multiscale
aggregation of filter responses and shape elements. In: ICCV (2003)
5. Kadir, T., Brady, M.: Unsupervised non-parametric region segmentation using level
sets. In: ICCV (2003)
6. Li, Y., Sun, J., Tang, C.K., Shum, H.Y.: Lazy snapping. ACM TOG (2004)
7. Rother, C., Kolmogorov, V., Blake, A.: “grabcut”: Interactive foreground extrac-
tion using iterated graph cuts. In: SIGGRAPH (2004)
8. Riklin-Raviv, T., Kiryati, N., Sochen, N.: Segmentation by level sets and symmetry.
In: CVPR (2006)
9. Borenstein, E., Ullman, S.: Class-specific, top-down segmentation. In: Heyden, A.,
Sparr, G., Nielsen, M., Johansen, P. (eds.) ECCV 2002. LNCS, vol. 2351. Springer,
Heidelberg (2002)
10. Leibe, B., Schiele, B.: Interleaved object categorization and segmentation. In:
BMVC (2003)
11. Levin, A., Weiss, Y.: Learning to combine bottom-up and top-down segmenta-
tion. In: Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3954.
12. Rother, C., Minka, T., Blake, A., Kolmogorov, V.: Cosegmentation of image pairs
by histogram matching - incorporating a global constraint into mrfs. In: CVPR
(2006)
13. Tu, Z., Chen, X., Yuille, A.L., Zhu, S.C.: Image parsing: Unifying segmentation,
detection, and recognition. IJCV (2005)
14. Boiman, O., Irani, M.: Similarity by composition. In: NIPS (2006)
15. Cover, T.M., Thomas, J.A.: Elements of information theory. Wiley, Chichester
(1991)
16. Martin, D.R., Fowlkes, C.C., Malik, J.: Learning to detect natural image bound-
aries using local brightness, color, and texture cues. PAMI (2004)
17. Boykov, Y., Veksler, O., Zabih, R.: Efficient approximate energy minimization via
graph cuts. PAMI (2001)
18. Boiman, O., Irani, M.: Detecting irregularities in images and in video. IJCV (2007)
19. Alpert, S., Galun, M., Basri, R., Brandt, A.: Image segmentation by probabilistic
bottom-up aggregation and cue integration. In: CVPR (2007)
20. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural
images and its application to evaluating segmentation algorithms and measuring
ecological statistics. In: ICCV (2001)
Light-Efficient Photography
Samuel W. Hasinoff and Kiriakos N. Kutulakos
Dept. of Computer Science, University of Toronto

{hasinoff,kyros}@cs.toronto.edu
Abstract. We consider the problem of imaging a scene with a given depth of

field at a given exposure level in the shortest amount of time possible. We show
that by (1) collecting a sequence of photos and (2) controlling the aperture, focus
and exposure time of each photo individually, we can span the given depth of
field in less total time than it takes to expose a single narrower-aperture photo.
Using this as a starting point, we obtain two key results. First, for lenses with
continuously-variable apertures, we derive a closed-form solution for the glob-
ally optimal capture sequence, i.e., that collects light from the specified depth of
field in the most efficient way possible. Second, for lenses with discrete aper-
tures, we derive an integer programming problem whose solution is the optimal
sequence. Our results are applicable to off-the-shelf cameras and typical photog-
raphy conditions, and advocate the use of dense, wide-aperture photo sequences
as a light-efficient alternative to single-shot, narrow-aperture photography.
1 Introduction
Two of the most important choices when taking a photo are the photo’s exposure level
and its depth of field. Ideally, these choices will result in a photo whose subject is
free of noise or pixel saturation [1,2], and appears in-focus. These choices, however,
come with a severe time constraint: in order to take a photo that has both a specific
exposure level and a specific depth of field, we must expose the camera’s sensor for
a length of time dictated by the optics of the lens. Moreover, the larger the depth of
field, the longer we must wait for the sensor to reach the chosen exposure level. In
practice, this makes it impossible to efficiently take sharp and well-exposed photos of
a poorly-illuminated subject that spans a wide range of distances from the camera. To
get a good exposure level, we must compromise something – accepting either a smaller
depth of field (incurring defocus blur [3,4,5,6]) or a longer exposure (incurring motion
blur [7,8,9]).
In this paper we seek to overcome the time constraint imposed by lens optics, by
capturing a sequence of photos rather than just one. We show that if the aperture, ex-
posure time, and focus setting of each photo is selected appropriately, we can span a
given depth of field with a given exposure level in less total time than it takes to expose
a single photo (Fig. 1). This novel observation is based on a simple fact: even though
wide apertures have a narrow depth of field (DOF), they are much more efficient than
narrow apertures in gathering light from within their depth of field. Hence, even though

This work was supported in part by the Natural Sciences and Engineering Research Council of
Canada under the RGPIN program and by an Ontario Premier’s Research Excellence Award.
46 S.W. Hasinoff and K.N. Kutulakos
2s 0.5 s 0.5 s
1 photo @ f /8 2 photos @ f /4 synthesized photo

total time: 2 s total time: 1 s with desired DOF
Fig. 1. Left: Traditional single-shot photography. The desired depth of field is shaded (red). Right:
Light-efficient photography. Two wide-aperture photos span the same DOF as a single-shot
narrow-aperture photo. Each wide-aperture photo requires 1/4 the time to reach the exposure
level of the single-shot photo, resulting in a 2× net speedup for the total exposure time.
it is not possible to span a wide DOF with a single wide-aperture photo, it is possible to
span it with several of them, and to do so very efficiently.
Using this observation as a starting point, we develop a general theory of light-
efficient photography that addresses four questions: (1) under what conditions is
capturing photo sequences with “synthetic” DOFs more efficient than single-shot pho-
tography? (2) How can we characterize the set of sequences that are globally optimal for
a given DOF and exposure level, i.e. whose total exposure time is the shortest possible?
(3) How can we compute such sequences automatically for a specific camera, depth of
field, and exposure level? (4) Finally, how do we convert the captured sequence into a
single photo with the specified depth of field and exposure level?
Little is known about how to gather light efficiently from a specified DOF. Research
on computational photography has not investigated the light-gathering ability of ex-
isting methods, and has not considered the problem of optimizing exposure time for
a desired DOF and exposure level. For example, even though there has been great
interest in manipulating a camera’s DOF through optical [10,11,12,13] or computa-
tional [5,14,15,16,17,18,2] means, current approaches do so without regard to exposure
time – they simply assume that the shutter remains open as long as necessary to reach
the desired exposure level. This assumption is also used for high-dynamic range pho-
tography [19,2], where the shutter must remain open for long periods in order to capture
low-radiance regions in a scene. In contrast, here we capture photos with camera set-
tings that are carefully chosen to minimize total exposure time for the desired DOF and
exposure level.
Since shorter total exposure times reduce motion blur, our work can be thought of
as complementary to recent synthetic shutter approaches whose goal is to reduce such
blur. Instead of controlling aperture and focus, these techniques divide a given exposure
interval into several shorter ones, with the same total exposure (e.g., n photos, each with
1/n the exposure time [9]; two photos, one with long and one with short exposure [8];
or one photo where the shutter opens and closes intermittently during the exposure [7]).
These techniques do not increase light-efficiency but can be readily combined with our
work, to confer the advantages of both methods.
Light-Efficient Photography 47
Moreover, our approach can be thought of as complementary to work on light field

cameras [17,18,13], which are based on an orthogonal tradeoff between resolution and
directional sampling. Compared to regular wide-aperture photography, these designs
do not have the ability to extend the DOF when their reduced resolution is taken into
account. Along similar lines, wavefront coding [11] exploits special optics to extend the
DOF with no change in exposure time by using another orthogonal tradeoff – accepting
lower signal-to-noise ratio for higher frequencies.
The final step in light-efficient photography involves merging the captured photos
to create a new one (Fig. 1). As such, our work is related to the well-known technique
of extended-depth-of-field imaging, which has found wide use in microscopy [18] and
macro photography [20,17].
Our work offers four contributions over the state of the art. First, we develop a theory
that leads to provably-efficient light-gathering strategies, and applies both to off-the-
shelf cameras and to advanced camera designs [7,9] under typical photography con-
ditions. Second, from a practical standpoint, our analysis shows that the optimal (or
near-optimal) strategies are very simple: for example, in the continuous case, a strategy
using the widest-possible aperture for all photos is either globally optimal or it is very
close to it (in a quantifiable sense). Third, our experiments with real scenes suggest
that it is possible to compute good-quality synthesized photos using readily-available
algorithms. Fourth, we show that despite requiring less total exposure time than a single
narrow-aperture shot, light-efficient photography provides more information about the
scene (i.e., depth) and allows post-capture control of aperture and focus.
2 The Exposure Time vs. Depth of Field Tradeoff

The exposure level of a photo is the total radiant energy integrated by the camera’s
entire sensor while the shutter is open. The exposure level can influence significantly
the quality of a captured photo because when there is no saturation or thermal noise,
a pixel’s signal-to-noise ratio (SNR) always increases with higher exposure levels [1].
For this reason, most modern cameras can automate the task of choosing an exposure
level that provides high SNR for most pixels and causes little or no saturation.
Lens-based camera systems provide only two ways to control exposure level – the di-
ameter of their aperture and the exposure time. We assume that all light passing through
the aperture will reach the sensor plane, and that the average irradiance measured over
this aperture is independent of the aperture’s diameter. In this case, the exposure level
L is equal to
L = τ D2 , (1)
where τ is exposure time, D is the effective aperture diameter, and the units of L are
chosen appropriately.
Now suppose that we have chosen a desired exposure level L∗ . How can we capture
a photo at this exposure level? Equation (1) suggests that there are only two general
strategies for doing this – either choose a long exposure time and a small aperture di-
ameter, or choose a large aperture diameter and a short exposure time. Unfortunately,
both strategies have important side-effects: increasing exposure time can introduce mo-
tion blur when we photograph moving scenes [8,9]; opening the lens aperture, on the
width of DOF in scene (mm)

150 5.0
aperture diameter (mm)

) ht
(10 7brig
)
10 5
100 7.4
y
ver
t(
3 )
h
0
ig
(1
br
rk
da
50 1) 14.8
0
k (1
dar
very
0 ∞
−4 −3 −2 −1 0 1
10 10 10 10 10 10
exposure time (s)
Fig. 2. Each curve represents all pairs (τ, D) for which τ D2 = L∗ in a specific scene. Shaded
zones correspond to pairs outside the camera limits (valid settings were τ ∈ [1/8000 s, 30 s] and
D ∈ [f /16, f /1.2] with f = 85 mm). Also shown is the DOF corresponding to each diameter D.
The maximum acceptable blur was set to c = 25 µm, or about 3 pixels in our camera. Different
curves represent scenes with different average radiance (relative units shown in brackets).
other hand, affects the photo’s depth of field (DOF), i.e., the range of distances where
scene points do not appear out of focus. These side-effects lead to an important tradeoff
between a photo’s exposure time and its depth of field (Fig. 2):
Exposure Time vs. Depth of field Tradeoff: We can either achieve a desired
exposure level L∗ with short exposure times and a narrow DOF, or with long
exposure times and a wide DOF.
In practice, the exposure time vs. DOF tradeoff limits the range of scenes that can be
photographed at a given exposure level (Fig. 2). This range depends on scene radiance,
the physical limits of the camera (i.e., range of possible apertures and shutter speeds),
as well as subjective factors (i.e., acceptable levels of motion blur and defocus blur).
Our goal is to “break” this tradeoff by seeking novel photo acquisition strategies that
capture a given depth of field at the desired exposure level L∗ much faster than tradi-
tional optics would predict. We briefly describe below the basic geometry and relations
governing a photo’s depth of field, as they are particularly important for our analysis.
2.1 Depth of Field Geometry

We assume that focus and defocus obey the standard thin lens model [3,21]. This model
relates three positive quantities (Eq. (A) in Table 1): the focus setting v, defined as the
distance from the sensor plane to the lens; the distance d from the lens to the in-focus
scene plane; and the focal length f , representing the “focusing power” of the lens.
Apart from the idealized pinhole, all apertures induce spatially-varying amounts of
defocus for points in the scene (Fig. 3a). If the lens focus setting is v, all points at
distance d from the lens will be in-focus. A scene point at distance d = d, however,
will be defocused: its image will be a circle on the sensor plane whose diameter b is
called the blur diameter. For any given distance d, the thin-lens model tells us exactly
what focus setting we should use to bring the plane at distance d into focus, and what the
blur diameter will be for points away from this plane (Eqs. (B) and (C), respectively).
sensor in-focus
plane plane
blur diameter (µm)

blur diameter (µm)
lens
b scene
D
DOF DOF
c c
v d
dn d df α v β
v d scene depth (cm) scene focus setting (mm)
(a) (b) (c)
Fig. 3. (a) Blur geometry for a thin lens. (b) Blur diameter as a function of distance to a scene
point. The plot is for a lens with f = 85 mm, focused at 117 cm with an aperture diameter
of 5.31 mm (i.e., an f /16 aperture in photography terminology). (c) Blur diameter and DOF
represented in the space of focus settings.
Table 1. Eqs. (A)–(F): Basic equations governing focus and DOFs for the thin-lens model
(A) Thin (B) Focus for (C) Blur diameter (D) Aper. diam. (E) Focus for (F) DOF for aper.
lens law distance d for distance d for DOF [α, β] DOF [α, β] diam. D, focus v
1 1 1 fd f |d − d| β +α 2αβ Dv
+ = v= b=D D=c v= α, β =
v d f d−f d (d − f ) β −α α+β D±c
For a given aperture and focus setting, the depth of field is the interval of distances
in the scene whose blur diameter is below a maximum acceptable size c (Fig. 3b).
Since every distance in the scene corresponds to a unique focus setting (Eq. (B)),
every DOF can also be expressed as an interval [α, β] in the space of focus settings.
This alternate DOF representation gives us especially simple relations for the aperture
and focus setting that produce a given DOF (Eqs. (D) and (E)) and, conversely, for
the DOF produced by a given aperture and focus setting (Eq. (F)). We adopt this DOF
representation for the rest of the paper (Fig. 3c).
A key property of the depth of field is that it shrinks when the aperture diameter
increases: from Eq. (C) it follows that for a given out-of-focus distance, larger apertures
always produce larger blur diameters. This equation is the root cause of the exposure
time vs. depth of field tradeoff.
3 The Synthetic DOF Advantage

Suppose that we want to capture a single photo with a specific exposure level L∗ and a
specific depth of field [α, β]. How quickly can we capture this photo? The basic DOF
geometry of Sect. 2.1 tells us we have no choice: there is only one aperture diameter
that can span the given depth of field (Eq. (D)), and only one exposure time that can
achieve a given exposure level with that diameter (Eq. (1)). This exposure time is1
1
The apertures and exposure times of real cameras span finite intervals and, in many cases, take
discrete values. Hence, in practice, Eq. (2) holds only approximately.
2
β−α
τ one = L∗ · . (2)
c (β + α)
The key idea of our approach is that while lens optics do not allow us to reduce this
time without compromising the DOF or the exposure level, we can reduce it by taking
more photos. This is based on a simple observation that takes advantage of the different
rates at which exposure time and DOF change: if we increase the aperture diameter
and adjust exposure time to maintain a constant exposure level, its DOF shrinks (at a
rate of about 1/D), but the exposure time shrinks much faster (at a rate of 1/D2 ). This
opens the possibility of “breaking” the exposure time vs. DOF tradeoff by capturing a
sequence of photos that jointly span the DOF in less total time than τ one (Fig. 1).
Our goal is to study this idea in its full generality, by finding capture strategies that
are provably time-optimal. We therefore start from first principles, by formally defining
the notion of a capture sequence and of its synthetic depth of field:
Definition 1 (Photo Tuple). A tuple D, τ, v that specifies a photo’s aperture di-
ameter, exposure time, and focus setting, respectively.
Definition 2 (Capture Sequence). A finite ordered sequence of photo tuples.
Definition 3 (Synthetic Depth of Field). The union of DOFs of all photo tuples in a
capture sequence.
We will use two efficiency measures: the total exposure time of a sequence is the sum
of the exposure times of all its photos; the total capture time, on the other hand, is the
actual time it takes to capture the photos with a specific camera. This time is equal to
the total exposure time, plus any overhead caused by camera internals (computational
and mechanical). We now consider the following general problem:
Light-Efficient Photography: Given a set D of available aperture diameters,
construct a capture sequence such that: (1) its synthetic DOF is equal to [α, β];
(2) all its photos have exposure level L∗ ; (3) the total exposure time (or capture
time) is smaller than τ one ; and (4) this time is a global minimum over all finite
capture sequences.
Intuitively, whenever such a capture sequence exists, it can be thought of as being opti-
mally more efficient than single-shot photography in gathering light. Below we analyze
three instances of the light-efficient photography problem. In all cases, we assume that
the exposure level L∗ , depth of field [α, β], and aperture set D are known and fixed.
Noise Properties. All photos we consider have similar noise, because most noise
sources (photon, sensor, and quantization noise) depend only on exposure level, which
we hold constant. The only exception is thermal noise, which increases with exposure
time [1], and so will be lower for light-efficient sequences with shorter exposures.
4 Theory of Light-Efficient Photography

4.1 Continuously-Variable Aperture Diameters
Many manual-focus SLR lenses allow their aperture diameter to vary continuously
within some interval D = [Dmin , Dmax ]. In this case, we prove that the optimal
capture sequence has an especially simple form – it is unique, it uses the same aperture
diameter for all tuples, and this diameter is either the maximum possible or a diameter
close to that maximum.
More specifically, consider the following special class of capture sequences:
Definition 4 (Sequences with Sequential DOFs). A capture sequence has sequential
DOFs if for every pair of adjacent photo tuples, the right endpoint of the first tuple’s
DOF is the left endpoint of the second.
The following theorem states that the solution to the light-efficient photography prob-
lem is a specific sequence from this class:
Theorem 1 (Optimal Capture √ Sequence for Continuous Apertures). (1) If the DOF
endpoints satisfy β < (7 + 4 3)α, the sequence that globally minimizes total exposure
time is a sequence with sequential DOFs whose tuples all have the same aperture. (2)
Define D(k) and n as follows:
⎢ ⎥
√ √ ⎢ ⎥
k
β+ α k ⎢ log α
⎥
D(k) = c √ √ , n = ⎣
β
⎦ . (3)
k
β− α k
log Dmax −cDmax +c
The aperture diameter D∗ and length n∗ of the optimal sequence is given by

⎧ ⎧
⎨D(n) if D(n) > n ⎨n if D(n)
> n
D∗ = Dmax n+1
n∗ = Dmax n+1
. (4)
⎩D otherwise. ⎩n + 1 otherwise.
max
Theorem 1 specifies the optimal sequence indirectly, via a “recipe” for calculating the
optimal length and the optimal aperture diameter (Eqs. (3) and (4)). Informally, this
calculation involves three steps. The first step defines the quantity D(k); in our proof of
Theorem 1 (see Appendix A), we show that this quantity represents the only aperture
diameter that can be used to “tile” the interval [α, β] with exactly k photo tuples of the
same aperture. The second step defines the quantity n; in our proof, we show that this
represents the largest number of photos we can use to tile the interval [α, β] with photo
tuples of the same aperture. The third step involves choosing between two “candidates”
for the optimal solution – one with n tuples and one with n + 1.
Theorem 1 makes explicit the somewhat counter-intuitive fact that the most light-
efficient way to span a given DOF [α, β] is to use images whose DOFs are very narrow.
This fact applies broadly, because Theorem 1’s inequality condition for α and β is
satisfied for all lenses for consumer photography that we are aware of (e.g., see [22]).2
See Fig. 4 for an application of this theorem to a practical example.
Note that Theorem 1 specifies the number of tuples in the optimal sequence and
their aperture diameter, but does not specify their exposure times or focus settings.
The following lemma shows that specifying those quantities is not necessary because
they are determined uniquely. Importantly, Lemma 1 gives us a recursive formula for
computing the exposure time and focus setting of each tuple in the sequence:
2
To violate the condition, the minimum focusing distance must be under 1.077f , measured
from the lens center.
n
17 1500 1 3 5 7 9 11 13 15
total exposure time (ms)

sequence length, n*
total capture time (ms)

13 1125 1500
s
fp ps
4.
5 10 f
9 750 944
20 fps
675
5 375 500 60 fps
300 no overhead
1 0 100
5.3 25 50 75 92
5.31 15.9 26.6 37.2 47.8 58.4 69.1 79.7
Dmax (mm) D(n) (mm)
(a) (b)
Fig. 4. (a) Optimal light-efficient photography of a “dark” subject spanning a DOF of

[110 cm, 124 cm], using an f = 85 mm lens with a continuously-variable aperture. In this exam-
ple, we can use a f /16 aperture (5.3 mm diameter) to cover the DOF with a single photo, which
requires a 1.5 s exposure to obtain the desired exposure level. The plot illustrates the optimal
sequences when the aperture diameter is restricted to a range [f /16, Dmax ]: for each value of
the maximum aperture, Dmax , Theorem 1 gives a unique optimal sequence. The graph shows the
number of images n∗ (red) and total exposure time (green) of this sequence. As Dmax increases,
the total exposure time of the optimal sequence falls dramatically: for lenses with an f /1.2 max-
imum aperture (71 mm), synthetic DOFs confer a 13× speedup over single-shot photography for
the same exposure level. (b) The effect of camera overhead for various frame-per-second (fps)
rates. Each point represents the total capture time of a sequence that spans the DOF and whose
photos all use the diameter D(n) indicated. Even though overhead reduces the efficiency of long
sequences, synthetic DOFs are faster than single-shot photography even for low fps rates.
Lemma 1 (Construction of Sequences with Sequential DOFs). Given a left DOF

endpoint α, every ordered sequence D1 , . . . , Dn of aperture diameters defines a unique
capture sequence with sequential DOFs whose n tuples are
L ∗ Di + c
Di , , αi , i = 1, . . . , n , (5)
Di 2 Di
with αi given by the following recursive relation:

α if i = 1 ,
αi = (6)
Di +c
α
Di −c i−1 otherwise.
4.2 Discrete Aperture Diameters

Modern auto-focus lenses often restrict the aperture diameter to a discrete set of choices,
D = {D1 , . . . , Dm }. These diameters form a geometric progression, spaced so that the
aperture area doubles every two or three steps. Unlike the continuous case, the optimal
capture sequence is not unique and may contain several distinct aperture diameters. To
find an optimal sequence, we reduce the problem to integer linear programming [23]:
Theorem 2 (Optimal Capture Sequence for Discrete Apertures). There exists an

optimal capture sequence with sequential DOFs whose tuples have a non-decreasing
speedup over single photo

relative DOF size α/β
0.993
frames per second

0.992
0.991
91.1 α focus setting (mm) 92.3 α focus setting (mm) β

(a) (b)
Fig. 5. Optimal light-efficient photography with discrete apertures, shown for a Canon EF85mm
1.2L lens (23 apertures, illustrated in different colors). (a) For a depth of field whose left endpoint
is α, we show optimal capture sequences for a range of relative DOF sizes α β
. These sequences
can be read horizontally, with subintervals corresponding to the apertures determined by Theo-
rem 2. Note that when the DOF is large, the optimal sequence approximates the continuous case.
The diagonal dotted line indicates the DOF to be spanned. (b) Visualizing the optimal capture se-
quence as a function of the camera overhead for the DOF [α, β]. Note that with higher overhead,
the optimal sequence involves fewer photos with larger DOFs (i.e., smaller apertures).
sequence of aperture diameters. Moreover, if ni is the number of times diameter Di

appears in the sequence, the multiplicities n1 , . . . , nm satisfy the integer program
m L∗
minimize i=1 ni Di 2 (7)
m Di −c
i=1 ni log Di +c ≤ log β
α
subject to (8)
ni ≥ 0 and integer . (9)
See [24] for a proof. As with Theorem 1, Theorem 2 does not specify the focus settings
in the optimal capture sequence. We use Lemma 1 for this purpose, which explicitly
constructs it from the apertures and their multiplicities.
While it is not possible to obtain a closed-form expression for the optimal sequence,
solving the integer program for any desired DOF is straightforward. We use a simple
branch-and-bound method based on successive relaxations to linear programming [23].
Moreover, since the optimal sequence depends only on the relative DOF size α β , we
pre-compute it for all possible DOFs and store the results in a lookup table (Fig. 5a).
4.3 Discrete Aperture Diameters Plus Overhead

Our treatment of discrete apertures generalizes easily to account for camera overhead.
We model overhead as a per-shot constant, τ over , that expresses the minimum delay
between the time that the shutter closes and the time it is ready to open again for the next
photo. To find the optimal sequence, we modify the objective function of Theorem 2 so
that it measures total capture time rather than total exposure time:
m ∗
i=1 ni [ τ + DLi 2 ] .
over
minimize (10)
Clearly, a non-negligible overhead penalizes long capture sequences and reduces the
synthetic DOF advantage. Despite this, Fig. 5b shows that synthetic DOFs offer
significant speedups even for current off-the-shelf cameras. These speedups will be
amplified further as camera manufacturers continue to improve frame-per-second rates.
5 Depth of Field Compositing and Resynthesis

DOF Compositing. To reproduce the desired DOF, we use a variant of the Photomon-
tage method [20], based on maximizing a simple “focus measure” that evaluates local
contrast according to the difference-of-Gaussians filter. In this method, each pixel in the
composite has a label that indicates the input photo for which the pixel is in-focus. These
labels are optimized with a Markov random field network that is biased toward piece-
wise smoothness. The resulting composite is a blend of the input photos, performed in
the gradient-domain to reduce artifacts at label boundaries.
3D Reconstruction. The DOF compositing operation produces a coarse depth map as

an intermediate step. This is because labels correspond to input photos, and each input
photo defines an in-focus depth according to the focus setting with which it was cap-
tured. We found this depth map to be sufficient for good-quality resynthesis, although a
more sophisticated depth-from-defocus analysis is also possible [6].
Synthesizing Photos for Novel Focus Settings and Aperture Diameters. To synthe-
size novel photos, we generalize DOF compositing and take advantage of the different
levels of defocus throughout the capture sequence. We proceed in four basic steps. First,
given a specific focus and aperture setting, we use Eq. (C) and the coarse depth map
to assign a blur diameter to each pixel in the final composite. Second, we use Eq. (C)
again to determine, for each pixel in the composite, the input photo whose blur diameter
that corresponds to the pixel’s depth matches most closely.3 Third, for each depth layer,
we synthesize a photo under the assumption that the entire scene is at that depth, and is
observed with the novel focus and aperture setting. To do this, we use the blur diameter
for this depth to define an interpolation between two of the input photos. We currently
interpolate using simple linear cross-fading, which we found to be adequate when the
DOF is sampled densely enough (i.e., with 5 or more images). Fourth, we generate the
final composite by merging all these synthesized images into one photo using the same
gradient-domain blending as in DOF compositing, with the same depth labels.
6 Experimental Results
Figure 6 shows results and timings for two experiments, performed with two differ-
ent cameras – a high-end digital SLR and a compact digital camera (see [24] for more
results and videos). All photos were captured at the same exposure level for each ex-
periment. In each case, we captured (1) a narrow-aperture photo and (2) the optimal
capture sequence for the equivalent DOF and the particular camera. To compensate for
the distortions that occur with changes in focus setting, we align the photos according
3
Note each blur diameter is consistent with two depths (Fig. 3b). We resolve the ambiguity by
choosing the matching input photo whose focus setting is closest to the synthetic focus setting.
Canon S3 IS (6MP)
photo 1 of 4 @ f /2.7 synthetic DOF composite 1 photo @ f /8

exposure time: 50 ms total exposure time: 200 ms exposure time: 500 ms
(a) (b) (c)
Canon Mark II 1Ds (16MP)
photo 3 of 14 @ f /1.2 synthetic DOF composite 1 photo @ f /16

exposure time: 5 ms total exposure time: 70 ms exposure time: 800 ms
(d) (e) (f)
coarse depth map, synthesized f /2.8 aperture, synthesized f /2.8 aperture,

labels from DOF composite same focus setting as (d) refocused further
(g) (h) (i)
Fig. 6. Light-efficient photography timings and synthesis, for several real scenes, captured using
a compact digital camera and a digital SLR. (a,d) Sample wide-aperture photo from the synthetic
DOF sequence. (b,e) DOF composites synthesized from this sequence. (c,f) Narrow-aperture
photos spanning an equivalent DOF, but with much longer exposure time. (g) Coarse depth map,
computed from the labeling we used to compute (e). (h) Synthetically changing aperture size,
focused at the same setting as (d). (i) Synthetically changing the focus setting as well.
to a one-time calibration method that fits a radial magnification model to focus setting
[25]. To determine the maximum acceptable blur diameter c for each camera, we eval-
uated focus using a resolution chart. The values we found, 5 µm (1.4 pixels) and 25 µm
(3.5 pixels) respectively, agree with standard values [21].
Timing Comparisons and Optimal Capture Sequences. To determine the optimal

capture sequences, we assumed zero camera overhead and applied Theorem 2 for the
chosen DOF and exposure level, according to the specifications of each camera and
lens. The optimal sequences involved spanning the DOF using the largest aperture in
both cases. As Fig. 6 shows, these sequences led to significant speedups in exposure
time – 2.5× and 11.9× for the compact digital camera and digital SLR, respectively.
DOF Compositing. Figures 6b and 6e show that despite the availability of just a coarse
depth map, our compositing scheme is able to reproduce high-frequency detail over the
whole DOF without noticeable artifacts, even in the vicinity of depth discontinuities.
Note that while the synthesized photos satisfy our goal of spanning a specific DOF,
objects outside that DOF will appear more defocused than in the corresponding narrow-
aperture photo (e.g., see the background in Figs. 6e–f). While increased background
defocus may be desirable (e.g., for portrait or macro photography), it is also possible to
capture sequences of photos to reproduce arbitrary levels of defocus outside the DOF.
Depth Maps and DOF Compositing. Despite being more efficient to capture, se-
quences with synthetic DOFs provide 3D shape information at no extra acquisition cost
(Fig. 6g). Figures 6h–i show results of using this depth map to compute novel images
whose aperture and focus setting was changed synthetically according to Sect. 5.
Implementation Details. Neither of our cameras provide the ability to control focus
remotely. For our compact camera we used modified firmware that enables scripting
[26], while for our SLR we used a computer-controlled motor to drive the focusing ring
mechanically. Both methods incur high overhead and limit us to about 1 fps.
While light-efficient photography is not practical in this context, it will become in-
creasingly so, as newer cameras begin to provide focus control and to increase frame-
per-second rates. For example, the Canon EOS-1Ds Mark III provides remote focus
control for all Canon EF lenses, and the Casio EX-F1 can capture 60 fps at 6MP.
7 Concluding Remarks
In this paper we studied the use of dense, wide-aperture photo sequences as a light-
efficient alternative to single-shot, narrow-aperture photography. While our emphasis
has been on the underlying theory, we believe our method has great practical potential.
We are currently investigating several extensions to the basic approach. These in-
clude designing light-efficient strategies (1) for spanning arbitrary defocus profiles,
rather than just the DOF; (2) improving efficiency by taking advantage of the camera’s
auto-focus sensor; and (3) operating under a highly-restricted time-budget, for which it
becomes important to weigh the tradeoff between noise and defocus.
References
1. Healey, G.E., Kondepudy, R.: Radiometric CCD camera calibration and noise estimation.
TPAMI 16(3), 267–276 (1994)
2. Hasinoff, S.W., Kutulakos, K.N.: A layer-based restoration framework for variable-aperture
photography. In: Proc. ICCV (2007)
3. Pentland, A.P.: A new sense for depth of field. TPAMI 9(4), 523–531 (1987)
4. Krotkov, E.: Focusing. IJCV 1(3), 223–237 (1987)
5. Hiura, S., Matsuyama, T.: Depth measurement by the multi-focus camera. In: CVPR, pp.
953–959 (1998)
6. Watanabe, M., Nayar, S.K.: Rational filters for passive depth from defocus. IJCV 27(3), 203–
225 (1998)
7. Raskar, R., Agrawal, A., Tumblin, J.: Coded exposure photography: motion deblurring using
fluttered shutter. In: SIGGRAPH, pp. 795–804 (2006)
8. Yuan, L., Sun, J., Quan, L., Shum, H.Y.: Image deblurring with blurred/noisy image pairs.
In: SIGGRAPH (2007)
9. Telleen, J., Sullivan, A., Yee, J., Gunawardane, P., Wang, O., Collins, I., Davis, J.: Synthetic
shutter speed imaging. In: Proc. Eurographics, pp. 591–598 (2007)
10. Farid, H., Simoncelli, E.P.: Range estimation by optical differentiation. JOSA A 15(7), 1777–
1786 (1998)
11. Cathey, W.T., Dowski, E.R.: New paradigm for imaging systems. Applied Optics 41(29),
6080–6092 (2002)
12. Levin, A., Fergus, R., Durand, F., Freeman, W.T.: Image and depth from a conventional
camera with a coded aperture. In: SIGGRAPH (2007)
13. Veeraraghavan, A., Raskar, R., Agrawal, A., Mohan, A., Tumblin, J.: Dappled photography:
Mask enhanced cameras for heterodyned light fields and coded aperture refocusing. In: SIG-
GRAPH (2007)
14. Aizawa, K., Kodama, K., Kubota, A.: Producing object-based special effects by fusing mul-
tiple differently focused images. In: TCSVT 10(2) (2000)
15. Chaudhuri, S.: Defocus morphing in real aperture images. JOSA A 22(11), 2357–2365
(2005)
16. Hasinoff, S.W., Kutulakos, K.N.: Confocal stereo. In: Leonardis, A., Bischof, H., Pinz, A.
(eds.) ECCV 2006. LNCS, vol. 3951, pp. 620–634. Springer, Heidelberg (2006)
17. Ng, R.: Fourier slice photography. In: SIGGRAPH, pp. 735–744 (2005)
18. Levoy, M., Ng, R., Adams, A., Footer, M., Horowitz, M.: Light field microscopy. In: SIG-
GRAPH, pp. 924–934 (2006)
19. Debevec, P., Malik, J.: Recovering high dynamic range radiance maps from photographs. In:
SIGGRAPH, pp. 369–378 (1997)
20. Agarwala, A., Dontcheva, M., Agrawala, M., Drucker, S., Colburn, A., Curless, B., Salesin,
D., Cohen, M.: Interactive digital photomontage. In: SIGGRAPH, pp. 294–302 (2004)
21. Smith, W.J.: Modern Optical Engineering, 3rd edn. McGraw-Hill, New York (2000)
22. Canon lens chart, http://www.usa.canon.com/app/pdf/lens/
23. Nocedal, J., Wright, S.J.: Numerical Optimization. Springer, Heidelberg (1999)
24. http://www.cs.toronto.edu/∼kyros/research/lightefficient/
25. Willson, R., Shafer, S.: What is the center of the image? JOSA A 11(11), 2946–2955 (1994)
26. CHDK, http://chdk.wikia.com/
A Proof of Theorem 1
Theorem 1 follows as a consequence of Lemma 1 and four additional lemmas. We first
state Lemmas 2–5 below and then prove a subset of them, along with a proof sketch the
theorem. All missing proofs can be found in [24].
Lemma 2 (Efficiency of Sequential DOFs). For every sequence S, there is a sequence

S with sequential DOFs that spans the same synthetic DOF and whose total exposure
time is no larger.
Lemma 3 (Permutation of Sequential DOFs). Given the left endpoint, α, every per-
mutation of D1 , . . . , Dn defines a capture sequence with sequential DOFs that has the
same synthetic DOF and the same total exposure time.
Lemma 4 (Optimality of Maximizing the Number of Photos). Among all sequences

with up to n tuples whose synthetic DOF is [α, β], the sequence that minimizes total
exposure time has exactly n of them.
√
Lemma 5 (Optimality of Equal-Aperture Sequences). If β < (7 + 4 3)α, then
among all capture sequences with n tuples whose synthetic DOF is [α, β], the sequence
that minimizes total exposure time uses the same aperture for all tuples. Furthermore,
this aperture is equal to √ √
n
β+ n α
D(n) = c √ n √
β− n α
. (11)
Proof of Lemma 1. We proceed inductively, by defining photo tuples whose DOFs

“tile” the interval [α, β] from left to right. For the base case, the left endpoint of the first
tuple’s DOF must be α1 = α. Now consider the i-th tuple. Equation (D) implies that
the left endpoint αi and the aperture diameter Di determine the DOF’s right endpoint
uniquely:
βi = D Di −c αi .
i +c
(12)
The tuple’s focus setting in Eq. (5) now follows by applying Eq. (E) to the interval
[αi , βi ]. Finally, since the DOFs of tuple i and i + 1 are sequential, we have αi+1 = βi .

Proof of Lemma 4. From Lemma 2 it follows that among all sequences up to length n
whose DOF is [α, β], there is a sequence S ∗ with minimum total exposure time whose
tuples have sequential DOFs. Furthermore, Lemmas 1 and 3 imply that this capture
sequence is fully determined by a sequence of n aperture settings, D1 ≤ D2 ≤ · · · ≤
Dn , for some n ≤ n. These settings partition the interval [α, β] into n sub-intervals,
whose endpoints are given by Eq. (6):
determined by S ∗

α = α1 < α2 < · · · < αn < βn = β . (13)
It therefore suffices to show that placing n − 1 points in [α, β] is most efficient when
n = n. To do this, we show that splitting a sub-interval always produces a more effi-
cient capture sequence.
Consider the case n = 2, where the sub-interval to be split is actually equal to [α, β].
Let x ∈ [α, β] be a splitting point. The exposure time for the sub-intervals [α, x] and
[x, β] can be obtained by combining Eqs. (D) and (1):

2
2
τ (x) = L
c2
x−α
x+α + L
c2
β−x
β+x , (14)
Differentiating Eq. (14) and evaluating it for x = α we obtain

dτ (β−α) β
dx = − 4L c2 (β+α)3 < 0 . (15)
x=α
dx is positive for x = β. Since τ (x) is continuous

Similarly, it is possible to show that dτ
in [α, β], it follows that the minimum of τ (x) occurs strictly inside the interval. Hence,
splitting the interval always reduces total exposure time. The general case for n intervals
follows by induction.

Proof Sketch of Theorem 1. We proceed in four steps. First, we consider sequences
whose synthetic DOF is equal to [α, β]. From Lemmas 4 and 5 it follows that the most
efficient sequence, S , among this set has diameter and length given by Eq. (3). Second,
we show that sequences with a larger synthetic DOF that are potentially more efficient
can have at most one more tuple. Third, we show that the most efficient of these se-
quences, S , uses a single diameter equal to Dmax . Finally, the decision rule in Eq. (4)
follows by comparing the total exposure times of S and S .
Flexible Depth of Field Photography
Hajime Nagahara1,2, Sujit Kuthirummal2 ,

Changyin Zhou2 , and Shree K. Nayar2
1
Osaka University
2
Columbia University
Abstract. The range of scene depths that appear focused in an image is

known as the depth of field (DOF). Conventional cameras are limited by
a fundamental trade-off between depth of field and signal-to-noise ratio
(SNR). For a dark scene, the aperture of the lens must be opened up to
maintain SNR, which causes the DOF to reduce. Also, today’s cameras
have DOFs that correspond to a single slab that is perpendicular to the
optical axis. In this paper, we present an imaging system that enables
one to control the DOF in new and powerful ways. Our approach is to
vary the position and/or orientation of the image detector, during the
integration time of a single photograph. Even when the detector motion
is very small (tens of microns), a large range of scene depths (several
meters) is captured both in and out of focus.
Our prototype camera uses a micro-actuator to translate the detec-
tor along the optical axis during image integration. Using this device,
we demonstrate three applications of flexible DOF. First, we describe
extended DOF, where a large depth range is captured with a very wide
aperture (low noise) but with nearly depth-independent defocus blur. Ap-
plying deconvolution to a captured image gives an image with extended
DOF and yet high SNR. Next, we show the capture of images with dis-
continuous DOFs. For instance, near and far objects can be imaged with
sharpness while objects in between are severely blurred. Finally, we show
that our camera can capture images with tilted DOFs (Scheimpflug imag-
ing) without tilting the image detector. We believe flexible DOF imaging
can open a new creative dimension in photography and lead to new ca-
pabilities in scientific imaging, vision, and graphics.
1 Depth of Field
The depth of field (DOF) of an imaging system is the range of scene depths that
appear focused in an image. In virtually all applications of imaging, ranging
from consumer photography to optical microscopy, it is desirable to control the
DOF. Of particular interest is the ability to capture scenes with very large DOFs.
DOF can be increased by making the aperture smaller. However, this reduces the
amount of light received by the detector, resulting in greater image noise (lower

Parts of this work were supported by grants from the National Science Foundation
(IIS-04-12759) and the Office of Naval Research (N00014-08-1-0329 and N00014-06-
1-0032.)

Flexible Depth of Field Photography 61
SNR). This trade-off gets worse with increase in spatial resolution (decrease in
pixel size). As pixels get smaller, DOF decreases since the defocus blur occupies
a greater number of pixels. At the same time, each pixel receives less light and
hence SNR falls as well. This trade-off between DOF and SNR is one of the
fundamental, long-standing limitations of imaging.
In a conventional camera, for any location of the image detector, there is
one scene plane – the focal plane – that is perfectly focused. In this paper, we
propose varying the position and/or orientation of the image detector during the
integration time of a photograph. As a result, the focal plane is swept through a
volume of the scene causing all points within it to come into and go out of focus,
while the detector collects photons.
We demonstrate that such an imaging system enables one to control the DOF
in new and powerful ways:
• Extended Depth of Field. Consider the case where a detector with a

global shutter (all pixels are exposed simultaneously and for the same du-
ration) is moved with uniform speed during image integration. Then, each
scene point is captured under a continuous range of focus settings, including
perfect focus. We analyze the resulting defocus blur kernel and show that it is
nearly constant over the range of depths that the focal plane sweeps through
during detector motion. Consequently, irrespective of the complexity of the
scene, the captured image can be deconvolved with a single, known blur
kernel to recover an image with significantly greater DOF. This approach is
similar in spirit to Hausler’s work in microscopy [1]. He showed that the DOF
of an optical microscope can be enhanced by moving a specimen of depth
range d, a distance 2d along the optical axis of the microscope, while filming
the specimen. The defocus of the resulting captured image is similar over
the entire depth range of the specimen. However, this approach of moving
the scene with respect to the imaging system is practical only in microscopy
and not suitable for general scenes. More importantly, Hausler’s derivation
assumes that defocus blur varies linearly with scene depth which is true only
for the small distances involved in microscopy.
• Discontinuous Depth of Field. A conventional camera’s DOF is a sin-
gle fronto-parallel slab located around the focal plane. We show that by
moving a global-shutter detector non-uniformly, we can capture images that
are focused for certain specified scene depths, but defocused for in-between
scene regions. Consider a scene that includes a person in the foreground, a
landscape in the background, and a dirty window in between the two. By
focusing the detector on the nearby person for some duration and the far
away landscape for the rest of the integration time, we get an image in which
both appear fairly well-focused, while the dirty window is blurred out and
hence optically erased.
• Tilted Depth of Field. Most cameras can only focus on a fronto-parallel
plane. An exception is the view camera configuration [2,3], where the image
detector is tilted with respect to the lens. When this is done, the focal plane
is tilted according to the well-known Scheimpflug condition [4]. We show
62 H. Nagahara et al.
that by uniformly translating an image detector with a rolling electronic

shutter (different rows are exposed at different time intervals but for the
same duration), we emulate a tilted image detector. As a result, we capture
an image with a tilted focal plane. Furthermore, by translating the image
detector non-uniformly (varying speed), we can emulate a non-planar image
detector. This allows us to focus on curved surfaces in the scene.
An important feature of our approach is that the focal plane of the camera can
be swept through a large range of scene depths with a very small translation
of the image detector. For instance, with a 12.5 mm focal length lens, to sweep
the focal plane from a distance of 450 mm from the lens to infinity, the detector
has to be translated only about 360 microns. Since a detector only weighs a few
milligrams, a variety of micro-actuators (solenoids, piezoelectric stacks, ultra-
sonic transducers, DC motors) can be used to move it over the required distance
within very short integration times (less than a millisecond if required). Note
that such micro-actuators are already used in most consumer cameras for focus
and aperture control and for lens stabilization. We present several results that
demonstrate the flexibility of our system to control DOF in unusual ways. We
believe our approach can open up a new creative dimension in photography and
lead to new capabilities in scientific imaging, computer vision, and computer
graphics.
2 Related Work
A promising approach to extended DOF imaging is wavefront coding, where
phase plates placed at the aperture of the lens cause scene objects within a cer-
tain depth range to be defocused in the same way [5,6,7]. Thus, by deconvolving
the captured image with a single blur kernel, one can obtain an all-focused im-
age. In this case, the effective DOF is determined by the phase plate used and is
fixed. On the other hand, in our system, the DOF can be chosen by controlling
the motion of the detector. Our approach has greater flexibility as it can even
be used to achieve discontinuous or tilted DOFs.
Recently, Levin et al. [8] and Veeraraghavan et al. [9] have used masks at
the lens aperture to control the properties of the defocus blur kernel. From a
single captured photograph, they aim to estimate the structure of the scene
and then use the corresponding depth-dependent blur kernels to deconvolve the
image and get an all-focused image. However, they assume simple layered scenes
and their depth recovery is not robust. In contrast, our approach is not geared
towards depth recovery, but can significantly extend DOF irrespective of scene
complexity. Also, the masks used in both these previous works attenuate some
of the light entering the lens, while our system operates with a clear and wide
aperture. All-focused images can also be computed from an image captured
using integral photography [10,11,12]. However, since these cameras make spatio-
angular resolution trade-offs to capture 4D lightfields in a single image, the
computed images have much lower spatial resolutions when compared to our
approach.
Aperture Lens Detector Motion

Scene Point
M
Scene
m m b
u v
p Integration Time
(a) (b)
Fig. 1. (a) A scene point M , at a distance u from the lens, is imaged in perfect focus
by a detector at a distance v from the lens. If the detector is shifted to a distance p
from the lens, M is imaged as a blurred circle with diameter b centered around m . (b)
Our flexible DOF camera translates the detector along the optical axis during the inte-
gration time of an image. By controlling the starting position, speed, and acceleration
of the detector, we can manipulate the DOF in powerful ways.
A related approach is to capture many images to form a focal stack [13,14,15].

An all-in-focus image as well as scene depth can be computed from a focal stack.
However, the need to acquire multiple images increases the total capture time
making the method suitable for only quasi-static scenes. An alternative is to
use very small exposures for the individual images. However, in addition to
the practical problems involved in reading out the many images quickly, this
approach would result in under-exposed and noisy images that are unsuitable
for depth recovery. Our approach does not recover scene depth, but can produce
an all-in-focus photograph from a single, well-exposed image.
There is similar parallel work on moving the detector during image integration
[16]. However, their focus is on handling motion blur, for which they propose
to move the detector perpendicular to the optical axis. Some previous works
have also varied the orientation or location of the image detector. Krishnan
and Ahuja [3] tilt the detector and capture a panoramic image sequence, from
which they compute an all-focused panorama and a depth map. For video super-
resolution, Ben-Ezra et al. [17] capture a video sequence by instantaneously
shifting the detector within the image plane, in between the integration periods
of successive video frames.
Recently, it has been shown that a detector with a rolling shutter can be used
to estimate the pose and velocity of a fast moving object [18]. We show how such
a detector can be used to focus on tilted scene planes.
3 Camera with Programmable Depth of Field

Consider Figure 1(a), where the detector is at a distance v from a lens with focal
length f and an aperture of diameter a. A scene point M is imaged in perfect
focus at m, if its distance u from the lens satisfies the Gaussian lens law:
1 1 1
= + . (1)
f u v
As shown in the figure, if the detector is shifted to a distance p from the lens
(dotted line), M is imaged as a blurred circle (the circle of confusion) centered
around m . The diameter b of this circle is given by
a
b = |(v − p)| . (2)
v
The distribution of light energy within the blur circle is referred to as the
point spread function (PSF). The PSF can be denoted as P (r, u, p), where r is
the distance of an image point from the center m of the blur circle. An idealized
model for characterizing the PSF is the pillbox function:
4 r
P (r, u, p) = Π( ), (3)
πb2 b
where, Π(x) is the rectangle function, which has a value 1, if |x| < 1/2 and
0 otherwise. In the presence of optical aberrations, the PSF deviates from the
pillbox function and is then often modeled as a Gaussian function:
2 2r2
P (r, u, p) = exp(− ), (4)
π(gb)2 (gb)2
where g is a constant.
We now analyze the effect of moving the detector during an image’s integration
time. For simplicity, consider the case where the detector is translated along the
optical axis, as in Figure 1(b). Let p(t) denote the detector’s distance from the
lens as a function of time. Then the aggregate PSF for a scene point at a distance
u from the lens, referred to as the integrated PSF (IPSF), is given by
T
IP (r, u) = P (r, u, p(t)) dt, (5)
0
where T is the total integration time. By programming the detector motion p(t)–
its starting position, speed, and acceleration – we can change the properties of
the resulting IPSF. This corresponds to sweeping the focal plane through the
scene in different ways. The above analysis only considers the translation of
the detector along the optical axis (as implemented in our prototype camera).
However, this analysis can be easily extended to more general detector motions,
where both its position and orientation are varied during image integration.
Figure 2(a) shows our flexible DOF camera. It consists of a 1/3” Sony CCD
(with 1024x768 pixels) mounted on a Physik Instrumente M-111.1DG transla-
tion stage. This stage has a DC motor actuator that can translate the detector
through a 15 mm range at a top speed of 2.7 mm/sec and can position it with
an accuracy of 0.05 microns. The translation direction is along the optical axis
of the lens. The CCD shown has a global shutter and was used to implement ex-
tended DOF and discontinuous DOF. For realizing tilted DOFs, we used a 1/2.5”
Micron CMOS detector (with 2592x1944 pixels) which has a rolling shutter.
Translation Lens Scene Required Maximum

Lens Focal Depth Detector Change in
Length Range Translation Image Position
Micro-actuator 1m - ∞ 81.7 μm 4.5 pixels
9.0mm .5m - ∞ 164.9 μm 5.0 pixels
.2m - 0.5m 259.1 μm 7.2 pixels
1m - ∞ 158.2 μm 3.6 pixels
Image Detector 12.5mm .5m - ∞ 320.5 μm 5.6 pixels
.2m - 0.5m 512.8 μm 8.5 pixels
(a) (b)
Fig. 2. (a) Prototype system with flexible DOF. (b) Translation of the detector re-
quired for sweeping the focal plane through different scene depth ranges. The maxi-
mum change in the image position of a scene point that results from this translation,
when a 1024x768 pixel detector is used, is also shown.
The table in Figure 2(b) shows detector translations (third column) required
to sweep the focal plane through various depth ranges (second column), using
lenses with two different focal lengths (first column). As we can see, the detector
has to be moved by very small distances to sweep very large depth ranges. Us-
ing commercially available micro-actuators, such translations are easily achieved
within typical image integration times (a few milliseconds to a few seconds).
It must be noted that when the detector is translated, the magnification of
the imaging system changes. The fourth column of the table in Figure 2(b) lists
the maximum change in the image position of a scene point for different trans-
lations of a 1024x768 pixel detector. For the detector motions we require, these
changes in magnification are very small. This does result in the images not being
perspectively correct, but the distortions are imperceptible. More importantly,
the IPSFs are not significantly affected by such a magnification change, since a
scene point will be in high focus only for a small fraction of this change and will
be highly blurred over the rest of it. We verify this in the next section.
4 Extended Depth of Field (EDOF)

In this section, we show that we can capture scenes with EDOF by translating
a detector with a global shutter at a constant speed during image integration.
We first show that the IPSF for an EDOF camera is nearly invariant to scene
depth for all depths swept by the focal plane. As a result, we can deconvolve the
captured image with the IPSF to obtain an image with EDOF and high SNR.
4.1 Depth Invariance of IPSF

Consider a detector translating along the optical axis with constant speed s, i.e.,
p(t) = p(0) + st. If we assume that the PSF of the lens can be modeled using
the pillbox function in Equation 3, the IPSF in Equation 5 simplifies to
1.000 1.000
750mm 0.08 2000mm 750mm 0.06 2000mm
9.996 550mm 450mm 450mm
9.996 550mm
0.012 1100mm 0.06 550mm 550mm
0.008 1100mm 0.04
0.008 2000mm 0.04 1100mm 1100mm
750mm 2000mm 750mm
450mm 0.004 0.02
0.004 0.02 450mm
0 0 0 0
-10 0 10 -10 0 10 -10 0 10 -10 0 10
(a) Normal Camera (b) EDOF Camera (c) Normal Camera (d) EDOF Camera
PSF (Pillbox) IPSF PSF (Gaussian) IPSF
Fig. 3. Simulated (a,c) normal camera PSFs and (b,d) EDOF camera IPSFs, obtained
using pillbox and Gaussian lens PSF models for 5 scene depths. Note that the IPSFs
are almost invariant to scene depth.

uf λ0 + λT 2λ0 2λT
IP (r, u) = − − , (6)
(u − f )πasT r b(0) b(T )
where, b(t) is the blur circle diameter at time t, and λt = 1 if b(t) ≥ 2r and 0
otherwise. On the other hand, if we use the Gaussian function in Equation 4 for
the lens PSF, we get

uf r r
IP (r, u) = √ erfc √ + erfc √ . (7)
(u − f ) 2πrasT 2gb(0) 2gb(T )
Figures 3(a) and (c) show 1D profiles of a normal camera’s PSFs for 5 scene
points with depths between 450 and 2000 mm from a lens with focal length
f = 12.5 mm and f /# = 1.4, computed using Equations 3 and 4 (with g = 1),
respectively. In this simulation, the normal camera was focused at a distance
of 750 mm. Figures 3(b) and (d) show the corresponding IPSFs of an EDOF
camera with the same lens, p(0) = 12.5 mm, s = 1 mm/sec, and T = 360
msec, computed using Equations 6 and 7, respectively. As expected, the normal
camera’s PSF varies dramatically with scene depth. In contrast, the IPSFs of
the EDOF camera derived using both pillbox and Gaussian PSF models look
almost identical for all 5 scene depths, i.e., the IPSFs are depth invariant.
To verify this empirical observation, we measured a normal camera’s PSFs and
the EDOF camera’s IPSFs for several scene depths, by capturing images of small
dots placed at different depths. Both cameras have f = 12.5 mm, f /# = 1.4,
and T = 360 msec. The detector motion parameters for the EDOF camera are
p(0) = 12.5 mm and s = 1 mm/sec. The first column of Figure 4 shows the
measured PSF at the center pixel of the normal camera for 5 different scene
depths; the camera was focused at a distance of 750 mm. (Note that the scale
of the plot in the center row is 50 times that of the other plots.) Columns 2-4
of the figure show the IPSFs of the EDOF camera for 5 different scene depths
and 3 different image locations. We can see that, while the normal camera’s
PSFs vary widely with scene depth, the EDOF camera’s IPSFs appear almost
invariant to both spatial location and scene depth. This also validates our claim
that the small magnification changes that arise due to detector motion (discussed
in Section 3) do not have a significant impact on the IPSFs.
Normal Camera PSF EDOF Camera IPSF

0.04 0.04 0.04 0.04
450mm
0.02 0.02 0.02 0.02
0 0 0 0
-10 0 10 -10 0 10 -10 0 10 -10 0 10
0.04 0.04 0.04 0.04
550mm
0.02 0.02 0.02 0.02
0 0 0 0
-10 0 10 -10 0 10 -10 0 10 -10 0 10
50x 0.04
Scene Depth 0.04 0.04 0.04
750mm
50x 0.02 0.02 0.02 0.02
0 0 0 0
-10 0 10 -10 0 10 -10 0 10 -10 0 10
0.04 0.04 0.04 0.04
1100mm
0.02 0.02 0.02 0.02
0 0 0 0
-10 0 10 -10 0 10 -10 0 10 -10 0 10
0.04 0.04 0.04 0.04
2000mm
0.02 0.02 0.02 0.02
0 0 0 0
-10 0 10 -10 0 10 -10 0 10 -10 0 10
(0,0) pix. (0,0) pix. (212,0) pix. (424,0) pix.
Center Center
Image Location (x,y)
Fig. 4. (Left column) The measured PSF of a normal camera shown for 5 different
scene depths. Note that the scale of the plot in the center row is 50 times that of
the other plots. (Right columns) The measured IPSF of our EDOF camera shown for
different scene depths (vertical axis) and image locations (horizontal axis). The EDOF
camera’s IPSFs are almost invariant to scene depth and image location.
4.2 Computing EDOF Images Using Deconvolution
Since the EDOF camera’s IPSF is invariant to scene depth and image location,
we can deconvolve a captured image with a single IPSF to get an image with
greater DOF. A number of techniques have been proposed for deconvolution,
Richardson-Lucy and Wiener [19] being two popular ones. For our results, we
have used the approach of Dabov et al. [20], which combines Wiener deconvolu-
tion and block-based denoising. In all our experiments, we used the IPSF shown
in the first row and second column of Figure 4 for deconvolution.
Figure 5(a) shows an image captured by our EDOF camera. It was captured
with a 12.5 mm Fujinon lens with f /1.4 and 0.36 second exposure. Notice that
the captured image looks slightly blurry, but high frequencies of all scene ele-
ments are captured. This scene spans a depth range of approximately 450 mm
to 2000 mm – 10 times larger than the DOF of a normal camera with identical
lens settings. Figure 5(b) shows the EDOF image computed from the captured
image, in which the entire scene appears focused1 . Figure 5(c) shows the image
1
Mild ringing artifacts in the computed EDOF images are due to deconvolution.
captured by a normal camera with identical f /# and exposure time. The near-
est scene elements are in focus, while the farther scene elements are severely
blurred. The image captured by a normal camera with the same exposure time,
but with a smaller aperture (f /8) is shown in Figure 5(d). The intensities of this
image were scaled up so that its dynamic range matches that of the correspond-
ing computed EDOF image. All scene elements look reasonably sharp, but the
image is very noisy as can be seen in the inset (zoomed). The computed EDOF
image has much less noise, while having comparable sharpness. Figures 5(e-h)
show another example, of a scene captured outdoors at night. As we can see, in a
normal camera, the tradeoff between DOF and SNR is extreme for such dimly lit
scenes. In short, our EDOF camera can capture scenes with large DOFs as well
as high SNR. High resolution versions of these images as well as other examples
can be seen at [21].
4.3 Analysis of SNR Benefits of EDOF Camera

Deconvolution using Dabov et al.’s method [20] produces visually appealing re-
sults, but since it has a non-linear denoising step, it is not suitable for analyzing
the SNR of deconvolved captured images. Therefore, we performed a simulation
that uses Wiener deconvolution [19]. Given an IPSF k, we convolve it with a nat-
ural image I, and add zero-mean white Gaussian noise with standard deviation
σ. The resulting image is then deconvolved with k to get the EDOF image I. ˆ
The standard deviation σ̂ of (I − I)
ˆ is a measure of the noise in the deconvolution
result when the captured image has noise σ.
The degree to which deconvolution amplifies noise depends on how much
the high frequencies are attenuated by the IPSF. This, in turn, depends on the
distance through which the detector moves during image integration – as the dis-
tance increases, so does the attenuation of high frequencies. This is illustrated in
Figure 6(a), which shows (in red) the MTF (magnitude of the Fourier transform)
for a simulated IPSF k1 , derived using the pillbox lens PSF model. In this case,
we use the same detector translation (and other parameters) as in our EDOF
experiments (Section 4.2). The MTF of the IPSF k2 obtained when the detector
translation is halved (keeping the mid-point of the translation the same) is also
shown (in blue). As expected, k2 attenuates the high frequencies less than k1 .
We analyzed the SNR benefits for these two IPSFs for different noise levels
in the captured image. The table in Figure 6(b) shows the noise produced by a
normal camera for different aperture sizes, given the noise level for the largest
aperture, f /1.4. (Image brightness is assumed to lie between 0 and 1.) The last
two rows show the effective noise levels for EDOF cameras with IPSFs k1 and
k2 , respectively. The last column of the table shows the effective DOFs realized;
the normal camera is assumed to be focused at a scene distance that corresponds
to the center position of the detector motion. One can see that, as the noise level
in the captured image increases, the SNR benefits of EDOF cameras increase.
As an example, if the noise of a normal camera at f /1.4 is 0.01, then the EDOF
camera with IPSF k1 has the SNR of a normal camera with f /2.8, but produces
the DOF of a normal camera with f /8.
(a) Captured Image ( f / 1.4, T =0.36sec) (b) Computed EDOF Image
(c) Image from Normal Camera (d) Image from Normal Camera
(f / 1. .4, T =0.36sec, Near Focus) ( f /8, T =0.36sec, Near Focus) with Scaling
(e) Captured Image ( f / 1.4, T=0.72sec) (f) Computed EDOF Image
(g) Image from Normal Camera (h) Image from Normal Camera
( f / 1.4, T =0.72sec, Near Focus) ( f /8, T =0.72sec, Near Focus) with Scaling
Fig. 5. (a,e) Images captured by the EDOF camera. (b,f) EDOF images computed
from images in (a) and (e), respectively. Note that the entire scene appears focused.
(c,g) Images captured by a normal camera with identical settings, with the nearest
object in focus. (d,h) Images captured by a normal camera at f /8.
1.0
Camera f /# Noise standard deviation DOF (mm)
IPSF k 1
0.8 IPSF k 2 Normal 1.4 0.001 0.002 0.005 0.010 0.020 140.98
0.6 Normal 2.8 0.004 0.008 0.020 0.040 0.080 289.57
MTF
Normal 4 0.008 0.016 0.041 0.082 0.163 429.77

0.4
Normal 5.6 0.016 0.032 0.080 0.160 0.320 649.21
0.2 Normal 8 0.033 0.065 0.163 0.327 0.653 1114.56
0 EDOF(k1 ) 1.4 0.013 0.021 0.036 0.049 0.062 1622.44
0 0.2 0.4 0.6 0.8 1.0 EDOF(k2 ) 1.4 0.0073 0.0133 0.0262 0.0399 0.0570 528.56
Spatial frequency [cycle/pixel]
(a) (b)
Fig. 6. (a) MTFs of simulated IPSFs, k1 and k2 , of an EDOF camera corresponding to

the detector traveling two different distances during image integration. (b) Comparison
of effective noise and DOF of a normal camera and a EDOF camera with IPSFs k1 and
k2 . The image noise of a normal camera operating at f /1.4 is assumed to be given.
In the above analysis, the SNR was averaged over all frequencies. However,
it must be noted that SNR is frequency dependent - SNR is greater for lower
frequencies than for higher frequencies in the deconvolved EDOF images. Hence,
high frequencies in an EDOF image would be degraded, compared to the high
frequencies in a perfectly focused image. However, in our experiments this degra-
dation is not strong, as can be seen in the full resolution images at [21].
5 Discontinuous Depth of Field
Consider the image in Figure 7(a), which shows two toys (cow and hen) in front
of a scenic backdrop with a wire mesh in between. A normal camera with a small
DOF can capture either the toys or the backdrop in focus, while eliminating the
mesh via defocusing. However, since its DOF is a single continuous volume, it
cannot capture both the toys and the backdrop in focus and at the same time
eliminate the mesh. If we use a large aperture and program our camera’s detector
motion such that it first focuses on the toys for a part of the integration time,
and then moves quickly to another location to focus on the backdrop for the
remaining integration time, we obtain the image in Figure 7(b). While this image
includes some blurring, it captures the high frequencies in two disconnected
DOFs - the foreground and the background - but almost completely eliminates
the wire mesh in between. This is achieved without any post-processing. Note
that we are not limited to two disconnected DOFs; by pausing the detector at
several locations during image integration, more complex DOFs can be realized.
6 Tilted Depth of Field
Normal cameras can focus on only fronto-parallel scene planes. On the other
hand, view cameras [2,3] can be made to focus on tilted scene planes by adjusting
the orientation of the lens with respect to the detector. We show that our flexible
(a) Image from Normal Camera (f /11) (b) Image from Our Camera (f /1.4)
Fig. 7. (a) An image captured by a normal camera with a large DOF. (b) An image
captured by our flexible DOF camera, where the toy cow and hen in the foreground
and the landscape in the background appear focused, while the wire mesh in between
is optically erased via defocusing.
(a) Image from Normal Camera (b) Image from our Camera
(f /1.4, T =0.03sec) (f /1.4, T =0.03sec)
Fig. 8. (a) An image captured by a normal camera of a table top inclined at 53◦ with
respect to the lens plane. (b) An image captured by our flexible DOF camera, where
the DOF is tilted by 53◦ . The entire table top (with the newspaper and keys) appears
focused. Observe that the top of the mug is defocused, but the bottom appears focused,
illustrating that the focal plane is aligned with the table top. Three scene regions of
both the images are shown at a higher resolution to highlight the defocus effects.
DOF camera can be programmed to focus on tilted scene planes by simply

translating (as in the previous applications) a detector with a rolling electronic
shutter. A large fraction of CMOS detectors are of this type – while all pixels
have the same integration time, successive rows of pixels are exposed with a
slight time lag. When such a detector is translated with uniform speed s, during
the frame read out time T of an image, we emulate a tilted image detector. If
this tilted detector makes an angle θ with the lens plane, then the focal plane
in the scene makes an angle φ with the lens plane, where θ and φ are related by
the well-known Scheimpflug condition [4]:

sT 2f tan(θ)
θ = tan−1 ( ) and, φ = tan−1 . (8)
H 2p(0) + H tan(θ) − 2f
Here, H is the height of the detector. Therefore, by controlling the speed s of
the detector, we can vary the tilt angle of the image detector, and hence the tilt
of the focal plane and its associated DOF.
Figure 8 shows a scene where the dominant scene plane – a table top with
a newspaper, keys and a mug on it – is inclined at an angle of approximately
53◦ with the lens plane. As a result, a normal camera is unable to focus on the
entire plane, as seen from Figure 8(a). By translating a rolling-shutter detector
(1/2.5” CMOS sensor with a 70msec exposure lag between the first and last row
of pixels) at 2.7 mm/sec, we emulate a detector tilt of 2.6◦ . This enables us to
achieve the desired DOF tilt of 53◦ (from Equation 8) and capture the table top
(with the newspaper and keys) in focus, as shown in Figure 8(b). Observe that
the top of the mug is not in focus, but the bottom appears focused, illustrating
the fact that the DOF is tilted to be aligned with the table top. It is interesting
to note that, by translating the detector with varying speed, we can emulate
non-planar detectors, that can focus on curved scene surfaces.
7 Discussion
In this paper we have proposed a camera with a flexible DOF. DOF is manip-
ulated in various ways by changing the position of the detector during image
integration. We have shown how such a system can capture arbitrarily com-
plex scenes with extended DOF and high SNR. We have also shown that we
can create DOFs that span multiple disconnected volumes. In addition, we have
demonstrated that our camera can focus on tilted scene planes. All of these func-
tionalities are achieved by simply controlling the motion of the detector during
the exposure of a single image.
While computing images with extended DOF, we have not explicitly modeled
occlusions at depth discontinuities or motion blur caused by object/camera mo-
tion. Due to defocus blur, images points that lie close to occlusion boundaries
can receive light from scene points at very different depths. However, since the
IPSF of the EDOF camera is nearly depth invariant, the aggregate IPSF for
such an image point can be expected to be similar to the IPSF of points far
from occlusion boundaries. With respect to motion blur, we have not observed
any visible artifacts in EDOF images computed for scenes with typical object
motion (see Figure 5). However, motion blur due to high-speed objects can be
expected to cause problems. In this case, a single pixel sees multiple objects with
possibly different depths. It is possible that neither of the objects are imaged
in perfect focus during detector translation. This scenario is an interesting one
that warrants further study.
In addition to the DOF manipulations shown in this paper, we have (a) cap-
tured extended DOF video by moving the detector forward one frame, backward
the next, and so on (the IPSF is invariant to the direction of motion), (b) cap-
tured scenes with non-planar DOFs, and (c) exploited the camera’s focusing
mechanism to capture extended DOF by manually rotating a SLR camera lens’

focus ring during image integration. For lack of space, we have not included these
results here; they can be seen at [21].
References
1. Hausler, G.: A Method to Increase the Depth of Focus by Two Step Image Pro-
cessing. Optics Communications, 38–42 (1972)
2. Merklinger, H.: Focusing the View Camera (1996)
3. Krishnan, A., Ahuja, N.: Range estimation from focus using a non-frontal imaging
camera. IJCV, 169–185 (1996)
4. Scheimpflug, T.: Improved Method and Apparatus for the Systematic Alteration
or Distortion of Plane Pictures and Images by Means of Lenses and Mirrors for
Photography and for other purposes. GB Patent (1904)
5. Dowski, E.R., Cathey, W.T.: Extended Depth of Field Through Wavefront Coding.
Applied Optics, 1859–1866 (1995)
6. George, N., Chi, W.: Extended depth of field using a logarithmic asphere. Journal
of Optics A: Pure and Applied Optics, 157–163 (2003)
7. Castro, A., Ojeda-Castaneda, J.: Asymmetric Phase Masks for Extended Depth of
Field. Applied Optics, 3474–3479 (2004)
8. Levin, A., Fergus, R., Durand, F., Freeman, B.: Image and depth from a conven-
tional camera with a coded aperture. SIGGRAPH (2007)
9. Veeraraghavan, A., Raskar, R., Agrawal, A., Mohan, A., Tumblin, J.: Dappled pho-
tography: mask enhanced cameras for heterodyned light fields and coded aperture.
SIGGRAPH (2007)
10. Adelson, E., Wang, J.: Single lens stereo with a plenoptic camera. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence, 99–106 (1992)
11. Ng, R., Levoy, M., Brdif, M., Duval, G., Horowitz, M., Hanrahan, P.: Light field
photography with a hand-held plenoptic camera. Technical Report Stanford Uni-
versity (2005)
12. Georgiev, T., Zheng, C., Curless, B., Salesin, D., Nayar, S.K., Intwala, C.: Spatio-
angular resolution tradeoff in integral photography. In: Eurographics Symposium
on Rendering, pp. 263–272 (2006)
13. Darrell, T., Wohn, K.: Pyramid based depth from focus. CVPR, 504–509 (1988)
14. Nayar, S.K.: Shape from Focus System. CVPR, 302–308 (1992)
15. Subbarao, M., Choi, T.: Accurate Recovery of Three-Dimensional Shape from Im-
age Focus. PAMI, 266–274 (1995)
16. Levin, A., Sand, P., Cho, T.S., Durand, F., Freeman, W.T.: Motion-Invarient Pho-
tography. SIGGRAPH, ACM Transaction on Graphics (2008)
17. Ben-Ezra, M., Zomet, A., Nayar, S.: Jitter Camera: High Resolution Video from a
Low Resolution Detector. CVPR, 135–142 (2004)
18. Ait-Aider, O., Andreff, N., Lavest, J.M., Martinet, P.: Simultaneous Object Pose
and Velocity Computation Using a Single View from a Rolling Shutter Camera.
In: Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3952, pp.
56–68. Springer, Heidelberg (2006)
19. Jansson, P.A.: Deconvolution of Images and Spectra. Academic Press, London
(1997)
20. Dabov, K., Foi, A., Katkovnik, V., Egiazarian, K.: Image restoration by sparse 3D
transform-domain collaborative filtering. SPIE Electronic Imaging (2008)
21. www.cs.columbia.edu/CAVE/projects/flexible dof
Priors for Large Photo Collections and
What They Reveal about Cameras
Sujit Kuthirummal1 , Aseem Agarwala2,

Dan B Goldman2 , and Shree K. Nayar1
1
Columbia University
2
Adobe Systems, Inc.
Abstract. A large photo collection downloaded from the internet spans

a wide range of scenes, cameras, and photographers. In this paper we
introduce several novel priors for statistics of such large photo collections
that are independent of these factors. We then propose that properties of
these factors can be recovered by examining the deviation between these
statistical priors and the statistics of a slice of the overall photo collection
that holds one factor constant. Specifically, we recover the radiometric
properties of a particular camera model by collecting numerous images
captured by it, and examining the deviation of this collection’s statistics
from that of a broader photo collection whose camera-specific effects have
been removed. We show that using this approach we can recover both
a camera model’s non-linear response function and the spatially-varying
vignetting of the camera’s different lens settings. All this is achieved using
publicly available photographs, without requiring images captured under
controlled conditions or physical access to the cameras. We also apply
this concept to identify bad pixels on the detectors of specific camera
instances. We conclude with a discussion of future applications of this
general approach to other common computer vision problems.
1 Introduction
Large publicly-available photo collections such as Flickr have recently spawned

new applications such as Photo Tourism [1] and Internet Stereo [2]. They have
also been exploited for filling in holes in images [3], inserting objects into scenes
[4], and object recognition [5]. These research efforts have demonstrated the
power of using large photo collections to develop novel applications as well as to
solve hard computer vision problems.
In this paper, we examine the statistics of such a large photo collection
and develop priors that are independent of the factors that influence any one
photograph: the scene, the camera, and the photographer. Statistical priors
for single images have already been used for a wide range of computer vision
tasks [6,7,8,9,10,11]. We argue that priors on the statistics of photo collec-
tions have the potential to be similarly powerful, since the statistics of a slice of
the photo collection that holds one factor constant should yield information as
to how that factor distorts the priors. We investigate this approach to recover

Priors for Large Photo Collections and What They Reveal about Cameras 75
camera properties. We first compute statistical priors from a photo collection

with camera-specific effects removed; that is, we use known camera calibration
profiles to remove radiometric distortion from a photo collection. As a result,
that collection becomes camera-independent. Then, we describe and experimen-
tally validate priors for (a) the spatial distribution of average image luminances
and (b) the joint histogram of irradiances at neighboring pixels. Next, we com-
pute these same statistics for a camera-model-specific photo collection whose
images have not had their distortion removed. We can then recover that cam-
era model’s radiometric properties – its non-linear response function and the
spatially-varying vignetting for different lens settings – by minimizing the devi-
ation of these statistics from the camera-independent priors. We also show how
the same concept can be used to identify bad pixels on the detectors of specific
camera instances.
Our approach to recovering properties of specific camera models assumes that
all instances of a model have the same properties. This is a reasonable assumption
to make for point and shoot cameras [12] since they do not have the variability
that arises from attaching different lenses to SLR camera bodies. Hence, in this
paper, we restrict ourselves to only point-and-shoot cameras. Also, the camera
model properties we recover are aggregate estimates over many instances of the
model; for most applications, these estimates are more than adequate. Thus,
our approach provides an attractive alternative to traditional camera calibra-
tion methods which are typically tedious. Also, since our approach can be used
to recover a camera’s properties using existing pictures, it provides a conve-
nient means to create a database of camera properties. Such a database would
be similar in spirit to the databases available with commercial products like
DxO [13] and PTLens [14], but with the important advantage that the cost
of creating it would be effectively zero — there would be no need to buy the
cameras and manually calibrate them. A photo-sharing website could use our
approach to leverage its growing image collection to continually update and add
to its database of profiles, and allow users to either undistort their images or
make photometrically-correct edits. More importantly, our results demonstrate
that the statistics of large photo collections contain significant information about
scenes, cameras, and photographers, and our work represents a first step towards
extracting and exploiting that information.
2 Related Work
A number of image priors have been proposed to describe the statistics of in-
dividual photographs, such as the sparsity of outputs of band-pass filters (e.g.
derivative filters) [6,7], biases in the distribution of gradient orientations [8,9],
and 1/f fall-off of the amplitude spectrum [10,11]. These priors have been ex-
ploited for applications such as deriving intrinsic images from image sequences
[15], super-resolution and image demosaicing [16], removing effects of camera
shake [17], and classifying images as belonging to different scene categories [9,18].
We focus on the aggregate statistics of large photo collections, which tend to have
76 S. Kuthirummal et al.
less variability than the statistics of a single image. We thus propose two new
priors for aggregate statistics of large photo collections and describe how they
can be exploited to recover radiometric properties of cameras.
The most popular method for estimating the camera response function in-
volves taking multiple registered images of a static scene with varying camera
exposures [19,20]. Grossberg and Nayar [21] relax the need for spatial corre-
spondences by using histograms of images at different exposures. If the exposure
cannot be varied, but can be locked, the response can be estimated by capturing
multiple registered images of a static scene illuminated by different combina-
tions of light sources [22]. All these methods require significant user effort and
physical access to the camera. Farid [23] assumes that the response function
has the form of a gamma curve and estimates it from a single image. However,
in practice response functions can differ significantly from gamma curves. Lin
et al. [24] also estimate the response from a single image by exploiting intensity
statistics at edges. Their results depend on the kinds of edges detected, and their
method employs a non-linear optimization which needs multiple initial guesses
for robustness. In contrast, we automatically and robustly estimate the response
function using numerous existing photographs.
Vignetting can be estimated by imaging a uniformly illuminated flat texture-
less Lambertian surface, and comparing the intensity of every pixel with that
of the center pixel (which is assumed to have no vignetting) [25,26]. Unfortu-
nately, realizing such capture conditions is difficult. One approach is to use a
device called an “integrating sphere,” but this specialized hardware is expen-
sive. Stumpfel et al. [27] capture many images of a known illuminant at different
locations in the image and fit a polynomial to the measured irradiances. The
same principle has been used to estimate vignetting from overlapping images
of an arbitrary scene [28,29,30] using measured irradiances of the same scene
point at different image locations. All these methods require the user to acquire
new images under controlled conditions. Some of the above approaches [28,29]
can be used to simultaneously estimate the vignetting and the response function
of a camera, but there are typically ambiguities in recovering this information.
Since we recover both properties independently, we do not have any ambigu-
ities. Recently, Zheng et al. [31] have proposed estimating vignetting from a
single image by assuming that a vignette-corrected image will yield an image
segmentation with larger segments. Their optimization algorithm, which con-
sists of many alternating image segmentation and vignetting estimation steps,
is highly non-linear and hence is likely to have local minima issues. In contrast,
we estimate vignetting linearly and efficiently.
During manufacturing, bad pixels are typically identified by exposing image
detectors to uniform illuminations. However, some pixels develop defects later
and it is difficult for consumers to create uniform environments to detect them.
Dudas et al. [32] detect such pixels by analyzing a set of images in a Bayesian
framework. However, they only show simulation results. We propose a simple
technique that is able to detect bad pixels, albeit using many images.
3 Aggregate Statistics of Photo Collections

We now describe how we collect various internet photo collections and how we
use them to form and experimentally validate two statistical priors that are
independent of specific scenes, cameras, and photographers. We form image col-
lections by downloading images from Flickr. Flickr supports searching for images
from a particular camera model; we chose five popular models and downloaded
thousands of images for each. We also manually calibrated these cameras using
HDRShop [19] for response functions and an integrating sphere for vignetting
(across different lens settings). To validate our approach, we then used the col-
lection of one camera model – Canon S1IS – as a training set to undistort its cor-
responding downloaded images and form camera-independent priors;1 the other
camera models and their downloaded images were used to test our hypotheses.
Internet photo collections can contain outliers that corrupt our aggregate
statistics. For example, images captured with flash, edited in Photoshop, or
cropped would add distortion beyond the radiometric properties that we are re-
covering. Fortunately, EXIF tags allow us to cull most outliers; we remove flash
images, images with certain Software fields, portrait-mode images, and images
that are not full resolution. Our resultant collections contain about 40,000 images
per camera model, which we then group using lens settings since camera prop-
erties vary with aperture and focal length. We would like to point out that since
Flickr does not support searching for all images with a particular camera-lens set-
ting, there are configurations for which we could not collect enough photographs
to compute robust statistics. However, as we will show, for configurations with
sufficient photographs, our approach gives uniformly good results.
3.1 Spatial Distribution of Average Image Luminances

Torralba and Oliva [18] and artist Jason Salavon (salavon.com) have made an
interesting observation: the average of a set of photographs of similar scenes is
not spatially stationary, but has a certain structure to it. So we ask: does the
average photograph obtained by pixel-by-pixel averaging of many photographs
captured with the same lens setting have a particular structure? To investigate
this question we computed the average of the log-luminance of the photographs
in the undistorted training set photo collection with the same lens setting.
Figures 1 (a,b) show the average log-luminance of two groups of photographs
captured with the same focal length, but different f-number. One can see that we
have averaged out particular scenes, but the average image is not uniform. This
is illustrated in Figures 1 (c) and (d) which show the contrast enhanced versions
of the images in Figures 1 (a) and (b), respectively. We can immediately make
two interesting observations. (i) The average images have a vertical gradient as
can also be seen in Figure 1 (e) which shows log-luminances along a column of the
1
We assume that undistortion is enough to make an image collection camera-
independent for the purpose of training priors. While this may not be true in all
cases, we have experimentally verified that our priors are accurate across all five
camera models (from four different manufacturers) that we used in our experiments.
(a) (b) (c) (d)

80
Focal Length: 5.8, F/#: 4.5 60
70 Focal Length: 5.8, F/#: 2.8
e (Log−Luminance)
e (Log−Luminance)
40
50
30
40
20
30
10
20
0 400 800 1200 1600 0 400 800 1200 1600 2000
Rows of the Average Image Columns of the Average Image
(e) (f)
Fig. 1. (a) Average log-luminance of 15,550 photographs captured by Canon S1IS

cameras with focal length 5.8 mm and f-number 4.5. The photographs, all 2048 × 1536,
were linearized and vignette corrected before averaging. (b) Average log-luminance of
13,874 photographs captured by Canon S1IS cameras with focal length 5.8 mm and
f-number 2.8. (c,d) Contrast-enhanced versions (for illustration only) of the images in
(a) and (b), respectively. (e,f) Plots of the average log-luminances of respectively the
1000th column and 1000th row for the two settings in (a) and (b). Response functions
were normalized so that luminance values were in the range (0,255) prior to averaging.
average images. This is possibly because illumination sources are typically above
– outdoors, from the sun and sky, while indoors, from ceiling-mounted light fix-
tures. (ii) The average images do not have a horizontal gradient, illustrated by
Figure 1 (f) which shows log-luminances along a row. We have found that these
two observations are general and they hold true for all camera models and lens
settings. In summary, in the absence of vignetting, average log-luminance images
have a vertical gradient, but no horizontal gradient. This observation serves as
the prior, which we exploit to recover vignetting in Section 4.2.
3.2 Joint Histogram of Irradiances at Neighboring Pixels

A prior on the distribution of gradients in a single image is commonly used in
computer vision estimation tasks [16,17]. However, the larger data set of a photo
collection allows us to measure how this gradient distribution varies as a function
of irradiance values. Therefore, we compute the joint histogram of irradiances at
neighboring pixels (where neighborhood is defined as 4-connected). Note that we
characterize the joint histogram only for a small block of pixels, since we know
from Section 3.1 that this statistic would also vary spatially.
We now describe how we compute the joint histogram of irradiances for a
color channel of a camera model. We assume that we know the inverse response
-5
200 200 200

-10
Irradiances
Irradiances
Irradiances
-15
100 100 100
-20
-25
0 100 200 0 100 200 0 100 200
Irradiances Irradiances Irradiances
(a) Red Joint Histogram (b) Green Joint Histogram (c) Blue Joint Histogram
Fig. 2. Log of the joint histograms of (a) red, (b) green, and (c) blue irradiances
computed from 15,550 photographs captured by Canon S1IS cameras with the extreme
lens setting – smallest focal length (5.8 mm) and largest f-number (4.5). The inverse
camera response functions used were normalized so that irradiance values were in the
range (0,255). When computing the histograms we ignored irradiances less than 5 and
greater than 250 to avoid the effects of under-exposure and saturation, respectively.
function, R, for that channel, where R(i) is the irradiance value corresponding
to intensity i. Using R we linearize that channel in photographs from that model
and compute a joint histogram, JH, where JH(i, j), gives the number of times
irradiances R(i) and R(j) occur in neighboring pixels in a desired pixel block. We
interpret the joint histogram as the joint probability distribution of irradiances
by assuming that the distribution is piecewise uniform within each bin. However,
since the values of R are typically non-uniformly spaced, the bins have different
areas. Therefore, to convert the joint histogram to a probability distribution, we
divide the value of each bin by its area. Note that the values of R determine the
sampling lattice, so to enable comparisons between joint histograms for different
response functions we resample the histogram on a regular grid in irradiance
space. Finally, we normalize the resampled distribution so that it sums to one.
We computed joint histograms of red, green, and blue irradiances for several
camera models using 31 × 31 pixel blocks at the center of photographs. Figure 3
shows the joint histograms for the Canon S1IS camera model computed from
photographs with the smallest focal length and largest f-number. These his-
tograms show that the probability of any two irradiances being incident on
neighboring pixels varies depending on the values of the irradiances. Also, the
probability of the same irradiance occuring at neighboring pixels is greater for
low irradiance values and decreases slowly as the irradiance value increases.
Finally, note that the histograms for different color channels differ slightly, il-
lustrating that the visual world has different distributions for different colors.
We have empirically observed that for any particular color channel, the joint
histogram looks very similar across camera models, especially when computed
for the extreme lens setting – smallest focal length and largest f-number. This is
not surprising, because the extreme setting is chosen by different camera models
for similar types of scenes. We quantified this similarity using the symmetric
Kullback-Leibler (KL) divergence between corresponding histograms. The sym-
metric KL divergence between distributions p and q is defined as
q(i) p(i)
KLDivSym(p, q) = Σi q(i) log( ) + Σi p(i) log( ), (1)
p(i) q(i)
where p(i) and q(i) are the samples. For the Canon S1IS and Sony W1 camera
models, the symmetric KL divergence between corresponding joint histograms
for the extreme lens setting were 0.059 (red channel), 0.081 (green channel),
and 0.068 (blue channel). These small numbers illustrate that the histograms are
very similar across camera models. Therefore, we can use the joint histograms
computed for any one camera model as non-parametric priors on these statistics.
4 Using the Priors for Radiometric Calibration

In this section we use these camera-independent statistical priors to recover
the response function of a camera model, the vignetting of a camera model for
different lens settings, and the bad pixels on the detector of a specific camera. We
use the same basic approach for all three applications; given a photo collection,
we estimate camera properties that minimize the difference between the statistics
of the photo collection and the priors defined in the previous section.
4.1 Estimating Camera Response Function

We estimate a camera model response function by minimizing the difference be-
tween the joint histogram of irradiances (Section 3.2) for the camera model and
the camera-independent prior joint histogram. To estimate the response for a
color channel of a camera model, we first compute the joint histogram, JH, of
intensities in a 31 × 31 pixel block at the center of photographs from a collec-
tion with the smallest focal length and largest f-number. Say R is an estimate
of the inverse response function. Since R is a one-to-one mapping from image
intensities to irradiances, JH can be used to compute the joint histogram of irra-
diances, as described in Section 3.2. We can then determine the ‘goodness’ of the
estimate R by computing the symmetric KL Divergence (Equation 1) between
this histogram and the prior histogram for that color channel. Therefore, we can
estimate the response function using an optimization over R that minimizes this
divergence. We use a simple polynomial [20] as the parametric representation

of R, and optimize over its coefficients. We define R(i) = 255 ∗ N k=1 αk ( 255 ) ,
i k
where R(i) is the irradiance corresponding to intensity i, αk are the coefficients,

and N is the degree of the polynomial. We normalize R(.) such that R(255) =
255. We have used N = 5 in our experiments, since we found it to be a good
fit for all inverse response functions in our data set; the mean RMS fitting error
was 0.41%. We use the Nelder-Mead Simplex method [33] for the optimization.
Note that the joint histogram of image intensities has to be computed only
once, though a resampling and normalization step must be performed at each
iteration.
We used the priors obtained from the Canon S1IS model to estimate the
inverse response functions of Sony W1, Canon G5, Casio Z120, and Minolta Z2
250 250
Ground Truth Ground Truth

200 200
Estimated Estimated
Initial Guess
Irradiance
Irradiance
150 150
100 100
50 50
0 0
0 50 100 150 200 250 0 50 100 150 200 250
Intensity Intensity
(a) Red Channel of Sony W1 (b) Green Channel of Canon G5
250 250
200 Ground Truth 200 Ground Truth

Estimated Estimated
Irradiance
Irradiance
150 150
100 100
50 50
0 0
0 50 100 150 200 250 0 50 100 150 200 250
Intensity Intensity
(c) Blue Channel of Casio Z120 (d) Red Channel of Minolta Z2
Sony W1 Canon G5 Casio Z120 Minolta Z2
Proposed [24] Proposed [24] Proposed [24] Proposed [24]
Red Channel 1.344% 2.587% 1.759% 2.553% 2.269% 1.518% 2.226% 4.914%
Green Channel 1.993% 1.243% 0.865% 3.396% 2.521% 1.155% 2.743% 3.237%
Blue Channel 1.164% 1.783% 2.523% 2.154% 2.051% 3.053% 2.653% 3.292%
(e)
Fig. 3. Estimated and ground truth inverse response functions of one channel for four
camera models – (a) Sony W1, (b) Canon G5, (c) Casio Z120, and (d) Minolta Z2. For
these estimates we used 17,819, 9,529, 1,315, and 3,600 photographs, respectively. (a)
also shows the initial guess used by our optimization. (e) RMS percentage errors of the
estimated inverse response functions for camera models from four different manufac-
turers obtained using our proposed method and the method of [24].
camera models. Due to space constraints, we only show the inverse responses of
one of their channels in Figures 3(a-d). For comparison we also show the ground
truth inverse response functions obtained using HDRShop [19]2 . As we can see,
the estimated curves are very close to the ground truth curves. The difference
between the two sets of curves is greater at higher image intensities, for which
HDRShop typically provides very noisy estimates.
The RMS estimation errors are shown in Figure 3(e). Even though our esti-
mation process uses a non-linear optimization, we have found it to be robust to
2
Inverse response functions can only be estimated up to scale. To compare the inverse
responses produced by our technique and HDRShop, we scaled the results from
HDRShop by a factor that minimizes the RMS error between the two curves.
(a) (b) (c) (d)
Fig. 4. (a) Average log-luminance of 15,550 photographs captured by Canon S1IS

cameras with focal length 5.8 mm and f-number 4.5. (b) Average log-luminance of
13,874 photographs captured by Canon S1IS cameras with focal length 5.8 mm and f-
number 2.8. (c,d) Contrast enhanced versions of the images in (a) and (b), respectively.
choices of the initial guess. For all our results we used the mean inverse response
from the EMoR database [34], shown in Figure 3(a), as the initial guess. For
comparison, Figure 3(e) also shows the estimation errors obtained when using
the method of Lin et al. [24] on large image sets (the same ones used by our
method) for robustness; the overall mean RMS error of their estimates is 28%
greater than ours. An interesting question to ask is: How many photographs
does our technique need to get a good estimate? We have found that only around
200 photographs are required to get an estimate with RMS error of about 2%.
In some cases, as few as 25 photographs are required. (See [35] for details.)
4.2 Determining Vignetting for a Lens Setting

Vignetting in a photograph depends on the lens setting – the focal length and the
f-number – used to capture it. In Section 3.1, we have seen that the average log-
luminance of a group of linearized and vignette-corrected photographs captured
with the same lens setting has a vertical gradient but no horizontal gradient.
Using the technique in Section 4.1, we can recover response functions, linearize
photographs and compute average log-luminance images. Figures 4 (a, b) show
the average log-luminances for two groups of linearized photographs captured
by Canon S1IS cameras with the same focal length, but different f-number. The
photographs used were not vignette-corrected. The contrast-enhanced versions of
these images are shown in Figures 4 (c) and (d), respectively. Note the darkening
of the corners, which suggests that vignetting information is embedded in the
average images. The average images now have a horizontal gradient in addition
to a vertical gradient. This observation coupled with our prior model (Section
3.1) leads to a simple vignetting estimation algorithm: find a vignetting function
that yields a corrected average log-luminance image with no horizontal gradient.
Since vignetting affects all color channels equally, we only need to analyze its
effect on luminance. The measured luminance m at pixel (x, y) in photograph i
can be written as:
mi (x, y) = v(x, y) ∗ li (x, y), (2)
where v(x, y) is the vignetting at that pixel and li (x, y) is the luminance that
would have been measured in the absence of vignetting. Taking the log on both
sides of Equation 2 and computing the average log-luminance in N photographs

with the same lens setting, we get
1 1
Σi log(mi (x, y)) = log(v(x, y)) + Σi log(li (x, y)). (3)
N N
Writing the measured average log-luminance, N1 Σi log(mi (x, y)), as M (x, y),
log(v(x, y)) as V (x, y), and the average log-luminance in the absence of vi-
gnetting, N1 Σi log(li (x, y)), as L(x, y), Equation 3 becomes
M (x, y) = V (x, y) + L(x, y). (4)
According to our prior model, in the absence of vignetting an average log-

luminance image does not have a horizontal gradient, i.e., all values in a row
are equal. This implies that Equation 4 can be rewritten as
M (x, y) = V (x, y) + L(y). (5)
Note that M is known, while V and L are unknown. We assume that vignetting
is radially symmetric about the center of the image. Therefore, vignetting at
pixel (x, y) can be expressed as a function of the distance, r, of the pixel from
the image center. We model the log of the vignetting as a polynomial in r:
V (x, y) = N k=1 βk r , where βk are the coefficients and N is the degree of the
k
polynomial. In our experiments we have used N = 9. Note that the value of V

is zero at the center of the image, modeling the fact that there is no vignetting
there. This model reduces Equation 5 to a set of linear equations in the unknowns
L(y) and the vignetting coefficients βk , which we can solve for efficiently.
The average log-luminance images in Figures 4 (a) and (b) can be used to esti-
mate vignetting. However, we have observed that the top halves of photographs
contain many saturated pixels, especially photographs taken with small focal
lengths and large f-numbers (typically used for outdoor scenes with lots of light).
For instance, photographs in our data set captured by Canon S1IS cameras with
such a setting had pixels in the top half that were saturated approximately 30%
of the time. This means that we significantly underestimate the average value
for pixels in the top half. Since statistics of the top half of the average images
are unreliable, we have used the bottom half to recover vignetting. Figures 5
(a-f) show the estimated vignetting curves obtained using this approach for two
lens settings each of three camera models – Canon S1IS, Sony W1, and Canon
G5. For comparison, ground truth vignetting curves obtained from photographs
captured in an integrating sphere are also shown. As one can see, the estimated
vignetting and ground truth curves are very close to each other. Figure 5(g)
shows the RMS and mean estimation errors. We have found that our technique
needs around 3000 photographs to get an estimate with RMS error of about 2%.
(See [35] for details.)
We have observed that statistics at the center of photographs differ slightly
from those of other portions of the image. We believe that this is due to a com-
positional bias – faces are usually captured in the center region. This deviation
Canon S1IS Sony W1 Canon G5

1 1 1
Relative Illuminance
0.9 0.9 0.9
0.8 0.8 0.8
0.7 0.7 0.7
0.6
Ground Truth 0.6
Ground Truth 0.6
Ground Truth
Estimated Estimated Estimated
0.5 0.5 0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Radial Distance Radial Distance Radial Distance
(a) f : 5.8 mm, N : 4.5 (b) f : 7.9 mm, N : 5.6 (c) f : 7.2 mm, N :4.0
1 1 1
0.9 0.9 0.9
0.8 0.8 0.8
0.7 0.7 0.7
0.6
Ground Truth 0.6
Ground Truth 0.6
Ground Truth
Estimated Estimated Estimated
0.5 0.5 0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Radial Distance Radial Distance Radial Distance
(d) f : 5.8 mm, N :2.8 (e) f : 7.9 mm, N :2.8 (f) f : 7.2 mm, N :2.0
Canon S1IS Sony W1 Canon G5

focal length: 5.8 mm focal length: 7.9 mm focal length: 7.18 mm
f/#:4.5 f/#:2.8 f/#:5.6 f/#:2.8 f/#:4 f/#:2
RMS Error % 0.989 1.399 0.594 2.324 0.664 1.723
Mean Error % 0.895 1.221 0.460 1.980 0.484 1.398
(g)
Fig. 5. (a-f) Vignetting estimated for two lens settings each of Canon S1IS, Sony W1,
and Canon G5 cameras, using the bottom half of their respective average log-luminance
images. 15,550, 13,874, 17,819, 15,434, 12,153, and 6,324 photographs, respectively were
used for these estimates. (f and N stand for focal length and f-number respectively.)
(g) RMS and mean percentage errors of the estimated vignetting for two lens settings
each of three camera models; estimation errors are typically less than 2%.
in statistics sometimes causes relative illuminance near the image center to be

incorrectly estimated as greater than one. We have handled this by clamping
the curves to have a maximum value of one. Note that for lens settings with
smaller f-numbers, the estimation is slightly poorer for a larger region near the
image center. Such a setting is usually chosen for indoor scenes, where people
are typically closer to the camera and their faces occupy a larger region near the
image center, thus accentuating this compositional bias.
It is interesting to note from Figure 5 that for these camera models, at lens
settings with small f-numbers (large apertures), the corners of the photograph
get about 40% less light than the center! This large difference becomes very no-
ticeable if overlapping photographs are stitched together without vignette correc-
tion. If photographs are corrected for vignetting, then the overlap seams become
barely visible as was shown by [28,29].
Camera Camera Mean Median

Model Instances Bad Pixels Bad Pixels
Canon G5 15 2.2 1
Canon SD 300 13 1.1 1
Sony W1 13 0.384 0
(a) (b) (c)
Fig. 6. (a) Contrast enhanced luminance of the average of 1,186 photographs from a
particular Canon S1IS camera. (b) Zoomed in portions of the image in (a) in which
we can clearly see bad pixels that have very different intensities from their neighbors.
(c) A comparative study of the number of bad detector pixels in a particular camera
instance for three different camera models.
4.3 Identifying Bad Pixels on a Camera Detector
During manufacturing, camera detectors are exposed to uniform illuminations

so that bad pixels – pixels with abnormal sensitivities and biases – stand out
and can be easily identified. However, some pixels develop defects later and it
is difficult for consumers to create uniform environments to detect them. In
Section 3.1 we saw that by averaging a large number of photographs, we average
out particular scenes and noise to get a smoothly varying image. Thus, a simple
prior for bad pixel detection is that the average image should be smooth; bad
pixels should be identifiable as causing discontinuities in the average image.
We grouped photographs by the Flickr users who uploaded them, so that
each group has pictures from the same camera instance. We then computed the
average of each group. Figure 6 (a) shows the contrast enhanced luminance of
the average of 1,186 photographs from a particular Canon S1IS camera. In this
image, bad pixels clearly stand out, as can be seen in the zoomed-in portions
shown in Figure 6 (b). We identify a pixel as bad if the difference between its
average value and the median of the average values in a neighborhood around
it is greater than a threshold (7 gray-levels). This technique can also be used
to rank camera models by the number of bad pixels in each instance. The table
in Figure 6(c) presents results from such a study, for which we picked camera
instances which had at least 500 photographs in our collection.
5 Conclusion
In this paper, we have presented priors on two aggregate statistics of large photo
collections, and exploited these statistics to recover the radiometric properties
of camera models entirely from publicly available photographs, without physi-
cal access to the cameras themselves. In future work, we would like to develop
statistics that reveal other camera properties such as radial distortion, chromatic
aberration, spatially varying lens softness, etc.. There are, of course, a number
of powerful and accurate approaches to camera calibration, and these existing
techniques have both advantages and disadvantages relative to ours. In that
light, our primary contribution is a new approach to exploiting the statistics of

large photo collections to reveal information about scenes, cameras, and photog-
raphers. Recovering camera properties is only one possible application, and we
hope that our work inspires others to exploit this approach in new and interesting
ways. For example, differences in scene-specific statistics and scene-independent
priors could yield information about a scene’s geometry, lighting, weather, and
motion. A photographer’s photo collection could yield information on propensity
for camera shake, typical field of view, and preferred camera orientation.
Statistical priors for single images have been useful for a number of computer
vision tasks [16,17]. We argue that priors on the statistics of photo collections
have the potential to be similarly powerful, since the deviation from these priors
of a slice of the photo collection that holds one factor constant should reveal
information about that factor. Computer vision problems that operate on a
single image are often ill-posed because they must tease apart the influence of
several confounding factors of the scene, the camera, and the photographer.
For example, vignetting calibration is challenging because it is hard to know
if darkening is caused by vignetting or changes in the scene. In effect, a photo
collection allows us to marginalize over the factors that confound the task at
hand. We believe that our work is only the first step in this exciting direction.
Acknowledgements. Thanks to William Freeman, Sylvain Paris, Anat Levin,

Antonio Torralba, and Brian Curless for helpful discussions and comments.
References
1. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: Exploring photo collections
in 3D. ACM Transactions on Graphics (SIGGRAPH), 835–846 (2006)
2. Goesele, M., Snavely, N., Curless, B., Hoppe, H., Seitz, S.M.: Multi-View Stereo
for Community Photo Collections. In: ICCV (2007)
3. Hays, J., Efros, A.A.: Scene Completion Using Millions of Photographs. ACM
Transactions on Graphics (SIGGRAPH) (2007)
4. Lalonde, J.-F., Hoiem, D., Efros, A.A., Rother, C., Winn, J., Criminisi, A.: Photo
Clip Art. ACM Transactions on Graphics (SIGGRAPH) (2007)
5. Torralba, A., Fergus, R., Freeman, W.: Tiny Images. MIT Tech Report (2007)
6. Olshausen, B.A., Field, D.J.: Emergence of simple-cell receptive field properties by
learning a sparse code for nature images. In: Nature, pp. 607–609 (1996)
7. Simoncelli, E.: Statistical Models for Images: Compression, Restoration and Syn-
thesis. In: Asilomar Conference on Signals, Systems and Computers, pp. 673–678
(1997)
8. Switkes, E., Mayer, M.J., Sloan, J.A.: Spatial frequency analysis of the visual envi-
ronment: anisotropy and the carpentered environment hypothesis. Vision Research,
1393–1399 (1978)
9. Baddeley, R.: The Correlational Structure of Natural Images and the Calibration
of Spatial Representations. Cognitive Science, 351–372 (1997)
10. Burton, G.J., Moorhead, I.R.: Color and spatial structure in natural scenes. Ap-
plied Optics, 157–170 (1987)
11. Field, D.: Relations between the statistics of natural images and the response
properties of cortical cells. J. of the Optical Society of America, 2379–2394 (1987)
12. Wackrow, R., Chandler, J.H., Bryan, P.: Geometric consistency and stability of
consumer-grade digital cameras for accurate spatial measurement. The Photogram-
metric Record, 121–134 (2007)
13. DxO Labs: www.dxo.com
14. PTLens: www.epaperpress.com/ptlens
15. Weiss, Y.: Deriving intrinsic images from image sequences. In: ICCV, pp. 68–75
(2001)
16. Tappen, M.F., Russell, B.C., Freeman, W.T.: Exploiting the sparse derivative prior
for super-resolution and image demosaicing. In: Workshop on Statistical and Com-
putational Theories of Vision (2003)
17. Fergus, R., Singh, B., Hertzmann, A., Roweis, S.T., Freeman, W.T.: Removing
Camera Shake From A Single Photograph. SIGGRAPH, 787–794 (2006)
18. Torralba, A., Oliva, A.: Statistics of Natural Images Categories. Network: Compu-
tation in Neural Systems 14, 391–412 (2003)
19. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from pho-
tographs. SIGGRAPH, 369–378 (1997)
20. Mitsunaga, T., Nayar, S.K.: Radiometric self calibration. CVPR, 1374–1380 (1999)
21. Grossberg, M.D., Nayar, S.K.: Determining the Camera Response from Images:
What is Knowable?. PAMI, 1455–1467 (2003)
22. Manders, C., Aimone, C., Mann, S.: Camera response function recovery from dif-
ferent illuminations of identical subject matter. ICIP, 2965–2968 (2004)
23. Farid, H.: Blind Inverse Gamma Correction. IEEE Transactions on Image Process-
ing, 1428–1433 (2001)
24. Lin, S., Gu, J., Yamazaki, S., Shum, H.-Y.: Radiometric Calibration Using a Single
Image. CVPR, 938–945 (2004)
25. Sawchuk, A.: Real-time correction of intensity nonlinearities in imaging systems.
IEEE Transactions on Computers, 34–39 (1977)
26. Kang, S.B., Weiss, R.: Can we calibrate a camera using an image of a flat textureless
lambertian surface? In: Vernon, D. (ed.) ECCV 2000. LNCS, vol. 1843, pp. 640–
653. Springer, Heidelberg (2000)
27. Stumpfel, J., Jones, A., Wenger, A., Debevec, P.: Direct HDR capture of the sun
and sky. Afrigraph, 145–149 (2004)
28. Goldman, D.B., Chen, J.H.: Vignette and exposure calibration and compensation.
In: ICCV, pp. 899–906 (2005)
29. Litvinov, A., Schechner, Y.Y.: Addressing radiometric nonidealities: A unified
framework. CVPR, 52–59 (2005)
30. Jia, J., Tang, C.K.: Tensor voting for image correction by global and local intensity
alignment. IEEE Transactions PAMI 27(1), 36–50 (2005)
31. Zheng, Y., Lin, S., Kang, S.B.: Single-Image Vignetting Correction. CVPR (2006)
32. Dudas, J., Jung, C., Wu, L., Chapman, G.H., Koren, I., Koren, Z.: On-Line Map-
ping of In-Field Defects in Image Sensor Arrays. In: International Symposium on
Defect and Fault-Tolerance in VLSI Systems, pp. 439–447 (2006)
33. Press, W., Teukolsky, S., Vetterling, W., Flannery, B.: Numerical Recipes in C:
The Art of Scientific Computing (1992)
34. Grossberg, M.D., Nayar, S.K.: What is the Space of Camera Response Functions?.
CVPR, 602–609 (2003)
35. http://www.cs.columbia.edu/CAVE/projects/photo priors/
Understanding Camera Trade-Offs
through a Bayesian Analysis of Light Field Projections
Anat Levin1 , William T. Freeman1,2 , and Frédo Durand1

1
MIT CSAIL
2
Adobe Systems
Abstract. Computer vision has traditionally focused on extracting structure,

such as depth, from images acquired using thin-lens or pinhole optics. The de-
velopment of computational imaging is broadening this scope; a variety of un-
conventional cameras do not directly capture a traditional image anymore, but
instead require the joint reconstruction of structure and image information. For
example, recent coded aperture designs have been optimized to facilitate the joint
reconstruction of depth and intensity. The breadth of imaging designs requires
new tools to understand the tradeoffs implied by different strategies.
This paper introduces a unified framework for analyzing computational imag-
ing approaches. Each sensor element is modeled as an inner product over the
4D light field. The imaging task is then posed as Bayesian inference: given the
observed noisy light field projections and a prior on light field signals, estimate
the original light field. Under common imaging conditions, we compare the per-
formance of various camera designs using 2D light field simulations. This frame-
work allows us to better understand the tradeoffs of each camera type and analyze
their limitations.
1 Introduction
The flexibility of computational imaging has led to a range of unconventional cam-
era designs. Cameras with coded apertures [1,2], plenoptic cameras [3,4], phase
plates [5,6], and multi-view systems [7] record different combinations of light rays. Re-
construction algorithms then convert the data to viewable images, estimate depth and
other quantites. These cameras involves tradeoffs among various quantites–spatial and
depth resolution, depth of focus or noise. This paper describes a theoretical framework
that will help to compare computational camera designs and understand their tradeoffs.
Computation is changing imaging in three ways. First, the information recorded at
the sensor may not be the final image, and the need for a decoding algorithm must be
taken into account to assess camera quality. Second, beyond 2D images, the new designs
enable the extraction of 4D light fields and depth information. Finally, new priors
can capture regularities of natural scenes to complement the sensor measurements and
amplify decoding algorithms. The traditional evaluation tools based on the image point
spread function (PSF) [8,9] are not able to fully model these effects. We seek tools for
comparing camera designs, taking into account those three aspects. We want to evaluate
the ability to recover a 2D image as well as depth or other information and we want to
model the decoding step and use natural-scene priors.

Understanding Camera Trade-Offs 89
A useful common denominator, across camera designs and scene information, is the
lightfield [7], which encodes the atomic entities (lightrays) reaching the camera. Light
fields naturally capture some of the more common photography goals such as high spa-
tial image resolution, and are tightly coupled with the targets of mid-level computer
vision: surface depth, texture, and illumination information. Therefore, we cast the re-
construction performed in computational imaging as light field inference. We then need
to extend prior models, traditionally studied for 2D images, to 4D light fields.
Camera sensors sum over sets of light rays, with the optics specifying the mapping
between rays and sensor elements. Thus, a camera provides a linear projection of the
4D light field where each projected coordinate corresponds to the measurement of one
pixel. The goal of decoding is to infer from such projections as much information as
possible about the 4D light field. Since the number of sensor elements is significantly
smaller than the dimensionality of the light field signal, prior knowledge about light
fields is essential. We analyze the limitations of traditional signal processing assump-
tions [10,11,12] and suggest a new prior on light field signals which explicitly accounts
for their structure. We then define a new metric of camera performance as follows:
Given a light field prior, how well can the light field be reconstructed from the data
measured by the camera? The number of sensor elements is of course a critical vari-
able, and we chose to standardize our comparisons by imposing a fixed budget of N
sensor elements to all cameras.
We focus on the information captured by each camera, and wish to avoid the con-
founding effect of camera-specific inference algorithms or the decoding complexity.
For clarity and computational efficiency we focus on the 2D version of the problem
(1D image/2D light field). We use simplified optical models and do not model lens
aberrations or diffraction (these effects would still follow a linear projection model and
can be accounted for with modifications to the light field projection function.)
Our framework captures the three major elements of the computational imaging
pipeline – optical setup, decoding algorithm, and priors – and enables a systematic
comparison on a common baseline.
1.1 Related Work

Approaches to lens characterization such as Fourier optics [8,9] analyze an optical ele-
ment in terms of signal bandwidth and the sharpness of the PSF over the depth of field,
but do not address depth information. The growing interest in 4D light field render-
ing has led to research on reconstruction filters and anti-aliasing in 4D [10,11,12], yet
this research relies mostly on classical signal processing assumptions of band limited
signals, and do not utilize the rich statistical correlations of light fields. Research on
generalized camera families [13,14] mostly concentrates on geometric properties and
3D configurations, but with an assumption that approximately one light ray is mapped
to each sensor element and thus decoding is not taken into account.
Reconstructing data from linear projections is a fundamental component in CT
and tomography [15]. Fusing multiple image measurements is also used for super-
resolution, and [16] studies uncertainties in this process.
90 A. Levin, W.T. Freeman, and F. Durand
a a
b plane b b
a plane
(a) 2D slice through a scene (b) Light field (c) Pinhole

a a a
b b b
(d) Lens (e) Lens, focus change (f) Stereo

a a a
b b b
(g) Plenoptic camera (h) Coded aperture lens (i) Wavefront coding
Fig. 1. (a) Flat-world scene with 3 objects. (b) The light field, and (c)-(i) cameras and the light
rays integrated by each sensor element (distinguished by color).
2 Light Fields and Camera Configurations

Light fields are usually represented with a two-plane parameterization, where each ray
is encoded by its intersections with two parallel planes. Figure 1(a,b) shows a 2D slice
through a diffuse scene and the corresponding 2D slice of the 4D light field. The color
at position (a0 , b0 ) of the light field in fig. 1(b) is that of the reflected ray in fig. 1(a)
which intersects the a and b lines at points a0 , b0 respectively. Each row in this light
field corresponds to a 1D view when the viewpoint shifts along a. Light fields typically
have many elongated lines of nearly uniform intensity. For example the green object in
fig. 1 is diffuse and the reflected color does not vary along the a dimension. The slope
of those lines corresponds to the object depth [10,11].
Each sensor element integrates light from some set of light rays. For example, with
a conventional lens, the sensor records an integral of rays over the lens aperture. We
review existing cameras and how they project light rays to sensor elements. We assume
that the camera aperture is positioned on the a line parameterizing the light field.
Pinhole Each sensor element collects light from a single ray, and the camera projection
just slices a row in the light field (fig 1(c)). Since only a tiny fraction of light is let in,
noise is an issue.
Lenses gather more light by focusing all light rays from a point at distance D to a sensor
point. In the light field, 1/D is the slope of the integration (projection) stripe (fig 1(d,e)).
An object is in focus when its slope matches this slope (e.g. green in fig 1(d)) [10,11,12].
Objects in front or behind the focus distance will be blurred. Larger apertures gather
more light but can cause more defocus.
Stereo [17] facilitate depth inference by recording 2 views (fig 1(g), to keep a constant
sensor budget, the resolution of each image is halved).
Plenoptic cameras capture multiple viewpoints using a microlens array [3,4]. If each
microlens covers k sensor elements one achieves k different views of the scene, but the
spatial resolution is reduced by a factor of k (k = 3 is shown in fig 1(g)).
Coded aperture [1,2] place a binary mask in the lens aperture (fig 1(h)). As with con-
ventional lenses, objects deviating from the focus depth are blurred, but according to
the aperture code. Since the blur scale is a function of depth, by searching for the code
scale which best explains the local image window, depth can be inferred. The blur can
also be inverted, increasing the depth of field.
Wavefront coding introduces an optical element with an unconventional shape so that
rays from any world point do not converge. Thus, integrating over a curve in light field
space (fig 1(i)), instead of the straight integration of lenses. This is designed to make
defocus at different depths almost identical, enabling deconvolution without depth in-
formation, thereby extending depth of field. To achieve this, a cubic lens shape (or phase
plate) is used. The light field integration curve, which is a function of the lens normal,
can be shown to be a parabola (fig 1(i)), which is slope invariant (see [18] for a deriva-
tion, also independently shown by M. Levoy and Z. Zhu, personal communication).
3 Bayesian Estimation of Light Field
3.1 Problem Statement
We model an imaging process as an integration of light rays by camera sensors, or in

an abstract way, as a linear projection of the light field
y = Tx + n (1)
where x is the light field, y is the captured image, n is an iid Gaussian noise n ∼
N (0, η 2 I) and T is the projection matrix, describing how light rays are mapped to
sensor elements. Referring to figure 1, T includes one row for each sensor element, and
this row has non-zero elements for the light field entries marked by the corresponding
color (e.g. a pinhole T matrix has a single non-zero element per row).
The set of realizable T matrices is limited by physical constraints. In particular,
the entries of T are all non-negative. To ensure equal noise conditions, we assume a
maximal integration time, and the maximal value for each entry of T is 1. The amount
of light reaching each sensor element is the sum of the entries in the corresponding T
row. It is usually better to collect more light to increase the SNR (a pinhole is noisier
because it has a single non-zero entry per row, while a lens has multiple ones).
To simplify notation, most of the following derivation will address a 2D slice in the
4D light field, but the 4D case is similar. While the light field is naturally continuous,
for simplicity we use a discrete representation.
Our goal is to understand how well we can recover the light field x from the noisy
projection y, and which T matrices (among the camera projections described in the
previous section) allow better reconstructions. That is, if one is allowed to take N mea-
surements (T can have N rows), which set of projections leads to better light field re-
construction? Our evaluation methodology can be adapted to a weight w which specifies
how much we care about reconstructing different parts of the light field. For example, if
the goal is an all-focused, high quality image from a single view point (as in wavefront
coding), we can assign zero weight to all but one light field row.
The number of measurements taken by most optical systems is significantly smaller
than the light field data, i.e. T contains many fewer rows than columns. As a result,
it is impossible to recover the light field without prior knowledge on light fields. We
therefore start by modeling a light field prior.
3.2 Classical Priors

State of the art light field sampling and reconstruction approaches [10,11,12] apply
signal processing techniques, typically assuming band-limited signals. The number of
non-zero frequencies in the signal has to be equal to the number of samples, and there-
fore before samples are taken, one has to apply a low-pass filter to meet the Nyquist
limit. Light field reconstruction is then reduced to a convolution with a proper low-pass
filter. When the depth range in the scene is bounded, these strategies can further bound
the set of active frequencies within a sheared rectangle instead of a standard square of
low frequencies and tune the orientation of the low pass filter. However, they do not
address inference for a general projection such as the coded aperture.
One way to express the underlying band limited assumptions in a prior terminology
is to think of an isotropic Gaussian prior (where by isotropic we mean that no direction
in the light field is favored). In the frequency domain, the covariance of such a Gaussian
is diagonal (with one variance per Fourier coefficient), allowing zero (or very narrow)
variance at high frequencies above the Nyqusit limit, and a wider one at the lower
frequencies. Similar priors can also be expressed in the spatial domain by penalizing
the convolution with a set of high pass filters:
1 1
P (x) ∝ exp(− |fk,i xT |2 ) = exp(− xT Ψ0−1 x) (2)
2σ0 2
k,i
where fk,i denotes the kth high pass filter centered at the ith light field entry. In sec 5,
we will show that band limited assumptions and Gaussian priors indeed lead to equiva-
lent sampling conclusions.
More sophisticated prior choices replace the Gaussian prior of eq 2 with a heavy-
tailed prior [19]. However, as will be illustrated in section 3.4, such generic priors ignore
the very strong elongated structure of light fields, or the fact that the variance along the
disparity slope is significantly smaller than the spatial variance.
3.3 Mixture of Gaussians (MOG) Light Field Prior

To model the strong elongated structure of light fields, we propose using a mixture of
oriented Gaussians. If the scene depth (and hence light field slope) is known we can
define an anisotropic Gaussian prior that accounts for the oriented structure. For this,
we define a slope field S that represents the slope (one over the depth of the visible
point) at every light field entry (fig. 2(b) illustrates a sparse sample from a slope field).
For a given slope field, our prior assumes that the light field is Gaussian, but has a
variance in the disparity direction that is significantly smaller than the spatial variance.
The covariance ΨS corresponding to a slope field S is then:
1 T 1 T 2
xT ΨS−1 x = |gS(i),i x|2 + |g0,i x| (3)
i
σ s σ 0
where gs,i is a derivative filter in orientation s centered at the ith light field entry (g0,i
is the derivative in the horizontal/spatial direction), and σs << σ0 , especially for non-
specular objects (in practice, we consider diffuse scenes and set σs = 0). Conditioning
on depth we have P (x|S) ∼ N (0, ΨS ).
We also need a prior P (S) on the slope field S. Given that depth is usually piecewise
smooth, our prior encourages piecewise smooth slope fields (like the regularization of
stereo algorithms). Note however that S and its prior are expressed in light-field space,
not image or object space. The resulting unconditional light field prior is an infinite
mixture of Gaussians (MOG) that sums over slope fields

P (x) = P (S)P (x|S) (4)
S
We note that while each mixture component is a Gaussian which can be evaluated in
closed form, marginalizing over the infinite set of slope fields S is intractable, and
approximation strategies are described below.
Now that we have modeled the probability of a light field x, we turn to the imaging
problem: Given a camera T and a noisy projection y we want to find a Baysian estimate
for the light field x. For this, we need to define P (x|y; T ), the probability that x is the
explanation of the measurement y. Using Bayes’ rule:

P (x|y; T ) = P (x, S|y; T ) = P (S|y; T )P (x|y, S; T ) (5)
S S
To express the above equation, we note that y should equal T x up to measurement

noise, that is, P (y|x; T ) ∝ exp(− 2η12 |T x − y|2 ) . As a result, for a given slope field S,
P (x|y, S; T ) ∝ P (x|S)P (y|x; T ) is also Gaussian with covariance and mean:
1 T 1
ΣS−1 = ΨS−1 + T T μS = ΣS T T y (6)
η2 η2
Similarly, P (y|S; T ) is also a Gaussian distribution measuring how well we can explain
y with the slope component S, or, the volume of light fields x which can explain the
measurement y, if the slope field was S. This can be computed by marginalizing over
light fields x: P (y|S; T ) = x P (x|S)P (y|x; T ). Finally, P (S|y; T ) is obtained from
Bayes’ rule: P (S|y; T ) = P (S)(y|S; T )/ S P (S)(y|S; T )
To recap, the probability P (x|y; T ) that a light field x explains a measurement y is
also a mixture of Gaussians (MOG). To evaluate it, we measure how well x can explain
y, conditioning on a particular slope field S, and weight it by the probability P (S|y)
that S is actually the slope field of the scene. This is integrated over all slope fields S.
Inference. Given a camera T and an observation y we seek to recover the light field
x. In this section we consider MAP estimation, while in section 4 we approximate the
variance as well in an attempt to compare cameras. Even MAP estimation for x is hard,
−3
x 10
7 lens isotropic gaussian prior
isotropic sparse prior
pinhole
6 light fields prior
band−pass assumption
5 wave
front
4 coding stereo
coded
3 aperture
2
plenoptic
1
(a) Test image (b) light field and slope field (c) SSD error in reconstruction
Fig. 2. Light field reconstruction
as the integral in eq 5 is intractable. We approximate the MAP estimate for the slope
field S, and conditioning on this estimate, solve for the MAP light field x.
The slope field inference is essentially inferring the scene depth. Our inference gener-
alizes MRF stereo algorithms [17] or the depth regularization of the coded aperture [1].
Details regarding slope inference are provided in [18], but as a brief summary, we model
slope in local windows as constant or having one single discontinuity, and we then reg-
ularize the estimate using an MRF.
Given the estimated slope field S, our light field prior is Gaussian, and thus the
MAP estimate for the light field is the mean of the conditional Gaussian μS in eq 6.
This mean minimizes the projection error up to noise, and regularize the estimate by
minimizing the oriented variance ΨS . Note that in traditional stereo formulations the
multiple views are used only for depth estimation. In contrast, we seek a light field
that satisfies the projection in all views. Thus, if each view includes aliasing, we obtain
“super resolution”.
3.4 Empirical Illustration of Light Field Inference

Figure 2(a,b) presents an image and a light field slice, involving depth discontinuities.
Fig 2(c) presents the numerical SSD estimation errors. Figure 3 presents the estimated
light fields and (sparse samples from) the corresponding slope fields. See [18] for more
results. Note that slope errors in the 2nd row often accompany ringing in the 1st row.
We compare the results of the MOG light field prior with simpler Gaussian priors (ex-
tending the conventional band limited signal assumptions [10,11,12]) and with modern
sparse (but isotropic) derivative priors [19]. For the plenoptic camera we also explic-
itly compare with signal processing reconstruction (last bar in fig 2(c))- as explained in
sec 3.2 this approach do not apply directly to any of the other cameras.
The prior is critical, and resolution is significantly reduced in the absence of a slope
model. For example, if the plenoptic camera includes aliasing, figure 3(left) demon-
strates that with our slope model we can super-resolve the measurements and the actual
information encoded by the recorded plenoptic data is higher than that of the direct
measurements.
The ranking of cameras also changes as a function of prior- while the plenoptic
camera produced best results for the isotropic priors, a stereo camera achieves a higher
Plenoptic camera Stereo Coded Aperture
Fig. 3. Reconstructing a light field from projections. Top row: reconstruction with our MOG light
field prior. Middle row: slope field (estimated with MOG prior), plotted over ground truth. Note
slope changes at depth discontinuities. Bottom row: reconstruction with isotropic Gaussian prior.
resolution under an MOG prior. Thus, our goal in the next section is to analytically
evaluate the reconstruction accuracy of different cameras, and to understand how it is
affected by the choice of prior.
4 Camera Evaluation Metric

We want to assess how well a light field x0 can be recovered from a noisy projection
y = T x0 + n, or, how much the projection y nails down the set of possible light field
interpretations. The uncertainty can be measured by the expected reconstruction error:

E(|W (x − x0 )|2 ; T ) = P (x|y; T )|W (x − x0 )|2 (7)
x
where W = diag(w) is a diagonal matrix specifying how much we care about different
light field entries, as discussed in sec 3.1.
Uncertainty computation. To simplify eq 7, recall that the average distance between
x0 and the elements of a Gaussian is the distance from the center, plus the variance:

E(|W (x − x0 )|2 |S; T ) = |W (μS − x0 )|2 + diag(W 2ΣS ) (8)
In a mixture model, the contribution of each component is weighted by its volume:

E(|W (x − x0 )|2 ; T ) = P (S|y)E(|W (x − x0 )|2 |S; T ) (9)
S
Since the integral in eq 9 can not be computed explicitly, we evaluate cameras using
synthetic light fields whose ground truth slope field is known, and evaluate an approxi-
mate uncertainty in the vicinity of the true solution. We use a discrete set of slope field
samples {S1 , ..., SK } obtained as perturbations around the ground truth slope field. We
approximate eq 9 using a discrete average:
1
E(|W (x − x0 )|2 ; T ) ≈ P (Sk |y)E(|W (x − x0 )|2 |Sk ; T ) (10)
K
k
Finally, we use a set of typical light fields x0t (generated using ray tracing) and eval-
uate the quality of a camera T as the expected squared error over these examples

E(T ) = E(|W (x − x0t )|2 ; T ) (11)
t
Note that this solely measures information captured by the optics together with the prior,
and omits the confounding effect of specific inference algorithms (like in sec 3.4).
5 Tradeoffs in Projection Design
Which designs minimize the reconstruction error?

Gaussian prior. We start by considering the isotropic Gaussian prior in eq 2. If the
x is Gaussian, we can integrate over x in eq 11 analytically
distribution of light fields
to obtain: E(T ) = 2 diag(1/η 2 T T T + Ψ0−1 )−1 . Thus, we reach the classical PCA
conclusion: to minimize the residual variance, T should measure the directions of max-
imal variance in Ψ0 . Since the prior is shift invariant, Ψ0−1 is diagonal in the frequency
domain, and the principal components are the lowest frequencies. Thus, an isotropic
Gaussian prior agrees with the classical signal processing conclusion [10,11,12] - to
sample the light field one should convolve with a low pass filter to meet the Nyquist
limit and sample both the directional and spatial axis, as a plenoptic camera does. (if
the depth in the scene is bounded, fewer directional samples can be used [10]). This
is also consistent with our empirical prediction, as for the Gaussian prior, the plenop-
tic camera achieved the lowest error in fig 2(c). However, this sampling conclusion is
conservative as the directional axis is clearly more redundant than the spatial one. The
second order statistics captured by a Gaussian distribution do not capture the high order
dependencies of light fields.
Mixture of Gaussian light field prior. We now turn to the MOG prior. While the
optimal projection under this prior cannot be predicted in closed-form, it can help us
understand the major components influencing the performance of existing cameras. The
score in eq 9 reveals two aspects which affect a camera quality - first, minimizing the
variance ΣS of each of the mixture components (i.e., the ability to reliably recover
the light field given the true slope field), and second, the need to identify depth and
make P (S|y) peaked at the true slope field. Below, we elaborate on these components.
5.1 Conditional Light Field Estimation – Known Depth
Fig 4 shows light fields estimated by several cameras, assuming the true depth (and
therefore slope field), was successfully estimated. We also display the variance of the
estimated light field - the diagonal of ΣS (eq 6).
In the right part of the light field, the lens reconstruction is sharp, since it averages
rays emerging from a single object point. On the left, uncertainty is high, since it av-
erages light rays from multiple points.In contrast, integrating over a parabolic curve
(wavefront coding) achieves low uncertainties for both slopes, since a parabola “cov-
ers” all slopes (see [18,20] for derivation). A pinhole also behaves identically at all
depths, but it collects only a small amount of light and the uncertainty is high due to the
small SNR. Finally, the uncertainty increases in stereo and plenoptic cameras due to the
smaller number of spatial samples.
The central region of the light field demonstrates the utility of multiple viewpoint in
the presence of occlusion boundaries. Occluded parts which are not measured properly
Pinhole
Lens
Wavefront coding
Stereo
Plenoptic
Fig. 4. Evaluating conditional uncertainty in light field estimate. Left: projection model. Middle:
estimated light field. Right: variance in estimate (equal intensity scale used for all cameras). Note
that while for visual clarity we plot perfect square samples, in our implementation samples were
convolved with low pass filters to simulate realistic optics blur.
lead to higher variance. The variance in the occluded part is minimized by the plenoptic
camera, the only one that spends measurements in this region of the light field.
Since we deal only with spatial resolution, our conclusions correspond to common
sense, which is a good sanity check. However, they cannot be derived from a naive Gaus-
sian model, which emphasizes the need for a prior such as as our new mixture model.
5.2 Depth Estimation

Light field reconstruction involves slope (depth) estimation. Indeed, the error in eq 9
also depends on the uncertainty in the slope field S. We need to make P (S|y) peaked
at the true slope field S 0 . Since the observation y is T x + n, we want the distributions
of projections T x to be as distinguishable as possible for different slope fields S. One
way to achieve this is to make the projections corresponding to different slope fields
concentrated within different subspaces of the N-dimensional space. For example, a
stereo camera yields a linear constraint on the projection- the N/2 samples from the
first view should be a shifted version (according to slope) of the other N/2. The coded
aperture camera also imposes linear constraints: certain frequencies of the defocused
signals are zero, and the location of these zeros shifts with depth [1].
To test this, we measure the probability of the true slope field, P (S 0 |y), aver-
aged over a set of test light fields (created with ray tracing). The stereo score is
= 0.95 (where = 1 means perfect depth discrimination)
compared to = 0.84 for coded aperture. This suggests that the disparity
constraint of stereo better distributes the projections corresponding to different slope
fields than the zero frequency subspace in coded aperture.
We can also quantitatively compare stereo with depth from defocus (DFD) - two
lenses with the same center of projection, focused at two different depths. As predicted
by [21], with the same physical size (stereo baseline shift doesn’t exceed aperture width)
both designs perform similarly, with DFD achieving = 0.92.
Our probabilistic treatment of depth estimation goes beyond linear subspace con-
straints. For example, the average slope estimation score of a lens was =
0.74, indicating that, while weaker than stereo, a single monocular image captured with
a standard lens contains some depth-from-defocus information as well. This result can-
not be derived using a disjoint-subspace argument, but if the full probability is consid-
ered, the Occam’s razor principle applies and the simpler explanation is preferred.
Finally, a pinhole camera-projection just slices a row out of the light field, and this
slice is invariant to the light field slope. The parabola filter of a wavefront coding lens
is also designed to be invariant to depth. Indeed, for these two cameras, the evaluated
distribution P (S|y) in our model is uniform over slopes.
Again, these results are not surprising but they are obtained within a general frame-
work that can qualitatively and quantitatively compare a variety of camera designs.
While comparisons such as DFD vs. stereo have been conducted in the past [21], our
framework encompasses a much broader family of cameras.
5.3 Light Field Estimation

In the previous section we gained intuition about the various parts of the expected error
in eq 9. We now use the overall formula to evaluate existing cameras, using a set of
diffuse light field generated using ray tracing (described in [18]). Evaluated configura-
tions include a pinhole camera, lens, stereo pair, depth-from-defocus (2 lenses focused
at different depths), plenoptic camera, coded aperture cameras and a wavefront coding
lens. Another advantage of our framework is that we can search for optimal parameters
within each camera family, and our comparison is based on optimized parameters such
as baseline length, aperture size and focus distance of the individual lens in a stereo pair,
and various choices of codes for coded aperture cameras (details provided in [18]).
By changing the weights, W on light field entries in eq 7, we evaluate cameras for
two different goals: (a) Capturing a light field. (b) Achieving an all-focused image from
a single view point (capturing a single row in the light field.)
We consider both a Gaussian and our new MOG prior. We consider different depth
complexity as characterized by the amount of discontinuities. We use slopes between
−45o to 45o and noise with standard deviation η = 0.01. Additionally, [18] evaluates
changes in the depth range and noise. Fig. 5(a-b) plot expected reconstruction error with
our MOG prior. Evaluation with a generic Gaussian prior is included in [18]. Source
code for these simulations is available on the authors’ webpage.
Full light field reconstruction. Fig. 5(a) shows full light field reconstruction with our
MOG prior. In the presence of depth discontinues, lowest light field reconstruction error
is achieved with a stereo camera. While a plenoptic camera improves depth informa-
tion our comparison suggests it may not pay for the large spatial resolution loss. Yet, as
discussed in sec 5.1 a plenoptic camera offers an advantage in the presence of complex
occlusion boundaries. For planar scenes (in which estimating depth is easy) the coded
aperture surpasses stereo, since spatial resolution is doubled and the irregular sampling
of light rays can avoid high frequencies losses due to defocus blur. While the perfor-
mance of all cameras decreases when the depth complexity increases, a lens and coded
−3 −3
x 10 x 10
4 2.5
No depth discontinuities pinhole No depth discontinuities
Modest depth discontinuities Modest depth discontinuities
3.5 Many depth discontinuities Many depth discontinuities
lens
coded 2 plenoptic
3 aperture lens coded
pinhole aperture
wave plenoptic DFD
2.5 front wave stereo
coding front
1.5 coding
DFD
2 stereo
1.5
1
(a) full light field (b) single view
Fig. 5. Camera evaluation. See [18] for enlarged plots
aperture are much more sensitive than others. While the depth discrimination of DFD
is similar to that of stereo (as discussed in sec 5.2), its overall error is slightly higher
since the wide apertures blur high frequencies.
The ranking in figs 5(a) agrees with the empirical prediction in fig 2(c). However,
while fig 5(a) measures inherent optics information, fig 2(c) folds-in inference errors as
well.
Single-image reconstruction. For single row reconstruction (fig 5(b)) one still has to
account for issues like defocus, depth of field, signal to noise ratio and spatial resolution.
A pinhole camera (recording this single row alone) is not ideal, and there is an advantage
for wide apertures collecting more light (recording multiple light field rows) despite not
being invariant to depth.
The parabola (wavefront coding) does not capture depth information and thus per-
forms very poorly for light field estimation. However, fig 5(b) suggests that for recov-
ering a single light field row, this filter outperforms all other cameras. The reason is
that since the filter is invariant to slope, a single central light field row can be recov-
ered without knowledge of depth. For this central row, it actually achieves high signal
to noise ratios for all depths, as demonstrated in figure 4. To validate this observation,
we have searched over a large set of lens curvatures, or light field integration curves,
parameterized as splines fitted to 6 key points. This family includes both slope sensitive
curves (in the spirit of [6] or a coded aperture), which identify slope and use it in the
estimation, and slope invariant curves (like the parabola [5]), which estimate the cen-
tral row regardless of slope. Our results show that, for the goal of recovering a single
light field row, the wavefront-coding parabola outperforms all other configurations. This
extends the arguments in previous wavefront coding publications which were derived
using optics reasoning and focus on depth-invariant approaches. It also agrees with the
motion domain analysis of [20], predicting that a parabolic integration curve provides
an optimal signal to noise ratio.
5.4 Number of Views for Plenoptic Sampling

As another way to compare the conclusions derived by classical signal processing ap-
proaches with the ones derived from a proper light field prior, we follow [10] and ask:
suppose we use a camera with a fixed N pixels resolution, how many different views
(N pixels each) do we actually need for a good ‘virtual reality’?
Figure 6 plots the expected reconstruction er-
ror as a function of the number of views for both 7 x 10 −3
MOG and naive Gaussian priors. While a Gaus- Gaussian prior

MOG prior
6
sian prior requires a dense sample, the MOG er-
ror is quite low after 2-3 views (such conclu- 5
Nyquist
sions depend on depth complexity and the range 4 Limit
of views we wish to capture). For comparison, 3
we also mark on the graph the significantly larger
views number imposed by an exact Nyquist limit 2
analysis, like [10]. Note that to simulate a re- 1
alistic camera, our directional axis samples are 0
0 10 20 30 40
aliased. This is slightly different from [10] which
blur the directional axis in order to properly elim- Fig. 6. Reconstruction error as a func-
inate frequencies above the Nyquist limit. tion number of views
6 Discussion
The growing variety of computational camera designs calls for a unified way to analyze
their tradeoffs. We show that all cameras can be analytically modeled by a linear mapping
of light rays to sensor elements. Thus, interpreting sensor measurements is the Bayesian
inference problem of inverting the ray mapping. We show that a proper prior on light
fields is critical for the successes of camera decoding. We analyze the limitations of tra-
ditional band-pass assumptions and suggest that a prior which explicitly accounts for the
elongated light field structure can significantly reduce sampling requirements.
Our Bayesian framework estimates both depth and image information, accounting
for noise and decoding uncertainty. This provides a tool to compare computational cam-
eras on a common baseline and provides a foundation for computational imaging. We
conclude that for diffuse scenes, the wavefront coding cubic lens (and the parabola light
field curve) is the optimal way to capture a scene from a single view point. For capturing
a full light field, a stereo camera outperformed other tested configurations.
We have focused on providing a common ground for all designs, at the cost of sim-
plifying optical and decoding aspects. This differs from traditional optics optimization
tools such as Zemax that provide fine-grain comparisons between subtly-different de-
signs (e.g. what if this spherical lens element is replaced by an aspherical one?). In
contrast, we are interested in the comparison between families of imaging designs (e.g.
stereo vs. plenoptic vs. coded aperture). We concentrate on measuring inherent informa-
tion captured by the optics, and do not evaluate camera-specific decoding algorithms.
The conclusions from our analysis are well connected to reality. For example, it
can predict the expected tradeoffs (which can not be derived using more naive light
field models) between aperture size, noise and spatial resolution discussed in sec 5.1. It
justifies the exact wavefront coding lens design derived using optics tools, and confirms
the prediction of [21] relating stereo to depth from defocus.
Analytic camera evaluation tools may also permit the study of unexplored camera
designs. One might develop new cameras by searching for linear projections that yield
optimal light field inference, subject to physical implementation constraints. While the
camera score is a very non-convex function of its physical characteristics, defining cam-
era evaluation functions opens up these research directions.
Acknowledgments. We thank Royal Dutch/Shell Group, NGA NEGI-1582-04-0004,

MURI Grant N00014-06-1-0734, NSF CAREER award 0447561. Fredo Durand ac-
knowledges a Microsoft Research New Faculty Fellowship and a Sloan Fellowship.
References
1. Levin, A., Fergus, R., Durand, F., Freeman, W.: Image and depth from a conventional camera
with a coded aperture. SIGGRAPH (2007)
2. Veeraraghavan, A., Raskar, R., Agrawal, A., Mohan, A., Tumblin, J.: Dappled photography:
Mask-enhanced cameras for heterodyned light fields and coded aperture refocusing. SIG-
GRAPH (2007)
3. Adelson, E.H., Wang, J.Y.A.: Single lens stereo with a plenoptic camera. PAMI (1992)
4. Ng, R., Levoy, M., Bredif, M., Duval, G., Horowitz, M., Hanrahan, P.: Light field photogra-
phy with a hand-held plenoptic camera. Stanford U. Tech. Rep. CSTR 2005-02 (2005)
5. Bradburn, S., Dowski, E., Cathey, W.: Realizations of focus invariance in optical-digital sys-
tems with wavefront coding. Applied optics 36, 9157–9166 (1997)
6. Dowski, E., Cathey, W.: Single-lens single-image incoherent passive-ranging systems. App.
Opt. (1994)
7. Levoy, M., Hanrahan, P.M.: Light field rendering. SIGGRAPH (1996)
8. Goodman, J.W.: Introduction to Fourier Optics. McGraw-Hill Book Company, New York
(1968)
9. Zemax: http://www.zemax.com
10. Chai, J., Tong, X., Chan, S., Shum, H.: Plenoptic sampling. SIGGRAPH (2000)
11. Isaksen, A., McMillan, L., Gortler, S.J.: Dynamically reparameterized light fields. SIG-
GRAPH (2000)
12. Ng, R.: Fourier slice photography. SIGGRAPH (2005)
13. Seitz, S., Kim, J.: The space of all stereo images. In: ICCV (2001)
14. Grossberg, M., Nayar, S.K.: The raxel imaging model and ray-based calibration. In: IJCV
(2005)
15. Kak, A.C., Slaney, M.: Principles of Computerized Tomographic Imaging
16. Baker, S., Kanade, T.: Limits on super-resolution and how to break them. PAMI (2002)
17. Scharstein, D., Szeliski, R.: A taxonomy and evaluation of dense two-frame stereo corre-
spondence algorithms. Intl. J. Computer Vision 47(1), 7–42 (2002)
18. Levin, A., Freeman, W., Durand, F.: Understanding camera trade-offs through a bayesian
analysis of light field projections. MIT CSAIL TR 2008-049 (2008)
19. Roth, S., Black, M.J.: Fields of experts: A framework for learning image priors. In: CVPR
(2005)
20. Levin, A., Sand, P., Cho, T.S., Durand, F., Freeman, W.T.: Motion invariant photography.
SIGGRAPH (2008)
21. Schechner, Y., Kiryati, N.: Depth from defocus vs. stereo: How different really are they. IJCV
(2000)
CenSurE: Center Surround Extremas for
Realtime Feature Detection and Matching
Motilal Agrawal1, Kurt Konolige2, and Morten Rufus Blas3

1
SRI International, Menlo Park CA 94025, USA
agrawal@ai.sri.com
2
Willow Garage, Menlo Park CA 94025, USA
konolige@willowgarage.com
3
Elektro/DTU University, Lyngby, Denmark
mrb@elektro.dtu.dk
Abstract. We explore the suitability of different feature detectors for

the task of image registration, and in particular for visual odometry,
using two criteria: stability (persistence across viewpoint change) and
accuracy (consistent localization across viewpoint change). In addition
to the now-standard SIFT, SURF, FAST, and Harris detectors, we intro-
duce a suite of scale-invariant center-surround detectors (CenSurE) that
outperform the other detectors, yet have better computational charac-
teristics than other scale-space detectors, and are capable of real-time
implementation.
1 Introduction
Image matching is the task of establishing correspondences between two images
of the same scene. This is an important problem in Computer Vision with appli-
cations in object recognition, image indexing, structure from motion and visual
localization – to name a few. Many of these applications have real-time constraints
and would benefit immensely from being able to match images in real time.
While the problem of image matching has been studied extensively for various
applications, our interest in it has been to be able to reliably match two images
in real time for camera motion estimation, especially in difficult off-road environ-
ments where there is large image motion between frames [1,2]. Vehicle dynamics
and outdoor scenery can make the problem of matching images very challenging.
The choice of a feature detector can have a large impact in the performance of
such systems.
We have identified two criteria that affect performance.
– Stability: the persistence of features across viewpoint change
– Accuracy: the consistent localization of a feature across viewpoint change

This material is based upon work supported by the United States Air Force under
Contract No. FA8650-04-C-7136. Any opinions, findings and conclusions or recom-
mendations expressed in this material are those of the author(s) and do not neces-
sarily reflect the views of the United States Air Force.

CenSurE: Center Surround Extremas 103
Stability is obviously useful in tracking features across frames. Accuracy of fea-

ture localization is crucial for visual odometry tasks, but keypoint operators
such as SIFT typically subsample the image at higher scales, losing pixel-level
precision.
Broadly speaking, we can divide feature classes into two types. Corner detec-
tors such as Harris (based on the eigenvalues of the second moment matrix [3,4])
and FAST [5] (analysis of circular arcs [6]) find image points that are well local-
ized, because the corners are relatively invariant to change of view. Both these
detectors can be implemented very efficiently and have been used in structure-
from-motion systems [2,7,8] because of their accuracy. However, they are not in-
variant to scale and therefore not very stable across scale changes, which happen
constantly with a moving camera. The Harris-Laplace and the Hessian-Laplace
features [9] combine scale-space techniques with the Harris approach. They use a
scale-adapted Harris measure [10] or the determinant of the Hessian to select the
features and the Laplacian to select the scale. Supposedly, visual odometry can
benefit from scale-space features, since they can be tracked for longer periods
of time, and should lead to improved motion estimates from incremental bundle
adjustment of multiple frames.
While we expect scale-space features to be more stable than simple corner
features, are they as accurate? The answer, at least for visual odometry, is “no”.
The reason is that, as typically implemented in an image pyramid, scale-space
features are not well localized at higher levels in the pyramid. Obviously, features
at high levels have less accuracy relative to the original image. The culprit in loss
of accuracy is the image pyramid. If the larger features were computed at each
pixel, instead of reducing the size of the image, accuracy could be maintained.
However, computing features at all scales is computationally expensive, which is
why SIFT features [11], one of the first scale-space proposals, uses the pyramid –
each level incurs only 1/4 the cost of the previous one. SIFT attempts to recover
some of the lost accuracy through subpixel interpolation.
Our proposal is to maintain accuracy by computing features at all scales at
every pixel in the original image. The extrema of the Laplacian across scale have
been shown to be very stable [12], so we consider this operator, or more generally,
extrema of a center-surround response (CenSurE, or Center Surround Extrema).
We explore a class of simple center-surround filters that can be computed in time
independent of their size, and show that, even when finding extrema across all
scales, they are suitable for real-time tasks such as visual odometry. CenSurE
filters outperform the best scale-space or corner features at this task in terms
of track length and accuracy, while being much faster to compute; and they are
also competitive in standard tests of repeatability for large-viewpoint changes.
While the main focus of this paper is on a novel feature detector, visual
odometry (and other motion estimation tasks) can benefit from matching using
a descriptor that is robust to viewpoint changes. In this paper, we develop a
fast variant of the upright SURF descriptor, and show that it can be used in
real-time tasks.
104 M. Agrawal, K. Konolige, and M.R. Blas
1.1 Related Work
The two scale-space detectors that are closest to our work, in technique and prac-
ticality, are SIFT [11] and SURF [13]. The main differences between approaches
is summarized in the table below.
CenSurE SIFT SURF
Spatial resolution at scale full subsampled subsampled
Scale-space operator Laplace Laplace Hessian
Approximation (Center-surround) (DOG) (DOB)
Edge filter Harris Hessian Hessian
Rotational invariance approximate yes no
The key difference is the full spatial resolution achieved by CenSurE at ev-
ery scale. Neither SIFT nor SURF computes responses at all pixels for larger
scales, and consquently do not detect extrema across all scales. Instead, they
consider each scale octave independently. Within an octave, they subsamples
the responses, and find extrema only at the subsampled pixels. At each suc-
cessive octave, the subsampling is increased, so that almost all computation is
spent on the first octave. Consequently, the accuracy of features at larger scales
is sacrificed, in the same way that it is for pyramid systems. While it would be
possible for SIFT and SURF to forego subsampling, it would then be inefficient,
with compute times growing much larger.
CenSurE also benefits from using an approximation to the Laplacian, which
has been shown to be better for scale selection [12]. The center-surround ap-
proximation is fast to compute, while being insensitive to rotation (unlike the
DOB Hessian approximation). Also, CenSurE uses a Harris edge filter, which
gives better edge rejection than the Hessian.
Several simple center-surround filters exist in the literature. The bi-level Lapla-
cian of Gaussian (BLoG) approximates the LoG filter using two levels. [14] de-
scribes circular BLoG filters and optimizes for the inner and outer radius to best
approximate the LoG filter. The drawback is that the cost of BLoG depends on
the size of the filter. Closer to our approach is that of Grabner et al. [15], who
describe a difference-of-boxes (DOB) filter that approximates the SIFT detector,
and is readily computed at all scales with integral images [16,17]. Contrary to the
results presented in [15], we demonstrate that our DOB filters outperform SIFT
in repeatability. This can be attributed to careful selection of filter sizes and using
the second moment matrix instead of the Hessian to filter out responses along a
line. In addition, the DOB filter is not invariant to rotation, and in this paper we
propose filters that have better properties.
The rest of the paper is organized as follows. We describe our CenSurE features
in detail in Section 2. We then discuss our modified upright SURF (MU-SURF)
in Section 3. We compare the performance of CenSurE against several other
feature detectors. Results of this comparison for image matching are presented
in Section 4.1 followed by results for visual odometry in Section 4.2. Finally,
Section 5 concludes this paper.
2 Center Surround Extrema (CenSurE) Features
Our approach to determining accurate large-scale features demands that we com-

pute all features at all scales, and select the extrema across scale and location.
Obviously, this strategy demands very fast computation, and we use simplified
bi-level kernels as center-surround filters. The main concern is finding kernels
that are rotationally invariant, yet easy to compute.
2.1 Finding Extrema
In developing affine-invariant features, Mikolajczyk and Schmid [18] report on

two detectors that seem better than others in repeatability – the Harris-Laplace
and Hessian-Laplace. Mikoljczyk and Schmid note that the Harris and Hessian
detectors (essentially corner detectors) are good at selecting a location within
a scale, but are not robust across scale. Instead, they show that the maximum
of Laplacian operator across scales gives a robust characteristic scale - hence
the hybrid operator, which they define as follows: first a peak in the Harris or
Hessian operator is used to select a location, and then the Laplacian selects the
scale at that location.
This strategy requires computing the Hessian/Harris measure at all locations
and all scales, and additionally calculating the Laplacian at all scales where
there are peaks in the corner detector. In our view, the Laplacian is easier to
compute and to approximate than the Hessian, as was discovered by Lowe for
SIFT features. So in our approach, we compute a simplified center-surround filter
at all locations and all scales, and find the extrema in a local neighborhood. In
a final step, these extrema are filtered by computing the Harris measure and
eliminating those with a weak corner response.
2.2 Bi-level Filters
While Lowe approximated the Laplacian with the difference of Gaussians, we

seek even simpler approximations, using center-surround filters that are bi-level,
that is, they multiply the image value by either 1 or −1. Figure 1 shows a
progression of bi-level filters with varying degrees of symmetry. The circular filter
is the most faithful to the Laplacian, but hardest to compute. The other filters
can be computed rapidly with integral images (Section 2.7), with decreasing
cost from octagon to hexagon to box filter. We investigate the two endpoints:
octagons for good performance, and boxes for good computation.
2.3 CenSurE Using Difference of Boxes
We replace the two circles in the circular BLoG with squares to form our
CenSurE-DOB. This results in a basic center-surround Haar wavelet. Figure
1(d) shows our generic center-surround wavelet of block size n. The inner box
is of size (2n + 1) × (2n + 1) and the outer box is of size (4n + 1) × (4n + 1).
Convolution is done by multiplication and summing. If In is the inner weight
Fig. 1. Progression of Center-Surround bi-level filters. (a) circular symmetric BLoG

(Bilevel LoG) filter. Successive filters (octagon, hexagon, box) have less symmetry.
and On is the weight in the outer box, then in order for the DC response of this
filter to be zero, we must have
On (4n + 1)2 = In (2n + 1)2 (1)
We must also normalize for the difference in area of each wavelet across scale.
In (2n + 1)2 = In+1 (2 (n + 1) + 1)2 (2)
We use a set of seven scales for the center-surround Haar wavelet, with block
size n = [1, 2, 3, 4, 5, 6, 7]. Since the block sizes 1 and 7 are the boundary, the
lowest scale at which a feature is detected corresponds to a block size of 2. This
roughly corresponds to a LoG with a sigma of 1.885. These five scales cover 2 12
octaves, although the scales are linear. It is easy to add more filters with block
sizes 8,9, and so on.
2.4 CenSurE Using Octagons

Difference of Boxes are obviously not rotationally invariant kernels. In particular,
DOBs will perform poorly for 45 degrees in-plane rotation. Octagons, on the
other hand are closer to circles and approximate LoG better than DOB.
In using octagons, the basic ideas of performing convolutions by inner and
outer weighted additions remain the same. As in DOB, one has to find weights
In and On such that the DC response is zero and all filters are normalized
according to the area of the octagons.
An octagon can be represented by the height of the vertical side (m) and
height of the slanted side (n) (Figure 1(b)). Table 1 shows the different octagon
sizes corresponding to the seven scales. These octagons scale linearly and were
experimentally chosen to correspond to the seven DOBs described in the previous
section.
2.5 Non-maximal Suppression

We compute the seven filter responses at each pixel in the image. We then
perform a non-maximal suppression over the scale space. Briefly, a response is
Table 1. CenSurE-OCT: inner and outer octagon sizes for various scales
scale n=1n=2n=3n=4n=5 n=6 n=7

inner (m, n) (3, 0) (3, 1) (3, 2) (5, 2) (5, 3) (5, 4) (5, 5)
outer (m, n) (5, 2) (5, 3) (7, 3) (9, 4) (9, 7) (13, 7) (15, 10)
suppressed if there is a response greater (maxima case) or a response less than

(minima case) its neighbors in a local neighborhood over the location and scales.
Pixels that are either maxima or minima in this neighborhood are the feature
point locations. We use a 3x3x3 neighborhood for our non-maximal suppression.
The magnitude of the filter response gives an indication of the strength of the
feature. The greater the strength, the more likely it is to be repeatable. Weak
responses are likely to be unstable. Therefore, we can apply a threshold to filter
out the weak responses.
Since all our responses are computed on the original image without subsam-
pling, all our feature locations are localized well and we do not need to perform
subpixel interpolation.
2.6 Line Suppression

Features that lie along an edge or line are poorly localized along it and therefore
are not very stable. Such poorly defined peaks will have large principal curva-
tures along the line but a small one in the perpendicular direction and therefore
can be filtered out using the ratio of principal curvatures. We use the second mo-
ment matrix of the response function at the particular scale to filter out these
responses. ! 2 "
L L L
H = x x2 y (3)
Lx Ly Ly
Lx and Ly are the derivatives of the response function L along x and y. The
summation is over a window that is linearly dependent on the scale of the par-
ticular feature point: the higher the scale, the larger the window size. Note that
this is the scale-adapted Harris measure [18,10] and is different from the Hessian
matrix used by SIFT [11,15] to filter out line responses. Once the Harris mea-
sure is computed, its trace and determinant can be used to compute the ratio of
principal curvatures. We use a threshold of 10 for this ratio and a 9 × 9 window
at the smallest scale of block size 2.
The Harris measure is more expensive to compute than the Hessian matrix
used by SIFT. However, this measure needs to be computed for only a small
number of feature points that are scale-space maxima and whose response is
above a threshold and hence does not present a computational bottleneck. In
our experience it does a better job than Hessian at suppressing line responses.
2.7 Filter Computation

The key to CenSurE is to be able to compute the bi-level filters efficiently at all
sizes. The box filter can be done using integral images [16,17]. An integral image
+
x’,y’
+
x,y x,y
Fig. 2. Using slanted integral images to con- Fig. 3. Regions and subregions for
struct trapezoidal areas. Left is a slanted in- MU-SURF descriptor. Each subregion
tegral image, where the pixel x, y is the sum (in blue) is 9x9 with an overlap of 2
of the shaded areas; α is 1. Right is a half- pixels at each boundary. All sizes are
trapezoid, from subtracting two slanted inte- relative to the scale of the feature s.
gral image pixels.
I is an intermediate representation for the image and contains the sum of gray
scale pixel values of image N with height y and width x, i.e.,

x
y
I(x, y) = N (x , y ) (4)
x =0 y =0
The integral image is computed recursively, requiring only one scan over the
image. Once the integral image is computed, it it takes only four additions to
calculate the sum of the intensities over any upright, rectangular area, indepen-
dent of its size.
Modified versions of integral images can be exploited to compute the other
polygonal filters. The idea here is that any trapezoidal area can be computed
in constant time using a combination of two different slanted integral images,
where the sum at a pixel represents an angled area sum. The degree of slant is
controlled by a parameter α:
x+α(y−y )

y
Iα (x, y) = N (x , y ). (5)
y =0 x =0
When α = 0, this is just the standard rectangular integral image. For α < 0, the
summed area slants to the left; for α > 0, it slants to the right (Figure 2, left).
Slanted integral images can be computed in the same time as rectangular ones,
using incremental techniques.
Adding two areas together with the same slant determines one end of a trape-
zoid with parallel horizontal sides (Figure 2, right); the other end is done sim-
ilarly, using a different slant. Each trapezoid requires three additions, just as
in the rectangular case. Finally, the polygonal filters can be decomposed into
1 (box), 2 (hexagon), and 3 (octagon) trapezoids, which is the relative cost of
computing these filters.
3 Modified Upright SURF (MU-SURF) Descriptor

Previously, we have demonstrated accurate visual odometry using ZNCC for
feature matching [1] (using a 11 × 11 region). However, it is relatively sensitive
to in-plane rotations (roll), larger changes in perspective, and inaccuracies in
keypoint localization. The problems related to rolls and perspective changes
become more significant as the region size increases. We have therefore decided
to switch to an upright SURF type descriptor [13].
The SURF descriptor builds on from the SIFT descriptor by encoding local
gradient information. It uses integral images to compute Haar wavelet responses,
which are then summed in different ways in 4 × 4 subregions of the region to
create a descriptor vector of length 64.
As pointed out by David Lowe [11], “it is important to avoid all boundary
effects in which the descriptor abruptly changes as a sample shifts smoothly
from being within one histogram to another or from one orientation to another.”
The SURF descriptor [13] weighs the Haar wavelet responses using a Gaussian
centered at the interest point. This single weighting scheme gave poor results
and we were unable to recreate the SURF descriptor results without accounting
for these boundary effects.
To account for these boundary conditions, each boundary in our descriptor
has a padding of 2s, thereby increasing our region size from 20s to 24s, s being
the scale of the feature. The Haar wavelet responses in the horizontal (dx ) and
vertical (dy ) directions are computed for each 24 × 24 point in the region with
filter size 2s by first creating a summed image, where each pixel is the sum of a
region of size s. The Haar wavelet output results in four fixed-size dx ,dy, |dx |,|dy |
images that have the dimensions 24 × 24 pixels irrespective of the scale.
Each dx ,dy, |dx |,|dy | image is then split into 4×4 square overlapping subregions
of size 9 × 9 pixels with an overlap of 2 pixels with each of the neighbors.
Figure fig:descriptor shows these regions and subregions. For each subregion
the values are then weighted with a precomputed Gaussian (σ1 = 2.5) centered
on the subregion center andsummed into the usual SURF descriptor vector for
each subregion: v = ( dx , dy , |dx | , |dy |). Each subregion vector is then
weighted using another Gaussian (σ2 = 1.5) defined on a mask of size 4 × 4 and
centered on the feature point. Like the original SURF descriptor, this vector is
then normalized.
The overlap allows each subregion to work on a larger area so samples that get
shifted around are more likely to still leave a signature in the correct subregion
vectors. Likewise, the subregion Gaussian weighting means that samples near
borders that get shifted out of a subregion have less impact on the subregion
descriptor vector.
From an implementation point of view the dynamic range of the vector was
small enough that the end results could be scaled into C++ shorts. This allows
for very fast matching using compiler vectorization.
CenSurE features themselves are signed based on their being dark or bright
blobs. This is similar to SURF and can also be used to speed up the matching
by only matching bright features to bright features and so forth.
We have compared the performance of MU-SURF with U-SURF for matching

and found them to be similar. As will be pointed out in Section 4.3, our imple-
mentation of MU-SURF is significantly faster than U-SURF. It is unclear to us
as to why MU-SURF is so much faster. We are currently looking into this.
We compare CenSurE-DOB and CenSurE-OCT to Harris, FAST, SIFT, and
SURF feature detectors for both image matching and visual odometry. Results
for image matching are presented in Section 4.1 and VO in Section 4.2.
4.1 Image Matching

For image matching, we have used the framework of [12] to evaluate repeatability
scores for each detector on the graffiti and boat sequences1 . We have used the
default parameters for each of these detectors. In addition, since each of these
detectors has a single value that represents the strength of the feature, we have
chosen a strength threshold such that each of these detectors results in the
same number of features in the common overlapping regions. Figure 4 (a) & (b)
shows a plot of the detector repeatability and number of correspondences for
each detector using 800 features and an overlap threshold of 40% for the graffiti
sequence. For Harris and FAST, the scale of all detected points was assumed to
be the same and set at 2.0.
Both versions of CenSurE are better than SIFT or SURF, although for large
viewpoint changes, the differences become only marginal. As can be expected,
CenSurE-OCT does better than CenSurE-DOB.
The boat sequence is more challenging because of large changes in rotation and
zoom. Figure 4 (c) & (d) shows the detector performance for this sequence for
800 features. On this challenging sequence, CenSurE performs slightly worse than
either SIFT or SURF, especially for the larger zooms. This can be attributed to
CenSurE’s non-logarithmic scale sampling. Furthermore, CenSurE filters cover
only 2 21 octaves and therefore has less degree of scale-invariance for large scale
changes.
To evaluate the matching performance, we used our MU-SURF descriptor
for each of those detectors and matched each detected point in one image to
the one with the lowest error using Euclidean distance. A correspondence was
deemed as matched if the true match was within a search radius r of its esti-
mated correspondence. Note that this is a different criterion than considering
overlap error and we have chosen this because this same criterion is used in vi-
sual odometry to perform image registration. Figure 5 shows the percentage of
correct matches as a function of search radius when the number of features is
fixed to 800.
1
Available from http://www.robots.ox.ac.uk/∼ vgg/research/affine/
100 700
Harris Harris
FAST FAST
90 SIFT SIFT
CenSurE−DOB 600 CenSurE−DOB
SURF SURF
80 CenSurE−OCT CenSurE−OCT
70 500
nb of correspondences
60
repeatibility %
400
50
300
40
30
200
20
100
10
0 0
10 20 30 40 50 60 70 20 25 30 35 40 45 50 55 60
viewpoint angle viewpoint angle
(a) (b)
100 600
Harris Harris
90 FAST FAST
SIFT SIFT
CenSurE−DOB 500 CenSurE−DOB
80
SURF SURF
CenSurE−OCT CenSurE−OCT
70
nb of correspondences
400
repeatibility %
60
50 300
40
200
30
20
100
10
0 0
1 2 3 4 5 6 7 2 2.5 3 3.5 4 4.5 5 5.5 6
image number image number
(c) (d)
Fig. 4. Repeatability and number of correspondences for different detectors for the
graffiti and boat sequences. The number of features is the same for each detector. (a)
& (b) graffiti sequence. (c) & (d) boat sequence.
4.2 Visual Odometry

We evaluate the performance of CenSurE for performing visual odometry in
challenging off-road environments. Because there can be large image motion
between frames, including in-plane rotations, the tracking task is difficult: es-
sentially, features must be re-detected at each frame. As usual, we compare our
method against Harris, FAST, SIFT, and SURF features. Note that this is a test
of the detectors; the same MU-SURF descriptor was used for each feature.
The Visual Odometry (VO) system derives from recent research by the authors
and others on high-precision VO [1,2] using a pair of stereo cameras. For each
new frame, we perform the following process.
1. Distinctive features are extracted from each new frame in the left image.
Standard stereo methods are used to find the corresponding point in the
right image.
2. Left-image features are matched to the features extracted in the previous
frame using our descriptor. We use a large area, usually around 1/5 of the
image, to search for matching features.
3. From these uncertain matches, we recover a consensus pose estimate us-
ing a RANSAC method [19]. Several thousand relative pose hypotheses are
generated by randomly selecting three matched non-collinear features, and
then scored using pixel reprojection errors.
Percentage of correct matches as a function of search radius. Number of features is fixed to 800
70
Relative performance of features
Harris
FAST 1
SIFT
60 CenSurE−DOB
SURF
CenSurE−OCT
0.8
Scale relative to best

50
FAST
Percentage inliers
0.6 Harris
40 SIFT
SURF
0.4 SURF+
30 DOB
OCT
0.2
20
0
10 percent inliers mean track length
1 2 3 4 5
Search radius
Fig. 5. Percentage of correct matches as a Fig. 6. Basic performance of operators in

function of search radius the VO dataset
4. If the motion estimate is small and the percentage of inliers is large enough,
we discard the frame, since composing such small motions increases error. A
kept frame is called a key frame. The larger the distance between key frames,
the better the estimate will be.
5. The pose estimate is refined further in a sparse bundle adjustment (SBA)
framework [20,21].
The dataset for this experiment consists of 19K frames taken over the course of
a 3 km autonomous, rough-terrain run. The images have resolution 512x384, and
were taken at a 10 Hz rate; the mean motion between frames was about 0.1m.
The dataset also contains RTK GPS readings synchronized with the frames, so
ground truth to within about 10 cm is available for gauging accuracy.
We ran each of the operators under the same conditions and parameters for
visual odometry, and compared the results. Since the performance of an operator
is strongly dependent on the number of features found, we set a threshold of 400
features per image, and considered the highest-ranking 400 features for each
operator. We also tried hard to choose the best parameters for each operator.
For example, for SURF we used doubled images and a subsampling factor of 1,
since this gave the best performance (labeled “SURF+” in the figures).
The first set of statistics shows the raw performance of the detector on two
of the most important performance measures for VO: the average percentage of
inliers to the motion estimate, and the mean track length for a feature (Figure 6).
In general, the scale-space operators performed much better than the simple
corner detectors. CenSurE-OCT did the best, beating out SURF by a small
margin. CenSurE-DOB is also a good performer, but suffers from lack of radial
symmetry. Surprisingly, SIFT did not do very well, barely beating Harris corners.
Note that the performance of the scale-space operators is sensitive to the
sampling density. For standard SURF settings (no doubled image, subsampling
of 2) the performance is worse than the corner operators. Only when sampling
densely for 2 octaves, by using doubled images and setting subsampling to 1, does
performance approach that of CenSurE-OCT. Of course, this mode is much more
expensive to compute for SURF (see Section 4.3).
Frames with fewer than N inliers STD from ground truth

300 14
FAST
FAST
Harris
12 Harris
250 SIFT
SIFT
SURF
SURF+
SURF+ 10 DOB
200 DOB
Number of frames
OCT
OCT
STD, meters
8
150
6
100
4
50
2
0 0
10 20 30 40 50 60 70 80 90 100 0 100 200 300 400 500
Inliers Interval distance, meters
Fig. 7. Accuracy statistics. Left: number of frames with inliers less than a certain
amount, out of 19K frames. For example, FAST and Harris both have around 50
frames with fewer than 30 inliers. Right: standard deviation from ground truth, over
trajectories of varying length.
The question to ask is: do these performance results translate into actual gains
in accuracy of the VO trajectory? We look at two measures of accuracy, the
number of frames with low inlier counts, and the deviation of the VO trajectory
from ground truth (Figure 7). The graph at the left of the figure can be used
to show how many frames are not matched, given a threshold for inliers. For
example, we typically use 30 inliers as a cutoff: any frames with fewer matches
are considered to have bad motion estimates. With this cutoff, SIFT, SURF+,
OCT, and DOB all have less than 10 missed frames, while Harris and FAST have
around 50. To show the influence of low-resolution localization, standard SURF
does very poorly here, as we expect from the previous performance graph.
Finally, we looked at the deviation of the VO estimates from ground truth,
for different trajectory lengths. At every 10 key frames along the VO trajectory,
we compared a trajectory of length N against the corresponding ground truth,
to give a dense sampling (about 1000 for each trajectory length). The standard
deviation is a measure of the goodness of the VO trajectory. Here, OCT, DOB
and Harris were all about equivalent, and gave the best estimates. Although
Harris does not do well in getting large numbers of inliers for difficult motions,
it is very well localized, and so gives good motion estimates. SIFT and SURF+
give equivalent results, and are penalized by their localization error.
Overall, CenSurE-OCT gives the best results in terms of accurate motion esti-
mates, and misses very few frames. Harris does very well in accuracy of motion,
but misses a large number of frames. SURF+ is a reasonable performer in terms
of missed frames, but is not as accurate as the CenSurE or Harris features.
4.3 Timing Results

Timing results for our CenSurE and MU-SURF implementations on an Intel
Pentium-M 2 GHz machine for a 512 × 384 image are presented in Table 2. For
comparison, SURF timings based on the original author’s implementations2 (on
the same computational platform and on the same images) are also included.
2
Available from http://www.vision.ee.ethz.ch/∼ surf/download.html
Table 2. Time in milliseconds for different feature detectors and descriptors
detector descriptor
SURF+ SURF-1 SIFT SURF OCT DOB Harris U-SURF MU-SURF
3408 292 304 75 23 17 10 308 16
SURF has default parameters (no doubled image, subsampling of 2), whereas
SURF-1 has subsampling set to 1, and SURF+ is SURF-1 with a doubled image.
For the descriptor, both U-SURF and MU-SURF are given the same features
(about 1000 in number).
For VO the best performance is with SURF+. In this case, CenSurE-OCT
yields more than a hundred-fold improvement in timing. Our MU-SURF is also
more than twenty times faster than U-SURF. It is clear that feature detection
using CenSurE features and matching using MU-SURF descriptors can be easily
accomplished in real time.
5 Conclusion
We have presented two variants of center-surround feature detectors (CenSurE)
that outperform other state-of-the-art feature detectors for image registration in
general and visual odometry in particular. CenSurE features are computed at
the extrema of the center-surround filters over multiple scales, using the original
image resolution for each scale. They are an approximation to the scale-space
Laplacian of Gaussian and can be computed in real time using integral images.
Not only are CenSurE features efficient, but they are distinctive, stable and
repeatable in changes of viewpoint. For visual odometry, CenSurE features result
in longer track lengths, fewer frames where images fail to match, and better
motion estimates.
We have also presented a modified version of the upright SURF descriptor
(MU-SURF). Although the basic idea is same as the original SURF descriptor,
we have modified it so as to handle the boundaries better, and it is also faster. It
has been our experience that MU-SURF is well suited for visual odometry and
performs much better than normalized cross-correlation without much compu-
tational overhead.
CenSurE is in constant use on our outdoor robots for localization; our goal
is to ultimately be able to do visual SLAM in real time. Toward this end, we
are exploiting CenSurE features to recognize landmarks and previously visited
places in order to perform loop closure.
References
1. Konolige, K., Agrawal, M., Solà, J.: Large scale visual odometry for rough terrain.
In: Proc. International Symposium on Robotics Research (November 2007)
2. Agrawal, M., Konolige, K.: Real-time localization in outdoor environments using
stereo vision and inexpensive GPS. In: ICPR (August 2006)
3. Harris, C., Stephens, M.: A combined corner and edge detector. In: Alvey Vision
Conference, pp. 147–151 (1988)
4. Shi, J., Tomasi, C.: Good features to track. In: Proc. Computer Vision and Pattern
Recognition (CVPR) (1994)
5. Rosten, E., Drummond, T.: Machine learning for high-speed corner detection. In:
European Conference on Computer Vision, vol. 1 (2006)
6. Rosten, E., Drummond, T.: Fusing points and lines for high performance tracking
2, 1508–1515 (2005)
7. Mouragnon, E., Lhuillier, M., Dhome, M., Dekeyser, F., Sayd, P.: Real time local-
ization and 3rd reconstruction. In: CVPR, vol. 1, pp. 363–370 (June 2006)
8. Nister, D., Naroditsky, O., Bergen, J.: Visual odometry. In: Proc. IEEE Conference
on Computer Vision and Pattern Recognition (June 2004)
9. Mikolajczyk, K., Schmid, C.: An affine invariant interest point detector. In: Heyden,
A., Sparr, G., Nielsen, M., Johansen, P. (eds.) ECCV 2002. LNCS, vol. 2350.
10. Lindeberg, T.: Feature detection with automatic scale selection. International Jour-
nal of Computer Vision 30(2) (1998)
11. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International
Journal of Computer Vision 60(2), 91–110 (2004)
12. Mikolajczyk, K., Tuytelaars, T., Schmid, C., Zisserman, A., Matas, J., Schaffal-
itzky, F., Kadir, T., Gool, L.V.: A comparison of affine region detectors. IJCV,
43–72 (2005)
13. Herbert Bay, T.T., Gool, L.V.: Surf: Speeded up robust features. In: European
Conference on Computer Vision (May 2006)
14. Pei, S.C., Horng, J.H.: Design of FIR bilevel Laplacian-of-Gaussian filter. Signal
Processing 82, 677–691 (2002)
15. Grabner, M., Grabner, H., Bischof, H.: Fast approximated SIFT. In: Narayanan,
P.J., Nayar, S.K., Shum, H.-Y. (eds.) ACCV 2006. LNCS, vol. 3851, pp. 918–927.
16. Viola, P., Jones, M.: Robust real-time face detection. In: ICCV 2001 (2001)
17. Lienhart, R., Maydt, J.: An extended set of Haar-like features for rapid object
detection. In: IEEE Conference on Image Processing (ICIP) (2002)
18. Mikolajczyk, K., Schmid, C.: Indexing based on scale invariant interest points. In:
International Conference on Computer Vision (ICCV) (2001)
19. Fischler, M., Bolles, R.: Random sample consensus: a paradigm for model fitting
with application to image analysis and automated cartography. Commun. ACM 24,
381–395 (1981)
20. Engels, C., Stewénius, H., Nister, D.: Bundle adjustment rules. Photogrammetric
Computer Vision (September 2006)
21. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment
- a modern synthesis. In: Vision Algorithms: Theory and Practice. LNCS, pp. 298–
Searching the World’s Herbaria: A System for
Visual Identification of Plant Species
Peter N. Belhumeur1 , Daozheng Chen2 , Steven Feiner1 , David W. Jacobs2,

W. John Kress3 , Haibin Ling4 , Ida Lopez3 , Ravi Ramamoorthi1 ,
Sameer Sheorey2 , Sean White1 , and Ling Zhang3
1
Dept. Of Computer Science, Columbia University, New York NY
2
Dept. Of Computer Science, University of Maryland, College Park MD
3
Department Of Botany, National Museum of Natural History, Smithsonian
Institution, Washington, DC
4
Information Science and Technology Center, Temple University, Philadelphia PA
Abstract. We describe a working computer vision system that aids in

the identification of plant species. A user photographs an isolated leaf on
a blank background, and the system extracts the leaf shape and matches
it to the shape of leaves of known species. In a few seconds, the sys-
tem displays the top matching species, along with textual descriptions
and additional images. This system is currently in use by botanists at
the Smithsonian Institution National Museum of Natural History. The
primary contributions of this paper are: a description of a working com-
puter vision system and its user interface for an important new applica-
tion area; the introduction of three new datasets containing thousands
of single leaf images, each labeled by species and verified by botanists
at the US National Herbarium; recognition results for two of the three
leaf datasets; and descriptions throughout of practical lessons learned in
constructing this system.
1 Introduction
We have built a hand-held botanical identification system for use by botanists at
the Smithsonian Institution. Employing customized computer vision algorithms,
our system significantly speeds up the process of plant species identification.
The system requires only that the user photograph a leaf specimen, returning
within seconds images of the top matching species, along with supporting data
such as textual descriptions and high resolution type specimen images. By using
our system, a botanist in the field can quickly search entire collections of plant
species—a process that previously took hours can now be done in seconds.
To date, we have created three datasets for the system: one that provides
complete coverage of the flora of Plummers Island (an island in the Potomac
River owned by the National Park Service); a second that covers all woody
plants in published flora of the Baltimore-Washington, DC area; and a nearly
complete third dataset that covers all the trees of Central Park in NYC. The

Searching the World’s Herbaria 117
Fig. 1. Left: A computer vision system for identifying temperate plants on the botani-
cally well-studied Plummers Island, Maryland, USA. Right: Congressman John Tanner
tries an augmented reality version of the system.
system is currently being used by botanists at the Smithsonian to help catalogue

and monitor plant species. Figure 1 shows the system and various versions of
the user interface (UI). Although a great deal of work remains to be done in
this ongoing collaboration between computer vision researchers and scientists at
the US National Herbarium, we hope that our system will serve as a model and
possible stepping stone for future mobile systems that use computer vision-based
recognition modules as one of their key components.
1.1 Motivation
Botanists in the field are racing to capture the complexity of the Earth’s flora
before climate change and development erase their living record. To greatly
speed up the process of plant species identification, collection, and monitoring,
botanists need to have the world’s herbaria at their fingertips. Tools are needed
to make the botanical information from the world’s herbaria accessible to anyone
with a laptop or cell phone, whether in a remote jungle or in NYC’s Central Park.
Only recently has the data required to produce these tools been made avail-
able. Volumes of biological information are just now going on-line: natural history
museums have recently provided on-line access to hundreds of thousands of im-
ages of specimens, including our own work in helping to digitize the complete
Type Specimen Collection of the US National Herbarium. These massive digiti-
zation efforts could make species data accessible to all sorts of people including
non-specialists, anywhere in the world.
Yet there is a critical shortfall in all these types of natural databases: finding
a species quickly requires that the searcher know in advance the name of the
species. Computer vision algorithms can remove this obstacle, allowing a user to
search through this data using algorithms that match images of newly collected
specimens with images of those previously discovered and described. Without
such tools, a dichotomous key must be painfully navigated to search the many
118 P.N. Belhumeur et al.
Photograph Leaf Segment Compute IDSC Show Top Matches
Fig. 2. A flow diagram of our plant identification system. A leaf from an unknown
species of plant is photographed by the user. The system then segments the leaf image
from its background, computes the IDSC shape representation used for matching, and
then displays the top matches, as they are computed.
branches and seemingly endless nodes of the taxonomic tree. The process of
identifying a single species using keys may take hours or days, even for specialists,
and is exceedingly difficult to impossible for non-scientists.
1.2 System Design and Contributions

Using our system, a botanist in the field can choose a leaf and photograph it
against a plain background. The leaf image is then compared to all possible
matches, and in a matter of seconds the botanist is shown information about
the best matching species. Figure 2 illustrates the process, with photographs of
our system in action. Figure 4 shows the performance of our system. On the
woody plants of the Baltimore-Washington, DC area (245 species) the system
returns the correct species in the top ten matches more than 97% of the time.
This paper makes several contributions. First and foremost, we describe a
complete working system for an important application that has received little
attention from the computer vision community. We hope the reader will take
from this paper an appreciation for the possible impact that computer vision can
have on the study of biodiversity. Also, while many individual components of our
system build on existing work, we have gained valuable experience getting these
pieces to work effectively together, and we want to pass these lessons on to others
in the field. Second, we describe several new datasets. Each dataset contains
thousands of images of isolated leaves, along with segmentation information
that extracts their shape. These each include leaves of about 150–250 different
species of plants, with about 30 different leaves per species. These are by far the
largest publicly available sets of leaf images and provide a unique challenge set for
researchers on shape understanding. Third, we demonstrate recognition results
for shape matching on two of these datasets (Figure 4). This can be viewed as a
high-performance baseline system for shape matching. In this context, we pose
a challenge problem to the computer vision community. We describe a set of
performance criteria and offer to include in our deployed system code for any
algorithm that can meet these criteria.
After describing prior work in Section 2, we describe in Section 3 exten-
sive datasets that we have collected for this project, which we are now making
publicly available. In Section 4, we address a number of practical considerations

needed to get a color-based EM algorithm to effectively segment images of iso-
lated leaves. In Section 5, we summarize the shape comparison algorithm we
use. In addition, we describe a nearest-neighbor method for metric spaces that
significantly speeds up the comparisons needed for this approach. In Section 6,
we describe the hardware and UIs that we have constructed to allow the user to
navigate the search results. We also describe our ongoing work on experimental
augmented reality (AR) UIs for the system. We present a challenge problem for
the computer vision community in Section 7 and describe our plans for a future
system in Section 8.
2 Related Work
2.1 Massive Digitization Efforts
The amount of digital information available on-line has recently increased dra-
matically. For example, our group has digitally photographed (at high
resolution) each of the 90,000 type specimens of vascular plants in the US
National Herbarium at the Smithsonian, where the images are now available
at http://botany.si.edu/types/. Complementary efforts include those of the
New York Botanical Garden (120,000 high resolution images), the Royal Botan-
ical Gardens, Kew (50,000 images, including 35,000 images of type specimens),
and the Missouri Botanical Garden (35,000 images of plants). Recently, a con-
sortium of museums and research institutions announced the creation of the
Encyclopedia of Life (http://www.eol.org) to someday house a webpage for
each species of organism on Earth.
2.2 New Means to Access Data

Traditionally, biologists use field guides and dichotomous keys to assist in species
identification. Field guides contain pictures and textual descriptions of known
species. Dichotomous keys provide a decision tree based on features of the organ-
ism, with species at the leaves of the tree. Although valuable, neither solves the
problem of identification, as field guides are difficult to search, and dichotomous
keys contain questions that are daunting to the non-expert and difficult even
for experts to answer with certainty. Electronic versions of these tools have been
available for a long time (Pankhurst [16]; Edwards and Morse [6]; Stevenson et al.
[21]). Electronic keys have been created through character databases (e.g., Delta:
http://delta-intkey.com, Lucid: http://www.lucidcentral.org). Some of
these guides are available on-line or for downloading onto PDAs (e.g., Heidorn
[11]), while active websites are being developed that can continually be re-
vised and updated (e.g., http://botany.si.edu/pacificislandbiodiversity
/hawaiianflora/index.htm). While valuable, these electronic systems do not
solve the fundamental problems faced by traditional tools.
2.3 Visual Search

Automatic recognition systems promise to greatly enhance access by using im-
ages as search keys—this, we believe, is the real key to making any such electronic
field guide truly groundbreaking. There has been a good deal of work on identify-
ing plants, primarily using leaf shape (see Nilsback and Zisserman [15], though,
for recent work using flowers). Abbasi et al. [1] and Mokhtarian and Abbasi [14]
present a method for classifying images of chrysanthemum leaves. Saitoh and
Kaneko [18] use a neural network to classify wild flowers based on shape and
color. Wang et al. [23] use what they call the centroid-contour distance, combined
with more standard, global descriptions of shape. Ling and Jacobs [13] introduce
shape descriptions based on the Inner Distance, which they combine with shape
contexts (Belongie et al. [5]), and show that the resulting IDSC outperforms
many other approaches on two large leaf datasets. More recently, Felzenszwalb
and Schwartz [8] have presented a hierarchical shape matching algorithm that
performs even better on a publicly available leaf dataset (Söderkvist [20]). How-
ever, since this method is significantly slower, a fast version of the IDSC seems to
be the best approach currently available for a large-scale, real-time identification
system. We present experiments with this algorithm using data sets that are ten
times the size of those used in Ling and Jacobs [13].
This paper is the first complete description of our system. A preliminary ver-
sion of our system was described in the botanical journal Taxon [2] to introduce
these ideas to biologists. Work on UIs for automated species identification has
been described in [24], [25]. Many components of the current system have not
appeared in any previous publication, including our segmentation algorithm and
our use of nearest neighbor algorithms. Finally, our datasets and experiments
are described here for the first time.
3 Datasets
An important objective of our project is the development of standard, compre-
hensive datasets of images of individual leaves. Currently, the only large leaf
image dataset available to vision researchers is a collection of 15 species with 75
leaf images per species (Söderkvist [20]). This dataset is useful, but insufficient
for testing large-scale recognition algorithms needed for species identification.
The datasets that we have collected have an order of magnitude more species
and are well suited for testing the scalability of recognition algorithms. They also
provide complete coverage of species in a geographical area. We have made them
available for research use at http://herbarium.cs.columbia.edu/data.php.
Leaves were collected by field botanists covering all plant species native to a
particular region, and entered in the collections of the US National Herbarium.
The number of leaves per species varied with availability, but averaged about
30. After collection, each leaf was flattened by pressing and photographed with
a ruler and a color chart for calibration. Each side of each leaf was photographed
with top and bottom lighting. The leaf images were then automatically resized to
a maximum side dimension of 512 pixels. Because manual processing of multiple,
large datasets is impractical, we developed systems to automatically crop images

to remove the ruler, color chart and empty space, and then to segment the images
to separate the leaf from the background, as described in the next section. The
results were inspected by hand, and a small number of erroneously processed
images were removed from the dataset. The datasets consist of the cropped
isolated leaf images, as well as the corresponding segmented binary images. To
date, we have collected the following three single leaf datasets, each representing
different regional flora with about 30 leaves per species:
Flora of Plummers Island. 5,013 leaves of 157 species. Provides complete

coverage of all vascular plant species of Plummers Island, MD, an island in
the Potomac River near Washington, DC, which has long been studied by
botanists.
Woody Plants of Baltimore-Washington, DC. 7,481 leaves of 245 species.
Provides complete coverage of all native woody plants (trees and shrubs) of
the Baltimore-Washington, DC area.
Trees of Central Park. 4,320 leaves of 144 species. Provides complete cover-
age of the trees of Central Park in New York City.
Finally, it is often critical for botanists to access more complete type specimens
when identifying species. When a new species is discovered, a cutting of branches,
leaves, and possibly flowers and fruit is collected. This specimen becomes the type
specimen that is then used as the definitive representative of the species. Type
specimens are stored in herbaria around the world. As part of this work, we have
helped to complete the digitization of the complete Type Specimen collection of
vascular plants at the US National Herbarium:
US National Herbarium Type Specimen Collection. 90,000 images, cov-

ering more than one quarter of all known plant species. Each specimen
has been digitally photographed under controlled lighting to produce an 18
megapixel image. These are online in lower resolution formats at http://
botany.si.edu/types/.
4 Segmentation
In our automatic identification system, a user photographs a leaf so that its
shape may be matched to known species. To extract leaf shape, we must begin
by segmenting the leaf from its background. While segmentation is a well-studied
and difficult problem, we can simplify it in our system by requiring the user to
photograph an isolated leaf on a plain white background. However, while we can
require users to avoid complex backgrounds and extreme lighting conditions, a
useful segmentation algorithm must still be robust to some lighting variations
across the image and to some shadows cast by leaves.
Unfortunately, there is no single segmentation algorithm that is universally
robust and effective for off-the-shelf use. We have experimented with a number
of approaches and achieved good performance using a color-based EM algorithm
Fig. 3. The first and third images show input to the system, to the right of each
are segmentation results. We first show a typical, clean image, and then show that
segmentation also works with more complex backgrounds.
(see, e.g., Forsyth and Ponce [9]). To begin, we map each pixel to HSV color
space. Interestingly, we find that it is best to discard the hue, and represent
each pixel with saturation and value only. This is because in field tests in the
forest, we find that the light has a greenish hue that dominates the hue of an
otherwise white background. We experimented with other representations, and
colored paper backgrounds of different hues, but found that they presented some
problems in separating leaves from small shadows they cast.
Once we map each pixel to a 2D saturation-value space, we use EM to separate
pixels into two groups. First, during clustering we discard all pixels near the
boundary of the image, which can be noisy. We initialize EM using K-means
clustering with k = 2. We initialize K-means by setting the background cluster
to the median of pixels near the boundary, and setting the foreground cluster
to the mean of the central pixels. Then, in order to make the segmentation
real-time, we perform EM using 5% of the image pixels. Finally, we classify all
pixels using the two resulting Gaussian distributions. The leaf was identified as
the largest connected component of the foreground pixels, excluding components
that significantly overlap all sides of the image (sometimes, due to lighting effects,
the foreground pixels consist of the leaf and a separate connected component that
forms a band around the image). In sum, to get effective results with an EM-
based approach has required careful feature selection, initialization, sampling,
and segment classification. Figure 3 shows sample results.
Although we did not rigorously evaluate competing segmentation algorithms,
we would like to informally mention that we did encounter problems when at-
tempting to apply graph-based segmentation algorithms to these images (e.g.,
Shi and Malik [19], Galun et al. [10]). One reason for this is that these algo-
rithms have a strong bias to produce compact image segments. While this is
beneficial in many situations, it can create problems with leaves, in which the
stems and small leaflets or branches are often highly non-compact. The seg-
mentation algorithm that we use goes to the other extreme, and classifies every
pixel independently, with no shape prior, followed by the extraction of a single
connected component. It is an interesting question for future research to devise
segmentation algorithms that have shape models appropriate for objects such as
leaves that combine compact and thin, wiry structures with a great diversity of
shape.
5 Shape Matching
Our system produces an ordered list of species that are most likely to match
the shape of a query leaf. It must be able to produce comparisons quickly for
a dataset containing about 8,000 leaves from approximately 250 species. It is
useful if we can show the user some initial results within a few seconds, and the
top ten matches within a few seconds more. It is also important that we produce
the correct species within the top ten matches as often as possible, since we are
limited by screen size in displaying matches.
To perform matching, we make use of the Inner Distance Shape Context
(IDSC, Ling and Jacobs [13]), which has produced close to the best published re-
sults for leaf recognition, and the best results among those methods quick enough
to support real-time performance. IDSC samples points along the boundary of
a shape, and builds a 2D histogram descriptor at each point. This histogram
represents the distance and angle from each point to all other points, along a
path restricted to lie entirely inside the leaf shape. Given n sample points, this
produces n 2D descriptors, which can be computed in O(n3 ) time, using an all
pairs shortest path algorithm. Note that this can be done off-line for all leaves
in the dataset, and must be done on-line only for the query. Consequently, this
run-time is not significant.
To compare two leaves, each sample point in each shape is compared to all
points in the other shape, and matched to the most similar sample point. A
shape distance is obtained by summing the χ2 distance of this match over all
sample points in both shapes, which requires O(n2 ) time.
Since IDSC comparison is quadratic in the number of sample points, we would
like to use as few sample points as possible. However, IDSC performance de-
creases due to aliasing if the shape is under-sampled. We can reduce aliasing ef-
fects and boost performance by smoothing the IDSC histograms. To do this, we
compute m histograms by beginning sampling at m different, uniformly spaced
locations, and average the results. This increases the computation of IDSC for a
single shape by a factor of m. However, it does not increase the size of the final
IDSC, and so does not affect the time required to compare two shapes, which is
our dominant cost.
We use a nearest neighbor classifier in which the species containing the most
similar leaf is ranked first. Because the shape comparison algorithm does not
imbed each shape into a vector space, we use a nearest neighbor algorithm de-
signed for non-Euclidean metric spaces. Our distance does not actually obey the
triangle inequality because it allows many-to-one matching, and so it is not really
a metric (eg., all of shape A might match part of C, while B matches a different
part of C, so A and B are both similar to C, but completely different from each
other). However, in a set of 1161 leaves, we find that the triangle inequality is
violated in only .025% of leaf triples, and these violations cause no errors in
the nearest neighbor algorithm we use, the AESA algorithm (Ruiz [17]; Vidal
[22]). In this method, we pre-compute and store the distance between all pairs
of leaves in the dataset. This requires O(N 2 ) space and time, for a dataset of N
leaves, which is manageable for our datasets. At run time, a query is compared
to one leaf, called a pivot. Based on the distance to the pivot, we can use the
triangle inequality to place upper and lower bounds on the distance to all leaves
and all species in the dataset. We select each pivot by choosing the leaf with
the lowest current upper bound. When one species has an upper bound distance
that is less than the lower bound to any other species, we can select this as the
best match and show it to the user. Continuing this process provides an ordered
list of matching species. In comparison to a brute force search, which takes nine
Washington−Baltimore woody dataset Plummers Island dataset

1 1
0.9 0.9
Correct result rate
Correct result rate

0.8 0.8
0.7 0.7
0.6 0.6
256 sample points 256 sample points
64x16 sample points 64x16 sample points
0.5 0.5
64 sample points 64 sample points
0.4 0.4
2 4 6 8 10 2 4 6 8 10
Top k matches Top k matches
Fig. 4. Experimental results for two datasets
seconds with a dataset of 2004 leaves from 139 species, this nearest-neighbor
algorithm reduces the time required to find the ten best matching species by
a factor of 3, and reduces the time required to find the top three species by a
factor of 4.4.
We have tested our algorithm using both the Plummers Island and Baltimore-
Washington Woody Plants datasets. We perform a leave-one-out test, in which
each leaf is removed from the dataset and used as a query. Figure 4 shows per-
formance curves that indicate how often the correct species for a query is placed
among the top k matches, as k varies. In this experiment, we achieve best per-
formance using n = 256 sample points for IDSC. We reach nearly the same
performance by computing the histograms using n = 64 sample points averaged
over m = 16 starting points. The figure also shows that using n = 64 points
without smoothing significantly degrades performance. Using 64 sample points
is approximately 16 times faster than using 256 sample points. The correct an-
swer appears in the top ten about 95%–97% of the time for woody plants of
Baltimore-Washington and somewhat less (about 90% of the time) for the flora
of Plummers Island. This is in part because shape matching is not very effec-
tive at discriminating between different species of grass (which are not woody
plants). Overall, these results demonstrate effective performance. It seems that
most errors occur for species in which the overall leaf shape is not sufficiently
distinctive. We plan to address these issues by using additional cues, such as
small scale features of the leaf margin (e.g., toothed or smooth) and the shape
of the venation (vascular structure).
6 User Interfaces and Hardware

We have developed several prototype UIs to integrate the individual pieces of the
matching system and investigate the performance of our interaction techniques
and vision algorithms in real world situations. These prototypes are the result
of collaboration with our botanist colleagues in an iterative process that has
included ethnographic study of botanical species identification and collection in
the field, user centered design, interaction technique development, and qualita-
tive and quantitative feedback and user studies. We have pursued two primary
research directions. The first focuses on existing mobile computing platforms for
ongoing botanical field studies. The second develops mobile AR systems that are
not appropriate for field use in their current form, but could provide significant
advantages as hardware and software mature.
The conceptual model we use in our mobile computing platform is an exten-
sion of existing paper field guides. The system provides access to a library of
knowledge about the physical world, and the physical leaf is the key to that
information. In the AR prototype, virtual images representing matched species
appear adjacent to the leaf in the physical world and can be manipulated directly
through tangible interaction. In this case, the conceptual model is enhanced per-
ception: the leaf anchors information embedded in the environment and accessed
through augmented reality.
6.1 Mobile Computing

Our initial prototype, LeafView (Figure 1), provides four tabbed panes for inter-
action: browse, sample, search results, and history. The browse pane provides a
zoomable UI (ZUI) (Bederson et al. [3]) with which the user can explore an entire
flora dataset. When the user photographs a leaf with the system, the image is
immediately displayed in the sample pane with contextual information including
time, date, GPS location, and collector. The segmented image is displayed next
to the captured leaf image to show the user what LeafView “sees” and provide
feedback about image quality. As results are found, they are displayed with the
Fig. 5. AR user interface viewed through a video see-through display

original image in the search results pane. Each species result provides access to
the matched leaf, type specimens, voucher images and information about the
species in a ZUI to support detailed visual inspection and comparison, which
is necessary when matching is imperfect. Selecting a match button associates a
given species with the newly collected specimen in the collection database. The
history pane displays a visual history of each collected leaf, along with access to
previous search results, also in a ZUI. This represents the collection trip, which
can be exported for botanical research, and provides a reference for previously
collected specimens. Making this data available improves the long term use of
the system by aiding botanists in their research.
LeafView was built with C#, MatLab, and Piccolo (Bederson, et al. [4]).
Our first versions of the hardware used a Tablet PC with a separate Wi-Fi
or Bluetooth camera and a Bluetooth WAAS GPS. However, feedback from
botanists during field trials made it clear that it would be necessary to trade
off the greater display area/processing power of the Tablet PC for the smaller
size/weight of an Ultra-Mobile PC (UMPC) to make possible regular use in the
field. We currently use a Sony VAIO VGN-UX390N, a UMPC with an integrated
camera and small touch-sensitive screen, and an external GPS.
6.2 Augmented Reality

AR can provide affordances for interaction and display that are not available in
conventional graphical UIs. This is especially true of Tangible AR (Kato et al.
[12]), in which the user manipulates physical objects that are overlaid with addi-
tional information. Tangible AR is well matched to the hands-on environmental
interaction typical of botanical field research. While current head-worn displays
and tracking cannot meet the demands of daily fieldwork, we are developing
experimental Tangible AR UIs to explore what might be practical in the future.
In one of our Tangible AR prototypes (Figure 5), a leaf is placed on a clipboard
with optical tracking markers and a hand-held marker is placed next to the leaf
to initiate a search. The results of matching are displayed alongside the physical
leaf as a set of individual leaf images representing virtual vouchers, multifaceted
representations of a leaf species that can be changed through tangible gestures.
As the user passes the hand-held marker over a leaf image, the card visually
transforms into that leaf’s virtual voucher. The visual representation can be
changed, through gestures such as a circular ”reeling” motion, into images of the
type specimen, entire tree, bark, or magnified view of the plant. Inspection and
comparison is thus achieved through direct spatial manipulation of the virtual
voucher—the virtual leaf in one hand and the physical leaf on the clipboard in
the other hand. To accept a match, the virtual voucher is placed below the leaf
and the system records the contextual data.
Different versions of our Tangible AR prototypes use a monoscopic Liteye-
500 display, fixed to a baseball cap, and a stereoscopic Sony LDI-D100B display,
mounted on a head-band, both of which support 800 × 600 resolution color
imagery. The system runs on a UMPC, which fits with the display electronics
into a fannypack. The markers are tracked in 6DOF using ARToolkit (Kato et al.
[12]) and ARTag (Fiala [7]), with a Creative Labs Notebook USB 2.0 camera
attached to the head-worn display.
6.3 System Evaluation
Our prototypes have been evaluated in several ways during the course of the
project. These include user studies of the AR system, field tests on Plummers
Island, and expert feedback, building on previous work (White et al. [24]). In
May 2007, both LeafView and a Tangible AR prototype were demonstrated
and used to identify plants during the National Geographic BioBlitz in Rock
Creek Park, Washington, DC, a 24-hour species inventory. Hundreds of people,
from professional botanists to amateur naturalists, school children to congress-
men, have tried both systems. While we have focused on supporting professional
botanists, people from a diversity of backgrounds and interests have provided
valuable feedback for the design of future versions.
7 Challenge Problem for Computer Vision
One goal of our project is to provide datasets that can serve as a challenge
problem for computer vision. While the immediate application of such datasets
is the identification of plant species, the datasets also provide a rich source of
data for a number of general 2D and silhouette recognition algorithms.
In particular, our website includes three image datasets covering more than
500 plant species, with more than 30 leaves per species on average. Algorithms
for recognition can be tested in a controlled fashion via leave-one-out tests, where
the algorithms can train on all but one of the leaf images for each species and test
on the one that has been removed. The web site also contains separate training
and test datasets in order to make fair comparisons. Our IDSC code can also be
obtained there, and other researchers can submit code and performance curves,
which we will post. We hope this will pose a challenge for the community, to
find the best algorithms for recognition in this domain.
Note that our system architecture for the electronic field guide is modular,
so that we can (and will, if given permission) directly use the best performing
methods for identification, broadening the impact of that work.
8 Future Plans
To date, we have focused on three regional floras. Yet, our goal is to expand the
coverage of our system in temperate climates to include all vascular plants of
the continental U.S. Other than the efforts involved in collecting the single leaf
datasets, there is nothing that would prevent us from building a system for the
U.S. flora. The visual search component of the system scales well: search can
always be limited to consider only those species likely to be found in the current
location, as directed by GPS.
In addition, we have begun to expand into the neotropics. The Smithsonian

Center for Tropical Forest Science has set up twenty 50-hectare plots in tropical
ecosystems around the world to monitor the changing demography of tropical
forests. We aim to develop versions of the system for three neotropical floras:
Barro Colorado Island, Panama; Yasuni National Park, Ecuador; and the Ama-
zon River Basin in Brazil. This domain demands algorithms that not only con-
sider leaf shape, but also venation (i.e., the leaf’s vascular structure). Initial
results are quite promising, but we have not yet developed a working system.
Finally, we have developed a prototype web-based, mobile phone version of
our system, allowing anyone with a mobile phone equipped with a camera and
browser to photograph leaves and submit them to a server version of our system
for identification. We hope to develop a touch-based version on an iPhone or
Android-based device in the near future. We feel that it should soon be possible
to create a mobile phone-based system that covers the entire U.S., usable by the
general population.
Acknowledgements
This work was funded in part by National Science Foundation Grant IIS-03-
25867, An Electronic Field Guide: Plant Exploration and Discovery in the 21st
Century, and a gift from Microsoft Research.
References
1. Abbasi, S., Mokhtarian, F., Kittler, J.: Reliable classification of chrysanthemum
leaves through curvature scale space. In: ter Haar Romeny, B.M., Florack, L.M.J.,
Viergever, M.A. (eds.) Scale-Space 1997. LNCS, vol. 1252, pp. 284–295. Springer,
Heidelberg (1997)
2. Agarwal, G., Belhumeur, P., Feiner, S., Jacobs, D., Kress, W.J., Ramamoorthi,
R., Bourg, N., Dixit, N., Ling, H., Mahajan, D., Russell, R., Shirdhonkar, S.,
Sunkavalli, K., White, S.: First steps towards an electronic field guide for plants.
Taxon 55, 597–610 (2006)
3. Bederson, B.: PhotoMesa: A zoomable image browser using quantum treemaps and
bubblemaps. In: Proc. ACM UIST 2001, pp. 71–80 (2001)
4. Bederson, B., Grosjean, J., Meyer, J.: Toolkit design for interactive structured
graphics. IEEE Trans. on Soft. Eng. 30(8), 535–546 (2004)
5. Belongie, S., Malik, J., Puzicha, J.: Shape Matching and Object Recognition Using
Shape Context. IEEE Trans. on Patt. Anal. and Mach. Intell. 24(4), 509–522 (2002)
6. Edwards, M., Morse, D.R.: The potential for computer-aided identification in bio-
diversity research. Trends in Ecology and Evolution 10, 153–158 (1995)
7. Fiala, M.: ARTag, a fiducial marker system using digital techniques. In: Proc.
CVPR 2005, pp. 590–596 (2005)
8. Felzenszwalb, P., Schwartz, J.: Hierarchical matching of deformable shapes. In:
Proc. CVPR 2007, pp. 1–8 (2007)
9. Forsyth, D., Ponce, J.: Computer vision: A modern approach. Prentice Hall, Upper
Saddle River (2003)
10. Galun, M., Sharon, E., Basri, R., Brandt, A.: Texture segmentation by multiscale
aggregation of filter responses and shape elements. In: Proc. CVPR, pp. 716–723
(2003)
11. Heidorn, P.B.: A tool for multipurpose use of online flora and fauna: The Biological
Information Browsing Environment (BIBE). First Monday 6(2) (2001),
http://firstmonday.org/issues/issue6 2/heidorn/index.html
12. Kato, H., Billinghurst, M., Poupyrev, I., Imamoto, K., Tachibana, K.: Virtual ob-
ject manipulation of a table-top AR environment. In: Proc. IEEE and ACM ISAR,
pp. 111–119 (2000)
13. Ling, H., Jacobs, D.: Shape Classification Using the Inner-Distance. IEEE Trans.
on Patt. Anal. and Mach. Intell. 29(2), 286–299 (2007)
14. Mokhtarian, F., Abbasi, S.: Matching shapes with self-intersections: Application
to leaf classification. Proc. IEEE Trans. on Image 13(5), 653–661 (2004)
15. Nilsback, M., Zisserman, A.: A visual vocabulary for flower classification. In: Proc.
CVPR, pp. 1447–1454 (2006)
16. Pankhurst, R.J.: Practical taxonomic computing. Cambridge University Press,
Cambridge (1991)
17. Ruiz, E.: An algorithm for finding nearest neighbours in (approximately) constant
average time. Patt. Rec. Lett. 4(3), 145–157 (1986)
18. Saitoh, T., Kaneko, T.: Automatic recognition of wild flowers. Proc. ICPR 2, 2507–
2510 (2000)
19. Shi, J., Malik, J.: Normalized Cuts and Image Segmentation. IEEE Trans. on Patt.
Anal. and Mach. Intell. 22(8), 888–905 (2000)
20. Söderkvist, O.: Computer vision classification of leaves from Swedish trees. Master
Thesis, Linköping Univ. (2001)
21. Stevenson, R.D., Haber, W.A., Morris, R.A.: Electronic field guides and user
communities in the eco-informatics revolution. Conservation Ecology 7(3) (2003),
http://www.consecol.org/vol7/iss1/art3
22. Vidal, E.: New formulation and improvements of the nearest-neighbour approx-
imating and eliminating search algorithm (AESA). Patt. Rec. Lett. 15(1), 1–7
(1994)
23. Wang, Z., Chi, W., Feng, D.: Shape based leaf image retrieval. IEE Proc. Vision,
Image and Signal Processing 150(1), 34–43 (2003)
24. White, S., Feiner, S., Kopylec, J.: Virtual vouchers: Prototyping a mobile aug-
mented reality user interface for botanical species identification. In: Proc. IEEE
Symp. on 3DUI, pp. 119–126 (2006)
25. White, S., Marino, D., Feiner, S.: Designing a mobile user interface for automated
species identification. In: Proc. CHI 2007, pp. 291–294 (2007)
A Column-Pivoting Based Strategy for
Monomial Ordering in Numerical Gröbner Basis
Calculations
Martin Byröd, Klas Josephson, and Kalle Åström
Centre For Mathematical Sciences,

Lund University, Lund, Sweden
{byrod,klasj,kalle}@maths.lth.se
Abstract. This paper presents a new fast approach to improving stabil-

ity in polynomial equation solving. Gröbner basis techniques for equation
solving have been applied successfully to several geometric computer vi-
sion problems. However, in many cases these methods are plagued by
numerical problems. An interesting approach to stabilising the compu-
tations is to study basis selection for the quotient space C[x]/I. In this
paper, the exact matrix computations involved in the solution proce-
dure are clarified and using this knowledge we propose a new fast basis
selection scheme based on QR-factorization with column pivoting. We
also propose an adaptive scheme for truncation of the Gröbner basis to
further improve stability. The new basis selection strategy is studied on
some of the latest reported uses of Gröbner basis methods in computer
vision and we demonstrate a fourfold increase in speed and nearly as
good over-all precision as the previous SVD-based method. Moreover,
we get typically get similar or better reduction of the largest errors1 .
1 Introduction
A large number of geometric computer vision problems can be formulated in
terms of a system of polynomial equations in one or more variables. A typical
example of this is minimal problems of structure from motion [1,2]. This refers to
solving a specific problem with a minimal number of point correspondences. Fur-
ther examples of minimal problems are relative motion for cameras with radial
distortion [3] or for omnidirectional cameras [4]. Solvers for minimal problems
are often used in the inner loop of a RANSAC engine to find inliers in noisy data,
which means that they are run repeatedly a large number of times. There is thus
a need for fast and stable algorithms to solve systems of polynomial equations.
Another promising, but difficult pursuit in computer vision (and other fields)
is global optimization for e.g. optimal triangulation, resectioning and funda-
mental matrix estimation. See [5] and references therein. In some cases these
1
This work has been funded by the Swedish Research Council through grant no. 2005-
3230 ’Geometry of multi-camera systems’ and grant no. 2004-4579 ’Image-Based
Localization and Recognition of Scenes’.

A Column-Pivoting Based Strategy for Monomial Ordering 131
optimization problems can be solved by finding the complete set of zeros of

polynomial equations [6,7].
Solving systems of polynomial equations is known to be numerically very
challenging and there exist no stable algorithm for the general case. Instead,
specific solver algorithms are developed for each application. The state-of-the-art
method for doing this is calculations with Gröbner bases. Typically, one obtains
a floating point version of Buchberger’s algorithm [8] by rewriting the various
elimination steps using matrices and matrix operations [9]. These techniques
have been studied and applied to vision problems in a number of cases [3,10,4].
However, for larger and more demanding problems Gröbner basis calculations
are plagued by numerical problems [11,6].
A recently introduced, interesting approach to stabilisation of Gröbner ba-
sis computations is to study basis selection for the quotient space C[x]/I [12],
where I is the ideal generated by the set of equations. The choice of basis has
been empirically shown to have a great impact on numerical performance and
by adaptively selecting the basis for each instance of a problem one can obtain
a dramatic increase in stability. In [12], a scheme based on singular value de-
composition (SVD) was used to compute an orthogonal change of basis matrix.
The SVD is a numerically very stable factorization method, but unfortunately
also computationally rather expensive. Since the involved matrices tend to be
large (around hundred rows and columns or more), the SVD computation easily
dominates the running time of the algorithm.
In this paper, we propose a new fast strategy for selecting a basis for C[x]/I
based on QR-factorization with column pivoting. The Gröbner basis like com-
putations employed to solve a system of polynomial equations can essentially
be seen as matrix factorization of an under-determined linear system. Based on
this insight, we combine the robust method of QR factorization from numer-
ical linear algebra with the Gröbner basis theory needed to solve polynomial
equations. More precisely, we employ QR-factorization with column pivoting
in a crucial elimination step and obtain a simultaneous selection of basis and
triangular factorization. With this method, we demonstrate an approximately
fourfold increase in speed over the previous SVD based method while retaining
good numerical stability.
Moreover, the technique of truncating the Gröbner basis to avoid large errors
introduced in [13] fits nicely within the framework of column pivoting. Since
the pivot elements are sorted in descending order, we get an adaptive criterion
for where to truncate the Gröbner basis by setting a maximal threshold for the
quotient between the largest and the smallest pivot element. When the quotient
exceeds this threshold we abort the elimination and move the remaining columns
into the basis. This way, we expand the basis only when necessary.
Factorization with column pivoting is a well studied technique and there ex-
ist highly optimized and reliable implementations of these algorithms in e.g.
LAPACK [14], which makes this technique accessible and straight forward to
implement. Matlab code for one of the applications, optimal triangulation from
three views, is available at http://www.maths.lth.se/vision/downloads.
132 M. Byröd, K. Josephson, and K. Åström
2 Review of Gröbner Basis Techniques for Polynomial

Equation Solving
Solving systems of polynomial equations is a challenging problem in many respects
and there exist no practical numerically stable algorithms for the general case.
Instead special purpose algorithms need to be developed for specific applications.
The state-of-the-art tool for doing this is calculations with Gröbner bases.
Our general goal is to find the complete set of solutions to a system
f1 (x) = 0, . . . , fm (x) = 0, (1)
of m polynomial equations in s variables x = (x1 , . . . , xs ). The polynomials

f1 , . . . , fm generate an ideal I in C[x], the ring of multivariate polynomials in x
over the field of complex numbers defined as the set
I = {g : g(x) = Σk hk (x)fk (x)}, (2)
where the hk ∈ C[x] are any polynomials. The reason for studying the ideal I is
that it has the same set of zeros as (1).
Consider now the space of equivalence classes modulo I. This space is denoted
C[x]/I and referred to as the quotient space. Two polynomials f and g are said
to be equivalent modulo I if f = g + h, where h ∈ I. The logic behind this
definition is that we get true equality, f (x) = g(x) on zeros of (1).
To do calculations in C[x]/I it will be necessary to compute unique repre-
sentatives of the equivalence classes in C[x]/I. Let [·] : C[x] → C[x]/I denote
the function that takes a polynomial f and returns the associated equivalence
class [f ]. We would now like to compose [·] with a mapping C[x]/I → C[x] that
associates to each equivalence class a unique representative in C[x]. The com-
posed map C[x] → C[x] should in other words take a polynomial f and return
the unique representative f for the equivalence class [f ] associated with f . As-
sume for now that we can compute such a mapping. This operation will here be
referred to as reduction modulo I.
A well known result from algebraic geometry now states that if the set of
equations (1) has r zeros, then C[x]/I will be a finite-dimensional linear space
with dimension r [8]. Moreover, an elegant trick based on calculations in C[x]/I
yields the complete set of zeros of (1) in the following way: Consider multipli-
cation by one of the variables xk . This is a linear mapping from C[x]/I to itself
and since we are in a finite-dimensional space, by selecting an appropriate basis,
this mapping can be represented as a matrix mxk . This matrix is known as the
action matrix and the eigenvalues of mxk correspond to xk evaluated at the
zeros of (1) [8]. Moreover, the eigenvectors of mxk correspond the vector of basis
monomials/polynomials evaluated at the same zeros and thus the complete set
of solutions can be directly read off from these eigenvectors. The action matrix
can be seen as a generalization of the companion matrix to the multivariate case.
Given a linear basis B = {[ei ]}ri=1 spanning C[x]/I, the action matrix mxk is
computed by calculating xk ei for each of the basis elements ei . Performing this
operation is the difficult part in the process. Traditionally, the reduction has
been done by fixing a monomial ordering and then computing a Gröbner basis
G for I, which is a canonical set of polynomials that generate I. Computing f
G
is then done by polynomial division by G (usually written f ).
We now make two important observations: (i) We are not interested in finding
the Gröbner basis per se; it is enough to get a well defined mapping f and (ii)
it suffices to calculate reduction modulo I on the elements xk ei , i.e. we do not
need to know what f is on all of C[x]. Note that if for some i, xk ei ∈ B then
nothing needs to be done for that element. With this in mind, we denote by
R = xk B \ B the set of elements f for which we need to calculate representatives
f of their corresponding equivalence classes [f ] in C[x]/I.
Calculating the Gröbner basis of I is typically accomplished by Buchberger’s
algorithm. This works well in exact arithmetic. However, in floating point arith-
metic Buchberger’s algorithm very easily becomes unstable. There exist some
attempts to remedy this [15,16], but for more difficult cases it is necessary to
study a particular class of equations (e.g. relative orientation for omnidirectional
cameras [4], optimal three view triangulation [6], etc.) and use knowledge of what
the structure of the Gröbner basis should be to design a special purpose Gröbner
basis solver [9].
In this paper we move away from the goal of computing a Gröbner basis for
I and focus on computing f for f ∈ R as mentioned above. However, it should
be noted that the computations we do much resemble those necessary to get a
Gröbner basis.
2.1 Computing Representatives for C[x]/I

In this section we show how representatives for C[x]/I can be efficiently calcu-
lated in floating point arithmetic. The reason why Buchberger’s algorithm breaks
down in floating arithmetic is that eliminations of monomials are performed suc-
cessively and this causes round-off errors to accumulate to the point where it is
impossible to tell whether a certain coefficient should be zero or not. The trick
introduced by Faugere [15] is to write the list of equations on matrix form
CX = 0, (3)
# $t
where X = xα1 . . . xαn is a vector of monomials with the notation xαk =
xα1
k1
· · · xα
s
ks
and C is a matrix of coefficients. Elimination of leading terms now
translates to matrix operations and we then have access to a whole battery of
techniques from numerical linear algebra allowing us to perform many elimina-
tions at the same time with control on pivoting etc.
By combining this approach with knowledge about a specific problem obtained
in advance with a computer algebra system such as Macaulay2 [17] it is possible
to write down a fixed number of expansion/elimination steps that will generate
the necessary polynomials.
In this paper, we use a linear basis of monomials B = {xα1 , . . . , xαr } for
C[x]/I. Recall now that we need to compute xk xαi for xk xαi ∈ / B, i.e. for R.
This is the aim of the following calculations.
Begin by multiplying the equations (1) by a large enough set of monomials

producing an equivalent (but larger) set of equations. We will come back to what
large enough means. Thereafter, stack the coefficients of the new equations in
an expanded coefficient matrix Cexp , yielding
Cexp Xexp = 0. (4)
Now partition the all monomials M occurring in the expanded set of

set of
equations as M = E R B and order them so that E > R > B holds for all
monomials in their respective sets. The monomials E (E for excessive) are simply
the monomials which are neither in R nor in B. This induces a corresponding
partitioning and reordering of the columns of Cexp :
⎡ ⎤
# $ XE
CE CR CB ⎣XR ⎦ = 0. (5)
XB
The E-monomials are not in the basis and do not need to be reduced so we
eliminate them by an LU decomposition on Cexp yielding
⎡ ⎤
! " X
UE1 CR1 CB1 ⎣ E ⎦
XR = 0, (6)
0 UR2 CB2
XB
where UE1 and UR2 are upper triangular. We can now discard the top rows of
the coefficient matrix producing
! "
# $ XR
UR2 CB2 = 0, (7)
XB
from which we get the elements of the ideal I we need since equivalently, if the
submatrix UR2 is of full rank, we have
XR = −U−1
R2 CB2 XB (8)
and then the R-monomials can be expressed uniquely in terms of the B-monomials.
As previously mentioned, this is precisely what we need to compute the action ma-
trix mxk in C[x]/I. In other words, the property of UR2 as being of full rank is
sufficient to get the operation f on the relevant part of C[x]. Thus, in designing
the set of monomials to multiply with (the first step in the procedure) we can use
the rank of UR2 as a criterion for whether the set is large enough or not. How-
ever, the main problem in these computations is that even if UR2 is in principle
invertible, it can be very ill conditioned.
A technique introduced in [12], which alleviates much of these problems uses
basis selection for C[x]/I. The observation is that the right linear basis for C[x]/I
induces a reordering of the monomials, which has the potential to drastically
improve the conditioning of UR2 . Since Cexp depends on the data, the choice of
linear basis cannot be made on beforehand, but has to be computed adaptively
each time the algorithm is run. This leads to the difficult optimisation problem
of selecting a linear basis so as to minimize the condition number of UR2 . In [12]
this problem was addressed by making use of SVD providing a numerically stable,
but computationally expensive solution.
The advantage of the above exposition is that it makes explicit the dependence
on the matrix UR2 , both in terms of rank and conditioning. In particular, the
above observations leads to the new fast strategy for basis selection which is the
topic of the next section and a major contribution of this paper.
3 Column Pivoting as Basis Selection Strategy

In the one-variable case the monomials are given a natural ordering by their
degree. In the multivariate case, there are several ways to order the monomials.
To specify representatives for C[x]/I, one traditionally fixes one of these. The
monomial order then automatically produces a linear basis for C[x]/I in form
of the set of monomials which are not divisible by the Gröbner basis in that
monomial order.
For Buchberger’s algorithm to make sense a monomial order is required to
respect multiplication, i.e. xα > xβ ⇒ xk xα > xk xβ . Interestingly, when we
relax the requirement of getting a strict Gröbner basis and compute f as out-
lined in the previous section, this property is unnecessarily strict. The crucial
observation is that we can choose any linear basis for C[x]/I we like, as long as
we are able to compute well defined representatatives for the equivalence classes
of C[x]/I. Thus, instead of letting the monomial order dictate the linear basis,
we would like to do it the other way around and start by choosing a set of basis
monomials B.
After noting that we have some freedom in choosing B, the first question is
which monomials P (for permissible) in M are eligible for inclusion in the linear
basis? Since we have to reduce the set xk B \ B to C[x]/I we obviously have to
require xk P ⊂ M. Moreover, by making the construction leading up to (8), but
replacing B by P we see that again the resulting UR2 needs to be of full rank
to be able to guarantee reduction modulo I for all elements.
With
# these properties
$ in place we aim at selecting P as large as possible and
form CE CR CP . Any selection of basis monomials B ⊂ P will then correspond
to a matrix CB consisting of a subset of the columns of CP .
By again performing an LU factorization and discarding the top rows to get
rid of the E-monomials, we get
! "! "
UR2 CP2 XR
= 0, (9)
0 CP3 XP
in analogy with (7), where we now get zeros below UR2 since the larger CP
means that we can still eliminate further. This is where the basis selection comes
to play.
As noted above we can choose which monomials of the p monomials in P to put
in the basis and which to reduce. This is equivalent to choosing a permutation
Π of the columns of CP3 so that
CP3 Π = [ cπ(1) ... cπ(p) ]. (10)

The goal must thus be to make this choice so as to minimize the condition
number κ([ cπ(1) ... cπ(p−r) ]) of the first p − r columns of the permuted matrix. In
its generality, this is a difficult combinatorial optimization problem. However,
the task can be approximately solved in an attractive way by QR factorization
with column pivoting [18]. With this algorithm, CP3 is factorized as
CP3 Π = QU, (11)
where Q is orthogonal and U is upper triangular. By solving for CP3 in % (11)&

and substituting into (9) followed by multiplication from the left with 0I Q0t
and from the right with [ 0I Π
0 ], we get
! "! "
UR2 CP2 Π XR
= 0, (12)
0 U Π t XP
# We observe
$ that U is not quadratic and emphasize this by writing U =
UP3 CB2 , where UP3 is quadratic upper triangular. We also write CP2 Π =
# $ # $t
CP4 CB1 and Π t XP2 = XP XB yielding
⎡ ⎤
! " XR
UR2 CP4 CB1 ⎣
XP ⎦ = 0 (13)
0 UP3 CB2
XB
and finally
! " ! "−1 ! "
XR UR2 CP4 CB 1
=− XB (14)
XP 0 UP3 CB 2
is the equivalent of (8) and amounts to solving r upper triangular equation
systems which can be efficiently done by back substitution.
The reason why QR factorization fits so nicely within this framework is that
it simultaneously solves the two tasks of reduction to upper triangular form and
numerically sound column permutation and with comparable effort to normal
Gaussian elimination.
Furthermore, QR factorization with column pivoting is a widely used and
well studied algorithm and there exist free, highly optimized implementations,
making this an accessible approach.
Standard QR factorization successively eliminates elements below the main
diagonal by multiplying from the left with a sequence of orthogonal matrices
(usually Householder transformations). For matrices with more columns than
rows (under-determined systems) this algorithm can produce a rank-deficient
U which would then cause the computations in this section to break down.
QR with column pivoting solves this problem by, at iteration k, moving the
column with greatest 2-norm on the last m − k + 1 elements to position k and
then eliminating the last m − k elements of this column by multiplication with
an orthogonal matrix Qk .
3.1 Adaptive Truncation
A further neat feature of QR factorization with column pivoting is that it pro-

vides a way of numerically estimating the conditioning of CP simultaneously
with the elimination. In [13], it was shown that for reductions with a Gröbner
basis, the Gröbner basis could be truncated yielding a larger representation of
C[x]/I (more than r basis elements), while retaining the original set of solutions.
The advantage of this is that the last elements of the Gröbner basis often are
responsible for a major part of the numerical instability and making use of the
observation in [13], the last elements do not have to be computed.
As discussed earlier we do not calculate exactly a Gröbner basis, but the
method of [13] is straightforward to adapt to the framework of this paper. How-
ever, both rank and conditioning of CP might depend on the data and we would
therefore like to decide adaptively where to truncate, i.e. when to abort the QR
factorization.
As a consequence of how the QR algorithm is formulated, the elements ukk
on the main diagonal of U will be sorted in decreasing absolute value. In exact
arithmetic, if the rank is q, then ukk = 0 for k > q. In floating point this
will not be the case due to round-off errors. However, we can set a threshold τ
and abort the elimination process once |ukk |/|u11 | < τ . The remaining columns
(monomials) are then transfered to the basis which is correspondingly expanded.
Apart from being numerically sound, this strategy also spares some compu-
tational effort compared to setting a fixed larger basis. Truncating the set of
polynomials means a higher dimensional representation of C[x]/I, which means
we have to solve a larger eigenvalue problem. As will be shown in the exper-
iments, the basis can usually be kept tight and only needs to be expanded in
some cases.
4 Experiments
The purpose of this section is to verify the speed and accuracy of the QR-
method. To this end, three different applications are studied. The first example is
relative pose for generalised cameras, first solved by Stewénius et al . in 2005 [19].
The second one is the previously unsolved minimal problem of pose estimation
with unknown focal length. The problem was formulated by Josephson et al .
in [20], but not solved in floating point arithmetic. The last problem is optimal
triangulation from three views [6].
Since the techniques described in this paper improve the numerical stability of
the solver itself, but do not affect the conditioning of the actual problem, there is
no point in considering the behavior under noise. Hence we will use synthetically
generated examples without noise to compare the intrinsic numerical stability of
the different methods.
In all three examples we compare with the “standard” method, by which we
mean to fix a monomial order (typically grevlex) and use the basis dictated
by that order together with straightforward gauss jordan elimination to express
monomials in terms of the basis. Previous works have often used several ex-
pansion / elimination rounds. We have found this to have a negative effect on
numerical stability so to make the comparison fair, we have implemented the
standard method using a single elimination step in all cases.
For the adaptive truncation method, the threshold τ for the ratio between the
k:th diagonal element and the first was set to 10−8 .
4.1 Relative Pose for Generalised Cameras
A generalised camera is a camera with no common focal point. This e.g. serves
as a useful model for several ordinary cameras together with fixed relative loca-
tions [21]. For generalised cameras there is a minimal case for relative pose with
two cameras and six points. This problem was solved in [19] and has 64 solu-
tions. In [12] this problem was used to show how the SVD-method improved the
numerics. We follow the methods of the later paper to get a single elimination
step. This gives an expanded coefficient matrix of size 101×165 with the columns
representing monomials up to degree eight in three variables. For details see [19]
and [12].
The examples for this experiment were generated by picking six points from
a normal distribution centered at the origin. Then six randomly chosen lines
through these point were associated to each camera. This made up two gener-
alised cameras with a relative orientation and translation.
Following this recipe, 10000 examples were generated and solved with the
standard, QR- and SVD-method. The angular errors between true and estimated
motion were measured. The results are shown in Figure 1.
The method with variable basis size was also implemented, but for this exam-
ple the UR2 (see Equation 7) part of the coefficient matrix was always reasonably
conditioned and hence the basis size was 64 in all 10000 test examples. There
were no large errors for neither the SVD nor the QR method.
0.4
Standard
0.35 SVD
QR
0.3
0.25
Frequency
0.2
0.15
0.1
0.05
0
−15 −10 −5 0
Log of angular error in degrees
10
Fig. 1. Error distributions for the problem of relative pose with generalised cameras.
The SVD-method yields the best results but the faster QR-method is not far behind
and also eliminates all large errors.
4.2 Localisation with Hybrid Features
This problem was introduced in [20]. The problem is to find the pose of a cali-
brated camera with unknown focal length. One minimal setup for this problem
is three point-correspondences with known world points and one correspondence
to a world line. The last feature is equivalent to having a point correspondence
with another camera. These types of mixed features are called hybrid features.
In [20], the authors propose a parameterisation of the problem but no solution
was given apart from showing that the problem has 36 solutions.
The parameterisation in [20] gives four equations in four unknowns. The un-
knowns are three quaternion parameters and the focal length. The equation
derived from the line correspondence is of degree 6 and those obtained from the
3D points are of degree 3. The coefficient matrix Cexp is then constructed by
expanding all equations up to degree 10. This means that the equation derived
from the line is multiplied with all monomials up to degree 4, but no single
variable in the monomials is of higher degree than 2. In the same manner the
point correspondence equations are multiplied with monomials up to degree 7
but no single variable of degree more than 5. The described expansion gives 980
equations in 873 monomials.
The next step is to reorder the monomials according to (5). In this problem
CP corresponds to all monomials up to degree 4 except f 4 where f is the focal
length, this gives 69 columns in CP . The part CR corresponds to the 5:th degree
monomials that appears when the monomials in B are multiplied with the first
of the unknown quaternion parameters.
For this problem, we were not able to obtain a standard numerical solver. The
reason for this was that even going to significantly higher degrees than mentioned
above, we did not obtain an invertible UR2 . In fact, with an exact linear basis
(same number of basis elements as solutions), even the QR and SVD methods
failed and truncation had to be used.
In this example we found that increasing the linear basis of C[x]/I by a few
elements over what was produced by the adaptive criterion was beneficial for the
stability. In this experiment, we added three basis elements to the automatically
produced basis. To get a working version of the SVD solver we had to adapt the
truncation method to the SVD case as well. We did this by looking at the ratio
of the singular values.
The synthetic experiments for this problem were generated by randomly draw-
ing four points from a cube with side length 1000 centered at the origin and two
cameras with a distance of approximately 1000 to the origin. One of these cam-
eras was treated as unknown and one was used to get the camera to camera
point correspondence. This gives one unknown camera with three point corre-
spondences and one line correspondence. The experiment was run 10000 times.
In Figure 2 (right) the distribution of basis sizes is shown for the QR-method.
For the SVD-method the basis size was identical to the QR-method in over 97%
of the cases and never differed by more than one element.
Figure 2 (left) gives the distribution of relative errors in the estimated focal
length. It can be seen that both the SVD-method and the faster QR-method
Fig. 2. Left: Relative error in focal length for pose estimation with unknown focal
length. Both the SVD- and QR-methods uses adaptive truncatation. Right: The size of
the adaptively chosen basis for the QR-method. For the SVD-method the size differs
from this in less than 3% of the cases and by at most one element.
give useful results. We emphasize that we were not able to construct a solver
with the standard method and hence no error distribution for that method is
available.
4.3 Optimal Triangulation from Three Views

The last experiment does not concern a geometrical minimal case, but instead
deals with an optimisation problem. Given noisy image measurements in three
views, the problem is to find the world point that minimises the sum of squares
of reprojection errors. This is the statistically optimal estimate under Gaussian
noise.
We find the global minimum by calculating the complete set of stationary
points of the reprojection error function. This was first done in [6], where the
standard Gröbner basis method was used. However, because of numerical prob-
lems they were forced to use extremely slow, emulated 128 bit numerics to get
accurate results. In [12] the SVD-method was later used to enable calculations
in standard double precision. It should be mentioned that a more recently intro-
duced and probably more practical global optimisation method for triangulation
is given in [22]. Still though, this problem serves as an interesting test bed for
equation solving.
For details on the construction of the coefficient matrix see [12,6]. The coeffi-
cient matrix constructed with this method is of size 225 × 209 and the number
of solutions is 50. The QR-method was implemented as described earlier and
the method with variable size basis was used. For reference, we implemented the
method of [6] in standard double precision, with some small modifications to get
a single elimination step (this made it slightly more stable).
The unknown point was randomly placed in a cubic box with side 1000 centered
around the origin. The three cameras were placed approximately on a sphere with
distance 1000 from origin and the focal lengths were also set to around 1000. The
error in 3D placement over 10000 iterations is shown in Figure 3. It can be seen
that the QR-method is almost as accurate as the SVD-method.
One important property of a solver is that the number of large errors is small.
Thus in Table 1 the number of large errors are shown. The results show that the
QR-method is better at suppressing large errors, probably due to the variable
size of the basis.
Fig. 3. The distribution of the error in 3D placement of the unknown point using op-
timal three view triangulation. The experiment was run 10000 times. The QR-method
gives nearly identical results compared to the SVD-method.
Table 1. Number of errors larger than some levels. This shows that the QR-method
gives fewer large errors probably due to the variable size of the basis.
Error > 1 > 10−1 > 10−2 > 10−3

QR 10 15 28 54
SVD 31 39 52 79
Table 2. Number of times a certain basis size appears in 10000 iterations. The largest
basis size obtained in the experiment was 66.
Basis size 50 51 52 53 54 55 ≥ 56
# 9471 327 62 34 26 17 58
4.4 Speed Comparison
In the problem of optimal three view triangulation the execution times for the
three different algorithms were measured. Since the implementations were done
in Matlab it was necessary to take care to eliminate the effect of Matlab being an
interpreted language. To do this only the time after construction of the coefficient
matrix was taken into account. This is because the construction of the coefficient
matrix essentially amounts to copying coefficients to the right places which can
be done extremely fast in e.g. a C language implementation.
In the routines that were measured no subroutines were called that were not
built-in functions in Matlab. The measurements were done with Matlab’s profiler.
The time measurements were done on an Intel Core 2 2.13 GHz machine
with 2 GB memory. Each algorithm was executed with 1000 different coefficient
matrices, these were constructed from the same type of scene setup as in the
previous section. The same set of coefficient matrices was used for each method.
The result is given in Table 3. Our results show that the QR-method with adap-
tive truncation is approximately four times faster than the SVD-method but
40% slower than the standard method. It should however be noted that here,
the standard method is by far too inaccurate to be of any practical value.
Table 3. Time consumed in the solver part for the three different methods. The time
is an average over 1000 calls.
Method Time per call / ms Relative time

SVD 41.685 1
QR 10.937 0.262
Standard 8.025 0.193
5 Conclusions
In this paper we have presented a new fast strategy for improving numerical
stability of Gröbner basis polynomial equation solvers. The key contribution is a
clarification of the exact matrix operations involved in computing an action ma-
trix for C[x]/I and the use of numerically sound QR factorization with column
pivoting to obtain a simultaneous basis selection for C[x]/I and reduction to up-
per triangular form. We demonstrate a nearly fourfold decrease in computation
time compared to the previous SVD based method while retaining good nu-
merical stability. Moreover, since the method is based on the well studied, freely
available QR algorithm it is reasonably simple to implement and not much slower
than using no basis selection at all.
The conclusion is thus that whenever polynomial systems arise and numerical
stability is a concern, this method should be of interest.
References
1. Chasles, M.: Question 296. Nouv. Ann. Math. 14 (1855)

2. Kruppa, E.: Zur Ermittlung eines Objektes aus Zwei Perspektiven mit innerer
Orientierung. Sitz-Ber. Akad. Wiss., Wien, math. naturw. Kl. Abt IIa, 1939–1948
(1913)
3. Kukelova, Z., Pajdla, T.: A minimal solution to the autocalibration of radial dis-
tortion. In: CVPR (2007)
4. Geyer, C., Stewénius, H.: A nine-point algorithm for estimating para-catadioptric

fundamental matrices. In: CVPR, Minneapolis, USA (2007)
5. Hartley, R., Kahl, F.: Optimal algorithms in multiview geometry. In: Yagi, Y.,
Kang, S.B., Kweon, I.S., Zha, H. (eds.) ACCV 2007, Part I. LNCS, vol. 4843, pp.
6. Stewénius, H., Schaffalitzky, F., Nistér, D.: How hard is three-view triangulation
really? In: Proc. Int. Conf. on Computer Vision, Beijing, China, pp. 686–693 (2005)
7. Hartley, R., Sturm, P.: Triangulation. Computer Vision and Image Understand-
ing 68, 146–157 (1997)
8. Cox, D., Little, J., O’Shea, D.: Ideals, Varieties, and Algorithms. Springer, Heidel-
berg (2007)
9. Stewénius, H.: Gröbner Basis Methods for Minimal Problems in Computer Vision.
PhD thesis, Lund University (2005)
10. Stewénius, H., Kahl, F., Nistér, D., Schaffalitzky, F.: A minimal solution for relative
pose with unknown focal length. In: Proc. Conf. Computer Vision and Pattern
Recognition, San Diego, USA (2005)
11. Kukelova, Z., Pajdla, T.: Two minimal problems for cameras with radial distortion.
In: Proceedings of The Seventh Workshop on Omnidirectional Vision, Camera
Networks and Non-classical Cameras (OMNIVIS) (2007)
12. Byröd, M., Josephson, K., Åström, K.: Improving numerical accuracy of gröbner
basis polynomial equation solvers. In: Proc.11th Int. Conf. on Computer Vision,
Rio de Janeiro, Brazil (2007)
13. Byröd, M., Josephson, K., Åström, K.: Fast optimal three view triangulation. In:
Asian Conference on Computer Vision (2007)
14. Anderson, E., et al.: LAPACK Users’ Guide. Third edn. Society for Industrial and
Applied Mathematics, Philadelphia, PA (1999)
15. Faugère, J.C.: A new efficient algorithm for computing gröbner bases (f4 ). Journal
of Pure and Applied Algebra 139, 61–88 (1999)
16. Faugère, J.C.: A new efficient algorithm for computing gröbner bases without re-
duction to zero (f5). In: ISSAC 2002, pp. 75–83. ACM Press, New York (2002)
17. Grayson, D., Stillman, M.: Macaulay 2 (1993)-2002),
http://www.math.uiuc.edu/Macaulay2
18. Golub, G.H., van Loan, C.F.: Matrix Computations, 3rd edn. The Johns Hopkins
University Press (1996)
19. Stewénius, H., Nistér, D., Oskarsson, M., Åström, K.: Solutions to minimal gener-
alized relative pose problems. In: OMNIVIS, Beijing, China (2005)
20. Josephson, K., Byröd, M., Kahl, F., Åström, K.: Image-based localization using
hybrid feature correspondences. In: BenCOS 2007 (2007)
21. Pless, R.: Using many cameras as one. In: Proc. Conf. Computer Vision and Pattern
Recognition, Madison, USA (2003)
22. Lu, F., Hartley, R.: A fast optimal algorithm for l2 triangulation. In: Yagi, Y.,
Kang, S.B., Kweon, I.S., Zha, H. (eds.) ACCV 2007, Part II. LNCS, vol. 4844, pp.
Co-recognition of Image Pairs
by Data-Driven Monte Carlo Image Exploration
Minsu Cho, Young Min Shin, and Kyoung Mu Lee
Department of EECS, ASRI, Seoul National University, 151-742, Seoul, Korea

minsucho@diehard.snu.ac.kr, shinyoungmin@gmail.com, kyoungmu@snu.ac.kr
Abstract. We introduce a new concept of ‘co-recognition’ for object-level

image matching between an arbitrary image pair. Our method augments
putative local region matches to reliable object-level correspondences with-
out any supervision or prior knowledge on common objects. It provides
the number of reliable common objects and the dense correspondences be-
tween the image pair. In this paper, generative model for co-recognition is
presented. For inference, we propose data-driven Monte Carlo image ex-
ploration which clusters and propagates local region matches by Markov
chain dynamics. The global optimum is achieved by a guiding force of our
data-driven sampling and posterior probability model. In the experiments,
we demonstrate the power and utility on image retrieval and unsupervised
recognition and segmentation of multiple common objects.
1 Introduction
Establishing correspondences between image pairs is one of the fundamental and
crucial issues for many vision problems. Although the development of various
kinds of local invariant features [1,2,3] have brought about notable progress
in this area, their local ambiguities remain hard to be solved. Thus, domain
specific knowledge or human supervision has been generally required for accurate
matching. Obviously, the best promising strategy to eliminate the ambiguities
from local feature correspondences is to go beyond locality [4,5,6,7]. The larger
image regions we exploit, the more reliable correspondences we can obtain. In
this work we propose a novel data-driven Monte Carlo framework to augment
naive local region correspondences to reliable object-level correspondences in an
arbitrary image pair. Our method establishes multiple coherent clusters of dense
correspondences to achieve recognition and segmentation of multiple common
objects without any prior knowledge of specific objects.
For the purpose, we introduce a perceptually meaningful entity, which can be
interpreted as a common object or visual pattern. We will refer to the entity in an
image pair as a Maximal Common Saliency (MCS) and define it as follows: (1) An
MCS is a semi-global region pair, composed of local region matches between the
image pair. (2) The region pair should be mutually consistent in geometry and
photometry. (3) Each region of the pair should be maximal in size. Now, the goal
of our work is defined to obtain the set of MCSs from an image pair. According
to the naming conventions of some related works [5,8], we term it co-recognition.

Co-recognition of Image Pairs 145
Fig. 1. Result of co-recognition on our dataset Mickey’s. Given an image pair, co-
recognition detects all Maximal Common Saliencies without any supervision or prior
knowledge. Each color represents each identity of the MCS, which means an object in
this case. Note that the book (blue) is separated by occlusion but identified as one
object. See the text for details.
As shown in Fig. 1, co-recognition is equivalent to recognizing and segmenting

multiple common objects in a given image pair under two conditions: (1) All
the common object appears mutually distinctive in geometry. (2) Each common
object lies on different backgrounds in photometry.1 Note that it can detect
separated regions by occlusion as a single object without any prior knowledge
or supervision. In this problem, local region correspondences can be established
more easily if reliable poses of common objects are known in advance, and the
converse is also true. We pose this chicken-and-egg problem in terms of data-
driven Monte Carlo sampling with reversible jump dynamics [9,10] over the intra-
and inter-image domain. A main advantage of our formulation is to combine
bottom-up and top-down processes in a integrated and principled way. Thus,
global MCS correspondences and their local region correspondences reinforce
each other simultaneously so as to reach global optimum.
Among recent works related to ours are co-segmentation [8], co-saliency [5], and
common visual pattern discovery [11]. Rother et al. [8] defined co-segmentation
as segmenting common regions simultaneously in two images. They exploited a
generative MRF-based graph model and the color histogram similarity measure.
Toshev et al. [5] defined co-saliency matching as searching for regions which have
strong intra-image coherency and high inter-image similarity. The method takes
advantage of the segmentation cue to address the ambiguity of local feature match-
ing. Yuan and Wu [11] used spatial random partition to discover common visual
patterns from a collection of images. The common pattern is localized by aggre-
gating the matched set of partitioned images. However, none of these methods
recognize multiple common objects as distinct entities. Moreover, [8] and [11] do
not consider geometrical consistency in the detected region.
1
With first condition unsatisfied, several distinct common objects can be recognized
as one. With second unsatisfied, objects can include a portion of similar background.
146 M. Cho, Y.M. Shin, and K.M. Lee
(a) Overview of our approach (b) Initial matching and latent regions
Fig. 2. (a) Given two images, data-driven Monte Carlo image exploration solves co-
recognition problem of the image pair. See the text for details. (b) Top: Several different
types of local features can be used for initial matches. Bottom: Overlapping circular
regions are generated covering the whole reference image for latent regions.
Our method has been inspired by the image exploration method for object
recognition and segmentation, proposed by Ferrari et al [6]. The method is based
on propagating initial local matches to neighboring regions by their affine homog-
raphy. Even with few true initial matches, their iterative algorithm expands in-
liers and contracts outliers so that the recognition can be highly improved.2 The
similar correspondence growing approaches were proposed also in [4,7] for non-
rigid image registration. Our new exploration scheme leads the image exploration
strategy of [6] to unsupervised multi-object image matching by the Bayesian for-
mulation and the DDMCMC framework [9]. Therefore, the co-recognition prob-
lem addressed by this paper can be viewed as a generalization of several other
problems reported in the literature [5,6,8,11].
2 Overview of the Approach

Given an image pair, the goal of co-recognition is to recognize and segment
all the MCSs and infer their dense correspondences in the pair simultaneously.
Figure 2(a) illustrates the overview of our algorithm. First, we obtain initial affine
region matches using several different types of local affine invariant features [3,2].
Then, each initial match forms an initial cluster by itself, which is a seed for an
MCS. Second, one of the pair is set to be the reference image, and we generate
2
Although recent object recognition and segmentation methods [12,13] based on local
region features demonstrate more accurate results in segmentation, they require
enough inliers to localize the object in their initialization step.
a grid of overlapping circular regions covering the whole reference image. All
the overlapping regions are placed into a latent region set Λ, in which each
element region waits to be included in one of the existing clusters (Fig. 2(b)).
After these initialization steps, our data-driven Monte Carlo image exploration
algorithm starts to search for the set of MCSs by two pairs of reversible moves;
expansion/contraction and merge/split. In the expansion/contraction moves, a
cluster obtains a new match or lose one. In merge/split moves, two clusters are
combined into one cluster, or one cluster is divided into two clusters. Utilizing all
these moves in a stochastic manner, our algorithm traverses the solution space
efficiently to find the set of MCSs. The final solution is obtained by eliminating
trivial MCSs from the result.
3 Generative Model of Co-recognition

We formulate co-recognition as follows. The set of MCSs is denoted by a vector
of unknown variables θ which consists of clusters of matches:
θ = (K, {Γi ; i = 1, ..., K}), (1)
where Γi represents a cluster of matches, K means the number of clusters. Γi
consists of local region matches across the image pair, expressed as follows:
Γi = {(Rj , Tj ); j = 1, ..., Li }, (2)
where Rj denotes a small local region of the reference image, Tj indicates an
affine transformation that maps the region Rj to the other image.3 Li denotes
the number of local regions included in the cluster Γi .
In the Bayesian framework, we denote the posterior probability p(θ|I) as the
probability of θ being the set of MCSs given an image pair I, which is pro-
portional to the product of the prior p(θ) and the likelihood p(I|θ). Therefore,
co-recognition is to find θ∗ that maximizes this posterior as follows.
θ∗ = arg max p(θ|I) = arg max p(I|θ)p(θ). (3)
θ θ
3.1 The Prior p(θ)

The prior p(θ) models the geometric consistency and the maximality of MCSs.
Geometric Consistency of MCSs. To formulate the geometric constraint of a
cluster Γi , we used the sidedness constraint of [6], and reinforced it with orienta-
tion consistency. Consider a triple (Rj , Rk , Rl ) of local regions in the reference im-
age and their corresponding regions (Rj , Rk , Rl )4 in the other image. Let cj , cj
be the centers of regions Rj , Rj , respectively. Then, the sidedness constraint,
sign((ck × cl )cj ) = sign((ck × cl )cj ) (4)
3
Registration of non-planar 3-d surfaces is approximated by a set of linear transfor-
mations of small local regions.
4
That is, Ri = Ti Ri.
(a) Sidedness constraint b) Reinforced with orientation consistency
Fig. 3. (a) 1 should be on the same side of the directed line from 2 to 3 in both images.
(b) 4,5 and 1 satisfies sidedness constraint, while it does not lies on the same object.
We can filter out this outlier triplet by checking if orientation(red arrow) changes in
the triplet are mutually consistent.
means that the side of cj w.r.t the directed line (ck × cl ) should be just the same
as the side of cj w.r.t the directed line (ck × cl ) (Fig. 3(a)). This constraint
holds for all correctly matching triplets of coplanar regions. Since the sidedness
constraint is valid even for most non-planar regions, it is useful for sorting out
triplets on a common surface. As illustrated in Fig. 3(b), we reinforce it with
orientation consistency to deal with multiple common surfaces for our problem
as follows:
∀(m, n) ∈ {(j, k), (j, l), (k, l)}, |angle(angle(om , om ), angle(on , on ))| < δori
(5)
where om means the dominant orientation of Rm in radian, while angle() denotes
the function which calculates the clockwise angle diffrence in radian. Hence, the
reinforced sidedness error with the orientation consistency is defined by

0 if (4) and (5) hold
errside (Rj , Rk , Rl ) = (6)
1 otherwise
A triple violating the reinforced sidedness constraint has higher chances of having
one or more mismatches in it. The geometric error of Rj (∈ Γi ) is defined by the
share of violations in its own cluster such that
1
errgeo (Rj ) = errside (Rj , Rk , Rl ), (7)
v
Rk ,Rl ∈Γi \Rj ,k>l
where v = (Li −1)(Li −2)/2 is the normalization factor that counts the maximum
number of violations. When Li < 3, errgeo (Rj ) is defined as 1 if the cluster
Γi ( Rj ) violates the orientation consistency, otherwise 0.
The geometric error of a cluster is then defined by the sum of errors for all
members in the cluster as follows:

Li
errgeo (Γi ) = errgeo (Rj ). (8)
j=1
Maximality of MCSs. To encode the degree of maximality of θ, the relative

area of each cluster should be examined. We approximate it by the number of
matches in each cluster since all the latent regions have the same area and the
number is constant after initialization. The maximality error is formulated as
K
Li 0.8 Li
errmaxi (θ) = ( ) − , (9)
i=1
N N
where N is the initial number of the latent region set Λ. The first term encourages
the clusters of θ to merge, and the second term makes each cluster of θ to expand.
3.2 Likelihood p(I|θ)
Photometric Consistency of MCSs. The likelihood encodes the photometric

consistency of θ using the observation of the given image pair. Let us define the
dissimilarity of two regions by
dRGB(R1 , R2 )
dissim(R1 , R2 ) = 1−NCC(R1 , R2 ) + , (10)
100
where NCC is the normalized cross-correlation between the gray patterns, while
dRGB is the average pixel-wise Euclidean distance in RGB color-space after in-
dependent normalization of the 3 colorbands for photometric invariance [6]. R1
and R2 are normalized to unit circles with the same orientation before compu-
tation. Since a cluster of matches should have low dissimilarity in each match,
the overall photometric error of a cluster is defined as follows.

Li
errphoto (Γi ) = dissim(Rj , Rj )2 . (11)
j=1
Visual patterns in each MCS are assumed to be mutually independent in our

model. Hence, the likelihood is defined as follows.

K
p(I|θ) ∝ exp −λphoto errphoto (Γi ) . (12)
i=1
3.3 Integrated Posterior p(θ|I)
From (8), (9), and (12), MCSs in a given image pair I can be obtained by
maximizing the following posterior probability:

K
K
p(θ|I) ∝ exp −λgeo errgeo (Γi ) − λmaxi errmaxi (θ) − λphoto errphoto (Γi ) .
i=1 i=1
(13)
This posterior probability reflects how well the solution generates the set of
MCSs from the given image pair.
4 Data-Driven Monte Carlo Image Exploration
The posterior probability p(θ|I) in (13) has a high-dimensional and complicated

landscape with a large number of local maxima. Moreover, maximizing the pos-
terior is a trans-dimensional problem because neither the number of MCSs nor
the number of matches in each MCS are known. To pursue the global optimum
of this complex trans-dimensional posterior p(θ|I), we propose a new image ex-
ploration algorithm based on the reversible jump MCMC [10] with data-driven
techniques [9].
The basic idea of MCMC is to design a Markov chain to sample from a
probability distribution p(θ|I). At each sampling step, we propose a candidate
state θ from a proposal distribution q(θ |θ). Through the Metropolis-Hastings
rule, the candidate state is accepted with the following acceptance probability.

q(θ|θ )p(θ |I)
α = min 1, . (14)
q(θ |θ)p(θ|I)
Theoretically, it is proven that the Markov chain constructed in this manner

has its stationary distribution as p(I|θ) irrespective of the choice of the pro-
posal q(θ |θ) and the initial state [10]. Nevertheless, in practice, the choice of
the proposal significantly affects the efficiency of MCMC. Recently in computer
vision area, data-driven MCMC [9] has been proposed and proven to improve
the efficiency by incorporating domain knowledge in proposing new states of the
Markov chain. In our algorithm, we adopt the data-driven techniques to guide
our Markov chain using the current observation obtained by local region matches
in the image pair. Our Markov chain kernel consists of two pairs of reversible
jump dynamics which perform expansion/contraction and merge/split, respec-
tively. At each sampling step, a move m ∈ {expand, contract, split, merge} is
selected with the constant probability q(m).
4.1 Expansion/Contraction Moves

Expansion is to increase the size of an existing cluster by picking a region out of
the latent region set Λ and propagating it with a support region in the cluster.
Conversely, contraction functions to decrease the size by taking a region out
of the members in the cluster and sending it back to Λ. Suppose, at a certain
sampling step, that a cluster Γi is expanded to Γi , or conversely that Γi is
contracted to Γi , then this process can be expressed as the following form without
loss of generality:

θ = (K, {Γi , ...}) ↔ K, {Γi , ...} = θ , where Γi ∪ (Rk , Tk ) = Γi . (15)
The Pathway to Propose Expansion. An expansion move is proposed by the

following stochastic procedure with data-driven techniques. Firstly, a cluster √ is
chosen among the current K clusters with the probability q(Γi |expand) ∝ Li ,
which reflects a preference to larger clusters. Secondly, among the matches in the
cluster, a support for propagation is selected with probability q(Rj |Γi , expand) ∝
(a) Propagation and refinement (b) Mergence tree
Fig. 4. (a) At the top, a support match (red dotted) propagates one of the latent regions
(blue solid) by affine homography F. At the bottom, by adjusting the parameter of the
ellipse, the initially propagated region (blue dotted) is refined into the more accurate
region (green solid). (b) Each of the present clusters has its own mergence tree, which
stores hierarchical information of the preceding clusters of itself. It helps to propose a
simple and reversible merge/split moves at low cost.

dist(R ,R)
R∈Λ exp − 2σ2 j , where dist() denotes the Euclidean distance between
expand
the region centers. In this stochastic selection, the supports that have more la-
tent regions at nearer distance are favored. Finally, a latent region to prop-
agate

by the support
is chosen with the probability q(Rk |Rj , Γi , expand) ∝
dist(Rk ,Rj )2
exp − 2
2σexpand
, which means a preference to closer ones.
Propagation Attempt and Refinement. The building block of expansion is

based on the propagation attempt and refinement in [6]. If an expansion move
is proposed, we perform a propagation attempt followed by the refinement. As
illustrated in Fig. 4(a), consider the case that a red dotted elliptical region R1
in the reference image is already matched to R1 in the other image. Each R1
and R1 has an affine transformation A and A , respectively, which transform the
regions onto the orientation normalized unit circles. Thus, we can get the affine
−1
homography F between R1 to R1 by F = (A ) A, satisfying F R1 = R1 . If a
latent region R2 is close enough to R1 and lie on the same physical surface, we
can approximate R2 in the other image by R2 = F R2 as shown in Fig. 4(a). In
that case, we state that the support match (R1 , R1 ) attempts to propagate the
latent region R2 . Next, by locally searching the parameter space of the current
affine homography F , the refiner adjusts it to find R2 with minimum dissimilarity
such that Fr = arg min dissim(R2 , F R2 ) as shown at the bottom of Fig. 4(a).
F
The Pathway to Propose Contraction. An previously expanded region is

proposed to contract by the following stochastic procedure with data-driven tech-
is chosen among the current K clusters with the prob-
niques. Firstly, a cluster√
ability q(Γi |contract) ∝ Li . Then, among matches supporting no other region

cluster,2 one match 2is selected with the probability q(Rk |Γi , contract) ∝
in the
errgeo (Rk ) +errphoto (Rk )
exp 2
2σcontract
, favoring the matches with higher error in geom-
etry and photometry.
4.2 Merge/Split Moves

This pair of moves is for merging two different clusters into a new one or splitting
one into two clusters. Suppose, at a certain sampling step, that a cluster Γi is
split into two cluster Γl and Γm , or conversely that that Γl and Γm is merged
into a cluster Γi , then the processes can be represented as the following form
without loss of generality.
θ = (K, {Γi , ...}) ↔ (K + 1, {Γl , Γm , ...}) = θ , where Γi = Γl ∪ Γm . (16)
The Pathway to Propose Merge. We propose the merge of two clusters

along the following stochastic procedure. Firstly, among the current K clusters,
one cluster is chosen with the probability q(Γl |merge) ∝ 1/K.
Then, another 2
cluster is selected with the probability q(Γm |Γl , merge) ∝ exp − dist(Γ
2
m ,Γl )
2σmerge ,
where dist() denotes the Euclidean distance between the cluster centroids. This
represents the sampling from a Gaussian Parzen window centered at the centroid
of the first cluster Γl .
Mergence Trees. Unlike merge, its reverse move, split, is complicated to propose
since it involves classifying all the member regions of a cluster into two poten-
tial clusters. Moreover, to satisfy the detailed balance condition of MCMC [10],
all the move sequences in dynamics should be reversible, which means that if a
merge move can be proposed, then the exact reverse split move should be pos-
sible. To design efficient and reversible merge/split, we construct mergence trees
for merge/split over all the process. Each cluster has its own mergence tree which
stores the information of all the constituent clusters of itself in the tree structure
(Fig. 4(b)). Utilizing the mergence trees, we can propose a simple but potential
split move at low cost, that is the move to the state just before the latest merge
move. Note that we always begin from the clusters with a single initial match, and
the clusters are grown up gradually by the accepted moves among four types of
proposals. Thus, one of the best split moves is simply tracing back to the past.
The Pathway to Propose Split. A previously merged cluster can be proposed
to split into two as follows using the mergence tree. Firstly, a cluster among the
current K clusters is chosen with the probability q(Γi |split) ∝ 1/K. Then, the
cluster is proposed to split into two clusters corresponding to child nodes in its
mergence tree, with the probability q(Γl , Γm |Γi , split, mergence trees) = 1.
4.3 Overall Markov Chains Dynamics and Criterion of Reliable

MCSs
Our DDMC image exploration algorithm simulates a Markov chain consisting of
two pairs of sub-kernels, which continuously reconfigures θ according to p(θ|I).
At each sampling step, the algorithm chooses a move m with probability q(m),
then the sub-kernel of the move m is performed. The proposed move along its
pathway is accepted with the acceptance probability (14). If the move is accepted,
the current state jumps from θ to θ . Otherwise, the current state is retained.
In the early stage of sampling, we perform only expansion/contration moves
without merge/split moves because the unexpanded clusters in the early stage
are prone to unreliable merge/split moves. After enough iterations, merge/split
moves incorporate with expansion/contraction moves, helping the Markov chains
to have better chances of proposing reliable expansion/contraction moves and
estimating correct MCSs.
To evaluate the reliability of MCSs in the best sample θ∗ , we define the ex-
pansion ratio of an MCS as the expanded area of the MCS divided by the entire
image area. Since a reliable MCS is likely to expand enough, we determine the
reliable MCSs as those expanded more than the threshold ratio in both of two
images. This criterion of our method eliminates the trivial or false correspon-
dences effectively.
4.4 Implementation Details

For initialization, we used Harris-Affine [3] and MSER [2] detectors with SIFT
as a feature descriptor. After nearest neighbor matching, potential outliers are
filtered out through the ratio test with threshold 0.8 [1]. In our experiments,
the grid for the latent region set is composed of regions of radius h/25, spaced
h/25, where h denotes the height of the reference image. The radius trades
correspondence density and segmentation quality for computational cost. It
can be selected based on the specific purpose. The parameters in the poste-
rior model were fixed as follows: δori = π/4, λgeo = 3, λphoto = 20, λmaxi = 6.
In the sampling stage, we set the probability of selecting each sub-kernel as
q(expand) = q(contr) = 0.4, q(split) = q(merge) = 0.1, and the parameters of
sub-kernels are set to σexpand = l/100, σcontract = 0.5, σmerge = l/10, where l
means the diagonal length of the reference image. The results were obtained af-
ter 7000 iteration runs. Only the expansion/constration moves are performed in
the first 1000 samplings. In most of our tests, the MAP θ∗ was generated within
about 5000 samplings. The expansion threshold ratio for reliable MCSs in all
our experiments is set to 2% of each image.
5 Experiments
We have conducted two experiments: (i) unsupervised recognition and segmen-
tation of multiple common objects and (ii) image retrieval for place recognition.
5.1 Unsupervised Recognition and Segmentation of Multiple

Common Objects
Since there is no available public dataset for this problem yet, we built a new
challenging dataset including multiple common objects with mutual occlusion
Fig. 5. Co-recognition results on Minnie’s, Jigsaws, Toys, Books, and Bulletins. We

built the datasets for evaluation of co-recognition except for Bulletins, which is bor-
rowed from [11] for comparison.
Table 1. Performance evaluation of segmentation
Dataset Mickey’s Minnie’s Jigsaws Toys Books Bulletins Average

Hit Ratio 80.7% 83.2% 80.0% 83.5% 94.6% 91.2% 85.5%
Bk Ratio 20.6% 37.4% 22.8% 25.2% 11.8% 16.8% 22.4%
and complex clutters. The ground truth segmentation of the common objects has
been achieved manually5 . Figure 5 and 1 show some of co-recognition results on
them. Each color of the boundary represents identity of each MCS. The inferred
MCSs, their segmentations (the 2nd column), and their dense correspondences
(the 3rd column) are of good quality in all pairs of the dataset. On the average,
the correct match ratio started from less than 5% in naive NN matches, growing
to 42.2% after initial matching step, and finally reached to 92.8% in final reliable
MCSs. The number of correct matches increased to 651%.
We evaluated segmentation accuracy by hit ratio hr and background ratio
br .6 The results are summarized in Table 1. It also shows high accuracy in
segmentation. For example, the dataset Bulletins is borrowed from [11], and our
result of hr = 0.91, br = 0.17 is much better than their result of hr = 0.76, br =
0.29 in [11]. Moreover, note that our method provides object-level identities and
5
The dataset with ground truth is available at http://cv.snu.ac.kr/∼corecognition.
6
hr = |GroundTruth∩Result|
|GroundTruth|
, br = |Result|−|Result∩GroundTruth|
|Result|
.
Fig. 6. Co-recognition on all combination pairs of 5 test images from the ETHZ Toys
dataset. Both the detection rate and the precision are 93%.
dense correspondences, which are not provided by the method of [11]. Most of
the over-expanded regions increasing the background ratio result from mutually
similar background regions.
To demonstrate the unsupervised detection performance of co-recognition in
view changes or deformation, we tested on all combination pairs of 5 complex
images from the ETHZ toys dataset7 . None of the model images in the dataset
are included in this experiment. As shown in Fig. 6, although this task is very
challenging even for human eyes, our method detected 13 true ones and 1 false
one among 14 common object correspondences in the combination pairs. The de-
tection rate and the precision are all 93%. Note that our method can recognize
the separate regions as one MCS if mutual geometry of the regions is consistent
according to the reinforced sidedness constraint (6). Thus, it can deal with com-
plex partial occlusion which separates the objects into fragments. This allows
us to estimate the correct number of identical entities of separate regions as in
result of Fig. 1 and Fig. 6.
7
http://www.robots.ox.ac.uk/∼ferrari/datasets.html
(b) Accuracy rate for Test4 and

(a) Co-recognition on ICCV2005 datasets Final5.
Fig. 7. (a) Co-recognition deals with object-level correspondence, which is higher

than segment-level correspondence. (b) Comparison with co-saliency matching [5] on
ICCV2005 datasets.
5.2 Image Retrieval for Place Recognition

For image retrieval, we have conducted the experiment as in [5] on ICCV 2005
Computer Vision Contest datasets8 . Each of two datasets (Test4 and Final5 ) has
been split into exemplar and query set. Test4 has 19 query images and 9 exemplar
images, while Final5 has 22 query images and 16 exemplar images. Each of query
images is compared with all exemplar images, and all the matched image pairs
are ranked according to the total area of reliable MCSs. For every query image
having at least k similar examplars, the accuracy rate is evaluated with how
many of them are included in top k ranks. The result in Fig. 7(b) reveals that our
co-recognition outperforms co-saliency matching [5] largely in this experiment.
The reason can be explained by comparing our result of the top in Fig. 7(a)
with the result of the same pair in [5]. Co-recognition deals with object-level
correspondences, which are higher than segment-level correspondences as [5],
our method generates larger, denser, and more accurate correspondence without
segmentation cue.
6 Conclusion
We have presented a novel notion of co-recognition and the algorithm, which
recognizes and segments all the common salient region pairs with their maximal
sizes in an arbitrary image pair. The problem is formulated as a Bayesian MAP
problem and the solution is obtained by our stochastic image exploration al-
gorithm using DDMCMC paradigm. Experiments on challenging datasets show
promising results on the problem, some of which even humans cannot achieve
easily. The proposed co-recognition has various applications for high-level image
matching such as object-driven image retrieval.
8
http://research.microsoft.com/iccv2005/Contest/
Acknowledgements
This research was supported in part by the Defense Acquisition Program Admin-
istration and Agency for Defense Development, Korea, through the Image Infor-
mation Research Center under the contract UD070007AD, and in part by the
MKE (Ministry of Knowledge Economy), Korea under the ITRC (Information
Technolgy Research Center) Support program supervised by the IITA (Institute
of Information Technology Advancement) (IITA-2008-C1090-0801-0018).
References
1. Lowe, D.G.: Object recognition from local scale-invariant features. In: ICCV, pp.
1150–1157 (1999)
2. Matas, J., Chum, O., Urban, M., Pajdla, T.: Robust wide baseline stereo from
maximally stable extremal regions. In: BMVC (2002)
3. Mikolajczyk, K., Schmid, C.: An affine invariant interest point detector. In: Heyden,
A., Sparr, G., Nielsen, M., Johansen, P. (eds.) ECCV 2002. LNCS, vol. 2350, pp.
4. Vedaldi, A., Soatto, S.: Local features, all grown up. In: CVPR, pp. 1753–1760
(2006)
5. Toshev, A., Shi, J., Daniilidis, K.: Image matching via saliency region correspon-
dences. In: CVPR (2007)
6. Ferrari, V., Tuytelaars, T., Gool, L.: Simultaneous object recognition and segmen-
tation from single or multiple model views. IJCV 67(2), 159–188 (2006)
7. Yang, G., Stewart, C.V., Michal Sofka, C.L.T.: Registration of challenging image
pairs:initialization, estimation, and decision. PAMI 29(11), 1973–1989 (2007)
8. Rother, C., Minka, T.P., Blake, A., Kolmogorov, V.: Cosegmentation of image pairs
by histogram matching - incorporating a global constraint into MRFs. In: CVPR,
pp. 993–1000 (2006)
9. Tu, Z., Chen, X., Yuille, A.L., Zhu, S.C.: Image parsing: unifying segmentation,
detection, and recognition. In: ICCV, vol. 1, pp. 18–25 (2003)
10. Green, P.: Reversible jump markov chain monte carlo computation and bayesian
model determination. Biometrica 82, 711–732 (1995)
11. Yuan, J., Wu, Y.: Spatial random partition for common visual pattern discovery.
In: ICCV, pp. 1–8 (2007)
12. Simon, I., Seitz, S.M.: A probabilistic model for object recognition, segmentation,
and non-rigid correspondence. In: CVPR (2007)
13. Cho, M., Lee, K.M.: Partially occluded object-specific segmentation in view-based
recognition. In: CVPR (2007)
Movie/Script: Alignment and Parsing
of Video and Text Transcription
Timothee Cour, Chris Jordan, Eleni Miltsakaki, and Ben Taskar
University of Pennsylvania, Philadelphia, PA 19104, USA

{timothee,wjc,elenimi,taskar}@seas.upenn.edu
Abstract. Movies and TV are a rich source of diverse and complex video of peo-
ple, objects, actions and locales “in the wild”. Harvesting automatically labeled
sequences of actions from video would enable creation of large-scale and highly-
varied datasets. To enable such collection, we focus on the task of recovering
scene structure in movies and TV series for object tracking and action retrieval.
We present a weakly supervised algorithm that uses the screenplay and closed
captions to parse a movie into a hierarchy of shots and scenes. Scene boundaries
in the movie are aligned with screenplay scene labels and shots are reordered
into a sequence of long continuous tracks or threads which allow for more ac-
curate tracking of people, actions and objects. Scene segmentation, alignment,
and shot threading are formulated as inference in a unified generative model and
a novel hierarchical dynamic programming algorithm that can handle alignment
and jump-limited reorderings in linear time is presented. We present quantitative
and qualitative results on movie alignment and parsing, and use the recovered
structure to improve character naming and retrieval of common actions in several
episodes of popular TV series.
1 Introduction
Hand-labeling images of people and objects is a laborious task that is difficult to scale
up. Several recent papers [1,2] have successfully collected very large-scale, diverse
datasets of faces “in the wild” using weakly supervised techniques. These datasets
contain a wide variation in subject, pose, lighting, expression, and occlusions which
is not matched by any previous hand-built dataset. Labeling and segmenting actions is
perhaps an even more painstaking endeavor, where curated datasets are more limited.
Automatically extracting large collections of actions is of paramount importance. In
this paper, we argue that using movies and TV shows precisely aligned with easily ob-
tainable screenplays can pave a way to building such large-scale collections. Figure 1
illustrates this goal, showing the top 6 retrieved video snippets for 2 actions (walk,
turn) in TV series LOST using our system. The screenplay is parsed into a temporally
aligned sequence of action frames (subject verb object), and matched to detected and
named characters in the video sequence. Simultaneous work[3] explores similar goals
in a more supervised fashion. In order to enable accurately localized action retrieval,
we propose a much deeper analysis of the structure and syntax of both movies and
transcriptions.

Movie/Script: Alignment and Parsing of Video and Text Transcription 159
Fig. 1. Action retrieval using alignment between video and parsed screenplay. For each ac-
tion verb (top: walk, bottom: turn), we display the top 6 retrieved video snippets in TV se-
ries LOST using our system. The screenplay and closed captions are parsed into a tempo-
rally aligned sequence of verb frames (subject-verb-object), and then matched to detected and
named characters in the video sequence. The third retrieval, second row (“Jack turns”) is
counted as an error, since the face shows Boone instead of Jack. Additional results appear under
www.seas.upenn.edu/∼{}timothee.
Movies, TV series, news clips, and nowadays plentiful amateur videos, are designed
to effectively communicate events and stories. A visual narrative is conveyed from mul-
tiple camera angles that are carefully composed and interleaved to create seamless ac-
tion. Strong coherence cues and continuity editing rules are (typically) used to orient
the viewer, guide attention and help follow the action and geometry of the scene. Video
shots, much like words in sentences and paragraphs, must fit together to minimize per-
ceptual discontinuity across cuts and produce a meaningful scene. We attempt to un-
cover elements of the inherent structure of scenes and shots in video narratives. This
uncovered structure can be used to analyze the content of the video for tracking objects
across cuts, action retrieval, as well as enriching browsing and editing interfaces.
We present a framework for automatic parsing of a movie or video into a hierarchy
of shots and scenes and recovery of the shot interconnection structure. Our algorithm
makes use of both the input image sequence, closed captions and the screenplay of
the movie. We assume a hierarchical organization of movies into shots, threads and
scenes, where each scene is composed of a set of interlaced threads of shots with smooth
transitions of camera viewpoint inside each thread. To model the scene structure, we
propose a unified generative model for joint scene segmentation and shot threading.
We show that inference in the model to recover latent structure amounts to finding
a Hamiltonian path in the sequence of shots that maximizes the “head to tail” shot
similarity along the path, given the scene boundaries. Finding the maximum weight
Hamiltonian path (reducible to the Traveling Salesman Problem or TSP) is intractable
in general, but in our case, limited memory constraints on the paths make it tractable.
In fact we show how to jointly optimize scene boundaries and shot threading in linear
time in the number of shots using a novel hierarchical dynamic program.
We introduce textual features to inform the model with scene segmentation, via tem-
poral alignment with screenplay and closed captions, see figure 2. Such text data has
been used for character naming [4,5] and is widely available, which makes our approach
applicable to a large number of movies and TV series. In order to retrieve temporally-
aligned actions, we delve deeper into resolving textual ambiguities with pronoun reso-
lution (determining whom or what ‘he’, ‘she’, ‘it’, etc. refer to in the screenplay) and
extraction of verb frames. By detecting and naming characters, and resolving pronouns,
we show promising results for more accurate action retrieval for several common verbs.
We present quantitative and qualitative results for scene segmentation/alignment, shot
160 T. Cour et al.
segmentation/threading, tracking and character naming across shots and action retrieval
in numerous episodes of popular TV series, and illustrate that shot reordering provides
much improved character naming.
The main contributions of the paper are: 1) novel probabilistic model and inference
procedure for shot threading and scene alignment driven by text, 2) extraction of verb
frames and pronoun resolution from screenplay, and 3) retrieval of the corresponding
actions informed by scene sctructure and character naming.
The paper is organized as follows. Section 2 proposes a hierarchical organization of
movies into shots, threads and scenes. Sections 3 and 4 introduce a generative model
for joint scene segmentation and shot threading, and a hierarchical dynamic program to
solve it as a restricted TSP variant. Section 5 addresses the textual features used in our
model. We report results in section 6 and conclude in section 7.
(a) (b)
Fig. 2. (a) Alignment between video, screenplay and closed captions; (b) Deconstruction pipeline
2 Movie Elements: Shots, Threads, Scenes

Movies and TV series are organized in distinctive hierarchical and continuity structures
consisting of elements such as scenes, threads and shots. Detecting and recovering these
elements is needed for uninterrupted tracking of objects and people in a scene across
multiple cameras, recovering geometric relationships of objects in a scene, intelligent
video browsing, search and summarization.
Shot boundaries. The aim of shot segmentation is to segment the input frames into a
sequence of shots (single unbroken video recordings) by detecting camera viewpoint
discontinuities. A popular technique is to compute a set of localized color histograms
for each image and use a histogram distance function to detect boundaries [6,7].
Shot threads. Scenes are often modeled as a sequence of shots represented as letters:
ABABAB represents a typical dialogue scene alternating between two camera points
of view A and B. More complex patterns are usually observed and in practice, the clus-
tering of the shots into letters (camera angles/poses) is not always a very well defined
problem, as smooth transitions between shots occur. Nevertheless we assume in our
case that each shot in a scene is either a novel camera viewpoint or is generated from
(similar to) a previous shot in the scene. This makes weaker assumptions about the
scene construction and doesn’t require reasoning about the number of clusters. In the
example above, the first A and B are novel viewpoints, and each subsequent A and B is
generated by the previous A or B. Figure 5 shows a more complex structure.
Scene boundaries. A scene consists of a set of consecutive semantically related shots

(coherence in action, location and group of actors is typical). The process of segmenting
a video sequence into scenes has received some attention in the video analysis litera-
ture [7]. An MCMC based clustering framework is used in [8]. Hierarchical cluster-
ing on a shot connectivity graph is proposed in [9]. In [10], the authors detect scene
boundaries as local minima of a backward shot coherence measure. As opposed to
shot boundaries, which correspond to strong visual discontinuity in consecutive frames,
scene boundaries are not detectable from purely local cues: the entire sequence of pre-
ceding and following shots must be considered. For example, ABCBABDEFEABD
shot sequence is one scene, while ABCBAB DEFEDEF can be two.
3 A (Simple) Generative Model of Movies
To capture the hierarchical and continuity structure, we propose a simple generative

model, where scenes are constructed independently of other scenes, while shots within
a scene are produced via an interleaved Markov (first order) structure.
We begin with some notation to define our model, assuming the video sequence has
already be segmented into shots:
– si : ith shot (interval of frames), with i ∈ [1, n]
– bj : j th scene boundary (index of its last shot), with j ≤ m; 1 ≤ b1 < ... < bm = n
– pj [i]: parent generating shot i in scene j (could be NULL), with j ≤ m, i ≤ n.
We assume the shots in a video sequence are generated as follows: first generate the
sequence of scene boundaries (bj ), then generate for each scene j a dependency struc-
ture pj defining a Markov chain on shots, and finally generate each shot i given its
parent pj [i]. The model is conditioned upon m and n, assumed to be known in advance.
This can be represented using the generative model in figure 3. For the scene bound-
ary model P (b), we investigate both a uniform model and an improved model, where
scene boundaries are informed by the screenplay (see section 5). The shot threading
model P (p|b) is uniformly distributed over valid Markov chains (shot orderings) on
Fig. 3. Graphical model for joint scene segmentation and shot reordering, see text for details
162 T. Cour et al.
each scene. The shot appearance model P (si |spj [i] ) is treated next (we set it to uni-
form for the root of scene j where pj [i] = NULL). This model encourages (1) smooth
shot transitions within a scene and (2) scene breaks between shots with low similarity,
since the model doesn’t penalize transitions across scenes.
Shot appearance model (P (si |si )). In order to obtain smooth transitions and al-
low tracking of objects throughout reordered shots, we require that P (si |si ) depends
on the similarity between the last frame of shot si (I = slast i ) and the first frame of
shot si (I = sfirst
i ). Treating each shot as a word in a finiteset, we parameterize the
shot similarity term as P (si |si ) = exp(−dshot (si , si ))/ i exp(−dshot (si , si ))
where dshot (si , si ) = dframe (I, I ) is the chi-squared distance in color histogram be-
tween frames I, I . Note, dshot (si , si ) is not symmetric, even though dframe (I, I ) is.
4 Inference in the Model

In this section we attempt to solve the MAP problem in figure 3. Let us first consider
the simplified case without scene transitions (when m = 1). In this case, maximizing
the log becomes:

max Wi,p[i] = max Wπt−1 ,πt (1)
p:Markov Chain π∈P[1,n]
i t
where Wii = log P (si |si ) and π ∈ P[1,n] denotes a permutation of [1, n] defined
recursively from the parent variable p as follows: p[πt ] = πt−1 , with π1 indicating
the root. This amounts to finding a maximum weight Hamiltonian Path or Traveling
Salesman Problem (TSP), with πt indicating which shot is visited at time t on a virtual
tour. TSPs are intractable in general, so we make one additional assumption restricting
the set of feasible permutations.
4.1 Memory-Limited TSPs

Given an integer k > 0 (memory width), and an initial ordering of shots (or cities by
analogy to TSP) 1, ..., n, we introduce the following limited memory constraint on our
hamiltonian path π = (πt ):
P[1,n]
k
= {π ∈ P[1,n] : ∀(i, i )i ≥ i + k ⇒ πi > πi } (2)
This is illustrated in figure 4 for k = 2 (k = 1 means π is the identity, and k = n
is fully unconstrained). There are two important consequences: (1) the MAP becomes
tractable (linear complexity in n), and (2) the problem becomes sparse, i.e., we can
restrict W.L.O.G. W to be sparse (banded):
πt ∈ [t − (k − 1), t + (k − 1)] (3)
Wii = −∞ except for i − (2k − 3) ≤ i ≤ i + 2k − 1 (4)
The first line comes from the pigeonhole principle, and the second one uses the first line:
−(2k − 3) ≤ πt+1 − πt ≤ 2k − 1. Note, this constraint is natural in a video sequence,
as video editing takes into account the limited memory span of humans consisting of a
few consecutive shots.
Fig. 4. Top: a feasible solution for the restricted TSP with k = 2. Bottom: an infeasible solution,
violating the precedence constraint (shaded cities). Middle: the constraint limits the range of the
permutation: πt ∈ [t − (k − 1), t + (k − 1)]. Right: the constraint implies a banded structure on
the similarity matrix W = (Wii ): i − (2k − 3) ≤ i ≤ i + 2k − 1.
4.2 Dynamic Programming Solution without Scene Breaks (P (p, s))

The solution to the simplified problem without scene breaks (1) under constraint (2)
has been addressed in [11] (it dealt with a hamiltonian cycle with π1 (1) = 1, but this
is easily adaptable to our case). We summarize the main points below. Let Ct (S, i ) be
the optimal cost of the paths π ∈ P[1,n]
k
satisfying πt = i and {π1 , ..., πt−1 } = S (set
of cities visited before time t). The dynamic programming solution uses the relation:
Ct (S, i ) = min Ct−1 (S − {i}, i) + Wii (5)

i∈S
Because of the precedence constraint, the pair (S, i ) can take at most (k + 1)2k−2 pos-
sible values at any given time t (instead of n−1t−1 n without the constraint). The idea
is to construct a directed weighted graph Gkn with n layers of nodes, one layer per
position in the path, with paths in the graph joining layer 1 to layer n corresponding
to feasible hamiltonian paths, and shortest paths joining layer 1 to n corresponding to
optimal hamiltonian paths. Since there are at most k incoming edges per node (corre-
sponding to valid transitions πt−1 → πt ), the total complexity of the dynamic program
is O(k(k + 1)2k−2 · n), exponential in k (fixed) but linear in n, see [11] for details.
4.3 Dynamic Programming Solution with Scene Breaks (P (b, p, s))

The general problem can be rewritten as:

max max Wπt−1 ,πt (6)
b k
π∈P(b
j j−1 ,bj ] t
Naive solution. One can solve (6) as follows: for each interval I ⊂ [1, n], pre-compute
the optimal path πI∗ ∈ PIk using 4, and then use a straightforward dynamic program-
ming algorithm to compute the optimal concatenation of m such paths to form the
optimal solution. Letting f (k) = k(k + 1)2k−2 , the complexity of this algorithm is
164 T. Cour et al.

O( 1≤i≤i ≤n f (k) · (i − i + 1)) = O(f (k)n(n + 1)(n + 2)/6) for the precomputation
and O(mn(n + 1)/2) for the dynamic program, which totals to O(f (k)n3 /6). The next
paragraph introduces our joint dynamic programming over scene segmentation and shot
threading, which reduces computational complexity by a factor n (number of shots).
Joint dynamic program over scene breaks and shot threading. We exploit the pres-
ence of overlapping subproblems. We construct a single tour π, walking over the joint
space of shots and scene labels. Our approach is based on the (categorical) product
graph Gkn × Cm where Gkn is the graph from 4.2 and Cm is the chain graph of order m.
A node (u, j) ∈ Gkn × Cm represents the node u ∈ Gkn in the j th scene. Given two
connected nodes u = (S, i, t) and u = (S , i , t + 1) in Gkn , there are two types of
connections in the product graph. The first connections correspond to shots i, i both
being in the j th scene:
(u, j) → (u , j), with weight Wii (7)
The second connections correspond to a scene transition:
(u, j) → (u , j + 1), with weight 0, (8)
and only happen when u = (S, i, t) satisfies max(i, max(S)) = t, to make sure the
tour decomposes into a tour of each scene (we can switch to the next scene when the
set of shots visited up to time t is exactly {1, ..., t}).
The solution to (6) similarly uses a dynamic program to find the shortest path in Gkn ×
Cm (and backtracking to recover the arg max). Since there are m times as many nodes
in the graph as in Gkn and at most twice as many incoming connections per node (nodes
from the previous scene or from the same scene), the total complexity is: O(2k(k +
1)2k−2 mn) = O(2f (k)mn).
Comparison. We manually labeled shot and scene breaks for a number of movies
and TV series and found that a typical scene contains on average about 11 shots,
i.e.m ≈ n/11. So the reduction in complexity between the naive algorithm and our
3
joint dynamic program is: O( f2f(k)n /6
(k)mn ) = O(n /(12m)) ≈ n, which is a huge gain,
2
especially given typical values of n = 600. The resulting complexity is linear in n and
m and in practice takes about 1 minute as opposed to 11 hours for an entire episode,
given pre-computed shot similarity.
5 Scene Segmentation Via Coarse Image to Text Alignment

(P (b))
We now assume we have some text data corresponding to the movie sequence, and
we focus on simultaneously segmenting/threading the video into scenes and aligning
the text with the video. The extra text media removes a lot of ambiguity for the scene
segmentation and, combined with our model, leads to improved scene segmentation
results as we shall see in section (6).
Fig. 5. Shot reordering to recover continuity in 3 scenes of LOST
5.1 Text Data: Screenplay and Closed Captions

We use two sources of text for our segmentation-alignment problem: the screenplay,
which narrates the actions and provides a transcript of the dialogues, and the closed
captions, which provide time-stamped dialogues, as in figure 2(a). Both sources are
essential since the screenplay reveals speaker identity, dialogues and scene transitions
but no time-stamps, and closed captions reveal dialogues with time-stamps but nothing
else. The screenplay and the closed captions are readily available for a majority of
movies and TV series produced in the US. A similar approach was used in [5] to align
faces with character names, with 2 differences: 1) they used the screenplay to reveal the
speaker identity as opposed to scene transitions, and 2) subtitles were used instead of
closed captions. Subtitles are encoded as bitmaps, thus require additional steps of OCR
and spell-checking to convert them to text[5], whereas closed captions are encoded as
ASCII text in DVDs, making our approach simpler and more reliable, requiring a simple
modification of mplayer (http://www.mplayerhq.hu/).
5.2 Screenplay/Closed Captions Alignment

The alignment between the screenplay and the closed captions is non-trivial since the
closed captions only contain the dialogues (without speaker) mentioned in the screen-
play, often with wide discrepancies between both versions. We extend the dynamic
time warping[12] approach in a straightforward way to time-stamp each element of the
screenplay (as opposed to just the dialogues as in [5]). The screenplay is first parsed into
a sequence of elements (either NARRATION, DIALOGUE, or SCENE-TRANSITION)
using a simple grammar, and the dynamic programming alignment of the words in the
screenplay and the closed captions provides a time interval [T start (i), T end (i)] for each
DIALOGUE element Ei . A NARRATION or SCENE-TRANSITION element Ej en-
closed between two DIALOGUE elements Ei1 , Ei2 is assigned the following conser-
vative time interval: [T start (i1 ), T end (i2 )].
5.3 Scene Segmentation Via Alignment

We determine the scene boundary term P (b) from section 3 by aligning each SCENE-
TRANSITION element mentioned in the screenplay to a scene start. P (b) is uniform
among the set of b satisfying the temporal alignment constraints:
166 T. Cour et al.
total verbs 25,000

distinct verbs 1,000
looks (most common) 2,000
turns 1,100
walks 800
takes 550
climbs 40
kisses 40
total dialogue lines 16,000
distinct speaker names 190
Jack (most common) 2,100
Fig. 6. Left: pronoun resolution and verb frames obtained from the parsed screenplay narrations.
Right: statistics collected from 24 parsed screenplays (1 season of LOST).
1 ≤ b1 < ... < bm = n (9)

t start
(j) ≤ bj−1 + 1 ≤ tend (j) (10)
where [tstart (j), [tend (j)] is the time interval of the j th SCENE-TRANSITION element,
converted into frame numbers, then to shot indexes.
Additional alignment constraints. Close inspection of a large number of screenplays

collected for movies and TV series revealed a fairly regular vocabulary used to describe
shots and scenes. One such example is FADE IN and FADE OUT corresponding to a
transition between a black shot (where each frame is totally black) and a normal shot,
and vice versa. Such black shots are easy to detect, leading to additional constraints in
the alignment problem, and a performance boost.
5.4 Pronoun Resolution and Verb Frames

Alignment of the screenplay to dialog in closed captions and scene boundaries in the
video helps to narrow down the scope of reference for other parts of the screenplay
that are interspersed – the narration or scene descriptions, which contain mentions of
actions and objects on the screen. In addition to temporal scope uncertainty for these de-
scriptions, there is also ambiguity with respect to the subject of the verb, since personal
pronouns (he, she) are commonly used. In fact, our analysis of common screenplays re-
veals there are more pronouns than occurences of character names in the narrations, and
so resolving those pronouns is an important task. We employed a simple, deterministic
scheme for pronoun resolution that uses a standard probabilistic context-free parser to
analyze sentences and determine verb frames (subject-verb-object) and then scans the
sentence for possible antecedents of each pronoun that agree in number and gender, see
figure 6. The details of the algorithm are given in supplemental materials. Here is an ex-
ample output of our implementation on a sentence extracted from screenplay narration
(pronoun resolution shown in parenthesis): On the side, Sun watches them. Jin reaches
out and touches Sun ’s chin, his (Jin’s) thumb brushes her (Sun’s) lips. She (Sun) looks
at him (Jin) and pulls away a little. He (Jin) puts his (Jin’s) hand down.
Output verb frames: (Sun - watches - something) (Jin - reaches out - ) (Jin - touches
- chin) (Sun - looks - at Jin) . (Sun - pulls away - ) (Jin - puts down - hand).
We report pronoun resolution accuracy on screenplay narrations of 3 different TV
series (about half a screenplay for each), see table 1.
Table 1. Pronoun resolution accuracy on screenplay narrations of 3 different TV series
TV series screenplay pronoun resolution accuracy # pronouns # sentences

LOST 75% 93 100
CSI 76 % 118 250
ALIAS 78% 178 250
6 Results
We experimented with our framework on a significant amount of data, composed of TV
series (19 episodes from one season of LOST, several episodes of CSI), one feature length
movie “The Fifth Element”, and one animation movie “Aladdin”, representing about 20
hours of video at DVD resolution. We report results on scene segmentation/alignment,
character naming and tracking, as well as retrieval of query action verbs.
Shot segmentation. We obtain 97% F-score (harmonic mean of precision and recall)
for shot segmentation, using standard color histogram based methods.
Scene segmentation and alignment. We hand labeled scene boundaries in one episode
of LOST and one episode of CSI based on manual alignment of the frames with the
screenplay. The accuracy for predicting the scene label of each shot was 97% for LOST
and 91% for CSI. The F-score for scene boundary detection was 86% for LOST and
75% for CSI, see figure 7. We used k = 9 for the memory width, a value similar to the
buffer size used in [10] for computing shot coherence. We also analyzed the effect on
performance of the memory width k, and report results with and without alignment to
screenplay in table 2. In comparison, we obtained an F-score of 43% for scene bound-
ary detection using a model based on backward shot coherence [10] uninformed by
screenplay, but optimized over buffer size and non-maximum suppression window size.
Scene content analysis. We manually labeled the scene layout in the same episodes
of LOST and CSI, providing for each shot in a scene its generating shot (including
Table 2. % F-score (first number) for scene boundary detection and % accuracy (second number)
for predicting scene label of shots (on 1 episode of LOST) as a function of the memory width k
used in the TSP, and the prior P (b). The case k = 1 corresponds to no reordering at all. Line 1:
P (b) informed by screenplay; line 2: P (b) uniform; line 3: total computation time.
P (b) k=1 k=2 k = 3 k = 9 k = 12

aligned 73/90 77/91 82/96 86/97 88/97
uniform 25/0 45/14 55/0 52/1 -/-
total time (s) < 0.1 < 0.1 0.1 5 68
168 T. Cour et al.
Fig. 7. Movie at a glance: scene segmentation-alignment and shot reordering for an episode of
LOST (only a portion shown for readability). Scene boundaries are in red, together with the set
of characters appearing in each scene, in blue.
Fig. 8. Character naming using screenplay alignment and shot threading. Top 3 rows: correctly
named faces; bottom row: incorrectly named faces. We detect face tracks in each shot and reorder
them according to the shot threading permutation. Some face tracks are assigned a name prior
based on the alignment between dialogues and mouth motion. We compute a joint assignment of
names to face tracks using an HMM on the reordered face tracks.
the special case when this is a new viewpoint). We obtain a precision/recall of 75%
for predicting the generating parent shot. See figure 5 for a sample of the results on 3
scenes. Note, to obtain longer tracks in figure 5, we recursively applied the memory
limited TSP until convergence (typically a few iterations).
Character identification on reordered shots. We illustrate a simple speaker iden-

tification based on screenplay alignment and shot threading, see figure 8. We use a
Viola-Jones[13] based face detector and tracking with normalized cross-correlation to
obtain face tracks in each shot. We build a Hiddden Markov Model (HMM) with states
Fig. 9. Top 10 retrieved video snippets for 15 query action verbs: close eyes, grab, kiss, kneel,
open, stand, cry, open door, phone, point, shout, sit, sleep, smile, take breath. Please zoom in to
see screenplay annotation (and its parsing into verb frames for the first 6 verbs).
corresponding to assignments of face tracks to character names. The face tracks are
ordered according to the shot threading permutation, and as a result there are much
fewer changes of character name along this ordering. Following [14], we detect on-
screen speakers as follows: 1) locate mouth for each face track using a mouth detector
based on Viola-Jones, 2) compute a mouth motion score based on the normalized cross
correlation between consecutive windows of the mouth track, averaged over temporal
segments corresponding to speech portions of the screenplay. Finally we label the face
tracks using Viterbi decoding for the Maximum a Posteriori (MAP) assignment (see
website for more details). We computed groundtruth face names for one episode of
LOST and compared our method against the following baseline that does not use shot
170 T. Cour et al.
reordering: each unlabeled face track (without a detected speaking character on screen)
is labeled using the closest labeled face track in feature space (position of face track and
color histogram). The accuracy over an episode of LOST is 76% for mainly dialogue
scenes and 66% for the entire episode, as evaluated against groundtruth. The baseline
model based using nearest neighbor performs at resp. 43% and 39%.
Retrieval of actions in videos. We consider a query-by-action verb retrieval task for
15 query verbs across 10 episodes of LOST, see figure 9. The screenplay is parsed into
verb frames (subject-verb-object) with pronoun resolution, as discussed earlier. Each
verb frame is assigned a temporal interval based on time-stamped intervening dialogues
and tightened with nearby shot/scene boundaries. Queries are further refined to match
the subject of the verb frame with a named character face. We report retrieval results
as follows: for each of the following action verbs, we measure the number of times
(out of 10) the retrieved video snippet correctly shows the actor on screen performing
the action (we penalize for wrong naming): close eyes (9/10), grab (9/10), kiss (8/10),
kneel (9/10), open (9/10), stand (9/10), cry (9/10), open door (10/10), phone (10/10),
point (10/10), shout (7/10), sit (10/10), sleep (8/10), smile (9/10), take breath (9/10).
The average is 90/100. Two additional queries are shown in figure 1 along with the de-
tected and identified characters. We created a large dataset of retrieved action sequences
combined with character naming for improved temporal and spatial localization, see
www.seas.upenn.edu/∼{}timothee for results and matlab code.
7 Conclusion
In this work we have addressed basic elements of movie structure: hierarchy of scenes
and shots and continuity of shot threads. We believe that this structure can be useful
for many intelligent movie manipulation tasks, such as semantic retrieval and indexing,
browsing by character or object, re-editing and many more. We plan to extend our work
to provide more fine-grained alignment of movies and screenplay, using coarse scene
geometry, gaze and pose estimation.
References
1. Huang, G., Jain, V., Learned-Miller, E.: Unsupervised joint alignment of complex images.
In: International Conference on Computer Vision, pp. 1–8 (2007)
2. Ramanan, D., Baker, S., Kakade, S.: Leveraging archival video for building face datasets. In:
International Conference on Computer Vision, pp. 1–8 (2007)
3. Laptev, I., Marszałek, M., Schmid, C., Rozenfeld, B.: Learning realistic human actions from
movies. In: IEEE Conference on Computer Vision and Pattern Recognition (2008),
http://lear.inrialpes.fr/pubs/2008/LMSR08
4. Sivic, J., Everingham, M., Zisserman, A.: Person spotting: video shot retrieval for face sets.
In: Leow, W.-K., Lew, M., Chua, T.-S., Ma, W.-Y., Chaisorn, L., Bakker, E.M. (eds.) CIVR
2005. LNCS, vol. 3568, Springer, Heidelberg (2005)
5. Everingham, M., Sivic, J., Zisserman, A.: Hello! my name is.. buffy – automatic naming of
characters in tv video. In: Proceedings of the British Machine Vision Conference (2006)
6. Lienhart, R.: Reliable transition detection in videos: A survey and practitioner’s guide. Int.
Journal of Image and Graphics (2001)
7. Ngo, C.-W., Pong, T.C., Zhang, H.J.: Recent advances in content-based video analysis. In-
ternational Journal of Image and Graphics 1, 445–468 (2001)
8. Zhai, Y., Shah, M.: Video scene segmentation using markov chain monte carlo. IEEE Trans-
actions on Multimedia 8, 686–697 (2006)
9. Yeung, M., Yeo, B.L., Liu, B.: Segmentation of video by clustering and graph analysis.
Comp. Vision Image Understanding (1998)
10. Kender, J., Yeo, B.: Video scene segmentation via continuous video coherence. In: IEEE
Conference on Computer Vision and Pattern Recognition (1998)
11. Balas, E., Simonetti, N.: Linear time dynamic programming algorithms for new classes of
restricted tsps: A computational study. INFORMS Journal on Computing 13, 56–75 (2001)
12. Myers, C.S., Rabiner, L.R.: A comparative study of several dynamic time-warping algorithms
for connected word recognition. The Bell System Technical Journal (1981)
13. Viola, P.A., Jones, M.J.: Robust real-time face detection. International Journal of Computer
Vision 57, 137–154 (2004)
14. Everingham, M.R., Sivic, J., Zisserman, A.: Hello! my name is buffy: Automatic naming of
characters in tv video. In: BMVC, vol. III, p. 899 (2006)
Using 3D Line Segments for Robust and Efficient
Change Detection from Multiple Noisy Images
Ibrahim Eden and David B. Cooper
Division of Engineering
Brown University
Providence, RI, USA
{ieden,cooper}@lems.brown.edu
Abstract. In this paper, we propose a new approach to change detection that is

based on the appearance or disappearance of 3D lines, which may be short, as
seen in a new image. These 3D lines are estimated automatically and quickly
from a set of previously-taken learning-images from arbitrary view points and
under arbitrary lighting conditions. 3D change detection traditionally involves
unsupervised estimation of scene geometry and the associated BRDF at each ob-
servable voxel in the scene, and the comparison of a new image with its predic-
tion. If a significant number of pixels differ in the two aligned images, a change
in the 3D scene is assumed to have occurred. The importance of our approach is
that by comparing images of lines rather than of gray levels, we avoid the compu-
tationally intensive, and some-times impossible, tasks of estimating 3D surfaces
and their associated BRDFs in the model-building stage. We estimate 3D lines
instead where the lines are due to 3D ridges or BRDF ridges which are compu-
tationally much less costly and are more reliably detected. Our method is widely
applicable as man-made structures consisting of 3D line segments are the main
focus of most applications. The contributions of this paper are: change detec-
tion based on appropriate interpretation of line appearance and disappearance in
a new image; unsupervised estimation of “short” 3D lines from multiple images
such that the required computation is manageable and the estimation accuracy is
high.
1 Introduction
The change detection problem consists of building an appearance model of a 3D scene

using n images, and then based on an n+1st image, determining whether a “significant”
change has taken place. A fundamental approach to this problem is to estimate a 3D
model for the scene and the associated BRDF; then based on the knowledge of the
n+ 1st image viewing position and scene illumination, a decision is made as to whether
there is a significant difference between the n + 1st image and its prediction by the n-
image based 3D geometry and BRDF (bidirectional reflectance distribution function)
estimates.
In its general form, all learning is done in the unsupervised mode, and the n-image
based learning is not done for a static 3D scene but rather for a functioning scene where
changes are often taking place. A complicating factor in the change detection problem

Using 3D Line Segments for Robust and Efficient Change Detection 173
Fig. 1. Our line segment based change detection result after training on a sequence of 5 images.
(A) A sample training image. (B) The test image. (C) Hand-marked ground truth for change
where the new object is shown in “red” and the disappeared object is shown in “blue”. (D) Result
of our method. Lines associated with the new object are shown in “red” and lines associated with
the disappeared object are shown in “blue”. Two major change regions are detected with only a
few false alarms due to specular highlights and object shadows. (This is a color image)
is that images can be taken at arbitrary time, under arbitrary lighting conditions and
from arbitrary view points. Furthermore, they are usually single images and not video.
For example, if they are taken from a flying aircraft, a 3D point in the scene is usually
seen in one image and not in the immediately preceding or succeeding images, and is
not seen again until the aircraft returns at some later time or until some other aircraft or
satellite or moving camera on the ground sees the point at some later time.
In this paper, we assume n images are taken of a scene, and we then look for a change
in the n + 1st image, and if one has occurred we try to explain its type (resulting from
the arrival or from the departure of a 3D object). The learning is done in an unsupervised
mode. We do not restrict ourselves to the case of buildings where the 3D lines are long,
easy to detect, easy to estimate and are modest in number. Rather, we are interested
in the case of many short lines where the lines can be portions of long curves or can
be short straight line segments associated with complicated 3D objects, e.g., vehicles,
scenes of damaged urban-scapes, natural structure, people, etc...
Why do we restrict this study to straight lines? We could deal with curves, but since
curves can be decomposed into straight lines, and since straight lines – especially short
174 I. Eden and D.B. Cooper
line segments - appear extensively in 3D scenes and in images, we decided to start with
those. The important thing is that estimating 3D structure and the associated BRDF
can often be done in theory, but this is usually difficult to do computationally. On the
other hand, estimating 3D line segments is much more tractable and can be considered
as a system in its own right or as contributing to applications that require efficient 3D
structure estimation.
Our paper consists of the following. Given n images, we estimate all 3D lines that
appear in three or more images. Our approach to 3D line estimation emphasizes com-
putational speed and accuracy. For very short lines, accuracy is greatly improved by
making use of incidence relations among the lines. For change detection we look for
the appearance or disappearance of one or more line segments in the n + 1st image.
This procedure depends on the camera position of the new image and the set of re-
constructed 3D line segments in the learning period, and therefore an interpretation of
whether a line is not seen because of self occlusion within the scene or because of a 3D
change. Usually, but not always, if an existing 3D line should be visible in the n + 1st
image and is not, the reason is because of occlusion by the arrival of a new object or
departure of an existing object. If a new object arrives, there will usually be new lines
that appear because of it, but it is possible that no new straight lines appear. Hence,
detecting and interpreting change, if it occurs, based on straight line segments is not
clear cut, and we deal with that problem in this paper.
2 Related Work
Some of the earlier work on change detection focuses on image sequences taken from
stationary cameras. The main drawback of these methods is their likelihood to cre-
ate false alarms in cases where pixel values are affected by viewpoint, illumination,
seasonal and atmospheric changes. This is the reason why pixel (intensity) and block
(histogram) based change detection algorithms such as image differencing [1,2] and
background modeling methods [3] fail in some applications.
Meanwhile, there exist change detection methods designed for non-stationary image
sequences. There has been a lot of work in the literature on methods based on detect-
ing moving objects [4,5], but these methods assume one or more moving objects in
a continuous video sequence. On the other hand, 3D voxel based methods [6] where
distributions of surface occupancy and associated BRDF are stored in each voxel can
manage complex and changing surfaces, but these methods suffer from sudden illumi-
nation changes, perform poorly around specular highlights and object boundaries.
To our knowledge, line segment based change detection methods have rarely been
studied in computer vision literature. Rowe and Grewe [7] make use of 2D line seg-
ments in their algorithm, but their method is specifically designed for aerial images
where the images can be registered using an affine transformation. Li et al. [8] provided
a method of detecting urban changes from a pair of satellite images by identifying
changed line segments over time. Their method does not estimate the 3D geometry
associated with the line segments and takes a pair of satellite (aerial) images as input
where line matching can be done by estimating the homography between the two im-
ages. The change detection method we propose in this work is more generic, it can
work on non-sequential image sequences where the viewpoint can change drastically
between pairs of images and it is not based on any prior assumptions on the set of
training images.
3 Multi-view Line Segment Matching and Reconstruction
Line segment matching over multiple images is known to be a difficult problem due to
its exponential complexity requirement and challenging inputs. As a result of imper-
fections in edge detection and line fitting algorithms, lines are fragmented into small
segments that diverge from the original line segments. When unreliable endpoints and
topological relationships are given as inputs, exponential complexity search algorithms
may fail to produce exact segment matching.
In this section, we present a generic, reliable and efficient method for multi-view
line matching and reconstruction. Although our method is also suitable for small base-
line problems (e.g. aerial images, continuous video sequences), such cases are not our
primary focus as their line ordering along the epipolar direction does not change much
and they can be solved efficiently by using planar homographies. In this paper, we fo-
cus on large baseline matching and reconstruction problems, where sudden illumination
changes and specular highlights make it more difficult obtain consistent line segments
in images of the same scene. These problems are more challenging as the line ordering
in different images change due to differences in viewing angles. The following subsec-
tions describe three steps of our 3D line segment reconstruction method: an efficient
line segment matching algorithm, reconstruction of single 3D lines segments and re-
construction of free form wire-frame structures.
3.1 An Efficient Line Segment Matching Algorithm
In general, the line matching problem is known to be exponential in the number of im-
ages. That is to say, given there are n images of the same scene and approximately m
lines in each image, the total complexity of the line matching problem (the size of the
search space) is O(mn ). One way to reduce the combinatorial expansion of the match-
ing problem is to use the epipolar beam [9,10]. Given the line l = (x1 , x2 ) in I, the
corresponding line in I should lie between l1 = F x1 and l2 = F x2 where F is the
fundamental matrix between I and I (see figure 2). While the epipolar beam reduces
the combinatorial expansion of the matching algorithm, this reduction highly depends
on the general alignment of line segments relative to epipolar lines. Plane sweep meth-
ods are also used to avoid the combinatorial expansion of the matching problem [11],
but these methods do not perform well when the endpoints of 2D line segments are not
consistent in different images of the same scene. Another way to increase the matching
efficiency is to use color histogram based feature descriptors for 2D line segments [12],
but these methods assume that colors only undergo slight changes and the data does not
contain specular highlights. Our work focuses on more challenging real world problems
where the above assumptions do not hold.
In this paper, we propose a new method that improves the multi-view line matching
efficiency. Our method is based on the assumption that the 3D region of interest (ROI)
Fig. 2. An example of epipolar beam. (A) I1 : “selected line” for matching is shown in “red”.
(B) I2 : the epipolar beam associated with the selected line in I1 is marked by “blue” and line
segments that lie inside the epipolar beam (i.e., candidates for matching) are shown in “red”. The
epipolar beam in image (B) reduces the search space by 5.55. (This is a color image)
is approximately known, however this assumption is not a limitation for most multi-
view applications, since the 3D ROI can be obtained by intersecting the viewing cones
of the input images. The basic idea of our approach is to divide the 3D ROI into smaller
cubes, and solve the matching problem for the line segments that lie inside each cube.
The matching algorithm iteratively projects each cube into the set of training images,
and extracts the set of 2D line segments in each image that lie (completely or partially)
inside the convex polygon associated with the 3D cube.
Assuming that lines are distributed on the cubes homogeneously, the estimated num-
ber of lines inside each cube is ( m
C ), where C is the total number of cubes, and the
mn
algorithmic complexity of the introduced matching problem is O( C n ). It must be noted
that under the assumption of homogenous distribution of lines segments in the 3D ROI
the total matching complexity is reduced by a factor of C1n , where the efficiency gain
is exponential. On the other hand, even in the existence of some dispersion over multi-
ple cubes, our proposed algorithm substantially reduces the computational complexity
of the matching algorithm. Figure 3 illustrates the quantitative comparison of different
matching algorithms for 4 images of the same scene. The matching method we use in
this work is a mixture of the algorithm described above and the epipolar beam method
(EB+Cubes).
3.2 3D Line Segment Reconstruction

3D line segments are reconstructed by using sets of corresponding 2D line segments
from different images obtained during the matching step. Each 3D line segment, L =
(X1 , X2 ) is represented by 6 parameters in 3D space, where X1 and X2 are 3D points
representing the end points. Here we assume that for each 3D line segment, correspond-
ing set of 2D line segments are available in different images. We use the Nelder-Mead
(Simplex) Method [13] to solve the minimization problem given in equation 1.

n
n
L∗ = arg min dl (li , li ) + β ds (li , li ) (1)
L∈{R3 ,R3 } i=1 i=1
Fig. 3. Quantitative comparison of four different line segment matching algorithms using four im-
ages of the same scene. Brute-force: the simplest matching method, all combinations are checked
over all images. EB: epipolar beam is used to reduce the search space. Cubes: the 3D space is
splitted into smaller sections and matching is done for each section separately. EB+Cubes: (the
method we use in this work) a combination of “Cubes” and “EB”. It is shown that “Cubes+EB”
method outperforms individual “EB” and “Cubes” methods and Brute-Force search. Notice that
the size of the search space is given in logarithmic scale.
where li = (Mi X1 ) × (Mi X2 ) is the projection of L to the ith image as an infinite line,
li = (Mi X1 , Mi X2 ) is the projection as a line segment, Mi is the projection matrix
for the ith image, dl is the distance metric between a line and a line segment and ds
is the distance metric between two line segments. The distance metrics dl and ds are
defined as

1 2
dl (l, l ) = dp (p, l )
|l|
p∈l

1 2
1 2
ds (l, l ) = dps (p, l ) + d (p , l)
|l| |l | ps
p∈l p ∈l
where dp (p, l) is the perpendicular distance of a point (p) to an infinite 2D line and
dps (p, l) is the distance of a point to a line segment.
Note that, β in equation 1, is used to control the convergence of the local search
algorithm. β is typically selected to be a number close to zero (0 < β << 1), so
the first part of the objective function dominates the local search until the algorithm
converges to the correct infinite line. Later, the second part (succeeded by β) of the
objective function starts to dominate the local search algorithm in order to find the
optimal end points for the 3D line segment.
3.3 Reconstruction of Wire-Frame Models

Next, we look into improving our line segment reconstruction results by exploiting the
pairwise constraints between 3D line segments. Even though the single line segment
reconstruction method explained in section 3.2 mostly provides good results, there ex-
ists degenerate cases that single line segment reconstruction fails to generate accurate
estimations. This usually occurs when a 3D line segment and camera centers lie on
the same plane. In this case, small errors in image space will lead to bigger errors in
3D space. One way to overcome this problem is to make use of incidence relations
between two line segments (L and Y junctions) to improve the estimation of the 3D
structure [14]. In the computer vision literature, models for special structures such as
building roofs have been developed [15,16] using aerial images with relatively small
baselines between pairs of images. These methods are not designed to work well for
reconstructing free form wire-frame models for arbitrary 3D objects where the viewing
angle between different images change drastically.
We use a pairwise checking algorithm between the reconstructed 3D line segments
for forming wire-frame model without prior information on the structure of the 3D
object. During the formation of the wire-frame model, we look at the closeness of the
endpoints of reconstructed 3D line segments as well as the closeness of the endpoints of
2D line segments in images that are associated with the reconstructed 3D line segment.
Two entities are joined if their endpoints are close enough both in the image space and
the 3D space. The wireframe model keeps growing unless there are no line segments
left that satisfy the closeness constraint.
Each wireframe model is formed as an undirected graph G = (V, E) where the set
of edges represent 3D line segments and vertices represent their endpoints. Instead of
minimizing the objective function for each line segment separately, here we minimize
an objective function for all lines (edges) in the wire-frame (graph) model. The criteria
for wire-frame (graph) minimization problem is given in equation 2.

n
n

G∗ = arg min dl (lie , lie ) + β ds (lie , lie ) (2)
V ∈R3NV e∈E i=1 i=1
where NV is the number of vertices in a wire-frame model and lie is the line in ith
image associated with edge e ∈ E in graph G = (V, E).
Note that, the wireframe model introduces more constraints, therefore the sum of
squares error of the individual line segment reconstruction is always smaller then the
error of the wireframe model reconstruction. On the other hand, additional constraints
provide better estimation of 3D line segments in terms of the actual 3D geometry. The
result of 3D line segment reconstruction using wire-frame models is given in figure 4.
4 Change Detection
Our change detection method is based on the appearance and disappearance of line
segments throughout an image sequence. We compare the geometry of lines rather than
the gray levels of image pixels to avoid the computationally intensive, and some-times
impossible, tasks of estimating 3D surfaces and their associated BRDFs in the model-
building stage. Estimating 3D lines is computationally much less costly and is more
Fig. 4. Reconstruction results for 3D line segments of a Jeep. (A)-(B) The Jeep is shown from two
different views. (C) Reconstruction results of single line segment reconstruction as explained in
section 3.2. (D) Reconstruction results of the wire-frame model reconstruction as proposed in
section 3.3.
reliable. Our method is widely applicable as man-made structures consisting of 3D line

segments are the main focus of most applications.
4.1 Definition of the Problem
The definition of the general change detection problem is the following: A 3D sur-
face model and BRDF are estimated for a region from a number of images taken by
calibrated cameras in various positions at distinctly different times. This permits the
prediction of the appearance of the entire region from a camera in an arbitrary new
position. When a new image is taken by a calibrated camera from a new position, the
computer must make a decision as whether a new object has appeared in the region or
the image is of the same region [17]. In our change detection problem, which differs
from the preceding, the algorithm must decide whether a new object has appeared in
the region or whether an object in the region has left. We predict the 3D model which
consists only of long and short straight lines, since estimating the complete 3D sur-
face under varying illumination conditions and in the existence of specular highlights
is often impractical. Moreover, image edges resulting from reflectance edges or from
3D surface ridges are less sensitive to image view direction and surface illumination,
hence so are the 3D curves associated with these edges. For man-made objects and for
general 3D curves, straight line approximations are usually appropriate and effective.
Our method detects changes by interpreting reconstructed 3D line segments and 2D line
segments detected in training and test images.
4.2 Decision Making Procedure
Our change detection method assigns a “state” to each 2D line segment in the test image
and each reconstructed 3D line segment from the training images. These states are:
“not-changed”, “changed (new)”, “changed (removed)” and “occluded”. The algorithm

has two independent components. The first one determines the “state” of each 2D line
segment l ∈ Vn+1 where Vn+1 is the set of all 2D line segments detected in the test
image. The second component estimates the “state” of each 3D line segment L ∈ Wn
where Wn is the set of reconstructed 3D line segments using the first n images. The
decision of whether or not a change has occurred is based on the appearance (checked
by tests T1 and T2 ) and disappearance (checked by tests T3 and T4 ) of a line segment
in a given test image.
Change Detection for 2D Lines in the Test Image. We estimate the “state” of each 2D
line segment in the new image using two statistical tests T1 and T2 . The classification
scheme for the 2D case is given in figure 5.
Fig. 5. General scheme for line segment based change detection. (Left) Change detection for 2D
line segment in the test image. (Right) Change detection for reconstructed 3D line segments from
training images. Threshold values t1 , t2 , t3 and t4 are selected to produce the desired change
detection rate.
First we apply T1 to test how well a 2D line segment in In+1 fits to the 3D model
Wn . Basically T1 is the distance of the 2D line segment in the new image to the closest
projection into In+1 of the 3D line segment in the 3D model.

T1 = min ds (l, ln+1 )
L∈Wn
The second step is, if necessary, to apply T2 to test if there exists a 2D line segment
in one of the past images that has taken from a similar viewing direction. Let’s define
the index of the image in the training set that has the closest camera projection matrix
compared to the test image as c∗ .
c∗ = arg min Mn+1 − Mi F

i∈{1,...,n}
where · F is the Frobenius norm. T2 is defined as:
T2 = min ds (l, lc∗ )

l∈Vc∗
Here we assume that for large training sets, there exists a camera close to the camera
that took the test image assuming that camera matrices are normalized.
Change Detection for Reconstructed 3D Line Segments. We estimate the “state” of

each reconstructed 3D line segment using two statistical tests T3 and T4 . The classifi-
cation scheme for the 3D case is given in figure 5.
First we apply T3 to test how well a 3D line segment in Wn fits to the 2D lines in
the new image. Basically T3 is the distance of a reconstructed 3D line segment to the
closest 2D line segment in the new image.

T3 = min ds (l, ln+1 )
l∈Vn+1

here ln+1 is the projection of L to the (n + 1)st image.
The second step is, if necessary, to apply T4 to test if there is an existing 3D line
segment in Wn that occludes the current 3D line L.

T4 = min ds (gn+1 , ln+1 )Z(Mn+1 , L, G)
G∈Wn \{L}

Here gn+1 is the projection of G into In+1 as a line segment and Z(Mn+1 , L, G) returns
1 if both endpoints of G are closer to the camera center of Mn+1 than both endpoints
of L, otherwise it returns ∞.
In this section, we present the results of our change detection algorithm in three dif-
ferent image sequences (experiments). It is assumed that the scene geometry does not
change during the training period. The reconstruction of the 3D model and the change
detection is done using the methods explained in section 3 and section 4. The aim of
each experiment is to show that our change detection method successfully detects the
changes and type of changes in test images in which the scene geometry is significantly
different than the training images.
The first sequence is a collection of 5 training images and 1 test image, all of which
are taken in a two hour interval. The result of this experiment is shown in figure 1. The
test image is taken at a different time of the day and from a different camera position;
hence the illumination and viewpoint direction is significantly different compared to
the training images. There are two important changes that have taken place in this test
image. The first is the disappearance of a vehicle that was previously parked in the
training images. The second change is the appearance of a vehicle close to the empty
parking spot. Both major changes (and their types) are detected accurately with a low
level of false alarm rate and main regions of change have been successfully predicted.
Notice that small “new lines” (shown in “red”) in the front row of the parking place are
due to the specular highlights that did not exist in the set of training images. There is
significant illumination difference between the test and training images, since the test
image is taken a few hours after the training images. The red line on the ground of the
empty parking spot is due to the shadow of another car. Geometrically, that line was
occluded by the car that has left the scene, so it can be thought as an existing line which
did not show up in training images due to self occlusion. The result of this experiment
shows that our method is robust to severe illumination and viewpoint changes.
Fig. 6. Change detection results for an urban area after training on a sequence of 20 images. (A)
A sample training image. (B) The test image. (C) Hand-marked ground truth for change where
the new objects are labeled with “red” and the removed objects are labeled with “blue”. (D) Line
segment based change detection results in which “new” lines are shown in “red” and “removed”
lines are shown in “blue”. Our method detects the permanent change regions successfully and
also recognizes a moving vehicle as an instance of temporal change. (This is a color image)
Unlike the first image sequence, the second and third image sequences do not have
significant viewpoint and illumination changes between the training and test images.
The result of the second experiment is shown in figure 6. To test our method, we man-
ually created a few changes using image manipulation tools. We removed the town bell
and added another chimney to a building. These change areas are marked with blue and
red respectively in figure 6-C. Also the building at the bottom right corner of the test
image is removed from the scene. This change is also marked with blue in figure 6-C.
These major changes are successfully detected in this experiment. Also, despite the
relatively small sizes of cars in this dataset, the change related to a moving vehicle
(temporal change) is detected successfully.
The results of the third experiment is shown in figure 7. In this experiment, two
new vehicles appear in the road and these changes are detected successfully. Similarly,
we created manual changes in the scene geometry using image manipulation tools. We
removed a few objects that existed on the terrace of the building in the bottom right part
of the test image. These changes are also detected successfully by our change detection
algorithm.
We also applied the Grimson change detection algorithm [3] to ground registered im-
ages for all sequences (see figure 8). Our 3D geometry based change detection method
performs well for the first image sequence under significant viewpoint and illumination
Fig. 7. Change detection results for an urban area after training on a sequence of 10 images. (A)
A sample training image. (B) The test image. (C) Hand-marked ground truth for change where
the new objects are labeled with “red” and the removed objects are labeled with “blue”. (D) Line
segment based change detection results in which “new” lines are shown in “red” and “removed”
lines are shown in “blue”. Our method detects the permanent change regions successfully and
also recognizes two moving vehicles as instances of temporal change. (This is a color image)
Fig. 8. Results of Grimson change detection algorithm applied to ground registered images. (A-B)
Ground truth change and the result of the Grimson algorithm for the first sequence. (C-D) Ground
truth change and the result of the Grimson algorithm for the second sequence. (E-F) Ground truth
change and the result of the Grimson algorithm for the third sequence. (This is a color image)
differences between the training and test images. On the other hand, the Grimson
method fails to detect changes reasonably due to viewpoint and illumination changes,
and the existence of specular highlights. For second and third image sequences, the
Grimson method successfully detects changes except insignificant false alarms caused
by the small viewpoint change between the training and test images.
To summarize our experimental results, we have shown that the significant change re-
gions are successfully detected in all 3 experiments. The first experimental setup shows
that our method can detect changes regardless of the changing viewpoint and illumi-
nation conditions. However, there are a few cases of insignificant false alarms possibly
caused by shadows, specular highlights and lack of 3D geometry of newly exposed
line segments in the test image. Also, it must be noted that unlike other change detec-
tion methods, our method detects the type of major changes successfully in almost all
experiments.
6 Conclusion and Future Work

In this paper, we present the first generally applicable 3D line segment based change
detection method for images taken from arbitrary viewing directions, at different times
and under varying illumination conditions. Our method has been shown to detect sig-
nificant changes with high accuracy on three different change detection experiments.
Experiments indicate that our algorithm is capable of efficiently matching and accu-
rately reconstructing small and large line segments, and successfully detecting changes
(and their types) by interpreting 2D and 3D line segments. We show that our multi-view
line segment matching algorithm works faster than other commonly used matching al-
gorithms, and our 3D line segment reconstruction algorithm, exploring the connectivity
of 2D and 3D line segments in order to constrain them in configurations, improve the
accuracy of existing individual 3D line segment reconstruction techniques.
Future work will involve detection of shadow and specular highlight regions to im-
prove the result of change detection by reducing the false alarm rate. Additionally, we
are investigating ways of propagating the uncertainty of line segments during the 3D
reconstruction process, and improving the change detection algorithm to work with un-
certain projective geometry. Additionally, some initial experiments have already been
performed to interpret the type of change associated with an existing pre-occluded line
segment that appears in the test image. It is necessary to build statistical models using
larger datasets to better interpret these changes related to newly appeared line segments
in the test image.
Acknowledgments. This work was funded, in part, by the Lockheed Martin Corpora-
tion. We are grateful to Joseph L. Mundy for many insightful discussions on this work.
We also thank Ece Kamar for her helpful comments on earlier drafts of this paper.
References
1. Bruzzone, L., Prieto, D.F.: Automatic analysis of the difference image for unsupervised
change detection. IEEE Transactions on Geoscience and Remote Sensing 38(3), 1171–1182
(2000)
2. Bruzzone, L., Prieto, D.F.: An adaptive semiparametric and context-based approach to un-
supervised change detection in multitemporal remote-sensing images. IEEE Transactions on
Image Processing 11(4), 452–466 (2002)
3. Stauffer, C., Grimson, W.E.L.: Adaptive background mixture models for real-time tracking.
In: CVPR, pp. 246–252 (1999)
4. Yalcin, H., Hebert, M., Collins, R.T., Black, M.J.: A flow-based approach to vehicle detection
and background mosaicking in airborne video. In: CVPR, vol. II, p. 1202 (2005)
5. Broadhurst, A., Drummond, T., Cipolla, R.: A probabilistic framework for space carving. In:
ICCV, pp. 388–393 (2001)
6. Pollard, T., Mundy, J.L.: Change detection in a 3-d world. In: CVPR, pp. 1–6 (2007)
7. Rowe, N.C., Grewe, L.L.: Change detection for linear features in aerial photographs us-
ing edge-finding. IEEE Transactions on Geoscience and Remote Sensing 39(7), 1608–1612
(2001)
8. Li, W., Li, X., Wu, Y., Hu, Z.: A novel framework for urban change detection using VHR
satellite images. In: ICPR, pp. 312–315 (2006)
9. Schmid, C., Zisserman, A.: The geometry and matching of lines and curves over multiple
views. International Journal of Computer Vision 40(3), 199–233 (2000)
10. Heuel, S., Förstner, W.: Matching, reconstructing and grouping 3D lines from multiple views
using uncertain projective geometry. In: CVPR, pp. 517–524 (2001)
11. Taillandier, F., Deriche, R.: Reconstruction of 3D linear primitives from multiple views for
urban areas modelisation. In: Photogrammetric Computer Vision, vol. B, p. 267 (2002)
12. Bay, H., Ferrari, V., Gool, L.J.V.: Wide-baseline stereo matching with line segments. In:
CVPR, vol. I, pp. 329–336 (2005)
13. Nelder, J.A., Mead, R.: A simplex method for function minimization. Computer Journal 7,
308–313 (1965)
14. Taylor, C.J., Kriegman, D.J.: Structure and motion from line segments in multiple images.
IEEE Transactions on Pattern Analysis and Machine Intelligence 17(11), 1021–1032 (1995)
15. Moons, T., Frère, D., Vandekerckhove, J., Gool, L.V.: Automatic modelling and 3d recon-
struction of urban house roofs from high resolution aerial imagery. In: Burkhardt, H.-J., Neu-
mann, B. (eds.) ECCV 1998. LNCS, vol. 1406, pp. 410–425. Springer, Heidelberg (1998)
16. Baillard, C., Schmid, C., Zisserman, A., Fitzgibbon, A.W.: Automatic line matching and 3D
reconstruction of buildings from multiple views. In: ISPRS Congress, pp. 69–80 (1999)
17. Radke, R.J., Andra, S., Al-Kofahi, O., Roysam, B.: Image change detection algorithms: a
systematic survey. IEEE Transactions on Image Processing 14(3), 294–307 (2005)
Action Recognition with a Bio–inspired
Feedforward Motion Processing Model: The
Richness of Center-Surround Interactions
Maria-Jose Escobar and Pierre Kornprobst
Odyssée project team, INRIA Sophia-Antipolis, France

{mjescoba,pkornp}@sophia.inria.fr
Abstract. Here we show that reproducing the functional properties of

MT cells with various center–surround interactions enriches motion rep-
resentation and improves the action recognition performance. To do so,
we propose a simplified bio–inspired model of the motion pathway in
primates: It is a feedforward model restricted to V1-MT cortical layers,
cortical cells cover the visual space with a foveated structure and, more
importantly, we reproduce some of the richness of center-surround inter-
actions of MT cells. Interestingly, as observed in neurophysiology, our
MT cells not only behave like simple velocity detectors, but also respond
to several kinds of motion contrasts. Results show that this diversity
of motion representation at the MT level is a major advantage for an
action recognition task. Defining motion maps as our feature vectors,
we used a standard classification method on the Weizmann database:
We obtained an average recognition rate of 98.9%, which is superior to
the recent results by Jhuang et al. (2007). These promising results en-
courage us to further develop bio–inspired models incorporating other
brain mechanisms and cortical layers in order to deal with more complex
videos.
1 Introduction
Action recognition in real scenes remains a challenging problem in computer

vision. Until recently, most proposed approaches considered simplified sequence
databases and relied on simplified assumptions or heuristics. Some examples
of these kind of approaches are [1,2,3,4,5], where one could find therein other
references and further information.
Motion is the key feature for a wide class of computer vision (CV) approaches:
Existing methods consider different motion representations or characteristics,
such as coarse motion estimation, global motion distribution, local motion fea-
ture detection or spatio-temporal structure learning [6,7,8,9,10,11,12]. Following
this general idea which is to consider motion as an informative cue for action
recognition (AR), we present a bio-inspired model for motion estimation and
representation. Interestingly, it is confirmed that in the visual system the mo-
tion pathway is also very much involved in the AR task [10], but of course other

The Richness of Center-Surround Interactions 187
brain areas (e.g., the form pathway) and mechanisms (e.g., top-down attentional
mechanisms) are also involved to analyze complex general scenes.
Among recent bio-inspired approaches for AR, [13] proposed a model for the
visual processing in the dorsal (motion) and ventral (form) pathways. They
validated their model in the AR task using stick figures constructed from real
sequences. More recently, [14] proposed a feedforward architecture, which can
be seen as an extension of [15]. In [14], the authors mapped their model to the
cortical architecture, essentially V1 (with simple and complex cells). The only
clear bio-inspired part is one of the models for S1 units and the pooling aspect.
The use of spatio-temporal chunks seems to be supported also but the authors
never claim any biological relevance for the corresponding subsequent processing
stages (from S2 to C3). The max operator is also controversial and not supported
in neurophysiology because it mainly does not allow feedbacks.
In this article, we follow the same objective as in [14], which is to propose a
bio-inspired model of motion processing for AR in real sequences. Our model
will be a connection-based network, in which a large number of neuron-like
processing units operate in parallel. Each unit neuron will have an ‘activation
level’ membrane potential that represents the strength of a particular feature
in the environment. Here, our main contribution will be to better account for
the visual system properties, and in particular, at MT layer level: We repro-
duce part of the variety of center-surround interactions [16,17]. Then, in order
to prove the relevance of this extended motion description, we will show its ben-
efits on the AR application, and compare our results with the ones obtained
by [14].
This article presents the model described in Fig. 1 and it is organized as fol-
lows. Section 2 presents the core of the approach which is a biologically-inspired
model of motion estimation, based on a feedforward architecture. As we pre-
viously mentioned, the aim of this article is to show how a bio-inspired model
can be used in a real application such as AR. Note that we also studied some
low-level properties of the model concerning motion processing [18] but those
studies are out of the scope of this article. The first stage (Section 2.1) is the
local motion extraction corresponding to the V1 layer, with a discrete foveated
organization. The output of this layer is fed to the MT layer (Section 2.2), which
is composed of a set of neurons whose dynamics are defined by a conductance-
based neuron model. We define the connectivity between V1 and MT layers
according to neurophysiology, which defines the center-surround interactions of
a MT neuron. The output of the MT layer is a set of neuron membrane po-
tentials, whose values indicate the presence of a certain velocity or contrasts of
velocities. Then, in Section 3, we consider the problem of AR based on the MT
layer activity. In this section we also present the experimental protocol, some
validations and a comparison with the approach presented by [14]. Interestingly,
we show how the variety of surround-interactions in MT cells found in physi-
ology allows the improvement of the recognition performances. We conclude in
Section 4.
188 M.-J. Escobar and P. Kornprobst
Fig. 1. Block diagram showing the different steps of our approach from the input
image sequence as stimulus until the motion map encoding the motion pattern. (a)
We use a real video sequence as input, the input sequences are preprocessed in order
to have contrast normalization and centered moving stimuli. To compute the motion
map representing the input image we consider a sliding temporal window of length Δt.
(b) Directional-selectivity filters are applied over each frame of the input sequence in
a log-polar distribution grid obtaining the activity of each V1 cell. (c) V1 outputs feed
the MT cells which integrate the information in space and time. (d) The motion map
is constructed calculating the mean activation of MT cells inside the sliding temporal
window. The motion map has a length of NL × Nc elements, where NL is the number
of MT layers of cells and Nc is the number of MT cells per layer. This motion map
characterizes and codes the action stimulus.
2 Bio–inspired Motion Analysis Model

Several bio-inspired motion processing models have been proposed in the litera-
ture [19,20,21,22,23], those models were validated considering certain properties
of primate visual systems, but none of them has been tested in a real application
such as AR. More complex motion processing models combining not only motion
information but also connections from different brain areas can be found in e.g.
[24,25].
2.1 V1 Layer: Local Motion Detectors

Our V1 model is defined by a bank of energy motion detectors as a local motion
estimation. The processing is done through energy filters which is a reliable and
biologically plausible method for motion information analysis [26]. Each energy
motion detector will emulate a complex cell, which is formed by a non-linear
combination of V1 simple cells (see [27] for V1 cells classification). Note that the
complex cells will be tuned for the direction of motion θ (and a range of speeds).
Simple Cells are characterized by linear receptive fields where the neuron re-
sponse is a weighted linear combination of the input stimulus inside its receptive
field. By combining two simple cells in a linear manner it is possible to get
direction-selective neurons.
The direction-selectivity (DS) refers to the property of a neuron to respond to
the direction of the motion of a stimulus. The way to model this selectivity is to
obtain receptive fields oriented in space and time (Fig. 1 (b.1)). Let us consider
two spatio-temporal oriented simple cells, Fθ,f a
and Fθ,f
b
, spatially oriented in
the direction θ, and spatio-temporal frequency oriented to f = (ξ, ¯ ω̄), where ξ̄
and ω̄ are the spatial and temporal maximal responses, respectively:
Fθ,f
a
(x, y, t) = Fθodd (x, y)Hf ast (t) − Fθeven (x, y)Hslow (t),
Fθ,f
b
(x, y, t) = Fθodd (x, y)Hslow (t) + Fθeven (x, y)Hf ast (t). (1)
The spatial parts Fθodd(x, y) and Fθeven (x, y) of each conforming simple cell are
formed using the first and second derivative of a Gabor function spatially ori-
ented in θ. The temporal contributions Hf ast (t) and Hslow (t) are defined by:
Hf ast (t) = T3,τ (t) − T5,τ (t),and Hslow (t) = T5,τ (t) − T7,τ (t), (2)
tη
t
where Tη,τ (t) is a Gamma function defined by Tη,τ (t) = τ η+1 η! exp − τ , which
models the series of synaptic and cellular delays in signal transmission, from
retinal photoreceptors to V1 afferents serving as a plausible approximation of
biological findings [28].
Remark that the causality of Hf ast (t) and Hslow (t) generates a more realistic
model than the one proposed by [22] (see also [14]), where the Gaussian proposed
as temporal profile is non-causal and inconsistent with V1 physiology.
The frequency analysis is required to a right design of our filter bank. For a
given speed, the filter covers a specified region of the spatio-temporal frequency
domain. The quotient between the highest temporal frequency activation (ω̄)
¯ is the speed of the filter. So, the filter will
and the highest spatial frequency (ξ)
be able to detect the motion for a stimulus whose spatial frequency lies inside
the energy spectrum of the filter. To pave all the space in a homogeneous way, it
is necessary to take more than one filter for the same spatio-temporal frequency
orientation (Fig. 1 (b.2)).
Complex Cells are also direction-selective neurons, however they include other
characteristics that cannot be explained by a linear combination of the input
stimulus. The complex cell property that we want to keep in this model is the
invariance to contrast polarity.
Based on [26], we define the ith V1 complex cell, located at xi = (xi , yi ), with
spatial orientation θi and spatio-temporal orientation fi = (ξ¯i , ω̄i ) as
# $2 # $2
Cxi ,θi ,fi (t) = Fθai ,fi ∗ I (xi , t) + Fθbi ,fi ∗ I (xi , t) , (3)
where the symbol ∗ represents the spatio-temporal convolution between the sim-
ple cells defined in (1) and the input sequence I(x, t). With this definition, the
cell response is independent of stimulus contrast sign and constant in time for a
drifting grating as input stimulus.
Finally, it is well known in biology that the V1 output shows several nonlinear-
ities due to: response saturation, response rectification, or contrast gain control
[29]. In order to obtain a nonlinear saturation in the V1 response, the V1 output
is passed through a sigmoid function S(·), where the respective parameters were
tuned to have a suitable response in the case of drifting gratings as inputs. So,
finally the V1 output will be given by:
riV 1 = S (Cxi ,θi ,fi (t)) . (4)
2.2 MT Layer: Higher Order Motion Analysis
Modeling Dynamics of MT Neurons. In this article, the dynamics of the

MT neurons are modeled by a simplified conductance-based neuron (without
input currents) [30]. Considering a MT neuron i, its membrane potential uMT
i (t)
evolves in time according to the conductance-driven equation:
duMT (t) exc inh

τ i
i (t) E
= Gexc − uMT (t) + Ginh
i (t − δ) E − uMT (t)
dt
i

i
+ g L E L − uMT
i (t) , (5)
where E exc , E inh and E L = 0 are constant which typical values of 70mV, -10mV
and
# inh0mV, respectively.
$ According to (5), uMT
i (t) will belong to the interval
E ,E exc
and it will be driven by several influences. The first term refers
to input pre-synaptic neurons and it will push the membrane potential uMT i (t)
towards E exc , with a strength defined by Gexc i (t). Similarly, the second term
also coming from pre-synaptic neurons will drive uMT i (t) towards E inh with a
strength Ginh
i (t). Finally, the last term will drive ui
MT
(t) towards the resting
potential E with a constant strength given by g . The constant δ, typically
L L
30ms, is the delay associated to the inhibitory effect.

The MT neuron i is a part of a neural network where the input conductances
i (t) and Gi (t) are obtained by pooling the activity of all the pre-synaptic
Gexc inh
neurons connected to it (Fig. 1). Each MT cell has a receptive field built from
the convergence of pre-synaptic afferent V1 complex cells (Fig. 1 (c.1)). The
excitatory inputs forming Gexci (t) are related with the activation of the classical
receptive field (CRF) of the MT cell; whereas Ginh i (t) afferents are the cells
forming the surround interactions that could modulate or not the response of
the CRF [16,17] (Fig. 1(c.2)). The surround does not elicit responses by itself,
it needs the CRF activation to be considered. According to this, the total input
conductances Gexci (t) and Gi (t) of the post-synaptic neuron i are defined by
inh

Gexc
i (t) = max 0, w r
ij j
V1
− w r
ij j
V1
, Ginh
i (t) = wij rjV 1 , (6)
j∈Ωi j∈Ωi j∈Φi
where Ωi = {j ∈ CRF | ϕij < π/2}, Ωi = {j ∈ CRF | ϕij > π/2} and
Φi = {j ∈ Surround | ϕij < π/2}, and where the connection weight wij is the
efficacy of the synapse from neuron j to neuron i, which is proportional to the
angle ϕij between the two preferred motion direction-selectivity of the V1 and
MT cell. It is important to remark that the values of the conductances will be
always greater or equal to zero, and their positive or negative contribution to
uMT
i (t) is due to the values of E exc and E inh .
The connection weights wij will be given by
wij = kc wcs (xi − xj ) cos(ϕij ), 0 ≤ ϕij ≤ π, (7)
where kc is an amplification factor, ϕij is the absolute angle between the pre-
ferred cell direction of the MT cell i and the preferred cell direction of the V1 cell
j. The weight wcs (·) is associated to the distance between the MT cell positioned
at xi = (xi , yi ) and the V1 cell positioned at xj = (xj , yj ), but also depends on
the CRF or surround associated to the MT cell.
Remark. Many studies on MT focused on motion direction selectivity (DS), but

very few on speed selectivity (see, e.g., [31,32,33]), showing that speed coding
relies on complex and unclear mechanisms. Based on this, here we only consid-
ered the motion direction and not the motion speed, as can be seen in (6): Our
MT cells pool V1 cells just considering their motion DS, and not their spatio-
temporal tuning. However, note that it is also possible to pool differently V1
cells in order to extract some speed information, as proposed for example in
[22,23,34]. As a result, one could obtain a velocity field qualitatively similar to
an optical flow (i.e., one velocity per position).
Modeling the Richness of Surround Modulations. The activation of a MT

neuron inside its CRF can be modulated by the activation of a surround area [16],
which is usually ignored in most MT-like models. In most cases this modulation
is inhibitory, but Huang et al. [35] showed that this interaction, depending on
the input stimulus, can be also integrative. The direction tuning of the surround
compared with the center tends to be either the same or opposite, but rarely
orthogonal.
Half of MT neurons have asymmetric receptive fields introducing anisotropies
in the processing of the spatial information [16]. The neurons with asymmetric
receptive fields seem to be involved in the encoding of important surfaces fea-
tures, such as slant and tilt or curvature. Their geometry is the main responsible
of the direction tuning of the MT cell and it changes along time.
Considering this, we included four types of MT cells (Fig. 2): One basic type
of cell just only activated by its CRF, and three other types with inhibitory
surrounds. We claim that inhibitory surrounds contain key information about
the motion characterization (such as motion contrasts), as we will illustrate in
Section 3. The tuning direction of the surround is always the same as the CRFs,
but their spatial geometry changes, from symmetric to asymmetric-unilateral
and asymmetric-bilateral surround interactions. It is important to mention that
this approach is a coarse approximation of the real receptive field shapes.
Fig. 2. MT center-surround interactions modeled in our approach. The classical recep-

tive field CRF (a) is modeled with a Gaussian. All the surrounds from (b) to (d) are also
modeled by Gaussians. In (b) the surround is symmetric. The two groups of cells with
asymmetric surrounds are represented in (c) and (d). (c) has a bilateral asymmetric
surround and (d) is a unilateral asymmetric surround. There is an important presence
of anisotropic surround interactions in MT cells: In [16,17], the authors showed that
within the MT cells with surround suppression, the configuration (b) is present only
in the 25% of the cells, while (c) and (d) cover the resting percentage with a presence
of 50% and 25%, respectively.
3 Action Recognition Based on MT Activity

3.1 Describing Motion Activity by a Motion Map
In this section, we use a standard supervised classification method which has
no biological inspiration. To do this, one needs to define the correspondence
between the input space (here the space of sequences) and a feature space, but
also a notion of distance between feature vectors. We considered the simpler
case of supervised classification which means that for some inputs, the class is
known (training set). Then, considering a new sequence to be analyzed, we will
estimate the corresponding feature vector and find the best class with a classifier.
Concerning our problem, we define below feature vectors as motion maps, which
represent averaged MT cells activity in a temporal window.
Motion Map as a Feature Vector. At time t, given a video stream I(x, t)

between [t − Δt, t], we define the feature vector (from now on called motion map,
see Fig. 1(c)) as the vector which represents the average membrane potential of
the MT neurons in a temporal window [t − t, t]:

HI (t, t) = γjI (t, t) j=1,...,N ×N , (8)
l c
t
with γjI (t, t) = Δt
1
t−t uj
MT
(s)ds, and where Nl is the number of MT layers
and Nc is the number of MT cells per layer.
The motion map defined in (8) is invariant to the sequence length and its
starting point (for Δt high enough depending on the scene). It is also includes
information regarding the temporal evolution of the activation of MT cells, re-
specting the causality in the order of events. The use of a sliding window allows
us to include motion changes inside the sequence.
Definition of a Distance Measure. We propose a measure discrimination to

evaluate the similarities between two motion maps HI (t, t) and HJ (t , t ),
defined by
Nl Nc
1 (γlI (t, t) − γlJ (t , t ))2
D(HI (t, t), HJ (t , t )) = . (9)
Nl Nc γlJ (t, t) + γlJ (t , t )
l=1
This measure refers to the triangular discrimination introduced by [36]. Other

measures derived from statistics, such as Kullback-Leiber (KL) could also be
used. The experiments done using, e.g., the KL measure showed no significant
improvements. Note that (9) and the motion representation (8) can be seen as
an extension of [37].
3.2 Experiments
Implementation Details. We considered luminosity and contrast normalized
videos of size 210×210 pixels, centered on the action to recognize. Given V1
cells modeled by (3), we consider 9 layers of V1 cells. Each layer is built with V1
cells tuned with the same spatio-temporal frequency and 8 different orientations.
The 9 layers of V1 cells are distributed in the frequency space in order to tile
the whole space of interest (maximal spatial frequency of 0.5 pixels/sec and a
maximal temporal frequency of 12 cycles/sec). The centers of the receptive fields
are distributed according to a radial log-polar scheme with a foveal uniform zone.
The limit between the two regions is given by the radius of the V1 fovea R0 (80
pixels). The cells with an eccentricity less than R0 have an homogeneous density
and receptive fields size. The cells with an eccentricity greater than R0 have a
density and a receptive field size depending on its eccentricity, giving a total of
4473 cells per layer.
The MT cells are also distributed in a log-polar architecture, but in this case
R0 is 40 pixels giving a total of 144 cells per layer. Different layers of MT cells
conform our model. Four different surround interactions were used in the MT
construction (see Fig. 2). Each layer, with a certain surround interaction, has 8
different directions.
gL = 0 gL = 0.25
TS = 4
TS = 6
Fig. 3. Recognition error rate obtained for Weizmann database using the four different
cells described in Fig. 2. We took all the combinations possibles considering 4 or 6
subjects in the training set (TS). For both cases, we ran the experiments with g L = 0
and g L = 0.25, and three surround-interactions: just CRF (black bars), CRF plus
isotropic surround suppression (gray bars) and CRF plus isotropic and anisotropic
surround suppression (red bars).
Experimental Protocol. In order to evaluate the performance of our algo-

rithm, we used the Weizmann Database1 : This database contains 9 different
samples of different people doing 9 actions: bending (bend ), jumping jack (jack ),
jumping forward on two legs (jump), jumping in place on two legs (pjump), run-
ning (run), galloping sideways (side), walking (walk ), waving one hand (wave1 )
and waving two hands (wave2 ). The number of frames per sequence is variable
and depends on the action.
We selected the actions of 4 or 6 (as in [14]) random subjects as training set
(total of 36 or 64 sequences, respectively) and use the remaining 5 or 3 subjects
for the test set (45 or 27 sequences, respectively). All the motion maps of the
training set were obtained and stored in a data container. We used a RAW
classifier2 : When a new input sequence belonging to the test set is presented
to the system, the motion map is calculated (with Δt covering here all the
1
http://www.wisdom.weizmann.ac.il/∼ vision/SpaceTimeActions.html
2
Note that we repeated the experiments with a standard SVM classifier but we did
not get significant improvements in the recognition performance.
Fig. 4. Histograms obtained from the recognition error rates of our approach using all
the cells defined in Fig. 2 for Weizmann database and the same experiment protocol
used in [14]. The gray bars are our histogram obtained for g L = 0.25. (a) Mean recog-
nition error rate obtained by [14] (GrC2 , dense C2 features): 8.9%/ ± 5.9. (b) Mean
recognition error rate obtained by [14] (GrC2 , sparse C2 features): 3.0%/ ± 3.0. (c)
Mean recognition error rate obtained with our approach: 1.1%/ ± 2.1.
dwalk (1) (2)

dside
(3) (4)
Fig. 5. Results obtained for the robustness experiments carried out for the three input
sequences represented by the snapshots shown for normal-walker (1), noisy sequence
(2), legs-occluded sequence (3) and moving-background sequence (4). In all the cases
the recognition was correctly performed as walk and the second closest distance was
to the class side. The red bars indicate the ratio between the distance to walk class
and the distance to side class (dwalk /dside ). The experiments were done for the three
configurations of surround-suppression: (a) just CRF, (b) CRF with isotropic surround
and (c) CRF with isotropic/anisotropic surround (g L = 0.25).
sequence) and it is compared using (9) to all motion maps stored in the training
set. The class of the sequence with the shortest distance is assigned as the match
class. The experiments were done considering every possible selection of 4 or
6 subjects, giving a total of 126 or 84 experiments. As output we obtained
histograms showing the frequency of the recognition error rates.
Results. In order to quantify the influence of the information coded by center-

surround interactions, we did the experiments with the different configurations
shown in Fig. 2. The cells were combined in order to create three different motion
maps: just considering the CRF, CRF plus the isotropic surround interaction,
and finally considering all the cells described in Fig. 2, i.e., with isotropic and
anisotropic surround interactions. Results are summarized in the histograms
shown in Fig. 3. Results show that g L > 0 significantly improves the performance
of our system, mainly because the constant leak term attracts the membrane
potential of the cell to its resting value (E L = 0), avoiding possible saturation.
It is also important to remark that in the case g L = 0, the effect of inhibitory
surrounds (either isotropic or anisotropic) is stronger than the case of g L = 0.25.
The explanation is that the inhibitory surround is the only mechanism to reduce
the activation of the cell. Maybe this effect can be compensated in the case of
g L = 0.25 by adding more relevance to the response of the cells with inhibitory
surround. Remark that the results have a strong variability and so that the
recognition performance highly depends on the sequences used to define the
training set.
In the case where 6 random subjects were taken to construct the training set,
we compared our results with [14]. As previously mentioned, we estimated the
performance of our approach based on all the possible combinations (84), and
not only on 5 random trials (as in [14]). In Fig. 4, we show the histogram with the
different recognition error rates obtained with our approach using the motion
maps generated for the CRF and isotropic/anisotropic surround interactions
cells. We obtained an average recognition rate of 98.9% (i.e., mean error rate of
1.1%), which exceeds the results obtained by [14].
To test the robustness of our approach, we considered input sequences with
different kinds of perturbations (Fig. 5): noise (case (2)), legs-occlusion (case
(3)) and moving textured background (case (4)). Both noisy and legs-occluded
sequences were created starting from the sequence shown in Fig. 5(1), which was
extracted from the training set for the robustness experiments. The legs-occluded
sequence was created placing a black box on the original sequence before the
centered cropping. The noisy sequence was created adding Gaussian noise. The
moving-background sequence was taken from [38]. For the original sequence and
the three modified input sequences the recognition was correctly performed as
walk. A graph with the ratio between the shortest distance to walk class and
the distance to the second closest class (side for the all the cases) is shown in
Fig. 5: the inclusion of the anisotropic surround interaction makes the model
less sensitive to occlusions or noise.
4 Conclusion
We proposed a feedforward bio-inspired model of V1-MT cortical layers that

can be used for solving several aspects of motion integration [18], but also high-
level tasks such as AR for natural scenes stimuli. Our model offers an efficient
platform to unveil the contribution of different components involved in visual
processing within a single experimental framework. One clear advantage of our

model is that it is generic: Unlike [13], there is no need to tune the properties
of local motion given the specific application of AR. Unlike optical-flow based
models, where a single velocity is assigned to each point, our model reproduces to
some extent the richness of center-surround interactions, giving different kinds of
motion contrasts for several orientations at every point. Interestingly, we showed
that taking into account this diversity of MT cells improves the recognition
performance. Our interpretation is that cells with inhibitory surrounds bring
information related to velocity opponency or singularities in the velocity field of
the input stimulus.
Future work will be focused on better exploiting the dynamical state of the MT
layer. Here, we defined the feature vector as the motion map, which represents
the average membrane potential of MT neurons in a temporal window. Since it
is averaged, this representation obviously misses the information about the fine
dynamical properties and the evolution of MT cells. For example, our goal will
be to detect and take into account synchronizations and temporal correlations
between cells.
Another perspective is about enriching the model with other brain functions
or cortical layers. Of course, the motion pathway is not the only actor for AR
in the visual system. Like every motion-based approach for AR, our approach
is likely to be limited. It will fail in complex situations such as those with large
occlusions, complex backgrounds or multiple persons. To do this, one has to
consider more complex processing corresponding to additional brain areas (e.g.,
V2, V4 or IT) and top-down mechanisms such as attention (e.g. [19]).
Acknowledgements
This work was partially supported by the EC IP project FP6-015879, FACETS

and CONICYT Chile. We also would like to thank John Tsotsos and Guillaume
S. Masson for their valuable comments.
References
1. Gavrila, D.: The visual analysis of human movement: A survey. Computer Vision
and Image Understanding 73(1), 82–98 (1999)
2. Goncalves, L., DiBernardo, E., Ursella, E., Perona, P.: Monocular tracking of the
human arm in 3D. In: Proceedings of the 5th International Conference on Computer
Vision, June 1995, pp. 764–770 (1995)
3. Mokhber, A., Achard, C., Milgram, M.: Recognition of human behavior by space-
time silhouette characterization. Pattern Recognition Letters 29(1), 81–89 (2008)
4. Seitz, S., Dyer, C.: View-invariant analysis of cyclic motion. The International
5. Collins, R., Gross, R., Shi, J.: Silhouette-based human identification from body
shape and gait. In: 5th Intl. Conf. on Automatic Face and Gesture Recognition, p.
366 (2002)
6. Zelnik-Manor, L., Irani, M.: Event-based analysis of video. In: Proceedings of

CVPR 2001, vol. 2, pp. 123–128 (2001)
7. Efros, A., Berg, A., Mori, G., Malik, J.: Recognizing action at a distance. In:
Proceedings of the 9th International Conference on Computer Vision, vol. 2, pp.
726–734 (October 2003)
8. Laptev, I., Capuo, B., Schultz, C., Lindeberg, T.: Local velocity-adapted motion
events for spatio-temporal recognition. Computer Vision and Image Understand-
ing 108(3), 207–229 (2007)
9. Dollar, P., Rabaud, V., Cottrell, G., Belongie, S.: Behavior recognition via sparse
spatio-temporal features. In: VS-PETS, pp. 65–72 (2005)
10. Michels, L., Lappe, M., Vaina, L.: Visual areas involved in the perception of human
movement from dynamic analysis. Brain Imaging 16(10), 1037–1041 (2005)
11. Niebles, J.C., Wang, H., Fei-Fei, L.: Unsupervised learning of human action cate-
gories using spatial–temporal words. Internation Journal of Computer Vision 79(3),
299–318 (2008)
12. Wong, S.F., Kim, T.K., Cipolla, R.: Learning motion categories using both seman-
tic and structural information. In: Proceedings of the International Conference on
Computer Vision and Pattern Recognition, pp. 1–6 (June 2007)
13. Giese, M., Poggio, T.: Neural mechanisms for the recognition of biological move-
ments and actions. Nature Reviews Neuroscience 4, 179–192 (2003)
14. Jhuang, H., Serre, T., Wolf, L., Poggio, T.: A biologically inspired system for action
recognition. In: Proceedings of the 11th International Conference on Computer
Vision, pp. 1–8 (2007)
15. Serre, T., Wolf, L., Poggio, T.: Object recognition with features inspired by visual
cortex. In: Proceedings of the International Conference on Computer Vision and
Pattern Recognition, pp. 994–1000 (June 2005)
16. Xiao, D.K., Raiguel, S., Marcar, V., Orban, G.A.: The spatial distribution of the
antagonistic surround of MT/V5 neurons. Cereb Cortex 7(7), 662–677 (1997)
17. Xiao, D., Raiguel, S., Marcar, V., Koenderink, J., Orban, G.A.: Spatial hetero-
geneity of inhibitory surrounds in the middle temporal visual area. Proceedings of
the National Academy of Sciences 92(24), 11303–11306 (1995)
18. Escobar, M., Masson, G., Kornprobst, P.: A simple mechanism to reproduce the
neural solution of the aperture problem in monkey area MT. Research Report 6579,
INRIA (2008)
19. Tsotsos, J., Liu, Y., Martinez-Trujillo, J., Pomplun, M., Simine, E., Zhou, K.:
Attending to visual motion. Computer Vision and Image Understanding 100, 3–40
(2005)
20. Nowlan, S., Sejnowski, T.: A selection model for motion processing in area MT of
primates. J. Neuroscience 15, 1195–1214 (1995)
21. Rust, N., Mante, V., Simoncelli, E., Movshon, J.: How MT cells analyze the motion
of visual patterns. Nature Neuroscience (11), 1421–1431 (2006)
22. Simoncelli, E.P., Heeger, D.: A model of neuronal responses in visual area MT.
Vision Research 38, 743–761 (1998)
23. Grzywacz, N., Yuille, A.: A model for the estimate of local image velocity by cells
on the visual cortex. Proc. R. Soc. Lond. B. Biol. Sci. 239(1295), 129–161 (1990)
24. Berzhanskaya, J., Grossberg, S., Mingolla, E.: Laminar cortical dynamics of visual
form and motion interactions during coherent object motion perception. Spatial
Vision 20(4), 337–395 (2007)
25. Bayerl, P., Neumann, H.: Disambiguating visual motion by form–motion interac-
tion – a computational model. International Journal of Computer Vision 72(1),
27–45 (2007)
26. Adelson, E., Bergen, J.: Spatiotemporal energy models for the perception of motion.
Journal of the Optical Society of America A 2, 284–299 (1985)
27. Carandini, M., Demb, J.B., Mante, V., Tollhurst, D.J., Dan, Y., Olshausen, B.A.,
Gallant, J.L., Rust, N.C.: Do we know what the early visual system does? Journal
of Neuroscience 25(46), 10577–10597 (2005)
28. Robson, J.: Spatial and temporal contrast-sensitivity functions of the visual system.
J. Opt. Soc. Am. 69, 1141–1142 (1966)
29. Albrecht, D., Geisler, W., Crane, A.: Nonlinear properties of visual cortex neurons:
Temporal dynamics, stimulus selectivity, neural performance, pp. 747–764. MIT
Press, Cambridge (2003)
30. Destexhe, A., Rudolph, M., Paré, D.: The high-conductance state of neocortical
neurons in vivo. Nature Reviews Neuroscience 4, 739–751 (2003)
31. Priebe, N., Cassanello, C., Lisberger, S.: The neural representation of speed in
macaque area MT/V5. Journal of Neuroscience 23(13), 5650–5661 (2003)
32. Perrone, J., Thiele, A.: Speed skills: measuring the visual speed analyzing proper-
ties of primate mt neurons. Nature Neuroscience 4(5), 526–532 (2001)
33. Liu, J., Newsome, W.T.: Functional organization of speed tuned neurons in visual
area MT. Journal of Neurophysiology 89, 246–256 (2003)
34. Perrone, J.: A visual motion sensor based on the properties of V1 and MT neurons.
Vision Research 44, 1733–1755 (2004)
35. Huang, X., Albright, T.D., Stoner, G.R.: Adaptive surround modulation in cortical
area MT. Neuron. 53, 761–770 (2007)
36. Topsoe, F.: Some inequalities for information divergence and related measures of
discrimination. IEEE Transactions on information theory 46(4), 1602–1609 (2000)
37. Zelnik-Manor, L., Irani, M.: Statistical analysis of dynamic actions. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence 28(9), 1530–1535 (2006)
38. Blank, M., Gorelick, L., Shechtman, E., Irani, M., Basri, R.: Actions as space-time
shapes. Proceedings of the 10th International Conference on Computer Vision 2,
1395–1402 (2005)
Linking Pose and Motion
Andrea Fossati and Pascal Fua
Computer Vision Laboratory

École Polytechnique Fédérale de Lausanne (EPFL)
1015 Lausanne, Switzerland
{andrea.fossati,pascal.fua}@epfl.ch
Abstract. Algorithms designed to estimate 3D pose in video sequences

enforce temporal consistency but typically overlook an important source
of information: The 3D pose of an object, be it rigid or articulated, has
a direct influence on its direction of travel.
In this paper, we use the cases of an airplane performing aerobatic
maneuvers and of pedestrians walking and turning to demonstrate that
this information can and should be used to increase the accuracy and
reliability of pose estimation algorithms.
1 Introduction
Temporal consistency is a key ingredient in many 3D pose estimation algorithms
that work on video sequences. However, the vast majority of methods we know of
neglect an important source of information: The direction in which most objects
travel is directly related to their attitude. This is just as true of the fighter plane
of Fig. 1(a) that tends to move in the direction in which its nose points as of the
pedestrian of Fig. 1(b) who is most likely to walk in the direction he is facing.
The relationship, though not absolute—the plane can slip and the pedestrian
can move sideways—provides nevertheless useful constraints.
There are very many Computer Vision papers on rigid, deformable, and ar-
ticulated motion tracking, as recent surveys can attest [1,2]. In most of these,
temporal consistency is enforced by regularizing the motion parameters, by re-
lating parameters in an individual frame to those estimated in earlier ones, or
by imposing a global motion model. However, we are not aware of any that ex-
plicitly take the kind of constraints we propose into account without implicitly
learning it from training data, as is done in [3].
In this paper, we use the examples of the plane and the pedestrian to show
that such constraints, while simple to enforce, effectively increase pose estimation
reliability and accuracy for both rigid and articulated motion. In both cases, we
use challenging and long video sequences that are shot by a single moving camera

This work has been funded in part by the Swiss National Science Foundation and
in part by the VISIONTRAIN RTN-CT-2004-005439 Marie Curie Action within the
EC’s Sixth Framework Programme. The text reflects only the authors’ views and the
Community is not liable for any use that may be made of the information contained
therein.

Linking Pose and Motion 201
(a) (b)
Fig. 1. Airplanes and people are examples of objects that exhibit a favored direction
of motion. (a) We project the 3D aircraft model using the recovered pose to produce
the white overlay. The original images are shown in the upper right corner. (b) We
overlay the 3D skeleton in the recovered pose, which is correct even when the person
is occluded.
that can zoom to keep the target object in the field of view, rendering the use
of simple techniques such as background subtraction impractical.
2 Related Work and Approach

Non-holonomic constraints that link direction of travel and position have been
widely used in fields such as radar-based tracking [4] or robot self-localization [5],
often in conjunction with Kalman filtering. However, these approaches deal with
points moving in space and do not concern themselves with the fact that they
are extended 3D objects, whether rigid or deformable, that have an orientation,
which conditions the direction in which they move. Such constraints have also
been adopted for motion synthesis in the Computer Graphics community [6],
but they are not directly applicable in a Computer Vision context since they
make no attempt at fitting model to data.
Tracking rigid objects in 3D is now a well understood problem and can rely
on many sources of image information, such as keypoints, texture, or edges [1]. If
202 A. Fossati and P. Fua
the image quality is high enough, simple dynamic models that penalize excessive
speed or acceleration or more sophisticated Kalman filtering techniques [7] are
sufficient to enforce temporal consistency. However, with lower quality data such
as the plane videos of Fig. 1(a), the simple quadratic regularization constraints [8]
that are used most often yield unrealistic results, as shown in Fig. 2.
(a) (b) (c)
Fig. 2. The first 50 frames of the first airplane sequence. The 3D airplane model is
magnified and plotted once every 5 frames in the orientation recovered by the algorithm:
(a) Frame by Frame tracking without regularization. (b) Imposing standard quadratic
regularization constraints. (c) Linking pose to motion produces a much more plausible
set of poses. Note for example the recovered depth of the brightest airplane: In (a) and
(b) it appears to be the frontmost one, which is incorrect. In (c) the relative depth is
correctly retrieved.
Tracking a complex articulated 3D object such as a human body is much more

complex and existing approaches remain brittle. Some of the problems are caused
by joint reflection ambiguities, occlusion, cluttered backgrounds, non-rigidity of
tissue and clothing, complex and rapid motions, and poor image resolution.
The problem is particularly acute when using a single video to recover the 3D
motion. In this case, incorporating motion models into the algorithms has been
shown to be effective [2]. The models can be physics-based [9] or learned from
training data [10,11,12,13]. However, all of these assume the joint angles, that
define the body pose, and the global motion variables are independent. As is the
case for rigid body tracking, they typically revert to second order Gauss-Markov
modeling or Kalman filtering to smooth the global motion. Again, this can lead
to unrealistic results as shown in Fig. 3. Some approaches implicitly take into
account the relationship between pose and direction of travel by learning from
training data a low-dimensional representation that includes both [3,14,15,16].
However, the set of motions that can be represented is heavily constrained by
the contents of the training database, which limits their generality.
To remedy these problems, we explicitly link pose and motion as follows: Given
an object moving along its trajectory as depicted by Fig. 4, the angle between
Ṗt , the derivative of its position, and its orientation Λt should in general be
small. We can therefore write that
(a) (b)
Fig. 3. Recovered 2D trajectory of the subject of Fig. 1(b). The arrows represent the
direction he is facing. (a) When pose and motion are not linked, he appears to walk
sideways. (b) When they are, he walks naturally. The underlying grid is made of 1
meter squares.
Ṗt · Λt
||Ṗt || · ||Λt ||
should be close to 1. To enforce this, we can approximate the derivative of the
locations using finite differences between estimated locations P̂ at different time
instants. This approximation is appropriate when we can estimate the location
at a sufficiently high frequency (e.g. 25 Hz).
Fig. 4. The continuous curve represents the real trajectory of the object, while the
dashed lines show its approximation by finite differences
Our constraint then reduces to minimizing the angle between the finite dif-
ferences approximation of the derivative of the trajectory at time t, given by
P̂t+1 − P̂t , and the object’s estimated orientation given by Λ̂t . We write this
angle, which is depicted as filled both at time t − 1 and t in Fig. 4, as
Ṗˆt · Λ̂t (P̂t+1 − P̂t ) · Λ̂t

φt→t+1 = acos = acos
ˆ
||Ṗt || · ||Λ̂t || ||(P̂t+1 − P̂t )|| · ||Λ̂t ||
and will seek to minimize it. It is important to note that the constraint we impose
is not a hard constraint, which can never be violated. Instead, it is a prior that
can be deviated from if the data warrants it. In the remainder of the paper we
will demonstrate the effectiveness of this idea for both rigid and articulated 3D
tracking.
3 Rigid Motion
In the case of a rigid motion, we demonstrate our approach using video sequences
of a fighter plane performing aerobatic maneuvers such as the one depicted by
Fig. 5. In each frame of the sequences, we retrieve the pose which includes
position expressed by cartesian coordinates and orientation defined by the roll,
pitch and yaw angles. We show that these angles can be recovered from single
viewpoint sequences with a precision down to a few degrees, and that linking
pose and motion estimation contributes substantially to achieving this level of
accuracy. This is extremely encouraging considering the fact that the videos we
have been working with were acquired under rather unfavorable conditions: As
can be seen in Fig. 5, the weather was poor, the sky gray, and the clouds many, all
of which make the plane less visible and therefore harder to track. The airplane
is largely occluded by smoke and clouds in some frames, which obviously has an
adverse impact on accuracy but does not result in tracking failure.
The video sequences were acquired using a fully calibrated camera that could
rotate around two axes and zoom on the airplane. Using a couple of encoders, it
could keep track of the corresponding values of the pan and tilt angles, as well
as the focal length. We can therefore consider that the intrinsic and extrinsic
camera parameters are known in each frame. In the remainder of this section,
we present our approach first to computing poses in individual frames and then
imposing temporal consistency, as depicted by Fig. 4, to substantially improve
the accuracy and the realism of the results.
3.1 Pose in Each Frame Independently

Since we have access to a 3D model of the airplane, our algorithm computes the
pose in each individual frame by minimizing an objective function Lr that is a
weighted sum of a color and an edge term:
– The color term is first computed as the Bhattacharyya distance [17] between
the color histogram of the airplane that we use as a model, whose pose was
captured manually in the first frame, and the color histogram of the image
area corresponding to its projection in subsequent frames. To this we add
a term that takes into account background information, also expressed as
a difference of color histograms, which has proved important to guarantee
robustness.
Fig. 5. Airplane video and reprojected model. First and third rows: Frames from
the input video. Note that the plane is partially hidden by clouds in some frames, which
makes the task more difficult. Second and fourth rows: The 3D model of the plane
is reprojected into the images using the recovered pose parameters. The corresponding
videos are submitted as supplemental material.
– The edge term is designed to favor poses such that projected model edges
correspond to actual image edges and plays an important role in ensuring
accuracy.
In each frame t, the objective function Lr is optimized using a particle-based
stochastic optimization algorithm [18] that returns the pose corresponding to
the best sample. The resulting estimated pose is a six-dimensional vector Ŝt =
(P̂t , Λ̂t ) = argminS Lr (S) where P̂t = (X̂t , Ŷt , Ẑt ) is the estimated position of
the plane in an absolute world coordinate system and Λ̂t = (ρ̂t , θ̂t , γ̂t ) is the
estimated orientation expressed in terms of roll, pitch and yaw angles. The esti-
mated pose Ŝt at time t is used to initialize the algorithm in the following frame
t + 1, thus assuming that the motion of the airplane between two consecutive
frames is relatively small, which is true in practice.
3.2 Imposing Temporal Consistency
Independently optimizing Lr in each frame yields poses that are only roughly
correct. As a result, the reconstructed motion is extremely jerky. To enforce
temporal consistency, we introduce a regularization term M defined over frames
t − 1, t, and t + 1 as
M (St ) = α1 ||A(Pt )||2 + α2 ||A(Λt )||2 + β(φ2t−1→t + φ2t→t+1 ) , (1)
A(Pt ) = Pt+1 − 2Pt + Pt−1 , (2)
A(Λt ) = Λt+1 − 2Λt + Λt−1 . (3)

The first two terms of (1) enforce motion smoothness. The third term is
the one of Fig. 4, which links pose to motion by forcing the orientation of the
airplane to be consistent with its direction of travel. In practice, α1 , α2 and β
are chosen to relate quantities that would otherwise be incommensurate and are
kept constant for all the sequences we used. For an N -frame video sequence,
ideally, we should minimize

N
N −1
fr (S1 , . . . , SN ) = Lr (St ) + M (St ) (4)
t=1 t=2
with respect to the poses in individual images. In practice, for long video
sequences, this represents a very large optimization problem. Therefore, in our
current implementation, we perform this minimization in sliding temporal 3-frame
windows using a standard simplex algorithm that does not require the computa-
tion of derivatives. We start with the first set of 3 frames, retain the resulting pose
in the first frame, slide the window by one frame, and iterate the process using the
previously refined poses to initialize each optimization step.
3.3 Tracking Results
The first sequence we use for the evaluation of our approach is shown in Fig. 5
and contains 1000 frames shot over 40 seconds, a time during which the plane
performs rolls, spins and loops and undergoes large accelerations.
In Fig. 6(a) we plot the locations obtained in each frame independently. In
Fig. 6(b) we imposed motion smoothness by using only the first two terms of (1).
In Fig 6(c) we link pose to motion by using all three terms of (1). The trajectories
are roughly similar in all cases. However, using the full set of constraints yields
a trajectory that is both smoother and more plausible.
In Fig. 2, we zoom in on a portion of these 3 trajectories and project the 3D
plane model in the orientation recovered every fifth frame. Note how much more
consistent the poses are when we use our full regularization term.
The plane was equipped with sophisticated gyroscopes which gave us mean-
ingful estimates of roll, pitch, and yaw angles, synchronized with the camera
(a) (b) (c)
Fig. 6. Recovered 3D trajectory of the airplane for the 40s sequence of Fig. 5: (a) Frame
by Frame tracking. (b) Imposing motion smoothness. (c) Linking pose to motion. The
coordinates are expressed in meters.
and available every third frame. We therefore use them as ground truth. Table 1
summarizes the deviations between those angles and the ones our algorithm
produces for the whole sequence. Our approach yields an accuracy improvement
over frame by frame tracking as well as tracking with simple smoothness con-
straint. The latter improvement is in the order of 5 %, which is significant if one
considers that the telemetry data itself is somewhat noisy and that we are there-
fore getting down to the same level of precision. Most importantly, the resulting
sequence does not suffer from jitter, which plagues the other two approaches, as
can be clearly seen in the videos given as supplemental material.
Table 1. Comparing the recovered pose angles against gyroscopic data for the sequence
of Fig. 5. Mean and standard deviation of the absolute error in the 3 angles, in degrees.
Roll Angle Error Pitch Angle Error Yaw Angle Error

Mean Std. Dev. Mean Std. Dev. Mean Std. Dev.
Frame by Frame 2.291 2.040 1.315 1.198 3.291 2.245
Smoothness Constraint only 2.092 1.957 1.031 1.061 3.104 2.181
Linking Pose to Motion 1.974 1.878 0.975 1.000 3.003 2.046
In Fig. 7 we show the retrieved trajectory for a second sequence, which lasts
20 seconds. As before, in Table 2, we compare the angles we recover against
gyroscopic data. Again, linking pose to motion yields a substantial improvement.
4 Articulated Motion
To demonstrate the effectiveness of the constraint we propose in the case of
articulated motion, we start from the body tracking framework proposed in [19].
In this work, it was shown that human motion could be reconstructed in 3D
(a) (b) (c)
Fig. 7. Recovered 3D trajectory of the airplane for a 20s second sequence: (a) Frame
by Frame tracking. (b) Imposing motion smoothness. (c) Linking pose to motion. The
coordinates are expressed in meters.
Table 2. Second sequence: Mean and standard deviation of the absolute error in the
3 angles, in degrees
Roll Angle Error Pitch Angle Error Yaw Angle Error

Frame by Frame 3.450 2.511 1.607 1.188 3.760 2.494
Smoothness Constraint only 3.188 2.445 1.459 1.052 3.662 2.237
by detecting canonical poses, using a motion model to infer the intermediate

poses, and then refining the latter by maximizing an image-based likelihood in
each frame independently. In this section, we show that, as was the case for
rigid motion recovery, relating the pose to the direction of motion leads to more
accurate and smoother 3D reconstructions.
In the remainder of the section, we first introduce a slightly improved version
of the original approach on which our work is based. We then demonstrate the
improvement that the temporal consistency constraint we advocate brings about.
4.1 Refining the Pose in Each Frame Independently
We rely on a coarse body model in which individual limbs are modeled as cylin-
ders. Let St = (Pt , Θt ) be the state vector that defines its pose at time t, where
Θt is a set of joint angles and Pt a 3D vector that defines the position and orien-
tation of the root of the body in a 2D reference system attached to the ground
plane.
In the original approach [19], a specific color was associated to each limb by
averaging pixel intensities in the projected area of the limb in the frames where
a canonical pose was detected. Then St was recovered as follows: A rough initial
state was predicted by the motion model. Then the sum-of-squared-differences
between the synthetic image, obtained by reprojecting the model, and the actual
one was minimized using a simple stochastic optimization algorithm.
Here, we replace the single color value associated to each limb by a histogram,
hereby increasing generality. As in Sect. 3.1, we define an objective function
La that measures the quality of the pose using the Bhattacharyya distance to
express the similarity between the histogram associated to a limb and that of
the image portion that corresponds to its reprojection. Optimizing La in each
frame independently leads, as could be expected, to a jittery reconstruction as
can be seen in the video given as supplemental material.
4.2 Imposing Temporal Consistency

In order to improve the quality of our reconstruction, we perform a global opti-
mization on all N frames between two key-pose detections, instead of minimizing
La independently in each frame. To model the relationship between poses we
learn a PCA model from a walking database and consider a full walking cycle as
a single data point in a low-dimensional space [20,11]. This lets us parameterize
all the poses Si between consecutive key-pose detections by n PCA coefficients
(α1 . . . αn ), plus a term, η, that represents possible variations of the walking
speed during the walking cycle (n = 5 in our experiments). These coefficients
do not take into account the global position and orientation of the body, which
needs to be parameterized separately. Since the walking trajectory can be ob-
tained by a 2D spline curve lying on the ground plane, defined by the position
and orientation of the root at the two endpoints of the sequence, modifying these
endpoints Pstart and Pend will yield different trajectories. The root position and
orientation corresponding to the different frames will then be picked along the
spline curve according to the value of η. It in fact defines where in the walk-
ing cycle the subject is at halftime between the two detections. For a constant
speed during a walking cycle the value of η is 0.5, but it can go from 0.3 to 0.7
depending on change in speed between the first and the second half-cycle.
We can now formulate an objective function that includes both the image
likelihood and a motion term, which, in this case, constrains the person to move
in the direction he is facing. This objective function is then minimized with
respect to the parameters introduced above (α1 , . . . , αn , Pstart , Pend , η) on the
full sequence between two consecutive key-pose detections. In other words, we
seek to minimize

N
N
fa (S1 , . . . , SN ) = La (St ) + β(φ2t−1→t ) (5)
t=1 t=2
with respect to (α1 , . . . , αn , Pstart , Pend , η), where the second term is defined the
same way as in the airplane case and β is as before a constant weight that
relates incommensurate quantities. The only difference is that in this case both
the estimated orientation and the expected motion, that define the angle φ, are
2-dimensional vectors lying on the ground plane. This term is the one that links
pose to motion. Note that we do not need quadratic regularization terms such as
the first two of (1) because our parameters control the entire trajectory, which
is guaranteed to be smooth.
4.3 Tracking Results

We demonstrate our approach on a couple of very challenging sequences. In the
sequence of Fig. 8, the subject walks along a circular trajectory and the camera
is following him from its center. At a certain point the subject undergoes a total
occlusion but the algorithm nevertheless recovers his pose and position thanks
to its global motion model. Since the tracking is fully 3D, we can also recover
the trajectory of the subject on the ground plane and his instantaneous speed
at each frame.
In Fig. 3 we examine the effect of linking or not pose to motion on the recov-
ered trajectory: That is, setting β to zero or not in (5). The arrows represent the
orientation of the subject on the ground plane. They are drawn every fifth frame.
Fig. 8. Pedestrian tracking and reprojected 3D model for the sequence of Fig. 1 First
and third rows: Frames from the input video. The recovered body pose has been
reprojected on the input image. Second and fourth rows: The 3D skeleton of the
person is seen from a different viewpoint, to highlight the 3D nature of the results. The
numbers in the bottom right corner are the instantaneous speeds derived from the re-
covered motion parameters. The corresponding videos are submitted as supplementary
material.
The images clearly show that, without temporal consistency constraints, the sub-
ject appears to slide sideways while when the constraints are enforced the motion
is perfectly consistent with the pose. This can best be evaluated from the videos
given as supplemental material.
Fig. 9. Pedestrian tracking and reprojected 3D model in a second sequence. First

and third rows: Frames from the input video. The recovered body pose has been
reprojected on the input image. Second and fourth rows: The 3D skeleton of the
person is seen from a different viewpoint, to highlight the 3D nature of the results.
The numbers in the bottom right corner are the instantaneous speeds derived from the
recovered motion parameters.
To validate our results, we manually marked the subject’s feet every 10 frames
in the sequence of Fig. 8 and used their position with respect to the tiles on the
ground plane to estimate their 3D coordinates. We then treated the vector joining
the feet as an estimate of the body orientation and the midpoint as an estimated
of its location. As can be seen in Table 3, linking pose to motion produces a small
improvement in the position estimate and a much more substantial one in the
orientation estimate, which is consistent with what can be observed in Fig. 3.
In the sequence of Fig. 9 the subject is walking along a curvilinear path and
the camera follows him, so that the viewpoint undergoes large variations. We
are nevertheless able to recover pose and motion in a consistent way, as shown
in Fig. 10 which represents the corresponding recovered trajectory.
Table 3. Comparing the recovered pose angles against manually recovered ground
truth data for the sequence of Fig. 8. It provides the mean and standard deviation
of the absolute error in the X and Y coordinates, in centimeters, and the mean and
standard deviation of the recovered orientation, in degrees.
X Error Y Error Orientation Error

Not Linking Pose to Motion 12.0 7.1 16.8 11.9 11.7 7.6
(a) (b)
Fig. 10. Recovered 2D trajectory of the subject of Fig. 9. As in Fig. 3, when orientation
and motion are not linked, he appears to walk sideway (a) but not when they are (b).
5 Conclusion
In this paper, we have used two very different applications to demonstrate that
jointly optimizing pose and direction of travel substantially improves the quality
of the 3D reconstructions that can be obtained from video sequences. We have
also shown that we can obtain accurate and realistic results using a single moving
camera.
This can be done very simply by imposing an explicit constraint that forces
the angular pose of the object or person being tracked to be consistent with their
direction of travel. This could be naturally extended to more complex interac-
tions between pose and motion. For example, when a person changes orientation,
the motion of his limbs is not independent of the turn radius. Similarly, the di-
rection of travel of a ball will be affected by its spin. Explicitly modeling these
subtle but important dependencies will therefore be a topic for future research.
References
1. Lepetit, V., Fua, P.: Monocular model-based 3d tracking of rigid objects: A survey.
Foundations and Trends in Computer Graphics and Vision (2005)
2. Moeslund, T.B., Hilton, A., Krüger, V.: A survey of advances in vision-based hu-
man motion capture and analysis. CVIU 104(2), 90–126 (2006)
3. Sidenbladh, H., Black, M.J., Sigal, L.: Implicit Probabilistic Models of Human Mo-
tion for Synthesis and Tracking. In: Heyden, A., Sparr, G., Nielsen, M., Johansen,
P. (eds.) ECCV 2002. LNCS, vol. 2350, pp. 784–800. Springer, Heidelberg (2002)
4. Bar-Shalom, Y., Kirubarajan, T., Li, X.R.: Estimation with Applications to Track-
ing and Navigation. John Wiley & Sons, Inc., Chichester (2002)
5. Zexiang, L., Canny, J.: Nonholonomic Motion Planning. Springer, Heidelberg
(1993)
6. Ren, L., Patrick, A., Efros, A.A., Hodgins, J.K., Rehg, J.M.: A data-driven ap-
proach to quantifying natural human motion. ACM Trans. Graph. 24(3) (2005)
7. Koller, D., Daniilidis, K., Nagel, H.H.: Model-Based Object Tracking in Monocular
Image Sequences of Road Traffic Scenes. IJCV 10(3), 257–281 (1993)
8. Poggio, T., Torre, V., Koch, C.: Computational Vision and Regularization Theory.
Nature 317 (1985)
9. Brubaker, M., Fleet, D., Hertzmann, A.: Physics-based person tracking using sim-
plified lower-body dynamics. In: CVPR (2007)
10. Urtasun, R., Fleet, D., Fua, P.: 3D People Tracking with Gaussian Process Dy-
namical Models. In: CVPR (2006)
11. Ormoneit, D., Sidenbladh, H., Black, M.J., Hastie, T.: Learning and tracking cyclic
human motion. In: NIPS (2001)
12. Agarwal, A., Triggs, B.: Tracking articulated motion with piecewise learned dy-
namical models. In: Pajdla, T., Matas, J(G.) (eds.) ECCV 2004. LNCS, vol. 3023,
pp. 54–65. Springer, Heidelberg (2004)
13. Taycher, L., Shakhnarovich, G., Demirdjian, D., Darrell, T.: Conditional Random
People: Tracking Humans with CRFs and Grid Filters. In: CVPR (2006)
14. Rosenhahn, B., Brox, T., Seidel, H.: Scaled motion dynamics for markerless motion
capture. In: CVPR (2007)
15. Brox, T., Rosenhahn, B., Cremers, D., Seidel, H.: Nonparametric density estima-
tion with adaptive, anisotropic kernels for human motion tracking. In: Workshop
on HUMAN MOTION Understanding, Modeling, Capture and Animation (2007)
16. Howe, N.R., Leventon, M.E., Freeman, W.T.: Bayesian reconstructions of 3D hu-
man motion from single-camera video. In: NIPS (1999)
17. Djouadi, A., Snorrason, O., Garber, F.: The quality of training sample estimates
of the bhattacharyya coefficient. PAMI 12(1), 92–97 (1990)
18. Isard, M., Blake, A.: CONDENSATION - conditional density propagation for visual
tracking. IJCV 29(1), 5–28 (1998)
19. Fossati, A., Dimitrijevic, M., Lepetit, V., Fua, P.: Bridging the Gap between De-
tection and Tracking for 3D Monocular Video-Based Motion Capture. In: CVPR
(2007)
20. Urtasun, R., Fleet, D., Fua, P.: Temporal Motion Models for Monocular and Mul-
tiview 3–D Human Body Tracking. CVIU 104(2-3), 157–177 (2006)
Automated Delineation of Dendritic Networks in Noisy
Image Stacks
Germán González1 , François Fleuret2, , and Pascal Fua1

1
Ecole Polytechnique Fédérale de Lausanne, Computer Vision Laboratory,
Bâtiment BC, CH-1015 Lausanne, Switzerland
{german.gonzalez,pascal.fua}@epfl.ch
2
IDIAP Research Institute, P.O. Box 592, CH-1920, Martigny, Switzerland
francois.fleuret@idiap.ch
Abstract. We present a novel approach to 3D delineation of dendritic networks

in noisy image stacks. We achieve a level of automation beyond that of state-
of-the-art systems, which model dendrites as continuous tubular structures and
postulate simple appearance models. Instead, we learn models from the data it-
self, which make them better suited to handle noise and deviations from expected
appearance.
From very little expert-labeled ground truth, we train both a classifier to recog-
nize individual dendrite voxels and a density model to classify segments connect-
ing pairs of points as dendrite-like or not. Given these models, we can then trace
the dendritic trees of neurons automatically by enforcing the tree structure of the
resulting graph. We will show that our approach performs better than traditional
techniques on brighfield image stacks.
1 Introduction
Full reconstruction of neuron morphology is essential for the analysis and understand-
ing of their functioning. In its most basic form, the problem involves processing stacks
of images produced by a microscope, each one showing a slice of the same piece of
tissue at a different depth.
Currently available commercial products such as Neurolucida1, Imaris2 , or Meta-
morph 3 provide sophisticated interfaces to reconstruct dendritic trees and rely heavily
on manual operations for initialization and re-initialization of the delineation proce-
dures. As a result, tracing dendritic trees in noisy images remains a tedious process. It
can take an expert up to 10 hours for each one. This limits the amount of data that can
be processed and represents a significant bottleneck in neuroscience research on neuron
morphology.
Automated techniques have been proposed but are designed to work on very high
quality images in which the dendrites can be modeled as tubular structures [1,2]. In

Supported by the Swiss National Science Foundation under the National Centre of Compe-
tence in Research (NCCR) on Interactive Multimodal Information Management (IM2).
1
http://www.microbrightfield.com/prod-nl.htm
2
http://www.bitplane.com/go/products/imaris
3
http://www.moleculardevices.com/pages/software/metamorph.html

Automated Delineation of Dendritic Networks in Noisy Image Stacks 215
(a) (b) (c)
Fig. 1. (a) Minimum intensity projection of an image stack. Each pixel value is the minimum in-
tensity value of the voxels that are touched by the ray cast from the camera through the pixel. (b)
3D tree reconstructed by our algorithm, which is best viewed in color. (c) Detail of the data vol-
ume showing the non-tubular aspect of a dendrite with the corresponding automatically generated
delineation.
practice, however, due to the underlying neuron structure, irregularities in the dyeing
process, and other sources of noise, the filaments often appear as an irregular series of
blobs surrounded by other non-neuron structures, as is the case of the brightfield image
stacks depicted by Fig. 1. Yet, such images are particularly useful for analyzing large
samples. More generally, very high resolution images take a long time to acquire and
require extremely expensive equipment, such as confocal microscopes. The ability to
automatically handle lower resolution and noisier ones is therefore required to make
these techniques more accessible. Ideally, the painstaking and data-specific tuning that
many existing methods require should also be eliminated.
In this paper, we therefore propose an approach to handling the difficulties that are
inherent to this imaging process. We do not assume an a priori dendrite model but
rely instead on supervised and unsupervised statistical learning techniques to construct
models as we go, which is more robust to unpredictable appearance changes. More
specifically, we first train a classifier that can distinguish dendrite voxels from others
using a very limited amount of expert-labeled ground truth. At run-time, it lets us detect
such voxels, some of which should be connected by edges to represent the dendritic
tree. To this end, we first find the minimum spanning tree connecting dendrite-like
voxels. We then use an Expectation-Maximization approach to learn an appearance
model for the edges that correspond to dendrites and those that do not. Finally, given
these appearance models, we re-build and prune the tree to obtain the final delineation,
such as the one depicted by Fig. 1(b), which is beyond what state-of-the-art techniques
can produce automatically.
To demonstrate the versatility of our approach, we also ran our algorithm on retinal
images, which we were able to do by simply training our classifier to recognize 2D
blood vessel pixels instead of 3D dendrite voxels.
216 G. González, F. Fleuret, and P. Fua
2 Related Work
Reconstructing networks of 3D filaments, be they blood vessels or dendrites, is an im-
portant topic in Biomedical Imaging and Computer Vision [3,4]. This typically involves
measuring how filament-like voxels are and an algorithm connecting those that appear
to be. We briefly review these two aspects below.
2.1 Finding Dendrite-Like Voxels

Most automated methods assume the filaments to be locally tubular and model them as
generalized cylinders. The most popular approach to detecting such cylindrical struc-
tures in image stacks involves computing the Hessian matrix at individual voxels by
convolution with Gaussian derivatives and relying on the eigenvalues of the Hessian to
classify voxels as filament-like or not [5,6,7]. The Hessians can be modified to create
an oriented filter in the direction of minimum variance, which should correspond to the
direction of any existing filament [8,9]. To find filaments of various widths, these meth-
ods perform the computation using a range of variances for the Gaussian masks and
select the most discriminant one. The fact that intensity changes inside and outside the
filaments has also been explicitly exploited by locally convolving the image with dif-
ferential kernels [1], finding parallel edges [10], and fitting superellipsoids or cylinders
to the vessel based on its surface integral [2,11].
All these methods, however, assume image regularities that are present in high-
quality images but not necessarily in noisier ones. Furthermore, they often require care-
ful parameter tuning, which may change from one data-set to the next. As a result,
probabilistic approaches able to learn whether a voxel belongs to a filament or not have
begun to be employed. Instead of assuming the filaments to be cylinders, they aim at
learning their appearance from the data. In [12], the eigenvalues of the structure tensor,
are represented by a mixture model whose parameters are estimated via E-M. Support
Vector Machines that operates on the Hessian’s eigenvalues have also been used to dis-
criminate between filament and non-filament voxels [13].
The latter approach [13] is closest to our dendrite detection algorithm. We however
go several steps further to increase robustness: First, we drop the Hessian and train our
classifier directly on the intensity data, thereby making fewer assumptions and being
able to handle structures that are less visibly tubular. Second, we also learn an appear-
ance model for the filament itself as opposed to individual voxels.
2.2 Reconstructing Filaments

Existing approaches to building the dendritic tree all rely on a dendritness measure of
how dendrite-like filaments look, usually based on the voxel-based measures discussed
above. They belong to one of two main classes.
The first class involves growing filaments from seed points [2,14,15,16]. This has
been successfully demonstrated for confocal fluorescent microscopy images. It is com-
putationally effective because the dendritness of filaments need only be evaluated in
a small subset of the voxels. However, it may easily fail in noisy data because of its
sequential nature. If the growing process diverges at one voxel, the rest of the dendritic
tree will be lost.
The second class requires optimizing the path between seed points, often provided
by the operator, to maximize the overall dendritness [8,11,17]. In these examples, the
authors use active contour models, geometrical constraints and the live-wire algorithm
between to connect the seeds.
By contrast to these methods that postulate an a priori cost function for connecting
voxels, our approach learns a model at run-time, which lets it deal with the potentially
changing appearance of the filaments depending on experimental conditions. Further-
more, we do this fully automatically, which is not the case for any of the methods
discussed above.
3 Methodology
Our goal is to devise an algorithm that is fully automatic and can adapt to noisy data in
which the appearance of the dendrites is not entirely predictable. Ideally we would like
to find the tree maximizing the probability of the image under a consistent generative
model. Because such an optimization is intractable, we propose an approximation that
involves the three following steps:
1. We use a hand-labeled training image stack to train once and for all a classifier that
computes a voxel’s probability to belong to a dendrite from its neighbors intensities.
2. We run this classifier on our stacks of test images, use a very permissive threshold to
select potential dendrite voxels, apply non-maximum suppression, and connect all
the surviving voxels with a minimum spanning tree. Some of its edges will corre-
spond to actual dendritic filaments and other will be spurious. We use both the correct
and spurious edges to learn filament appearance models in an EM framework.
3. Under a Markovian assumption, we combine these edge appearance models to
jointly model the image appearance and the true presence of filaments. We then
optimize the probability of the latter given the former and prune spurious branches.
As far as detecting dendrite voxels is concerned, our approach is related to the Hessian-
based approach of [13]. However, dropping the Hessian and training our classifier di-
rectly on the intensity data lets us relax the cylindrical assumption and allows us to
handle structures that are less visibly tubular. As shown in Fig. 2, this yields a marked
improvement over competing approaches.
In terms of linking, our approach can be compared to those that attempt to find opti-
mal paths between seeds [11,8] using a dendrite appearance model, but with two major
improvements: First our seed points are detected automatically instead of being manu-
ally supplied, which means that some of them may be spurious and that the connectivity
has to be inferred from the data. Second we do not assume an a priori filament model but
learn one from the data as we go. This is much more robust to unpredictable appearance
changes. Furthermore, unlike techniques that model filaments as tubular structures [1,2],
we do not have to postulate regularities that may not be present in our images.
3.1 Notations
Given the three step algorithm outlined above, we now introduce the notations we will
use to describe it in more details.
(a) Training volume (b) Grayscale (c) Hessian (d) Boosting
Fig. 2. (a) Training data. On top: image stack representing one neuron. Bellow: Manually delin-
eated filaments overlaid in white. (b,c,d) Voxels labeled as potentially belonging to a dendrite. (b)
By thresholding the grayscale images. (c) By using the Hessian. (d) By using our classifier. Note
that the seed points obtained with our method describe better the underlying neuron structure.
Let Z1 , . . . , ZN be the voxels corresponding to the local maxima of the classifier

response and will serve as vertices for the dendritic tree we will build. For 1 ≤ n ≤ N ,
let Xn be a Boolean random variable standing for whether or not there truly is a filament
at location Zn . Finally, Let x = (x1 , . . . , xN ) and x\i = (x1 , . . . , xi−1 , xi+1 , xN ).
For 1 ≤ i ≤ N and 1 ≤ j ≤ N , let Ji,j denote a random variable standing for the
appearance of the edge going from Zi to Zj and let Li,j = ||Zi − Zj || be its length. Ji,j
is obtained by sampling the voxel response of the classifier in a regular lattice between
(Zi , Zj ). Let Ai,j be a vector composed by the projection of Ji,j in a latent space and
Li,j .
Let T denote the true dendritic tree we are trying to infer. It is a graph whose vertices
are a subset of Z1 , . . . , ZN and whose edges are defined by G, a set of pairs of indexes
in {1, . . . , N } × {1, . . . , N }.
3.2 Local Dendrite Model
As discussed in Section 2, the standard approach to deciding whether voxels are inside
a dendrite or not is to compute the Hessian of the intensities and look at its eigenvalues.
This however implicitly makes strong assumptions on the expected intensity patterns.
Instead of using such a hand-designed model, we train a classifier from a small quantity
of hand-labeled neuron data with AdaBoost [18], which yields superior classification
performance as shown in Fig. 2.
More specifically, the resulting classifier f is a linear combination of weak
learners hi :

N
f (x, y, z) = αi hi (x, y, z) , (1)
i=1
where the hi represent differences of the integrals of the image intensity over two cubes
in the vicinity of (x, y, z) and Ti is the weak classifier threshold. We write
⎛ ⎞

hi (x, y, z) = σ⎝ I(x , y , z )− I(x , y , z ) − Ti ⎠ (2)
Vi1 Vi2
where σ is the sign function, Vi1 , Vi2 are respectively the two volumes defining hi , trans-
lated according to (x, y, z). These weak classifiers can be calculated with just sixteen
memory accesses by using precomputed integral cubes, which are natural extensions of
integral images.
During training, we build at each iteration 103 hi weak learners by randomly picking
volume pairs and finding an optimal Ti threshold for each. After running Adaboost,
N = 1000 weak learners are retained in the f classifier of 1. The training samples are
taken from the manual reconstruction of Fig. 2. They consist of filaments at different
orientations and of a certain width. The final classifier responds to filaments of the pre-
defined width, independently of the orientation.
At run time, we apply f on the whole data volume and perform non-maximum sup-
pression by retaining only voxels that maximize it within a 8 × 8 × 20 neighborhood,
such as those shown in Fig. 2. The anisotropy on the neighborhood is due to the low
resolution of the images in the z axis, produced by the point spread function of the
microscope.
3.3 Learning an Edge Appearance Model

The process described above yields Z1 , . . . , ZN , a set of voxels likely, but not guar-
anteed to belong to dendrites. To build an edge appearance model, we compute their
minimum spanning tree. Some of its edges will correspond to filaments and some not.
We therefore create a low dimensional descriptor for the edges, and use it to learn a
gaussian mixture model that we can use to distinguish the two classes of edges.
To obtain an edge descriptor, we first sample the voxel response on a regular lattice
centered around each edge and perform PCA on the resulting set of vectors. For each
edge, we retain the first N PCA components. We construct a N + 1-D edge feature
vector, Ai,j by appending the edge length Li,j to this N -D vector.
This population of N + 1-D vectors is a mixture of edges truly located on filaments,
and of edges located elsewhere. We therefore apply an E-M procedure to derive both a
prior and a Gaussian model for both. The only specificity of this unsupervised training is
to force the covariance between the length and the other N components to be zero, since
the length of an edge is only weakly correlated with its length-normalized appearance.
Hence, given a subgraph G with a population of edges that are both in the dendrite
and elsewhere, this E-M procedure produces two Gaussian models μ0 and μ1 on RN +1
that represent respectively the edges truly on filaments and those elsewhere.
3.4 Building and Pruning the Tree

We can now use the edge appearance model to reconstruct the dendritic tree. To this
end we first compute the maximum spanning tree using as weight for the edges their

8I

8J 8K 8L

8M

í
í !IJ !JK

í
!KL
í í í !JM
(a) (b)
Fig. 3. (a) First two dimensions of the PCA space of the edge appearance models. The Gaussian
models are shown as contour lines. The two small figures at the top represent the projection of
the means in the original lattice. The top-left one represents the model μ1 for filaments, which
appear as a continuous structure. The top-right one represents the non-filament model μ0 . Since,
by construction the endpoints of the edges are local maxima, the intensity there is higher than
elsewhere. (b) Hidden Markov Model used to estimate the probability of a vertex to belong to the
dendritic tree.
likelihood to be part of a dendrite. Nevertheless, the tree obtained with this procedure
is over-complete, spanning vertices that are not part of the dendrites, Fig. 4(b). In order
to eliminate the spurious branches, we use the tree to evaluate the probability that in-
dividual vertices belong to a dendrite, removing those with low probability. We iterate
between the tree reconstruction and vertex elimination until convergence, Fig. 4(c).
We assume that the relationship between the hidden state of the vertices and the
edge appearance vectors can be represented in terms of a hidden Markov model such
(a) (b) (c) (d)
Fig. 4. Building and pruning the tree. (a) Image stack (b) Initial maximum spanning tree. (c) After
convergence of the iterative process. (d) Manually delineated ground truth. Red solid lines denote
edges that are likely to be dendrites due to their appearance. Blue dashed lines represent edges
retained by the minimum spanning tree algorithm to guarantee connectivity.The main filaments
are correctly recovered. Note that our filament detector is sensitive to filaments thinner than the
ones in the ground truth data. This produces the structures in the right part of the images that are
not part of the ground truth data.
as the one depicted by Fig. 3(b). More precisely, we take N (G, i) to be the neighboring
vertices of i in G and assume that
P (Xi | X \i , (Ak,l )(k,l)∈G ) = P (Xi | (Xk )k∈N (G,i) , (Ai,k )k∈N (G,i) ) , (3)
P (Ai,j | X, (Ak,l )(k,l)∈G\(i,j) ) = P (Ai,j | Xi , Xj ) . (4)
Under these assumptions, we are looking for a tree consistent with the edge appearance
model of section 3.3. This means that the labels of its vector of maximum posterior
probabilities x are all 1s. To do so we alternate the building of a tree spanning the
vertices currently labeled 1 and the re-labeling of the vertices to maximize the posterior
probability. The tree we are looking for is a fixed point of this procedure.
Building the Tree. We are looking for maximum likelihood tree that spans all vertices.
Formally:
argmax {log P (T = T | (Ai,j )1≤i,j≤N )}

T
μ1 (Ai,j )
= argmax {log P ((Ai,j )1≤i,j≤N | T = T )} = argmax log .
T T μ0 (Ai,j )
i,j∈T
To this end, we use a slightly modified version of the minimum spanning tree algorithm.
Starting with an empty graph, we add to it at every iteration the edge (i, j) that does not
create a cycle and maximizes
log(μ1 (Ai,j )/μ0 (Ai,j )) .

While this procedure is not guaranteed to find a global optimum, it gives good results
in practice. The main weakness we have to deal with is the over-completeness of the
resulting tree. While it is very rare to miss an important vertex or part of filament, we
have to discard many spurious branches spanned on non-filaments.
Eliminating Unlikely Vertices. From the appearance models μ0 and μ1 learned in sec-
tion 3.3, and the Markovian assumption of Section 3.3, we can estimate for any graph G
the most probable subset of nodes truly on filaments. More specifically, we are looking
for the labeling x of maximum posterior probability given the appearance, defined as
follow
argmax P (X = x | (Ai,j )i,j∈G )

x
Since full optimization is intractable we propose an iterative greedy search. We loop
through each point i, flipping the value of xi if it increases the posterior probability.
This can be seen as a component-wise optimization where the updating rule consists of
fixing all xj , j = i and applying the following update to xi
xi ← argmax P (Xi = x, X \i = x\i | (Ai,j )i,j∈G )

x
= argmax P (Xi = x | X \i = x\i , (Ai,j )i,j∈G ),
x
and under assumptions (3) and (4), we have

P (Xi = x | X \i = x\i , (Ai,j )i,j∈G )

= P (Xj = xj | Xi = x)P (Ai,j | Xi = x, Xj = xj ),
j∈N (G,i)
where P (Xj = 0 | Xi = 0) = P (Xj = 1 | Xi = 1) = 1− and P (Xj = 1 | Xi = 0) =

P (Xj = 0 | Xi = 1) = . is a parameter chosen to be 0.2. P (Ai,j | Xi = x, Xj = xj )
comes from our appearance model, with the assumption that the only true filaments
correspond to Xi = Xj = 1.
The initialization of each xi is done according to the posterior probability of the
edges going through it. If there is an edge with μ1 (ai,j ) > μ0 (ai,j ), then xi = 1.
The termination condition for the loop is that all points are visited without any flip, or
that the number of flips excess ten times the number of points. In practice the second
condition is never met, and only 10−20% of the points flip their hidden variable.
4 Results
In this section we first describe the images we are using. We then compare the dis-
criminative power of our dendrite model against simple grayscale thresholding and the
baseline Hessian based method [6]. Finally, we validate our automated tree reconstruc-
tion results by comparing them against a manual delineation.
4.1 Image Data
Our image database consists of six neuron image stacks, in two of which the dendritic
tree has been manually delineated. We use one of those trees for training and the other
for validation purposes.
The neurons are taken from the somatosensory cortex of Wistar-han rats. The image
stacks are obtained with a standard brightfield microscope. Each image of the stack
shows a slice of the same piece of tissue at a different depth. The tissue is transparent
enough so that these pictures can be acquired by simply changing the focal plane.
Each image stack has an approximate size of 5 ∗ 109 voxels, and is downsampled
to a size of 108 voxels to make the evaluation of the image functional in every voxel
computationally tractable. After down-sampling, each voxel has the same width, height
and depth, of 0.8 μm.
4.2 Image Functional Evaluation
The f classifier of 1 is trained using the manual delineation of Fig. 2. As positive sam-
ples, we retain 500 voxels belonging to filaments of width ranging from two to six
voxels and different orientations. As negative samples, we randomly pick 1000 voxels
that are no closer to a neuron than three times the neuron width and are representative of
the image noise. Since the training set contains filaments of many different orientations,
Adaboost produces a classifier that is orientation independent.
Fig. 2 depicts the candidate dendrite voxels obtained by performing non maxima sup-
pression of images calculated by simply thresholding the original images, computing
1
1 0.9 b
True Positive Rate

0.9 0.8
c
7UXH3RVLWLYH5DWH
0.8 0.7
0.7 (a) 0.6
0.6
0.5
0.5
0.4
0.4 (b) 0.3
0.3
0.2
0.2
Boosting Threshold on the Posterior
Grayscale Threshold 0.1
Trees Iterated
0.1
Hessian (c) 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 True Positives True Negatives
False Positive Rate False Positives False Negatives False Positive Rate
(a) (b) (c)
Fig. 5. (a) ROC curve for all three measures using the validation data of figure 4(d). The boosting
classifier outperforms the baseline hessian method of [6] in noisy brightfield images. (b) Defining
a metric to compare our results against a manual delineation. Top: portion of a manual delineation
in which the vertices are close to each other and the tolerance width painted in red. Middle:
Portion of the tree found by our algorithm at the same location. Bottom: The fully-connected
graph we use to evaluate our edge appearance model and plot the corresponding ROC curves.
(c) ROC curve for the detection of edges on filament obtained by thresholding the individual
estimated likelihood of the edges of the graph of (b). The individual points represent the iterations
of the tree reconstruction algorithm. Two of them are depicted by Fig. 4(b,c). After five iterations
we reach a fixed point, which is our final result.
a Hessian-based measure [6], or computing the output of our classifier at each voxel.
The same procedure is applied in the validation data of Fig 4(d). Considering correct
the vertices that are within 5 μm (6 voxels) of the neuron, we can plot the three ROC
curves of Fig. 5(a) that show that our classifier outperforms the other two.
4.3 Tree Reconstruction

To evaluate the quality of the tree, we compare it against the validation data of Fig. 4(d),
which is represented as a set of connected points. As shown in Fig. 5(a,b), performing
this comparison is non-trivial because in the manual delineation the vertices are close
to each other whereas our algorithm allows for distant points to be connected.
To overcome this difficulty, we introduce a measure of whether an edge linking Xi
to Xj is present in the manual delineation. First, we use the manually labeled points to
construct a volume in which every voxel closer than 5 μm to one such point is assigned
the value 1, and 0 otherwise. We then compute the average value in the straight line
linking Xi and Xj in that volume. If it is greater than a threshold, we consider that the
edge is described by the graph. Here, we take the threshold to be 0.8.
Given this measure, labeling the edges of the tree returned by our algorithm as true
or false positives is straightforward. However, since we also need to compute rates of
true and false negatives to build ROC curves such as the one of Fig. 5, we create graphs
such as the one depicted by Fig. 5(c) in which each vertex is connected to all its nearest
neighbors.
In Fig. 5, we plot a ROC curve obtained by thresholding the likelihood that the edges
of the graph of Fig. 5(c) belong to a neuron based on the edge appearance model of Sec-
tion 3.3. Note that this model is not very discriminative by itself. The individual points
in Fig. 5 represent true and false positive rates for the successive trees built by the pro-
cedure of Section 3.4 and depicted by Fig. 4(b,c,d). As the iterations proceed, the false
Fig. 6. Three additional reconstructions without annotations. Top row: Image stacks. Bottom row:
3D Dendritic tree built by our algorithm. As in Fig. 4, the edges drawn with a solid red lines are
those likely to belong to a dendrite given their appearance. The edges depicted with dashed blue
lines are kept to enforce the tree structure through all the vertices. This figure is best viewed in
color.
positive rate is progressively reduced. Unfortunately, so is the true positive rate as we

loose some of the real dendrite edges. However, the main structure remains and cleaning
up this result by hand is much faster than manually delineating the tree of Fig. 4(e).
In Fig. 6, we show reconstruction results in four more image stacks. Our algorithm
recovers the main dendrites despite their irregularities and the high noise level and,
again, cleaning up this tree is much easier than producing one from scratch. Some of
incorrect edges are also retained because the minimum spanning algorithm enforces
connectivity of all the vertices, even when it is not warranted.
4.4 From Dendrites to Blood Vessels

Since we learn filament models as we go, one of the strengths of our approach is its
generality. To demonstrate it, we ran our algorithm on the retina images of Fig. 7 and 8
without any changes, except for the fact that we replaced the 3D weak classifiers of
Section 3.2 by 2D ones, also based on Haar wavelets. The algorithm learned both a
local blood-vessel model and 2D filament model.
In Fig. 7(b), we evaluate the performance of our boosted classifier against that of
other approaches discussed in [19]. It performs similarly to most of them, but a bit worse
than the best. This can be attributed to the fact that it operates at a single scale and is
optimized to detect large vessels, whereas the others are multiscale. As a consequence,
when we run the full algorithm we obtain the results of Fig. 8 in which the large vessels
are correctly delineated, but some of the small ones are missed. This would be fixed by
training our initial classifier to handle different widths.
0.9
True Positive Fraction

0.8
0.7
0.6
0.5 boosted
2nd observer
0.4
Staal
0.3 Niemeijer
Zana
0.2 Jiang
Martinez-Perez
0.1
Chaudhuri
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
False Positive Fraction
(a) (b)
Fig. 7. (a) Top: image of the retina. Bottom: response of our boosting classifier in this image. (b)
Comparison of our classifier against other algorithms evaluated in the DRIVE database [19]. It
performs similarly to most of them, but worse than algorithms designed specifically to trace blood
vessels in images of the retina. This can be attributed to the fact that our boosted classifier operates
at a single scale and is optimized to detect large vessels, whereas the others are multiscale.
Fig. 8. Retinal trees reconstructed with our method. Top row: original image with the recon-
structed tree overlay. As in Fig. 6, edges likely to belong to filaments are drawn in red, while
edges kept to enforce the tree structure are colored in blue. Bottom row: manually obtained
ground truth. Note that thick filaments are correctly delineated, whereas thin filaments are prone
to errors because our classifier is trained only for the thick ones.
5 Conclusion
We have proposed a novel approach to fully-automated 3D delineation of dendritic

networks in noisy brightfield images based on statistical machine learning techniques
and tree-optimization methods.
By contrast to state-of-the-art methods, we do not postulate a priori models for ei-
ther the dendrite or the edge model between dendrite-like voxels. Instead, we generate
the dendrite measure using discriminative machine learning techniques. We model the
edges as a gaussian mixture model, whose parameters are learned using E-M on neuron-
specific samples.
To demonstrate the generality of the approach, we showed that it also works for
blood vessels in retinal images, without any parameter tuning.
Our current implementation approximates the maximum likelihood dendritic tree
under the previous models by means of minimum spanning trees and markov random
fields. Those techniques are very easy to compute, but tend to produce artifacts. In
future work we will replace them by more general graph optimization techniques.
References
1. Al-Kofahi, K., Lasek, S., Szarowski, D., Pace, C., Nagy, G., Turner, J., Roysam, B.: Rapid
automated three-dimensional tracing of neurons from confocal image stacks. IEEE Transac-
tions on Information Technology in Biomedicine (2002)
2. Tyrrell, J., di Tomaso, E., Fuja, D., Tong, R., Kozak, K., Jain, R., Roysam, B.: Robust 3-
d modeling of vasculature imagery using superellipsoids. Medical Imaging 26(2), 223–237
(2007)
3. Kirbas, C., Quek, F.: Vessel extraction techniques and algorithms: A survey. In: Proceedings
of the Third IEEE Symposium on BioInformatics and BioEngineering, p. 238 (2003)
4. Krissian, K., Kikinis, R., Westin, C.F.: Algorithms for extracting vessel centerlines. Technical
Report 0003, Department of Radiology, Brigham and Women’s Hospital, Harvard Medical
School, Laboratory of Mathematics in Imaging (September 2004)
5. Sato, Y., Nakajima, S., Atsumi, H., Koller, T., Gerig, G., Yoshida, S., Kikinis, R.: 3d multi-
scale line filter for segmentation and visualization of curvilinear structures in medical images.
Medical Image Analysis 2, 143–168 (1998)
6. Frangi, A.F., Niessen, W.J., Vincken, K.L., Viergever, M.A.: Multiscale vessel enhance-
ment filtering. In: Wells, W.M., Colchester, A.C.F., Delp, S.L. (eds.) MICCAI 1998. LNCS,
vol. 1496, pp. 130–137. Springer, Heidelberg (1998)
7. Streekstra, G., van Pelt, J.: Analysis of tubular structures in three-dimensional confocal im-
ages. Network: Computation in Neural Systems 13(3), 381–395 (2002)
8. Meijering, E., Jacob, M., Sarria, J.C.F., Steiner, P., Hirling, H., Unser, M.: Design and valida-
tion of a tool for neurite tracing and analysis in fluorescence microscopy images. Cytometry
Part A 58A(2), 167–176 (2004)
9. Aguet, F., Jacob, M., Unser, M.: Three-dimensional feature detection using optimal steerable
filters. In: Proceedings of the 2005 IEEE International Conference on Image Processing (ICIP
2005), Genova, Italy, September 11-14, 2005, vol. II, pp. 1158–1161 (2005)
10. Dima, A., Scholz, M., Obermayer, K.: Automatic segmentation and skeletonization of neu-
rons from confocal microscopy images based on the 3-d wavelet transform. IEEE Transaction
on Image Processing 7, 790–801 (2002)
11. Schmitt, S., Evers, J.F., Duch, C., Scholz, M., Obermayer, K.: New methods for the
computer-assisted 3d reconstruction of neurons from confocal image stacks. NeuroImage 23,
1283–1298 (2004)
12. Agam, G., Wu, C.: Probabilistic modeling-based vessel enhancement in thoracic ct scans.
In: CVPR 2005: Proceedings of the 2005 IEEE Computer Society Conference on Computer
Vision and Pattern Recognition (CVPR 2005), vol. 2, pp. 684–689. IEEE Computer Society,
Washington (2005)
13. Santamarı́a-Pang, A., Colbert, C.M., Saggau, P., Kakadiaris, I.A.: Automatic centerline ex-
traction of irregular tubular structures using probability volumes from multiphoton imaging.
In: Ayache, N., Ourselin, S., Maeder, A. (eds.) MICCAI 2007, Part II. LNCS, vol. 4792, pp.
14. Al-Kofahi, K.A., Can, A., Lasek, S., Szarowski, D.H., Dowell-Mesfin, N., Shain, W., Turner,
J.N., et al.: Median-based robust algorithms for tracing neurons from noisy confocal micro-
scope images (December 2003)
15. Flasque, N., Desvignes, M., Constans, J., Revenu, M.: Acquisition, segmentation and track-
ing of the cerebral vascular tree on 3d magnetic resonance angiography images. Medical
Image Analysis 5(3), 173–183 (2001)
16. McIntosh, C., Hamarneh, G.: Vessel crawlers: 3d physically-based deformable organisms
for vasculature segmentation and analysis. In: CVPR 2006: Proceedings of the 2006 IEEE
Computer Society Conference on Computer Vision and Pattern Recognition, pp. 1084–1091.
IEEE Computer Society Press, Washington (2006)
17. Szymczak, A., Stillman, A., Tannenbaum, A., Mischaikow, K.: Coronary vessel trees from
3d imagery: a topological approach. Medical Image Analisys (08 2006)
18. Freund, Y., Schapire, R.: Experiments with a New Boosting Algorithm. In: International
Conference on Machine Learning, pp. 148–156. Morgan Kaufmann, San Francisco (1996)
19. Staal, J., Abramoff, M., Niemeijer, M., Viergever, M., van Ginneken, B.: Ridge based vessel
segmentation in color images of the retina. IEEE Transactions on Medical Imaging 23, 501–
509 (2004)
Calibration from Statistical Properties
of the Visual World
Etienne Grossmann1, , José António Gaspar2 , and Francesco Orabona3

1
Tyzx, Inc., Menlo Park, USA
2
ISR, Instituto Superior Técnico, Lisbon, Portugal
3
Idiap Research Institute, Martigny, Switzerland
Abstract. What does a blind entity need in order to determine the geometry of
the set of photocells that it carries through a changing lightfield? In this paper, we
show that very crude knowledge of some statistical properties of the environment
is sufficient for this task.
We show that some dissimilarity measures between pairs of signals produced
by photocells are strongly related to the angular separation between the photo-
cells. Based on real-world data, we model this relation quantitatively, using dis-
similarity measures based on the correlation and conditional entropy. We show
that this model allows to estimate the angular separation from the dissimilarity.
Although the resulting estimators are not very accurate, they maintain their per-
formance throughout different visual environments, suggesting that the model
encodes a very general property of our visual world.
Finally, leveraging this method to estimate angles from signal pairs, we show
how distance geometry techniques allow to recover the complete sensor geometry.
1 Introduction
This paper departs from traditional computer vision by not considering images or image
features as input. Instead, we take signals generated by photocells with unknown ori-
entation and a common center of projection, and explore the information these signals
can shed on the sensor and its surrounding world.
We are particularly interested in determining whether the signals allow to determine
the geometry of the sensor, that is, to calibrate a sensor like the one shown in Figure 1.
Psychological experiments [1] showed that a person wearing distorting glasses for a
few days, after a very confusing and disturbing period, could learn the necessary image
correction to restart interacting effectively with the environment. Can a computer do the
same when, rather than distorted images, it is given the signals produced by individual
photocells? In this situation, it is clear that traditional calibration techniques [2,3] are
out of the question.
Less traditional non-parametric methods that assume a smooth image mapping and
smooth motion [4] can obviously not be applied either. Using controlled-light stimuli

This work was partially supported by TYZX, Inc, by the Portuguese FCT POS_C program that
includes FEDER funds, and by the EU-project URUS FP6-EU-IST-045 062.

Calibration from Statistical Properties of the Visual World 229
Fig. 1. A discrete camera consists of a number of photocells (pixels) that measure the light trav-
eling along pencil of lines
or known calibration, matches could be obtained, allowing to use match-based non-

parametric techniques [5]. In this study however, we wish to exclude known calibration
objects and other controlled stimuli.
Our approach is inspired from the work of Pierce and Kuipers [6], who measure the
dissimilarity, or distance, between sensor elements that are not necessarily light sensors.
The elements are then embedded in a metric space using metric scaling [7], which
also determines the dimension of the space. A relaxation method then improves this
embedding, so that the Euclidean distance between sensor elements better matches the
dissimilarity between the sensor inputs. Getting close to the problem addressed in the
present paper, the authors use this method to reconstitute the geometry of a rectangular
array of visual sensors that scans a fronto-parallel image.
Going further, Olsson et al. [8] use the information distance of [9] as a more appropri-
ate method to measure the distance between visual or other sensor elements. They also
show how visual sensors -the pixels of the camera of a mobile robot- can be mapped
to a plane, either using the method of [6], or their own, that embeds sensor elements
specifically in a square grid.
The works of Olsson et al. and of Pierce and Kuipers are very interesting to com-
puter vision researchers, but they cannot calibrate an arbitrary discrete camera, since the
embedding space is either abstract or fixed to a grid. In both cases, it lacks an explicit
connection to the geometry of the sensor.
Grossmann et al [10] partially fill this gap by showing that the information distance
can be used to estimate the angular separation between pairs of photocells, and from
there, estimate the geometry of a sensor of limited angular radius.
Because the present work exploits statistical properties of the light-field of the world
surrounding a light sensor, it is also related to research on the statistical properties of
real-world images. In that area, a model of image formation is used, but images, rather
than sequences, are studied. That research has put in evidence fundamental properties,
in terms of local, global and spectral statistics, of real-world images, and found ways
to exploit these properties for computer vision tasks, such as classification [11], image
restoration [12] and 3D inference [13]. Although these results are of great interest, they
are not directly applicable in our case, mainly because we lack images.
230 E. Grossmann, J.A. Gaspar, and F. Orabona
Moreover, these statistics are about planar images, which is a hindrance in our case:
first, we do not want to exclude the case of visual sensor elements that are separated
by more than 180 degrees, such as the increasingly popular omnidirectional cameras.
Also, the local statistical properties of perspective images depend of the orientation of
the image plane with respect to the scene, except in special constrained cases such as the
fronto-parallel “leaf world” of Wu et al. [14]. Defining images on the unit sphere thus
appears as a natural way to render image statistics independent of the sensor orientation,
at least with proper assumptions on the surrounding world and/or the motion of the
sensor.
The present article elaborates and improves over our previous work [10]. We in-
novate by showing that the correlation, like the information distance, can be used to
provide geometric information about a sensor. Also, we use a simpler method to model
to relation between angles and signal statistics.
More important, we go much beyond [15] in showing that this model generalizes well
to diverse visual environments, and can thus be considered to be a reliable characteristic
of our visual world. In addition, we show that the presented calibration method performs
much better, for example by allowing to calibrate sensors that cover more than one
hemisphere.
1.1 Proposed Approach
The present work relies on statistical properties of the data streams produced by pairs
of sensor elements that depend only on the angular separation between the photocells.
For example, if the sampled lightfield is a homogeneous random field defined on the
sphere [16], then the covariance between observations depends only on the angular
separation between the sampled points.
This assumption does not hold in general in our anisotropic world, but it does hold,
e.g. if the orientation of the sensor is uniformly distributed amongst all unitary transfor-
mations of the sphere, that is, if the sensor is randomly oriented, so that each photocell
is just as likely to sample the light-field in any direction.
Fig. 2. The process of estimating the geometry of an unknown discrete camera
This assumption of homogeneity -or isotropy- of the sampled lightfield is of great

practical utility, in conjunction with a few other assumptions of good behavior: in this
work, we only use statistics that converge properly (e.g. in probability or more strongly)
when signal lengths tend to infinity.
Perhaps more importantly we are only interested in statistics that have an expectancy
that is a strictly monotonous function of the angular separation of the pair of photocells.
That is, if x, y are two signals (random variables) generated by two photocells separated
by an angle θ, and d (x, y) is the considered statistic, then the expectancy of d (x, y) is
a strictly monotonous function of θ, for 0 ≤ θ ≤ π. The importance of this last point is
that this function can be inverted, resulting in a functional model that links the value of
the statistic to the angle.
The statistic-to-angle graph of such statistics is the a-priori knowledge about the
world that we leverage to estimate the geometry of discrete cameras. In the present
work, we use discrepancy measures based on the correlation or conditional entropy,
defined in Section 3. In Section 4, we show how to build the considered graph.
Having obtained angle estimates, we recover the sensor geometry, in Section 5.1, by
embedding the angles in a sphere. This is done using simple techniques from distance
geometry [17]. Experimental results are presented in Section 5.2. Finally, Section 6
presents some conclusions and possible directions for future research. The calibration
process considered in the present work is outlined in Figure 2. The statistic-to-angle
modeling produces the crucial functional relation used in the third-from right element
of Figure 2.
2 Discrete Camera Model and Simulation

Before entering into the details of our methodology for estimating the sensor geometry,
we define the discrete camera and explain how to simulate it using an omnidirectional
image sensor.
We define a discrete camera [10] as a set of N photocells indexed by i ∈ {1, . . . , N },
pointing in directions Xi ∈ R3 and having a unique center of projection. These photo-
cells acquire along the time t, brightness measurements x (i, t) in the range {0, . . . , 255}.
The directions of the light rays, contrarily to conventional cameras, are not necessarily
organized in a regular grid. Many examples of cameras can be found under these defini-
tions. One example is the linear camera, where all the Xi are co-planar. Another example
is the conventional perspective camera which comprises a rectangular grid of photocells
that are enumerated in our model by a single index i,
, % i%W & -
Xi | Xi ∼ K −1 i/W 1
, 0 ≤ i < HW
where W , H are the image width and height, K is the intrinsic parameters matrix,
% represents the integer modulo operation and . is the lower-rounding operation.
Cameras equipped with fisheye lenses, or having log- polar sensors, can also be modeled
again by setting Xi to represent the directions of the light-rays associated to the image
pixels. In the same vein, omnidirectional cameras having a single projection center,
as the ones represented by the unified projection model [18], also fit in the proposed
model. In this paper we use a calibrated omnidirectional camera to simulate various
discrete cameras.
2.1 Image Sensor

We simulate a discrete camera with known Euclidean geometry by sampling a cal-
ibrated panoramic image with unique projection center at fixed locations. Since the
camera is calibrated, it is straightforward to locate the position (u, v) in the panoramic
image corresponding to the 3D direction X of a photocell that is part of the simu-
lated discrete camera. In the present work, we use bilinear interpolation to measure the
graylevel value at non-integer coordinates (u, v).
Fig. 3. Left: The camera used to sample omnidirectional images (image mirrored). Right: A
calibrated omnidirectional image mapped to a sphere.
Images are acquired by a VStone catadiopric camera consisting of a perspective

camera fitted to a hyperbolic mirror, shown in Figure 3, left. This system is modeled as
single projection center camera [18] with a 360◦ × 210◦ field of view and a ∼ 45◦ blind
spot at the south pole (Fig. 3, right). The mirror occupies a 453 × 453 pixel region of
the image. The angular separation between neighboring pixels in the panoramic image
is usually slightly smaller than 0.5◦ . Also, some mild vignetting occurs, that could be
corrected. Apart for these minor inconveniences, simulating a discrete camera by an
omnidirectional camera presents many advantages: no other specialized hardware is
needed and each omnidirectional image can be used to simulate many discrete camera
“images”, as in Fig. 4, right. With respect to perspective cameras, the available field of
view allows to study very-wide-angle discrete cameras.
3 Distances between Pairs of Signals

In this section, we define the measures of distance between signals, correlation and
information distance, that will later be used to estimate angles.
3.1 Correlation Distance

We call correlation distance between signals x (t) and y (t), 1 ≤ t ≤ T , the quantity
1
dc (x, y) = (1 − C (x, y)) ,
2
where C (x, y) is the correlation between the signals. It is easy to verify that dc (., .) is
a distance.
For the task considered in this paper, it is natural to prefer the correlation distance
over the variance or the (squared) Euclidean distance x − y2 , because both vary
with signal amplitude (and offset, for the latter), whereas dc (., .) is offset- and scale-
invariant.
3.2 Information Distance
Given two random variables x and y (in our case, the values produced by individual
pixels of a discrete camera) taking values in a discrete set {1, . . . , Q}, the information
distance between x and y is [9]:
d (x, y) = H (x|y) + H (y|x) = 2H (x, y) − H (y) − H (x) , (1)
where H (x, y) is the Shannon entropy of the paired random variable (x, y), and H (x)
and H (y) are the entropies of x and y, respectively. It is easy to show that Eq. (1) de-
fines a distance over random variables. This distance is bounded by H (x, y) ≤ log2 Q,
and is conveniently replaced thereafter by the normalized information distance :
dI (x, y) = d (x, y) /H (x, y) , (2)
which is bounded by 1, independently of Q [9].

It should be noted that estimating the information distance is non-trivial: naively re-
placing unknown probabilities px (q) by sample frequencies pˆx (q) =|{t|x (t) = q}| /T ,
where T is the signal length and |.| denotes the set cardinal, yields a biased estimator
Ĥ (x). This estimator has expectancy
1
, - Q − 1 1 − q px (q) 1
E Ĥ = H − + +O . (3)
2T 12T 2 T3
This expression shows the slow convergence rate and strong bias of Ĥ (x). We some-
what alleviate these problems by first, correcting for the first bias term (Q − 1) /2T ,
i.e. applying the Miller-Madow correction; and by re-quantizing the signal to a much
smaller number of bins, Q = 4. Extensive benchmarking in [15] has shown these
choices to be beneficial.
4 Estimating Angular Separation from Inter-signal Distance
As explained earlier, our a-priori knowledge of the world will be encoded in a graph
mapping a measure of discrepancy between two signals, to the angular separation be-
tween the photocells that generated the signals. We now show how to build this graph,
and assess its effectiveness at estimating angles.
For this purpose, we use the 31-pixel planar discrete camera (or “probe”) shown in
Fig. 4, left. This probe design allows to study the effect of angular separations rang-
ing from 0.5 to 180 degrees and each sample provides 465=31(31-1)/2 pixel pairs. In
Sampled Directions
0.8
0.6
0.4
0.2
-1 -0.5 0 0.5 1
Fig. 4. Left: Geometry of a discrete camera consisting of a planar array of thirty one (31) pixels,
spanning 180◦ in the plane. The first two pixels are separated by 0.5◦ , the separation between
consecutive photocells increases geometrically (ratio
1.14), so that the 31st photocell is an-
tipodal with respect to the first. Right: Two instances of the linear discrete camera, inserted in an
omnidirectional image. Pixels locations are indicated by small crosses connected by white lines.
the “tighter” part of the discrete camera layout, there exists a slight linear dependence
between the values of consecutive pixels due to aliasing.
The camera is hand-held and undergoes “random” general rotation and translation,
according to the author´s whim, while remaining near the middle of the room, at 1.0 to
1.8 meters from the ground. We acquired three sequences consecutively, in very similar
conditions and joined them in a single sequence totaling 1359 images, i.e. approxi-
mately 5 minutes of video at ˜4.5 frames per second.
To simulate the discrete camera, we randomly choose an orientation (i.e. half a great
circle) such that all pixels of the discrete camera fall in the field of view of the panoramic
camera. Figure 4 shows two such choices of orientations. For each choice of orientation,
we produce a sequence of 31 samples x (i, t), 1 ≤ i ≤ 31, 1 ≤ t ≤ 1359, where each
x (i, t) ∈ {0, . . . , 255}. Choosing 100 different orientations, we obtain 100 discrete
sensors and 100 arrays of data xn (i, t), 1 ≤ n ≤ 100. Appending these arrays we
obtain 31 signals x (i, t) of length to 135900.
We then compute, for each pair of pixels (indices) 1 ≤ i, j ≤ 31, the correlation
and information distances, dc (i, j) and dI (i, j). Joining to these the known angular
separations θi,j , we obtain a set of pairs (θi,j , d (i, j)), 1 ≤ i, j ≤ 31.
From this dataset, we build a constant by parts model of the expectancy of the dis-
tance, knowing the angle. For the correlation distance, we limit the abscissa to values
in [0, 1/2]. After verifying and, if needed enforcing, the monotonicity of this model,
we invert it, obtaining a graph of angles as a function of (correlation or information)
distances. Strict monotonicity has to be enforced for the correlation-based data, owing
to the relatively small number of data points used for each quantized angle.
Figure 5 shows the resulting graphs. This figure shows one of the major issues that
appear when estimating the angular separation between pixels from the correlation or
information distance: the graphs become very steep for large values of the distance,
indicating that small changes of the distance result in large changes in the estimated an-
gle. On the other hand, for small distance values, the curves are much flatter, suggesting
Correlation-Distance-to-Angle Model Information-Distance-to-Angle Model

120 120
100 100
Predicted Angle
Predicted Angle
80 80
60 60
40 40
20 20
0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Correlation Distance Information Distance
Fig. 5. Models relating correlation (left) or information distance (right) to angular separation
between photocells. These models were build from simulated signals produced by the linear
probe of Fig. 4, left. Signals of length T = 135900, acquired indoors were used.
that small angles can be determined with greater accuracy. Both trends are particularly
true for the information distance.
4.1 Experimental Validation

We now assess how well angles can be estimated from the graphs obtained in the pre-
vious section. For this purpose, we use 100 sets of 31 signals xn (i, t), 1 ≤ n ≤ 100,
1 ≤ i ≤ 31, 1 ≤ t ≤ 1359 acquired in the same conditions as above. We compute the
correlation and information distances of pairs of signals dc (n, i, j) and dI (n, i, j) and,
using the models in Fig. 5, angular estimates θ̂c (n, i, j) and θ̂I (n, i, j).
Figure 6 shows the precision and accuracy of the estimated angles. This figure shows
that the estimated angles are fairly accurate for angular separations smaller than 5◦ , but
degrades sharply for greater values. As could be expected from our comments at the
beginning of the section, the curves confirm that the information distance yields better
estimates of small angles, while correlation distance does best (but still not very well)
for larger angles.
We now turn to the generalization ability of the models in Fig. 5. For this purpose,
we use 100 31-uplets of signals of length 2349, taken from an out- and indoor sequence,
four images of which are shown in Fig. 7. In this sequence, and contrarily to the previous
sequence, the camera remains mostly horizontal. Also, the scene is usually farther away
and more textured. A lot of saturation is also apparent.
Following the previous procedure, we estimate angles from these new signals and
show the precision and accuracy statistics in Figure 8.
The striking resemblance between Figures 8 and 6 indicates that the models in Fig. 5
generalize pretty well to outdoors scenes. We surmise that the fact that the correlation
distance yields more accurate estimates outdoors than indoors is due to the extra tex-
ture, which increases the correlation distance for small angles, and corrects the bias in
angular estimates observed near the origin of the top left curve of Fig. 6.
Angles from Correlation Distance Angles from Information Distance

180 180
150 150
Estimated Angular Separation

120 120
90 90
60 60
30 30
20 20
10 10
5 5
1 1
0 0
0 1 5 10 20 30 60 90 120 150 180 0 1 5 10 20 30 60 90 120 150 180
True Angular Separation True Angular Separation
Accuracy of Angles From Corr. Distance Accuracy of Angles From Info. Distance
90 90
80 80
Mean Absolute Error (deg)

70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180
True Angle (deg) True Angle (deg)
Fig. 6. Precision and accuracy of angles estimated from correlation (left) or information distance
(right). The boxplots at the top show the 5th percentile, first quartile, median, third quartile and
95th percentile of the estimated angles, plotted against the true angles. The bottom curves show
the mean absolute error in the estimated angles. These statistics were generated from 100 planar
probes (Fig. 4, left) and signals of length T = 1359. The angles were estimated using the models
of Fig. 5. The signals were acquired in the same conditions as those used to build the models.
Fig. 7. Four images from a sequence of 2349 images acquired indoors and outdoors at approxi-
mately 4.5FPS
5 Calibrating a Discrete Camera
Having seen the qualities and shortcomings of the proposed angle estimators, we now
show how to use them to calibrate a discrete camera.
To stress the generalization ability of the angle estimators, all the reconstructions
produced by the above method are obtained from the in- and outdoors sequence of
Fig. 7, rather than from the indoors sequence used to build the distance-to-angle models.

180 180
150 150

120 120
90 90
60 60
30 30
20 20
10 10
5 5
1 1
0 0
0 1 5 10 20 30 60 90 120 150 180 0 1 5 10 20 30 60 90 120 150 180
90 90
80 80

70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180
Fig. 8. Precision and accuracy of angles estimated in the same conditions as in Fig. 6, except that
signals extracted from an indoors-and-outdoors sequence (Fig. 7) were used. These figures show
that the models in Fig. 5 generalize fairly well to signals produced in conditions different from
that in which the models were produced. In particular, the angles estimated from the correlation
distance are improved w.r.t. those of Fig. 6 (see text).
5.1 Embedding Points in the Sphere

The last step we take to calibrate a discrete camera requires solving the problem:
Problem 1) Spherical embedding problem: Given angle estimates θij , 1 ≤ i, j ≤

N , find points Xi on the unit sphere, separated by angles approximately equal to
θij , i.e. Xi Xj cos θij , for all i, j.
This problem can be reduced to the classical problem of distance geometry [17]:
Problem 2) Euclidean embedding problem: Given distance estimates Dij , 1 ≤ i, j

≤ N , find points Yi in a metric vector space, such that, for all i, j, Yi − Yj Dij
.
Indeed, by defining an extra point Y0 = (0, 0, 0), and distances Dij = 2 − 2 cos θij
for i, j = 0 and Doi = 1, the mapping of the first problem to the second is imme-
diate. Solutions to both problems (with exact equality, rather than approximate) were
published in 1935 [19]1 . Schoenberg´s Theorem 2 [19] states that if the matrix C with
terms Cij = cos θij is positive semidefinite with rank r ≥ 1, then there exist points on
1
Schoenberg cites previous work by Klanfer and by Menger, to which we did have access.

180 180
150 150

120 120
90 90
60 60
30 30
20 20
10 10
5 5
1 1
0 0
0 1 5 10 20 30 60 90 120 150 180 0 1 5 10 20 30 60 90 120 150 180
90 90
80 80

70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180
Fig. 9. Precision and accuracy of angles estimated in the same conditions as in Fig. 8, except that
the planar probes are constrained to remain approximately horizontal. These figures show that
the models in Fig. 5 are usable even if the isotropy assumption of the moving entity is not valid.
the unit (r − 1) −dimensional sphere that verify Xi Xj = Cij for all i, j. This result
directly suggests the following method for embedding points in the 2-sphere:
1. Build the matrix C with terms Cij = cos θij , 1 ≤ i, j ≤ N .

2. Compute, using the SVD decomposition, the rank-3 approximation C̃ = U U of
C, where U is N × 3.
3. Define Xi = (Ui1 , Ui2 , Ui3 ) / (Ui1 , Ui2 , Ui3 ).
One should note that this very simple algorithm is not optimal in many ways. In par-
ticular, it does not take into account that the error in the angles θij is greater in some
cases than in others. It is easy to verify that the the problem is not directly tractable by
variable-error factorization methods used in computer vision.
Noting that the error in the estimated angles is approximately proportional to the
actual angle suggests an embedding method that weighs less heavily large angular esti-
mates. One such method is Sammon´s algorithm [20], which we adapt and modify for
the purpose of spherical embedding from our noisy data. In this paper, we minimize the
sum
, -
2 max 0, 1−C 1
− 1
if Cij = 1
wi,j Xi Xj − Cij , where wij = 1
ij 1−Co
i,j η otherwise.
Fig. 10. Calibrations of two different sensors covering more than one hemisphere. On the left,
a band-like sensor consisting of 85 photocells, calibrated from correlations (estimated: smaller,
true: bigger). On the right, a discrete camera covering more than 180×360◦ , of 168 photocells,
calibrated from the information distance (estimated: smaller, true: bigger). Each ball represents a
photocell except the big black balls, representing the optical center.
To reflect the fact that big angles are less well estimated, we set C0 = 0.9, so that
estimates greater than acos (0.9) 25◦ be ignored. The other parameter, η is set to 1,
allowing the points Xi to stray a little bit away from the unit sphere. Our
implementation is inspired by the second-order iterative method of Cawley and
Talbot (http://theoval.sys.uea.ac.uk/~gcc/matlab/default.html). For
initialization, we use an adaptation of [21] to the spherical metric embedding problem,
which will be described in detail elsewhere.
5.2 Sensor Calibration
We now evaluate the results of this embedding algorithm on data produced by the angle-
estimating method of Sec. 4. For this purpose, we produce sequences of pixel signals
in the same conditions as previously, using the outdoors and indoors sequence shown
in Figure 7, except that the sensor shape is different. The information and correlation
distances between pixels is then estimated from these signals, the angular separation
between the pixels is estimated using Sec. 4, and the embedding method of Sec. 5.1 is
applied to these angle estimates.
Figure 10 shows the results of our calibration method on sensors covering more than
a hemisphere, which thus cannot be embedded in a plane without significant distortion.
It should be noted that, although the true sensor is each time more than hemispheric,
the estimated calibration is in both cases smaller. This shrinkage is a known effect of
some embedding algorithms, which we could attempt to correct.
Figure 11 shows how our method applies to signals produced by a different sensor
from the one used to build the distance-to-angle models, namely an Olympus Stylus 300
camera. An 8-by-8 square grid pixels spanning 34 degrees was sampled along a 22822
image sequence taken indoors and outdoors. From this sequence, the estimated angles
were generally greater than the true angles, which explains the absence of shrinkage.
The higher angle estimates were possibly due to higher texture contents of the sequence.
The estimated angles were also fairly noisy, possibly due to the sequence length, and
we surmise that longer sequences would yield better results.
Reconstruction using correlations (New sensor). Reconstruction using information distances (New sensor).
Signal length: 22822 Signal length: 22822
0.3 0.3
Estimated Estimated
True True
0.2 0.2
0.1 0.1
0 0
-0.1 -0.1
-0.2 -0.2
-0.3 -0.3
-0.2 -0.1 0 0.1 0.2 -0.2 -0.1 0 0.1 0.2
Correlation-Based Information Distance-Based
Fig. 11. Reconstructed and true pixel layouts of a discrete camera consisting of photocells lying
on a rectangular grid. The sensor used differs from that with which the models of Fig 5 were
built. The reconstructions are obtained by first estimating the pairwise angular distances, then
embedding the angles in the sphere (see text). For visualization, the reconstructions are aligned
by the usual procrustes method, mapped to the plane by projective mapping with unit focal length.
Added line segments show the true pixel neighborhood relations. The left plot is obtained from
the correlation distance, and the right from the information distance.
These results represent typical results that researchers reproducing our method may
encounter. Results from other experiments will be presented elsewhere.
6 Discussion
In this paper, we have shown that simple models exist that relate signal discrepancy
to angular separation, and are valid in indoors and outdoors scenes. This suggests the
existence of near-universal properties of our visual world, in line with other work show-
ing statistical properties of natural images. Contrarily to previous works, we consider
statistics of the lightfield taken as a function defined on the sphere, rather than the plane,
a choice that allows us to consider fields of view greater than 180 degrees.
We addressed the problem of determining the geometry of a set of photocells in a
very general setting. We have confirmed that a discrete camera can be calibrated to a
large extent, using just two pieces of data: a table relating signal distances to angles;
and a long enough signal produced by the camera.
The presented results are both superior and of a much wider scope than that of [15]:
we have shown that it is necessary neither to strictly enforce the assumptions that the
camera directs each pixel uniformly in all directions, nor that statistically similar en-
vironments be used to build the statistic-to-angle table and to calibrate the discrete
camera. This flexibility reinforces the impression that models such as those shown in
Figure 5 have a more general validity than the context of calibration.
We showed also that angle estimators based on correlation and information distance
(entropy) have different performance characteristics. It would be very interesting to
apply machine learning techniques to leverage the power of many such weak estimators.
Finally a more curious question is worth asking in the future: can the problem of
angle estimation be altogether bypassed in a geometrically meaningful calibration pro-
cedure? Embedding methods based on rank or connectivity [17,22], e.g. correlation or
information distance, suggest that this is possible.
References
1. Kohler, I.: Experiments with goggles. Scientific American 206, 62–72 (1962)
2. Tsai, R.: An efficient and accurate camera calibration technique for 3D machine vision. In:
IEEE Conf. on Computer Vision and Pattern Recognition (1986)
3. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cambridge Uni-
versity Press, Cambridge (2000)
4. Nistér, D., Stewenius, H., Grossmann, E.: Non-parametric self-calibration. In: Proc. ICCV
(2005)
5. Ramalingam, S., Sturm, P., Lodha, S.: Towards complete generic camera calibration. In:
Proc. CVPR, vol. 1, pp. 1093–1098 (2005)
6. Pierce, D., Kuipers, B.: Map learning with uninterpreted sensors and effectors. Artificial
Intelligence Journal 92(169–229) (1997)
7. Krzanowski, W.J.: Principles of Multivariate Analysis: A User’s Perspective. Statistical Sci-
ence Series. Clarendon Press (1988)
8. Olsson, L., Nehaniv, C.L., Polani, D.: Sensory channel grouping and structure from uninter-
preted sensor data. In: NASA/NoD Conference on Evolvable Hardware (2004)
9. Crutchfield, J.P.: Information and its metric. In: Lam, L., Morris, H.C. (eds.) Nonlinear Struc-
tures in Physical Systems–Pattern Formation, Chaos and Waves, pp. 119–130. Springer, Hei-
delberg (1990)
10. Grossmann, E., Orabona, F., Gaspar, J.A.: Discrete camera calibration from the informa-
tion distance between pixel streams. In: Proc. Workshop on Omnidirectional Vision, Camera
Networks and Non-classical Cameras, OMNIVIS (2007)
11. Torralba, A., Oliva, A.: Statistics of natural image categories. Network: Computation in Neu-
ral Systems 14, 391–412 (2003)
12. Freeman, W.T., Pasztor, E.C., Carmichael, O.T.: Learning low-level vision. International
13. Potetz, B., Lee, T.S.: Scaling laws in natural scenes and the inference of 3d shape. In: NIPS –
Advances in Neural Information Processing Systems, pp. 1089–1096. MIT Press, Cambridge
(2006)
14. Wu, Y.N., Zhu, S.C., Guo, C.E.: From information scaling of natural images to regimes of
statistical models. Technical Report 2004010111, Department of Statistics, UCLA (2004)
15. Grossmann, E., Gaspar, J.A., Orabona, F.: Discrete camera calibration from pixel streams.
In: Computer Vision and Image Understanding (submitted, 2008)
16. Roy, R.: Spectral analysis for a random process on the sphere. Annals of the institute of
statistical mathematics 28(1) (1976)
17. Dattorro, J.: Convex Optimization & Euclidean Distance Geometry. Meboo Publishing
(2005)
18. Geyer, C., Daniilidis, K.: A unifying theory for central panoramic systems and practical
applications. In: Vernon, D. (ed.) ECCV 2000. LNCS, vol. 1843, pp. 445–461. Springer,
Heidelberg (2000)
19. Schoenberg, I.J.: Remarks to Maurice Fréchet’s article “Sur la définition axiomatique d’une
classe d’espaces distanciés vectoriellement applicable sur l’espace de Hilbert”. Annals of
Mathematics 36(3), 724–732 (1935)
20. Sammon, J.W.J.: A nonlinear mapping for data structure analysis. IEEE Transactions on
Computers C-18, 401–409 (1969)
21. Lee, R.C.T., Slagle, J.R., Blum, H.: A triangulation method for the sequential mapping of
points from n-space to two-space. IEEE Trans. Computers 26(3), 288–292 (1977)
22. Shang, Y., Ruml, W., Zhang, Y., Fromherz, M.P.J.: Localization from mere connectivity. In:
MobiHoc 2003: Proc. ACM Intl. Symp. on Mobile Ad Hoc Networking & Computing, pp.
201–212. ACM Press, New York (2003)
Regular Texture Analysis as Statistical Model
Selection
Junwei Han, Stephen J. McKenna, and Ruixuan Wang
School of Computing, University of Dundee, Dundee DD1 4HN, UK

{jeffhan,stephen,ruixuanwang}@computing.dundee.ac.uk
http://www.computing.dundee.ac.uk
Abstract. An approach to the analysis of images of regular texture is

proposed in which lattice hypotheses are used to define statistical mod-
els. These models are then compared in terms of their ability to explain
the image. A method based on this approach is described in which lattice
hypotheses are generated using analysis of peaks in the image autocor-
relation function, statistical models are based on Gaussian or Gaussian
mixture clusters, and model comparison is performed using the marginal
likelihood as approximated by the Bayes Information Criterion (BIC).
Experiments on public domain regular texture images and a commer-
cial textile image archive demonstrate substantially improved accuracy
compared to two competing methods. The method is also used for clas-
sification of texture images as regular or irregular. An application to
thumbnail image extraction is discussed.
1 Introduction
Regular texture can be modelled as consisting of repeated texture elements, or

texels. The texels tesselate (or tile) the image (or more generally a surface). Here
we consider so-called wallpaper patterns. Wallpaper patterns can be classified
into 17 groups depending on their symmetry [1]. Translationally symmetric reg-
ular textures can always be generated by a pair of shortest vectors (two linearly
independent directions), t1 and t2 , that define the size, shape and orientation
(but not the position) of the texel and the lattice which the texel generates. The
lattice topology is always then quadrilateral. Geometric deformations, varying
illumination, varying physical characteristics of the textured surface, and sensor
noise all result in images of textured patterns exhibiting approximately regu-
lar, as opposed to exactly regular, texture. This paper considers the problem of
automatically inferring texels and lattice structures from images of planar, ap-
proximately regular textures viewed under orthographic projection. While this
might at first seem restrictive, this problem is, as will become apparent, far from
solved. There exists no fully automatic and robust algorithm to the best of the
authors’ knowledge. Furthermore, solutions will find application, for example in
analysis, retrieval and restoration of images of printed textiles, wallpaper and
tile designs.

Regular Texture Analysis as Statistical Model Selection 243
1.1 Related Work
Extraction of periodicity plays an important role in understanding texture and

serves as a key component in texture recognition [2], synthesis [3] and segmenta-
tion [4]. Previous work proposed for texel and lattice extraction can be grouped
broadly into two categories: the local feature-based approach [5,6,7,8,9,10,11]
and the global structure-based approach [1,12,13,14,15,16]. All texture analysis
is necessarily both local and global. The categorisation is in terms of the compu-
tational approach: whether it starts by identifying local features and proceeds to
analyse global structure, or starts with a global analysis and proceeds by refining
estimates of local structure.
The local feature-based approach starts by identifying a number of texel can-
didates. Matching based on visual similarity between these potential texels and
their neighbours is then performed. Successful matching leads to the connection
of texels into a lattice structure. The approach iterates until no more new texels
are found. Methods vary in the way they initialise texel candidates and in the
parametric models used to cope with geometric and photometric variation. Lin
et al. [6] asked users to provide an initial texel. Interest points and edges have
been used to generate texel candidates automatically [7,8,9]. However, Hays et
al. [5] pointed out that interest points often fail to find texel locations and in-
stead initialized by combining interest points and normalized cross correlation
patches. Affine models have been adopted to deal with local variation among
texels [7,10,11]. Global projective transformation models have also been used,
taking advantage of the spatial arrangement of texels [8,9]. Hays et al. [5] formu-
lated lattice detection as a texel correspondence problem and performed texel
matching based on visual similarity and geometric consistency. Lin et al. [6]
proposed a Markov random field model with a lattice structure to model global
topological relationships among texels and an image observation model able to
handle local variations.
The global structure-based approach [1,12,13,14,15,16] tries to extract texels
using methods that emphasise the idea of periodic patterns as global processes.
Starovoitov et al. [16] used features derived from cooccurrence matrices to ex-
tract texels. Charalampidis et al. [15] used a Fourier transform and made use of
peaks corresponding to fundamental frequencies to identify texels. The autocor-
relation (AC) function is generally more robust than the Fourier transform for
the task of texel extraction especially in cases in which a regular texture image
contains only a few texel repetitions [1,12]. Peaks in the AC function of a regular
texture image can identify the shape and arrangement of texels. Chetverikov [13]
developed a regularity measure by means of finding the maximum over all di-
rections on the AC function. Leu [14] used the several highest peaks in the AC
function computed on the gradient field of the image to capture translation vec-
tors. A promising approach was presented by Lin et al. [12] in which salient peaks
were identified using Gaussian filters to iteratively smooth the AC function. The
generalized Hough transform was then applied to find translation vectors, t1 and
t2 . Liu et al. [1] highlighted the fact that spurious peaks often result in incorrect
lattice vectors. Therefore, they proposed a “region of dominance” operator to
244 J. Han, S.J. McKenna, and R. Wang
select a list of dominant peaks. The translation vectors were estimated based on
these dominant peaks. However, the important problem of how to determine the
number of dominant peaks was not addressed. Whilst it is usually relatively easy
for a human to select an appropriate subset of peaks, automating this process is
difficult. Fig. 1 shows three different texels obtained similarly to Lin et al. [12]
from the same image by using different numbers of peaks. The peaks were ob-
tained using the region of dominance method [1]. Whilst using only the first ten
peaks can result in success, the method is rather sensitive to this choice.
(a) (b) (c)
Fig. 1. Texels obtained using (a) ten, (b) forty, and (c) seventy dominant peaks in the
autocorrelation function. The peak locations are marked with white dots.
Available local feature-based methods can be effective under significant tex-

ture surface deformation and are more suited to such situations. However, they
require texels that can be identified based on local features (such as corners) and
perform matching between individual texels. Therefore they often fail to detect
larger, non-homogeneous texels. Fig. 2 shows examples of such failures.
Global structure-based methods are suitable for textures that do not exhibit
large geometric deformation and often successfully identify larger texels with
more complicated appearances. However, existing methods have free parameters
Fig. 2. Two examples of a local feature-based method [5] extracting incorrect lattices
for which a fixed value that works on a wide range of images can often not be
found. Methods based on finding peaks in an AC function often yield many un-
reliable peaks and the number which are reliable can vary dramatically between
images. This serious drawback currently makes these methods difficult to apply
to large image collections.
1.2 Contributions
We propose a novel model comparison framework to test texel hypotheses and

find the optimal one. Hypotheses can be constructed using existing methods ac-
cording to different subsets of AC peaks by varying the number of peaks used.
A statistical model is defined for each lattice hypothesis. The most probable hy-
pothesis given the image observation will be selected. The design of the statistical
model takes account of photometric and (to a lesser extent) geometric variations
between texels. Hence, our method is robust and completely automatic.
The contributions of this paper can be summarized as follows. (i) A Bayesian
model comparison framework is proposed to extract texels from regular texture
images based on statistical models defined to handle variations between tex-
els. (ii) Lattice comparison is also used to classify texture images as regular or
irregular. (iii) Empirical comparison of the proposed method with two exist-
ing methods is performed on a challenging regular texture image database. (iv)
The method is applied to generate smart thumbnails for an image browsing and
retrieval system.
The rest of this paper is organized as follows. Section 2 presents the Bayesian
model comparison framework. Section 3 describes details of lattice model compar-
ison. Section 4 describes the method used in our experiments for generating lattice
hypotheses. Experimental results are given in Section 5. An application in which
the proposed method is used to generate smart thumbnails for regular texture im-
ages is reported in Section 6. Finally, conclusions are drawn in Section 7.
2 Bayesian Model Comparison Framework
Our approach is to formulate texel hypotheses as statistical models and then

compare these models given the image data. It is not sufficient for a model to be
able to fit the data well. The best texel hypothesis under this criterion would be
the image itself whereas our purpose is to extract the smallest texture element.
Therefore, overfitting must be guarded against by penalising model complexity.
Texel hypothesis comparison can be regarded as a typical model comparison
problem for unsupervised statistical modelling of data. Such a problem can be
formulated as Bayesian model comparison which naturally penalises complexity
(Occam’s razor).
Let I = {x1 , x2 , . . . , xN } be an image with N pixels. Here, xn , 1 ≤ n ≤ N
is the intensity of the nth pixel. Let H ≡ (t1 , t2 ) denote a texel hypothesis for
I, Hk the k th in a set of hypotheses, and Mk a statistical model defined based
on Hk with parameters θk . Texel extraction can be formulated as choosing the
most probable texel hypothesis given the image. According to Bayes’ theorem,
the posterior probability is proportional to the likelihood of the hypothesis times
a prior:
p(I|Hk )p(Hk )
p(Hk |I) = ∝ p(I|Hk )p(Hk ) (1)
p(I)
In the absence of prior knowledge favouring any of the texel hypotheses, the
(improper) prior is taken to be uniform. For each Hk , we define a unique Mk
deterministically so p(Mk |Hk ) is a delta function. Hence,

p(Hk |I) ∝ p(I|Mk ) = p(I|θk , Mk )p(θk |Mk )dθk (2)
Texel hypotheses can be compared by comparing the marginal likelihoods,

p(I|Mk ), for their models. Here p(I|θk , Mk ) is the probability density function
of the image data given the model Mk and its parameters θk , and p(θk |Mk ) is
the prior probability density function of parameters θk given the model Mk .
The integral in Equation (2) can only be computed analytically in certain
cases such as exponential likelihoods with conjugate priors. Otherwise, approxi-
mations can be obtained using sampling methods, for example. While it would
be interesting to explore these alternatives in future work, this paper uses the
Bayes Information Criterion (BIC) as a readily computable approximation. BIC
approximates the marginal likelihood integral via Laplace’s method and the
reader is referred to the papers by Schwarz [17] and Raftery [18] for full details
of its derivation. Given a maximum likelihood parameter estimate, θ̂, we have
d d 1
log 2π − log N − log |i| + O(N −1/2 )
log p(I|M ) ≈ log p(I|θ̂, M ) + log p(θ̂) +
2 2 2
(3)
where d is the number of parameters and i is the expected Fisher information
matrix for one observation. The subscript k has been dropped here for clarity.
The term log p(I|θ̂, M ) is of order O(N ), (d/2) log N is of order O(log N ), and
the remaining terms are of order O(1) or less. The log marginal likelihood can
be approximated by removing all terms of order O(1) or less. The BIC for the
model is then
BIC(M ) = − log p(I|θ̂, M ) + (d/2) log N ≈ − log p(I|Mk ) (4)
The first term can be interpreted as an error of fit to the data while the second
term penalises model complexity.
The proposed approach to regular texture analysis involves (i) generation of
multiple texel hypotheses, and (ii) comparison of hypotheses based on statistical
models. The hypothesis with the model that has the largest marginal likelihood
is selected. Using the BIC approximation, hypothesis Hk̂ is selected where,
k̂ = arg max{p(Hk |I)} = arg min{BIC(Mk )} (5)

k k
This method can also be used to classify textures as regular or irregular. If a

‘good’ lattice can be detected in an image then it should be classified as regular.
The proposed lattice comparison framework can be adopted for this purpose
by comparing the most probable lattice found with a reference hypothesis in
which the entire image is a single ‘texel’. If the reference hypothesis has a higher
BIC value then the image is classified as regular. Otherwise, it is classified as
irregular, i.e.
BIC(MR ) ≤ BIC(Mk̂ ) Irregular texture

BIC(MR ) > BIC(Mk̂ ) Regular texture (6)
where MR refers to the model corresponding to the reference lattice and Mk̂ is
the best lattice hypothesis selected by Equation (5).
3 Lattice Models
The lattice model should be able to account for both regularity from periodic
arrangement and statistical photometric and geometric variability. Let us first
suppose a regular texture image I with N pixels x1 , x2 , . . . , xN , and a hypothesis
H with Q pixels per texel. Based on H, each pixel of the image is assigned to
one of Q positions on the texel according to the lattice structure. Thus, the N
pixels are partitioned into Q disjoint sets, or clusters. If we choose to assume
that the N pixels are independent given the model, we have,

N
Q
p(I|M ) = p(xn |M ) = p(xn |M ) (7)
n=1 q=1 n:f (n,H)=q
where f (n, H) ∈ {1, . . . , Q} maps n to its corresponding index in the texel. Fig. 3
illustrates this assigment of pixels to clusters.
Fig. 3. An example of cluster allocation according to a texel hypothesis, H ≡ (t1 , t2 ).

The value of f (n, H) is the same for each of the highlighted pixels. There are Q pixels
in each parallelogram.
Modelling each of the Q clusters as Gaussian with fixed variance gives:

Q
BIC(M ) = (Q/2) log N − log p(xn |μˆq , σ 2 ) (8)
q=1 n:f (n,H)=q
1
Q
= (Q/2) log N + C1 + (xn − μˆq )2 (9)
2σ 2 q=1
n:f (n,H)=q
where C1 is a constant that depends on σ 2 , and μˆq is a maximum likelihood

estimate of the mean of the q th cluster.
Alternatively, a more heavy-tailed distribution can be used for each cluster.
This might better model outliers due to physical imperfections in the texture
surface and variations due to small geometric deformations. For example, a clus-
ter can be modelled as a mixture of two Gaussians with the same mean but
different variances, (σ12 , σ22 ), and a mixing weight, π1 that places greater weight
on the low variance Gaussian. In that case,

Q
BIC(M ) = − log p(xn |μˆq , σ12 , σ22 , π1 ) + (Q/2) log N (10)
q=1 n:f (n,H)=q
= (Q/2) log N + C2 (11)

Q π1 −(xn − μˆq )2 1 − π1 −(xn − μˆq )2
− log( exp + exp )
q=1 n:f (n,H)=q
σ1 2σ12 σ2 2σ22
where C2 is a constant.
4 Lattice Hypothesis Generation
In principle, there is an unlimited number of lattice hypotheses. However, prob-

ability density will be highly concentrated at multiple peaks in the hypothesis
space. The posterior distribution can therefore be well represented by only con-
sidering a, typically small, number of hypotheses at these peaks. In the maxi-
mum a posteriori setting adopted here, the approach taken is to identify multiple
hypotheses in a data-driven manner and then compare these hypotheses using
BIC. The approach is general in that any algorithms that generate a variety of
reasonable hypotheses can be used.
In the experiments reported here, aspects of the methods of Lin et al. [12] and
Liu et al. [1] were combined to generate hypotheses. Peaks in AC functions are
associated with texture periodicity but automatically deciding which peaks can
characterize the arrangement of texels is problematic and has not been properly
addressed in the literature [1,12,13,14]. In particular, changing the number of
peaks considered can result in different lattice hypotheses. Since the total number
of peaks is limited, we can only obtain a limited number of hypotheses.
Given a grey-scale image I(x, y), 1 ≤ x ≤ L, 1 ≤ y ≤ W where L and W are

image height and width, its AC function can be computed as follows:
L W
i=1 j=1 I(i, j)I(i + x, j + y)
AC(x, y) = L W 2 (12)
i=1 j=1 I (i, j)
Applying the fast Fourier transform (FFT) to calculate the AC function is a

more efficient alternative.
AC(x, y) = F −1 [F [I(x, y)]∗ F [I(x, y)]] (13)

−1
where F and F denote FFT and inverse FFT, respectively.
Lin et al. [12] used iterative smoothing with Gaussian filters to obtain salient
peaks. However, Liu et al. [1] advised to take into account the spatial rela-
tionships among peaks and used a “region of dominance” operator. The basic
idea behind this operator is that peaks that dominate large regions of the AC
function are more perceptually important. In this paper, we combine these two
algorithms. First, we apply Gaussian filters to iteratively smooth the AC func-
tion. Then, salient peaks obtained from the first stage are ranked according to
their dominance. The most highly ranked peaks are selected as input for lattice
hypothesis construction using a Hough transform [12]. The number of peaks in
the rank-ordered list to use was varied in order to generate multiple hypotheses.
Typically a few tens of the generated hypotheses will be distinct.
5 Experiments
A dataset of 103 regular texture images was used for evaluation, comprising 68
images of printed textiles from a commercial archive and 35 images taken from
three public domain databases (the Wikipedia Wallpaper Groups page, a Corel
database, and the CMU near regular texture database). These images ranged in
size from 352 × 302 pixels to 2648 × 1372 pixels. The number of texel repeats
per image ranged from 5 to a few hundreds. This data set includes images that
are challenging because of (i) appearance variations among texels, (ii) small
geometric deformations, (iii) texels that are not distinctive from the background
and are large non-homogeneous regions, (iv) occluding labels, and (v) stains,
wear and tear in some of the textile images.
Systematic evaluations of lattice extraction are lacking in the literature. We
compared the proposed method with two previously published algorithms. Two
volunteers (one male and one female) qualitatively scored and rank ordered
the algorithms. In cases of disagreement, they were forced to reach agreement
through discussion. (Disagreement happened in very few cases).
When the proposed method used Gaussians to model clusters, the only free
parameter was the variance, σ 2 . A suitable value for σ 2 was estimated from a
set of 20 images as follows. Many texel hypotheses were automatically generated
using different numbers of AC peaks and a user then selected from them the
best translation vectors, t1 , t2 . Pixels were allocated to clusters according to
the resulting lattice and a maximum likelihood estimation of σ 2 was computed.

The result was σ 2 = 264. Since this semi-automatic method might not be using
precise texel estimates, it might overestimate the variance compared to that
which would be obtained using optimal lattices. Therefore, further values for σ 2
(100, 144 and 196) were also used for evaluation in order to test the sensitivity of
the method. In any particular experiment, σ 2 was fixed for all 103 test images.
The method was also evaluated using a Gaussian mixture to model each cluster,
with free parameters set to σ12 = 60, σ22 = 800, and π1 = 0.9.
The observers were shown lattices overlaid on images and were asked to label
each lattice as obviously correct (OC), obviously incorrect (OI), or neutral. They
were to assign OC if the lattice was exactly the same or very close to what
they expected, OI if the result was far from their expectations, and neutral
otherwise. The presentation of results to the observers was randomised so as to
hide from them which algorithms produced which results. The proposed method
was compared with two related algorithms [12,1]. Liu et al. [1] did not specify
how to determine the number of peaks in the autocorrelation function. Results
are reported here using three different values for the number of peaks, namely
10, 40, and 70. Table 1 summarises the results. It seems clear that the method
proposed in this paper has superior accuracy to the two other methods. The
value of σ 2 had little effect on the results. Fig. 4 shows some examples of lattices
obtained. The two images displayed in the first row have clear intensity variations
between texels. The two examples in the second row have labels in the image and
appearance varies among texels. Examples shown in rows 3 to 5 contain large
non-homogenous texels. The left example in the last row is a neutral result.
This example has a significant geometric deformation among texels. The right
example in the last row is an OI result since it did not find the smallest texel.
Table 1. Comparison of proposed algorithm with related algorithms. Accuracy is de-

fined as the number of OC results divided by the total number of test images.
Algorithm variant # OC results # OI results # Neutral results Accuracy

Gaussian (σ 2 = 100) 83 9 11 0.81
Gaussian (σ 2 = 144) 83 14 6 0.81
Gaussian (σ 2 = 196) 82 14 7 0.80
Gaussian (σ 2 = 264) 79 18 6 0.77
Gaussian mixture 81 17 5 0.79
Liu et al. [1] (10 peaks) 45 54 4 0.44
Liu et al. [1] (40 peaks) 50 47 6 0.49
Liu et al. [1] (70 peaks) 28 70 5 0.27
Lin et al. [12] 22 70 11 0.21
A further experiment was performed to compare the proposed method to

the two other methods. For each image, lattice results from our algorithm using
Gaussians, our algorithm using Gaussian mixtures, the algorithm of Liu et al. [1],
and the algorithm of Lin et al. [12], respectively, were shown on the screen
simultaneously. The two subjects rank ordered those four results. Algorithms
Fig. 4. Results from the proposed algorithm using Gaussian models

shared the same rank if they yielded equally good results. For example, if three
of the algorithms gave good lattices of equal quality and the fourth algorithm
gave a poor lattice then three algorithms shared rank 1 and the other algorithm
was assigned rank 4. Table 2 summarizes the rankings. For the Gaussian model,
we set σ 2 = 264 which yields the worst accuracy of the variance values tried.
For the algorithm of Liu et al. [1], we set the number of dominant peaks to
40, which achieved the best performance of the values tried. Even with these
parameter settings which disadvantage the proposed method, Table 2 shows
that it is superior to the other algorithms.
Table 2. Comparisons by ranking results of different algorithms
Algorithm # Rank 1 # Rank 2 # Rank 3 # Rank 4

Gaussian, σ 2 = 264 83 12 6 2
Gaussian mixture 86 11 5 1
Liu et al. [1] (# peaks = 40) 56 5 23 19
Lin et al. [12] 18 2 24 59
The method was also used to classify texture images as regular or irregular as
described in Equation (6). A set of 62 images was selected randomly from a mu-
seum fine art database and from the same commercial textile archive as used ear-
lier. Figure 5 shows some examples of these images. A classification experiment
Fig. 5. Examples of images to be classified as having irregular texture
0.4
0.35
0.3
False negative rate
0.25
0.2
0.15
0.1
0.05
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
False positive rate
Fig. 6. Classification of texture as regular or irregular. The curve was plotted by vary-
ing the value of σ 2 and characterises the trade-off between the two types of error.
was performed using these images as negative examples and the 103 regular tex-
ture images as positive examples. Figure 6 shows the ROC curve obtained by vary-
ing the value of σ 2 in the Gaussian model (σ 2 ∈ {49, 64, 81, 100, 144}). The equal
error rate was approximately 0.22.
The computational speed depends on the number of lattice hypotheses (and
many different subsets of peaks lead to the same lattice hypothesis). A Matlab
implementation typically takes a few minutes per image on a 2.4GHz, 3.5GB
PC which is adequate for off-line processing.
6 Smart Thumbnail Generation for Regular Texture

Images
Thumbnail images are widely used when showing lots of images on a display
device of limited size. Most traditional approaches generate thumbnails by di-
rectly sub-sampling the original image which often reduces the recognisability of
meaningful objects and patterns in the image. Suh et al. [19] developed a novel
thumbnail generation method by taking into account human visual attention.
A saliency map and a face detector were used to identify regions expected to
attract visual attention. Although this method is effective for many images, it
is not appropriate for images with regular texture that often comprise abstract
patterns.
In an informal experiment, 9 human observers of varied age were asked to
draw a rectangle on each of 14 regular texture images to delineate the region
they would like to see as a thumbnail on a limited display. Most users tended
to select regions a little larger than a single texel, or containing a few texels.
This suggests that thumbnails might usefully be generated from regular tex-
ture images automatically by cropping based on texel extraction. Currently, we
are exploring the use of such thumbnails for content-based image browsing and
retrieval. Thumbnails are generated by cropping a rectangular sub-image that
bounds a region a little larger than a texel,(1.5t1 , 1.5t2 ). Fig. 7 compares two
thumbnails generated in this way with the standard method of directly reducing
Fig. 7. Comparisons of two thumbnail generation methods. In each set, the first image
is the original image, the second image is the thumbnail generated by our method, and
the third image is the thumbnail generated by the standard method.
the resolution. Thumbnails extracted using knowledge of the texels can convey
more detailed information about the pattern design.
7 Conclusions
A fully automatic lattice extraction method for regular texture images has been
proposed using a framework of statistical model selection. Texel hypotheses were
generated based on finding peaks in the AC function of the image. BIC was
adopted to compare various hypotheses and to select a ‘best’ lattice. The exper-
iments and comparisons with previous work have demonstrated the promise of
the approach. Various extensions to this work would be interesting to investigate
in future work. Alternative methods for generating hypotheses could be explored
in the context of this approach. Further work is needed to explore the relative
merits of non-Gaussian models. This should enable better performance on im-
ages of damaged textiles, for example. BIC can give poor approximations to the
marginal likelihood and it would be worth exploring alternative approximations
based on sampling methods, for example. Finally, it should be possible in princi-
ple to extend the approach to analysis of near-regular textures on deformed 3D
surfaces by allowing relative deformation between texels. This could be formu-
lated as a Markov random field over texels, for example. Indeed, Markov random
field models have recently been applied to regular texture tracking [6].
Acknowledgments. The authors thank J. Hays for providing his source code,
and Chengjin Du and Wei Jia for helping to evaluate the algorithm. This re-
search was supported by the UK Technology Strategy Board grant “FABRIC:
Fashion and Apparel Browsing for Inspirational Content” in collaboration with
Liberty Fabrics Ltd., System Simulation Ltd. and Calico Jack Ltd. The Technol-
ogy Strategy Board is a business-led executive non-departmental public body,
established by the government. Its mission is to promote and support research
into, and development and exploitation of, technology and innovation for the
benefit of UK business, in order to increase economic growth and improve the
quality of life. It is sponsored by the Department for Innovation, Universities
and Skills (DIUS). Please visit www.innovateuk.org for further information.
References
1. Liu, Y., Collins, R.T., Tsin, Y.: A computational model for periodic pattern percep-
tion based on frieze and wallpaper groups. IEEE Transactions on Pattern Analysis
and Machine Intelligence 26, 354–371 (2004)
2. Leung, T., Malik, J.: Recognizing surfaces using three-dimensional textons. In:
IEEE International Conference on Computer Vision, Corfu, Greece, pp. 1010–1017
(1999)
3. Liu, Y., Tsing, Y., Lin, W.: The promise and perils of near-regular texture. Inter-
national Journal of Computer Vision 62, 145–159 (2005)
4. Malik, J., Belongie, S., Shi, J., Leung, T.: Textons, contours and regions: cue
integration in image segmentation. In: IEEE International Conference of Computer
Vision, Corfu, Greece, pp. 918–925 (1999)
5. Hays, J., Leordeanu, M., Efros, A., Liu, Y.: Discovering texture regularity as a
higher-order correspondance problem. In: European Conference on Computer Vi-
sion, Graz, Austria, pp. 533–535 (2006)
6. Lin, W., Liu, Y.: A lattice-based MRF model for dynamic near-regular texture
tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence 29,
777–792 (2007)
7. Leung, T., Malik, J.: Detecting, localizing and grouping repeated scene elements
from an image. In: European Conference on Computer Vision, Cambridge, UK,
pp. 546–555 (1996)
8. Tuytelaars, T., Turina, A., Gool, L.: Noncombinational detection of regular repeti-
tions under perspective skew. IEEE Transactions on Pattern Analysis and Machine
Intelligence 25, 418–432 (2003)
9. Schaffalitzky, F., Zisserman, A.: Geometric grouping of repeated elements within
images. In: Shape, Contour and Grouping in Computer Vision. Lecture Notes In
Computer Science, pp. 165–181. Springer, Heidelberg (1999)
10. Forsyth, D.A.: Shape from texture without boundries. In: European Conference in
Computer Vision, Copenhagen, Denmark, pp. 225–239 (2002)
11. Lobay, A., Forsyth, D.A.: Recovering shape and irradiance maps from rich dense
texton fields. In: Computer Vision and Pattern Recognition, Washington, USA,
pp. 400–406 (2004)
12. Lin, H., Wang, L., Yang, S.: Extracting periodicity of a regular texture based on
autocorrelation functions. Pattern Recognition Letters 18, 433–443 (1997)
13. Chetverikov, D.: Pattern regularity as a visual key. Image and Vision Comput-
ing 18, 975–985 (2000)
14. Leu, J.: On indexing the periodicity of image textures. Image and Vision Comput-
ing 19, 987–1000 (2001)
15. Charalampidis, D.: Texture synthesis: Textons revisited. IEEE Transactions on
Image Processing 15, 777–787 (2006)
16. Starovoitov, V., Jeong, S.Y., Park, R.: Texture periodicity detection: features, prop-
erties, and comparisons. IEEE Transactions on Systems, Man, and Cybernetics-
A 28, 839–849 (1998)
17. Schwarz, G.: Estimating the dimensions of a model. Annals and Statistics 6, 461–
464 (1978)
18. Raftery, A.E.: Bayesian model selection in social research. Sociological Methodol-
ogy 25, 111–163 (1995)
19. Suh, B., Ling, H., Benderson, B.B., Jacobs, D.W.: Automatic thumbnail cropping
and its effectiveness. In: ACM Symposium on User Interface Software and Tech-
nology, pp. 95–104 (2003)
Higher Dimensional Affine Registration
and Vision Applications
Yu-Tseh Chi1 , S.M. Nejhum Shahed1 , Jeffrey Ho1 , and Ming-Hsuan Yang2
1
CISE Department, University of Florida, Gainesville, 32607
{ychi,smshahed,jho}@csie.ufl.edu
2
EECS, University of California, Merced, CA 95344
mhyang@ucmerced.edu
Abstract. Affine registration has a long and venerable history in computer vi-
sion literature, and extensive work have been done for affine registrations in IR2
and IR3 . In this paper, we study affine registrations in IRm for m > 3, and to
justify breaking this dimension barrier, we show two interesting types of match-
ing problems that can be formulated and solved as affine registration problems
in dimensions higher than three: stereo correspondence under motion and image
set matching. More specifically, for an object undergoing non-rigid motion that
can be linearly modelled using a small number of shape basis vectors, the stereo
correspondence problem can be solved by affine registering points in IR3n . And
given two collections of images related by an unknown linear transformation of
the image space, the correspondences between images in the two collections can
be recovered by solving an affine registration problem in IRm , where m is the
dimension of a PCA subspace. The algorithm proposed in this paper estimates
the affine transformation between two point sets in IRm . It does not require con-
tinuous optimization, and our analysis shows that, in the absence of data noise,
the algorithm will recover the exact affine transformation for almost all point sets
with the worst-case time complexity of O(mk2 ), k the size of the point set. We
validate the proposed algorithm on a variety of synthetic point sets in different
dimensions with varying degrees of deformation and noise, and we also show
experimentally that the two types of matching problems can indeed be solved
satisfactorily using the proposed affine registration algorithm.
1 Introduction
Matching points, particularly in low-dimensional settings such as 2D and 3D, has been
a classical problem in computer vision. The problem can be formulated in a variety of
ways depending on the allowable and desired deformations. For instance, the orthogonal
and affine cases have been studied already awhile ago, e.g., [1][2], and recent research
activities have been focused on non-rigid deformations, particularly those that can be
locally modelled by a family of well-known basis functions such as splines, e.g., [3]. In
this paper, we study the more classical problem of matching point sets1 related by affine
transformations. The novel viewpoint taken here is the emphasis on affine registrations
1
In this paper, the two point sets are assumed to have the same size.

Higher Dimensional Affine Registration and Vision Applications 257
in IRm for m > 3, and it differs substantially from the past literature on this subject,
which has been overwhelmingly devoted to registration problems in IR2 and IR3 .
To justify breaking this dimension barrier, we will demonstrate that two important
and interesting types of matching problems can be formulated and solved as affine reg-
istration problems in IRm with m > 3: stereo correspondence under motion and image
set matching (See Figure 1). In the stereo correspondence problem, two video cameras
are observing an object undergoing some motion (rigid or non-rigid), and a set of k
points on the object are tracked consistently in each view. The problem is to match
the tracking results across two views so that the k feature points can be located and
identified correctly. In the image set matching problem, two collections of images are
given such that the unknown transformation between corresponding pairs of images

can be approximated by some linear transformation F : IRm → IRm between two
(high-dimensional) image spaces. The task is to compute the correspondences directly
from the images. Both problems admit quick solutions. For example, for stereo corre-
spondence under motion, one quick solution would be to select a pair of corresponding
frames and compute the correspondences directly between these two frames. This ap-
proach is clearly since there is no way to know a priori which pair of frames is optimal
for computing the correspondences. Furthermore, if the baseline between cameras is
large, direct stereo matching using image features does not always produce good re-
sults, even when very precise tracking result are available. Therefore, there is a need
for a principled algorithm that can compute the correspondences directly using all the
tracking results simultaneously instead of just a pair of frames.
Fig. 1. Left: Stereo Correspondence under Motion. A talking head is observed by two (affine)
cameras. Feature points are tracked separately on each camera and the problem is to compute
the correspondences between observed feature points across views. Center and Right: Image
Set Matching. Two collections (432 images each) of images are given. Each image on the right
is obtained by rotating and down-sizing an image on the left. The problem is to recover the
correspondences. These two problems can be formulated as affine registration problems in IRm
with m > 3.
An important point to realize is that in each problem there are two linear subspaces
that parameterize the input data. For nonrigid motions that can be modelled using linear
shape basis vectors, this follows immediately from the work of [4][5]. For image set
matching, each set of images can usually be approximated by a linear subspace with
dimension that is considerably smaller than that of the ambient image space. We will
258 Y.-T. Chi et al.
show that the correspondences can be computed (or be approximated) by affine regis-
tering point sets in these two linear subspaces. Therefore, instead of using quantities
derived from image intensities, our solution to these two matching problems is to first
formulate them as affine point set matching problems in IRm , with m > 3, and solve
the resulting affine registration problems.
Let P = {p1 , · · · , pk } and Q = {q1 , · · · , qk } denote two point sets in IRm with equal
number of points. The affine registration problem is typically formulated as an opti-
mization problem of finding an affine transformation A and a correspondence map π be-
tween points in P, Q such that the following registration error function is
minimized
k
E(A, π) = d2 (Api , qπ(i) ), (1)
i=1
where d(Api , qπ(i) ) denotes the usual L2 -distance between Api and qπ(i) . The vener-
able iterative closest point (ICP) algorithm [6][7] can be easily generalized to handle
high-dimensional point sets, and it gives an algorithm that iteratively solves for corre-
spondences and affine transformation. However, the main challenge is to produce good
initial correspondences and affine transformation that will guarantee the algorithm’s
convergence and the quality of the solution. For dimensions two and three, this is al-
ready a major problem and the difficulty increases exponentially with dimension. In this
paper, we propose an algorithm that can estimate the affine transformation (and hence
the correspondences π) directly from the point sets P, Q. The algorithm is algebraic in
nature and does not require any optimization, which is its main strength. Furthermore, it
allows for a very precise analysis showing that for generic point sets and in the absence
of noise, it will recover the exact affine transformation and the correspondences. For
noisy data, the algorithm’s output can serve as a good initialization for the affine-ICP
algorithm. While the algorithm is indeed quite straightforward, it is to the best of our
knowledge that there has not been published algorithm which is similar to ours in its
entirety. In this paper, we will provide experimental results that validate the proposed
affine registration algorithm and show that both the stereo correspondence problem un-
der motion and image set matching problem can be solved quite satisfactorily using the
proposed affine registration algorithm.
2 Affine Registrations and Vision Applications

In this section, we provide the details for formulating the stereo correspondence under
motion and image set matching problems as affine registration problems.
2.1 Stereo Correspondences under Motion

For clarity of presentation, we will first work out the simpler case of rigid motions.
We assume two stationary affine cameras C1 , C2 observing an object O undergoing
some (rigid or nonrigid) motion. On each camera, we assume that some robust tracking
algorithm is running so that a set {X1 , · · · , Xk } of k points on O are tracked over
T frames separately on both cameras. Let ( xtij , yij t
) 1 ≤ i ≤ 2, 1 ≤ j ≤ k, 1 ≤
t ≤ T denote the image coordinates of Xj ∈ O in the tth frame from camera i. For

each camera, the tracker provides the correspondences (xtij , yij t
) ↔ (xtij , yij
t
) across

different frames t and t . Our problem is to compute correspondences across two views
so that the corresponding points (xt1j , y1j
t
) ↔ (xt2j , y2j
t
) are the projections of the scene
point Xj in the images. We show next that it is possible to compute the correspondences
directly using only the high-dimensional geometry of the point sets ( xtij , yij t
) without
referencing to image features such as intensities.
For each view, we can stack the image coordinates of one tracked point over T frames
vertically into a 2T -dimensional vector:
pj = ( x11j y1j
1
· · · , xT1j y1j
T t
) , qj = ( x12j y2j
1
· · · , xT2j y2j
T t
) (2)
In motion segmentation (e.g., [8]), the main objects of interest are the 4-dimensional
subspaces Lp , Lq spanned by these 2T -dimensional vectors
P = {p1 , · · · , pk }, Q = {q1 , · · · , qk },
and the goal is to cluster motions by determining the subspaces Lp , Lq given the set of
vectors P ∪ Q. Our problem, on the hand, is to determine the correspondences between
points in P and Q. It is straightforward to show that there exists an affine transformation
L : Lp → Lq that produces the correct correspondences, i.e., L(pi ) = qi for all i. To
see this, we fix an arbitrary world frame with respect to which we can write down the
camera matrices for C1 and C2 . In addition, we also fix an object coordinates system
with orthonormal basis {i, j, k} centered at some point o ∈ O. Since O is undergoing
a rigid motion, we denote by ot , it , jt , kt , the world coordinates of o, i, j, k at frame t.
The point Xj , at frame t, with respect to the fixed world frame is given by
Xjt = ot + αj it + βj jt + γj kt , (3)
for some real coefficients αj , βj , γj that are independent of time t. The corresponding
image point is then given as
( xtij , yij
t t
) = õit + αj ĩit + βj j̃it + βj k̃it ,
where õit , ĩit , j̃it , k̃it are the projections of the vectors ot , it , jt , kt onto camera i. In par-
ticular, if we define the 2T -dimensional vectors Oi , Ii , Ji , Ki by stacking the vectors
õit , ĩit , j̃it , k̃it vertically as before, we have immediately,
pj = O1 + αj I1 + βj J1 + γj K1 , qj = O2 + αj I2 + βj J2 + γj K2 . (4)
The two linear subspaces Lp , Lq are spanned by the basis vectors {O1 , I1 , J1 , K1 },
{O2 , I2 , J2 , K2 }, respectively. The linear map that produces the correct correspon-
dences is given by the linear map L such that L(O1 ) = O2 , L(I1 ) = I2 , L(J1 ) = J2
and L(K1 ) = K2 . A further reduction is possible by noticing that the vectors pj , qj be-
long to two three-dimensional affine linear subspaces Lp , Lq in IR2T , affine subspaces
that pass through the points O1 , O2 with bases {I1 , J1 , K1 } and {I2 , J2 , K2 }, respec-
tively. These two subspaces can be obtained by computing the principle components
for the collections of vectors P, Q. By projecting points in P, Q onto Lp , Lq , respec-
tively, it is clear that the two sets of projected points are now related by an affine map
A : Lp → Lq . In other words, the correspondence problem can now be solved by
solving the equivalent affine registration problem for these two sets of projected points
(in IR3 ).
Non-Rigid Motions. The above discussion generalizes immediately to the types of non-
rigid motions that can be modelled (or approximated) using linear shape basis [2,5,9].
In this model, for k feature points, a shape basis element Bl is a 3 × k matrix. For a
model that employs m linear shape basis elements, the 3D world coordinates of the k
feature points at tth frame can be written as a linear combination of these shape basis
elements:
# t $ m
X1 · · · Xkt = atl Bl , (5)
l=1
for some real numbers atl . Using affine camera model, the imaged points (disregarding
the global translation) are given by the following equation [9]
# t $
x1 · · · xtk = (a ⊗ P )B, (6)
where at = (at1 , · · · , atm ), P is the first 2 × 3 block of the affine camera matrix and B is
the 3m × k matrix formed by vertically stacking the shape basis matrices Bl . The right
factor in the above factorization is independent of the camera (and the images), and we
have the following equations similar to Equations 4:

m
m
pj = O1 + (αjl I1l + βjl J1l + γjl K1l ), qj = O2 + (αjl I2l + βjl J2l + γjl K2l ),
l=1 l=1
(7)
where Iil , Jil Kil are the projections of the three basis vectors in the lth shape basis
element Bl onto camera i. The numbers αjl , βjl and γjl are in fact entries in the matrix
Bl . These two equations then imply, using the same argument as before, that we can re-
cover the correspondences directly using a 3m-dimensional affine registration provided
that the vectors Oi , Iil , Jil , Kil are linearly independent for each i, which is typically
the case when the number of frames is sufficiently large.
2.2 Image Set Matching
In the image set matching problem, we are given two sets of images P = {I1 , · · · , Ik }

⊂ IRm , Q = {I1 , · · · , Ik } ⊂ IRm and the corresponding pairs of images Ii , Ii are

related by a linear transformation F : IRm → IRm between two high-dimensional
image spaces:
Ik ≈ F(Ik ).
Examples of such sets of images are quite easy to come by, and Figure 1 gives an
example in which Ii is obtained by rotating and downsizing Ii . It is easy to see that
many standard image processing operations such as image rotation and down-sampling
can be modelled as (or approximated by) a linear map F between two image spaces.
The problem here is to recover the correspondences Ii ↔ Ii without actually computing
the linear transformation F , which will be prohibitively expensive since the dimensions
of the image spaces are usually very high.
Many interesting sets of images can in fact be approximated well by low-dimensional
linear subspaces in the image space. Typically, such linear subspaces can be computed
readily using principal component analysis (PCA). Let Lp , Lq denote two such low-
dimensional linear subspaces approximating P, Q, respectively and we will use the
same notations P, Q to denote their projections onto the subspace Lp , Lq . A natural
question to ask is how are the (projected) point sets P, Q related? Suppose that F is
orthogonal and Lp , Lq are the principle subspaces of the same dimension. If the data
is “noiseless”, i.e., Ik = F (Ik ), it is easy to show that P, Q are then related by an
orthogonal transformation. In general, F may not be orthogonal and data points are
noisy, the point sets P, Q are related by a transformation T = A + r, which is a sum
of an affine transformation A and a nonrigid transformation r. If the nonrigid part is
small, we can recover the correspondences by affine registering the two point sets P, Q.
Note that this gives an algorithm for computing the correspondences without explicitly
using the image contents, i.e., there is no feature extraction. Instead, it works directly
with the geometry of the point sets.
3 Affine Registrations in IRm

The above discussion provides the motivation for studying affine registration in IRm for
m > 3. Let P = {p1 , · · · , pk } and Q = {q1 , · · · , qk } be two point sets in IRm related
by an unknown affine transformation
qπ(i) = Api + t, (8)
where A ∈ GL(m), t ∈ IRm the translational component of the affine transformation

and π : P → Q, the unknown correspondence to be recovered. We assume that the
point sets P, Q have same number of points and π is a bijective correspondence.
Iterative closest point (ICP) algorithm is a very general point registration algorithm
that generalizes easily to higher dimensions. Several papers have been published re-
cently [10,11,12,13,14] on ICP-related point registration algorithms in IR2 and IR3 .
While these works concern exclusively with rigid transformations, it is straightfor-
ward to incorporate affine transformation into ICP algorithm, which iterative solves for
correspondences and affine transformation2. Given an assignment (correspondences)
π : {1, · · · , k} → {1, · · · , k} the optimal affine transformation A in the least squares
sense can be solved by minimizing

k
E(A, t, π) = d2 (Api + t, qπ(i) ). (9)
i=1
Solving A, t separately while holding π fixed, the above registration error function
gives a quadratic programming problem in the entries of A, and the optimal solution
can be computed readily by solving a linear system. With a fixed A, t can be solved
immediately. On the hand, given an affine transformation, a new assignment π can be
defined using closest points:
π(i) = arg min d2 (Api + t, qj ).

1≤j≤k
2
We will call this algorithm affine-ICP.
Once an initial affine transformation and assignment is given, affine-ICP is easy to

implement and very efficient. However, the main difficulty is the initialization, which
can significantly affect the algorithm’s performance. With a poor initialization, the al-
gorithm almost always converges to an undesirable local minimum and as the group
of affine transformations is noncompact, it is also possible that it diverges to infinity,
i.e., the linear part of the affine transformation converges to a singular matrix. One
way to generate an initial affine transformation (disregarding t) is to randomly pick m
pairs of points from P, Q, {(x1 , y1 ), · · · , (xm , ym )}, xi ∈ P, yi ∈ Q and define A as
yi = A(xi ). It is easy to see that the probability of picking a good set of pairs that will
yield good initialization is roughly in the order of 1/C(k, m). For small dimensions
m = 2, 3 and medium-size point sets (k in the order of hundreds), it is possible to ex-
haustively sample all these initial affine transformations. However, as C(k, m) depends
exponentially on the dimension m, this approach becomes impractical once m > 3.
Therefore, for affine-ICP approach to work, we need a novel way to generate good
initial affine transformation and correspondences.
Our solution starts with a novel affine registration algorithm. The outline of the al-
gorithm is straightforward: we first reduce the problem to orthogonal case and spectral
information is then used to narrow down the correct orthogonal transformation. This
algorithm does not require continuous optimization (e.g., solving linear systems) and
we can show that for generic point sets without noise, it will recover the exact affine
transformation. This latter property suggests that for noisy point sets, the affine trans-
formation estimated by the proposed algorithm should not be far from the optimal one.
Therefore, the output of our proposed algorithm can be used as the initial affine trans-
formation for the affine-ICP algorithm.
3.1 Affine Registration Algorithm
Let P, Q be two point sets as above related by an unknown affine transformation as in

Equation 8. By centering the point sets with respect to their respective centers of mass
mp , mq ,
1 1
k k
mp = pi , mq = qi ,
k i=1 k i=1
the centered point sets P c = {p1 − mp , · · · , pk − mp } and Qc = {q1 − mq , · · · , qk −

mq } are related by the same A: qπ(i) − mq = A(pi − mp ). That is, we can work with
centered point sets P c and Qc . Once A and π have been recovered from the point sets
P c and Qc , the translational component t can be estimated easily. In the absence of
noise, determining the matrix A is in fact a combinatorial search problem. We can se-
lect m linearly independent points {pi1 , · · · , pim } from P. For every ordered m points
ω = {qi1 , · · · , qim } in Q, there is a (nonsingular) matrix Bω sending pij to qij for
1 ≤ j ≤ m. The desired matrix A is among the set of such matrices, which numbers
roughly k m (k is the number of points). For generic point sets, this exponential depen-
dence on dimension can be avoided if A is assumed to be orthogonal. Therefor, we
will first use the covariance matrices computed from P and Q to reduce the problem
to the ‘orthogonal case’. Once the problem has been so reduced, there are various ways
to finish off the problem by exploiting invariants of the orthogonal matrices, namely,
distances. Let SP and SQ denote the covariance matrices for P and Q, respectively:

k
k
SP = pi pti , SQ = qi qit .
i=1 i=1
We make simple coordinates changes using their inverse square-roots:

−1 −1
pi → SP 2 pi , qi → SQ 2 qi . (10)
We will use the same notations to denote the transformed points and point sets. If the
original point sets are related by A, the transformed point sets are then related by Ā =
−1 1
SQ 2 ASP2 . The matrix Ā can be easily shown to be orthogonal:
Proposition 1. Let P and Q denote two point sets (of size k) in IRm , and they are
related by an unknown linear transformation A. Then, the transformed point sets (using
Equation 10) are related by a matrix Ā, whose rows are orthonormal vectors in IRm .
The proof follows easily from the facts that 1) the covariance matrices SP and SQ
are now identity matrices for the transformed point sets, and 2) SQ = ĀSP Āt . They
together imply that the rows of Ā must be orthonormal.
3.2 Determining the Orthogonal Transformation Ā

Since the point sets P, Q have unit covariance matrices, the invariant approach in [1]
cannot be applied to solve for the orthogonal transformation Ā. Nevertheless, there are
other invariants that can be useful. For example, if the magnitudes of points in P are all
different, registration becomes particularly easy: each point pij is matched to the point
qij with the same magnitude. Of course, one does not expect to encounter such nice
point sets very often. However, for orthogonal matrices, there is a very general way to
produce a large number of useful invariants.
Let p1 , p2 be any two points in P and q1 , q2 their corresponding points in Q. Since
Ā is orthogonal, the distance d(p1 , p2 ) between p1 and p2 equals the distance d(q1 , q2 )
between q1 and q2 . Although we do not know the correspondences between points in
P and Q, the above observation naturally suggests the idea of canonically constructing
two symmetric matrices, LP and LQ , using pairwise distances between points in P and
Q, respectively. The idea is that the matrices so constructed differ only by an unknown
permutation of their columns and rows. Their eigenvalues, however, are not effected by
such permutations, and indeed, the two matrices LP and LQ have the same eigenval-
ues. Furthermore, there are also correspondences between respective eigenspaces EλP
and EλQ associated with eigenvalue λ. If λ is a non-repeating eigenvalue, we have two
associated (unit) eigenvectors vPλ
and vQ λ
of LP and LQ , respectively. The vector vP λ
differs from vQ by a permutation of its components and a possible multiplicative factor

λ
of −1.
There are many ways to construct the matrices LP and LQ . Let f (x) be any function.
We can construct a k × k symmetric matrix LP (f ) from pairwise distances using the
formula
⎛ ⎞
f (d(p1 , p1 )) · · · f (d(p1 , pk ))
⎜ .. .. ⎟
LP (f ) = Ik − μ ⎝ . ··· . ⎠, (11)
f (d(pk , p1 )) · · · f (d(pk , pk ))
where Ik is the identity matrix and μ some real constant. One common choice of f
that we will use here is the Gaussian exponential f (x) = exp(−x2 /σ 2 ), and the result-
ing symmetric matrix LP is related to the well-known (unnormalized) discrete Lapla-
cian associated with the point set P [15]. Denote Up Dp Upt = LP , Uq Dq Uqt = LQ
the eigen-decompositions of LP and LQ . When the eigenvalues are all distinct, up to
sign differences, Up and Uq differ only by some unknown row permutation if we order
the columns according to the eigenvalues. This unknown row permutation is exactly
the desired correspondence π. In particular, we can determine m correspondences by
matching m rows of Up and Uq , and from these m correspondences, we can recover the
orthogonal transformation Ā. The complexity of this operation is O(mk 2 ) and we have
the following result
Proposition 2. For a generic pair of point sets P, Q with equal number of points in
IRm related by some orthogonal transformation L and correspondences π such that
qπ(i) = Lpi , the above method will recover L and π exactly for some choice of σ.
The proof (omitted here) is an application of Sard’s theorem and transversality in dif-
ferential topology [16]. The main idea is to show that for almost all point sets P, the
symmetric matrix LP will not have repeating eigenvalues for some σ. This will guaran-
tee that the row-matching procedure described above will find the m needed correspon-
dences after examining all rows of Up m times. Since the time complexity for matching
one row is O(k), the total time complexity is no worse than O(mk 2 ).
3.3 Dealing with Noises

The above method breaks down when noise is present. In this case, the sets of eigenval-
ues for LP , LQ are in general different, and the matrices Up , Uq are no longer expected
to differ only by a row permutation. Nevertheless, for small amount of noise, one can
expect that the matrices LP , LQ are small perturbations of two corresponding matrices
for noiseless data. For example, up to a row permutation, Uq is a small perturbation
of Up . For each eigenvalue λp of LP , there should be an eigenvalue λq of LQ such
that the difference |λp − λq | is small, and this will allow us to establish correspon-
dences between eigenvalues of LP , LQ . The key idea is to define a reliable matching
measure M using eigenvectors of LP , LQ , e.g., if p, q are two corresponding points,
M(p, q) will tend to be small. Otherwise, it is expected to be large. Once a matching
measure M is defined, it will allow us to establish tentative correspondences pi ←→ qj :
qj = arg mini M(pi , qi ). Similar to the homography estimation in structure from mo-
tion [2], some of the tentative correspondences so established are incorrect while a good
portion of them are expected to be correct. This will allow us to apply RANSAC [17]
to determine the orthogonal transformation: generate a small number of hypotheses (or-
thogonal matrices from sets of randomly generated m correspondences) and pick the
one that gives the smallest registration error. We remark that in our approach, the tenta-
tive correspondences are computed from the geometry of the point sets P, Q embedded
in IRm . In stereo matching and homography estimation [2], they are computed using
image features such as image gradients and intensity values.
More precisely, let λp1 < λp2 < · · · < λpl (l ≤ k) be l non-repeating eigenvalues of
LP and likewise, λq1 < λq2 < · · · < λql the l eigenvalues of LQ such that |λqi − λpi | <
for some threshold value . Let vλP1 , vλP2 , · · · , vλPl , and vλQ1 , vλQ2 , · · · , vλQl denote the
corresponding eigenvectors. We stack these eigenvectors horizontally to form two k × l
matrices VP and VQ:
VP = [vλP1 vλP2 · · · vλPl ], VQ = [vλQ1 vλQ2 · · · vλQl ]. (12)
Denote the i, j-entry of VP (and also VQ) by VP(i, j). We define the matching measure
M as

l
M(pi , qj ) = min{ (VP(i, h) − VQ(j, h))2 , (VP(i, h) + VQ(j, h))2 }.
h=1
Note that if l = k, M is comparing the ith row of LP with j th row of LQ . For ef-
ficiency, one does not want to compare the entire row; instead, only a small fragment
of it. This would require us to use those eigenvectors that are most discriminating for
picking the right correspondences. For discrete Laplacian, eigenvectors associated with
smaller eigenvalues can be considered as smooth functions on the point sets, while those
associated with larger eigenvalues are the non-smooth ones since they usually exhibit
greater oscillations. Typically, the latter eigenvectors provide more reliable matching
measures than the former ones and in many cases, using one or two such eigenvectors
(l = 2) is already sufficient to produce good results.
Table 1. Experimental Results I. For each dimension and each noise setting, one hundred trials,
each with different point sets and matrix A, were performed. The averaged relative error and
percentage of mismatched points as well as standard deviations (in parenthesis) are shown.
Dim → 3 3 5 5 10 10
Noise ↓ Matrix Error Matching Error Matrix Error Matching Error Matrix Error Matching Error
0% 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
1% 0.001 (0.0005) 0 (0) 0.002 (0.0006) 0 (0) 0.004 (0.0008) 0 (0)
2% 0.003 (0.001) 0 (0) 0.004 (0.001) 0 (0) 0.008 (0.001) 0 (0)
5% 0.008 (0.003) 0 (0) 0.01 (0.003) 0 (0) 0.02 (0.003) 0 (0)
10% 0.017 (0.01) 0.008 (0.009) 0.05 (0.05) 0.009 (0.04) 0.04 (0.009) 0 (0)
4 Experiments
In this section, we report four sets of experimental results. First, with synthetic point
sets, we show that the proposed affine registration algorithm does indeed recover exact
affine transformations and correspondences for noiseless data. Second, we show that
the proposed algorithm also works well for 2D point sets. Third, we provide two se-
quences of nonrigid motions and show that the feature point correspondences can be
Table 2. Experimental Results II. Experiments with point sets of different sizes with 5% noise
added. All trials match point sets in IR10 with settings similar to Table 1. Average errors for one
hundred trials are reported with standard deviations in parenthesis.
# of Pts → 100 150 200 250 300 400

Errors ↓ Points Points Points Points Points Points
Matrix Error 0.02 (0.003) 0.05(0.008) 0.05 (0.009) 0.05 (0.01) 0.05 (0.01) 0.04 (0.009)
Matching Error 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)
satisfactorily solved using affine registration in IR9 . And finally, we use images from
COIL database to show that the image set matching problem can also be solved using
affine registration in IR8 . We have implemented the algorithm using MATLAB without
any optimization. The sizes of the point sets range from 20 to 432, and on a DELL
desktop with single 3.1GHz processor, each experiment does not run longer than one
minute.
4.1 Affine Registration in IRm

In this set of experiments, our aim is to give a qualitative as well as quantitative analy-
sis on the accuracy and robustness of the proposed method. We report our experimental
results on synthetic data in several different dimensions and using various different
noise settings. Tables 1 and 2 summarize the experimental results. In Table 1, the al-
gorithm is tested in three dimensions, 3, 5 and 10, and five different noise settings,
0%, 1%, 2%, 5%, 10%. For each pair of dimension and noise setting, we ran 100 trials,
each with a randomly generated non-singular matrix A and a point set containing 100
points. In trials with x% noise setting, we add a uniform random noise (±x%) to each
coordinate of every point independently. Let A denote the estimated matrix. A point
p ∈ P is matched to the point q ∈ Q if q = minqi ∈Q dist(A p, qi ). For each trial,
we report the percentage of mismatched points and the relative error of the estimated

matrix A : AA −A
, using the Frobenius norm.
The number of (RANSAC) samples drawn in each trial has been fixed at 800 for
the results reported in Table 1. This is the number of samples needed to produce zero
mismatch for dimension 10 with 10% noise setting. In general, for lower dimensions,
a much smaller number of samples (around 200) would also have produced similar re-
sults. In Table 2, we vary the sizes of the point sets and work in IR10 . The setting is
similar to that of Table 1 except with fixed 5% noise setting for all trials. The results
clearly show that the proposed algorithm consistently performs well with respect to the
sizes of the point sets. Note also that for noiseless point sets the exact affine transfor-
mations are always recovered.
4.2 2D Point Sets

In the second set of experiments, we apply the proposed algorithm to 2D image registra-
tion. It is known that the effect of a small view change on an image can be approximated
by a 2D affine transformation of the image [2]. Using images from COIL database, we
manually click feature points on pairs of images with 15◦ to 30◦ difference in view
Fig. 2. 2D Image Registration. 1st column: Source images (taken from COIL database) with
feature points marked in red. 2nd and 4th column: Target images with feature points marked in
blue. 3rd and 5th column: Target images with corresponding feature points marked in blue. The
affine transformed points from the source images are marked in red. Images are taken with 15◦
and 30◦ differences in viewpoint. The RMS errors for these four experiments (from left to right)
2.6646, 3.0260, 2.0632, 0.7060, respectively.
Fig. 3. Top: Sample frames from two video sequences of two objects undergoing nonrigid mo-
tions. Bottom: Sample frames from another camera observing the same motions.
point. The registration results for four pairs of images are shown in Figure 2. Notice the
small RMS registration errors for all these results given that the image size is 128×128.
4.3 Stereo Correspondences under Nonrigid Motions

In this experiment, we apply affine registration algorithm to compute correspondences
between tracked feature points in two image sequences. We gathered four video se-
quences from two cameras observing two objects undergoing nonrigid motions
(Figure 3). One is a talking head and the other is a patterned tatoo on a man’s belly.
A simple correlation-based feature point tracker is used to track twenty and sixty points
for these two sequences, respectively. Seventy frames were tracked in both sequences
and manual intervention was required several times in both sequences to correct and
adjust tracking results. We use three shape basis for both sequences [5], and to com-
pute the correspondences, we affine register two point sets in IR9 as discussed before.
For the two point sets P, Q ⊂ IR9 , we applied the proposed algorithm to obtain ini-
tial correspondences and affine transformation. This is followed by running an affine-
ICP algorithm with fifty iterations. For comparison, the affine-ICP algorithm initialized
using closest points3 is run for one hundred iterations. For the talking sequence, the
3
Given two point sets in IR9 , the initial correspondence pi ↔ qj is computed by taking qj to
be the point in Q closest to pi .
proposed algorithm recovers all the correspondences correctly, while for the tatoo se-
quence, among the recovered sixty feature point correspondences, nine are incorrect.
This can be explained by the fact that in several frames, some of the tracked feature
points are occluded and missing and the subsequent factorizations produce relatively
noisy point sets in IR9 . On the other hand, affine-ICP with closest point initialization
fails poorly for both sequences. In particular, more than three quarters of the estimated
correspondences are incorrect.
4.4 Image Set Matching

In this experiment, images from the first six objects in the COIL database are used.
They define the image set A with 432 images. Two new sets B, C of images are gener-
ated from A: the images are 80% down-sampled and followed by 45◦ and 90◦ rotations,
respectively. The original images have size 128 × 128 and the images in the two new
sets have size 100 × 100. An eight-dimensional PCA subspace is used to fit each set of
images with relative residue smaller than 1%. Images in each set are projected down to
their respective PCA subspaces and the correspondences are automatically computed
by affine registering the projected point sets. The two experiments shown in Figure 4
match point sets A, B and A, C. We apply the proposed affine registration algorithm to
obtain an initial estimate on correspondences and affine transformation. Since the data
is noisy, we follow this with the affine-ICP algorithm running fifty iterations as above.
For comparison, we apply the affine-ICP algorithm using closest points as initializa-
tion. In both experiments, the affine-ICP algorithm, not surprisingly, performs poorly
with substantial L2 -registration errors (Equation 9) and large number of incorrect cor-
respondences. The proposed algorithm recovers all correspondences correctly and it
yields small L2 -registration errors.
Fig. 4. Image Set Matching. The original image set A is shown in Figure 1. Image sets B, C are
shown above. The plots on the right show the L2 -registration error for each of the fifty iterations
of running affine-ICP algorithm using different initializations. Using the output of the proposed
affine registration as the initial guess, the affine-ICP algorithm converges quickly to the desired
transformation (blue curves) and yields correct correspondences. Using closest points for initial
correspondences, the affine-ICP algorithm converges (red curves) to incorrect solutions in both
experiments.

In this paper, we have shown that the stereo correspondence problem under motion and
image set matching problem can be solved using affine registration in IRm with m > 3.
We have also proposed an algorithm for estimating an affine transformation directly
from two point sets without using continuous optimization. In the absence of noise, it
will recover the exact affine transformation for generic pairs of point sets in IRm . For
noisy data, the output of the proposed algorithm often provides good initializations for
the affine-ICP algorithm. Together, they provide us with an efficient and effective algo-
rithm for affine registering point sets in IRm with m > 3. We have applied the proposed
algorithm to the two aforementioned problems. Preliminary experimental results are
encouraging and they show that these two problems can indeed be solved satisfactorily
using the proposed affine registration algorithm.
References
1. Scott, G., Lonquiet-Higgins, C.: An algorithm for associating the features of two images.
Proc. of Royal Society of London B244, 21–26 (1991)
2. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cambridge Uni-
versity Press, Cambridge (2003)
3. Chui, H., Rangarajan, A.: A new algorithm for non-rigid point matching. In: Proc. IEEE
Conf. on Comp. Vision and Patt. Recog, vol. 2, pp. 44–51 (2000)
4. Toamsi, C., Kanade, T.: Shape and motion from image streams under orthography—a factor-
ization method. Int. J. Computer Vision 9(2), 137–154 (1992)
5. Bregler, C., Hertzmann, A., Biermann, H.: Recovering non-rigid 3d shape from image
streams. In: Proc. IEEE Conf. on Comp. Vision and Patt. Recog., pp. 2690–2696 (2000)
6. Besel, P.J., Mckay, H.D.: A method for registration of 3-d shapes. PAMI 14, 239–256 (1992)
7. Zhang, Z.: Iterative point matching for registration of free-form curves and surfaces. Int. J.
Computer Vision 13, 119–152 (1994)
8. Kanatani, K.: Motion segmentation by subspace separation and model selection. In: Proc.
Int. Conf. on Computer Vision, vol. 2, pp. 586–591 (2001)
9. Brand, M.: Morphable 3d models from video. In: Proc. IEEE Conf. on Comp. Vision and
Patt. Recog., vol. 2, pp. 456–463 (2001)
10. Fitzgibbon, A.W.: Robust registration of 2d and 3d point sets. Computer Vision and Image
Understanding 2, 1145–1153 (2003)
11. Sharp, G.C., Lee, S.W., Wehe, D.K.: Icp registration using invariant features. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence 24, 90–102 (2002)
12. Rusinkiewicz, S., Levoy, M.: Efficient variants of the icp algorithm. In: Proc. Third Interna-
tional Conference on 3D Digital Imaging and Modeling (3DIM), pp. 145–152 (2001)
13. Granger, S., Pennec, X.: Multi-scale em-icp: A fast and robust approach for surface registra-
tion. In: Proc. European Conf. on Computer Vision, vol. 3, pp. 418–432 (2002)
14. Makadia, A., Patterson, A.I., Daniilidis, K.: Fully automatic registration of 3d point clouds.
In: Proc. IEEE Conf. on Comp. Vision and Patt. Recog., vol. 1, pp. 1297–1304 (2006)
15. Chung, F.R.K.: Spectral Graph Theory. American Mathematical Society (1997)
16. Hirsch, M.: Differential Topology. Springer, Heidelberg (1976)
17. Fischler, M., Bolles, R.: Random sample consensus: A paradigm for model fitting with ap-
plications to image analysis and automated cartography. Communications of the ACM 24,
381–395 (1981)
Semantic Concept Classification by Joint
Semi-supervised Learning of Feature Subspaces
and Support Vector Machines
Wei Jiang1 , Shih-Fu Chang1 , Tony Jebara1, and Alexander C. Loui2

1
Columbia University, New York, NY 10027, USA
2
Eastman Kodak Company, Rochester, NY 14650, USA
Abstract. The scarcity of labeled training data relative to the high-

dimensionality multi-modal features is one of the major obstacles for
semantic concept classification of images and videos. Semi-supervised
learning leverages the large amount of unlabeled data in developing effec-
tive classifiers. Feature subspace learning finds optimal feature subspaces
for representing data and helping classification. In this paper, we present
a novel algorithm, Locality Preserving Semi-supervised Support Vector
Machines (LPSSVM), to jointly learn an optimal feature subspace as well
as a large margin SVM classifier. Over both labeled and unlabeled data,
an optimal feature subspace is learned that can maintain the smoothness
of local neighborhoods as well as being discriminative for classification.
Simultaneously, an SVM classifier is optimized in the learned feature sub-
space to have large margin. The resulting classifier can be readily used to
handle unseen test data. Additionally, we show that the LPSSVM algo-
rithm can be used in a Reproducing Kernel Hilbert Space for nonlinear
classification. We extensively evaluate the proposed algorithm over four
types of data sets: a toy problem, two UCI data sets, the Caltech 101 data
set for image classification, and the challenging Kodak’s consumer video
data set for semantic concept detection. Promising results are obtained
which clearly confirm the effectiveness of the proposed method.
1 Introduction
Consider one of the central issues in semantic concept classification of images

and videos: the amount of available unlabeled test data is large and growing, but
the amount of labeled training data remains relatively small. Furthermore, the
dimensionality of the low-level feature space is generally very high, the desired
classifiers are complex and, thus, small sample learning problems emerge.
There are two primary techniques for tackling above issues. Semi-supervised
learning is a method to incorporate knowledge about unlabeled test data into
the training process so that a better classifier can be designed for classifying
test data [1], [2], [3], [4], [5]. Feature subspace learning, on the other hand, tries
to learn a suitable feature subspace for capturing the underlying data manifold
over which distinct classes become more separable [6], [7], [8], [9].

Semantic Concept Classification 271
One emerging branch of semi-supervised learning methods is graph-based

techniques [2], [4]. Within a graph, the nodes are labeled and unlabeled sam-
ples, and weighted edges reflect the feature similarity of sample pairs. Under the
assumption of label smoothness on the graph, a discriminative function f is of-
ten estimated to satisfy two conditions: the loss condition – it should be close to
given labels yL on the labeled nodes; and the regularization condition – it should
be smooth on the whole graph, i.e., close points in the feature space should have
similar discriminative functions. Among these graph-based methods, Laplacian
Support Vector Machines (LapSVM ) and Laplacian Regularized Least Squares
(LapRLS ) are considered state-of-the-art for many tasks [10]. They enjoy both
high classification accuracy and extensibility to unseen out-of-sample data.
Feature subspace learning has been shown effective for reducing data noise
and improving classification accuracy [6], [7], [8], [9]. Finding a good feature
subspace can also improve semi-supervised learning performance. As in classifi-
cation, feature subspaces can be found by supervised methods (e.g., LDA [8]),
unsupervised methods (e.g., graph-based manifold embedding algorithms [6],
[9]), or semi-supervised methods (e.g., generalizations of graph-based embedd-
ing by using the ground-truth labels to help the graph construction process [7]).
In this paper, we address both issues of feature subspace learning and semi-
supervised classification. We pursue a new way of feature subspace and classifier
learning in the semi-supervised setting. A novel algorithm, Locality Preserving
Semi-supervised SVM (LPSSVM ), is proposed to jointly learn an optimal feature
subspace as well as a large margin SVM classifier in a semi-supervised manner.
A joint cost function is optimized to find a smooth and discriminative feature
subspace as well as an SVM classifier in the learned feature subspace. Thus,
the local neighborhoods relationships of both labeled and unlabeled data can be
maintained while the discriminative property of labeled data is exploited. The
following highlight some aspects of the proposed algorithm:
1. The target of LPSSVM is both feature subspace learning and semi-supervised

classification. A feature subspace is jointly optimized with an SVM classifier so
that in the learned feature subspace the labeled data can be better classified
with the optimal margin, and the locality property revealed by both labeled and
unlabeled data can be preserved.
2. LPSSVM can be readily extended to classify novel unseen test examples.
Similar to LapSVM and LapRLS and other out-of-sample extension methods
[5], [10], this extends the algorithm’s flexibility in real applications, in contrast
with many traditional graph-based semi-supervised approaches [4].
3. LPSSVM can be learned in the original feature space or in a Reproducing
Kernel Hilbert Space (RKHS). In other words, a kernel-based LPSSVM is for-
mulated which permits the method to handle real applications where nonlinear
classification is often needed.
To evaluate the proposed LPSSVM algorithm, extensive experiments are car-
ried out over four different types of data sets: a toy data set, two UCI data sets
[11], the Caltech 101 image data set for image classification [12], and the large
scale Kodak’s consumer video data set [13] from real users for video concept
272 W. Jiang et al.
detection. We compare our algorithm with several state of the arts, including
the standard SVM [3], semi-supervised LapSVM and LapRLS [10], and the naive
approach of first learning a feature subspace (unsupervised) and then solving an
SVM (supervised) in the learned feature subspace. Experimental results demon-
strate the effectiveness of our LPSSVM algorithm.
2 Related Work
Assume we have a set of data points X = [x1, . . ., xn ], where xi is represented

by a d-dimensional feature vector, i.e., xi ∈ Rd . X is partitioned into labeled
subset XL (with nL data points) and unlabeled subset XU (with nU data points),
X=[XL , XU ]. yi is the class label of xi , e.g., yi ∈ {−1, +1} for binary classification.
2.1 Supervised SVM Classifier
The SVM classifier [3] has been a popular approach to learn a classifier based on
the labeled subset XL for classifying the unlabeled set XU and new unseen test
samples. The primary goal of an SVM is to find an optimal separating hyperplane
that gives a low generalization error while separating the positive and negative
training samples. Given a data vector x, SVMs determine the corresponding
label by the sign of a linear decision function f (x) = wT x+b. For learning non-
linear classification boundaries, a kernel mapping φ is introduced to project data
vector x into a high dimensional feature space as φ(x), and the corresponding
class label is given by the sign of f (x) = wT φ(x) + b. In SVMs, this optimal
hyperplane is determined by giving the largest margin of separation between
different classes, i.e. by solving the following problem:
1
1 nL
min Qd = min ||w||22 +C i , s.t. yi (wTφ(xi )+b) ≥ 1−i, i ≥ 0, ∀ xi ∈XL . (1)
w,b, w,b, 2
i=1
where = 1 , . . . , nL are the slack variables assigned to training samples, and C
controls the scale of the empirical error loss the classifier can tolerate.
2.2 Graph Regularization
To exploit the unlabeled data, the idea of graph Laplacian [6] has been shown
promising for both subspace learning and classification. We briefly review the
ideas and formulations in the next two subsections. Given the set of data points
X, a weighted undirected graph G = (V, E, W ) can be used to characterize the
pairwise similarities among data points, where V is the vertices set and each
node vi corresponds to a data point xi ; E is the set of edges; W is the set of
weights measuring the strength of the pairwise similarity.
Regularization for feature subspace learning. In feature subspace learning,
the objective of graph Laplacian [6] is to embed original data graph into an m-
dimensional Euclidean subspace which preserves the locality property of original
data. After embedding, connected points in original G should stay close. Let X̂
be the m×n dimensional embedding, X̂= [x̂1 , . . . , x̂n ], the cost function is:
⎧ ⎫
⎨ n ⎬ , -
min ||x̂i − x̂j ||22 Wij , s.t.X̂DX̂ T= I ⇒ min tr(X̂LX̂ T) , s.t.X̂DX̂ T = I.(2)
X̂ ⎩ i,j=1
⎭ X̂
where L is the Laplacian matrix and L = D−W , D is the diagonal weight matrix
whose entries are defined as Dii = j Wij . The condition X̂DX̂ T = I removes
an arbitrary scaling factor in the embedding [6]. The optimal embedding can be
obtained as the matrix of eigenvectors corresponding to the lowest eigenvalues
of the generalized eigenvalue problem: Lx̂=λDx̂. One major issue of this graph
embedding approach is that when a novel unseen sample is added, it is hard
to locate the new sample in the embedding graph. To solve this problem, the
Locality Preserving Projection (LPP ) is proposed [9] which tries to find a linear
projection matrix a that maps data points xi to aT xi , so that aT xi can best
approximate graph embedding x̂i . Similar to Eq(2), the cost function of LPP is:

mina Qs = mina tr(aT XLX T a) , s.t. aT XDX T a = I . (3)
We can get the optimal projection as the matrix of eigenvectors corresponding to

the lowest eigenvalues of generalized eigenvalue problem: XLX T a = λXDX T a.
Regularization for classification. The idea of graph Laplacian has been used
in semi-supervised classification, leading to the development of Laplacian SVM
and Laplacian RLS [10]. The assumption is that if two points xi , xj ∈ X are close
to each other in the feature space, then they should have similar discriminative
functions f (xi ) and f (xj ). Specifically the following cost function is optimized:
1 nL
min V (xi , yi , f ) + γA ||f ||22 + γI f T Lf . (4)
f nL i=1
where V(xi ,yi ,f ) is the loss function, e.g., the square loss V(xi ,yi ,f )=(yi−f(xi ))2
for LapRLS and the hinge loss V(xi ,yi ,f ) = max(0, 1 − yi f (xi )) for LapSVM;
f is the vector of discriminative functions over the entire data set X, i.e., f =
[f (x1 ), . . . , f (xnU +nL )]T . Parameters γA and γI control the relative importance
of the complexity of f in the ambient space and the smoothness of f according
to the feature manifold, respectively.
2.3 Motivation
In this paper, we pursue a new semi-supervised approach for feature subspace
discovery as well as classifier learning. We propose a novel algorithm, Locality
Preserving Semi-supervised SVM (LPSSVM ), aiming at joint learning of both an
optimal feature subspace and a large margin SVM classifier in a semi-supervised
manner. Specifically, the graph Laplacian regularization condition in Eq(3) is
adopted to maintain the smoothness of the neighborhoods over both labeled
and unlabeled data. At the same time, the discriminative constraint in Eq(1)
274 W. Jiang et al.
is used to maximize the discriminative property of the learned feature subspace

over the labeled data. Finally, through optimizing a joint cost function, the semi-
supervised feature subspace learning and semi-supervised classifier learning can
work together to generate a smooth and discriminative feature subspace as well
as a large-margin SVM classifier.
In comparison, standard SVM does not consider the manifold structure pre-
sented in the unlabeled data and thus usually suffers from small sample learn-
ing problems. The subspace learning methods (e.g. LPP) lack the benefits of
large margin discriminant models. Semi-supervised graph Laplacian approaches,
though incorporating information from unlabeled data, do not exploit the ad-
vantage of feature subspace discovery. Therefore, the overarching motivation
of our approach is to jointly explore the merit of feature subspace discovery
and large-margin discrimination. We will show through four sets of experiments
such approach indeed outperforms the alternative methods in many classification
tasks, such as semantic concept detection in challenging image/video sets.
3 Locality Preserving Semi-supervised SVM

In this section we first introduce the linear version of the proposed LPSSVM
technique then show it can be readily extended to a nonlinear kernel version.
3.1 LPSSVM
The smooth regularization term Qs in Eq(3) and discriminative cost function Qd
in Eq(1) can be combined synergistically to generate the following cost function:
5 nL 6
1
min Q = min {Qs + γQd} = min tr(aTXLX Ta)+γ[ ||w||22+C i ] (5)
a,w,b, a,w,b, a,w,b, 2 i=1
s.t. aT XDX T a = I, yi (wTaTxi +b) ≥ 1−i, i ≥ 0, ∀ xi ∈ XL .
Through optimizing Eq(5) we can obtain the optimal linear projection a and
classifier w, b simultaneously. In the following, we develop an iterative algorithm
to minimize over a and w, b, which will monotonically reduce the cost Q by
coordinate ascent towards a local minimum. First, using the method of Lagrange
multipliers, Eq(5) can be rewritten as the following:
5 6
1
min Q = min max tr(a XLX a)+γ[ ||w||2 −F (XLaw−B)+M ] , s.t.aTXDX Ta=I.
T T 2 T T
a,w,b, a,w,b, α,μ 2
where
nL we have nL defined quantities:nL F =[α1 y1 , . . . , αnL ynL]T , B= [b, . . . , b]T , M =
C i=1 i + i=1 αi (1−i )− i=1 μi i , and non-negative Lagrange multipliers α=
α1 , . . . , αnL , μ=μi , . . . , μnL . By differentiating Q with respect to w, b, i we get:
∂Q nL
=0⇒w= αi yi aT xi = aT XL F . (6)
∂w i=1
∂Q nL ∂Q
=0⇒ αi yi = 0, = 0 ⇒ C − αi − μi = 0 . (7)
∂b i=1 ∂i
Note Eq(6) and Eq(7) are the same as those seen in SVM optimization [3], with
the only difference that the data points are now transformed by a as x̃i = aT xi .
That is, given a known a, the optimal w can be obtained through the standard
SVM optimization process. Secondly, by substituting Eq(6) into Eq(5), we get:
, γ -
min Q = min tr(aTXLX Ta)+ F T XLT aaT XL F , s.t. aT XDX T a = I . (8)
a a 2
∂Q γ
= 0 ⇒ (XLX + XL F F T XLT )a = λXDX T a .
T
(9)
∂a 2
It is easy to see that XLX T + γ2 XL F F T XLT is positive semi-definite and we can

update a by solving the generalized eigenvalue problem described in Eq(9).
Combining the above two components, we have a two-step interative process
to optimize the combined cost function:
Step-1. With the current projection matrix at at the t-th iteration, train an
SVM classifier to get wt and α1,t , . . . , αnL ,t .
Step-2. With the current wt and α1,t , . . . , αnL ,t , update the projection matrix
at+1 by solving the generalized eigenvalue problem in Eq(9).
3.2 Kernel LPSSVM
In this section, we show that the LPSSVM method proposed above can be ex-
tended to a nonlinear kernel version. Assume that φ(xi ) is the projection function
which maps the original data point xi into a high-dimension feature space. Sim-
ilar to the approach used in Kernel PCA [14] or Kernel LPP [9], we pursue the
projection matrix a in the span of existing data points, i.e.,
n
a= φ(xi )vi = φ(X)v . (10)
i=1
where v = [v1 , . . . , vn ]T . Let K denote the kernel matrix over the entire
! data set
"
KL KLU
X = [XL , XU ], where Kij = φ(xi )·φ(xj ). K can be written as: K = ,
KUL KU
where KL and KU are the kernel matrices over the labeled subset XL and the
unlabeled subset XU respectively; KLU is the kernel matrix between the labeled
data set and the unlabeled data set and KUL is the kernel matrix between the
unlabeled data and the labeled data (KLU = KUL T
).
In the kernel space, the projection updating equation (i.e., Eq(8)) turns to:
, γ -
minQ = min tr(aTφ(X)LφT(X)a)+ F TφT(XL)aaTφ(XL)F , s.t.aTφ(X)DφT(X)a = I .
a a 2
By differentiating Q with respect to a, we can get:
γ
φ(X)LφT(X)a+ φ(XL)F F TφT(XL)a = λφ(X)DφT(X)a

2
γ LU|L
⇒ KLK + K F F T (K LU|L )T v = λKDKv . (11)
2
276 W. Jiang et al.
where K LU|L = [KLT , KUL

T T
] . Eq(11) plays a role similar to Eq(9) that it can be
used to update the projection matrix.
Likewise, similar to Eq(6) and Eq(7) for the linear case, we can find the
maximum margin solution in the kernel space by solving the dual problem:
nL 1 nL nL
Q̃dual
svm = αi − αi αj yi yj φT(xi)aaT φ(xj)
i=1 2 i=1 j=1
nL 1 nL nL %n &%n &
L|LU LU|L
= αi − αi αj yi yj Kig vg Kgj vg .
i=1 2 i=1 j=1 g=1 g=1
where K L|LU=[KL , KLU ]. This is the same with the original SVM dual problem
[3], except that the kernel matrix is changed from original K to:
% &% &
K̂ = K L|LU v vT K LU|L . (12)
Combining the above two components, we can obtain the kernel-based two-step
optimization process as follows:
Step-1: With the current projection matrix vt at iteration t, train an SVM to
get wt and α1,t , . . . , αnL ,t with the new kernel described in Eq(12).
Step-2: With the current wt , α1,t , . . . , αnL ,t , update vt+1 by solving Eq(11).
In the testing stage, given a test example xj (xj can be an unlabeled training
sample, i.e., xj ∈ XU or xj can be an unseen test sample), the SVM classifier
gives classification prediction based on the discriminative function:
7 n 87 n 8T
nL
nL L|LU
f (xj ) = w a φ(xj ) = αi yi φ(xi )aa φ(xj ) = αi yi
T T T
Kig vg K(xg , xj )vg .
i=1 i=1 g=1 g=1
Thus the SVM classification process is also similar to that of standard SVM [3],
with the difference that the kernel function between
# labeled
$# training $data and
test data is changed from K L|test to: K̂ L|test = K L|LU v vT K LU|test . v plays
the role of modeling kernel-based projection a before computing SVM.
3.3 The Algorithm

The LPSSVM algorithm is summarized in Fig. 1. Experiments show usually
LPSSVM converges within 2 or 3 iterations. Thus in practice we may set T = 3. γ
controls the importance of SVM discriminative cost function in feature subspace
learning. If γ = 0, Eq(11) becomes traditional LPP. In experiments we set γ = 1 to
balance two cost components. The dimensionality of the learned feature subspace
is determined by controlling the energy ratio of eigenvalues kept in solving the
eigenvalue problem of Eq(11). Note that in LPSSVM, the same Gram matrix is
used for both graph construction and SVM classification, and later (Sec.4) we will
see without extensive tuning of parameters LPSSVM can get good performance.
For example, the default parameter setting in LibSVM [15] may be used. This is
very important in real applications, especially for large-scale image/video sets.
Repeating experiments to tune parameters can be time and resource consuming.
Input: nL labeled data XL , and nU unlabeled data XU .

1 Choose a kernel function K(x, y), and compute Gram matrix Kij = K(xi , xj ), e.g.
RBF kernel K(xi ,xj ) = exp{−θ||xi −xj ||22} or Spatial Pyramid Match Kernel [16].
2 Construct data adjacency graph over entire XL ∪XU using kn nearest neighbors. Set
edge weights Wij based on the kernel matrix described in step 1.
3 Compute graph Laplacian matrix: L = D−W where D is diagonal, Dii = j Wij .
4 Initialization: train SVM over Gram matrix of labeled XL , get w0 and α1,0 , . . . , αnL ,0 .
5 Iteration: for t = 1, . . . , T
– Update vt by solving problem in Eq(11) with wt−1 and α1,t−1 , . . . , αnL ,t−1 .
– Calculate new kernel by Eq(12) using vt . Train SVM to get wt , α1,t , . . . , αnL ,t .
– Stop iteration if n i=1 (αi,t−1 − αi,t ) < τ .
L 2
Fig. 1. The LPSSVM algorithm
In terms of speed, LPSSVM is very fast in the testing stage, with complexity
similar to that of standard SVM classification. In training stage, both steps of
LPSSVM are fast. The generalized eigenvalue problem in Eq(11) has a time com-
plexity of O(n3 ) (n = nL+nU ). It can be further reduced by exploiting the sparse
implementation of [17]. For step 1, the standard quadratic programming opti-
mization for SVM is O(n3L ), which can be further reduced to linear complexity
(about O(nL )) by using efficient solvers like [18].
4 Experiments
We conduct experiments over 4 data sets: a toy set, two UCI sets [11], Caltech
101 for image classification [12], and Kodak’s consumer video set for concept de-
tection [13]. We compare with some state-of-the-arts, including supervised SVM
[3], semi-supervised LapSVM and LapRLS [10]. We also compare with a naive
LPP+SVM: first apply kernel-based LPP to get projection and then learn SVM
in projected space. For fair comparison, all SVMs in different algorithms use
RBF kernels for classifying UCI data, Kodak’s consumer videos, and toy data,
and use the Spatial Pyramid Match (SPM) kernel [16] for classifying Caltech
101 (see Sec.4.3 for details). This is motivated by the promising performance in
classifying Caltech 101 in [16] by using SPM kernels. In LPSSVM, γ = 1 in Eq(5)
to balance the consideration on discrimination and smoothness, and θ = 1/d in
RBF kernel where d is feature dimension. This follows the suggestion of the pop-
ular toolkit LibSVM [15]. For all algorithms, the error control parameter C = 1
for SVM. This parameter setting is found robust for many real applications [15].
Other parameters: γA , γI in LapSVM, LapRLS [10] and kn for graph construc-
tion, are determined through cross validation. LibSVM [15] is used for SVM,
and source codes from [17] is used for LPP.
4.1 Performance over Toy Data

We construct a “three suns” toy problem in Fig. 2. The data points with each
same color (red, blue or cyan) come from one category, and we want to separate
278 W. Jiang et al.
“three suns”toy data SVM classification result 10%

9% Standard SVM
LapSVM
8%
LapRLS
7%
Error Rate (%)

LPP+SVM
6% LPSSVM
5%
LapSVM classification result LPSSVM classification result
4%
3%
2%
1%
0%
5 10 15 20 25 30 35 40 45 50 55 60
Labeled Ratio (%)
Fig. 2. Performance over toy data. Compared with others, LPSSVM effectively dis-
criminates 3 categories. Above results are generated by using the SVM Gram matrix
directly for constructing Laplacian graph. With more deliberate tuning of the Lapla-
cian graph, LapSVM, LapRLS, and LPSSVM can give better results. Note that the
ability of LPSSVM to maintain good performance without graph tuning is important.
the three categories. This data set is hard since data points around the class
boundaries from different categories (red and cyan, and blue and cyan) are close
to each other. This adds great difficulty to manifold learning. The one-vs.-all
classifier is used to classify each category from others, and each test data is
assigned the label of the classifier with the highest classification score. Fig. 2
gives an example of the classification results using different methods with 10%
samples from each category as labeled data (17 labeled samples in total). The
averaged classification error rates (over 20 randomization runs) when varying the
number of labeled data are also shown. The results clearly show the advantage
of our LPSSVM in discriminative manifold learning and classifier learning.
4.2 Performance over UCI Data

This experiment is performed on two UCI data sets [11]: Johns Hopkins Iono-
sphere (351 samples with 34-dimension features), and Sonar (208 samples with
60-dimension features). Both data sets are binary classification problems. In Fig.
3 we randomly sample N points from each category (2N points in total) as la-
beled data and treat the rest data as unlabeled data as well as test data for
evaluation. The experiments are repeated for 20 randomization runs, and the
averaged classification rates (1 - error rates) are reported. From the result, our
LPSSVM consistently outperforms all other competing methods over different
numbers of labeled data in both data sets.
4.3 Performance over Caltech 101

The Caltech 101 set [12] consists of images from 101 object categories and an
additional background class. This set contains some variations in color, pose and
lighting. The bag-of-features representation [19] with local SIFT descriptors [20]
has been proven effective for classifying this data set by previous works [16]. In
this paper we adopt the SPM approach proposed in [16] to measure the image
similarity and compute the kernel matrix. In a straightforward implementation
Classification Rate (1-Error Rate)
Classification Rate (1-Error Rate)

Number of Labeled Data (2N) Number of Labeled Data (2N)
(a) Sonar (b) Johns Hopkins Ionosphere
Fig. 3. Classification rates over UCI data sets. The vertical dotted line over each point
shows the standard deviation over 20 randomization runs.
of SPM, only the labeled data is fed to the kernel matrix for standard SVM.
For other methods, the SPM-based measure is used to construct kernel matrices
for both labeled and unlabeled data (i.e., KL , KU , KLU ) before various semi-
supervised learning methods are applied. Specifically, for each image category,
5 images are randomly sampled as labeled data and 25 images are randomly
sampled as unlabeled data for training. The remaining images are used as novel
test data for evaluation (we limit the maximum number of novel test images
in each category to be 30). Following the procedure of [16], a set of local SIFT
features of 16 × 16 pixel patches are uniformly sampled from these images over
a grid with spacing of 8 pixels. Then for each image category, a visual codebook
is constructed by clustering all SIFT features from 5 labeled training images
into 50 clusters (codewords). Local features in each image block are mapped
to the codewords to compute codeword histograms. Histogram intersections are
calculated at various locations and resolutions (2 levels), and are combined to
estimate similarity between image pairs. One-vs.-all classifiers are built for classi-
fying each image category from the other categories, and a test image is assigned
the label of the classifier with the highest classification score.
Table 1 (a) and (b) give the average recognition rates of different algorithms
over 101 image categories for the unlabeled data and the novel test data, respec-
tively. From the table, over the unlabeled training data LPSSVM can improve
baseline SVM by about 11.5% (on a relative basis). Over the novel test data,
LPSSVM performs quite similarly to baseline SVM1 .
It is interesting to notice that all other competing semi-supervised meth-
ods, i.e., LapSVM, LapRLS, and naive LPP+SVM, get worse performance than
LPSSVM and SVM. Please note that extensive research has been conducted for
supervised classification of Caltech 101, among which SVM with SPM kernels
gives one of the top performances. To the best of our knowledge, there is no
report showing that the previous semi-supervised approaches can compete this
state-of-the-art SPM-based SVM in classifying Caltech 101. The fact that our
LPSSVM can outperform this SVM, to us, is very encouraging.
1
Note the performance of SPM-based SVM here is lower than that reported in [16].
This is due to the much smaller training set than that in [16]. We focus on scenarios
of scarce training data to access the power of different semi-supervised approaches.
280 W. Jiang et al.
Table 1. Recognition rates for Caltech 101. All methods use SPM to compute image
similarity and kernel matrices. Numbers shown in parentheses are standard deviations.
(a) Recognition rates (%) over unlabeled data

SVM LapSVM LapRLS LPP+SVM LPSSVM
30.2(±0.9) 25.1(±1.1) 28.6(±0.8) 14.3(±4.7) 33.7(±0.8)
(b) Recognition rates (%) over novel test data
SVM LapSVM LapRLS LPP+SVM LPSSVM
29.8(±0.8) 24.5(±0.9) 26.1(±0.8) 11.7(±3.9) 30.1(±0.7)
The reason other competing semi-supervised algorithms have a difficult time

in classifying Caltech 101 is because of the difficulty in handling small sam-
ple size in high dimensional space. With only 5 labeled and 25 unlabeled high
dimensional training data from each image category, curse of dimensionality
usually hurts other semi-supervised learning methods as the sparse data mani-
fold is difficult to learn. By simultaneously discovering lower-dimension subspace
and balancing class discrimination, our LPSSVM can alleviate this small sample
learning difficulty and achieve good performance for this challenging condition.
4.4 Performance over Consumer Videos
We also use the challenging Kodak’s consumer video data set provided in [13],
[21] for evaluation. Unlike the Caltech images, content in this raw video source
involves more variations in imaging conditions (view, scale, lighting) and scene
complexity (background and number of objects). The data set contains 1358
video clips, with lengths ranging from a few seconds to a few minutes. To avoid
shot segmentation errors, keyframes are sampled from video sequences at a 10-
second interval. These keyframes are manually labeled to 21 semantic concepts.
Each clip may be assigned to multiple concepts; thus it represents a multi-label
corpus. The concepts are selected based on actual user studies, and cover several
categories like activity, occasion, scene, and object.
To explore complementary features from both audio and visual channels, we
extract similar features as [21]: visual features, e.g., grid color moments, Gabor
texture, edge direction histogram, from keyframes, resulting in 346-dimension
visual feature vectors; Mel-Frequency Cepstral Coefficients (MFCCs) from each
audio frame (10ms) and delta MFCCs from neighboring frames. Over the video
interval associated with each keyframe, the mean and covariance of the audio
frame features are computed to generate a 2550-dimension audio feature vector
. Then the visual and audio feature vectors are concatenated to form a 2896-
dimension multi-modal feature vector. 136 videos (10%) are randomly sampled
as training data, and the rest are used as unlabeled data (also for evaluation). No
videos are reserved as novel unseen data due to the scarcity of positive samples
for some concepts. One-vs.-all classifiers are used to detect each concept, and
average precision (AP) and mean of APs (MAP) are used as performance metrics,
which are official metrics for video concept detection [22].
0.5
Standard SVM
0.45
LapSVM
0.4 LapRLS
0.35 LPP+SVM
Average Precision
0.3 LPSSVM
0.25
0.2
0.15
0.1
0.05
0
on
par ng
yg ic
pla picn k
dan ing
sun rts
on ni m
cro oat
b y
bir each
spo w
b y
we set
ade
ers t
P
gro _3+
muup_2
MA i
sho d
bab l
up d
e_p gh
sk
a
par
a
n
gro w
seu
im
thd
ci
rou
dd
an
Fig. 4. Performance over consumer videos: per-concept AP and MAP. LPSSVM gets
good performance over most concepts with strong cues from both visual and audio chan-
nels, where LPSSVM can find discriminative feature subspaces from multi-modalities.
Fig. 4 gives the per-concept AP and the overall MAP performance of different
algorithms2 . On average, the MAP of LPSSVM significantly outperforms other
methods - 45% better than the standard SVM (on a relative basis), 42%, 41% and
92% better than LapSVM, LapRLS and LPP+SVM, respectively. From Fig. 4,
we notice that our LPSSVM performs very well for the “parade” concept, with
a 17-fold performance gain over the 2nd best result. Nonetheless, even if we
exclude “parade” and calculate MAP over the other 20 concepts, our LPSSVM
still does much better than standard SVM, LapSVM, LapRLS, and LPP+SVM
by 22%, 15%, 18%, and 68%, respectively.
Unlike results for Caltech 101, here semi-supervised LapSVM and LapRLS
also slightly outperform standard SVM. However, the naive LPP+SVM still per-
forms poorly - confirming the importance of considering subspace learning and
discriminative learning simultaneously, especially in real image/video classifica-
tion. Examining individual concepts, LPSSVM achieves the best performance
for a large number of concepts (14 out of 21), with a huge gain (more than
100% over the 2nd best result) for several concepts like “boat”, “wedding”, and
“parade”. All these concepts generally have strong cues from both visual and
the audio channels, and in such cases LPSSVM takes good advantage of finding
a discriminative feature subspace from multiple modalities, while successfully
harnessing the challenge of the high dimensionality associated with the multi-
modal feature space. As for the remaining concepts, LPSSVM is 2nd best for 4
additional concepts. LPSSVM does not perform as well as LapSVM or LapRLS
for the rest 3 concepts (i.e., “ski”, “park”, and “playground”), since there are
no consistent audio cues associated with videos in these classes, and thus it is
difficult to learn an effective feature subspace. Note although for “ski” visual
2
Note the SVM performance reported here is lower than that in [21]. Again, this is
due to the much smaller training set than that used in [21].
282 W. Jiang et al.
0.35 0.24
LPSSVM
0.3 LPSSVM
Standard SVM
0.18 Standard SVM
0.25
0.2
AP
AP
0.12
0.15
0.1
0.06
0.05
0 0
10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Energy Ratio Energy Ratio
(a) parade (b) crowd
Fig. 5. Effect of varying energy ratio (subspace dimensionality) on the detection per-
formance. There exists a reasonable range of energy ratio that LPSSVM performs well.
features have consistent patterns, the performance may be influenced more by

high-dimension audio features than by visual features.
Intriguing by the large performance gain for several concepts like “parade”,
“crowd”, and “wedding”, we analyze the effect of varying dimensionality of the
subspace on the final detection accuracy. The subspace dimensionality is deter-
mined by the energy ratio of eigenvalues kept in solving the generalized eigen-
value problem. As shown in Fig. 5, even if we keep only 10% energy, LPSSVM
still gets good performance compared to standard SVM - 73% gain for “pa-
rade” and 20% gain for “crowd”. On the other hand, when we increase the
subspace dimensionality by setting a high energy ratio exceeding 0.7 or 0.8, the
performances start to decrease quickly. This further indicates that there exist
effective low-dimension manifolds in high-dimension multi-modal feature space,
and LPSSVM is able to take advantage of such structures. In addition, there
exists a reasonable range of energy ratio (subspace dimension) that LPSSVM
will outperform competing methods. How to automatically determine subspace
dimension is an open issue and will be our future work.
5 Conclusion
We propose a novel learning framework, LPSSVM, and optimization methods

for tackling one of the major barriers in large-scale image/video concept clas-
sification - combination of small training size and high feature dimensionality.
We develop an effective semi-supervised learning method for exploring the large
amount of unlabeled data, and discovering subspace structures that are not only
suitable for preserving local neighborhood smoothness, but also for discrimi-
native classification. Our method can be readily used to evaluate unseen test
data, and extended to incorporate nonlinear kernel formulation. Extensive ex-
periments are conducted over four different types of data: a toy set, two UCI
sets, the Caltech 101 set and the challenging Kodak’s consumer videos. Promis-
ing results with clear performance improvements are achieved, especially under
adverse conditions of very high dimensional features with very few training sam-
ples where the state-of-the-art semi-supervised methods generally tend to suffer.
Future work involves investigation of automatic determination of the opti-
mal subspace dimensionality (as shown in Fig. 5). In addition, there is another
way to optimize the proposed joint cost function in Eq(5). With relaxation
aTXDX Ta − I 0 instead of aTXDX Ta − I = 0, the problem can be solved
via SDP (Semidefinite Programming), where all parameters can be recovered
without resorting to iterative processes. In such a case, we can avoid the local
minima, although the solution may be different from that of the original problem.
References
1. Joachims, T.: Transductive inference for text classification using support vector
machines. In: ICML, pp. 200–209 (1999)
2. Chapelle, O., et al.: Semi-supervised learning. MIT Press, Cambridge (2006)
3. Vapnik, V.: Statistical learning theory. Wiley-Interscience, New York (1998)
4. Zhu, X.: Semi-supervised learning literature survey. Computer Sciences Technique
Report 1530. University of Wisconsin-Madison (2005)
5. Bengio, Y., Delalleau, O., Roux, N.: Efficient non-parametric function induction in
semi-supervised learning. Technique Report 1247, DIRO. Univ. of Montreal (2004)
6. Belkin, M., Niyogi, P.: Laplacian eigenmaps for dimensionality reduction and data
representation. Neural Computation 15, 1373–1396 (2003)
7. Cai, D., et al.: Spectral regression: a unified subspace learning framework for
content-based image retrieval. ACM Multimedia (2007)
8. Duda, R.O., et al.: Pattern classification, 2nd edn. John Wiley and Sons, Chichester
(2001)
9. He, X., Niyogi, P.: Locality preserving projections. Advances in NIPS (2003)
10. Belkin, M., Niyogi, P., Sindhwani, V.: Manifold regularization: A geometric frame-
work for learning from labeled and unlabeled examples. Journal of Machine Learn-
ing Research 7, 2399–2434 (2006)
11. Blake, C., Merz, C.: Uci repository of machine learning databases (1998),
http://www.ics.uci.edu/∼ mlearn/MLRepository.html
12. Li, F., Fergus, R., Perona, P.: Learning generative visual models from few training
examples: An incremental bayesian approach tested on 101 object categories. In:
CVPR Workshop on Generative-Model Based Vision (2004)
13. Loui, A., et al.: Kodak’s consumer video benchmark data set: concept definition
and annotation. In: ACM Int’l Workshop on Multimedia Information Retrieval
(2007)
14. Schölkopf, B., Smola, A., Müller, K.: Nonlinear component analysis as a kernel
eigenvalue problem. Neural Computation 10, 1299–1319 (1998)
15. Hsu, C., Chang, C., Lin, C.: A practical guide to support vector classification,
http://www.csie.ntu.edu.tw/∼ cjlin/libsvm/
16. Lazebnik, S., Schmid, C., Ponce, J.: Beyond bags of features: Spatial pyramid match-
ing for recognizing natural scene categories. In: CVPR, vol, 2, pp. 2169–2178
17. Cai, D., et al.: http://www.cs.uiuc.edu/homes/dengcai2/Data/data.html
18. Joachims, T.: Training linear svms in linear time. ACM KDD, 217–226 (2006)
19. Fergus, R., et al.: Object class recognition by unsupervised scale-invariant learning.
In: CVPR, pp. 264–271 (2003)
20. Lowe, D.: Distinctive image features from scale-invariant keypoints. International
21. Chang, S., et al.: Large-scale multimodal semantic concept detection for consumer
video. In: ACM Int’l Workshop on Multimedia Information Retrieval (2007)
22. NIST TRECVID (2001 – 2007),
http://www-nlpir.nist.gov/projects/trecvid/
Learning from Real Images to Model Lighting
Variations for Face Images
Xiaoyue Jiang1,2 , Yuk On Kong3 , Jianguo Huang1 ,

Rongchun Zhao1 , and Yanning Zhang1
1
School of Computer Science, Northwestern Polytechnical University,
Xi’an, 710072, China
2
School of Psychology, University of Birmingham, Birmingham B15 2TT, UK
3
Department of Electronics and Informatics, Vrije Universiteit Brussel,
Brussels 1050, Belgium
Abstract. For robust face recognition, the problem of lighting varia-

tion is considered as one of the greatest challenges. Since the nine points
of light (9PL) subspace is an appropriate low-dimensional approxima-
tion to the illumination cone, it yielded good face recognition results
under a wide range of difficult lighting conditions. However building the
9PL subspace for a subject requires 9 gallery images under specific light-
ing conditions, which are not always possible in practice. Instead, we
propose a statistical model for performing face recognition under vari-
able illumination. Through this model, the nine basis images of a face
can be recovered via maximum-a-posteriori (MAP) estimation with only
one gallery image of that face. Furthermore, the training procedure re-
quires only some real images and avoids tedious processing like SVD
decomposition or the use of geometric (3D) or albedo information of a
surface. With the recovered nine dimensional lighting subspace, recogni-
tion experiments were performed extensively on three publicly available
databases which include images under single and multiple distant point
light sources. Our approach yields better results than current ones. Even
under extreme lighting conditions, the estimated subspace can still rep-
resent lighting variation well. The recovered subspace retains the main
characteristics of 9PL subspace. Thus, the proposed algorithm can be
applied to recognition under variable lighting conditions.
1 Introduction
Face recognition is difficult due to variations caused by pose, expression, occlu-
sion and lighting (or illumination), which make the distribution of face object
highly nonlinear. Lighting is regarded as one of the most critical factors for robust
face recognition. Current attempt to handle lighting variation by either finding
the invariant features or modeling the variation. The edge based algorithm [1]
and the algorithm based on quotient image [2,3,4]belong to the first type. But
these methods cannot extract sufficient features for accurate recognition.
Early work on modeling lighting variation [5,6] showed that a 3D linear
subspace can represent the variation of a Lambertian object under a fixed

Learning from Real Images to Model Lighting Variations for Face Images 285
pose when there is no shadow. With the same Lambertian assumption, Bel-
humeur and Kriegman [7] showed that images illuminated by an arbitrary num-
ber of point light sources formed a convex polyhedral cone, i.e. the illumination
cone. In theory, the dimensionality of the cone is finite. They also pointed out
that the illumination cone can be approximated by a few properly chosen im-
ages. Good recognition results of the illumination cone in [8] demonstrated its
representation for lighting variation. [9] indicated that lighting subspace of Lam-
bertian object can be approximated by a linear subspace with dimension be-
tween three and seven. Recent research is mainly focused on the application
of low-dimensional subspace to lighting variation modeling. With the assump-
tion of Lambertian surface and non-concavity, Ramamoorith and Hanrahan [10]
and Basri and Jacobs[11] independently introduced the spherical harmonic (SH)
subspace to approximate the illumination cone. However, the harmonic images
(basis images of SH subspace) are computed from the geometric and albedo in-
formation of the subject’s surface. In order to use the SH subspace theory, a
lot of algorithms applied the 3D model of faces to handling lighting variations
[12,13,14,15,16]. However, recovering the 3D shape from images is still an open
problem in computer vision.
Lee et al.[19] built up a subspace that is nearest to the SH subspace and has
the largest intersection with the illumination cone, called the nine points of light
(9PL) subspace. It has a universal configuration for different subjects, i.e. the
subspace is spanned by images under the same lighting conditions for different
subjects. In addition, the basis images of 9PL subspace can be duplicated in
real environments, while those of the SH subspace cannot because its the basis
images contain negative values. Therefore the 9PL subspace can overcome the
inherent limitation of SH subspace. Since the human face is neither completely
Lambertain nor entirely convex, SH subspace can hardly represent the specular-
ities or cast shadows (not to mention inter-reflection). The basis images of 9PL
subspace are taken from real environment, they already contain all the compli-
cated reflections of the objects. Therefore the 9PL subspace can give a more
detailed and accurate description of lighting variation.
In practice, the requirement of these nine real images cannot always be ful-
filled. Usually there are fewer gallery images (e.g. one gallery image) per subject,
which can be taken under arbitrary lighting conditions. In this paper, we pro-
pose a statistical model for recovering the 9 basis images of the 9PL subspace
from only one gallery image. Zhang and Samaras [12] presented a statistical
method for recovering the basis images of SH subspace instead. In their training
procedure, geometric and albedo information is still required for synthesizing
the harmonic images. In contrast, the proposed method requires only some real
images that can be easily obtained in real environment. Since the recovered ba-
sis images of the 9PL subspace contain all the reflections caused by the shape
of faces, such as cast shadows, specularities, and inter-reflections, better recog-
nition results are obtained, even under extreme lighting conditions. Compared
with other algorithms based on 3D model [12,15,16], the proposed algorithm is
entirely a 2D algorithm, which has much lower computational complexity. The
286 X. Jiang et al.
proposed algorithm also has comparable recognition results. Note that we do

not consider pose variation in this paper and assume that all subjects are in the
frontal pose.
This paper is organized as follows. In Section 2, we briefly summarize the
methods of low-dimensional linear approximation of the illumination cone, in-
cluding the SH subspace and the 9PL subspace. The training of our statistical
model and the application of the model for recovering basis images from only
one gallery image are described in Sections 3 and 4 respectively. Section 5 is
dedicated to the experimental results. The conclusion is given in Section 6.
2 Approximation of the Illumination Cone

Belhumeur and Kriegman [7] proved that the set of n-pixel images of a con-
vex object that had a Lambertian surface illuminated by an arbitrary number
of point light sources at infinity formed a convex polyhedral cone, called the
illumination cone C in Rn . Each point in the cone is an image of the object
under a particular lighting condition, and the entire cone is the set of images of
the object under all possible lighting conditions. Any images in the illumination
cone C (including the boundary) can be determined as a convex combination of
extreme rays (images) given by
Iij = max(B̃s̃ij , 0) (1)
where s̃ij = b˜i × b˜j and B̃ ∈ !n×3 . Every row b˜i of B̃ is a three element row
vector determined by the product of the albedo with the inward pointing unit
normal vector of a point on the surface. There are at most q(q − 1) extreme rays
for q ≤ n distinct surface normal vectors. Therefore the cone can be constructed
with finite extreme rays and the dimensionality of the lighting subspace is finite.
However, building the full illumination cone is tedious, and the low dimensional
approximation of the illumination cone is applied in practice.
From the view of signal processing, the reflection equation can be considered
as the rotational convolution of incident lighting with the albedo of the surface
[10]. The spherical harmonic functions Ylm (θ, φ) are a set of orthogonal basis
functions defined in the unit sphere, given as follows,
Ylm (θ, φ) = Nlm Plm (cos θ) expimφ (2)

(l−m)!
where Nlm = 2l+1 4π (l+m)! , (θ, φ)is the spherical coordinate (θ is the elevation
angle, which is the angle between the polar axis and the z-axis with range 0 ≤
θ ≤ 180o , and φ is the azimuth angle with the range −180o ≤ φ ≤ 180o). Plm is
the associated Legendre function, and the two indices meet the conditions l ≥ 0
and l ≥ m ≥ −l. Then functions in the sphere, such as the reflection equation,
can be expanded by the spherical harmonic functions, which are basis functions
on the sphere. Images can be represented as a linear combination of spherical
harmonic functions. The first three order (l ≤ 3) basis can account for 99%
energy of the function. Therefore the first three order basis functions (altogether
9) can span a subspace for representing the variability of lighting. This subspace
is called the spherical harmonic (SH) subspace .
Good recognition results reported in [11] indicates that the SH subspace H
is a good approximation to the illumination cone C. Given the geometric infor-
mation of a face, its spherical harmonic functions can be calculated with Eq.(2).
These spherical harmonic functions are synthesized images, also called harmonic
images. Except the first harmonic image, all the others have negative values,
which cannot be obtained in reality. To avoid the requirement of geometric in-
formation, Lee et al.[19] found a set of real images which can also serve as a low
dimensional approximation to illumination cone based on linear algebra theory.
Since the SH subspace H is good for face recognition, it is reasonable to
assume that a subspace R close to H would be likewise good for recognition. R
should also intersect with the illumination cone C as much as possible. Hence a
linear subspace R which is meant to provide a basis for good face recognition
will also be a low dimensional linear approximation to the illumination cone C.
Thus subspace should satisfy the following two conditions [19]:
1. The distance between R and H should be minimized.
2. The unit volume (vol(C ∩ R)) of C ∩ R should be maximized ( the unit
volume is defined as the volume of the intersection of C ∩ R with the unit ball)
Note that C ∩R is always a subcone of C; therefore maximizing its unit volume
is equivalent to maximize the solid angle subtended by the subcone C ∩ R. If
{I˜1 , I˜2 , · · · , I˜k }are the basis images of R. The cone Rc ⊂ R is defined by I˜k ,

M
Rc = {I|I ∈ R, I = αk I˜k , αk ≥ 0} (3)
k=1
is always a subset of C ∩ R. In practice the subcone C ∩ R is taken as Rc and

the subtended angle of Rc is maximized. R is computed as a sequence of nested
linear subspace R0 ⊆ R1 ⊆ · · · ⊆ Ri ⊆ · · · ⊆ R9 = R, with Rk (k > 0) a linear
subspace of dimension i and R0 = ∅. First, EC denotes the set of (normalized)
extreme rays in the illumination cone C; and ECk denotes the set obtained by
deleting k extreme rays from EC, where EC0 = EC. With Rk−1 and ECk−1 ,
the sets ECk and Rk can be defined iteratively as follows:
dist(I, Rk−1 )
I˜k = arg max (4)
I∈ECk−1 dist(I, H)
where I˜k denotes the element in ECk−1 . Rk is defined as the space spanned by
Rk−1 and I˜k . ECk = ECk−1 \I˜k . The algorithm stops when R9 ≡ R is reached.
The result of Eq.(4) is a set of nine extreme rays that span R and there are
nine directions corresponding to these nine extreme rays. For different subjects,
the nine lighting directions are qualitatively very similar. By averaging Eq.(4)
of different subjects and maximizing this function as follows:

N
dist(I p , Rpk−1 )
I˜k = arg max (5)
I∈ECk−1
p=1
dist(I p , Hp )
288 X. Jiang et al.
where I p denotes the image of subject p taken under a single light source. H p
is the SH subspace of subject p. Rpk−1 denotes the linear subspace spanned by
images {I˜1p , · · · , I˜kp }of subject p. The universal configuration of nine light source
direction is obtained. They are (0, 0), (68, −90), (74, 108), (80, 52), (85, −42),
(85, −137), (85, 146), (85, −4), (51, 67)[14]. The directions are expressed in spher-
ical coordinates as pairs of (φ, θ), Figure 1(a) illustrates the nine basis images
of a person from the Yale Face Database B [8].
Fig. 1. the basis images of 9PL subspace. (a) images taken under certain lighting
conditions can serve as the basis images of the object. (b) the mean images of the basis
images estimated from the bootstrap data set.
3 Statistical Model of Basis Images
According to the universal configuration of lighting directions, we can apply nine

images taken under controlled environment to spanning the 9PL linear subspace.
However, even these nine images may not be available in some situations. Thus,
we propose a statistical method for estimating the basis images from one gallery
image. To build the statistical model, we must find the probability density func-
tion (pdf) of basis images and the pdf of the error term. Due to the limited
amount of the training data, we use the bootstrap method to estimate the statis-
tics of basis images. The recovering step is to estimate the corresponding basis
images from one single image of a novel subject under arbitrary lighting condi-
tions. For a given image, we first estimate its lighting coefficient. Then according
to the maximum a posteriori (MAP) estimation, we obtain an estimation of the
basis images. Finally, we apply the recovered subspace to face recognition. The
probe image is identified as the face whose lighting subspace is closest in distance
to the image.
Given nine basis images, we can reconstruct images under arbitrary lighting
conditions as follows,
I = Bs + e(s) (6)
where I ⊂ !d×1 is the image vector. B ⊂ !d×9 is the matrix of nine basis images,
every column of which is the vector of the basis image. s ⊂ !d×1 is the vector
of lighting coefficients which denotes the lighting conditions of the image. Error
term e(s) ⊂ !d×1 is related to the pixels’ position and lighting conditions.
For a novel image, we estimate its basis images through the maximum a
posterior (MAP) estimation. That is
BMAP = arg max P (B|I) (7)
B
According to the Bayes rule

P (I|B)P (B)
P (B|I) = (8)
P (I)
where P (I) is the evidence factor which guarantees that posterior probabilities
would sum to one. Then Eq.(7) can become
BMAP = arg max(P (I|B)P (B)) (9)
B
In order to recover basis images from an image with Eq.(9), one should know
the pdf of the basis images, i.e. P (B), and the pdf of the likelihood, i.e. P (I|B).
Assuming the error term of Eq.(6) is normally distributed with mean μe (s) and
variance σe2 (s), we can deduce that the pdf of the likelihood P (I|B) is also
Gaussian with mean Bs + μe (s) and variance σe2 (s) according to Eq.(6).
We assume that the pdf of the basis images B are Gaussians of means μB
and covariances CB as in [12,20]. The probability P (B) can be estimated from
the basis images in the training set. In our experiments, the basis images of 20
different subjects from the extented Yale face database B [8] are introduced to
the bootstrap set. Note that, the basis images of every subject are real images
which were taken under certain lighting conditions. The lighting conditions are
determined by the universal configurations of the 9PL subspace. The sample
mean μB and sample covariance matrix CB are computed. Figure 1(b) shows
the mean basis images, i.e. μB .
The error term e(s) = I − Bs models the divergence between the real image
and the estimated image which is reconstructed by the low dimensional subspace.
The error term is related to the lighting coefficients. Hence, we need to know
the lighting coefficients of different lighting conditions. In the training set, there
are 64 different images that taken under different lighting condition for every
subject. Under a certain lighting condition, we calculate the lighting coefficients
of every subject’s image, i.e. spk (the lighting coefficients of the pth subject’s image
under the lighting condition sk ). For a training image, its lighting coefficients
can be estimated by solving the linear equation I = Bs. The mean value of
different subjects’ lighting coefficients can be the estimated coefficients (s̄k ) for
N p
that lighting condition, i.e. s̄k = p=1 sk /N . Then, under a certain lighting
condition, the error term the of the p subject’s image is
th
ep (s̄k ) = Ikp − Bp s̄k (10)

where Ikp is the training image of the pth subject under lighting condition sk and
Bp is the basis images of the pth subject. Following the above assumption, we
estimate the mean μe (s̄k )and variance σe2 (s̄k ) of the error term.
290 X. Jiang et al.
4 Estimating the Basis Images
As described in the previous section, the basis images of a novel image can
be recovered by using the MAP estimation. Since the error term is related to
lighting condition, we need to estimate the lighting condition, i.e. the lighting
coefficients, of every image before calculating its basis images.
4.1 Estimating Lighting Coefficients
Lighting influences greatly the appearance of an image. Under similar illumina-

tion, images of different subjects will appear almost the same. The difference
between the images of the same subject under different illuminations is always
larger than that between the images of different subjects under the same illumi-
nation [21]. Therefore we can estimate the lighting coefficients of a novel image
with an interpolation method. The kernel regression is a smooth interpolation
method [22]. It is applied to estimating the lighting coefficients. For every train-
ing image, we have their corresponding lighting coefficients. For a novel image
In , its lighting coefficient is given by
M
wk spk
s = k=1
M
(11)
k=1 wk
[D(In , Ikp )]2

wk = exp(− ) (12)
2(σIkp )2
where D(In , Ikp ) = In − Ikp 2 is the L2 norm of the image distance. σIkp deter-
mines the weight of test image Ikp in the interpolation. In the training set, every
subject has 64 different images and there are altogether 20 different subjects.
Thus, for a novel image, there are 20 images with similar illumination. In our
experiment, we assign the farthest distance of these 20 images from the probe
image to σIkp . spk is the lighting coefficient of image Ikp .
4.2 Estimating the Error Term
The error term denotes the difference between the reconstructed image and the
real image. This divergence is caused by the fact that the 9PL subspace is the
low-dimensional approximation to the lighting subspace, and it only accounts for
the low frequency parts of the lighting variance. The statistics of the error under
a new lighting condition can be estimated from those of the error under known
illumination, i.e. μe (s̄k ), σe2 (s̄k ), also via the kernel regression method [20].
M
k=1 wk μe (s̄k )
μe (s) = M (13)
k=1 wk
M
k=1 wk σe (s̄k )
2
σe2 (s) = M (14)
k=1 wk
[D(s, s̄k )]2

wk = exp(− ) (15)
2[σs̄k ]2
where D(s, s̄k ) = s − s̄2 2 is the L2 norm of the lighting coefficient distance.
Like σIkp , σs̄k determines the weight of the error term related to the lighting
coefficients s̄k . Also, we assign the farthest lighting coefficient distance of these
20 images from the probe image to σs̄k .
4.3 Recovering the Basis Images

Given the estimated lighting coefficients s and the corresponding error term
μe (s), σe2 (s), we can recover the basis images via the MAP estimation. If we
apply the log probability, omit the constant term, and drop s for compactness,
Eq.(9) can become

1 I − Bs − μe 2 1 −1
arg max − ( ) − (B − μB )CB (B − μB ) T
(16)
B 2 σe 2
To solve Eq.(16), we estimate the derivatives,

2 −1
− (I − Bs − μe )sT + 2(B − μB )CB =0 (17)
σe2
Then we rewrite Eq.(17) as a linear equation,
AB = b (18)
−1 −1
where A = ss
T
σe2 + CB and b = σe2 s + CB μB .The solution of the linear

I−μe
equation is B = A−1 b. Using the Woodbury’s identity [25], we can obtain an

explicit solution
BMAP = A−1 b (19)

CB ssT CB I − μe −1
= CB − 2 s + CB μB
σe + sT CB s σe2

I − μB s − μe
= CB s + μB
σe2 + sT CB s
From
Eq.(19), the estimated
basis image is composed of the term of characteris-
I−μB s−μe
tics, σ2 +sT CB s CB s , and the term of mean, μB . In the term of characteristics,
e
(I − μB s − μe ) is the difference between the probe image and the image recon-
structed by the mean basis images.
4.4 Recognition
The most direct way to perform recognition is to measure the distance between
probe images and the subspace spanned by the recovered basis images. Every
column of B is one basis image. However, the basis images are not orthonormal
292 X. Jiang et al.
vectors. Thus we perform the QR decomposition on B to obtain a set of or-

thonormal basis, i.e. the matrix Q. Then the projection of probe image I to the
subspace spanned by B is QQT I, and the distance between the probe image I
and the subspace spanned by B can be computed as QQT I − I2 . In the recog-
nition procedure, the probe image is identified as the subspace with minimum
distance from it.
5 Experiments
The statistical model is trained by images from the extended Yale Face Database
B. With the trained statistical model, we can reconstruct the lighting subspace
from only one gallery image. This estimation is insensitive to lighting variation.
Thus, recognition can be achieved across illumination conditions.
5.1 Recovered Basis Images

To recover the basis images from a single image, the lighting coefficients of the
image are estimated first. Then we estimate the error terms of the image Finally,
the basis images of the image can be obtained with Eq.(19).
Fig. 2. Recovered basis images. (a)∼(d) are images in subset 1∼4 of Yale Face Database
B respectively. (e)∼(h) are recovered basis images from image (a)∼(d) respectively. (i)
are the reconstruction results: from left to right, the columns are the original im-
ages, the reconstruction results from the real basis images and the estimated basis
images(e)∼(h), respectively.
Although the images of the same object are under different lighting condi-
tions, the recovered basis images should be similar. The probe images are from
the Yale face database B. There are 10 subjects and 45 probe images per subject.
According to the lighting conditions of the probe images, they can be grouped
into 4 subsets as in [8]. The details can be found in Table 1. From subset1 to
subset4, the lighting conditions become extreme. For every subject, we recover
its basis images from only one of its probe images each time. Then we can ob-
tain 45 sets of basis images for every subject. Fig.2(e)∼(h) are the basis images
recovered from an image of each subset. σ̄basis (the mean standard deviation of
the 45 sets of basis images of 10 subjects) is 7.76 intensity levels per pixel, while
σ̄image (the mean standard deviation of the original 45 probe images of 10 sub-
jects) is 44.12 intensity levels per pixel. From the results, we can see that the
recovered basis images are insensitive to the variability of lighting. Thus we can
recover the basis images of a subject from its images under arbitrary lighting
conditions. Fig.2(i) are the reconstruction results from different basis images.The
reconstructed images also contain shadows and inter-reflections because the re-
covered basis images contain detailed reflection information. As a result, good
recognition results can be obtained.
Table 1. The subsets of Yale Face Database B
subset1 subset2 subset3 subset4

illumination 0∼12 13∼25 26∼50 50∼77
Number of images 70 120 120 140
5.2 Recognition
Recognition is performed on the Yale Face Database B [8] first. We take the
frontal images (pose 0) as the probe set, which is composed of 450 images (10
subjects, 45 images per subject). For every subject, one image is used for recov-
ering its lighting subspace and the 44 remaining images are used for recognition.
The comparison of our algorithm with the reported results is shown in Table 2.
Our algorithm reconstructed the 9PL subspace for every subject. The recov-
ered basis images also contained complicated reflections on faces, such as cast
shadows, specularities, and inter-reflection. Therefore the recovered 9PL sub-
space can give a more detailed and accurate description for images under differ-
ent lighting conditions. As a result, we can get good recognition results on images
with different lighting conditions. Also, the reported results of ’cone-cast’, ’har-
monic images-cast’ and ’9PL-real’ showed that better results can be obtained
when cast shadows were considered. Although [15,16] also use only one image to
adjust lighting conditions, they need to recover the 3D model of the face first.
The performance of our algorithm is comparable to that of these algorithms,
which are based on high-resolution rendering [15,16] and better than that of
those algorithms based on normal rendering [14]. Our algorithm is a completely
2D-based approach. Computationally, it is much less expensive compared with
294 X. Jiang et al.
Table 2. The Recognition Error Rate of Different Recognition Algorithms on Yale

Face Database B
Algorithms subset1&2 subset3 subset4

Correlation[8] 0.0 23.3 73.6
Eigenfaces[8] 0.0 25.8 75.7
Linear Subspace[8] 0.0 0.0 15.0
Cones-attached[8] 0.0 0.0 8.6
Cones-cast[8] 0.0 0.0 0.0
harmonic images-cast[8] 0.0 0.0 2.7
3D based SH model [12] 0.0 0.3 3.1
BIM(30 Bases)[15] 0.0 0.0 0.7
Wang et al.[16] 0.0 0.0 0.1
Chen et al.[17] 0.0 0.0 1.4
9PL-real[19] 0.0 0.0 0.0
our algorithm 0.0 0.0 0.72
those 3D based methods. The basis images of a subject can be directly com-
puted with Eq.19 while the recognition results are comparable to those from the
3D-based methods.
5.3 Multiple Lighting Sources

An image taken under multiple lighting sources can be considered as images
taken under a single lighting source being superimposed. Through interpolation,
the lighting coefficients of images taken under single lighting are linearly com-
bined to approximate those of the image taken under multiple-lighting. Here we
also apply the statistical model trained on the extended Yale Database B to ba-
sis images estimation. Similarly the lighting coefficients of images are estimated
through interpolation. Then the error term can be estimated according to the
lighting coefficients. Finally, the basis images are recovered.
In the PIE face database [23], there are 23 images per subject taken under
multiple lighting sources, and altogether 69 subjects. We recover 23 sets of the
basis images from the 23 images of every subject respectively. With these esti-
mated basis images, we perform recognition on the 1587 images (23 images per
person) 23 times. We also estimate basis for images in the AR database [24]. We
select randomly 4 images under different illumination per subject (image 1, 5,
6, 7) and recover the respective basis images from those images. Recognition is
performed on 504 images (126 subjects and 4 images per subject) 4 times. Sam-
ples of the recovered basis images from images in the PIE and AR databases
are shown in Fig.3. The average recognition rates, the mean standard deviation
of the recovered basis images (σ̄basis ) and the mean standard deviation of the
gallery images (σ̄images ) are presented in Table 3. Also [12] reported a recogni-
tion rate of 97.88% on part of PIE and [18] reported his recognition rate as 97%
on PIE database. Our recognition results are better. The results show that the
statistical model trained by images taken under a single lighting source can also
be generalized to images taken under multiple lighting sources.
Table 3. Recognition Rate on Different Databases
Face Database PIE AR

σ̄basis 11.01 11.34
σ̄image 285 38.59
Recognition rate 98.21% 97.75%
Fig. 3. Recovered basis images. (a) and (b) are images in PIE database, (e) and (f)
are estimated basis images from image (a) and (b), respectively. (c) and (d)are im-
ages in AR database, (g) and (h) are estimated basis images from image (c) and (d),
respectively.
6 Conclusion
The 9PL provides a subspace which is useful for recognition and is spanned by real
images. Based on this framework, we built a statistical model for these basis im-
ages. With the MAP estimation, we can recover the basis images from one gallery
image under arbitrary lighting conditions, which could be single lighting source
or multiple lighting sources. The experimental results based on the recovered sub-
space are comparable to those from other algorithms that require lots of gallery
images or the geometric information of the subjects. Even in extreme lighting con-
ditions, the recovered subspace can still appropriately represent lighting variation.
The recovered subspace retains the main characteristics of the 9PL subspace.
Based on our statistical model, we can build the lighting subspace of a sub-
ject from only one gallery image. It avoids the limitation of requiring tedious
training or complex training data, such as many gallery images or the geometric
information of the subject. After the model has been trained well, the computa-
tion for recovering the basis images is quite simple and without the need of 3D
models. The proposed framework can also potentially be used to deal with pose
and lighting variations together, with training images in different poses taken
under different lighting for building the statistical model.
Acknowledgement
This work is funded by China Postdoctoral Science Foundation(No.20070421129).
296 X. Jiang et al.
References
1. Guo, X., Leung, M.: Face recognition using line edge map. IEEE Trans. Pattern
Recognition and Machine Intelligence 24(6), 764–799 (2002)
2. Shashua, A., Tammy, R.: The quotient images: class-based rendering and recog-
nition with varying illuminations. IEEE Trans. Pattern Recognition and Machine
Intelligence 23(2), 129–139 (2001)
3. Gross, R., Brajovic, V.: An image processing algorithm for illumination invari-
ant face recognition. In: 4th International Conference on Audio and Video Based
Biometric Person Authentication, pp. 10–18 (2003)
4. Wang, H., Li, S.Z., Wang, Y.: Generalized quotient image. In: Proc. IEEE Conf.
Computer Vision and Pattern Recognition (2004)
5. Hallinan, P.: A low-dimensional representation of human faces for arbitrary lighting
conditions. In: Proc. IEEE Conf. Computer Vision and Pattern Recognition, pp.
995–999 (1994)
6. Nayar, S., Murase, H.: Dimensionality of illumination in appearance matching. In:
Proc. IEEE Conf. Robotics and Automation, pp. 1326–1332 (1996)
7. Belhumeur, P., Kriegman, D.J.: What is set of images of an object under all possible
lighting conditions? In: Proc. IEEE Conf. Computer Vision and Pattern Recogni-
tion, pp. 270–277 (1996)
8. Georghiads, A., Belhumeur, P., Kriegman, D.: From few to many: illumination cone
models for face recognition under variable lighting and pose. IEEE Trans. Pattern
Recognition and Machine Intelligence 23(6), 643–660 (2001)
9. Yuille, A., Snow, D., Epstein, R., Belhumeur, P.: Determing generative models of
objects under varying illumination: shape and albedo from multiple images using
SVD and integrability. International Journal of Computer Vision 35(3), 203–222
(1999)
10. Ramamoorthi, R., Hanrahan, P.: On the relationship between radiance and irra-
diance: determine the illumination from images of a convex Lambertian object. J.
Optical. Soc. Am. A 18(10), 2448–2459 (2001)
11. Basri, R., Jacobs, D.: Lambertian reflectance and linear subspaces. IEEE Trans.
Pattern Recognition and Machine Intelligence 25(2), 218–233 (2003)
12. Zhang, L., Samaras, D.: Face recognition under variable lighting using harmonic
image exemplars. In: Proc. IEEE Conf. Computer Vision and Pattern Recognition
(2003)
13. Wen, Z., Liu, Z., Huang, T.: Face relighting with radiance environment map. In:
Proc. IEEE Conf. Computer Vision and Pattern Recognition (2003)
14. Zhang, L., Wang, S., Samaras, D.: Face synthesis and recognition from a single im-
age under arbitrary unknown lighting using a spherical harmonic basis morphable
model. In: Proc. IEEE Conf. Computer Vision and Pattern Recognition (2005)
15. Lee, J., Moghaddam, B., Pfister, H., Machiraju, R.: A bilinear illumination model
for robust face recognition. In: Proc. IEEE International Conference on Computer
Vision (2005)
16. Wang, Y., Liu, Z., Hua, G., et al.: Face re-lighting from a single image under
harsh lighting conditions. In: Proc. IEEE Conf. Computer Vision and Pattern
Recognition (2007)
17. Chen, H.F., Belhumeur, P.N., Jacobs, D.W.: In search of illumination invariants.
In: Proc. IEEE Conf. Computer Vision and Pattern Recognition (2000)
18. Zhou, S., Chellappa, R.: Illuminating light field: image-based face recognition across
illuminations and poses. In: Proc. IEEE Intl. Conf. on Automatic Face and Gesture
Recognition (May 2004)
19. Lee, K., Ho, J., Kriegman, D.: Acquiring linear subspaces for face recognition under
variable lighting. IEEE Trans. Pattern Recognition and Machine Intelligence 27(5),
684–698 (2005)
20. Sim, T., Kanade, T.: Combining models and exemplars for face recognition: an
illumination example. In: Proc. Of Workshop on Models versus Exemplars in Com-
puter Vision, CVPR 2001 (2001)
21. Adini, Y., Moses, Y., Ullman, S.: Face recognition: the problem of compensating
for changes in illumination directions. IEEE Trans. Pattern Analysis and Machine
Intelligence 19(7), 721–733 (1997)
22. Atkenson, C., Moore, A., Schaal, S.: Locally weighted learning. Artificial Intelli-
gence Review (1996)
23. Sim, T., Baker, S., Bsat, M.: The CMU pose, illumination, and expression (pie)
database. In: Proc. IEEE International Conference on Automatic Face and Gesture
Recognition (May 2002)
24. Martinez, A.M., Benavente, R.: The AR face database. CVC Tech. Report No.24
(1998)
25. Scharf, L.: Statistical signal processing: detection, estimation and time series anal-
ysis, p. 54. Addison-Wesley, Reading (1991)
Toward Global Minimum through Combined
Local Minima
Ho Yub Jung, Kyoung Mu Lee, and Sang Uk Lee
Department of EECS, ASRI, Seoul National University, 151-742, Seoul, Korea

hoyub@diehard.snu.ac.kr, kyoungmu@snu.ac.kr, sanguk@ipl.snu.ac.kr
Abstract. There are many local and greedy algorithms for energy min-
imization over Markov Random Field (MRF) such as iterated condition
mode (ICM) and various gradient descent methods. Local minima so-
lutions can be obtained with simple implementations and usually re-
quire smaller computational time than global algorithms. Also, methods
such as ICM can be readily implemented in a various difficult problems
that may involve larger than pairwise clique MRFs. However, their short
comings are evident in comparison to newer methods such as graph cut
and belief propagation. The local minimum depends largely on the ini-
tial state, which is the fundamental problem of its kind. In this paper,
disadvantages of local minima techniques are addressed by proposing
ways to combine multiple local solutions. First, multiple ICM solutions
are obtained using different initial states. The solutions are combined
with random partitioning based greedy algorithm called Combined Lo-
cal Minima (CLM). There are numerous MRF problems that cannot be
efficiently implemented with graph cut and belief propagation, and so
by introducing ways to effectively combine local solutions, we present a
method to dramatically improve many of the pre-existing local minima
algorithms. The proposed approach is shown to be effective on pairwise
stereo MRF compared with graph cut and sequential tree re-weighted be-
lief propagation (TRW-S). Additionally, we tested our algorithm against
belief propagation (BP) over randomly generated 30×30 MRF with 2×2
clique potentials, and we experimentally illustrate CLM’s advantage over
message passing algorithms in computation complexity and performance.
1 Introduction
Recently, there are great interests in energy minimization methods over MRF.
The pairwise MRF is currently the most prominent MRF which became most
frequent subject of study in computer vision. Also, in the forefront, there is a
movement toward 2×2 and higher clique potentials for de-noising and segmenta-
tion problems [1,2,3,4,5,6]. They claim better performance through larger clique
potentials that can give more specified constraints.
However, the conventional belief propagation which has been so effective in
the pairwise MRF, is shown to have severe computational burden over large
cliques. In a factor graph belief propagation, the computational load increases
exponentially as the size of clique increases, although for the linear constraint

Toward Global Minimum through Combined Local Minima 299
MRFs, the calculation can be reduced to time linear [6,3]. Graph cut based
methods are also introduced for energy functions with global constraints and
larger clique potentials with pair-wise elements [5,4]. However, these methods
are targeted toward a specific category of energy functions and the applicability
limitations are high.
A practical and proven method for minimizing even the higher order MRFs
is simulated annealing. Gibbs sampler, generalized Gibbs sampler, data-driven
Markov chain Monte Carlo and Swendsen-Wang cut were respectively applied
to de-noising, texture synthesizing and segmentation problems that involved
large clique potentials [7,8,9,10]. However, simulated annealing is considered im-
practically slow compared to belief propagation and graph cut even in pairwise
MRFs [10,11]. More recently, simulated annealing has been modified by local-
ized temperature scheduling and additional window scheduling to increase its
effectiveness [12,13].
Another approach that is often being ignored is the greedy local minimum
algorithms. With the introductions of theoretically sound graph cut and belief
propagation over pairwise MRF, older methods such as ICM [14] and various
gradient descent methods are often disregarded as an under-performing alter-
natives [11]. However, methods like ICM and other local minimum algorithms
do not have any constraints over the size of cliques in MRF. Gradient descent
method was readily implemented over 5 × 5 and 3 × 2 clique potential in the de-
noising problem [2,1]. Texture synthesis and segmentation problems were model
by high order MRF and the energy was minimized using ICM [15]. Thus, when
considering both the computational time and performance, local greedy methods
that depend largely on the initial states are still viable in many of the high order
MRFs.
In this paper we propose a new algorithm to effectively combine these local
minima to obtain a solution that is closer to the global minimum state. First,
local solutions are calculated from various initial states. Then, they are com-
bined by random partitioning process such that the energy is minimized. The
proposed Combined Local Minima (CLM) approach is very simple but it can
effectively find lower energy state than graph cut and belief propagation. CLM
is tested on the pairwise stereo MRFs provided by [16,17,18,19,20], and it is
shown that the performance can be better than graph cut [21] and sequential
tree reweighted belief propagation (TRWS) [22]. We also performed tests over
randomly generated 2 × 2 clique MRFs, and showed that the proposed method
converges not only faster but finds lower energy state than belief propagation.
However, the biggest advantage of the proposed algorithm is that it can bring
further improvement over various local minima algorithms that are applicable
to general energy functions.
Section 2 will review ICM algorithm. Section 3 presents proposed CLM. In
the experiment section, CLM is shown to be competative over pairwise MRF
and superior over 2 × 2 clique MRF. The paper will close with conclusion and
possible future work.
300 H.Y. Jung, K.M. Lee, and S.U. Lee
(a) ICM 0 (b) ICM 5 (c) ICM 8
(d) ICM 10 (e) ICM 14 (f) CLM
Fig. 1. (a) to (e) show ICM solutions from different initial states. Homogeneous states
of disparity 0, 5, 8, 10, and 14 are respectively used as the initial states of (a), (b),
(c), (d), and (e). Combined local minima algorithm effectively combines these ICM
solutions into an lower energy state (f).
2 Iterated Conditional Mode (ICM)

For obtaining the local minima states, there are various different methods to
choose from. However, in this section, iterated conditional mode will be reviewed
for discrete MRF.
MRF consists of a set of nodes V = {v1 , v2 , ...vN }. For each nodes v ∈ V , a
label l can be assigned from a set L, producing a state x. The number of nodes
in V is denoted as N , and the number of labels in L is Q. In a discrete labelling
problem, the number of possible states will be QN . The energy function ϕ(x) is
a function of N dimension vector x = (x1 , x2 , x3 , ..., xN ).
ICM is a simple algorithm that determines the minimum energy at each ver-
tex v ∈ V . For a non-convex energies (such as pairwise energy function), ICM
produces a local minima solutions that depends upon the starting state. Follow-
ing pseudo code minimizes the energy function ϕ(x) in a labelling problem with
nodes vi ∈ {v1 , v2 , ...vN } and labels lj ∈ L = {l1 , l2 , ...lQ }.
Iterated Conditional Modes: ICM
1. Determine the initial state x
2. Repeat until no further energy minimization.
3. For i = 1 to i = N
4. For j = 1 to j = Q
5. Assign lj to vi
if ϕ(x1 , ..., xi = lj , .., xN ) < ϕ(x1 , ..., xN ).
The problem of choosing the right initial state is the big disadvantage of ICM.
Figure 1 (a) to (e) show the ICM solutions for Tsukuba stereo MRF. The solutions
in Figure 1 are obtained with different initial homogeneous states. Even though
the energy minimization cannot be low as graph cut or belief propagation, the
computational time is very small because the comparative inequality of step 5 can
be evaluated in O(1) for most of the energy functions, including pairwise functions.
ICM guarantees to converge but the performance is very poor as shown in figure 1.
Also, because of its simplicity, ICM can be applied to high order MRF with larger
cliques where graph cut and BP are having problems with.
3 Combined Local Minima

The simplest way to overcome the initial state dilemma of greedy algorithm is
to take multiple initial states. Among the multiple local minima obtained from
ICM, the lowest energy state can be chosen as the final solution. However, this
approach is problematic for MRF with very large dimensions, and obtaining
comparable solutions to graph cut and belief propagation is near impossible.
Thus, greedy algorithms are not often used for the MRF problems. In this sec-
tion, however, we assume that each local minima solution has a subset that is a
match to a subset of global minima state. We believe that a random partition
combination of local minima solution can be used to obtain energy level closer
to global minima.
3.1 Combined Space

In a typical labelling problems such as segmentation and stereo, the nodes are
presented by the pixel positions. The number of all possible states for such set
up will be QN . However, the combination of local minima will produce a smaller
space. In this section, the general notations will be defined for the proposed
algorithm.
The solution space for N number of nodes and set of labels L = {l1 , l2 , ..., lQ }
is Ω = {L×L×, ..., ×L}, where Ω is N dimension space. However, we are propos-
ing to minimize energy over reduced solution space that is obtained from the
combinations of local solutions. First, k number of local minima set {s1 , s2 , ..., sk }
are found using ICM such that each si is N dimension vector having following
labels.
si = ls1i , ls2i , ls3i , ..., lsNi . (1)
lsji is the label value for vj ∈ V node of si local minima state. ΩS ⊆ Ω is the
new solution space composed of the new sets of labels Lj ⊆ L.
ΩS = (L1 × L2 × L3 ×, ..., ×LN ). (2)
Lj is obtained from the set of the local solutions such that Lj = {lsj1 , lsj2 , lsj3 , ..., lsjk }.
The search for the minimum energy state will be over ΩS , although there
is no guarantee that the global minima is in the reduced space. Choosing the
right combinations of local minima for CLM will admittedly be heuristic for
each problem. More on the choices of local minima will be discussed in the later
sections. However, when the sufficient number and variety of local minima are
present in the proposed CLM, the solution space will be the original Ω.
3.2 Combined Local Minima

The proposed combinatorial algorithm for local minima is very simple and intu-
itive, however, it is shown to be very effective over traditional pairwise MRF and
randomly generated 2 × 2 clique MRF. For the pairwise MRF, QPBO algorithm
can effectively combine two minima solutions together [23]. However, QPBO al-
gorithm is viable only for pairwise MRF, thus we rely on random partitioning
technique which is simpler and can be applicable to higher order MRF.
We propose following algorithm to minimize energy from a set of local minima.
CLM partitions both current state and local minima states and replaces a part
of current state to one of the local minima states’ such that energy is reduced
for current state. It is a basic greedy algorithm over partitioned states.
Combined Local Minima: CLM
1. Given k number of local minima states from k different initial states,
s1 = ls11, ls21 , ls31 , ..., lsN1 , s2 = ls12 , ls22 , ls32 , ..., lsN2 ,
...,sk = ls1k , ls2k , ls3k , ..., lsNk .
and the current state x = lx1 , lx2 , lx3 , ..., lxN ,
repeat for specified number of iterations.
2. Randomly partition both the current state x and local minima states
s1 , s2, ..., sk into same m number of partitions such that
x = Vx1 , Vx2 , Vx3 , ..., Vxm ,s1 = Vs11, Vs21 , Vs31 , ..., Vsm 1
,
s2 = Vs12 , Vs22 , Vs32 , ..., Vsm 2
,...,sk = V 1
sk , V sk
2
, V 3
sk , ..., V sk .
m
3. Repeat for i = 1 to i = m.
4. Make k + 1 proposal states {x0 , x1 , x2 , ..., xk } in combinations
of current state x and s1 , ..., sk such that Vxi vector partition of
x is replaced 1by 2the 3Vs of local
i
minima 1states. See below.
x0 = x = V
1 x2 x i x , V , V , ..., V m
x , x 1 = x 1 x 2 Vs1i, ..., Vx m ,
V , V 2
, ..., i m
x2 = Vx , Vx , ..., Vs2 , ..., Vx ,.., xk = Vx , Vx , ..., Vsk , ..., Vx .

m
Among set S = {x0 , x1 , ..., xk }, take the lowest energy state as the
current state.
The computational complexity of CLM depends largely on the complexity of
evaluating ϕ(xi ). If ϕ(xi ) is needed to be calculated in O(N ), ICM’s complexity
will be O(kmN ). If m is randomly chosen, the worst case would be for m = N ,
and the time complexity will be O(kN 2 ) per iteration. However, if the maximum
clique size is small compared to MRF size, both the worst and best complexity
will be O(kN ) because only V i and areas around V i are needed to be evaluated
to find the lowest energy among S = {x0 , x1 , ..., xk }. Also, the complexity can
still be lowered using various computation techniques such as integral image
method [24].
The proposed algorithm is greedy and guarantees that the energy does not
increase for each iteration. Figure 2 shows the iterative results of the proposed
(a) Initial state (b) Iteration 1 (c) Iteration 2
(d) Iteration 3 (e) Iteration 4 (f) Iteration 6
Fig. 2. (a) shows the initial state of CLM. (b), (c), (d), (e), and (f) show respectively
the first, second, third, fourth, and sixth iterations of combined local minima algorithm.
CLM over Tsukuba stereo pair MRF. k = 16 number of local minima were
used. Few of local minima are shown in Figure 1. With only a small number
of iterations, CLM can output energy minimization result far superior to ICM
method, and with enough iterations it can be effective as the message passing
and graph cut algorithms.
However, there are two heuristics that must be resolved for CLM. First, it is
unclear how current state x and {s1 , s2 , ..., sk } should be randomly partitioned
in step 2 of the algorithm. Second, the choice of local minima and the value of k
are subject to question. These two issues are important to the performance of the
proposed algorithm and the basic guidelines are provided in next subsections.
3.3 Obtaining k Local Minima

It is intuitive to assert that if large number of local minima is used for CLM,
the obtained energy will be lower. However, for the price of lower energy, more
computational time and memory are required. The right tradeoff between com-
putation resources and desired energy level is essential to CLM. This is both
advantage and disadvantage of proposed method because by using CLM, you
can control the level of performance and computing resources.
Another factor that contributes to the performance of CLM is the variety of
local minima. For example, if all the local minima solutions are same, the energy
will not be lowered no matter how many times they are combined. Usually, vari-
ety of initial states for ICM result in the variety of minima solutions. However,
some heuristics may be needed for obtaining different local minima. We have
empirically developed few precept for both of these issues.
Thus, in order to have different local minima states, ICM with different ho-
mogeneous initial states were used. See experimental section and Figure 4 and 5.
In both of the comparison tests, the number of local minima are set to Q, the
number of labels. {s1 , ..., sQ } are obtained from ICM with homogeneous initial
state, respectively having labels l1 , l2 , ..., lQ . In both stereo MRF and randomly
generated MRF, such initial states resulted in the energy minimization compa-
rable to message passing algorithms. Thus, the rule of thumb is to use Q number
of local minima derived from the respective homogeneous initial states.
However, by increasing the number of local minima as shown in Figure 5,
much lower energy can be achieved with incremental addition to computation
time. In Figure 5, CLM200 minimizes energy using total of 200 local minima
composed of Q homogeneous initial states and 200 − Q number of ICM solutions
obtained from random initial states. CLM200 achieves much lower energy than
belief propagation. Although, random initial states are used here, more adaptive
initial states can also be applied for different problems.
3.4 Random Partition

In this paper, we use rectangular partition method for step 2 of CLM algorithm,
much like window annealing of [13]. See Figure 3 (a). 4 integers are randomly
chosen, and MRF can be partitioned accordingly. Such method is used because of
simplicity of computation and the fact that it can commodate the square lattice
structure of digital images. Furthermore, by having rectangular partitions, the
energy value of state can be obtained very fast using integral image technique
[24], which was used for the stereo pair experiment. However, the integral image
technique is not essential to the CLM. In Figure 5, integral image technique is
not used during the operations of CLM, and it has superior performance over
belief propagation.
For MRFs with random structure, rectangular partition can not be applied.
A possible random partitioning algorithm that can be used is the one that was
applied in Swendsen-Wang cut algorithm [10]. In Swendsen-Wang cut, the edges
between the nodes are probabilistically cut, and the connected nodes after the ran-
dom cut would make a single cluster. This method was not used in the experimen-
tation section because of needless complexity over square lattice MRF. Again, the
partition method can be specified to each problem at hand. However, V i should
be no larger than N obviously, and there should be a positive probability that size
of V i could be 1, so that the optimization can be occur over single nodes.
4 Experiments
In order to show the effectiveness of the proposed CLM, we compared it’s per-
formance with graph cut and TRW-S over pairwise stereo MRF. Additionally,
window annealing (WA) [13] results are included in the test. Pairwise stereo
MRF is known to be effectively optimized by alpha expansion graph cut (GC)
and TRW-S [21,22], but very ill posed for greedy algorithms such as ICM. The
experiments were performed over stereo pairs provided by [16,17,18,20,19].
V1 V2
Vm
(a) Rectangular Partition (b) 2 × 2 clique MRF
Fig. 3. (a) shows an example of rectangular partitioning of a square lattice MRF. A

state can be partitioned into rectangular clusters V = {V 1 , V 2 , ..., V m }. Such partition
method allows simple calculation of energy function by integral image technique which
was used for the pairwise MRF test. In this Figure (b), 4 × 4 MRF with 2 × 2 clique
potentials is depicted. The circle nodes are v ∈ V . The square factor nodes define the
cliques of MRF by connected the neighbors, Ng . MRF is built by assigning random
clique potentials from an uniform distribution. In the randomly generated 2 × 2 clique
MRF, integral image technique is not used for computational speed up.
Also recently, larger than pairwise clique models are often proposed for vision
problems. Gradient descent and belief propagation are used over 2 × 2 and larger
clique MRF to attack such problems as de-noising, and shape from shading
[6,1,2,3]. Thus, we tested our algorithm over randomly generated MRF with
2 × 2 clique potentials, see Figure 5 (a). Alpha expansion algorithms cannot
deal with randomly generated larger than pairwise MRF, and it was excluded
from the test. CLM reaches a lower energy faster than belief propagation (BP)
and WA methods. The computational complexity of proposed method is O(kN ),
allowing CLM to be a practical minimization scheme over large clique MRFs.
All computations are done over a 3.4GHz desktop.
4.1 Pair-Wise Stereo MRF

Pairwise and sumodular MRF is most common MRF used in computer vision.
Also, it has been the subject to many comparative tests. Particularly, the stereo
MRF has been an frequent in comparison tests of energy minimization meth-
ods [11,25,22]. However, the performance differences between two state of art
methods, graph cut and message passing algorithms, are still not clear when
the computational time is an issue. Although, TRW-S may eventually find lower
energy than graph cut, it can take many more iterations to do so. In some cases,
TRW-S is faster and finds lower energy than graph cut. In this test, we tried to
present energy functions that are fair to both graph cut and TRW-S. As shown in
Figure 4, for Cones and Bowling2 MRF, TRW-S clearly outperforms the graph
cut. Otherhand, for Teddy and Art MRF, the graph cut finds lower energy
much faster. The performance of each methods seem to depend largely upon the
strength of discontinuity costs. Simulated annealing, otherhand, depends large
on the temperature scheduling. Awhile WA is competitive with previous methods
in speed, usually it could not find lower energy. Although it is possible to tweak
725000 1e+06
CLM 60 CLM 60
720000 WA WA
Graph Cut 980000 Graph Cut
715000 TRW-S TRW-S
Low Bound Low Bound
960000
710000
705000 940000
Energy
Energy
700000
920000
695000
900000
690000
685000 880000
680000
860000
675000
0 20 40 60 80 100 0 20 40 60 80 100
Time(seconds) Time(seconds)
(a) Cones - truncated quadratic (b) Teddy - Potts

760000 1.25e+06
CLM 67 CLM 76
WA WA
Graph Cut Graph Cut
740000 TRW-S TRW-S
Low Bound 1.2e+06 Low Bound
720000
Energy
Energy
1.15e+06
700000
1.1e+06
680000
1.05e+06
660000
0 50 100 150 200 250 300 350 0 20 40 60 80 100 120 140
(c) Bowling2 - truncated linear (d) Art - Potts
Fig. 4. (a)Cones uses truncated quadratic discontinuity cost. (c) Bowling2 is the result
for truncated linear discontinuity cost. (b) Teddy and (d) Art use Potts discontinu-
ity cost. CLM 60 means 60 local minima are used in CLM algorithm. The CLM’s
performance is shown to be in-between TRW-S and GC. The performance difference
to state-of-art methods are very small, however, CLM performance does not seem to
strongly vary according to the discontinuity model apposed to TRW-S and graph cut.
the annealing scheduling for lower minimization, we kept the same temperature
and window scheduling of [13].
For the energy function, gray image Birchfield and Tomasi matching costs [26]
and Potts, truncated linear and truncated quadratic discontinuity cost are used.

ϕ(x) = D(p) + V (p, q). (3)
p∈V (p,q)∈Ng
D(p) is a pixel-wise matching cost between left and right image. V (p, q) is
pair-wise discontinuity costs. The implementations of graph cut and TRW-S
by [11,21,27,28,29,22] are used in this experiment.
For the implementation of CLM, Q number of local minima ICM solutions
are obtained from following set of initial states {(0, 0, ..., 0), (1, 1, ..., 1), ..., (Q −
1, Q − 1, ..., Q − 1)}. As mentioned before, a rule of thumb seems to be Q number
of local minima with homogeneous initial states, especially if the MRF is known
to have smoothness constraint. For the state partition technique of step 2 of
CLM, a simple rectangular partitioning method is used, see Figure 3.
500 500
CLM 2 CLM 3
CLM 200 CLM 200
450 WA 450 WA
ICM ICM
BP 400 BP
400
350
350
Energy
Energy
300
300
250
250
200
200 150
150 100
0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 2 4 6 8 10
(a) Q = 2 (b) Q = 3
500 500
CLM 4 CLM 5
CLM 200 CLM 200
450 WA 450 WA
ICM ICM
400 BP 400 BP
350 350
Energy
Energy
300 300
250 250
200 200
150 150
100 100
50 50
0 5 10 15 20 25 30 35 40 45 50 0 50 100 150 200 250
(c) Q = 4 (d) Q = 5
Fig. 5. Energy versus time results of max product BP, ICM, WA, and CLM over
30 × 30 randomly generated MRF. Figure (a), (b), (c), and (d) respectively have label
size Q = 2, Q = 3, Q = 4, and Q = 5. CLM using k = Q and k = 200 number of
local minima are performed for each random MRF. The increase in the local minima
allows lower energy state to be achieved in exchange for computation time and memory.
However, such price is very small compared to the computation time of BP.
Figure 4 shows energy vs time graph results using Potts, truncated linear,
truncated quadratic discontinuity model. Qualitatively, there is a very small
difference between TRW-S, graph cut, WA, and CLM, see Figure 6. However,
the energy versus time graphs show more edifying comparison. The first iteration
of the CLM takes much longer time than the other iterations because all the local
solutions are needed to be computed. Overall performance of the proposed CLM
stands in the middle of graph cut and TRW-S. However, compared with window
annealing, CLM outperforms it everywhere except for the initial calculations.
4.2 Randomly Generated 2 × 2 Clique MRF

However, the biggest advantage of proposed CLM is that the computational
complexity does not increase exponentially. Belief propagation based methods,
however, the time complexity of message calculation goes up exponentially as
the number of clique size increases [6]. In this section, the proposed CLM is
tested over randomly generated 2 × 2 clique MRF. Below equation describes the
energy function as sum of clique potentials.
(a-1) Cones (a-2) Teddy (a-3) Bowling2 (a-4) Art
(b-1) CLM (b-2) CLM (b-3) CLM (b-4) CLM
(c-1) WA (c-2) WA (c-3) WA (c-4) WA
(d-1) Graph cut (d-2) Graph cut (d-3) Graph cut (d-4) Graph cut
(e-1) TRW-S (e-2) TRW-S (e-3) TRW-S (e-4) TRW-S
Fig. 6. This figure shows the qualitative stereo energy minimization results at roughly
at same computation time. (a-1) to (a-4) are left reference stereo images. (b-1) to (b-
4) are the results of proposed CLM. (c), (d), and (e) respectively show the results
of window annealing, graph cut and TRW-S. For each stereo pair, the same energy
function is used. The qualitative differences between 4 methods are very small, except
for Teddy image where graph cut’s lower energy makes a difference over the roof area of
image. Otherwise, the energy difference between 4 methods are small enough to make
no visible differences.

ϕ(x) = V (p, q, r, s) (4)
(p,q,r,s)∈Ng
The clique potential V (p, q, r, s) is a function of 4 dimensioned vector. The

value of each V (p, q, r, s) is randomly assigned from an uniform distribution
[0, 1]. In Figure 3 (b), the square factor nodes are connected to 4 variable nodes
p, q, r, s. 30 × 30 variable nodes with 2 × 2 clique potentials are generated for
the comparison test. The energy minimization results of CLM, ICM, and BP are
shown in Figure 5; (a) to (d) are the results obtained for MRFs with label size
Q = 2 to Q = 5, respectively.
For the implementation of belief propagation, the factor nodes are transformed
into variable nodes with Q4 number of labels having corresponding V (p, q, r, s)
as the unary costs. The pairwise potentials are assigned either 0 or ∞ based on
the consistency requirement. CLM is implemented using k = Q local minima
and also for k = 200 local minima. For k = Q, local minima are found by ICM
over Q homogeneous states like the stereo problem, even though the smoothness
assumption is no longer viable in this problem. For CLM200, additional 200 − Q
ICM minima obtained from random initial states are used. Same rectangular
partitioning is used but the integral image technique is not used.
In these 4 tests, it is clear that the proposed CLM converges faster than
BP and WA. The difference from BP is more evident for Figure 5 (c) and (d),
because even though BP is fast as CLM200 for Q = 2 Figure 5 (a), but as
the label size increases, BP could not keep up with speed of CLM. Thus, as
the number of labels and clique size become larger, message passing algorithms
will become practically ineffective awhile proposed CLM can maintain reasonable
computational time. Furthermore, with larger number of local minima, the CLM
can reach much lower energy than BP and WA with comparably insignificant
addition to computation resources.

In this paper, we propose a new a method to combine local minima solutions
toward more global minimum by random partition method. CLM’s performance
is compared with state-of-art energy minimization methods over most well known
pairwise stereo MRF. Combined local minima is shown to be effective as graph
cut and TRW-S. Furthermore, tests over randomly generated 2 × 2 clique MRFs
show that the computation complexity of CLM is much smaller than traditional
message passing algorithms as the clique and label size become larger.
Additionally, we included window annealing method in the experiment. How-
ever, due to heuristics of simulated annealing and the proposed method, it is hard
to say which method is better. Nevertheless, both algorithms show clear advantages
over the high ordered MRF compared to existing methods awhile maintaining com-
petitiveness in the pairwise MRFs. We hope that such conclusion will encourage
other computer vision researchers to explore more complex MRFs involving larger
clique potentials. In the future, MRF with random structure (non square lattice)
will be studied using Swendsen Wang cut like partition method.
Acknowledgement
This research was supported in part by the Defense Acquisition Program Admin-
istration and Agency for Defense Development, Korea, through the Image Infor-
mation Research Center under the contract UD070007AD, and in part by the
MKE (Ministry of Knowledge Economy), Korea under the ITRC (Information
Technolgy Research Center) Support program supervised by the IITA (Institute
of Information Technology Advancement) (IITA-2008-C1090-0801-0018).
References
1. Roth, S., Black, M.J.: Steerable random fields. In: ICCV (2007)
2. Roth, S., Black, M.J.: Field of experts: A framework for learning image priors. In:
CVPR (2005)
3. Potetz, B.: Efficient belief propagation for vision using linear constraint nodes. In:
CVPR (2007)
4. Kohli, P., Mudigonda, P., Torr, P.: p3 and beyond: Solving energies with higher
order cliques. In: CVPR (2007)
5. Rother, C., Kolmogorov, V., Minka, T., Blake, A.: Cosegmenation of image pairs
by histogram matching- incorporating a global constraint into mrfs. In: CVPR
(2006)
6. Lan, X., Roth, S., Huttenlocher, D., Black, M.J.: Efficient belief propagation with
learned higher-order markov random fields. In: Leonardis, A., Bischof, H., Pinz, A.
(eds.) ECCV 2006. LNCS, vol. 3952, pp. 269–282. Springer, Heidelberg (2006)
7. Geman, S., Geman, D.: Stochastic relaxation, gibbs distributions, and the bayesian
restoration of images. PAMI 6 (1984)
8. Zhu, S.C., Liu, X.W., Wu, Y.N.: Exploring texture ensembles by efficent markov
chain monte carlo: Toward a trichromacy theory of texture. PAMI 22(6) (2000)
9. Tu, Z., Zhu, S.C.: Image segmentation by data-driven markov chain monte carlo.
PAMI 24 (2002)
10. Barbu, A., Zhu, S.C.: Generalizing swendsen-wang cut to sampling arbitrary pos-
terior probabilities. PAMI 27 (2005)
11. Szeliski, R., Zabih, R., Scharstein, D., Veksler, O., Kolmogorov, V., Agarwala, A.,
Tappen, M., Rother, C.: A comparative study of energy minimization methods for
markov random fields. In: Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006.
LNCS, vol. 3952, pp. 16–29. Springer, Heidelberg (2006)
12. Woodford, O.J., Reid, I.D., Torr, P.H.S., Fitzgibbon, A.W.: Field of experts for
image-based rendering. BMVC (2006)
13. Jung, H.Y., Lee, K.M., Lee, S.U.: Window annealing over square lattice markov
random field. ECCV (2008)
14. Besag, J.: On the statistical analysis of dirty pictures (with discussion). Journal of
the Royal Statistical Society Series B 48 (1986)
15. Mignotte, M.: Nonparametric multiscale energy-based model and its application in
some imagery problems. PAMI 26 (2004)
16. http://vision.middlebury.edu/stereo/
17. Scharstein, D., Szeliski, R.: A taxonomy and evaluation of dense two-frame stereo
correspondence algorithms. In: IJCV (2002)
18. Scharstein, D., Szeliski, R.: High-accuracy stereo depth maps using structured light.
In: CVPR (2003)
19. Hirshmuller, H., Szeliski, R.: Evaluation of cost functions for stereo matching. In:
CVPR (2007)
20. Scharstein, D., Pal, C.: Learning conditional random fields for stereo. In: CVPR
(2007)
21. Boykov, Y., Veksler, O., Zabih, R.: Fast approximate energy minimization via
graph cuts. PAMI 23 (2001)
22. Kolmogorov, V.: Convergent tree-reweighted message passing for energy minimiza-
tion. PAMI 28 (2006)
23. Lempitsky, V., Rother, C., Blake, A.: Logcut - efficient graph cut optimization for
markov random fields. In: ICCV (2007)
24. Crow, F.: Summed-area tables for texture mapping. SIGGRAPH (1984)
25. Tappen, M.F., Freeman, W.T.: Comparison of graph cuts with belief propagation
for stereo, using identical mrf parameters. In: ICCV (2003)
26. Birchfield, S., Tomasi, C.: A pixel dissimilarity measure that is insensiitive to image
samplin. PAMI 20 (1998)
27. Kolmogorov, V., Zabih, R.: What energy functions can be minimized via graph
cuts? PAMI 26 (2004)
28. Boykov, Y., Kolmogorov, V.: An experimental comparison of min-cut/max-flow
algorithms for energy minimization in vision. PAMI 26 (2004)
29. Wainwright, M.J., Jaakkola, T.S., Willsky, A.S.: Map estimation via agreement on
trees: Message-passing and linear-programming approaches. IEEE Trans. Informa-
tion Theory 51(11) (2005)
Differential Spatial Resection - Pose Estimation
Using a Single Local Image Feature
Kevin Köser and Reinhard Koch
Institute of Computer Science

Christian-Albrechts-University of Kiel
24098 Kiel, Germany
{koeser,rk}@mip.informatik.uni-kiel.de
Abstract. Robust local image features have been used successfully in

robot localization and camera pose estimation; region tracking using
affine warps is considered state of the art also for many years. Although
such correspondences provide a warp of the local image region and are
quite powerful, in direct pose estimation they are so far only consid-
ered as points and therefore three of them are required to construct a
camera pose. In this contribution we show how it is possible to directly
compute a pose based upon one such feature, given the plane in space
where it lies. This differential correspondence concept exploits the tex-
ture warp and has recently gained attention in estimation of conjugate
rotations. The approach can also be considered as the limiting case of
the well-known spatial resection problem when the three 3D points ap-
proach each other infinitesimally close. We show that the differential
correspondence is more powerful than conic correspondences while its
exploitation requires nothing more complicated than the roots of a third
order polynomial. We give a detailed sensitivity analysis, a comparison
against state-of-the-art pose estimators and demonstrate real-world ap-
plicability of the algorithm based on automatic region recognition.
1 Introduction
Since the first description of spatial resection from 3 points by Grunert[7] in

1841, many people have worked on pose estimation or the so called P3P problem
[5,31,6,8]. PnP stands for pose estimation from n points and is underconstrained
for n < 3 unless further information is incorporated. In this work we derive how
a variation of the problem may be solved, namely when only a single affine im-
age feature (cf. to [24] for a discussion) can be identified with a known 3D space
surface with orthophoto texture. Additionally to the traditionally used 2D-3D
point correspondence, such an image-model relation provides a local linear tex-
ture warp between the image and an orthophoto of the surface. This warp can be
interpreted as the Jacobian of the perspectivity between the image and the 3D
surface’s tangent plane and we show that it determines the open degrees of free-
dom. The novel approach allows to estimate a perspective camera’s pose based
upon only one image-model correspondence, which is particularly interesting in

Differential Spatial Resection 313
Fig. 1. Differential Spatial Resection exploiting Perspectivity. This figure

shows an aerial camera observing a ground plane (left image). If the internal camera
calibration is removed, the two images are related by a perspectivity. The projection of
some point on the plane and the linear transform of the surrounding region provide 6
constraints for the 6 degrees of freedom for pose estimation. In the right part we see an
MSER feature correspondence between an orthophoto and the unknown camera image
providing a locally affine texture transform. The primitive for correspondence can be
imagined as an infinitesimally small ellipse with orientation.
robot localization [29], initialization or recovery in camera tracking[3,32] or de-

termining the pose of a detected object[30]. In these applications, often SIFT[18]
or MSER[21] features are used nowadays, which cover some image region ideally
corresponding to a surface in the scene. In [3] even the normal of such local
surface regions is estimated and also [29] performs stereo from three cameras on
a robot. However, in all of the above cited approaches, the correspondences are
geometrically handled as points when it comes to initialization or direct pose
estimation, although they carry much more information. Therefore, by now at
least three of such robust feature correspondences were required to directly esti-
mate a camera or object pose. In contrast, in this contribution we demonstrate
how one affine image-model correspondence is already sufficient to estimate
the pose.
The exploited primitive can also be seen as the limiting case where the three
3D points of Grunert’s solution come infinitesimally close, allowing for what we
call differential spatial resection. The concept of such correspondences has lately
been proposed in [15] for estimation of the infinite homography and is displayed
in fig.1. The question we answer is: Given a local affine transform between a
region in some view and an orthophoto, how can we compute a homography
with this transform as its local linearization and what camera pose belongs to
it, given that the homography maps from a known world plane to the camera
plane ? Furthermore, we show in section 4.1 the relation to pose estimation
from conics, which is essentially a squared formulation of our approach although
providing one degree of freedom less.
The proposed algorithm belongs to the set of minimal solvers, which exploit
n DOF (degrees of freedom) in some observation to estimate a model with also
n DOF. Such solutions are not targeted to produce ultimate optimal estimates
314 K. Köser and R. Koch
but initial start values from as little data as possible, suitable for further pro-
cessing. For instance, when dealing with small or hand-clicked data sets or when
RANSAC-like estimators[6] are used, it is often desirable to obtain a minimal
solution, which requires as little of the data as possible. In RANSAC, the prob-
ability of picking an all-inlier-set from correspondences with many mismatches
depends exponentially on the number of samples required to construct a solution
hypothesis. Using our novel approach, it is now possible to obtain an estimate
of a camera’s pose from as little as one e.g. MSER[21] or comparable feature
(cf. to [24] for a discussion) or e.g. one suitable photogrammetric ground control
point (cf. to [22], p.1111) in an image, given the local plane in 3D space where
it is located and its texture.
For instance, when a feature descriptor is recognized in an unknown image, the
6 DOF camera or object pose can be obtained by the methods given here. To im-
prove the pose estimation result, gradient based optimization techniques[19,16]
can be applied between the current view and a reference texture. The reference
texture can either be an orthophoto (cf. to [22], p.758) or any other view with
sufficient resolution for which the warp to an orthophoto is known. When sev-
eral such feature correspondences and the camera poses are optimized at once,
this is similar to the approach of Jin et al.[11]. However, their approach is for-
mulated in a nonlinear fashion only and requires an initialization, comparable
to the requirements for bundle adjustment. Since we exploit the perspectiv-
ity concept, a plane-to-plane mapping in euclidian space, in section 3 we also
present the related work in homography estimation [33,13,10] and projective
reconstruction[27], which did not inspect the differential constraints on the per-
spectivity, because often the calibrated camera case is not considered in projec-
tive approaches. The exploitation of the Jacobian of the texture warp has been
proposed though for the estimation of a conjugate rotation in [15].
Notation. To improve the readability of the equations we use the following
notation: Boldface italic serif letters x denote Euclidean vectors while boldface
upright serif letters x denote homogeneous vectors. For matrices we do not use
serifs, so that Euclidean matrices are denoted as A and homogeneous matrices
are denoted as A, while functions H [x] appear in typewriter font.
2 Perspectivity
The contribution is based on estimating a transformation between two theoreti-
cal planes: The first plane is tangent to a textured surface in 3D and the second
plane is orthogonal to the optical axis of a camera. The estimation of the pose is
then formulated as the problem of obtaining a perspectivity between these two
planes (see figure 1). A 2D perspectivity is a special kind of homography (cf.
also to [9], pp. 34), which has only 6 degrees of freedom and which is particularly
important for mappings between planes in Euclidian space. We assume a locally
planar geometry at the origin of 3D space facing into z-direction and attach
x, y-coordinates onto it, which coincide with the x, y coordinates in 3D space.
If we now move a perspective pinhole camera to position C with orientation R
(which has rows rT

i ) and with internal camera calibration K, a point ps in space
is mapped to an image point pi by the camera as follows(cf. to [9], p. 157 for
details):
pi = K(R T | − R T C)ps (1)
We assume the internal parameters of our camera to be known and without loss
of generality set K to the identity in the following. The method is not restricted
to straight-line preserving ideal cameras, but can also be applied with real lenses
with distortion, fish-eye lenses or even omni-directional cameras, as long as they
have a single center of projection and the equivalent function of the matrix K,
which maps rays in the camera coordinate system to positions in the image, is
differentiable and invertible.
Now we have a look at the points on our z = 0 plane to derive the
perspectivity:
pi = (R T | − R T C)ps,z=0 = (r1 r2 r3 − R T C)(x y 0 1)T (2)
= (r 1 r2 − R T C)(x y 1)T (r˜1 r˜2 t)(x y 1)T = H pp (3)

r˜i are scaled versions of ri such that tz = 1 and means equality up to scale.
Obviously, the homography H maps points pp of the plane coordinate system
to points pi in the image coordinate system. H is a perspectivity and depends
only on 6 parameters, the pose of the camera. Since H is an object of projective
space, it can be scaled without changing the actual transformation. While the
perspectivity H acts linearly in projective space P2 , in Euclidian 2D space H is a
nonlinear mapping from R2 → R2 because of the nonlinear homogenization:
# $ (Hpp )|1..2
H pp = pi = (4)
(Hpp )|3
In the next section we describe the differential correspondence and how it can
be exploited to obtain constraints on H.
3 Differential Correspondence
Progress in robust local features (cf. to [24,23] for a thorough discussion) al-
lows automatic matching of images in which appearance of local regions un-
dergoes approximately affine changes of brightness and/or of shape, e.g. for
automated panorama generation[1], scene reconstruction[30] or wide-baseline
matching[18,21]. The idea is that interesting features are detected in each im-
age and that the surrounding region of each feature is normalized with respect
to the local image structure in this region, leading to about the same normal-
ized regions for correspondences in different images, which can be exploited for
matching. The concatenation of the normalizations provides affine correspon-
dences between different views, i.e. not only a point-to-point relation but also
a relative transformation of the local region (e.g. scale, shear or rotation). Al-
though such correspondences carry more information than the traditional point
correspondence used in estimation of multiple view geometry [9], this additional

information is rarely used. Approaches not using point correspondences deal
with conic correspondences [14,12], which typically lead to systems of quadratic
equations or require lots of matrix factorizations. Schmid and Zisserman[28] in-
vestigated the behavior of local curvature under homography mapping. Chum
et al. noted in [2] that an affine correspondence is somehow equivalent to three
point correspondences: in addition to the center point two further points can be
detected in the feature coordinate system (the local affine frame). This allowed
the estimation of a fundamental matrix from 3 affine feature correspondences
(from which 9 point correspondence were generated). A similar idea was also
exploited recently in projective reconstruction, where the projection matrix was
locally linearized[27] leading to additional constraints in non-linear optimization.
The “local sampling” of the affine feature concept on the other hand was also
adopted for other epipolar geometry problems, e.g. in [26]. In contrast to the
latter we do not sample but use a compact analytic expression for the whole
correspondence: We observe that the concatenation of the normalization trans-
formations provides a good approximation to the first order Taylor expansion
of the perspectivity, i.e. that the resulting affine transform is the local lineariza-
tion of the perspectivity, as it has been recently proposed for estimation of the
infinite homography[15]:

∂H
H [x] = H [x0 ] + (x − x0 ) + . . . (5)
∂x x0

∂H
A≈ A ∈ R2×2 (6)
∂x x0
Here H : R2 → R2 is the homography mapping between the image and the
orthophoto in Euclidean coordinates and A represents local shear, scale and
rotation between the two corresponding features. This fact has been exploited in
matching for quite some time but has not been used for pose estimation before.
The considerations so far apply to affine features (e.g. MSER[21]). However, if
matches result from weaker features (e.g. DoG/SIFT[18]), the proposed method
can also be applied. The main insight is that if a correct match has been estab-
lished such that the local regions are approximately aligned the affine transform
based upon the relative parameters is already nearly correct.
However, since we need an accurate estimate of the Jacobian of the im-
age transformation, it is reasonable even for already affine features to apply
a gradient-based optimization of A using the Lucas-Kanade approach [19,16].
When using affine trackers, e.g. such as [3], the optimized information is read-
ily available. We will call the point correspondence plus the local linear warp a
differential correspondence in the remainder.
4 Pose Estimation from a Differential Correspondence

Having obtained a differential correspondence between a camera image and
the textured plane in the origin, the local warp equals the derivative of the
perspectivity. This derivative ∂H/∂pp tells us something about the relative scal-
ing of coordinates between the plane in the origin and the image, e.g. if C is
large and the camera is far away from the origin ∂H/∂pp will be small, because
a large step on the origin plane will result in a small step in the image far away.
Actually, ∂H/∂pp carries information about rotation, scale and shear through
perspective effects. Since H can be scaled arbitrarily without changing H, we set
H3,3 = 1 without loss of generality1 and compute the derivative at the origin:

∂H r̃11 − r̃13 t1 r̃12 − r̃13 t1 a11 a12
= = (7)
∂pp 0
r̃21 − r̃23 t2 r̃22 − r̃23 t2 a21 a22
Also, we compute where the origin is projected in our image:
porigin = H(0 0 1)T = −R T C t (8)
Given a differential correspondence, the derivative as well as the projection of

the origin are given by the relative parameters of the detected features. This
can determine all degrees of freedom of the camera pose, however the over-
parameterization of the rotation must be resolved: Since R̃ is a scaled rotation
matrix, r̃ 1 and r̃2 must be of same length and orthogonal:
r̃11
2
+ r̃12
2
+ r̃13
2
= r̃21
2
+ r̃22
2
+ r̃23
2
∧ r̃T
1 r̃ 2 = 0 (9)
We can now compute H by first substituting t into eq. (7), then solving for
r̃11 , r̃21 , r̃12 and r̃22 and substituting into eq.(9), leaving us with two quadratic
equations in the two unknowns r̃13 and r̃23 :
(r̃13 t1 + a11 )2 + (r̃13 t1 + a12 )2 + r̃13

2
= (r̃23 t2 + a21 )2 + (r̃23 t2 + a22 )2 + r̃23
2
(10)
(r̃13 t1 + a11 )(r̃23 t2 + a21 ) + (r̃13 t1 + a12 )(r̃23 t2 + a22 ) + r̃13 r̃23 = 0 (11)
The first equation is about the length and the second about the orthogonal-
ity of the r̃-vectors as typical for constraints on rotation matrices. We find it
instructive to interpret them as the intersection problem of two planar conics,
the length conic Cl and the orthogonality conic Co :
(r̃13 r̃23 1)Cl (r̃13 r̃23 1)T = 0 (12)

T
(r̃13 r̃23 1)Co (r̃13 r̃23 1) = 0 (13)
⎛ ⎞
2t21 + 1 0 t1 (a11 + a12 )
Cl = ⎝ 0 −2t22 − 1 −t2 (a21 + a22 ) ⎠ (14)
t1 (a11 + a12 ) −t2 (a21 + a22 ) a11 + a12 − a21 − a22
2 2 2 2
⎛ ⎞
0 t1 t2 + 12 (a21 + a22 )t1
Co = ⎝ t1 t2 + 12 0 (a11 + a12 )t2 ⎠ (15)
(a21 + a22 )t1 (a11 + a12 )t2 a11 a21 + a12 a22
1
This is not a restriction because the only unrepresented value H3,3 = 0 maps the
origin to the line at infinity and therefore such a feature would not be visible.
Solving for the Pose Parameters. Two conics cannot have more than four
intersection points, therefore, we can obtain at most four solutions for our camera
pose. To solve the intersection of the two conics we use the elegant method of
Finsterwalder and Scheufele[5], which proved also to be the numerically most
stable method of the six different 3-point algorithms for spatial resection [8]:
Since a common solution of equations (12) and (13) must also fulfill any linear
combination of both, we construct a linear combination of both conics, which
does not have full rank (zero determinant), but which still holds all solutions.
This creates a third order polynomial, which has at least one real root and which
can be solved easily:
det(λCo + (1 − λ)Cl ) = 0 (16)
The resulting degenerate conic will in general consist of two lines. The inter-
section of these lines with the original conics is only a quadratic equation and
determines the solutions. The resulting R and C have to be selected and normal-
ized in such a way that we obtain an orthonormal rotation matrix (determinant
+1) and the camera looks towards the plane. We have now obtained up to four
hypotheses for the pose of the camera in the object coordinate system (relative
to the feature). If there is a world coordinate system, in which the plane is not at
the origin, the rigid world transformation has to be appended to the computed
pose of the camera. Computing the relative pose in the object coordinate system
in general also improves conditioning since the absolute numbers of the object’s
pose in the world become irrelevant.
Optimization and Tracking. Once initial parameters are obtained it is

straightforward to use a 6-parametric gradient-based minimization technique
[19,16] to further optimize the camera pose. Note that if we are using a pin-
hole camera and the feature in 3D is locally planar, instead of optimizing an
approximate affine transform we might as well use a 6-parametric homography.
Thus measurements may be incorporated from a larger region without making
a mistake or an approximation. Even better, since it is possible to use global
camera pose parameters, it is easy to optimize even multiple rigidly coupled
features (e.g. in a rigid scene). Or, if robustness against outliers is a concern,
each of the features provides an individual pose estimate and robust estimation
techniques such as RANSAC[6] can be used to obtain a fused solution. If video
data is available, the parameters can directly be used for tracking the regions,
objects or camera pose over time similar to what is proposed in [11]. However,
in this contribution we focus on the geometric aspects of the minimal solution,
i.e. where we see a single feature in a single image, without prior knowledge.
4.1 Relation to Conic Correspondence

In this section the differential feature concept is shown to be a simplified version
of correspondences of conics, providing more constraints in a linear (instead of
quadratic) fashion: In [20] Ma derived a way to determine the pose of a camera
from two conics. He noted that a conic has only 5 DOF and thus a single conic
is not sufficient to determine the 6 DOF of the camera pose uniquely. A conic
CS on the space plane of the previous section maps to a conic CI in the image
with the equation
CI = HT CS H, (17)
where H is the perspectivity of the previous sections. First, we show how the
two primitives used in our differential correspondence can be related to conic
representations: For each affine feature, e.g. MSER, there exists a local image
coordinate system, the local affine frame[2], such that coordinates can be speci-
fied relative to the size, shear, position and orientation of a feature. Imagine that
L takes (projective) points from local feature coordinates to image coordinates:
xI = LxLAF (18)
If the same feature is seen in two images, points with identical feature (LAF)
coordinates will have the same grey value. The local affine frames of the features
in the different images are then called L1 and L2 and their concatenation is the
first order Taylor approximation HTaylor of the texture warp (e.g. a homography)
between the two images at the feature positions:
HTaylor = L1 L−1
2 (19)
If we now just think of a single image and imagine a small ellipse through
the points (0; λ)T ,(λ; 0)T ,(0; −λ)T and (−λ; 0)T of the local feature coordinate
system, this ellipse can be represented by a conic equation in homogeneous co-
ordinates such that points at the ellipse contour fulfill the quadratic constraint:
⎛ ⎞
1
0 = xTLAF
⎝ 1 ⎠ xLAF (20)
−λ2
The LAF described as a conic matrix in image coordinates therefore is

⎛ ⎞ ⎛ ⎞
1 1
Cλ = LT ⎝ 1 ⎠ L = LT RT ⎝ 1 ⎠ RL (21)
−λ2
−λ2
where R is an arbitrary (homogeneous 2D) rotation matrix, which cancels out.

Therefore the first thing to observe is that 2D orientation of the feature is lost
in conic representation. A conic has only five degrees of freedom and a conic
correspondence therefore imposes at most five constraints on any H. Further-
more, these constraints are quadratic in the entries of H as can be seen from eq.
(17). This equation is also essentially a squared version of equation (19). On the
other hand, the differential correspondence is only valid locally and introduces
inacurracies for larger regions, but it is available when sufficient texture is in the
image, while a conic may have any size. However, conics traditionally exploit
a special geometric shape (typically an ellipse contour) and ideal perspective
cameras and ideal planes, because conic curve estimation in distorted cameras
is more involved. In contrast, the differential feature concept can also directly
be applied in fish-eye or omnidirectional cameras.
5 Evaluation
In this section the differential correspondence-based pose estimation is evaluated
first using synthetic sensitivity experiments. Next, rendered images with known
ground truth information are used to evaluate the real-world applicability, where
everything has to be computed from image data. In the final experiments, object
pose estimation from one feature is shown qualitatively using non-ideal cameras.
Sensitivity to Noise and Internal Calibration Errors. Our evaluation

starts with an analysis of the sensitivity to different disturbances. Since the
algorithm provides a minimal solution, which translates a 6 DOF differential
correspondence into a 6 DOF pose, the pose will adapt to noise in the corre-
spondence. In figure (5) it is shown that for localization accuracies better than 1
pixel in a camera with focal length 500 pixel the camera orientation is on average
better than 1 degree and also the direction of the camera center is better than 1
degree. The orientation error is computed from the axis-angle representation of
the rotation which transforms the ground truth orientation into the estimated
orientation and therefore incorporates all directions. The center error is the an-
gle between the ground truth camera center and the estimated camera center as
seen from the 3D feature’s position.
To obtain a reasonable noise magnitude for the differential correspondence
parameters, we assume that the center of a patch can be localized with a Gaus-
sian uncertainty of zero mean and variance σp2 and that the corners of a square
patch of size (2w + 1) × (2w + 1) pixels can be localized with about the same
uncertainty, which can then be propagated to uncertainty for the affine parame-
ters. When creating noisy 6D affine features, we therefore sample the noise from
a Gaussian distribution with diagonal covariance depending on one parameter
σp , which is printed on the x-axis of figure 5. It is remarkable that the errors in
orientation and position are highly correlated. This can be explained from the
fact that a slightly different differential correspondence results in a slightly differ-
ent camera orientation. However, since the feature must be projected to about
the same position, the camera center has to adapt accordingly. As figure 5
shows, the pose estimation is stable even when the camera is not calibrated cor-
rectly, although it can be seen that the resulting pose is disturbed as inherent
in minimal solutions. In particular it is clear that an error in principal point
results in an error in the pose when the reference feature in 3D is correct. Keep
in mind that at focal length 500 a principal point error of ten pixel means that
the optical axis is more than 1◦ mis-calibrated.
Solid Angle, Approximation by 3 Points and Comparison with Spa-

tial Resection/POSIT. Using the proposed approach, the affine warp must
be measured between the orthophoto and the image under inspection and this
requires a region upon which this is done. If the alignment is done using an affine
warp, the region should be chosen as small as possible, particularly when the fea-
ture is seen from an oblique angle, because in the affine warp model it is assumed
that the warp (the Jacobian of the homography) does not change between the
Fig. 2. Sensitivity with Respect to Noise, Calibration and Feature Area. In

these experiments, 105.000 random camera poses in front of the z = 0 plane have
been synthesized (providing the ground truth differential correspondences). In the two
top graphs, focal length (500) and principal point (100;100) have been disturbed up
to 10% and the resulting error in the best pose is displayed as well as the number
of cases, where no solution was possible (or the best pose was more than 5◦ off).
In the lower left graph, Gaussian noise has been added to the 6 parameters of the
differential correspondence, where we assume that the position accuracy σp of the
center of the patch is the same as for the corners of √
a patch of half window size w and
therefore disturb the 4 affine parameters with σp /( 2w), where we assume a 21×21
window. The error bars indicate the size of the standard deviation. In the bottom right
figure, we compare the 3-point solution proposed in the Manual of Photogrammetry[22,
pp.786] (P3P), the planar POSIT algorithm [25] based on the 4 patch corners (which
already includes the parallel projection approximation by [17] in the POS step) and our
novel solution applied to the case that we use 3 or more close points: The differential
correspondence is approximated using the four patch corners only, while we vary the
size (given as the solid angle) of the patch and fix σp for the corners at 0.5. The error
bars show 1/3 standard deviation. As expected, it can be seen that for large solid angle
spatial resection performs best while for decreasing solid angles the novel solution gets
better and better, outperforming the other approaches for very narrow constellations.
corners of the local patch. On the other hand, when the 3 individual 3D points
of Grunert’s solution approach each other, the standard spatial resection can
become unstable, because it is based on the difference of the distances to the 3
points. To overcome this issue, Kyle [17] proposed an approximate initial guess
Fig. 3. Camera Pose From Noisy Images. A ground plane has been textured with
an aerial image serving as an orthophoto and a series of 40 views have been rendered
with different levels of noise (upper row: sample views with low noise). A reference
MSER feature with orientation has been chosen in the orthophoto. This feature is
then detected in the other views and refined using a simple 6-parametric affine warp
(see ellipses in bottom left image) according to [16] based upon a half window size of
10 pixels. From such differential correspondences, the camera pose is estimated and
compared against the known ground truth value as explained earlier. Whenever the
error was above 20◦ or the algorithm did not come up with a solution a failure was
recorded. The bottom right graph shows the average pose errors in dependence of the
added image noise. When adding much more image noise, the MSER detector is no
longer able to find the feature. This experiment is particularly interesting because it
shows that the concept does still work when the ellipse is not infinitely small.
for narrow angle images, which is the same as the POS (Pose from Orthography
and Scaling) in the POSIT[4] algorithm: Both require 4 non-coplanar points.
For the POSIT algorithm however, there exists also a planar variant[25], which
copes with planar 3D points.
Therefore we compare our novel algorithm (well-suited for small solid an-
gles) to the spatial resection[7,8] implemented as proposed in the Manual of
Photogrametry[22, pp. 786] and the planar POSIT[25] algorithm kindly pro-
vided on the author’s homepage, which are both desiged for larger solid angles.
We vary the size of a local square image patch from ten to several hundred
pixels and use the corners as individual 2D-3D correspondences in the existing
algorithms. For our new method the patch corner points are used to compute
a virtual local affine transform which approximates the required Jacobian. An
evaluation of the quality of the approximation can be seen in the bottom right
of fig. 5, which shows that for small solid angles the novel solution outper-
forms spatial resection, while for large solid angles - as expected - the affine
Fig. 4. Object Pose Estimation from a Single Feature. This figure shows that
in a real camera with radial distortion object pose estimation is possible from a single
feature. The orthophoto of the object is displayed in the right image with the local fea-
ture region enlarged. The two left images show cluttered views with the object partially
occluded. The “M” has been detected using MSER and refined, the resulting object
poses from this single differential correspondence are then displayed by augmenting a
contour model (white).
approximation is not suitable. It is however still better in average than the or-
thographic approximation in the planar POSIT algorithm. Particularly, when
the solid angle approaches zero, the error in the novel solution tends to zero,
while for the other algorithms no solution can be obtained or the best solution
is worse than the robust error threshold of 10◦ .
Normal or Pose Error of the Local Plane. An error of the normal of the
3D reference plane, for which the orthophoto exists or an error of the pose of this
plane cannot be detected within the algorithm. The pose is computed relative
to this plane and an error of the plane in global coordinates will consequently
result in a relative error of the camera pose in global coordinates.
Warp Measured From Real Texture. In the next experiment, we rendered

views from a ground plane and apply automatic matching and pose estimation
based upon a single prominent image feature. Since we do have the ground
truth data the estimated pose can be analyzed in this case. The experiments are
explained in fig.3 and show that even in presence of image noise using automatic
matching and refinement approaches the pose can be estimated quite reliably,
given the minimal local texture data which is used.
Images with Radial Distortion. In the final experiment we took photographs

of an office scene, where we detect a cereal box, which is partially occluded. As
in the previous experiment, an MSER feature is obtained from an orthophoto
of the cereal box. Next this feature is automatically found in the test image and
refined using gradient-based (affine) optimization. Again, from this differential
correspondence the object pose is estimated, which might be interesting in ap-
plications where only small portions of an object are visible. The augmentation
in fig.4 shows quite good results.
6 Conclusion
A method for estimating a camera pose based upon a single local image feature
has been proposed which exploits the often readily available local affine warp
between two images. This differential correspondence provides more constraints
than a point or a conic and can be used easily in calibrated cameras even if they
deviate from the linear projection model. The algorithm proved to be stable
under several kinds of disturbance and can also be applied when the 3 individual
3D points of a general spatial resection problem come very close because the
novel formulation avoids directly computing the 3 distances, which can lead to
numerical difficulties in practise. Another benefit of the novel minimal solution
is that it allows now for computing the pose from a single image-model match
of common robust features which could reduce RANSAC complexity compared
to the previously required set of 3 correspondences.
References
1. Brown, M., Lowe, D.G.: Automatic panoramic image stitching using invariant fea-
tures. International Journal of Computer Vision 74(1), 59–73 (2007)
2. Chum, O., Matas, J., Obdrzalek, S.: Epipolar geometry from three correspon-
dences. In: Computer Vision Winter Workshop, Prague, pp. 83–88 (2003)
3. Davison, A.J., Reid, I.D., Molton, N.D., Stasse, O.: Monoslam: Real-time sin-
gle camera slam. IEEE Transactions on Pattern Analysis and Machine Intelli-
gence 29(6), 1052–1067 (2007)
4. DeMenthon, D., Davis, L.S.: Model-based object pose in 25 lines of code. Interna-
tional Journal of Computer Vision 15, 123–141 (1995)
5. Finsterwalder, S., Scheufele, W.: Das Rückwärtseinschneiden im Raum. In: Bay-
erische, K., der Wissenschaften, A. (eds.) Sitzungsberichte der mathematisch-
physikalischen Klasse, vol. 23/4, pp. 591–614 (1903)
6. Fischler, M., Bolles, R.: RANdom SAmpling Consensus: a paradigm for model
fitting with application to image analysis and automated cartography. Communi-
cations of the ACM 24(6), 381–395 (1981)
7. Grunert, J.A.: Das Pothenot’sche Problem, in erweiterter Gestalt; nebst Bemerkun-
gen über seine Anwendung in der Geodäsie. In: Archiv der Mathematik und Physik,
vol. 1, pp. 238–248, Greifswald. Verlag C.A. Koch (1841)
8. Haralick, B., Lee, C., Ottenberg, K., Nölle, M.: Review and analysis of solutions
of the three point perspective pose estimation problem. International Journal of
Computer Vision 13(3), 331–356 (1994)
9. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision, 2nd edn.
Cambridge University Press, Cambridge (2004)
10. Irani, M., Rousso, B., Peleg, S.: Recovery of ego-motion using region alignment.
Transact. on Pattern Analysis and Machine Intelligence 19(3), 268–272 (1997)
11. Jin, H., Favaro, P., Soatto, S.: A semi-direct approach to structure from motion.
The Visual Computer 19(6), 377–394 (2003)
12. Kahl, F., Heyden, A.: Using conic correspondence in two images to estimate the
epipolar geometry. In: Proceedings of ICCV, pp. 761–766 (1998)
13. Kähler, O., Denzler, J.: Rigid motion constraints for tracking planar objects. In:
Hamprecht, F.A., Schnörr, C., Jähne, B. (eds.) DAGM 2007. LNCS, vol. 4713, pp.
14. Kannala, J., Salo, M., Heikkila, J.: Algorithms for computing a planar homography
from conics in correspondence. In: Proceedings of BMVC 2006 (2006)
15. Koeser, K., Beder, C., Koch, R.: Conjugate rotation: Parameterization and esti-
mation from an affine feature corespondence. In: Proceedings of IEEE Conference
on Computer Vision and Pattern Recognition (CVPR) (2008)
16. Koeser, K., Koch, R.: Exploiting uncertainty propagation in gradient-based image
registration. In: Proc. of BMVC 2008 (to appear, 2008)
17. Kyle, S.: Using parallel projection mathematics to orient an object relative to a
single image. The Photogrammetric Record 19, 38–50 (2004)
18. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International
19. Lucas, B.D., Kanade, T.: An iterative image registration technique with an appli-
cation to stereo vision. In: IJCAI 1981, pp. 674–679 (1981)
20. De Ma, S.: Conics-based stereo, motion estimation, and pose determination. Inter-
national Journal of Computer Vision 10(1), 7–25 (1993)
21. Matas, J., Chum, O., Urban, M., Pajdla, T.: Robust wide baseline stereo from
maximally stable extremal regions. In: Proceedings of BMVC 2002 (2002)
22. McGlone, J.C. (ed.): Manual of Photogrammetry, 5th edn. ASPRS (2004)
23. Mikolajczyk, K., Schmid, C.: A performance evaluation of local descriptors. Trans-
act. on Pattern Analysis and Machine Intell. 27(10), 1615–1630 (2005)
itzky, F., Kadir, T., van Gool, L.: A comparison of affine region detectors. Inter-
national Journal of Computer Vision 65(1-2), 43–72 (2005)
25. Oberkampf, D., DeMenthon, D., Davis, L.S.: Iterative pose estimation using copla-
nar feature points. CVGIP 63(3) (1996)
26. Riggi, F., Toews, M., Arbel, T.: Fundamental matrix estimation via TIP - transfer
of invariant parameters. In: Proceedings of the 18th International Conference on
Pattern Recognition, Hong Kong, August 2006, pp. 21–24 (2006)
27. Rothganger, F., Lazebnik, S., Schmid, C., Ponce, J.: Segmenting, modeling, and
matching video clips containing multiple moving objects. IEEE Transactions on
28. Schmid, C., Zisserman, A.: The geometry and matching of lines and curves over
multiple views. International Journal of Computer Vision 40(3), 199–234 (2000)
29. Se, S., Lowe, D.G., Little, J.: Vision-based global localization and mapping for
mobile robots. IEEE Transactions on Robotics 21(3), 364–375 (2005)
30. Skrypnyk, I., Lowe, D.G.: Scene modelling, recognition and tracking with invari-
ant image features. In: IEEE and ACM International Symposium on Mixed and
Augmented Reality, pp. 110–119 (2004)
31. Thompson, E.H.: Space resection: Failure cases. The Photogrammetric
Record 5(27), 201–207 (1966)
32. Williams, B., Klein, G., Reid, I.: Real-time slam relocalisation. In: Proceedings of
ICCV, Rio de Janeiro, Brazil, pp. 1–8 (2007)
33. Zelnik-Manor, L., Irani, M.: Multiview constraints on homographies. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence 24(2), 214–223 (2002)
Riemannian Anisotropic Diffusion for Tensor Valued
Images
Kai Krajsek1 , Marion I. Menzel1 , Michael Zwanger2 , and Hanno Scharr1

1
Forschungszentrum Jülich, ICG-3, 52425 Jülich, Germany
{k.krajsek,m.i.menzel,h.scharr}@fz-juelich.de
2
Siemens AG, Healthcare Sector
MR Application Development, 91052 Erlangen, Germany
Michael.Zwanger@siemens.com
Abstract. Tensor valued images, for instance originating from diffusion tensor
magnetic resonance imaging (DT-MRI), have become more and more important
over the last couple of years. Due to the nonlinear structure of such data it is
nontrivial to adapt well-established image processing techniques to them. In this
contribution we derive anisotropic diffusion equations for tensor-valued images
based on the intrinsic Riemannian geometric structure of the space of symmet-
ric positive tensors. In contrast to anisotropic diffusion approaches proposed so
far, which are based on the Euclidian metric, our approach considers the nonlin-
ear structure of positive definite tensors by means of the intrinsic Riemannian
metric. Together with an intrinsic numerical scheme our approach overcomes
a main drawback of former proposed anisotropic diffusion approaches, the so-
called eigenvalue swelling effect. Experiments on synthetic data as well as real
DT-MRI data demonstrate the value of a sound differential geometric formula-
tion of diffusion processes for tensor valued data.
1 Introduction
In this paper anisotropic diffusion driven by a diffusion tensor is adapted to tensor-
valued data in a way respecting the Riemannian geometry of the data structure. Nonlin-
ear diffusion has become a widely used technique with a well understood theory (see
e.g. [1,2] for overviews). It was introduced in [3] and has been frequently applied to
scalar-, color- or vector-valued data. Anisotropic diffusion1 driven by a diffusion ten-
sor [2] is the most general form of diffusion processes. Tensor-valued data frequently
occur in image processing, e.g. covariance matrices or structure tensors in optical flow
estimation (see e.g. [4]). Due to rapid technological developments in magnetic reso-
nance imaging (MRI) also interest in tensor-valued measurement data increases. Due
to the increasing need of processing tensor valued data, the development of appropri-
ate regularization techniques become more and more important (e.g. see [5,6,7,8] and
[9] as well as references therein). Riemannian geometry refers to the fact that the set
of positive definite tensors P (n) of size n does not form a vector space but a nonlin-
ear manifold embedded in the vector space of all symmetric matrices. The nonlinear
1
Please note that the term ’anisotropic diffusion’ is not uniquely defined in literature. In this
contribution we use the term in accordance with the definition given in [2].

Riemannian Anisotropic Diffusion for Tensor Valued Images 327
structure of P (n) is studied from a differential geometric point of view for a long time
[10]. Due to the nonlinear structure of P (n), well established image processing tech-
niques for scalar and vector valued data might destroy the positive definiteness of the
tensors. Approaches for processing tensor valued images can be classified into two
groups: using extrinsic [5,11,12,13,14] or intrinsic view [15,16,17,18,19,20,21,7,22].
Methods using the extrinsic point of view consider the space of positive definite sym-
metric tensors as an embedding in the space of all symmetric tensors which constitute
a vector space. Distances, as e.g. required for derivatives, are computed with respect
to the flat Euclidian metric of the space of symmetric matrices. To keep tensors on the
manifold of positive definite tensors, solutions are projected back onto the manifold
[5], selected only on the manifold in a stochastic sampling approach [11], or process-
ing is restricted to operations not leading out of P (n), e.g. convex filters [12,13,14].
Although then tensors stay positive definite the use of a flat metric is not appropriate to
deal with P (n). For instance in regularization, the processed tensors become deformed
when using the flat Euclidian metric [7] known as eigenvalue swelling effect [5,6,7,8].
Tschumperlé and Deriche [5] avoid the eigenvalue swelling effect by applying a spectral
decomposition and regularizing eigenvalues and eigenvectors separatively. Chefd’hotel
et al. [6] proposed to take the metric of the underlying manifold for deriving evolution
equations from energy functionals that intrinsically fulfill the constraints upon them
(e.g. rank or eigenvalue preserving) as well as for the numerical solution scheme. How-
ever, they consider the Euclidian metric for measuring distances between tensors such
that their methods suffer from the eigenvalue swelling effect for some of the proposed
evolution equations. Methods using the intrinsic point of view consider P (n) as a Rie-
mannian symmetric space (see [23] and Sect. 3 for an introduction in symmetric Rie-
mannian spaces) equipped with an affine invariant metric on the tangent space at each
point. Consequently, using this metric the eigenvalue swelling effect is avoided. The
symmetry property of the Riemannian manifold easily allows to define evolution equa-
tions on the tangent spaces, approximate derivatives by tangent vectors as well as con-
struct intrinsic gradient descent schemes as we will show for anisotropic diffusion in the
following.
Related work. Differential geometric approaches have been introduced to different

fields in image processing and computer vision [24,25,26]. Only quite recently, meth-
ods based on the Riemannian geometry of P (n) have been introduced independently
by different authors [16,17,18,19,20,21,7,22]. For instance, in [20,7] a ’Riemannian
framework for tensor computing’, has been proposed in which several well established
image processing approaches including interpolation, restoration and isotropic nonlin-
ear diffusion filtering have been generalized to P (n) in an intrinsic way. Furthermore,
an anisotropic regularization approach has been proposed by adapting the isotropic
Laplace-Beltrami operator that can be identified with a second order Markov Random
field approach. A quite similar approach has been proposed in [27] by formulating dif-
fusion filtering directly on a discrete graph structure. In [8], a weighted mean has been
proposed that allows to smooth the image in an anisotropic way. However, all these ap-
proaches [7,27,8], do not allow to construct diffusion tensors from model based struc-
ture estimation, as common in literature for scalar data [2]. To do so in an intrinsic
way, one cannot do without a numerical scheme for mixed second order derivatives,
328 K. Krajsek et al.
first introduced in the current paper. A computational more efficient approach than the
framework of Pennec et al. [7] based on the so called log-Euclidean metric has been
introduced in [28]. There, the positive definite tensors are mapped onto the space of
symmetric matrices by means of the matrix logarithmic map. In this new space com-
mon vector valued approaches can be applied. The final result is obtained by mapping
the transformed symmetric matrices back onto the space of positive definite matrices
using the matrix exponential map. However, the log-Euclidean metric is not affine in-
variant. As a consequence the approach might suffer from a change of coordinates.
However, the formulation of anisotropic diffusion for tensor valued data based on the
log-Euclidean metric might be a computational efficient alternative not proposed in lit-
erature so far. In [22,29] a Riemannian framework based on local coordinates has been
proposed (see also in [30] for a variational framework for general manifolds). Although,
the authors in [22,29] consider the affine invariant metric their approach may only be
classified as intrinsic in a continuous formulation. For computing discrete data, a sim-
ple finite difference approximation is applied. Inferring from a continuous formulation
without a proof to a discrete approximation can be misleading as constraints holding in
the continuous case may be relaxed by discretization. As a consequence, the proposed
approaches not necessarily preserve positive definiteness of tensors (for a detailed dis-
cussion of this topic for scalar valued signals we refer to [2]). Furthermore, the approach
of [29] shows no significant difference with the log Euclidean framework whereas our
approach clearly outperforms it. We refer to our approach as the full intrinsic scheme
in order to distinguish it from schemes that are only intrinsic in the continuous setting.
Anisotropic diffusion based on an extrinsic view [12,31] and by means of the exponen-
tial map [6] has been proposed. In both cases the Euclidian metric is used to measure
distances between tensors. As a consequence, both approaches suffer from the eigen-
value swelling effect.
Our contribution. We derive an intrinsic anisotropic diffusion equation for the mani-
fold of positive definite tensors. To this end, second order derivatives in the continuous
as well as discrete approximations are derived as they occur in the anisotropic diffusion
equation. The derived numerical scheme could also be used to generalize other PDEs
involving mixed derivatives from scalar valued images to the manifold P (n) without
the need of local coordinates. In the experimental part, we provide a study in which we
compare different state of the art regularization approaches with our approach.
2 Diffusion for Scalar Valued Images
We review diffusion filtering which is a well established image processing technique

for scalar valued images [3,32,2]. We formulate the diffusion equation by means of a
gradient descent of some energy functional that later allows us to generalize this con-
cept to tensor valued data. Let f be a scalar valued image defined on a N -dimensional
domain. Diffusion filtering image processing creates a family of images {u(x, t)|t ≥ 0}
from the solution of the physical diffusion equation
∂t u = div (D∇u) (1)

with initial condition f = u(x, 0) and diffusion tensor D with components dij . Note
that we could also formulate the image restoration task as a solution of a diffusion reac-
tion equation by adding a data depending term to (1). We will discuss the pure diffusion
process only. All following results keep valid also for a formulation with data depend-
ing reaction terms. The diffusion equation can be reformulated applying the chain rule
in the form ∂t u = i,j (∂i dij )(∂j u) + dij ∂i ∂j u which will be more convenient for the
formulation on tensor valued data. The diffusion process can be classified according to
the diffusion tensor D. If the diffusion tensor does not depend upon the evolving im-
age, the diffusion process is denoted as linear due to the linearity of (1) otherwise it
is termed nonlinear. The diffusion process can furthermore be classified into isotropic
when the diffusion tensor is proportional to the identity matrix otherwise it is denoted as
anisotropic. Except for the nonlinear anisotropic diffusion scheme, the diffusion equa-
tion can be derived from a corresponding energy functional E(u) via calculus of varia-
tion, i.e. the gradient descent scheme of these energy functionals can be identified with
a diffusion equation. Let L(u) denote the energy density such that E(u) = L(u) dx,
w : IRN → IR a test function and ε a real valued variable. The functional derivative

δE := δE(u+εw)δε of an energy functional E(u) can be written as
ε=0

δE = ∇L(u), wu dx , (2)
where ∇L(u) defines the gradient of the energy density and ∇L(u), wu denotes the
scalar product of the energy density gradient ∇L(u) and the test function evaluated at
x. Note that w as well as ∇L(u) are elements of the tangent space at u which is the
Euclidian space itself for scalar valued images. As we will see in Sect. 4, this formu-
lation allows a direct generalization to the space of symmetric positive definite tensors.
The gradient descent scheme of the energy functional leads to the diffusion equation in
terms of the energy density
∂t u = −∇L(u) . (3)
Let us now consider the linear anisotropic diffusion equation (1), i.e. D not depending
on the evolving signal. The corresponding energy function is known to be

1
E(u) = ∇uT D∇u dx . (4)
2
The functional derivative of (4) can be brought into the form

δE(u) = −div (D∇u) , wu dx (5)
assuming homogenous Neumann boundary conditions and applying Green’s formula.

Comparing (5) with (2) gives together with (3) the diffusion equation (1). Our objective
is now to generalize the linear anisotropic diffusion process to the space of positive def-
inite tensors by means of the energy functional formulation. The nonlinear anisotropic
diffusion equation on P (n), can then be deduced from the linear one.
3 The Space of Positive Definite Tensors
In the following we review the structure of the space of positive definite tensors P (n)
and introduce the differential geometric tools necessary for deriving anisotropic diffu-
sion equations for P (n). By introducing a basis, any tensor can be identified with its
corresponding matrix representation A ∈ Rn×n . The space of n × n matrices
consti-
tutes a vector space embodied with a scalar product A, B = Tr AT B , inducing
.
the norm ||A|| = A, A. However, tensors Σ frequently occurring in computer vi-
sion and image processing applications, e.g. covariance matrices and DT-MRI tensors,
embody further structure on the space of tensors: they are symmetric Σ T = Σ and
positive definite, i.e. it holds xT Σx > 0 for all nonzero x ∈ Rn . The approach to
anisotropic diffusion presented here, measures distances between tensors by the length
of the shortest path, the geodesic, with respect to GL(n) (affine) invariant Riemannian
metric on P (n). This metric takes the nonlinear structure of P (n) into account and it
has demonstrated in several other application its superiority over the flat Euclidean ma-
tric [17,18,20,21,7,22]. Such an intrinsic treatment requires the formulation of P (n) as
a Riemannian manifold, i.e. each tangent space is equipped with an inner product that
smoothly varies from point to point. A geodesic Γ X (t) parameterized by the ’time’ t
and going through the tensor Γ (0) = Σ at time t = 0 is uniquely defined by its tan-
gent vector X at Σ. This allows one to describe each geodesic by a mapping from the
subspace A = (tX), t ∈ R spanned by the tangent vector onto the manifold P (n). The
GL(n) invariant metric is induced by the scalar product

W1 , W2 Σ = Tr Σ − 2 W1 Σ −1 W2 Σ − 2 ,
1 1
(6)
as one can easily verify. The GL(n) invariant metric allows to derive an expression of
the geodesic equation going through Σ by tangent vectors X [7]
Γ Σ (t) = Σ 2 exp(tΣ − 2 XΣ − 2 )Σ 2 .
1 1 1 1
(7)
For t = 1 this map is denoted as the exponential map which is one to one in case of the
space of positive definite tensors. Its inverse, denoted as the logarithmic map, reads

1
X = Σ 2 log Σ − 2 Γ Σ (1)Σ − 2 Σ 2 .
1 1 1
(8)
As the gradient of any energy density ∇L is element of the tangent space [33], we can
formulate a diffusion process as ∂t Σ = −∇L on the tangent space. The evolution of
the tensor Σ is obtained by going a small step in the negative direction of the gradient
−dt∇L and mapping this point back on the manifold using the geodesic equation (7).
The energy density is then computed for the tangent vector at Γ Σ (dt) which in turn can
then be used for finding the next tensor in the evolving scheme as described above. This
is a gradient descent approach, denoted as the geodesic marching scheme, for energy
densities defined on P (n) and which per construction assures that we cannot leave the
manifold.
4 Riemannian Anisotropic Diffusion

After reviewing the necessary differential geometric tools, we will derive anisotropic
diffusion equations for a tensor field P (n) over RN . As done for the diffusion equation
for the scalar valued signals (Sect. 2), we derive the linear diffusion equation by vari-
ation of the corresponding energy functional and infer from the linear equation to the
nonlinear counterpart. Let ∂i Σ(x), i = 1, ..., N denote partial derivative of the tensor
field in direction i, elements of the tangent space at Σ. We define the energy functional

E(Σ) = dij ∂i Σ, ∂j ΣΣ dx (9)
i,j

with ∂i Σ, ∂j ΣΣ = Tr (∂i Σ)Σ −1 (∂j Σ)Σ −1 . (10)
The components of the diffusion tensor dij (please do not confuse dij with the elements
of the tensor field) locally controls the direction of smoothing and for the moment being
does not depend on the evolving tensor field. The gradient of the energy functional is
then derived by defining a ’test function’ W that is actually a tangent vector in the
tangent space at Σ and computing the functional derivative

δE = 2 dij Tr((∂i W ) Σ −1 (∂j Σ) Σ −1 (11)
ij
− (∂i Σ) Σ −1 (∂j Σ) Σ −1 W Σ −1 ) dx (12)

In order to get rid of the derivatives on the ’test function’ W we integrate by parts with
respect to xj . Assuming homogenous Neumann boundary conditions the functional
derivative can be brought in the form

δE = −2 W, Σ∂i (dij Σ −1 (∂j Σ)Σ −1 )Σ (13)
i,j
+(∂i Σ)Σ −1 (∂j Σ)Σ dx (14)

Comparing the inner product with the general form in (2) identifies the gradient of the
energy density

∇L = −2 Σ∂i (dij Σ −1 (∂j Σ)Σ −1 )Σ + (∂i Σ)Σ −1 (∂j Σ) . (15)
i,j
Inserting this energy density in (3) results in the desired diffusion equation. Using the
identity ∂i Σ −1 = −Σ −1 (∂i Σ)Σ −1 the energy density gradient can be simplified to

∇L = −2 ∂i ∂j Σ − (∂i Σ)Σ −1 (∂j Σ) − 2 (∂i dij )(∂j Σ) (16)
i,j i,j
The terms on the right side of (16) for which i = j hold Δi Σ= ∂i2 Σ − (∂i Σ)Σ −1
(∂i Σ) are components of the Laplace Beltrami operator Δ = i Δi derived in [7]. In
addition to the work in [20,7], we also derived mixed components
Δij Σ = ∂i ∂j Σ − (∂i Σ)Σ −1 (∂j Σ), i = j (17)
needed for the linear anisotropic diffusion equation. The nonlinear anisotropic diffu-
sion equation is defined exchanging the diffusion tensor components in (4) with com-
ponents depending on the evolved tensor field. So we have all components to define an
anisotropic diffusion equation on the space of positive definite matrices in an intrinsic
way. To this end, only the second order derivatives ∂i2 and ∂i ∂j occurring in (1) need
to be exchanged by their counterparts Δi and Δij . So far we have not specified the
explicit form of the diffusion tensor which should be made up here. We generalize the
structure tensor to the nonlinear space and afterwards, as in the case of scalar valued
images, construct the diffusion tensor from the spectral decomposition of the structure
tensor. Let ∇Σ = (∂1 Σ, ..., ∂N Σ)T denote the gradient and a a unite vector in RN
such that we can express the derivative in direction a as ∂a = aT ∇. The direction of
less variation in the tensor space can then analogous to the structure tensor in linear
spaces, be estimated by minimizing the local energy

E(a) = ∂a Σ, ∂a ΣΣ dx = aT Ja , (18)
V
where we defined the components of the structure tensors J on P (n) by Jij =

V
∂i Σ, ∂j ΣΣ dx.The diffusion tensor D is then designed as usual by exchanging
the eigenvalues λj of the structure tensor by a decreasing diffusivity
. function g(λj ).
For our numerical experiments (in 2D) we choose g(λl ) = 1/ 1 + λl /β 2 for the
larger eigenvalue and g(λs ) = 1 for the smaller eigenvalue with the heuristically cho-
sen contrast parameter β = 0.05.
5 Numerical Issues
So far we have assumed the tensor to be defined on a continuous domain. In the ex-
periential setting we are confronted with tensor fields defined on a discrete grid. The
application of Riemanian anisotropic diffusion requires a discrete approximation for
the derivatives derived in Sect. 4. In principle, we could use matrix differences to ap-
proximate the derivatives but this would contradict our effort to derive an intrinsic ex-
pression of the anisotropic diffusion equation. The finite differences are extrinsic since
they are based on Euclidian differences between tensors, i.e. they use the difference in
the space of symmetric matrices and not the Riemannian metric of the space P (n). In
order to approximate the gradient ∇L in (16) on a discrete grid, we need discrete ap-
proximations of derivatives of first and second order. Intrinsic approximations to first
order derivatives have already proposed in [20] and is reviewed here with the following
−−−−−−−−−−−−→
preposition. Let us denote with T Σ exj := Σ(x)Σ(x + εej ) the tangent vector defined
by the logarithmic map as

1
T Σ exj = Σ 2 log Σ − 2 Σ(x + εej )Σ − 2 Σ 2
1 1 1
(19)
Preposition 1. The first order discrete approximation of the first order derivative of Σ
in direction j reads
1
−−−−−−−−−−−−→ −−−−−−−−−−−−→
∂j Σ = Σ(x)Σ(x + εej ) − Σ(x)Σ(x − εej ) + O(ε) (20)
2ε
A second order discrete approximation scheme to the second order derivative in di-
rection ej has been derived in [7]. We state it here as a second preposition, for the proof
see [7].
Preposition 2. The second order discrete approximation of the second order derivative
in direction ej is
1 −−−−−−−−−−−−→ −−−−−−−−−−−−→
Δj Σ = (Σ(x)Σ(x + εej ) + Σ(x)Σ(x − εej )) + O(ε2 ) . (21)
ε2
For the anisotropic diffusion equation we also need mixed derivatives Δij Σ that can be
approximated according to preposition 3.
Preposition 3. The second order discrete approximation of the second order mixed
derivative in direction i and j is given by
Δij Σ + Δji Σ 1 −−−−−−−−−−−−→ −−−−−−−−−−−−→
= 2 (Σ(x)Σ(x + εen ) + Σ(x)Σ(x − εen ) (22)
2 ε
−−−−−−−−−−−−→ −−−−−−−−−−−−→
−Σ(x)Σ(x + εep ) − Σ(x)Σ(x − εep )) + O(ε2 ) ,
with the abbreviation en = √1 (ei
2
+ ej ), ep = √1 (ei
2
− ej ).
Proof. We expend the tangent vector as
ε2 2 ε2
∂n Σ − (∂n Σ)Σ − 2 (∂n Σ) + O(ε3 ) .
1
T Σ exn = ε∂n Σ + (23)
2 2
Now, we express the derivative in direction n by derivatives along the coordinate axes
in i and j direction , ∂n = √12 ∂i + √12 ∂j , yielding
ε ε2
T Σ exn = √ (∂i Σ + ∂j Σ) + ( ∂i2 Σ + ∂j2 Σ + 2∂i ∂j Σ
2 4
−(∂i Σ)Σ − 2 (∂i Σ) − (∂j Σ)Σ − 2 (∂j Σ)
1 1
−(∂i Σ)Σ − 2 (∂j Σ) − (∂j Σ)Σ − 2 (∂i Σ)) + O(ε3 ) .

1 1
Computing the sum T Σ Δex

n
:= T Σ exn +T Σ −e
x
n
becomes a fourth order approximation
as all uneven terms cancel out
ε2
= ( ∂i2 Σ + ∂j2 Σ + 2∂i ∂j Σ − (∂i Σ)Σ − 2 (∂i Σ) −
1
T Σ Δe
x
n
(24)
4
(∂j Σ)Σ − 2 (∂j Σ) − (∂i Σ)Σ − 2 (∂j Σ) − (∂j Σ)Σ − 2 (∂i Σ)) + O(ε4 )
1 1 1
Expanding T Σ Δe
x
p
:= T Σ exp + T Σ −e
x
p
in the same way yields
ε2 2
( ∂i Σ + ∂j2 Σ − 2∂i ∂j Σ − (∂i Σ)Σ − 2 (∂i Σ) −
1
T Σ Δe
x
p
= (25)
4
(∂j Σ)Σ − 2 (∂j Σ) + (∂i Σ)Σ − 2 (∂j Σ) + (∂j Σ)Σ − 2 (∂i Σ)) + O(ε4 )
1 1 1
By subtracting (25) from (24) and dividing the square of the grid size ε2 we obtained
the claimed second order approximation for the mixed derivatives which concludes the
proof.
6 Experiments
Performance of our Riemannian anisotropic diffusion (RAD) approach is demonstrated

on synthetic tensor fields and real DT-MRI data. We compare our Riemannian anisotropic
diffusion scheme with three state of the art tensor valued regularization schemes: the
anisotropic diffusion (EAD) scheme based on the flat Euclidean metric [12,31], the in-
trinsic nonlinear isotropic diffusion (RID) scheme [20] and the nonlinear isotropic diffu-
sion (LEID) scheme based on the log-Euclidean metric [34]. As a computational effective
alternative to our Riemannian anisotropic diffusion scheme, we propose to combine the
diffusion scheme proposed in [12,31] with the log-Euclidean metric [34] which is con-
sidered as a fourth reference method (LEAD). As a performance measure for the reg-
ularized tensor field, we choose the fractional anisotropy (FA) [35]. Measures derived
from DT-MRI such as the FA are used to generate additional image contrast required
for detection of brain lesions, or to delineate white matter (highly directional structures)
from non-white matter tissue, which is important for surgery. FA takes on values between
0 (corresponding to perfect isotropy) and 1 indicating maximal anisotropy. For solving
the diffusion equations, we used the same time step of dt = 0.01 for all experiments and
computed the evolving tensor field for 1000 time steps. As shown in [21], the linear gradi-
ent descent scheme realizes a first order approximation to the intrinsic marching scheme,
such that for small time steps diffusion processes based on different metrics should be
comparable for distinct times.
Fig. 1. Line reconstruction experiment; upper row (from left to right): original tensor field, EAD
scheme, LEAD scheme; lower row (from left to right): LEID scheme, RID scheme, RAD scheme
Fig. 2. Denoising experiment; upper row (from left to right): noise corrupted tensor field, EAD
scheme, our LEAD scheme; lower row (from left to right): LEID scheme, RID scheme, RAD
scheme
6.1 Synthetic Data
Experiment 1. In the first experiment on synthetic data we examine the ability of the
different diffusion processes to complete interrupted line structures. To this end, we
generate a 32 × 32 large tensor field of 3 × 3 tensors (see Fig. 2 upper left; in order to
visualize details more precise only a cutout of the tensor field is shown). Each tensor is
represented by an ellipsoid and the orientation of its main axis is additional color coded
whereas the FA is encoded in the saturation of the depicted tensors. The line structure is
interrupted by isotropic tensors with small eigenvalues (λj = 0.05) that are hardly vis-
ible due to the saturation encoding of the FA. The results for all diffusion processes are
shown in Fig. 1. The nonlinear isotropic processes LEID and RID stops at the line inter-
ruption and is not able to complete the line. This results from the fact that, although the
smoothing process is also anisotropic for nonlinear isotropic diffusion processes [20],
the diffusivity function depends only on its direct neighbors and therefore does not ’see’
the line behind the gap. The anisotropic diffusion schemes are steered by the diffusion
tensor which encodes the directional information of a neighborhood depending on the
average region for the structure tensor. The anisotropic diffusion approaches fill the gap
and reconstruct the line. However, again the EAD-process suffers from the eigenvalue
swelling effect and only one tensor connects both interrupted line structures. However
increasing the average region of the structure tensor might fill the gap more clearly. Our
RAD and LEAD schemes reconstruct the line structure. However, we observe a small
decreasing of the anisotropy for the log-Euclidean metric, whereas the anisotropy for
the affine invariant metric increases in the vicinity of image borders.
Experiment 2. In this experiment we examine the ability of the different diffusion

schemes to reconstruct the tensor field from noisy data. To this end, we corrupt the tan-
gent vector of each tensor by Gaussian noise (with standard deviations σ = 0.6). Fig. 2
shows the noise corrupted field (the noise free tensor field is the same as in experiment
1) and the evolved tensor fields for the different diffusion schemes. The anisotropic
schemes manage (more or less) to close the gap in the line structure despite the noise
whereas the isotropic schemes does not. The schemes based on the log Euclidean met-
ric lead to a slight decrease of the anisotropy whereas the RAD schemes leads to an
increase of the anisotropy in the tensor field. How this effect influences further process-
ing steps, e.g. fiber tracking algorithm, is left to be examined for future research.
6.2 Real Data

Experiment 3. In our last experiment, the different algorithms were applied to DT-
MRI data measured from a human brain in-vivo. DT-MRI of the brain of a healthy vol-
unteer (written informed consent was obtained) was performed on a 1.5 T Magnetom
Avanto scanner (Siemens Medical Solutions). A single-shot diffusion-weighted twice-
refocused spin-echo planar imaging sequence was used. Measurement parameters were
Fig. 3. Denoising experiment 3; (upper row, from left to right): noisy DT-MRI image, LEID
scheme, RID scheme; (lower row, from left to right) EAD scheme, LEAD scheme, RAD scheme
as follows : TR = 6925 ms / TE=104ms / 192 matrix with 6/8 phase partial fourier, 23
cm field of view (FOV), and 36 2.4-mm-thick contiguous axial slices. The in-plane res-
olution was 1.2 mm/pixel. We estimate a volumetric tensor field of size 192 × 192 × 36
and take one slice for further processing. For evaluation purposes we recorded tensor
fields of the brain with 6 different signal-to-noise ratios (SNR), denoted as DTI1-6 in
the following. Thus, we can use the DT-MRI-images (DTI6) from the long measure-
ment (i.e. good SNR) as a reference data set, where we compare the FA of the tensor
with the results obtained from the lower SNR data set (DTI1-5), which can be obtained
in a clinical feasible measurement time. We compute, starting from the five different
noisy tensor fields, the evolved tensor fields for all considered diffusion schemes (Fig. 3
shows cutouts of the noisy field and evolved fields) and compare its FA with the ref-
erence field. All schemes lead to rather smooth tensor fields. However, the anisotropic
diffusion schemes (EAD, LEAD and RAD) lead to an enhancement of orientated struc-
tures within in the tensor fields which is most distinct for our RAD scheme. As in the
previous experiments, the eigenvalue swelling effect in case of the EAD scheme can be
observed. Our RAD/LEAD schemes yield the best results among anisotropic regular-
ization schemes with respect to the FA measure as shown in Tab. 1.
Table 1. Results of experiment 3: The average and standard deviation of the the fractional
anisotropy error |F A − F A| (FA belongs to the reference tensor field) over 1000 time steps
for each diffusion scheme as well as for five different noise levels are computed
Method DTI1 DTI2 DTI3 DTI4 DTI5

EAD 0.098 ± 0.007 0.100 ± 0.008 0.103 ± 0.008 0.109 ± 0.009 0.112 ± 0.010
RID 0.112 ± 0.016 0.119 ± 0.015 0.116 ± 0.013 0.114 ± 0.012 0.113 ± 0.013
LEID 0.099 ± 0.017 0.108 ± 0.017 0.107 ± 0.014 0.106 ± 0.012 0.105 ± 0.012
LEAD 0.078 ± 0.005 0.079 ± 0.006 0.081 ± 0.006 0.084 ± 0.007 0.086 ± 0.007
RAD 0.089 ± 0.004 0.089 ± 0.005 0.093 ± 0.007 0.096 ± 0.007 0.098 ± 0.009
7 Conclusion
We generalized the concept of anisotropic diffusion to tensor valued data with respect
to the affine invariant Riemannian metric. We derived the intrinsic mixed second order
derivatives as they are required for the anisotropic diffusion process. Furthermore, we
derived a discrete intrinsic approximation scheme for the mixed second order deriva-
tives. Since mixed second order derivatives appear also in other methods based on par-
tial differential equation, this contribution could also serve as a basis for generalizing
these methods in an intrinsic way in a discrete formulation. Experiments on synthetic
as well as real world data demonstrate the value of our full intrinsic differential geo-
metrical formulation of the anisotropic diffusion concept. As a computational effective
alternative, we proposed an anisotropic diffusion scheme based on the log-Euclidean
metric. Summing up, our proposed anisotropic diffusion schemes show promising re-
sults on the given test images. Further work might examine the reconstruction proper-
ties of other tensor characteristics as well as the influence on so far heuristically chosen
parameters, e.g. the diffusivity function.
References
1. Berger, M.-O., Deriche, R., Herlin, I., Jaffré, J., Morel, J.-M. (eds.): Icaos 1996: Images and
wavelets and PDEs. Lecture Notes in Control and Information Sciences, vol. 219 (1996)
2. Weickert, J.: Anisotropic diffusion in image processing. Teubner, Stuttgart (1998)
3. Perona, P., Malik, J.: Scale-space and edge detection using anisotropic diffusion. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence 12 (1990)
4. Bigün, J., Granlund, G.H.: Optimal orientation detection of linear symmetry. In: ICCV, Lon-
don, UK, pp. 433–438 (1987)
5. Tschumperlé, D., Deriche, R.: Diffusion tensor regularization with constraints preservation.
In: Proc. IEEE Computer Society Conference on Computer Vision and Pattern Recognition
(CVPR 2001), pp. 948–953 (2001)
6. Chefd’hotel, C., Tschumperlé, D., Deriche, R., Faugeras, O.: Regularizing flows for con-
strained matrix-valued images. J. Math. Imaging Vis. 20(1-2), 147–162 (2004)
7. Pennec, X., Fillard, P., Ayache, N.: A Riemannian framework for tensor computing. Interna-
tional Journal of Computer Vision 66(1), 41–66 (2006)
8. Castano-Moraga, C.A., Lenglet, C., Deriche, R., Ruiz-Alzola, J.: A Riemannian approach to
anisotropic filtering of tensor fields. Signal Processing 87(2), 263–276 (2007)
9. Weickert, J., Hagen, H.: Visualization and Processing of Tensor Fields (Mathematics and
Visualization). Springer, New York (2005)
10. Rao, C.: Information and accuracy attainable in estimation of statistical parameters. Bull.
Calcutta Math. Soc. 37, 81–91 (1945)
11. Martin-Fernandez, M., San-Jose, R., Westin, C.F., Alberola-Lopez, C.: A novel Gauss-
Markov random field approach for regularization of diffusion tensor maps. In: Moreno-Dı́az
Jr., R., Pichler, F. (eds.) EUROCAST 2003. LNCS, vol. 2809, pp. 506–517. Springer, Hei-
delberg (2003)
12. Weickert, J., Brox, T.: Diffusion and regularization of vector- and matrix-valued images. In:
Inverse Problems, Image Analysis, and Medical Imaging. Contemporary Mathematics, pp.
251–268 (2002)
13. Westin, C.-F., Knutsson, H.: Tensor field regularization using normalized convolution. In:
Moreno-Dı́az Jr., R., Pichler, F. (eds.) EUROCAST 2003. LNCS, vol. 2809, pp. 564–572.
14. Burgeth, B., Didas, S., Florack, L., Weickert, J.: A generic approach to the filtering of matrix
fields with singular PDEs. In: Sgallari, F., Murli, A., Paragios, N. (eds.) SSVM 2007. LNCS,
15. Gur, Y., Sochen, N.A.: Denoising tensors via Lie group flows. In: Paragios, N., Faugeras, O.,
Chan, T., Schnörr, C. (eds.) VLSM 2005. LNCS, vol. 3752, pp. 13–24. Springer, Heidelberg
(2005)
16. Moakher, M.: A differential geometric approach to the geometric mean of symmetric
positive-definite matrices. SIAM J. Matrix Anal. Appl (2003)
17. Fletcher, P., Joshi, S.: Principle geodesic analysis on symmetric spaces: Statistics of diffusion
tensors. In: Computer Vision and Mathematical Methods in Medical and Biomedical Image
Analysis, ECCV 2004 Workshops CVAMIA and MMBIA, pp. 87–98 (2004)
18. Lenglet, C., Rousson, M., Deriche, R., Faugeras, O.D., Lehericy, S., Ugurbil, K.: A Rie-
mannian approach to diffusion tensor images segmentation. In: Christensen, G.E., Sonka, M.
(eds.) IPMI 2005. LNCS, vol. 3565, pp. 591–602. Springer, Heidelberg (2005)
19. Batchelor, P.G., Moakher, M., Atkinson, D., Calamante, F., Connelly, A.: A rigorous frame-
work for diffusion tensor calculus. Magn. Reson. Med. 53(1), 221–225 (2005)
20. Fillard, P., Arsigny, V., Ayache, N., Pennec, X.: A Riemannian framework for the processing
of tensor-valued images. In: Fogh Olsen, O., Florack, L.M.J., Kuijper, A. (eds.) DSSCV
2005. LNCS, vol. 3753, pp. 112–123. Springer, Heidelberg (2005)
21. Lenglet, C., Rousson, M., Deriche, R., Faugeras, O.: Statistics on the manifold of multivariate
normal distributions: Theory and application to diffusion tensor MRI processing. J. Math.
Imaging Vis. 25(3), 423–444 (2006)
22. Zéraı̈, M., Moakher, M.: Riemannian curvature-driven flows for tensor-valued data. In: Sgal-
lari, F., Murli, A., Paragios, N. (eds.) SSVM 2007. LNCS, vol. 4485, pp. 592–602. Springer,
Heidelberg (2007)
23. Helgason, S.: Differential Geometry, Lie groups and symmetric spaces. Academic Press,
London (1978)
24. El-Fallah, A., Ford, G.: On mean curvature diffusion in nonlinear image filtering. Pattern
Recognition Letters 19, 433–437 (1998)
25. Sochen, N., Kimmel, R., Malladi, R.: A geometrical framework for low level vision. IEEE
Transaction on Image Processing, Special Issue on PDE based Image Processing 7(3), 310–
318 (1998)
26. Begelfor, E., Werman, M.: Affine invariance revisited. In: CVPR ’06: Proceedings of the
2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp.
2087–2094. IEEE Computer Society Press, Washington (2006)
27. Zhang, F., Hancock, E.: Tensor MRI regularization via graph diffusion. In: BMVC 2006, pp.
578–589 (2006)
28. Arsigny, V., Fillard, P., Pennec, X., Ayache, N.: Log-Euclidean metrics for fast and simple
calculus on diffusion tensors. Magnetic Resonance in Medicine 56(2), 411–421 (2006)
29. Gur, Y., Sochen, N.A.: Fast invariant Riemannian DT-MRI regularization. In: Proc. of IEEE
Computer Society Workshop on Mathematical Methods in Biomedical Image Analysis (MM-
BIA), Rio de Janeiro, Brazil, pp. 1–7 (2007)
30. Mémoli, F., Sapiro, G., Osher, S.: Solving variational problems and partial differential equa-
tions mapping into general target manifolds. Journal of Computational Physics 195(1), 263–
292 (2004)
31. Brox, T., Weickert, J., Burgeth, B., Mrázek, P.: Nonlinear structure tensors. Revised version
of technical report no. 113. Saarland University, Saarbrücken, Germany (2004)
32. Nielsen, M., Johansen, P., Olsen, O., Weickert, J. (eds.): Scale-Space 1999. LNCS, vol. 1682.
33. Maaß, H.: Siegel’s Modular Forms and Dirichlet Series. Lecture notes in mathematics,
vol. 216. Springer, Heidelberg (1971)
34. Arsigny, V., Fillard, P., Pennec, X., Ayache, N.: Fast and simple calculus on tensors in the log-
Euclidean framework. In: Duncan, J.S., Gerig, G. (eds.) MICCAI 2005. LNCS, vol. 3749,
35. Bihan, D.L., Mangin, J.F., Poupon, C., Clark, C.A., Pappata, S., Molko, N., Chabriat, H.:
Diffusion tensor imaging: Concepts and applications. Journal of Magnetic Resonance Imag-
ing 13(4), 534–546 (2001)
FaceTracer: A Search Engine for
Large Collections of Images with Faces
Neeraj Kumar1, , Peter Belhumeur1 , and Shree Nayar1
Columbia University
Abstract. We have created the first image search engine based entirely
on faces. Using simple text queries such as “smiling men with blond hair
and mustaches,” users can search through over 3.1 million faces which
have been automatically labeled on the basis of several facial attributes.
Faces in our database have been extracted and aligned from images down-
loaded from the internet using a commercial face detector, and the num-
ber of images and attributes continues to grow daily. Our classification
approach uses a novel combination of Support Vector Machines and Ad-
aboost which exploits the strong structure of faces to select and train on
the optimal set of features for each attribute. We show state-of-the-art
classification results compared to previous works, and demonstrate the
power of our architecture through a functional, large-scale face search
engine. Our framework is fully automatic, easy to scale, and computes
all labels off-line, leading to fast on-line search performance. In addition,
we describe how our system can be used for a number of applications,
including law enforcement, social networks, and personal photo manage-
ment. Our search engine will soon be made publicly available.
1 Introduction
We have created the first face search engine, allowing users to search through
large collections of images which have been automatically labeled based on the
appearance of the faces within them. Our system lets users search on the basis
of a variety of facial attributes using natural language queries such as, “men
with mustaches,” or “young blonde women,” or even, “indoor photos of smiling
children.” This face search engine can be directed at all images on the internet,
tailored toward specific image collections such as those used by law enforcement
or online social networks, or even focused on personal photo libraries.
The ability of current search engines to find images based on facial appear-
ance is limited to images with text annotations. Yet, there are many problems
with annotation-based search of images: the manual labeling of images is time-
consuming; the annotations are often incorrect or misleading, as they may refer
to other content on a webpage; and finally, the vast majority of images are

Supported by the National Defense Science & Engineering Graduate Fellowship.

FaceTracer: A Search Engine for Large Collections of Images with Faces 341
(a) (b)
Fig. 1. Results for the query “smiling asian men with glasses,” using (a) the Google
image search engine and (b) our face search engine. Our system currently has over
3.1 million faces, automatically detected and extracted from images downloaded from
the internet, using a commercial face detector [1]. Rather than use text annotations
to find images, our system has automatically labeled a large number of different facial
attributes on each face (off-line), and searches are performed using only these labels.
Thus, search results are returned almost instantaneously. The results also contain links
pointing back to the original source image and associated webpage.
simply not annotated. Figures 1a and 1b show the results of the query, “smil-
ing asian men with glasses,” using a conventional image search engine (Google
Image Search) and our search engine, respectively. The difference in quality of
search results is clearly visible. Google’s reliance on text annotations results in
it finding images that have no relevance to the query, while our system returns
only the images that match the query.
Like much of the work in content-based image retrieval, the power of our
approach comes from automatically labeling images off-line on the basis of a
large number of attributes. At search time, only these labels need to be queried,
resulting in almost instantaneous searches. Furthermore, it is easy to add new
images and face attributes to our search engine, allowing for future scalability.
Defining new attributes and manually labeling faces to match those attributes
can also be done collaboratively by a community of users.
Figures 2a and 2b show search results of the queries, “young blonde women”
and “children outdoors,” respectively. The first shows a view of our extended
interface, which displays a preview of the original image in the right pane when
the user holds the mouse over a face thumbnail. The latter shows an example of
a query run on a personalized set of images. Incorporating our search engine into
photo management tools would enable users to quickly locate sets of images and
then perform bulk operations on them (e.g., edit, email, or delete). (Since current
tools depend on manual annotation of images, they are significantly more time-
consuming to use.) Another advantage of our attribute-based search on personal
collections is that with a limited number of people, simple queries can often find
images of a particular person, without requiring any form of face recognition.
342 N. Kumar, P. Belhumeur, and S. Nayar
(a) (b)
Fig. 2. Results of queries (a)“young blonde women” and (b) “children outside,” using
our face search engine. In (a), search results are shown in the left panel, while the right
panel shows a preview of the original image for the selected face. (b) shows search
results on a personalized dataset, displaying the results as thumbnails of the original
images. Note that these results were correctly classified as being “outside” using only
the cropped face images, showing that face images often contain enough information
to describe properties of the image which are not directly related to faces.
Our search engine owes its superior performance to the following factors:
– A large and diverse dataset of face images with a significant subset
containing attribute labels. We currently have over 3.1 million aligned
faces in our database – the largest such collection in the world. In addition to
its size, our database is also noteworthy for being a completely “real-world”
dataset. The images are downloaded from the internet and encompass a wide
range of pose, illumination, imaging conditions, and were taken using a large
variety of cameras. The faces have been automatically extracted and aligned
using a commercial face and fiducial point detector [1]. In addition, 10 at-
tributes have been manually labeled on more than 17,000 of the face images,
creating a large dataset for training and testing classification algorithms.
– A scalable and fully automatic architecture for attribute classi-
fication. We present a novel approach tailored toward face classification
problems, which uses a boosted set of Support Vector Machines (SVMs) [2]
to form a strong classifier with high accuracy. We describe the results of this
algorithm on a variety of different attributes, including demographic infor-
mation such as gender, age, and race; facial characteristics such as eye wear
and facial hair; image properties such as blurriness and lighting conditions;
and many others as well. A key aspect of this work is that classifiers for
new attributes can be trained automatically, requiring only a set of labeled
examples. Yet, the flexibility of our framework does not come at the cost of
reduced accuracy – we compare against several state-of-the-art classification
methods and show the superior classification rates produced by our system.
We will soon be releasing our search engine for public use.
2 Related Work
Our work lies at the intersection of several fields, including computer vision,
machine learning, and content-based image retrieval. We present an overview of
the relevant work, organized by topic.
Attribute Classification. Prior works on attribute classification have focused

mostly on gender and ethnicity classification. Early works such as [3]used neural
networks to perform gender classification on small datasets. The Fisherfaces
work of [4] showed that linear discriminant analysis could be used for simple
attribute classification such as glasses/no glasses. More recently, Moghaddam
and Yang [5] used Support Vector Machines (SVMs) [2] trained on small “face-
prints” to classify the gender of a face, showing good results on the FERET
face database [6]. The works of Shakhnarovich et al. [7] and Baluja & Rowley
[8] used Adaboost [9] to select a linear combination of weak classifiers, allowing
for almost real-time classification of faces, with results in the latter case again
demonstrated on the FERET database. These methods differ in their choice of
weak classifiers: the former uses the Haar-like features of the Viola-Jones face
detector [10], while the latter uses simple pixel comparison operators.
In contrast, we develop a method that combines the advantages of SVMs and
Adaboost (described in Sect. 4). We also present results of an extensive com-
parison against all three of these prior methods in Sect. 5. Finally, we note that
this is an active area of research, and there are many other works on attribute
classification which use different combinations of learning techniques, features,
and problem formulations [11,12]. An exploration of the advantages and disad-
vantages of each is beyond the scope of this paper.
Content-Based Image Retrieval (CBIR). Our work can also be viewed as

a form of CBIR, where our content is limited to images with faces. Interested
readers can refer to the work of Datta et al. [13] for a recent survey of this
field. Most relevant to our work is the “Photobook” system [14], which allows
for similarity-based searches of faces and objects using parametric eigenspaces.
However, their goal is different from ours. Whereas they try to find objects
similar to a chosen one, we locate a set of images starting only with simple
text queries. Although we use vastly different classifiers and methods for feature
selection, their division of the face into functional parts such as the eyes, nose,
etc., is echoed in our approach of training classifiers on functional face regions.
3 Creating the Face Database

To date, we have built a large database of over 3.1 million face images extracted
from over 6.2 million images collected from the internet. This database con-
tinues to grow as we automatically collect, align, and assign attributes to face
images daily. An overview of the database creation process is illustrated in Fig. 3.
We download images using two different methods – keyword searches and ran-
dom downloads. The first allows us to build datasets related to particular terms
Fig. 3. Overview of database creation. See text for details.
(e.g., celebrity names and professions). The latter allows us to sample from the
more general distribution of images on the internet. In particular, it lets us
include images that have no corresponding textual information, i.e., that are
effectively invisible to current image search engines. Our images are downloaded
from a wide variety of online sources, such as Google Images, Microsoft Live
Image Search, and Flickr, to name a few. Relevant metadata such as image and
page URLs are stored in the EXIF tags of the downloaded images.
Next, we apply the OKAO face detector [1] to the downloaded images to
extract faces. This detector also gives us the pose angles of each face, as well as
the locations of six fiducial points (the corners of both eyes and the corners of
the mouth). We filter the set of faces by resolution and face pose (±10◦ from
front-center). Finally, the remaining faces are aligned to a canonical pose by
applying an affine transformation. This transform is computed using linear least
squares on the detected fiducial points and corresponding points defined on a
template face. (In future work, we intend to go beyond near frontal poses.)
We present various statistics of our current face database in Table 1, divided
by image source. We would like to draw attention to three observations about
our data. First, from the statistics of randomly downloaded images, it appears
that a significant fraction of them contain faces (25.7%), and on average, each
image contains 0.5 faces. Second, our collection of aligned faces is the largest
such collection of which we are aware. It is truly a “real-world” dataset, with
completely uncontrolled lighting and environments, taken using unknown cam-
eras and in unknown imaging conditions, with a wide range of image resolutions.
In this respect, our database is similar to the LFW dataset [15], although ours is
larger by 2 orders of magnitude and not targeted specifically for face recognition.
In contrast, existing face datasets such as Yale Face A&B [16], CMU PIE [17],
and FERET [6] are either much smaller in size and/or taken in highly controlled
settings. Even the more expansive FRGC version 2.0 dataset [18] has a limited
number of subjects, image acquisition locations, and all images were taken with
the same camera type. Finally, we have labeled a significant number of these im-
ages for our 10 attributes, enumerated in Table 2. In total, we have over 17,000
attribute labels.
4 Automatic Attribute Classification for Face Images

Our approach to image search relies on labeling each image with a variety of
attributes. For a dataset as large as ours, it is infeasible to manually label every
Table 1. Image database statistics. We have collected what we believe to be the largest
set of aligned real-world face images (over 3.1 million so far). These faces have been
extracted using a commercial face detector [1]. Notice that more than 45% of the
downloaded images contain faces, and on average, there is one face per two images.
Total # Average #
Image Source # Images # Images % Images Faces Found
Faces
Downloaded With Faces With Faces Per Image
Found
Randomly Downloaded 4,289,184 1,102,964 25.715 2,156,287 0.503
Celebrities 428,312 411,349 96.040 285,627 0.667
Person Names 17,748 7,086 39.926 10,086 0.568
Face-Related Words 13,028 5,837 44.804 14,424 1.107
Event-Related Words 1,658 997 60.133 1,335 0.805
Professions 148,782 75,105 50.480 79,992 0.538
Series 7,472 3,950 52.864 8,585 1.149
Camera Defaults 895,454 893,822 99.818 380,682 0.425
Miscellanous 417,823 403,233 96.508 194,057 0.464
Total 6,219,461 2,904,343 46.698 3,131,075 0.503
Table 2. List of labeled attributes. The labeled face images are used for training our
classifiers, allowing for automatic classification of the remaining faces in our database.
Note that these were labeled by a large set of people, and thus the labels reflect a group
consensus about each attribute rather than a single user’s strict definition.
Attribute/ Number Attribute/ Number Attribute/ Number

Options Labeled Options Labeled Options Labeled
Gender 1,954 Smiling 1,571 Race 1,309
Male 867 True 832 White 433
Female 1,087 False 739 Black 399
Age 3,301 Mustache 1,947 Asian 477
Baby 577 True 618 Eye Wear 2,360
Child 636 False 1,329 None 1,256
Youth 784 Blurry 1,763 Eyeglasses 665
Middle Aged 815 True 763 Sunglasses 439
Senior 489 False 1,000 Environment 1,583
Hair Color 1,033 Lighting 633 Outdoor 780
Black 717 Flash 421 Indoor 803
Blond 316 Harsh 212 Total 17,454
image. Instead, we use our large sets of manually-labeled images to build accurate
classifiers for each of the desired attributes.
In creating a classifier for a particular attribute, we could simply choose all
pixels on the face, and let our classifier figure out which are important for the
task and which are not. This, however, puts too great a burden on the classifier,
confusing it with non-discriminative features. Instead, we create a rich set of local
feature options from which our classifier can automatically select the best ones.
Each option consists of four choices: the region of the face to extract features
from, the type of pixel data to use, the kind of normalization to apply to the
data, and finally, the level of aggregation to use.
Face Regions. We break up the face into a number of functional regions, such as
the nose, mouth, etc., much like those defined in the work on modular eigenspaces
Fig. 4. The face regions used for automatic feature selection. On the left is one region
corresponding to the whole face, and on the right are the remaining regions, each
corresponding to functional parts of the face. The regions are large enough to be robust
against small differences between individual faces and overlap slightly so that small
errors in alignment do not cause a feature to go outside of its region. The letters in
parentheses denote the code letter for the region, used later in the paper.
[19]. The complete set of 10 regions we use are shown in Fig. 4. Our coarse divi-
sion of the face allows us to take advantage of the common geometry shared by
faces, while allowing for differences between individual faces, as well as robust-
ness to small errors in alignment.
Types of Pixel Data. We include different color spaces and image derivatives
as possible feature types. These can often be more discriminative than standard
RGB values for certain attributes. Table 3 lists the various options.
Normalizations. Normalizations are important for removing lighting effects,
allowing for better generalization across images. We can remove illumination
gains by using mean normalization, x̂ = μx , or both gains and offsets by using
σ . In these equations, x refers to the input value, μ
energy normalization, x̂ = x−μ
and σ are the mean and standard deviation of all the x values within the region,
and x̂ refers to the normalized output value.
Aggregations. For some attributes, aggregate information over the entire re-
gion might be more useful than individual values at each pixel. This includes
histograms of values over the region, or simply the mean and variance.
To concisely refer to a complete feature option, we define a shorthand nota-
tion using the format, “Region:pixel type.normalization.aggregation.” The re-
gion notation is shown in Fig. 4; the notation for the pixel type, normalization,
and aggregation is shown in Table 3.
4.1 Classifier Architecture
In recent years, Support Vector Machines (SVMs) [2] have been used success-
fully for many classification tasks [20,21]. SVMs aim to find the linear hyper-
plane which best separates feature vectors of two different classes, so as to
Table 3. Feature type options. A complete feature type is constructed by first convert-
ing the pixels in a given region to one of the pixel value types from the first column,
then applying one of the normalizations from the second column, and finally aggregat-
ing these values into the output feature vector using one of the options from the last
column. The letters in parentheses are used as code letters in a shorthand notation for
concisely designating feature types.
Pixel Value Types Normalizations Aggregation

RGB (r) None (n) None (n)
HSV (h) Mean-Normalization (m) Histogram (h)
Image Intensity (i) Energy-Normalization (e) Statistics (s)
Edge Magnitude (m)
Edge Orientation (o)
simultaneously minimize the number of misclassified examples (training error)

and maximize the distance between the classes (the margin).
As with many classification algorithms, SVMs perform best when given only
the relevant data – too many extraneous inputs can confuse or overtrain the
classifier, resulting in poor accuracy on real data. In particular, if we would like
to train a classifier for an attribute that is only dependent on a certain part
of the face (e.g., “is smiling?”), giving the SVM a feature vector constructed
from all the pixels of the face is unlikely to yield optimal results. Given the
large number of regions and feature types described in the previous section,
an efficient and automatic selection algorithm is needed to find the optimal
combination of features for each attribute. Following the successes of [10,7,8,11],
we use Adaboost [9] for this purpose.
Adaboost is a principled, iterative approach for building strong classifiers out
of a collection of “weak” classifiers. In each iteration of Adaboost, the weak
classifier that best classifies a set of weighted examples is greedily picked to
form part of the final classifier. The weights on the examples are then adjusted
to make misclassified examples more important in future iterations, and the
process is repeated until a given number of weak classifiers has been picked. A
major advantage of Adaboost is that it is resistant to overtraining [22,23].
We combine the strengths of these two methods by constructing a number of
“local” SVMs and letting Adaboost create an optimal classifier using a linear
combination of them. We create one SVM for each region, feature type, and SVM
parameter combination, using the LibSVM library [24]. Normally, Adaboost is
performed using weak classifiers, which need to be retrained at the beginning
of each round. However, we rely on the fact that our local SVMs will either be
quite powerful (if created using the relevant features for the current attribute),
or virtually useless (if created from irrelevant features). Retraining will not sig-
nificantly improve the classifiers in either case.
Accordingly, we precompute the results of each SVM on all examples, one
SVM at a time. Thus, our classifiers remain fixed throughout the Adaboost
process, and we do not need to keep a large number of SVMs in memory. Once
all SVM outputs have been computed, we run our Adaboost rounds to obtain the
Table 4. Error rates and top feature combinations for each attribute, computed by
training on 80% of the labeled data and testing on the remaining 20%, averaging over
5 runs (5-fold cross-validation). Note that the attribute-tuned global SVM performs as
well as, or better than, the local SVMs in all cases, and requires much less memory and
computation than the latter. The top feature combinations selected by our algorithm
are shown in ranked order from more important to less as “Region:feature type” pairs,
where the region and feature types are listed using the code letters from Fig. 4 and
Table 3. For example, the first combination for the hair color classifier, “H:r.n.s,” takes
from the hair region (H) the RGB values (r) with no normalization (n) and using only
the statistics (s) of these values.
Error Rates Error Rates Top Feature Combinations in

for Attribute- for Attribute- Ranked Order
Attribute
Tuned Local Tuned Global Each combination is represented
SVMs SVM as Region:pixtype.norm.aggreg
W:i.m.n | W:o.n.n | W:i.n.n |
Gender 9.42% 8.62%
W:i.e.n
W:i.m.n | W:i.n.n | H:r.e.n |
Age 17.34% 16.65%
E:r.m.n | H:r.e.s | W:o.n.n
W:i.m.n | E:r.e.n | C:o.n.n |
Race 7.75% 6.49%
M:r.m.n | W:o.n.n
H:r.n.s | W:i.m.n | E:r.m.n |
Hair Color 7.85% 5.54%
H:r.n.n | W:i.n.n | H:r.m.n
W:m.n.n | W:i.n.n | K:o.n.h |
Eye Wear 6.22% 5.14%
W:m.m.n | N:r.n.n
Mustache 6.42% 4.61% U:r.e.n | M:r.m.n
M:r.m.n | M:r.n.n | M:r.e.n |
Smiling 4.60% 4.60%
W:i.n.n | W:i.e.n | M:i.n.n
W:m.m.n | H:m.n.n | W:m.n.n |
Blurry 3.94% 3.41%
H:m.m.n | M:m.m.n
W:i.n.n | W:i.e.n | K:r.n.n |
Lighting 2.82% 1.61%
C:o.n.n | E:o.n.n
N:r.m.n | K:r.e.n | K:r.m.n |
Environment 12.25% 12.15%
W:r.m.n | E:r.m.n
weights on each SVM classifier. We use the formulation of Adaboost described in

[8], with the modification that errors are computed in a continuous manner (using
the confidence values obtained from the SVM classifier), rather than discretely
as is done in [8]. We found this change improves the stability of the results,
without adversely affecting the error rates.
The error rates of these “attribute-tuned local SVMs” are shown in the second
column of Table 4. The rates were computed by dividing the labeled examples
for each attribute into 5 parts, using 4 parts to train and the remaining one to
test, and then rotating through all 5 sets (5-fold cross-validation). Note that in
most cases, our error rates are below 10%, and for many attributes, the error rate
is under 5%. (The higher error rates for age are due to the fact that different
people’s labels for each of the age categories did not match up completely.)
(a) (b)
(c) (d)
Fig. 5. Illustrations of automatically-selected region and feature types for (a) gender,
(b) smiling, (c) environment, and (d) hair color. Each face image is surrounded by
depictions of the top-ranked feature combinations for the given attribute, along with
their corresponding shorthand label (as used in Table 4). Notice how each classifier
uses different regions and feature types of the face.
We emphasize the fact that these numbers are computed using our real-world
dataset, and therefore reflect performance on real images.
A limitation of this architecture is that classification will require keeping a
possibly large number of SVMs in memory, and each one will need to be evalu-
ated for every input image. Furthermore, one of the drawbacks of the Adaboost
formulation is that different classifiers can only be combined linearly. Attributes
which might depend on non-linear combinations of different regions or feature
types would be difficult to classify using this architecture.
We solve both of these issues simultaneously by training one “global” SVM
on the union of the features from the top classifiers selected by Adaboost. We do
this by concatenating the features from the N highest-weighted SVMs (from the
output of Adaboost), and then training a single SVM classifier over these features
(optimizing over N ). In practice, the number of features chosen is between 2
(for “mustache”) and 6 (e.g., for “hair color”). Error rates for this algorithm,
denoted as “Attribute-Tuned Global SVM,” are shown in the third column of
Table 4. Notice that for each attribute, these rates are equal to, or less than,
the rates obtained using the combination of local SVMs, despite the fact that
these classifiers run significantly faster and require only a fraction of the memory
(often less by an order of magnitude).
The automatically-selected region and feature type combinations for each at-
tribute are shown in the last column of Table 4. Listed in order of decreasing
importance, the combinations are displayed in a shorthand notation using the
codes given in Fig. 4 and Table 3. In Fig. 5, we visually illustrate the top feature
Table 5. Comparison of classification performance against prior methods. Our

attribute-tuned global SVM performs better than prior state-of-the-art methods. Note
the complementary performances of both Adaboost methods versus the full-face SVM
method for the different attributes, showing the strengths and weaknesses of each
method. By exploiting the advantages of each method, our approach achieves the best
performance.
Classification Method Gender Error Rate Smiling Error Rate

Attribute-Tuned Global SVM 8.62% 4.60%
Adaboost (pixel comparison feats.) [9] 13.13% 7.41%
Adaboost (Haar-like feats.) [8] 12.88% 6.40%
Full-face SVM [6] 9.52% 13.54%
combinations chosen for the gender, smiling, environment, and hair color at-
tributes. This figure shows the ability of our feature selection approach to iden-
tify the relevant regions and feature types for each attribute.
5 Comparison to Prior Work

While we have designed our classifier architecture to be flexible enough to handle
a large variety of attributes, it is important to ensure that we have not sacrificed
accuracy in the process. We therefore compare our approach to three state-
of-the-art methods for attribute classification: full-face SVMs using brightness
normalized pixel values [5], Adaboost using Haar-like features [7], and Adaboost
using pixel comparison features [8]. Since these works have mostly focused on
gender classification, we use that attribute as our first testing criteria.
The error rates for gender classification using our training and testing data on
all methods are shown in the second column of Table 5. We note that our method
performs slightly better than the prior SVM method and significantly better than
both Adaboost methods. The difference between the Adaboost and SVM methods
may reflect one limitation of using linear combinations of weak classifiers – the
classifiers might be too weak to capture all the nuances of gender differences.
To see how these methods do on a localized attribute, we also applied each
of them to the “smiling” attribute. Here, while once again our method has the
lowest error rate, we see that the Adaboost methods perform significantly better
than the prior SVM method. This result highlights the power of Adaboost to
correctly find the important features from a large set of possibilities, as well as
the degradation in accuracy of SVMs when given too much irrelevant data.
6 The FaceTracer Engine

We have trained attribute-tuned global SVM classifiers for each attribute listed
in Table 4. In an offline process, all images in our database are sent through the
classifiers for each attribute, and the resulting attribute labels are stored for fast
online searches using the FaceTracer engine.
(a) (b)
Fig. 6. Results of queries (a) “older men with mustaches” and (b) “dark-haired people
with sunglasses” on our face search engine. The results are shown with aligned face
images on the left, and a preview of the original image for the currently selected face
on the right. Notice the high quality of results in both cases.
For a search engine, the design of the user interface is important for enabling
users to easily find what they are looking for. We use simple text-based queries,
since these are both familiar and accessible to most internet users. Search queries
are mapped onto attribute labels using a dictionary of terms. Users can see the
current list of attributes supported by the system on the search page, allowing
them to construct their searches without having to guess what kinds of queries
are allowed. This approach is simple, flexible, and yields excellent results in prac-
tice. Furthermore, it is easy to add new phrases and attributes to the dictionary,
or maintain separate dictionaries for different languages.
Results are ranked in order of decreasing confidence, so that the most relevant
images are shown first. (Our classifier gives us confidence values for each labeled
attribute.) For searches with multiple query terms, we combine the confidences
of different labels such that the final ranking shows images in decreasing order of
relevance to all search terms. To prevent high confidences for one attribute from
dominating the search results, we convert the confidences into probabilities, and
then use the product of the probabilities as the sort criteria. This ensures that
the images with high confidences for all attributes are shown first.
Example queries on our search engine are shown in Figs. 1b, 2, and 6. The
returned results are all highly relevant, and the user can view the results in a
variety of ways, as shown in the different examples. Figure 2b shows that we can
learn useful things about an image using just the appearance of the faces within
it – in this case determining whether the image was taken indoors or outdoors.
Our search engine can be used in many other applications, replacing or aug-
menting existing tools. In law enforcement, eyewitnesses to crimes could use our
system to quickly narrow a list of possible suspects and then identify the actual
criminal from this reduced list, saving time and increasing the chances of finding
the right person. On the internet, our face search engine is a perfect match for
social networking websites such as Facebook and Myspace, which contain large
numbers of images with people. Additionally, the community aspect of these
websites would allow for collaborative creation of new attributes. Finally, users
can utilize our system to more easily organize and manage their own personal
photo collections. For example, searches for blurry or other poor-quality images
can be used to find and remove all such images from the collection.
7 Discussion
In this work, we have described a new approach to searching for images in large
databases and have constructed the first face search engine using this approach.
By limiting our focus to images with faces, we are able to align the images to a
common coordinate system. This allows us to exploit the commonality of facial
structures across people to train accurate classifiers for real-world face images.
Our approach shows the power of combining the strengths of different algorithms
to create a flexible architecture without sacrificing classification accuracy.
As we continue to grow and improve our system, we would also like to ad-
dress some of our current limitations. For example, to handle more than just
frontal faces would require that we define the face regions for each pose bin.
Rather than specifying the regions manually, however, we can define them once
on a 3D model, and then project the regions to 2D for each pose bin. The other
manual portion of our architecture is the labeling of example images for train-
ing classifiers. Here, we can take advantage of communities on the internet by
offering a simple interface for both defining new attributes and labeling example
images. Finally, while our dictionary-based search interface is adequate for most
simple queries, taking advantage of methods in statistical natural language pro-
cessing (NLP) could allow our system to map more complex queries to the list
of attributes.
Acknowledgements. We are grateful to Omron Technologies for providing us

the OKAO face detection system. This work was supported by NSF grants IIS-
03-08185 and ITR-03-25867.
References
1. Omron: OKAO vision (2008), http://www.omron.com/rd/vision/01.html
2. Cortes, C., Vapnik, V.: Support-vector networks. Machine Learning 20(3) (1995)
3. Golomb, B.A., Lawrence, D.T., Sejnowski, T.J.: Sexnet: A neural network identifies
sex from human faces. NIPS, 572–577 (1990)
4. Belhumeur, P.N., Hespanha, J., Kriegman, D.J.: Eigenfaces vs. fisherfaces: Recog-
nition using class specific linear projection. In: Buxton, B.F., Cipolla, R. (eds.)
ECCV 1996. LNCS, vol. 1065, pp. 45–58. Springer, Heidelberg (1996)
5. Moghaddam, B., Yang, M.-H.: Learning gender with support faces. TPAMI 24(5),
707–711 (2002)
6. Phillips, P., Moon, H., Rizvi, S., Rauss, P.: The FERET evaluation methodology
for face-recognition algorithms. TPAMI 22(10), 1090–1104 (2000)
7. Shakhnarovich, G., Viola, P.A., Moghaddam, B.: A unified learning framework for
real time face detection and classification. ICAFGR, 14–21 (2002)
8. Baluja, S., Rowley, H.: Boosting sex identification performance. IJCV (2007)
9. Freund, Y., Shapire, R.E.: Experiments with a new boosting algorithm. In: ICML
(1996)
10. Viola, P., Jones, M.: Rapid object detection using a boosted cascade of simple
features. In: CVPR (2001)
11. Bartlett, M.S., Littlewort, G., Fasel, I., Movellan, J.R.: Real time face detection and
facial expression recognition: Development and applications to human computer
interaction. CVPRW 05 (2003)
12. Wang, Y., Ai, H., Wu, B., Huang, C.: Real time facial expression recognition with
adaboost. In: ICPR, pp. 926–929 (2004)
13. Datta, R., Li, J., Wang, J.Z.: Content-based image retrieval: Approaches and trends
of the new age. Multimedia Information Retrieval, 253–262 (2005)
14. Pentland, A., Picard, R., Sclaroff, S.: Photobook: Content-based manipulation of
image databases. IJCV, 233–254 (1996)
15. Huang, G.B., Ramesh, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild:
A database for studying face recognition in unconstrained environments. Technical
Report 07-49 (2007)
16. Georghiades, A.S., Belhumeur, P.N., Kriegman, D.J.: From few to many: Il-
lumination cone models for face recognition under variable lighting and pose.
TPAMI 23(6), 643–660 (2001)
17. Sim, T., Baker, S., Bsat, M.: The CMU pose, illumination, and expression (PIE)
database. In: ICAFGR, pp. 46–51 (2002)
18. Phillips, P.J., Flynn, P.J., Scruggs, T., Bowyer, K.W., Chang, J., Hoffman, K.,
Marques, J., Min, J., Worek, W.: Overview of the face recognition grand challenge.
CVPR, 947–954 (2005)
19. Pentland, A., Moghaddam, B., Starner, T.: View-based and modular eigenspaces
for face recognition. CVPR, 84–91 (1994)
20. Huang, J., Shao, X., Wechsler, H.: Face pose discrimination using support vector
machines (SVM). In: ICPR, pp. 154–156 (1998)
21. Osuna, E., Freund, R., Girosi, F.: Training support vector machines: An application
to face detection. CVPR (1997)
22. Schapire, R., Freund, Y., Bartlett, P., Lee, W.S.: Boosting the margin: A new
explanation for the effectiveness of voting methods. The Annals of Statistics 26(5),
1651–1686 (1998)
23. Drucker, H., Cortes, C.: Boosting decision trees. NIPS, 479–485 (1995)
24. Chang, C.C., Lin, C.J.: LIBSVM: a library for support vector machines (2001),
http://www.csie.ntu.edu.tw/cjlin/libsvm/
What Does the Sky Tell Us about the Camera?
Jean-François Lalonde, Srinivasa G. Narasimhan, and Alexei A. Efros
School of Computer Science, Carnegie Mellon University

http://graphics.cs.cmu.edu/projects/sky
Abstract. As the main observed illuminant outdoors, the sky is a rich

source of information about the scene. However, it is yet to be fully
explored in computer vision because its appearance depends on the sun
position, weather conditions, photometric and geometric parameters of
the camera, and the location of capture. In this paper, we propose the
use of a physically-based sky model to analyze the information available
within the visible portion of the sky, observed over time. By fitting this
model to an image sequence, we show how to extract camera parameters
such as the focal length, and the zenith and azimuth angles. In short, the
sky serves as a geometric calibration target. Once the camera parameters
are recovered, we show how to use the same model in two applications:
1) segmentation of the sky and cloud layers, and 2) data-driven sky
matching across different image sequences based on a novel similarity
measure defined on sky parameters. This measure, combined with a rich
appearance database, allows us to model a wide range of sky conditions.
1 Introduction
When presented with an outdoor photograph (such as images on Fig. 1), an

average person is able to infer a good deal of information just by looking at the
sky. Is it morning or afternoon? Do I need to wear a sunhat? Is it likely to rain?
A professional, such as a sailor or a pilot, might be able to tell even more: time
of day, temperature, wind conditions, likelihood of a storm developing, etc. As
the main observed illuminant in an outdoor image, the sky is a rich source of
information about the scene. However it is yet to be fully explored in computer
vision. The main obstacle is that the problem is woefully under-constrained. The
appearance of the sky depends on a host of factors such as the position of the
sun, weather conditions, photometric and geometric parameters of the camera,
and location and direction of observation. Unfortunately, most of these factors
remain unobserved in a single photograph; the sun is rarely visible in the picture,
the camera parameters and location are usually unknown, and worse yet, only a
small fraction of the full hemisphere of sky is actually seen.
However, if we were to observe the same small portion of the sky over time, we
would see the changes in sky appearance due to the sun and weather that are not
present within a single image. In short, this is exactly the type of problem that
might benefit from observing a time-lapse image sequence. Such a sequence is
typically acquired by a static camera looking at the same scene over a period of

What Does the Sky Tell Us about the Camera? 355
Fig. 1. The sky appearance is a rich source of information about the scene illumination
time. When the scene is mostly static, the resulting sequence of images contains a
wealth of information that has been exploited in several different ways, the most
commonly known being background subtraction, but also shadow detection and
removal [1], video factorization and compression [2], radiometric calibration [3],
camera geo-location [4], temporal variation analysis [5] and color constancy [6].
The main contribution of this paper is to show what information about the cam-
era is available in the visible portion of the sky in a time-lapse image sequence,
and how to extract this information to calibrate the camera.
The sky appearance has long been studied by physicists. One of the most pop-
ular physically-based sky model was introduced by Perez et al [7]. This model has
been used in graphics for relighting [8] and rendering [9]. Surprisingly however,
very little work has been done on extracting information from the visible sky.
One notable exception is the work of Jacobs et al [10] where they use the sky to
infer the camera azimuth by using a correlation-based approach. In our work, we
address a broader question: what does the sky tell us about the camera? We show
how we can recover the viewing geometry using an optimization-based approach.
Specifically, we estimate the camera focal length, its zenith angle (with respect
to vertical), and its azimuth angle (with respect to North). We will assume that
a static camera is observing the same scene over time, with no roll angle (i.e.
the horizon line is parallel to the image horizontal axis). Its location (GPS co-
ordinates) and the times of image acquisition are also known. We also assume
that the sky region has been segmented, either manually or automatically [5].
Once the camera parameters are recovered, we then show how we can use our
sky model in two applications. First, we present a novel sky-cloud segmentation
algorithm that identifies cloud regions within an image. Second, we show how
we can use the resulting sky-cloud segmentation in order to find matching skies
across different cameras. To do so, we introduce a novel bi-layered sky model
which captures both the physically-based sky parameters and cloud appearance,
and determine a similarity measure between two images. This distance can then
be used for finding images with similar skies, even if they are captured by differ-
ent cameras at different locations. We show qualitative cloud segmentation and
sky matching results that demonstrate the usefulness of our approach.
In order to thoroughly test our algorithms, we require a set of time-lapse
image sequences which exhibit a wide range of skies and cameras. For this, we
use the AMOS (Archive of Many Outdoor Scenes) database [5], which contains
image sequences taken by static webcams over more than a year.
356 J.-F. Lalonde, S.G. Narasimhan, and A.A. Efros
Fig. 2. Geometry of the problem, when a camera is viewing a sky element (blue patch
in the upper-right). The sky element is imaged at pixel (up , vp ) in the image, and the
camera is rotated by angles (θc , φc ). The camera focal length fc , not shown here, is
the distance between the origin (center of projection), and the image center. The sun
direction is given by (θs , φs ), and the angle between the sun and the sky element is γp .
Here (up , vp ) are known because the sky is segmented.
2 Physically-Based Model of the Sky

First, we introduce the physically-based model of the sky that lies at the foun-
dation of our approach. We will first present the model in its general form, then
in a useful simplified form, and finally demonstrate how it can be written as a
function of camera parameters. We will consider clear skies only, and address
the more complicated case of clouds at a later point in the paper.
2.1 All-Weather Perez Sky Model

The Perez sky model [7] describes the luminance of any arbitrary sky element
as a function of its elevation, and its relative orientation with respect to the
sun. It is a generalization of the CIE standard clear sky formula [11], and it has
been found to be more accurate for a wider range of atmospheric conditions [12].
Consider the illustration in Fig. 2. The relative luminance lp of a sky element is
a function of its zenith angle θp and the angle γp with the sun:
# $
lp = f (θp , γp ) = [1 + a exp(b/ cos θp )] × 1 + c exp(dγp ) + e cos2 γp , (1)
where the 5 constants (a, b, c, d, e) specify the current atmospheric conditions. As

suggested in [9], those constants can also be expressed as a linear function of a
single parameter, the turbidity t. Intuitively, the turbidity encodes the amount
of scattering in the atmosphere, so the lower t, the clearer the sky. For clear
skies, the constants take on the following values: a = −1, b = −0.32, c = 10,
d = −3, e = 0.45, which corresponds approximately to t = 2.17.
The model expresses the absolute luminance Lp of a sky element as a function
of another arbitrary reference sky element. For instance, if the zenith luminance
Lz is known, then
f (θp , γp )
Lp = Lz , (2)
f (0, θs )
where θs is the zenith angle of the sun.
2.2 Clear-Weather Azimuth-Independent Sky Model

By running synthetic experiments, we were able to determine that the influence
of the second factor in (1) becomes negligible when the sun is more than 100◦
away from a particular sky element. In this case, the sky appearance can be
modeled by using only the first term from (1):
lp = f (θp ) = 1 + a exp(b/ cos θp ) . (3)
This equation effectively models the sky gradient, which varies from light to dark
from horizon to zenith on a clear day. Lp is obtained in a similar fashion as in (2):
f (θp )
Lp = Lz . (4)
f (0)
2.3 Expressing the Sky Model as a Function of Camera Parameters

Now suppose a camera is looking at the sky, as in Fig. 2. We can express the gen-
eral (1) and azimuth-independent (3) models as functions of camera parameters.
Let us start with the simpler azimuth-independent model.
If we assume that the camera
zenith
angle θc is independent of its azimuth
v
angle φc , then θp ≈ θc − arctan fp . This can be substituted into (3):

b
lp = g (vp , θc , fc ) = 1 + a exp , (5)
cos(θc − arctan(vp /fc )
where, vp is the v-coordinate of the sky element in the image, and fc is the
camera focal length.
In the general sky model case, deriving the equation involves expressing γp as
a function of camera parameters:
γp = arccos (cos θs cos θp + sin θs sin θp cos Δφp ) , (6)

u
where Δφp ≈ φc −φs −arctan fp , and up is the sky element u-coordinate in the
image. We substitute (6) into (1) to obtain the final equation. For succinctness,
we omit writing it in its entirety, but do present its general form:
lp = g(up , vp , θc , φc , fc , θs , φs ) , (7)
where θc , φc (θs , φs ) are the camera (sun) zenith and azimuth angles.
Luminance as a function of pixel height and field of view Luminance as a function of pixel height and camera azimuth
1 1
fov = 40° θc = 70°
° °
fov = 60 θc = 80
0.9 ° 0.9
fov = 80 θc = 90°
°
fov = 100
° θ = 100°
fov = 120 c
0.8 0.8 θ = 110°
c
Scaled luminance
Scaled luminance
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Pixel height of sky element in the image (vp) Pixel height of sky element in the image (vp)
(a) (b)
Fig. 3. Luminance profiles predicted by the azimuth-independent model (5). For clear
skies, intensity diminishes as pixel height (x-axis) increases. (a) The camera zenith
angle is kept constant at θc = 90◦ , while the field of view is varied. (b) The field of
view is kept constant at 80◦ , while the camera zenith angle is varied. Both parameters
have a strong influence on the shape and offset of the predicted sky gradient.
Before we present how we use the models presented above, recall that we are
dealing with ratios of sky luminances, and that a reference element is needed.
Earlier, we used the zenith luminance Lz as a reference in (2) and (4), which
unfortunately is not always visible in images. Instead, we can treat this as an
additional unknown in the equations. Since the denominators in (2) and (4) do
not depend on camera parameters, we can combine them with Lz into a single
unknown scale factor k.
3 Using the Clear Sky as a Calibration Target

In the previous section, we presented a physically-based model of the clear sky
that can be expressed as a function of camera parameters. Now if we are given a
set of images taken from a static camera, can we use the clear sky as a calibration
target and recover the camera parameters, from the sky appearance only?
3.1 Recovering Focal Length and Zenith Angle

Let us first consider the simple azimuth-independent model (5). If we plot the
predicted luminance profile for different focal lengths as in Fig. 3-(a) (or, equiv-
alently, for different fields of view), we can see that there is a strong dependence
between the focal length fc and the shape of the luminance gradient. Similarly,
the camera azimuth θc dictates the vertical offset, as in Fig. 3-(b). From this
intuition, we devise a method of recovering the focal length and zenith angle
of a camera from a set of images where the sun is far away from its field of
view (i.e. at least 100◦ away). Suppose we are given a set I of such images, in
which the sky is visible at pixels in set P, also given. We seek to find the camera
parameters (θc , fc ) that minimize

2
min yp(i) − k (i) g (vp , θc , fc ) , (8)
θc ,fc ,k(i)
i∈I p∈P
(i)
where yp is the observed intensity of pixel p in image i, and k (i) are unknown
scale factors (Sect. 2.3), one per image. This non-linear least-squares minimiza-
tion can be solved iteratively using standard optimization techniques such as
Levenberg-Marquadt, or fminsearch in Matlab. fc is initialized to a value
corresponding to a 35◦ field of view, and θc is set such that the horizon line is
aligned with the lowest visible sky pixel. All k (i) ’s are initialized to 1.
3.2 Recovering Azimuth Angle
From the azimuth-independent model (5) and images where the sun is far from
the camera field of view, we were able to estimate the camera focal length fc
and its zenith angle θc . Now if we consider the general model (7) that depends
on the sun position, we can also estimate the camera azimuth angle using the
same framework as before.
Suppose we are given a set of images J where the sky is clear, but where the
sun is now closer to the camera field of view. Similarly to (8), we seek to find
the camera azimuth angle which minimizes

2
min yp(j) − k (j) g(up , vp , θc , φc , fc , θs , φs ) . (9)
φc ,k(j)
j∈J p∈P
We already know the values of fc and θc , so we do not need to optimize over

them. Additionally, if the GPS coordinates of the camera and the time of capture
of each image are known, the sun zenith and azimuth (θs , φs ) can be computed
using [13]. Therefore, the only unknowns are k (j) (one per image), and φc . Since
this equation is highly non-linear, we have found that initializing φc to several
values over the [−π, π] interval and keeping the result that minimizes (9) works
the best.
4 Evaluation of Camera Parameters Estimation
In order to thoroughly evaluate our model, we have performed extensive tests on

synthetic data generated under a very wide range of operating conditions. We
also evaluated our model on real image sequences to demonstrate its usefulness
in practice.
4.1 Synthetic Data
We tested our model and fitting technique on a very diverse set of scenarios
using data synthetically generated by using the original Perez sky model in (1).
During these experiments, the following parameters were varied: the camera
focal length fc , the camera zenith and azimuth angles (θc , φc ), the number of
Table 1. Camera calibration from the sky on 3 real image sequences taken from the
AMOS database [5]. Error in focal length, zenith and azimuth angle estimation is shown
for each sequence. The error is computed with respect to values obtained by using the
sun position to estimate the same parameters [14].
Focal length Zenith angle Azimuth angle

Sequence name
error (%) error (◦ ) error (◦ )
257 1.1 < 0.1 2.6
414 3.1 < 0.1 2
466 2.5 < 0.1 4.5
input images used in the optimization, the number of visible sky pixels, and the
camera latitude (which effects the maximum sun height). In all our experiments,
1000 pixels are randomly selected from each input image, and each experiment
is repeated for 15 random selections.
The focal length can be recovered with at most 4% error even in challenging
conditions: 30% visibility, over a wide range of field of view ([13◦ , 93◦ ] interval),
zenith angles ([45◦ , 135◦]), azimuth angles ([−180◦, 180◦ ]), and sun positions
(entire hemisphere). We note a degradation in performance at wider fields of
view (> 100◦ ), because the assumption of independent zenith and azimuth angles
starts to break down (Sect. 2.3). Less than 0.1◦ error for both zenith and azimuth
angles is obtained in similar operating conditions.
4.2 Real Data

Although experiments on synthetic data are important, real image sequences
present additional challenges, such as non-linear camera response functions, non-
gaussian noise, slight variations in atmospheric conditions, etc. We now evaluate
our method on real image sequences and show that our approach is robust to
these noise sources and can be used in practice.
First, the camera response function may be non-linear, so we need to radio-
metrically calibrate the camera. Although newer techniques [3] might be more
suitable for image sequences, we rely on [15] which estimates the inverse response
function by using color edges gathered from a single image. For additional ro-
bustness, we detect edges across several frames. Recall that the optimization
procedures in (8) and (9) requires clear sky image sets I and J , where the
sun is far or close to the camera respectively. We approximate (5) by a vertical
quadratic in image space, and automatically build set I by keeping images with
low residual fitting error. Similarly, set J is populated by finding images with
a good fit to horizontal quadratic. It is important that the effect of the moving
sun be visible in the selected images J .
We present results from applying our algorithm on three image sequences
taken from the AMOS database [5]. Since ground truth is not available on those
sequences, we compare our results with those obtained with the method de-
scribed in [14], which uses hand-labelled sun positions to obtain high-accuracy es-
timates. Numerical results are presented in Table 1, and Fig. 4 shows a
Seq. 257 Seq. 414 Seq. 466
Fig. 4. Illustration of estimated camera parameters. First row: Example image for the
three sequences in Table 1. The horizon line is drawn in red. Note that the horizon line
in sequence 414 is found to be just below the image. Second row: Graphical illustration
of all three estimated parameters: focal length, zenith and azimuth angles. The sun is
drawn at the position corresponding to the image in the first row.
visualization of the recovered camera parameters. The results are consistent

with image data: for instance, sun flares are visible in the first image (Seq. 257),
which indicate that the sun must be above the camera, slightly to its left. This
matches the visualization below the image.
5 Application: Separation of Sky and Cloud Layers

Now that we have recovered camera parameters, we demonstrate how to use the
same physically-based model for two applications. Until now, we have only dealt
with clear skies, but alas, this is not always true! In this section, we present a
novel cloud segmentation algorithm, which will later be used for sky matching.
Clouds exhibit a wide range of textures, colors, shapes, and even trans-
parencies. Segmenting the clouds from the sky cannot be achieved with simple
heuristics such as color-based thresholding as they are easily confounded by the
variation in their appearances. On the other hand, our physically-based model
predicts the sky appearance, so any pixel that differs from it is an outlier and is
likely to correspond to a cloud. Using this intuition, we now consider two ways
of fitting our model to skies that may contain clouds. Note that we perform all
processing in the xyY color space as recommended in [9].
5.1 Least-Squares Fitting

The first idea is to follow a similar approach as we did previously and fit the
model (5) in a non-linear least-squares fashion, by adjusting the coefficients
(a, b, c, d, e) and the unknown scale factor k independently in each color chan-
nel, and treating the outliers as clouds. To reduce the number of variables,
we follow [9] and express the five weather coefficients as a linear function of
a single
# value, the turbidity
$ t. Strictly speaking, this means minimizing over
x = t k (1) k (2) k (3) :
3
2
min yp(i) − k (i) g(up , vp , θs , φs , τ (i) (t)) , (10)
x
i=1 p∈P
where i indexes the color channel. Here the camera parameters are fixed, so
we omit them for clarity. The vector τ (i) (t) represents the coefficients (a, . . . , e)
obtained by multiplying the turbidity t with the linear transformation M (i) :
# $T
τ (i) (t) = M (i) t 1 . The entries of M (i) for the xyY space are given in the
appendix in [9]. The k (i) are initialized to 1, and t to 2 (low turbidity).
Unfortunately, solving this simplified minimization problem does not yield
satisfying results because the L2-norm is not robust to outliers, so even a small
amount of clouds will bias the results.
5.2 Regularized Fitting
In order to increase robustness to outliers, we compute a data-driven prior model

of clear skies xc , which we use to add 2 terms to (10): 1) we assign more weight
to pixels we believe are part of the sky; and 2) we penalize parameters that differ
from the prior in an L2 sense. Equation (10) becomes

3
2
min wp yp(i) − k (i) g(up , vp , θs , φs , τ (i) (t)) + βx − xc 2 , (11)
x
i=1 p∈P
where, wp ∈ [0, 1] is a weight given to each pixel, and β = 0.05 controls the
importance of the prior term in the optimization. We initialize x to the prior xc .
Let us now look at how xc is obtained. We make the following observation:
clear skies should have low turbidities, and they should be smooth (i.e. no patchy
clouds). Using this insight, if minimizing (10) on a given image yields low residual
error and turbidity, then the sky must be clear. We compute a database of clear
skies by keeping all images with turbidity less than a threshold (we use 2.5), and
keep the best 200 images, sorted by residual error. Given an image, we compute
xc by taking the mean over the K nearest neighbors in the clear sky database,
using the angular deviation between sun positions as distance measure (we use
K = 2). This allows us to obtain a prior model of what the clear sky should look
like at the current sun position. Note that we simply could have used the values
for (a, . . . , e) from Sect. 2 and fit only the scale factors k (i) , but this tends to
over-constrain, so we fit t as well to remain as faithful to the data as possible.
To obtain the weights wp in (11), the color distance λ between each pixel and
the prior model is computed and mapped to the [0, 1] interval with an inverse
exponential: wp = exp{−λ2 /σ 2 } (we use σ 2 = 0.01 throughout this paper). After
the optimization is over, we re-estimate wp based on the new parameters x, and
repeat the process until convergence, or until a maximum number of iterations
Fig. 5. Sky-cloud separation example results. First row: input images (radiometrically
corrected). Second row: sky layer. Third row: cloud segmentation. The clouds are color-
coded by weight: 0 (blue) to 1 (red). Our fitting algorithm is able to faithfully extract
the two layers in all these cases.
is reached. The process typically converges in 3 iterations, and the final value
wp is used as the cloud segmentation. Cloud coverage is then computed as
for
p∈P wp .
1
|P|
5.3 Segmentation Results

Figure 5 shows typical results of cloud layers extracted using our approach. Note
that unweighted least-squares (10) fails on all these examples because the clouds
occupy a large portion of the sky, and the optimization tries to fit them as much as
possible, since the quadratic loss function is not robust to outliers. A robust loss
function behaves poorly because it treats the sky pixels as outliers in the case of
highly-covered skies, such as the examples shown in the first two columns of Fig. 6.
Our approach injects domain knowledge into the optimization by using a data-
driven sky prior, forcing it to fit the visible sky. Unfortunately, since we do not model
sunlight, the estimation does not converge to a correct segmentation when the sun
is very close to the camera, as illustrated in the last two columns of Fig. 6.
6 Application: Matching Skies across Image Sequences

After obtaining a sky-cloud segmentation, we consider the problem of finding
matching skies between images taken by different cameras. Clearly, appearance-
based matching algorithms such as cross-correlation would not work if the cam-
eras have different parameters. Instead, we use our sky model along with cloud
Fig. 6. More challenging cases for the sky-cloud separation, and failure cases. First
row: input images (radiometrically corrected). Second row: sky layer. Third row: cloud
layer. The clouds are color-coded by weight: 0 (blue) to 1 (red). Even though the sky
is more than 50% occluded in the input images, our algorithm is able to recover a good
estimate of both layers. The last two columns illustrate a failure case: the sun (either
when very close or in the camera field of view) significantly alters the appearance of
the pixels such that they are labeled as clouds.
statistics in order to find skies that have similar properties. We first present our
novel bi-layered representation for sky and clouds, which we then use to define
a similarity measure between two images. We then present qualitative matching
results on real image sequences.
6.1 Bi-layered Representation for Sky and Clouds
Because clouds can appear so differently due to weather conditions, a generative

model such as the one we are using for the sky is likely to have a large number
of parameters, and thus be difficult to fit to image data. Instead, we propose a
hybrid model: our physically-based sky model parameterized by the turbidity t
for the sky appearance, and a non-parametric representation for the clouds.
Taking inspiration from Lalonde et al [16], we represent the cloud layer by a
joint color histogram in the xyY space over all pixels which belong to the cloud
regions. While they have had success with color histograms only, we have found
this to be insufficient on our richer dataset, so we also augment the representation
with a texton histogram computed over the same regions. A 1000-word texton
dictionary is built from a set of skies taken from training images different than
the ones used for testing. In our implementation, we choose 213 bins for the color
histograms.
Input Nearest-neighbors
Sequence 466 Sequence 257 Sequence 407 Sequence 414
Fig. 7. Sky matching results across different cameras. The left-most column shows
several images taken from different days of sequence 466 in the AMOS database. The
three other columns are the nearest-neighbor matches in sequences 257, 407 and 414
respectively, obtained using our distance measure. Sky conditions are well-matched,
even though cameras have different parameters.
Once this layered sky representation is computed, similar images can be re-
trieved by comparing their turbidities and cloud statistics (we use χ2 distance
for histogram comparison). A combined distance is obtained by taking the sum
of cloud and turbidity distance, with the relative importance between the two
determined by the cloud coverage.
6.2 Qualitative Evaluation

The above algorithm was tested on four sequences from the AMOS database.
Since we do not have ground truth to evaluate sky matching performance, we
provide qualitative results in Fig. 7. Observe that sky conditions are matched
correctly, even though cameras have different horizons, focal lengths, and camera
response functions. A wide range of sky conditions can be matched successfully,
including clear, various amounts of clouds, and overcast conditions. We provide
additional segmentation and matching results on our project website.
7 Summary
In this paper, we explore the following question: what information about the
camera is available in the visible sky? We show that, even if a very small portion
of the hemisphere is visible, we can reliably estimate three important camera
parameters by observing the sky over time. We do so by expressing a well-
known physically-based sky model in terms of the camera parameters, and by
fitting it to clear sky images using standard minimization techniques. We then
demonstrate the accuracy of our approach on synthetic and real data. Once the
camera parameters are estimated, we show how we can use the same model to
segment out clouds from sky and build a novel bi-layered representation, which
can then be used to find similar skies across different cameras.
We plan to use the proposed sky illumination model to see how it can help
us predict the illumination of the scene. We expect that no parametric model
will be able to capture this information well enough, so data-driven methods will
become even more important.
Acknowledgements
This research is supported in parts by an ONR grant N00014-08-1-0330 and NSF
grants IIS-0643628, CCF-0541307 and CCF-0541230. A. Efros is grateful to the
WILLOW team at ENS Paris for their hospitality.
References
1. Weiss, Y.: Deriving intrinsic images from image sequences. In: IEEE International
Conference on Computer Vision (2001)
2. Sunkavalli, K., Matusik, W., Pfister, H., Rusinkiewicz, S.: Factored time-lapse
video. ACM Transactions on Graphics (SIGGRAPH 2007) 26(3) (August 2007)
3. Kim, S.J., Frahm, J.M., Polleyfeys, M.: Radiometric calibration with illumination
change for outdoor scene analysis. In: IEEE Conference on Computer Vision and
Pattern Recognition (2008)
4. Jacobs, N., Satkin, S., Roman, N., Speyer, R., Pless, R.: Geolocating static cameras.
In: IEEE International Conference on Computer Vision (2007)
5. Jacobs, N., Roman, N., Pless, R.: Consistent temporal variations in many outdoor
scenes. In: IEEE Conference on Computer Vision and Pattern Recognition (2007)
6. Sunkavalli, K., Romeiro, F., Matusik, W., Zickler, T., Pfister, H.: What do color
changes reveal about an outdoor scene? In: IEEE Conference on Computer Vision
and Pattern Recognition (2008)
7. Perez, R., Seals, R., Michalsky, J.: All-weather model for sky luminance distribution
– preliminary configuration and validation. Solar Energy 50(3), 235–245 (1993)
8. Yu, Y., Malik, J.: Recovering photometric properties of architectural scenes from
photographs. Proceedings of ACM SIGGRAPH 1998 (July 1998)
9. Preetham, A.J., Shirley, P., Smits, B.: A practical analytic model for daylight.
Proceedings of ACM SIGGRAPH 1999 (August 1999)
10. Jacobs, N., Roman, N., Pless, R.: Toward fully automatic geo-location and geo-
orientation of static outdoor cameras. In: Workshop on applications of computer
vision (2008)
11. Committee, C.T.: Spatial distribution of daylight – luminance distributions of var-
ious reference skies. Technical Report CIE-110-1994, International Commission on
Illumination (1994)
12. Ineichen, P., Molineaux, B., Perez, R.: Sky luminance data validation: comparison
of seven models with four data banks. Solar Energy 52(4), 337–346 (1994)
13. Reda, I., Andreas, A.: Solar position algorithm for solar radiation applications.
Technical Report NREL/TP-560-34302, National Renewable Energy Laboratory
(November 2005)
14. Lalonde, J.F., Narasimhan, S.G., Efros, A.A.: Camera parameters estimation from
hand-labelled sun positions in image sequences. Technical Report CMU-RI-TR-08-
32, Robotics Institute. Carnegie Mellon University (July 2008)
15. Lin, S., Gu, J., Yamazaki, S., Shum, H.Y.: Radiometric calibration from a single
image. In: IEEE Conference on Computer Vision and Pattern Recognition (2004)
16. Lalonde, J.F., Hoiem, D., Efros, A.A., Rother, C., Winn, J., Criminisi, A.: Photo
clip art. ACM Transactions on Graphics (SIGGRAPH 2007) 26(3) (August 2007)
Three Dimensional Curvilinear Structure Detection
Using Optimally Oriented Flux
Max W.K. Law and Albert C.S. Chung
Lo Kwee-Seong Medical Image Analysis Laboratory,

Department of Computer Science and Engineering,
The Hong Kong University of Science and Technology, Hong Kong
{maxlawwk,achung}@cse.ust.hk
Abstract. This paper proposes a novel curvilinear structure detector, called Op-
timally Oriented Flux (OOF). OOF finds an optimal axis on which image gradi-
ents are projected in order to compute the image gradient flux. The computation
of OOF is localized at the boundaries of local spherical regions. It avoids con-
sidering closely located adjacent structures. The main advantage of OOF is its
robustness against the disturbance induced by closely located adjacent objects.
Moreover, the analytical formulation of OOF introduces no additional computa-
tion load as compared to the calculation of the Hessian matrix which is widely
used for curvilinear structure detection. It is experimentally demonstrated that
OOF delivers accurate and stable curvilinear structure detection responses under
the interference of closely located adjacent structures as well as image noise.
1 Introduction
Analysis of curvilinear structures in volumetric images has a wide range of appli-

cations, for instance centerline extraction [1,3], detection and segmentation [7,15,9],
vascular image enhancement [12,8,11] or visualization [2]. In particular, low-level
detectors which are sensitive to curvilinear structures are the foundations of the afore-
mentioned applications. One classic low-level detector is the multiscale based image
intensity second-order statistics. Lindeberg [10] conducted in depth research regard-
ing the use of the Gaussian smoothing function with various scale factors for extracting
multiscale second-order statistics. Koller et al. [7] exploited the image intensity second-
order statistics to form Hessian matrices for the analysis of curvilinear structures in
three dimensional image volumes. Frangi et al. [6] introduced the vesselness measure
based on eigenvalues extracted from the Hessian matrix in a multiscale fashion. Kris-
sian et al. [9] studied the relation between the Hessian matrix and the image gradient
computed in multiple scales for the detection of tubular structures. Manniesing et al.
[11] made use of the multiscale Hessian matrix based features to devise a nonlinear
scale space representation of curvilinear structures for vessel image enhancement.
Another recently developed low-level detector for the curvilinear structure analysis
is the image gradient flux. It is a scalar measure which quantifies the amount of image
gradient flowing in or out of a local spherical region. A large magnitude of the image
gradient flux is an indication of the presence of a curvilinear structure disregarding the

Three Dimensional Curvilinear Structure Detection Using Optimally Oriented Flux 369
structure direction. Bouix et al. proposed to compute the image gradient flux for ex-
tracting centerlines of curvilinear structures [3]. Siddqi et al. [15] showed promising
vascular segmentation results by evolving an image gradient flux driven active surface
model. But the major disadvantage of the image gradient flux is its regardless of direc-
tional information.
Grounded on the multiscale based Hessian matrix, Sato et al. [12] presented a thor-
ough study on the properties of the eigenvalues extracted from the Hessian matrix in
different scales, and their performance in curvilinear structure segmentation and visu-
alization. The study showed that the eigenvalues extracted from the Hessian matrix can
be regarded as the results of convolving the image with the second derivative of a Gaus-
sian function. This function offers differential effects which compute the difference
between the intensity inside an object and in the vicinity of the object. However, if the
intensity around the objects is not homogeneous due to the presence of closely located
adjacent structures, the differential effect given by the second derivatives of Gaussian
is adversely affected.
In this paper, we propose a novel detector of curvilinear structures, called optimally
oriented flux (OOF). Specifically, the oriented flux encodes directional information by
projecting the image gradient along some axes, prior to measuring the amount of the
projected gradient that flows in or out of a local spherical region. Meanwhile, OOF dis-
covers the structure direction by finding an optimal projection axis which minimizes
the oriented flux. OOF is evaluated for each voxel in the entire image. The evaluation
of OOF is based on the projected image gradient at the boundary of a spherical region
centered at a local voxel. When the local spherical region boundary touches the object
boundary of a curvilinear structure, the image gradient at the curvilinear object bound-
ary produces an OOF detection response. Depending on whether the voxels inside the
local spherical region have stronger intensity, the sign of the OOF detection response
varies. It can be utilized to distinguish between regions inside and outside curvilinear
structures.
The major advantage of the proposed method is that the OOF based detection is lo-
calized at the boundary of the local spherical region. Distinct from the Hessian matrix,
OOF does not consider the region in the vicinity of the structure where a nearby ob-
ject is possibly present. As such, OOF detection result is robust against the disturbance
introduced by closely located objects. With this advantage, utilizing OOF for curvilin-
ear structure analysis is highly beneficial when closely located structures are present.
Moreover, the computation of OOF does not introduce additional computation load
compared to the Hessian matrix. Validated by a set of experiments, OOF is capable of
providing more accurate and stable detection responses than the Hessian matrix, with
the presence of closely located adjacent structures.
2 Methodology
2.1 Optimally Oriented Flux (OOF)
The notion of oriented flux along a particular direction refers to the amount of image
gradient projected along that direction at the surface of an enclosed local region. The
image gradient can flow either in or out of the enclosed local region. Without loss of
370 M.W.K. Law and A.C.S. Chung
generality, our elaboration focuses on the situation where the structures have stronger
intensity than background regions. As such, optimally oriented flux (OOF) aims at find-
ing an optimal projection direction that minimizes the inward oriented flux for the de-
tection of curvilinear structure.
The outward oriented flux along a direction ρ̂ is calculated by projecting the image
gradient v(·) along the direction of ρ̂ prior to the computation of flux in a local spherical
region Sr with radius r. Based on the definition of flux [13], the computation of the
outward oriented flux along the direction of ρ̂ is,

f (x; r, ρ̂) = (v(x + h) · ρ̂)ρ̂ · n̂dA, (1)
∂Sr
where dA is the infinitesimal area on ∂Sr , n̂ is the outward unit normal of ∂Sr at the
position ĥ. As ∂Sr is a sphere surface, h = rn̂, thus
3 3 1

f (x; r, ρ̂) = vk (x + rn̂)ρk ρl nl dA = ρ̂T Qr,x ρ̂, (2)
∂Sr k=1 l=1
where ρ̂ = (ρ1 , ρ2 , ρ3 ) , v(x) = (v1 (x), v2 (x), v3 (x))T , n̂ = (n1 , n2 , n3 )T , Qr,x is

T
a matrix that the entry at the ith row and jth column (i, j ∈ {1, 2, 3}) is,

qr,x
i,j
= vi (x + rn̂)nj dA. (3)
∂Sr
2.2 Analytical Computation of OOF

The idea of OOF is to identify the direction p̂ that the inward oriented flux attains the
minimum. It is not easy to discretize any one of the surface integrals of Equations 1
and 3 to estimate oriented flux and find the optimal axis which minimizes the inward
oriented flux. Nevertheless, computation of OOF can be achieved analytically by ac-
quiring the values of the entries of Qr,x , that only involves convolving an image with a
set of filters ψr,i,j ,
qr,x
i,j
= ψr,i,j (x) ∗ I(x). (4)
The above formulation avoids discretization and reduces computation complexity as
compared with the discretization of either Equation 1or Equation 3. By using fast Fourier
transform, the complexity of evaluating Equation 4 and thus Qr,x is O(N log N ), where
∀x ∈ Ω and Ω is the image domain having N voxels. The proposed method introduces
no additional computation load compared to some traditional approaches, such as Hes-
sian matrix based methods [12,9,6].
We begin the elaboration of the filters ψr,i,j (x) from Equation 3,

qr,x =
i,j
vi (x + rn̂)nj dA = [vi (x + rn̂)aˆj ] · n̂dA, (5)
∂Sr ∂Sr
where aˆ1 , aˆ2 and aˆ3 are the unit vectors along the x-, y- and z-directions respectively.
Assuming that v is continuous, by the divergence theorem,

∂
qr,x =
i,j
∇ · [vi (x + y)âj ]dV = vi (x + y)dV, (6)
Sr Sr aˆj
∂
where y is the position vector inside the sphere Sr and dV is the infinitesimal volume in
Sr . The continuous image gradient v(x) is acquired by convolving the discrete image
with the first derivatives of Gaussian with a small scale factor, i.e. vi (x) = (gaî ,σ ∗
I)(x), where ∗ is the convolution operator, gaî ,σ is the first derivative of Gaussian
along the direction of aî and σ = 1 in all our implementations. Furthermore, the volume
integral of Equation5 6 is extended to the entire image domain Ω by employing a step
1, ||x|| ≤ r,
function, br (x) = hence
0, otherwise ,

qr,x =
i,j
br (y)((gaî aˆj ,σ ∗ I)(x + y))dV = (br ∗ gaî aˆj ,σ )(x) ∗ I(x), (7)
Ω
where gaî aˆj ,σ is the second derivative of Gaussian along the axes aî and aˆj . Therefore,
the set of linear filters of Equation 4 is ψr,i,j (x) = (br ∗ gaî ,aˆj ,σ )(x). The next step is to
obtain the analytical Fourier expression of ψr,i,j (x) in order to compute the convolution
in Equation 4 by Fourier coefficient multiplication. Denote Ψr,i,j (u) be the Fourier
expression of ψr,i,j (x), where u = (u1 , u2 , u3 )T is the position vector in the frequency
domain. The values of u1 , u2 and u3 are in ”cycle per unit voxel” and in a range of
[−0.5, 0.5). By employing Fourier transforms on gaî aˆj ,σ and Hankel transforms [4] on
br (x),

−2(π||u||σ)2 1 sin(2πr||u||)
Ψr,i,j (u) = 4πrui uj e cos(2πr||u||) − . (8)
||u||2 2πr||u||
Based on the above formulation, the optimal projection axis which minimizes inward
oriented flux can be computed analytically. Denote the optimal direction is ωr,x , min-
imizing inward oriented flux is equivalent to maximizing fr (x; ωr,x ) subject to the
constraint ||ω r,x || = [ω r,x ]T ω r,x = 1. The solution is found by taking the first deriva-
tive on the Lagrange equation,
L(ω r,x ) = [ω r,x ]T Qr,x ω r,x + λr,x (1 − [ω r,x ]T ωr,x ), (9)
for ∇L(ω r,x ) = 0, and since qr,x
i,j
= qr,x
j,i
(see Equation 7), and thus Q = QT ,
Qr,x ω r,x = λr,x ω r,x . (10)
Equation 10 is in turn solved as a generalized eigenvalue problem. For volumetric im-
ages, there are at most three distinct pairs of λr,x and w r,x . The eigenvalues can be pos-
itive, zero or negative. These eigenvalues are denoted as λi (x; r), for λ1 (·) ≤ λ2 (·) ≤
λ3 (·), and the corresponding eigenvectors are ωi (x; r). Inside a curvilinear structure
having stronger intensity than the background, the first two eigenvalues would be much
smaller than the third one, λ1 (·) ≤ λ2 (·) << λ3 (·) and λ3 (·) ≈ 0. The first two eigen-
vectors span the normal plan of the structure and the third eigenvector is the structure
direction.
2.3 Eigenvalues and Eigenvectors

The major difference between the eigenvalues and eigenvectors extracted from OOF
and those from the Hessian matrix is that the computation of OOF and thus, its eigen-
values and eigenvectors are grounded on the analysis of image gradient on the local
sphere surface (∂Sr in Equation 3). In contrast, as pointed out by Sato et al. [12], the
computation of the Hessian matrix is closely related to the results of applying the sec-
ond derivative of Gaussian function on the image. This function computes the weighted
intensity average difference between the regions inside the structure and in the vicinity
of the structure. As such, the coverage of this function extends beyond the boundary
of target structures and possibly includes structures nearby. As a result, the weighted
intensity average difference computed by the second derivative of Gaussian function
can be affected by the adjacent objects. It can be harmful to the detection accuracy of
the Hessian matrix when closely located adjacent structures are present.
On the contrary, the evaluation of OOF is performed on the boundary of a local spher-
ical region ∂Sr . Detection response of OOF is induced from the intensity discontinuities
at the object boundary when the local sphere surface touches the object boundary of the
structure. The detection of OOF is localized at the boundary of the local spherical re-
gion. The localized detection avoids the inclusion of objects nearby. Therefore, the OOF
based detection is robust against the disturbance introduced by closely located adjacent
structures.
The eigenvalues extracted from OOF are the values of oriented flux along the corre-
sponding eigenvectors,
λi (x; r) = [ω i (x; r)]T Qr,x ω i (x; r) = f (x; r, ωi (x; r)). (11)
The image gradient at the object boundary of a strong intensity curvilinear structure
points to the centerline of the structure. Inside the structure, when the local spherical
region boundary ∂Sr (see Equation 1) touches the object boundary, at the contacting
position of these two boundaries, the image gradient v(·) is aligned in the opposite
direction of the outward normal n̂, hence λ1 (·) ≤ λ2 (·) << 0. On the other hand, the
image gradient is perpendicular to the structure direction, the projected image gradient
along ω 3 (·) has zero or very small magnitude, thus λ3 (·) ≈ 0. In contrast, if OOF
is computed for a voxel which is just outside the curvilinear structure, at the position
where ∂Sr touches the curvilinear structure boundary, the image gradient v(·) is in the
same direction as the outward normal n̂. It results in a large positive eigenvalue, that is
λ3 (·) >> 0.
Combining multiple eigenvalues to tailor a measure for identifying structures in a
specific shape is now possible. For instance Λ12 (x; r) = λ1 (x; r) + λ2 (x; r) can pro-
vide responses at the curvilinear object centerline with circular cross section. According
to Equations 1 and 11,
0.8
0.6
0.4
0.2
0
(a) (b) (c) (d)
Fig. 1. (a, b, c) The values of ||[W12 (·)]T n̂||. (d) The intensity scale of the images in (a-c).

Λ12 (x; r) = [W12 (x; r)]T v(x + h) · [W12 (x; r)]T n̂ dA.
∂Sr
where W12 (x; r) = [ω 1 (x; r) ω 2 (x; r)]. The term involving the projection of n̂ in
the second half of the surface integral of the above equation is independent to the im-
age gradient. This term varies along the boundary of the spherical region ∂Sr . It is a
weighting function that makes the projected image gradients at various positions on the
sphere surface contribute differently to the resultant values of Λ12 (x; r). The values of
||[W12 (x; r)]T n̂|| on the local spherical region surface are shown in Figures 1a-c. A
large value of ||[W12 (x; r)]T n̂|| represents the region where Λ12 (x; r) is sensitive, as
the projected image gradient at that region receives a higher weight for the computa-
tion of Λ12 (x; r). The large valued regions of ||[W12 (x; r)]T n̂|| are distributed in a
ring shape around the axis ω 3 (x; r). In a curvilinear structure having circular cross sec-
tion, the image gradient at the object boundary points to the centerline of the structure.
Therefore, at the centerline of the structure, Λ12 (x; r) delivers the strongest response if
r and the radius of the structure are matched.
Finally, it is worth mentioning that the elaboration of Λ12 (·) merely demonstrates
a possibility to integrate different eigenvalues to facilitate the analysis of curvilinear
structures. It is possible to devise other combinations of eigenvalues of the proposed
method analogous to those presented in [12] and [6].
2.4 Regarding Multiscale Detection
Multiscale detection is an essential technique for handling structures with various sizes.
The multiscale detection of OOF involves repetitive computations of OOF using a set
of radii (r in Equation 1). The radius set should cover both the narrowest and the widest
curvilinear structures in an image volume. Since the evaluation of OOF is localized
at the spherical region boundary, the spherical region has to touch the target structure
boundary to obtain detection responses of OOF. As such, linear radius samples should
be taken for OOF with the consideration of the voxel length in order to properly detect
vessels in a given range of radii. It also ensures that a structure with non-circular cross
section can induce detection responses of OOF obtained in at least one radius sample.
We suggest that radius samples are taken in every 0.5 voxel length according to the
Nyquist sampling rate.
For different values of r, the area covered by the surface integral of Equation 1
varies. Dividing the computation result of Equation 1 by 4πr2 (the surface area of the
spherical region) is an appropriate mean to normalize the detection response over radii
and hence, the computation of Equation 1 is scale-invariant. Such normalization is es-
sential to aggregating OOF responses in a multiple scale setting. For the same rea-
son, the eigenvalues of Qr,x , λi (r, x) are divided by 4πr2 prior to being utilized in
any multiscale framework. This OOF normalization scheme is distinct to the average-
outward-flux (AOF) measure [5], which divides the outward flux by the surface area of
the spherical region to attain the AOF-limiting-behavior. The AOF measure works only
on the gradient of a distance function of a shape with its boundary clearly delineated.
OOF, in contrast, is applied to a gradient of a gray-scale image, where no explicit shape
boundary is embedded and noise is possibly present.
(a) (b)
Intensity Intensity
1 1
0.5 0.5
0 0
L1 C1 R1 L2 C2 R2 L1 C1 R1 L2 C2 R2
(c) (d)
Tr(Qr,x)/4πr 2
0.1 0.1
0 0
−0.1 1
−0.2 2
r=1 r=2 r=3 r=4 r=5 r=6 r=1 r=2 r=3 r=4 r=5 r=6
L1 C1 R1 L2 C2 R2 L1 C1 R1 L2 C2 R2
(e) (f)
Fig. 2. Examples of evaluating OOF using multiple radii. (a, b) The slices of z = 0 (left) and
x = 0 (right) of two synthetic image volumes consisting of synthetic tubes with a radius of 4
voxel length. C1 and C2 are the positions of the centers of the tubes. L1, R1 and L2, R2 are the
positions of the boundaries of the tubes centered at C1 and C2 respectively. (b) The width of the
separation between the closely located tubes is 2 voxel length. (c, d) The intensity profiles along
the line x = 0, z = 0 of the synthetic image volumes shown in (a) and (b) respectively. (e, f) The
normalized trace of Qr,x along the line x = 0, z = 0 of the image volumes shown in (a) and (b)
respectively.
In Figures 2a-f, we show two examples of evaluating OOF on image volumes con-
sisting of one synthetic tube (Figures 2a and c) and two closely located synthetic tubes
(Figures 2b and d) using multiple radii. The normalized trace of the matrix Qr,x (Equa-
tions 9), which is equal to the sum of the normalized eigenvalues of Qr,x , is utilized
to quantify the detection response strength of OOF. The normalized trace of the matrix
Qr,x is computed using multiple radii in both of the synthetic image volumes. In Fig-
ures 2e and f, it is observed that the normalized trace of Qr,x is negative for all radii
inside the tubes. It attains its maximal negative values at the tube centers and with the
radius r matching the tube radius, i.e. r = 4. The magnitudes of the normalized trace
of Qr,x with r = 4 decline at positions away from the tube centers. In these positions,
it attains its maximal magnitudes with smaller values of r when approaching the tube
boundaries. Therefore, making use of the normalized trace of Qr,x as well as the nor-
malized eigenvalues of Qr,x , (the trace of Qr,x is equal to the sum of its eigenvalues),
with maximal negative values or maximal magnitudes over radii is capable of delivering
a strong detection responses inside curvilinear structures.
When OOF is computed using multiple radii, the spherical regions of OOF with
large radii possibly overshoot the narrow structure boundaries. The computation of OOF
with overshot radii can include the objects nearby and adversely affects the detection
responses of OOF (see Figure 2e, r = 5 and 6 versus Figure 2f, r = 5 and 6). In which,
utilizing the normalized eigenvalues or the normalized trace of the matrix Qr,x with
the maximal negative values or maximal magnitudes over radii as mentioned above can
eliminate the responses obtained by using overshot radii. Furthermore, it excludes the
OOF responses associated with undersized radii at the center of curvilinear structures
(see Figures 2e and f, r = 1, 2 and 3). In the case that the radius of the spherical region
r matches the target structures, OOF avoids the inclusion of objects nearby. It therefore
reports the same response at the centerlines of the tubes with r = 4 despite the presence
of closely located structures (see Figure 2e, r = 4 versus Figure 2f, r = 4).
In this section, we compare the performance of OOF and the Hessian matrix by using
both synthetic data and real clinical cases. The differential terms of the Hessian matrix
are obtained by employing the central mean difference scheme on the image smoothed
by a Gaussian kernel with scale factor !. The eigenvalues and eigenvectors extracted
from the Hessian matrix and Q for OOF (Equation 10) are represented as λH i (x; r),
Q Q
ωHi (x; r) and λi (x; r), ω i (x; r), respectively. The order of the eigenvalues and the
Q
notation of sums of the first two eigenvalues (ΛH 12 (x; r) and Λ12 (x; r)) are analogous
to those described in Section 2.2.
3.1 Synthetic Data

The proposed method, OOF, is examined in this section using synthetic images con-
taining tori with various sizes. There are 10 synthetic volumetric images in the size of
100 × 100 × 100 voxels being generated for the synthetic experiments. The main pur-
pose is to verify the performance of OOF and compare OOF with the Hessian matrix
when closely located structures are present.
The configurations of the tori in the synthetic images are shown in Figure 3. The
number of tori in different synthetic images varies and depends on the values of d and
R. The tori are placed in a layer fashion along the z-direction. The strategy to generate
the first layer of tori is to place a torus with D = 10 at the altitude z = 8. The center
z-direction
The center of the tori

in the (i+1)th layer
The tori in the

(i+1)th layer
The tori in the d d+2R
ith layer R
D
The center of the tori
in the ith layer d
Fig. 3. The description of the tori. These tori have been used in the synthetic data experiments.
The center of the tori in each layer is randomly selected from the positions of (x = 35, y =
35), (x = 45, y = 35), (x = 35, y = 45) and (x = 45, y = 45). The values of d and R are
fixed to generate a torus image. In the experiments, there are 10 torus images generated by using
10 pairs of {d, R}, {2, 1}, {2, 2}, {2, 3}, {2, 4}, {2, 5}, {5, 1}, {5, 2}, {5, 3}, {5, 4} and {5, 5}.
of that torus is randomly selected among the positions (x = 45, y = 45, z = 8),
(x = 35, y = 45, z = 8), (x = 45, y = 35, z = 8) and (x = 35, y = 35, z = 8). We
keep deploying adjacent tori centered at the same position of the first torus but having
larger values of D in an interval of 2R + d until D ≤ 42. Each successive layer of tori
is generated in a 2R + d interval of altitude z for z ≤ 90. The center of each layer of
tori is randomly selected among the positions of (x = 35, y = 35), (x = 45, y = 35),
(x = 35, y = 45) and (x = 45, y = 45). The background intensity of these images
is 0 and the intensity inside the tori is assigned to 1. The torus images are smoothed
by a Gaussian kernel with scale factor 1 to mimic the smooth intensity transition from
structures to background. Each synthetic image is corrupted by two levels of additive
Gaussian noise, with standard deviations of σnoise = {0.75, 1}. Finally, 20 testing cases
are generated for this experiment.
The experiment results are based on the measures obtained in the estimated object
scales of the both methods. For the testing objects with circular cross section such as the
tori used in this experiment, computing the sums of the first two eigenvalues ΛH 12 (·) and
ΛQ12 (·) at structure centerlines is useful to determine the structure scales. The reason is
that ΛH 12 (·) of the Hessian matrix quantifies the second order intensity change occurred
along the radial direction of a circle on the normal plane of the structure. Meanwhile, for
OOF, ΛQ 12 (·) evaluates the amount of gradient pointing to the centerlines of tubes with
circular cross section. Based on the above observation, the object scale is obtained as
2
SxH = arg max(− s3 ΛH √s
12 (x; 3 )) for the Hessian matrix (see [7,14] for details regarding
s∈E
the structure scale detection and [10] for Hessian matrix based feature normalization
Q
over scales) and SxQ = arg max(− 4πs1
2 Λ12 (x; s)) for OOF. The set of discrete detection
s∈F
scales of OOF and detection scales of the Hessian matrix are represented as F and E
respectively. These scales cover the structure radii ranged from 1 to 6 voxel length. The
radii of OOF are taken for each 0.5 voxel length and there are in total 11 different radii in
F . Meanwhile, the same number of scales are logarithmically sampled for the Hessian
matrix scale set E so as to minimize the detection error of the Hessian matrix [12].
There are two measures being studied for the comparison of OOF and the Hessian
matrix, “Angular discrepancy” and “Response fluctuation”. For objects with circular
cross section and having stronger intensity than the background, the third eigenvector
represents the structure direction. At the estimated structure scales, we measure the
angular discrepancy of the Hessian matrix and OOF by
Q Q
arccos(|Gt · ω H
3 (x; St )|), arccos(|Gt · ω 3 (x; St )|),
H
(12)
respectively, where Gt is the ground truth direction, which is defined as the tangent
direction of the torus inner-tube centerline at the position t, t ∈ T , where T is a set
of samples taken in every unit voxel length at the inter-tube centerlines of the tori.
Bilinear interpolation is applied if t does not fall on an integer coordinate. The value
of the angular discrepancy is in a range of [0, π/2] and a small value of the angular
discrepancy represents an accurate estimation of structure direction.
The second measure, “Response fluctuation” for the tori having circular cross section
is defined as the ratio between the variance and the mean absolute value of Λ12 (·). The
“Response fluctuation” of the Hessian matrix and OOF are defined as
Table 1. The performance of optimally oriented flux and the Hessian matrix obtained in the syn-
thetic data experiments. The entries in the columns of ”Angular discrepancy” include two values,
the mean and the standard deviation (the bracketed values) of the resultant values of Equation 12.
The values in the columns of ”Response fluctuation” are the results based on Equation 13.
d = 5, σnoise = 0.75 d = 2, σnoise = 0.75

Angular discrepancy Response fluctuation Angular discrepancy Response fluctuation
R Hessian matrix OOF Hessian matrix OOF R Hessian matrix OOF Hessian matrix OOF
1 0.406 (0.250) 0.309 (0.176) 0.270 0.246 1 0.408 (0.260) 0.304 (0.178) 0.283 0.252
2 0.232 (0.197) 0.180 (0.093) 0.166 0.160 2 0.305 (0.215) 0.227 (0.129) 0.218 0.195
3 0.109 (0.111) 0.110 (0.065) 0.092 0.095 3 0.162 (0.155) 0.135 (0.072) 0.133 0.117
4 0.063 (0.068) 0.062 (0.054) 0.059 0.054 4 0.098 (0.127) 0.087 (0.055) 0.092 0.085
5 0.054 (0.075) 0.059 (0.027) 0.052 0.056 5 0.079 (0.125) 0.065 (0.033) 0.086 0.069
d = 5, σnoise = 1 d = 2, σnoise = 1
Angular discrepancy Response fluctuation Angular discrepancy Response fluctuation
R Hessian matrix OOF Hessian matrix OOF R Hessian matrix OOF Hessian matrix OOF
1 0.518 (0.288) 0.409 (0.239) 0.321 0.291 1 0.532 (0.305) 0.414 (0.243) 0.338 0.298
2 0.331 (0.252) 0.246 (0.148) 0.210 0.200 2 0.435 (0.278) 0.319 (0.192) 0.272 0.239
3 0.204 (0.218) 0.169 (0.109) 0.129 0.105 3 0.279 (0.243) 0.200 (0.132) 0.177 0.134
4 0.112 (0.158) 0.110 (0.080) 0.089 0.080 4 0.181 (0.220) 0.125 (0.095) 0.127 0.108
5 0.107 (0.159) 0.082 (0.044) 0.073 0.061 5 0.157 (0.217) 0.097 (0.088) 0.107 0.085

Var ΛH 12 (x; St )
H
Var ΛQ Q
12 (x; St )
t∈T
, t∈T
, (13)
Mean ΛH 12 (x; S H )
t

Mean ΛQ (x; S Q )

t∈T 12 t
t∈T
respectively. A small value of fluctuation implies a stable response, which is robust

against the adverse effects introduced by the interference of closely located structures
as well as image noise.
The results based on the above measurements for different combinations of noise
levels and torus separations are presented and listed in Table 1. In Table 1, it is ob-
served that both the Hessian matrix and OOF perform better when the inner-tube radii
of tori rise. It is because structures having low curvature surfaces such as large inner-
tube radius tori are easier to be detected than the tori having small inner-tube radii.
To evaluate the performance drops of OOF and the Hessian matrix in handling images
having closely located structures, the changes of the mean angular discrepancy and re-
sponse fluctuation in various cases are investigated in Table 2. In the entries of Table 2,
a small value represents high robustness against the reduction of torus separation (Table
2a); the increment of noise level (Table 2b); and both of them (Table 2c).
As previously mentioned, the detection of OOF is localized at the boundary of local
spherical regions. The OOF detection responses are merely induced from the intensity
discontinuities taken place at the structure boundary, when the local sphere surface of
OOF touches the structure boundary. In contrast to OOF, the Hessian matrix based de-
tection relies on the computation of the weighted intensity average difference between
the regions inside the structure and in the vicinity of the structure, where a nearby ob-
ject is possibly present. As the correct detection scale of the Hessian matrix increases
for recognizing large scale structures, the detection coverage of the correct scale of the
Hessian matrix expands. It increases the chances to include adjacent structures. Hence,
the increments of mean angular discrepancies and response fluctuations of the Hessian
Table 2. The changes of mean angular discrepancy and response fluctuation from the case of
”d = 5, σnoise = 0.75” to other three cases presented in Table 1
(a)
From ”d = 5, σnoise = 0.75” to ”d = 2, σnoise = 0.75”
Changes of mean angular discrepancy Changes of response fluctuation
R Hessian matrix OOF Hessian matrix OOF
1 +0.002 -0.005 +0.013 +0.006
2 +0.073 +0.048 +0.052 +0.035
3 +0.053 +0.025 +0.041 +0.023
4 +0.035 +0.024 +0.033 +0.031
5 +0.025 +0.005 +0.034 +0.012
(b)
From ”d = 5, σnoise = 0.75” to ”d = 5, σnoise = 1”
1 +0.112 +0.100 +0.050 +0.045
2 +0.099 +0.067 +0.044 +0.040
3 +0.095 +0.059 +0.036 +0.010
4 +0.049 +0.047 +0.030 +0.026
5 +0.053 +0.023 +0.021 +0.004
(c)
From ”d = 5, σnoise = 0.75” to ”d = 2, σnoise = 1”
1 +0.126 +0.104 +0.068 +0.052
2 +0.203 +0.139 +0.106 +0.079
3 +0.170 +0.090 +0.085 +0.039
4 +0.118 +0.062 +0.068 +0.054
5 +0.103 +0.037 +0.054 +0.029
matrix are larger than those of OOF, especially when R increases, in the cases that the
torus separation is reduced from 5 voxel length to 2 voxel length (the second and the
forth columns versus the first and the third columns of Table 2a).
Moreover, in the situation where noise is increased (Table 2b), it is observed that
OOF (the second and the forth columns) has less increment of the mean angular dis-
crepancies than the Hessian matrix (the first and the third columns), particularly when
R increases. Although the Gaussian smoothing taken by the Hessian matrix partially
eliminates noise from the image volume, the smoothing process also reduces the edge
sharpness of the structure boundaries. In particular, the scale factor of the Gaussian
smoothing process of the Hessian matrix has to rise to deal with large scale structures.
Consequently, the Hessian matrix performs detection based on the smoothed object
boundaries which are easier to be corrupted by image noise. For OOF, the detection
does not require Gaussian smoothing using a large scale factor (σ = 1 for OOF). It re-
tains the edge sharpness of the structure boundaries. Therefore, the OOF detection has
higher robustness against image noise than the Hessian matrix. As expected, when the
torus separation is reduced to 2 voxel length and the noise level is raised to σnoise = 1,
OOF has higher robustness than the Hessian matrix, against the presence of both closely
located adjacent structures and high level noise than the Hessian matrix (Table 2c).
To summarize the results of the synthetic data experiments (Tables 1 and 2), OOF
is validated in several aspects, the structure direction estimation accuracy, the stabil-
ity of responses, the robustness against the disturbance introduced by closely located
structures and the increment of noise levels. In some applications, an accurate structure
direction estimation is vital. For instance, vascular image enhancement, the estimated
structure direction is to avoid smoothing along the directions across object boundaries.
Furthermore, for tracking curvilinear structure centerlines (a centerline tracking exam-
ple is in [1]), estimated structure direction is to guide the centerline tracking process.
Also, small response fluctuation facilitates the process to extract curvilinear structures
or locate object centerlines by discovering the local maxima or ridges of the response.
On the other hand, the structure direction estimation accuracy and the stability of
structure responses of OOF are robust against the reduction of structure separation and
the increment of noise levels. As such, employing OOF to provide information of curvi-
linear structures is highly beneficial for curvilinear structure analysis.
3.2 Application Example - Blood Vessel Extraction

In this section, we demonstrate an example on utilizing OOF to supply information of
curvilinear structures for extracting vessels in a vascular image. The vascular image
utilized in this example is a phase contrast magnetic resonance angiographic (PCMRA)
image volume (Figure 4a) and the image intensity represents the blood flow speed inside
the vasculature. The challenges to extraction algorithms are the presence of closely
located vessels due to the complicated geometry of vascular structures, and the small
and low intensity vessels in images with relatively high background noise level.
To perform comparison between OOF and the Hessian matrix, we replace the Hes-
sian matrix based information used by a vessel extraction measure with the similar
information extracted from OOF. It is reminded that the main goal of this paper is to
propose OOF as a general curvilinear structure detector. Therefore, measures having
heuristic parameters which involve different values for certain kinds of structures are
not preferred in this example, such as the vesselness measure [6] or majority of tech-
niques in [12] for integrating multiple eigenvalues which involve heuristic parameters.
On the other hand, the sum of the first two eigenvalues employed in the synthetic ex-
periments is designed to provide responses at centerlines of curvilinear structures. It
is not suitable for vessel extraction, which requires a measure to give vessel detection
responses in the entire image region. We make use of the geometric mean of the first
two eigenvalues, which was proposed for the detection of vessels in [12,7],
5.
|λ1 (x; s)λ2 (x; s)|, λ1 (x; s) ≤ λ2 (x; s) < 0,
M(x; s) = (14)
0, otherwise,
This measure

2 is computedin a set of discrete scales to obtain the object scales, SxH =

arg max s3 MH (x; √s3 ) for Hessian matrix and SxQ = arg max 4πs 1
2 M (x; s)
Q
s∈E s∈F
for OOF. There are 15 radii and scales being employed for F and E respectively to
cover the vessel radii ranged from 1 to 8 voxel length. Linear radius samples for F
and logarithmic scale samples for E are utilized analogous to those described in the
synthetic experiments. The vessel measure response is retrieved as the resultant values
of Equation 14 obtained in the estimated object scales. The binary extraction results
are obtained by thresholding the vessel measure responses. The thresholding value is
found empirically so that neither over-segmentation nor under-segmentation of major
vessels is observed and the same amount of voxels for both the methods are selected.
Finally, 4% of voxels having the highest vessel measure responses among all voxels are
thresholded as the extraction results.
The vessel extraction results are shown in Figures 4b and c. The interesting positions
in the results are highlighted by the numbered arrows in Figures 4b and c. In the regions
pointed at by the fifth and sixth arrows in Figures 4b and c, the Hessian based method
misidentifies closely located vessels as merged structures. On the contrary, the OOF
based method is capable of discovering the small separation between the closely located
vessels. This result is consistent with the findings in the synthetic experiments, where
OOF is more robust than the Hessian matrix when handling closely located structures
(Table 2a).
In Figure 4c, it is found that several vessels with weak intensity (arrows 1, 2, 3, 4
and 7) are missed by the Hessian based method where the OOF based method has
no problem to extract them (Figure 4b). The reason is that the noise level relative to
the weak intensity structures is higher than those relative to strong intensity structures.
Axial view Sagittal view
Coronal view
(a)
(b) (c)
Fig. 4. (a) A phase contrast magnetic resonance angiographic image volume with the size of
213 × 143 × 88 voxels. (b) The vessel extraction results obtained by using the optimally oriented
flux based method. (c) The vessel extraction results obtained by using the Hessian matrix based
method.
Coherent to the synthetic experiments, in which OOF shows higher robustness against
image noise as compared to Hessian matrix (see Table 2b). The vessel extraction results
in this real case experiment reflects that robustness against image noise is important on
extracting vessels with weak intensity.
4 Future Developments and Conclusion
In this paper, we have presented the use of optimally oriented flux (OOF) for detecting
curvilinear structures. With the aid of the analytical Fourier expression of OOF, no dis-
cretization and orientation sampling are needed. It therefore leads to a highly efficient
computation of OOF. Computation-wise, it has the same complexity as in the compu-
tation of the commonly used approach, Hessian matrix. Furthermore, computation of
OOF is based on the image gradient at the boundary of local spheres. It focuses on the
detection of intensity discontinuities occurred at the object boundaries of curvilinear
structures.
The OOF based detection avoids including the adjacent objects. Thus, it exhibits the
robustness against the interference introduced by closely located adjacent structures.
This advantage is validated and demonstrated by a set of experiments on synthetic and
real image volumes. In addition, in the experiments, it is observed that OOF has higher
structure direction estimation accuracy and stable detection responses under the distur-
bance of high level image noise. With the aforementioned high detection accuracy and
robustness, OOF as opposed to the Hessian matrix, to supply information of curvilinear
structures, is more beneficial for curvilinear structure analysis.
In this paper, our current focus is on formulating OOF as a general detector for
extracting reliable information of curvilinear structures. Identifying branches, high cur-
vature curvilinear structures or distinguishing between blob-like, sheet-like and tubular
structures would involve post-processing steps of the information extracted by the curvi-
linear structure detector, such as those presented in [12]. Considering the robustness of
OOF against image noise and interference of closely located adjacent structures, tailor-
ing appropriate post-processing steps of OOF for various kinds of structures will be an
interesting direction for the future developments of this work.
References
1. Aylward, S., Bullitt, E.: Initialization, noise, singularities, and scale in height ridge traversal
for tubular object centerline extraction. TMI 21(2), 61–75 (2002)
2. Bouix, S., Siddiqi, K., Tannenbaum, A.: Flux driven fly throughs. CVPR 1, 449–454 (2003)
3. Bouix, S., Siddiqi, K., Tannenbaum, A.: Flux driven automatic centerline extraction. Me-
dIA 9(3), 209–221 (2005)
4. Bracewell, R.: The Fourier Transform and Its Application. McGraw-Hill, New York (1986)
5. Dimitrov, P., Damon, J.N., Siddiqi, K.: Flux invariants for shape. CVPR 1, I–835–I–
841(2003)
6. Frangi, A., Niessen, W., Viergever, M.: Multiscale vessel enhancement filtering. In: Wells,
W.M., Colchester, A.C.F., Delp, S.L. (eds.) MICCAI 1998. LNCS, vol. 1496, pp. 130–137.
7. Koller, T., Gerig, G., Szekely, G., Dettwiler, D.: Multiscale detection of curvilinear structures
in 2-d and 3-d image data. In: IEEE International Conference on Computer Vision, pp. 864–
869 (1995)
8. Krissian, K.: Flux-based anisotropic diffusion applied to enhancement of 3-d angiogram.
TMI 21(11), 1440–1442 (2002)
9. Krissian, K., Malandain, G., Ayache, N., Vaillant, R., Trousset, Y.: Model-based multiscale
detection of 3d vessels. CVPR 3, 722–727 (1998)
10. Lindeberg, T.: Edge detection and ridge detection with automatic scale selection. IJCV 30(2),
117–156 (1998)
11. Manniesing, W.N.R., Viergever, M.A.: Vessel enhancing diffusion a scale space representa-
tion of vessel structures. MedIA 10(6), 815–825 (2006)
12. Sato, Y., Nakajima, S., Shiraga, N., Atsumi1, H., Yoshida, S., Koller, T., Gerig, G., Kikinis,
R.: Three-dimensional multi-scale line filter for segmentation and visualization of curvilinear
structures in medical images. MedIA 2(2), 143–168 (1998)
13. Schey, H.M.: div, grad, curl, and all that, 3rd edn. W.W.Norton & Company (1997)
14. Steger, C.: An unbiased detector of curvilinear structures. PAMI 20(2), 113–125 (1998)
15. Vasilevskiy, A., Siddiqi, K.: Flux maximizing geometric flows. PAMI 24(12), 1565–1578
(2002)
Scene Segmentation for Behaviour Correlation
Jian Li, Shaogang Gong, and Tao Xiang
Department of Computer Science

Queen Mary College, University of London, London, E1 4NS, UK
{jianli,sgg,txiang}@dcs.qmul.ac.uk
Abstract. This paper presents a novel framework for detecting ab-

normal pedestrian and vehicle behaviour by modelling cross-correlation
among different co-occurring objects both locally and globally in a given
scene. We address this problem by first segmenting a scene into semantic
regions according to how object events occur globally in the scene, and
second modelling concurrent correlations among regional object events
both locally (within the same region) and globally (across different re-
gions). Instead of tracking objects, the model represents behaviour based
on classification of atomic video events, designed to be more suitable for
analysing crowded scenes. The proposed system works in an unsupervised
manner throughout using automatic model order selection to estimate
its parameters given video data of a scene for a brief training period. We
demonstrate the effectiveness of this system with experiments on public
road traffic data.
1 Introduction
Automatic abnormal behaviour detection has been a challenging task for visual
surveillance. Traditionally, anomaly is defined according to how individuals be-
have in isolation over space and time. For example, objects can be tracked across
a scene and if a trajectory cannot be matched by a set of known trajectory model
templates, it is considered to be abnormal [1,2]. However, due to scene complex-
ity, many types of abnormal behaviour are not well defined by only analysing
how individuals behave alone. In other words, many types of anomaly definition
are only meaningful when behavioural interactions/correlations among differ-
ent objects are taken into consideration. In this paper, we present a framework
for detecting abnormal behaviour by examining correlations of behaviours from
multiple objects. Specifically, we are interested in subtle multiple object abnor-
mality detection that is only possible when behaviours of multiple objects are
interpreted in correlation as the behaviour of each object is normal when viewed
in isolation. To that end, we formulate a novel approach to representing visual
behaviours and modelling behaviour correlations among multiple objects.
In this paper, a type of behaviour is represented as a class of visual events
bearing similar features in position, shape and motion information [3]. However,
instead of using per frame image events, atomic video events as groups of image
events with shared attributes over a temporal window are extracted and utilised

384 J. Li, S. Gong, and T. Xiang
as the basic units of representation in our approach. This reduces the sensitivity
of events to image noise in crowded scenes. The proposed system relies on both
globally and locally classifying atomic video events. Behaviours are inherently
context-aware, exhibited through constraints imposed by scene layout and the
temporal nature of activities in a given scene. In order to constrain the number of
meaningful behavioural correlations from potentially a very large number of all
possible correlations of all the objects appearing everywhere in the scene, we first
decompose semantically the scene into different spatial regions according to the
spatial distribution of atomic video events. In each region, events are re-clustered
into different groups with ranking on both types of events and their dominating
features to represent how objects behave locally within each region. As shown in
Section 5, by avoiding any attempt to track individual objects over a prolonged
period in space, our representation provides an object-independent representa-
tion that aims to capture categories of behaviour regardless contributing objects
that are associated with scene location. We demonstrate in our experiments
that such an approach is more suitable and effective for discovering unknown
and detecting subtle abnormal behaviours attributed by unusual presence of
and correlation among multiple objects.
Behavioural correlation has been studied before, although it is relatively new
compared to the more established traditional trajectory matching based tech-
niques. Xiang and Gong [3] clustered local events into groups and activities
are modelled as sequential relationships among event groups using Dynamic
Bayesian Networks. Their extended work was shown to have the capability of
detecting suspicious behaviour in front of a secured entrance [4]. However, the
types of activities modelled were restricted to a small set of events in a small
local region without considering any true sense of global context. Brand and
Kettnaker [5] attempted modelling scene activities from optical flows using a
Multi-Observation-Mixture+Counter Hidden Markov Model (MOMC-HMM). A
traffic circle at a crossroad is modelled as sequential states and each state is a
mixture of multiple activities (observations). However, their anomaly detection
is based only on how an individual behaves in isolation. How activities inter-
act in a wider context is not considered. Wang et al [6] proposed hierarchical
Bayesian models to learn visual interactions from low-level optical flow features.
However, their framework is difficult to be extended to model behaviour cor-
relation across different type of features, in which adding more features would
significantly increase complexity of their models.
In our work, we model behaviour correlation by measuring the frequency of
co-occurrence of any pairs of commonly occurred behaviours both locally and re-
motely over spatial locations. An accumulated concurrence matrix is constructed
for a given training video set and matched with an instance of this matrix cal-
culated for any testing video clip in order to detect irregular object correlations
in the video clip both within the same region and across different regions in the
scene. The proposed approach enables behaviour correlation to be modelled be-
yond a local spatial neighbourhood. Furthermore, representing visual behaviours
using different dominant features at different spatial locations makes it possible
Scene Segmentation for Behaviour Correlation 385
Fig. 1. Semantic scene segmentation and behaviour correlation for anomaly detection
to discover subtle unusual object behaviour correlations that either human prior
knowledge is unaware of or it is difficult to be defined by human analysis. An
overall data flow of the system is shown in Fig. 1.
2 Event Detection and Clustering

2.1 Image Events
We define an image event as a group of foreground neighbouring pixels detected
using background subtraction. Different background models can be adopted.
When only moving objects are of interest, we can use a dynamic Gaussian-
Mixture background model [7]. As we also want to extract those long-staying
objects, an alternative background model [8] is preferred.
Detected foreground pixels are grouped into blobs using connected compo-
nents, with each blob corresponding to an image event given by a rectangular
bounding box. An image event vf is represented by a set of 10 features given
the membership of a group as follows:
vf = [x, y, w, h, rs , rp , u, v, ru , rv ], (1)
where (x, y) and (w, h) are the centroid position and the width and height of the
bounding box respectively, rs = w/h is the ratio between width and height, rp is
the percentage of foreground pixels in a bounding box, (u, v) is the mean optic
flow vector for the bounding box, ru = u/w and rv = v/h are the scaling features
between motion information and blob shape. Clearly, some of these features are
more dominant for certain image events depending on their loci in a scene, as
they are triggered by the presence and movement of objects in those areas of the
scene. However, at this stage of the computation, we do not have any information
about the scene therefore all 10 features are used at this initial step to represent
all the detected image events across the entire scene.
Given detected image events, we wish to seek a behavioural grouping of these
image events with each group associated with a similar type of behaviour. This
shares the spirit with the work of Xiang and Gong [3]. However, direct grouping
of these image events is unreliable because they are too noisy due to their spread
over a wide-area outdoor scene under variable conditions. It has been shown
by Gong and Xiang [9] that precision of feature measurements for events af-
fects strongly the performance of event grouping. When processing video data of
crowded outdoor scenes of wide-areas, variable lighting condition and occlusion
can inevitably introduce significant noise to the feature measurement. Instead
of directly grouping image events, we introduce an intermediate representation
of atomic video event which is less susceptible to scene noise.
2.2 Atomic Video Events

Derived from image events, an atomic video event is defined as a spatio-temporal
group of image events with similar features. To generate atomic video events, a
video is cut into short non-overlapping clips and image events within a single
clip are clustered into groups using K-means. Each group then corresponds to an
atomic video event. In our system, we segment a video into clips of equal frame
length Nf , where Nf is between 100 to 300 depending on the nature of a scene.
For K-means clustering in each clip, the number of clusters is set to the average
number of image event across all the frames in this clip. An atomic video event is
represented by both the mean feature values of all the membership image events
in its cluster, and their corresponding variances, resulting in a 20 components
feature vector for each atomic video event, consisting of:
v = [v̄f , v̄s ], (2)
where v̄f = mean(vf ) and v̄s = var(vf ), vf given by Eqn. (1).
2.3 Event Grouping

We seek a behavioural grouping of all the atomic video events detected in the
scene in a 20 dimensional feature space. Here we assume an atomic video event
being a random variable following a Mixture of Gaussian (MoG) distribution.
We need to determine both the number of Gaussian components in the mix-
ture (model order selection) and their parameters. To automatically determine
the model order, we adopt the Schwarz’s Bayesian Information Criterion (BIC)
model selection method [10]. Given the number of Gaussians K being deter-
mined, the Gaussian parameters and priors are computed using Expectation-
Maximisation [11]. Each atomic video event is associated with the kth Gaussian
representing a behaviour class in the scene, 1 ≤ k ≤ K, which gives the maxi-
mum posterior probability.
3 Scene Segmentation
This behavioural grouping of atomic video events gives a concise and semanti-
cally more meaningful representation of a scene (top middle plot in Fig. 1). We
consider that each group represents a behaviour type in the scene. However, such
a behaviour representation is based on a global clustering of all the atomic video
events detected in the entire scene without any spatial or temporal restriction.
It thus does not provide a good model for capturing behaviour correlations more
selectively, both in terms of spatial locality and temporal dependency. In order
to impose more contextual constraints, we segment a scene semantically into
regions according to event distribution with behaviour labelling, as follows.
We treat the problem similar to an image segmentation problem except that
we represent each image position by a multivariate feature vector instead of RGB
values. To that end, we introduce a mapping procedure transferring features from
event domain to image domain. We assign each image pixel location of the scene
a feature vector p with K components, where K is the number of groups of
atomic video events estimated for a given scene, i.e. the number of behaviour
types automatically determined by the BIC algorithm (Section 2.3). The value
of the kth component pk is given as the count of the kth behaviour type occurred
at this image position throughout the video. In order to obtain reliable values
of p, we use the following procedure. First of all, the behavioural type label for
an atomic video event is applied to all image events belonging to this atomic
video event. Secondly, given an image event, its label is applied to all pixels
within its rectangular bounding box. In other words, each image position is
assigned with a histogram of different types of behaviours occurred at that pixel
location for a given video. Moreover, because we perform scene segmentation by
activities, those locations without or with few activities are to be removed from
the segmentation procedure. For doing this, we apply a lower bound threshold
T Hp to the number of events happened at each pixel location, i.e. the sum of
component values of p. Finally the value of this K dimensional feature vector p
at each pixel location is scaled to [0, 1] for scene segmentation.
With this normalised behavioural histogram representation in the image do-
main, we employ a spectral clustering technique modified from the method pro-
posed by Zelnik-Manor and Perona [12]. Given a scene in which N locations with
activities, an N × N affinity matrix A is constructed and the similarity between
the features at the ith position and the jth position is computed according to
Eqn. (3),
⎧

⎨ (d(pi ,pj ))2 (d(xi ,xj ))2
exp − σi σj exp − , if xi − xj ≤ r
A(i, j) = σx 2
(3)
⎩
0, otherwise
where pi and pj are feature vectors at the ith and the jth locations, d represents
Euclidean distance, σi and σj correspond to the scaling factors for the feature
vectors at the ith and the jth positions, xi and xj are the coordinates and σx
is the spatial scaling factor. r is the radius indicating a circle only within which,
similarity is computed.
Proper computation of the scaling factors is a key for reliable spectral cluster-
ing. The original Zelnik-Perona’s method computes σi using the distance between
the current feature and the feature for a specific neighbour. This setting is very
arbitrary and we will show that it suffers from under-fitting in our experiment. In
order to capture more accurate statistics of local feature similarities, we compute
σi as the standard deviation of feature distances between the current location
and all locations within a given radius r. The scaling factor σx is computed as
the mean of the distances between all positions and the circle center within the
radius r. The affinity matrix is then normalised according to:
Ā = L− 2 AL− 2
1 1
(4)
N
where L is a diagonal matrix with L(s, s) = t=1 (A(s, t)). Ā is then used as
the input to the Zelnik-Perona’s algorithm which automatically determines the
number of segments and performs segmentation. This procedure groups those
pixel locations with activities into M regions for a given scene.
4 Behaviour Concurrence Modelling

4.1 Regional Event Classification
Recall that due to the lack of any prior information at the initial behavioural
grouping stage for scene segmentation, all 10 features together with their corre-
sponding variances were used to represent atomic video events. These settings are
not necessarily optimal for accurately describing behaviours once the scene has
been segmented semantically into regions. To address the problem, we re-classify
behaviours in each region. Essentially, we follow the same procedure described
in Section 2 but perform an additional computation to refine the grouping of
atomic video events in each individual region as follows.
Given a segmented scene, we determine the most representative features in
each region by computing entropy values for the features in vf in each region and
select the top five features with high entropy values. The selected features are
then used for grouping image events in each video clip into atomic video events.
When representing atomic video events, their corresponding 5 variances are also
considered. This results in different and smaller set of features being selected for
representing events in different regions. After atomic video event clustering, we
obtain Km regional event classes in each region m, where 1 ≤ m ≤ M .
4.2 Behaviour Correlation Modelling

Suppose we have now obtained M in total Ko clusters of atomic video events in
all regions, i.e. Ko = m=1 K m , we wish to examine the frequency of con-
currence among all pairs of behaviours happened in the scene throughout a
video. Given a training video F which is segmented into Nc non-overlapping
clips F = [f1 , · · · , fNc ], each atomic video event in a single clip fn , 1 ≤ n ≤ Nc ,
has been clustered to a specific regional event class bi , where 1 ≤ i ≤ Ko . To
indicate the concurrence of a pair of regional event classes bi and bj occurred

in clip n, we construct a Ko × Ko dimension binary matrix Cn so that
5
1, if bi = TRUE and bj = TRUE
Cn (i, j) = (5)
0, otherwise
An accumulated concurrence matrix C over all the clips in the video is then
computed as:
Nc
C= Cn (6)
n=1
It is clear that the diagonal components of C indicate the number of occur-
rence of event class bi throughout the video and each other component C(i, j)
corresponds to the total number of concurrence of event classes bi and bj . To
normalise the accumulated concurrence matrix C, components in each row of C
is divided by the diagonal component in this row. This results in a non-symmetric
normalised matrix Ce . The final symmetric concurrence matrix is computed as:
1
Cf = (Ce + CTe ), (7)
2
where T is transpose. After re-scaling the values in Cf to [0, 1], Cf is then used
as the model to recognise irregular behaviour labelled atomic video event con-
currence. It is worth pointing out that in practice, a measurement of concurrent
frequency between a pair of atomic video event classes bi and bj is meaningful
only when bi and bj individually occur sufficiently frequently. In order to remove
those rarely occurred regional event classes from the concurrence matrix during
training, we set a lower bound threshold T Hb to the diagonal components of
accumulated concurrence matrix C. If C(i, i) < T Hb , the ith row and the ith
column are removed from C. The rectified matrix C is then used for generating
the concurrence matrix Cf .
4.3 Anomaly Detection

A test video is segmented into clips in the same way as the training video set.
Image events are grouped into atomic video events using K-means. Each atomic
video event is then assigned to a regional event class. In order to detect anomaly
due to unexpected multi-object behaviour concurrence, we identify abnormal
video clips as those with unexpected pairs of concurrences of regional event
classes when compared with the concurrence matrix constructed from the train-
ing video set. More precisely, for a test video Q with Nq clips: Q = [q1 , · · · , qNq ],
we generate a binary concurrence matrix Ct for each clip qt by Eqn. (5). We
then generate a matrix CTt according to Eqn. (8).
5
1 − Cf (i, j), if Ct (i, j) = 1 and Cf (i, j) ≤ T Hc
CTt (i, j) = (8)
0, otherwise
where T Hc is a threshold. Given matrix CTt for clip qt , a score St is computed
as the mean of all the non-zero values in CTt . Based on the values of St , t =
1, · · · , Nq , those clips with unexpected behavioural concurrence can be identified

if the corresponding St values are higher than a threshold T Hs . In the identified
irregular video clips, pairs of unexpected concurrent regional event classes can
be further detected as the pairs whose values in Cf are lower than T Hc .
5 Experiments
We evaluated the performance of the proposed system using video data captured
from two different public road junctions (Scene-1 and Scene-2). Example frames
are shown in Fig. 2. Scene-1 is dominated by three types of traffic patterns:
the vertical traffic, the leftward horizontal traffic and the rightward traffic, from
multiple entry and exit points. In addition, vehicles are allowed to stop between
the vertical traffic lanes waiting for turning right or left. In Scene-2, vehicles
usually move in from the entrances near the left boundary and near the right
bottom corner. They move towards the exits located on the top, at left bottom
corner and near the right boundary. Both videos were recorded at 25Hz and have
a frame size of 360×288 pixels.
Failure Mode For Tracking: We first highlight the inadequacy of tracking

based representation for behaviour modelling in a crowded scene such as Scene-
1. Fig. 3 (a) shows the trajectories extracted from a two-minute video clip. In
(b), we plot a histogram of the durations of all the tracked object trajectories
(red), 331 in total and compare it to that of the ground-truth (blue), which
was exhaustively labelled manually for all the objects appeared in the scene (in
total 114 objects). It is evident that inevitable and significant fragmentation
of object trajectories makes a purely trajectory based representation unsuitable
for accurate behaviour analysis in this type of scenes. Moreover, it is equally
important to point out that monitoring object in isolation even over a prolonged
period of time through tracking does not necessarily facilitate the detection and
discovery of unexpected and previously unknown anomaly in a complex scene.
(a) Scene-1 (b) Scene-2
Fig. 2. Two public road scenarios for experiment

(a) Trajectories (b) Histogram of duration
Fig. 3. Trajectory analysis
Event Clustering and Scene Segmentation: In this section, we show the

performance of semantic event clustering and scene segmentation. In Scene-1,
22000 frames were used for training, in which 121583 image events were de-
tected and grouped into 2117 atomic video events using K-means. In Scene-2,
415637 image events were detected from 45000 frames and grouped into 4182
atomic video events. The global atomic video events were automatically grouped
into 13 and 19 clusters using the EM algorithm where the number of clus-
ters in each scene was automatically determined by the BIC model selection
method. The clustering results are shown in Fig. 4 (a) and (d) where clusters are
(a) Scene-1 (b) Proposed: T Hp = 300 (c) Original: T Hp = 300
(d) Scene-2 (e) Proposed: T Hp = 200 (f) Original: T Hp = 200
Fig. 4. Atomic video event classification and semantic scene segmentation

Table 1. Regional feature selection
x y w h rs rp u v ru rv
√ √ √ √√
R1
√ √ √ √ √
R2
√ √ √√ √
R3
√ √ √√ √
R4
√ √ √ √ √
R5
√ √√√ √
R6
distinguished by colour and labels. After mapping from feature domain to image
domain, the modified Zelnik-Manor and Perona’s image segmentation algorithm
was then used to segment Scene-1 and Scene-2 into 6 regions and 9 regions,
respectively, as shown in Fig. 4 (b) and (e). For comparison, we also segmented
the scenes using Zelnik-Manor and Perona’s original algorithm (ZP) which re-
sulted in 4 segments for Scene-1 and 2 segments for Scene-2 (Fig. 4 (c) and
(f)). It is evident that Zelnik-Manor and Perona’s original algorithm suffered
from under-fitting severely and was not able to segment those scenes correctly
according to expected traffic behaviours. In contrast, our approach provides a
more meaningful semantic segmentation of both scenes.
Anomaly Detection: We tested the performance of anomaly detection using

Scene-1. Comparing to Scene-2, Scene-1 contains more complex behaviour corre-
lations that also subject to frequent deviations from normal correlations. Given
the labelled scene segmentation shown in Fig. 4 (b), we re-classified atomic video
events in each region. We performed a feature selection procedure which selected
the 5 dominant features in each region with largest entropy values. The selected
features in each region are shown in Table 1.
Atomic video events were then clustered in each region. From region 1 to
region 6, the BIC determined 6, 5, 6, 4, 5 and 4 classes of events (behaviours)
respectively. The clustering resulted in 30 local clusters of atomic video events
in total (see Fig. 5 (a)). The number of concurrence for each pair of atomic event
classes was then accumulated using the 73 clips in the training data to construct
a 30 × 30 dimension accumulating concurrence matrix C. By removing those
behaviour which occurred less than 10 times (i.e. T Hb = 10), the dimension of
the matrix C was reduced to 25 × 25. The concurrence matrix Cf was then
computed by normalising and re-scaling C which is shown in Fig. 5 (b).
According to the scores shown in Fig. 5 (c), 7 clips had been picked out of
a testing video consisting of 12000 frames (39 clips) as being abnormal with
irregular concurrences shown in Fig. 6, in which objects with irregular concur-
rence are bounded by red and green boxes and the corresponding segments are
highlighted using colour. Clip 4 detected a situation when a fire engine suddenly
appeared and the surrounding moving vehicles had to stop unexpectedly. In Clip
28, another fire engine appeared. Although the fire engine did not significantly
interrupt the normal traffic, it did caused a white van to stop in Region 3 which
was not expected to be concurrent with horizontal traffic. A typical example was
(a) Distributions of Local (b) Concurrence Matrix (c) Anomaly Scores: T Hc =

Behaviours 0.12, T Hs = 0.9
Fig. 5. Local events classification and anomaly detection. In (a), the mean and covari-
ance of the location of different classes of regional events are illustrated using ellipses
in different colour.
(a) Clip 4 (b) Clip 9 (c) Clip 27
(d) Clip 28 (e) Clip 30 (f) Clip 37
(g) Clip 38 (h) Clip 38
Fig. 6. Detected irregular concurrences

(a) Anomaly scores (b) False Alarm 1 (c) False Alarm 2
Fig. 7. False detections without scene segmentation
detected in Clip 30. Moreover, the second fire engine also caused strange driving
behaviour for another car labelled in Clip 28 which strongly conflicted with the
normal traffic. In Clip 9 and 37, two right-turn vehicles were detected in Region 2
and Region 5 respectively showing that they were quite close to each other which
were not observed in the training data. Clip 27 indicates a false alarm mainly due
to the imperfect blob detection which resulted in regional events being classified
into wrong classes. In Clip 38, the irregular atomic events were detected in the
same clip without frame overlapping (Fig. 6 (g) and (h)). This is an example
that when the size of objects are large enough to cover two regions, error could
also be introduced as most of vehicles in the training data have smaller size.
For comparison, we performed irregular concurrence detection without scene
segmentation, i.e. only using globally clustered behaviours. The results are shown
in Fig. 7. Compared with the proposed scheme, the scheme without scene seg-
mentation gave much more false alarms (comparing (a) of Fig. 7 with (c) of
Fig. 5). From the examples of false detections in Fig. 7 (b) and (c), it can be
seen that using global behaviours without scene decomposition cannot accu-
rately represent how objects behave locally. In other words, each of the global
behaviour categories for the vehicles and pedestrians may not truly reflect the
local behaviours of the objects and this would introduce more errors in detect-
ing such abnormal correlations of subtle and short-duration behaviours. On the
other hand, true irregular incidents were missed, e.g. the interruption from the
fire engine was ignored. To summarise, when only using global classification,
contextual constraints on local behaviour is not described accurately enough
and general global correlation is too arbitrary. This demonstrates the advantage
in behaviour correlation based on contextual constraint from semantic scene
segmentation.
6 Conclusion
This paper presented a novel framework for detecting abnormal pedestrian and
vehicle behaviour by modelling cross-correlation among different co-occurring
objects both locally and globally in a given scene. Without tracking objects, the
system was built based on local image events and atomic video events, which
made the system more suitable for crowded scenes. Based on globally classified
atomic video events, a scene was semantically segmented into regions and in
each region, more detailed local events were re-classified. Local and global events
correlations were learned by modelling event concurrence within the same region
and across different regions. The correlation model was then used for detecting
anomaly.
The experiments with public traffic data have shown the effectiveness of the
proposed system on scene segmentation and anomaly detection. Compared with
the scheme which identified irregularities only using atomic video events clas-
sified globally, the proposed system provided more detailed description of local
behaviour, and showed more accurate anomaly detection and less false alarms.
Furthermore, the proposed system is entirely unsupervised which ensures its gen-
eralisation ability and flexibility on processing video data with different scene
content and complexity.
References
1. Hu, W., Xiao, X., Fu, Z., Xie, D., Tan, T., Maybank, S.: A system for learning
statistical motion patterns. PAMI 28 (9), 1450–1464 (2006)
2. Johnson, N., Hogg, D.: Learning the distribution of object trajectories for event
recognition. BMVC 2, 583–592 (1995)
3. Xiang, T., Gong, S.: Beyond tracking: Modelling activity and understanding be-
haviour. IJCV 67 (1), 21–51 (2006)
4. Xiang, T., Gong, S.: Video behavior profiling for anomaly detection. PAMI 30(5),
893–908 (2008)
5. Brand, M., Kettnaker, V.: Discovery and segmentation of activities in video.
PAMI 22(8), 844–851 (2000)
6. Wang, X., Ma, X., Grimson, W.E.L.: Unsupervised activity perception by hierar-
chical bayesian models. In: CVPR, Minneapolis, USA, June 18-23, pp. 1–8 (2007)
7. Stauffer, C., Grimson, W.E.L.: Adaptive background mixture models for real-time
tracking. In: CVPR, vol. 2, pp. 246–252 (1999)
8. Russell, D., Gong, S.: Minimum cuts of a time-varying background. In: BMVC,
Edinburgh, UK, 1–10 (September 2006)
9. Gong, S., Xiang, T.: Scene event recognition without tracking. Special issue on
visual surveillance, Acta Automatica Sinica 29(3), 321–331 (2003)
10. Schwarz, G.: Estimating the dimension of a model. Annals of Statistics 6(2), 461–
464 (1978)
11. Dempster, A., Laird, N., Rubin, D.: Maximum likelihood from incomplete data
via the em algorithm. Journal of the Royal Statistical Society, series B 39(1), 1–38
(1977)
12. Zelnik-Manor, L., Perona, P.: Self-tuning spectral clustering. In: NIPS (2004)
Robust Visual Tracking Based on an Effective
Appearance Model
Xi Li1 , Weiming Hu1 , Zhongfei Zhang2, and Xiaoqin Zhang1

1
National Laboratory of Pattern Recognition, CASIA, Beijing, China
{lixi,wmhu,xqzhang}@nlpr.ia.ac.cn
2
State University of New York, Binghamton, NY 13902, USA
zhongfei@cs.binghamton.edu
Abstract. Most existing appearance models for visual tracking usually construct
a pixel-based representation of object appearance so that they are incapable of
fully capturing both global and local spatial layout information of object ap-
pearance. In order to address this problem, we propose a novel spatial Log-
Euclidean appearance model (referred as SLAM) under the recently introduced
Log-Euclidean Riemannian metric [23]. SLAM is capable of capturing both the
global and local spatial layout information of object appearance by constructing
a block-based Log-Euclidean eigenspace representation. Specifically, the process
of learning the proposed SLAM consists of five steps—appearance block division,
online Log-Euclidean eigenspace learning, local spatial weighting, global spatial
weighting, and likelihood evaluation. Furthermore, a novel online Log-Euclidean
Riemannian subspace learning algorithm (IRSL) [14] is applied to incrementally
update the proposed SLAM. Tracking is then led by the Bayesian state inference
framework in which a particle filter is used for propagating sample distributions
over the time. Theoretic analysis and experimental evaluations demonstrate the
promise and effectiveness of the proposed SLAM.
1 Introduction
For visual tracking, handling appearance variations of an object is a fundamental and

challenging task. In general, there are two types of appearance variations: intrinsic and
extrinsic. Pose variation and/or shape deformation of an object are considered as the
intrinsic appearance variations while the extrinsic variations are due to the changes
resulting from different illumination, camera motion, camera viewpoint, and occlusion.
Consequently, effectively modeling such appearance variations plays a critical role in
visual tracking.
Hager and Belhumeur [1] propose a tracking algorithm which uses an extended
gradient-based optical flow method to handle object tracking under varying illumina-
tion conditions. In [3], curves or splines are exploited to represent the appearance of
an object to develop the Condensation algorithm for contour tracking. Due to the sim-
plistic representation scheme, the algorithm is unable to handle the pose or illumination
change, resulting in tracking failures under a varying lighting condition. Zhao et al.[18]
present a fast differential EMD tracking method which is robust to illumination changes.
Silveira and Malis [16] present a new algorithm for handling generic lighting changes.

Robust Visual Tracking Based on an Effective Appearance Model 397
Black et al.[4] employ a mixture model to represent and recover the appearance
changes in consecutive frames. Jepson et al.[5] develop a more elaborate mixture model
with an online EM algorithm to explicitly model appearance changes during tracking.
Zhou et al.[6] embed appearance-adaptive models into a particle filter to achieve a ro-
bust visual tracking. Wang et al.[20] present an adaptive appearance model based on the
Gaussian mixture model (GMM) in a joint spatial-color space (referred to as SMOG).
SMOG captures rich spatial layout and color information. Yilmaz [15] proposes an
object tracking algorithm based on the asymmetric kernel mean shift with adaptively
varying the scale and orientation of the kernel. Nguyen et al.[17] propose a kernel-
based tracking approach based on maximum likelihood estimation.
Lee and Kriegman [7] present an online learning algorithm to incrementally learn
a generic appearance model for video-based recognition and tracking. Lim et al.[8]
present a human tracking framework using robust system dynamics identification and
nonlinear dimension reduction techniques. Black et al.[2] present a subspace learning
based tracking algorithm with the subspace constancy assumption. A pre-trained, view-
based eigenbasis representation is used for modeling appearance variations. However,
the algorithm does not work well in the scene clutter with a large lighting change due
to the subspace constancy assumption. Ho et al.[9] present a visual tracking algorithm
based on linear subspace learning. Li et al.[10] propose an incremental PCA algorithm
for subspace learning. In [11], a weighted incremental PCA algorithm for subspace
learning is presented. Limy et al.[12] propose a generalized tracking framework based
on the incremental image-as-vector subspace learning methods with a sample mean
update. Chen and Yang [19] present a robust spatial bias appearance model learned
dynamically in video. The model fully exploits local region confidences for robustly
tracking objects against partial occlusions and complex backgrounds. In [13], li et al.
present a visual tracking framework based on online tensor decomposition.
However, the aforementioned appearance-based tracking methods share a problem
that their appearance models lack a competent object description criterion that captures
both statistical and spatial properties of object appearance. As a result, they are usually
sensitive to the variations in illumination, view, and pose. In order to tackle this prob-
lem, Tuzel et al. [24] and Porikli et al.[21] propose a covariance matrix descriptor for
characterizing the appearance of an object. The covariance matrix descriptor, based on
several covariance matrices of image features, is capable of fully capturing the infor-
mation of the variances and the spatial correlations of the extracted features inside an
object region. In particular, the covariance matrix descriptor is robust to the variations
in illumination, view, and pose. Since a nonsingular covariance matrix is a symmetric
positive definite (SPD) matrix lying on a connected Riemannian manifold, statistics for
covariance matrices of image features may be computed through Riemannian geome-
try. Nevertheless, most existing algorithms for statistics on a Riemannian manifold rely
heavily on the affine-invariant Riemannian metric, under which the Riemannian mean
has no closed form. Recently, Arsigny et al.[23] propose a novel Log-Euclidean Rie-
mannian metric for statistics on SPD matrices. Under this metric, distances and Rieman-
nian means take a much simpler form than the widely used affine-invariant Riemannian
metric.
398 X. Li et al.
Based on the Log-Euclidean Riemannian metric [23], we develop a tracking frame-

work in this paper. The main contributions of the developed framework are as follows.
First, the framework does not need to know any prior knowledge of the object, and only
assumes that the initialization of the object region is provided. Second, a novel block-
based spatial Log-Euclidean appearance model (SLAM) is proposed to fully capture
both the global and local spatial properties of object appearance. In SLAM, the object
region is first divided into several p × q object blocks, each of which is represented
by the covariance matrix of image features. A low dimensional Log-Euclidean Rie-
mannian eigenspace representation for each block is then learned online, and updated
incrementally over the time. Third, we present a spatial weighting scheme to capture
both the global and local spatial layout information among blocks. Fourth, while the
Condensation algorithm [3] is used for propagating the sample distributions over the
time, we develop an effective likelihood function based on the learned Log-Euclidean
eigenspace model. Last, the Log-Euclidean Riemannian subspace learning algorithm
(i.e., IRSL) [14] is applied to update the proposed SLAM as new data arrive.
2 The Framework for Visual Tracking

2.1 Overview of the Framework
The tracking framework includes two stages:(a) online SLAM learning; and (b)Bayesian
state inference for visual tracking.
In the first stage, five steps are needed. They are appearance block division, online
Log-Euclidean eigenspace learning, local spatial weighting, global spatial weighting,
and likelihood evaluation, respectively. A brief introduction to these five steps is given
as follows. First, the object appearance is uniformly divided into several blocks. Sec-
ond, the covariance matrix feature from Eq. (2) in [14] is extracted for representing
each block. After the Log-Euclidean mapping from Eq. (5) in [14], a low dimensional
Log-Euclidean Riemannian eigenspace model is learned online. The model uses the
incremental Log-Euclidean Riemannian subspace learning algorithm (IRSL) [14] to
find the dominant projection subspaces of the Log-Euclidean unfolding matrices. Third,
the block-specific likelihood between a candidate block and the learned Log-Euclidean
eigenspace model is computed to obtain a block related likelihood map for object ap-
pearance. Fourth, the likelihood map is filtered by local spatial weighting into a new
one. Fifth, the filtered likelihood map is further globally weighted by a spatial Gaussian
kernel into a new one. Finally, the overall likelihood between a candidate object region
and the learned SLAM is computed by multiplying all the block-specific likelihoods
after local and global spatial weighting.
In the second stage, the object locations in consecutive frames are estimated by max-
imum a posterior (MAP) estimation within the Bayesian state inference framework in
which a particle filter is applied to propagate sample distributions over the time. Af-
ter MAP estimation, we just use the block related Log-Euclidean covariance matrices
of image features inside the affinely warped image region associated with the highest
weighted hypothesis to update the SLAM.
These two stages are executed repeatedly as time progresses. Moreover, the frame-
work has a strong adaptability in the sense that when new image data arrive, the Log-
Fig. 1. The architecture of the tracking framework
Euclidean Riemannian eigenspace observation model follows the updating online. The
architecture of the framework is shown in Fig. 1.
2.2 Spatial Log-Euclidean Appearance Model (SLAM)
The process of learning the SLAM consists of five steps—appearance block division, on-
line Log-Euclidean eigenspace learning, local spatial weighting, global spatial weight-
ing, and likelihood evaluation. The details of these five steps are given as follows.
(1) Appearance block division. Given an object appearance tensor F = {F t ∈

Rm×n }t=1,2,...,N , we divide the object appearance F t at any time t into several p × q
blocks (m = n = 36 and p = q = 6 in the paper), as illustrated in Figs. 2(a) and
(b). For each block Fijt ∈ Rp×q , the covariance matrix feature from Eq. (2) in [14]
is extracted for representing Fijt , i.e., Ctij ∈ Rd×d . We call the covariance matrix Ctij
as the block-(i, j) covariance matrix. In this case, the block-(i, j) covariance matri-
ces {Ctij }t=1,2,...,N constitute a block-(i, j) covariance tensor Aij ∈ Rd×d×N . If Ctij
is a singular matrix, we replace Ctij with Ctij + Id , where is a very small positive
constant ( = 1e − 18 in the experiments), and Id is a d × d identity matrix. By the
Log-Euclidean mapping from Eq. (5) in [14], as illustrated in Fig. 2(c), the block-(i, j)
covariance subtensor Aij is transformed into a new one:
LAij = {log(C1ij ), . . . , log(Ctij ), . . . , , log(CN

ij )} (1)
where is a very small positive constant, and Id is a d × d identity matrix. We call

LAij as the block-(i, j) Log-Euclidean covariance subtensor, as illustrated in Fig. 2(d).
Denote [·] as the rounding operator, m∗ as [ m ∗
p ], and n as [ q ]. Consequently, all the Log-
n
Euclidean covariance subtensors {(LAij )m∗ ×n∗ }t=1,2,...,N forms a Log-Euclidean co-
variance tensor LA associated with the object appearance tensor F ∈ Rm×n×N . With
the emergence of new object appearance subtensors, F is extended along the time axis t
(i.e., N increases gradually), leading to the extension of each Log-Euclidean covariance
subtensor LAij along the time axis t. Consequently, we need to track the changes of
LAij , and need to identify the dominant projection subspace for a compact representa-
tion of LAij as new data arrive.
400 X. Li et al.
Fig. 2. Illustration of appearance block division, Log-Euclidean mapping, and Log-Euclidean

unfolding. A face image F t at time t is shown in the upper part of (a) while a 3-order face tensor
F = {F t }t=1,2,...,N (i.e., face image ensemble) is displayed in the lower one of (a). The results
of appearance block division are exhibited in (b). The Log-Euclidean mapping results are shown
in (c). An example of the block-(i, j) Log-Euclidean mapping is given in (d). (e) displays the
results of Log-Euclidean unfolding.
Due to the vector space structure of log(Ctij ) under the Log-Euclidean Riemannian
metric, log(Ctij ) is unfolded into a d2 -dimensional vector vect(i) which is formulated as:
vect(i) = UT(log(Ctij )) = (ct1 , ct2 , . . . , ctd2 )T (2)
where UT(·) is an operator unfolding a matrix into a column vector. The unfolding
process can be illustrated by Figs. 2(e) and 3(a). In Fig. 3(a), the left part displays the
covariance tensor Aij ∈ Rd×d×N , the middle part corresponds to the Log-Euclidean
covariance tensor LAij , and the right part is associated with the Log-Euclidean unfold-
ing matrix LAij with the t-th column being vectij for 1 ≤ t ≤ N . As a result, LAij is
formulated as:
LAij = vec1ij vec2ij · · · vectij · · · vecN
ij . (3)
The next step of the SLAM is to learn an online Log-Euclidean eigenspace model
for LAij . Specifically, we will introduce an incremental Log-Euclidean Riemannian
subspace learning algorithm (IRSL) [14] for the Log-Euclidean unfolding matrix LAij .
IRSL applies the online learning technique (R-SVD [12,27]) to find the dominant pro-
jection subspaces of LAij . Furthermore, a new operator CVD(·) used in IRSL is de-
fined as follows. Given a matrix H = {K1 , K2 , . . . , Kg } and its column mean K,
we let CVD(H) denote the SVD (i.e., singular value decomposition) of the matrix
{K1 − K, K2 − K, . . . , Kg − K}.
(2) Online Log-Euclidean eigenspace learning. For each Log-Euclidean covariance

subtensor LAij , IRSL [14] is used to incrementally learn a Log-Euclidean eigenspace
model (i.e., LAij ’s column mean L̄ij and CVD(LAij ) = Uij Dij VTij ) for LAij . For
convenience, we call L̄ij and CVD(LAij ) as the block-(i, j) Log-Euclidean eigenspace
Fig. 3. Illustration of Log-Euclidean unfolding and IRSL. (a) shows the generative process of the
Log-Euclidean unfolding matrix; (b) displays the incremental learning process of IRSL.
model. For a better understanding of IRSL, Fig. 3(b) is used to illustrate the incremental
learning process of IRSL. Please see the details of IRSL in [14].
The distance between a candidate sample Bi,j and the learned block-(i, j) Log-
Euclidean eigenspace model (i.e. LAij ’s column mean L̄ij and CVD(LAij ) =
Uij Dij VTij ) is determined by the reconstruction error norm:
REij =(vecij −L̄ij )−U(j) · U(j)

T
· (vecij −L̄ij )2 (4)
where · is the Frobenius norm, and vecij = UT(log(Bi,j )) is obtained from Eq. (2).
Thus, the block-(i, j) likelihood pij is computed as: pij ∝ exp(−REij ). The smaller
the REij , the larger the likelihood pij . As a result, we can obtain a likelihood map
∗ ∗
M = (pij )m∗ ×n∗ ∈ Rm ×n for all the blocks.
(3) Local spatial weighting. In this step, the likelihood map M is filtered into a new
∗ ∗
one Ml ∈ Rm ×n . The details of the filtering process are given as follows. Denote the
Fig. 4. Illustration of local spatial weighting for the i-th and j-th block. (a) shows the original
likelihood map while (b) displays the filtered map by local spatial weighting for the i-th and j-th
block.
402 X. Li et al.
original map M = (pij )m∗ ×n∗ , and the filtered map Ml = (plij )m∗ ×n∗ . After filtering
by local spatial weighting, the entry plij of Ml is formulated as:
9 :
Nij+ − Nij−
plij ∝ pij · exp , (5)
σij
% & % |puv −pij |−(puv −pij ) &
|p −p |+(p −p )
where Nij+ = k1ij sgn uv ij 2 uv ij , Nij−= k1ij sgn 2 ,
u,v∈Nij u,v∈Nij
|·| is a function returning the absolute value of its argument, sgn[·] is a sign function, σij
is a positive scaling factor (σij = 8 in the paper), Nij denotes the neighbor elements
of pij , and ki,j stands for the number of the neighbor elements. In this paper, if all
the 8-neighbor elements of pij exist, ki,j = 8; otherwise, ki,j is the number of the
valid 8-neighbor elements of pij . A brief discussion on the theoretical properties of
Eq. (5) is given as follows. The second term of Eq. (5)(i.e., exp(·)) is a local spatial
weighting factor. If Nij+ is smaller than Nij− , the factor will penalize pij ; otherwise it
will encourage pij . The process of local spatial weighting is illustrated in Fig. 4.
(4) Global spatial weighting. In this step, the filtered likelihood map Ml =(plij )m∗ ×n∗
is further globally weighted by a spatial Gaussian kernel into a new one Mg = (pgij ) ∈
∗ ∗
Rm ×n . The global spatial weighting process is formulated as follows.

pgij ∝ plij · exp −posij − poso 2 /2σp2ij

+ −
Nij −Nij (6)
∝ pij · exp −posij − poso /2σpij · exp
2 2
σij
where posij is the block-(i, j) positional coordinate vector, poso is the positional co-
ordinate vector associated with the center O of the likelihood map Ml , and σpij is a
scaling factor (σpij = 3.9 in the paper). The process of global spatial weighting can
be illustrated by Fig. 5, where the likelihood map Ml (shown in Fig. 5(a)) is spatially
weighted by the Gaussian kernel (shown in Fig. 5(b)).
Fig. 5. Illustration of global spatial weighting. (a) shows the original likelihood map Ml while
(b) exhibits the spatial weighting kernel for Ml .
(5) Likelihood evaluation for SLAM. In this step, the overall likelihood between
a candidate object region and the learned SLAM is computed by multiplying all the
block-specific likelihoods after local and global spatial weighting. Mathematically, the
likelihood is formulated as:

LIKI ∝ pgij
∗ ∗
1≤i≤m 1≤j≤n

+ −
Nij −Nij (7)
∝ pij · exp −posij − poso 2 /2σp2ij · exp σij
i j
2.3 Bayesian State Inference for Visual Tracking
For visual tracking, a Markov model with a hidden state variable is used for motion
estimation. In this model, the object motion between two consecutive frames is usually
assumed to be an affine motion. Let Xt denote the state variable describing the affine
motion parameters (the location) of an object at time t. Given a set of observed images
Ot = {O1 , . . . , Ot }, the posterior probability is formulated by Bayes’ theorem as:

p(Xt |Ot )∝p(Ot |Xt ) p(Xt |Xt−1 )p(Xt−1 |Ot−1 )dXt−1 (8)
where p(Ot | Xt ) denotes the observation model, and p(Xt | Xt−1 ) represents the dy-
namic model. p(Ot |Xt ) and p(Xt |Xt−1 ) decide the entire tracking process. A particle
filter [3] is used for approximating the distribution over the location of the object using
a set of weighted samples.
In the tracking framework, we apply an affine image warping to model the object
motion of two consecutive frames. The six parameters of the affine transform are used
to model p(Xt | Xt−1 ) of a tracked object. Let Xt = (xt , yt , ηt , st , βt , φt ) where
xt , yt , ηt , st , βt , φt denote the x, y translations, the rotation angle, the scale, the aspect
ratio, and the skew direction at time t, respectively. We employ a Gaussian distribution
to model the state transition distribution p(Xt | Xt−1 ). Also the six parameters of the
affine transform are assumed to be independent. Consequently, p(Xt |Xt−1 ) is formu-
lated as:
p(Xt |Xt−1 ) = N (Xt ; Xt−1 , Σ) (9)
where Σ denotes a diagonal covariance matrix whose diagonal elements are σx2 , σy2 , ση2 ,
σs2 , σβ2 , σφ2 , respectively. The observation model p(Ot | Xt ) reflects the similarity be-
tween a candidate sample and the learned SLAM. In this paper, p(Ot |Xt ) is formulated
as: p(Ot |Xt ) ∝ LIKI, where LIKI is defined in Eq. (7). After maximum a posterior
(MAP) estimation, we just use the block related Log-Euclidean covariance matrices of
features inside the affinely warped image region associated with the highest weighted
hypothesis to update the block related Log-Euclidean eigenspace model.
3 Experiments
In order to evaluate the performance of the proposed tracking framework, four videos
are used in the experiments. The first three videos are recorded with moving cameras
404 X. Li et al.
while the last video is taken from a stationary camera. The first two videos consist of
8-bit gray scale images while the last two are composed of 24-bit color images. Video
1 consists of dark gray scale images, where a man moves in an outdoor scene with
drastically varying lighting conditions. In Video 2, a man walks from left to right in a
bright road scene; his body pose varies over the time, with a drastic motion and pose
change (bowing down to reach the ground and standing up back again) in the middle of
the video stream. In Video 3, a girl changes her facial pose over the time in a color scene
with varying lighting conditions. Besides, the girl’s face is severely occluded by a man
in the middle of the video stream. In the last video, a pedestrian moves along a corridor
in a color scene. In the middle of the video stream, his body is severely occluded by the
bodies of two other pedestrians.
During the visual tracking, the size of each object region is normalized to 36 × 36
pixels. Then, the normalized object region is uniformly divided into thirty-six 6 × 6
blocks. Further, a block-specific SLAM is online learned and online updated by IRSL
every three frames. The maintained dimension rij of the block-(i, j) Log-Euclidean
eigenspace model (i.e., Uij referred in Sec. 2.2) learned by IRSL is obtained from the
experiments. For the particle filtering in the visual tracking, the number of particles is
set to be 200. The six diagonal elements (σx2 , σy2 , ση2 , σs2 , σβ2 , σφ2 ) of the covariance
matrix Σ in Eq. (9) are assigned as (52 , 52 , 0.032, 0.032 , 0.0052, 0.0012 ), respectively.
Three experiments are conducted to demonstrate the claimed contributions of the
proposed SLAM. In these four experiments, we compare tracking results of SLAM with
those of a state-of-the-art Riemannian metric based tracking algorithm [21], referred
here as CTMU, in different scenarios including drastic illumination changes, object pose
variation, and occlusion. CTMU is a representative Riemannian metric based tracking
algorithm which uses the covariance matrix of features for object representation. By us-
ing a model updating mechanism, CTMU adapts to the undergoing object deformations
and appearance changes, resulting in a robust tracking result. In contrast to CTMU,
SLAM constructs a block-based Log-Euclidean eigenspace representation to reflect the
appearance changes of an object. Consequently, it is interesting and desirable to make a
comparison between SLAM and CTMU. Furthermore, CTMU does not need additional
parameter settings since CTMU computes the covariance matrix of image features as
the object model. More details of CTMU are given in [21].
The first experiment is to compare the performances of the two methods SLAM and
CTMU in handling drastic illumination changes using Video 1. In this experiment, the
maintained eigenspace dimension rij in SLAM is set as 8. Some samples of the final
tracking results are demonstrated in Fig. 6, where rows 1 and 2 are for SLAM and
CTMU, respectively, in which five representative frames (140, 150, 158, 174, and 192)
of the video stream are shown. From Fig. 6, we see that SLAM is capable of tracking
the object all the time even in a poor lighting condition. In comparison, CTMU is lost
in tracking from time to time.
The second experiment is for a comparison between SLAM and CTMU in the sce-
narios of drastic pose variation using Video 2. In this experiment, rij in SLAM is set as
6. Some samples of the final tracking results are demonstrated in Fig. 7, where rows 1
and 2 correspond to SLAM and CTMU, respectively, in which five representative frames
(142, 170, 178, 183, and 188) of the video stream are shown. From Fig. 7, it is clear
Fig. 6. The tracking results of SLAM (row 1) and CTMU (row 2) over representative frames with
drastic illumination changes
Fig. 7. The tracking results of SLAM (row 1) and CTMU (row 2) over representative frames with
drastic pose variation
that SLAM is capable of tracking the target successfully even with a drastic pose and
motion change while CTMU gets lost in tracking the target after this drastic pose and
motion change.
The last experiment is to compare the tracking performance of SLAM with that of
CTMU in the color scenarios with severe occlusions using Videos 3 and 4. The RGB
color space is used in this experiment. rij for Videos 3 and 4 are set as 6 and 8, re-
spectively. We show some samples of the final tracking results for SLAM and CTMU
in Fig. 8, where the first and the second rows correspond to the performances of SLAM
and CTMU over Video 3, respectively, in which five representative frames (158, 160,
162, 168, and 189) of the video stream are shown, while the third and the last rows
correspond to the performances of SLAM and CTMU over Video 4, respectively, in
which five representative frames (22, 26, 28, 32, and 35) of the video stream are shown.
Clearly, SLAM succeeds in tracking for both Video 3 and Video 4 while CTMU fails.
In summary, we observe that SLAM outperforms CTMU in the scenarios of illu-
mination changes, pose variations, and occlusions. SLAM constructs a block-based
Log-Euclidean eigenspace representation to capture both the global and local spatial
properties of object appearance. The spatial correlation information of object appear-
ance is incorporated into SLAM. Even if the information of some local blocks is partially
lost or drastically varies, SLAM is capable of recovering the information using the cues
of the information from other local blocks. In comparison, CTMU only captures the
statistical properties of object appearance in one mode, resulting in the loss of the local
406 X. Li et al.
Fig. 8. The tracking results of SLAM and CTMU over representative frames in the color scenarios
of severe occlusions. Rows 1 and 2 show the tracking results of SLAM and CTMU for Video
4, respectively. Rows 3 and 4 display the tracking results of SLAM and CTMU for Video 5,
respectively.
spatial correlation information inside the object region. In particular, SLAM constructs a
robust Log-Euclidean Riemannian eigenspace representation of each object appearance
block. The representation fully explores the distribution information of covariance ma-
trices of image features under the Log-Euclidean Riemannian metric, whereas CTMU
relies heavily on an intrinsic mean in the Lie group structure without considering the
distribution information of the covariance matrices of image features. Consequently,
SLAM is an effective appearance model which performs well in modeling appearance
changes of an object in many complex scenarios.
4 Conclusion
In this paper, we have developed a visual tracking framework based on the proposed
spatial Log-Euclidean appearance model (SLAM). In this framework, a block-based
Log-Euclidean eigenspace representation is constructed by SLAM to reflect the appear-
ance changes of an object. Then, the local and global spatial weighting operations on
the block-based likelihood map are performed by SLAM to capture the local and global
spatial layout information of object appearance. Moreover, a novel criterion for the
likelihood evaluation, based on the Log-Euclidean Riemannian subspace reconstruc-
tion error norms, has been proposed to measure the similarity between the test image
and the learned subspace model during the tracking. SLAM is incrementally updated by
the proposed online Log-Euclidean Riemannian subspace learning algorithm (IRSL).
Experimental results have demonstrated the robustness and promise of the proposed
framework.
Acknowledgment
This work is partly supported by NSFC (Grant No. 60520120099, 60672040 and
60705003) and the National 863 High-Tech R&D Program of China (Grant No.
2006AA01Z453). Z.Z. is supported in part by NSF (IIS-0535162). Any opinions, find-
ings, and conclusions or recommendations expressed in this material are those of the
authors and do not necessarily reflect the views of the NSF.
References
1. Hager, G., Belhumeur, P.: Real-time tracking of image regions with changes in geometry and
illumination. In: Proc. CVPR, pp. 410–430 (1996)
2. Black, M.J., Jepson, A.D.: Eigentracking: Robust matching and tracking of articulated ob-
jects using view-based representation. In: Buxton, B.F., Cipolla, R. (eds.) ECCV 1996.
3. Isard, M., Blake, A.: Contour tracking by stochastic propagation of conditional density. In:
Buxton, B.F., Cipolla, R. (eds.) ECCV 1996. LNCS, vol. 1065, pp. 343–356. Springer, Hei-
delberg (1996)
4. Black, M.J., Fleet, D.J., Yacoob, Y.: A framework for modeling appearance change in image
sequence. In: Proc. ICCV, pp. 660–667 (1998)
5. Jepson, A.D., Fleet, D.J., El-Maraghi, T.F.: Robust Online Appearance Models for Visual
Tracking. In: Proc. CVPR, vol. 1, pp. 415–422 (2001)
6. Zhou, S.K., Chellappa, R., Moghaddam, B.: Visual Tracking and Recognition Using
Appearance-Adaptive Models in Particle Filters. IEEE Trans. on Image Processing 13, 1491–
1506 (2004)
7. Lee, K., Kriegman, D.: Online Learning of Probabilistic Appearance Manifolds for Video-
based Recognition and Tracking. In: Proc. CVPR, vol. 1, pp. 852–859 (2005)
8. Lim, H., Morariu3, V.I., Camps, O.I., Sznaier1, M.: Dynamic Appearance Modeling for Hu-
man Tracking. In: Proc. CVPR, vol. 1, pp. 751–757 (2006)
9. Ho, J., Lee, K., Yang, M., Kriegman, D.: Visual Tracking Using Learned Linear Subspaces.
In: Proc. CVPR, vol. 1, pp. 782–789 (2004)
10. Li, Y., Xu, L., Morphett, J., Jacobs, R.: On Incremental and Robust Subspace Learning.
Pattern Recognition 37(7), 1509–1518 (2004)
11. Skocaj, D., Leonardis, A.: Weighted and Robust Incremental Method for Subspace Learning.
In: Proc. ICCV, pp. 1494–1501 (2003)
12. Limy, J., Ross, D., Lin, R., Yang, M.: Incremental Learning for Visual Tracking. In: NIPS,
pp. 793–800. MIT Press, Cambridge (2005)
13. Li, X., Hu, W., Zhang, Z., Zhang, X., Luo, G.: Robust Visual Tracking Based on Incremental
Tensor Subspace Learning. In: Proc. ICCV (2007)
14. Li, X., Hu, W., Zhang, Z., Zhang, X., Luo, G.: Visual Tracking Via Incremental Log-
Euclidean Riemannian Subspace Learning. In: Proc. CVPR (2008)
15. Yilmaz, A.: Object Tracking by Asymmetric Kernel Mean Shift with Automatic Scale and
Orientation Selection. In: Proc. CVPR (2007)
16. Silveira, G., Malis, E.: Real-time Visual Tracking under Arbitrary Illumination Changes. In:
Proc. CVPR (2007)
17. Nguyen, Q.A., Robles-Kelly, A., Shen, C.: Kernel-based Tracking from a Probabilistic View-
point. In: Proc. CVPR (2007)
18. Zhao, Q., Brennan, S., Tao, H.: Differential EMD Tracking. In: Proc. ICCV (2007)
408 X. Li et al.
19. Chen, D., Yang, J.: Robust Object Tracking Via Online Dynamic Spatial Bias Appearance
Models. IEEE Trans. on PAMI 29(12), 2157–2169 (2007)
20. Wang, H., Suter, D., Schindler, K., Shen, C.: Adaptive Object Tracking Based on an Effective
Appearance Filter. IEEE Trans. on PAMI 29(9), 1661–1667 (2007)
21. Porikli, F., Tuzel, O., Meer, P.: Covariance Tracking using Model Update Based on Lie Al-
gebra. In: Proc. CVPR, vol. 1, pp. 728–735 (2006)
22. Tuzel, O., Porikli, F., Meer, P.: Human Detection via Classification on Riemannian Mani-
folds. In: Proc. CVPR (2007)
23. Arsigny, V., Fillard, P., Pennec, X., Ayache, N.: Geometric Means in a Novel Vector Space
Structure on Symmetric Positive-Definite Matrices. SIAM Journal on Matrix Analysis and
Applications (2006)
24. Tuzel, O., Porikli, F., Meer, P.: Region Covariance: A Fast Descriptor for Detection and
Classification. In: Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3952,
25. Pennec, X., Fillard, P., Ayache, N.: A Riemannian Framework for Tensor Computing. In:
IJCV, pp. 41–66 (2006)
26. Rossmann, W.: Lie Groups: An Introduction Through Linear Group. Oxford Press (2002)
27. Levy, A., Lindenbaum, M.: Sequential Karhunen-Loeve Basis Extraction and Its Application
to Images. IEEE Trans. on Image Processing 9, 1371–1374 (2000)
Key Object Driven Multi-category Object Recognition,
Localization and Tracking Using Spatio-temporal
Context
Yuan Li and Ram Nevatia
University of Southern California

Institute for Robotics and Intelligent Systems
Los Angeles, CA, USA
{yli8,nevatia}@usc.edu
Abstract. In this paper we address the problem of recognizing, localizing and

tracking multiple objects of different categories in meeting room videos. Dif-
ficulties such as lack of detail and multi-object co-occurrence make it hard to
directly apply traditional object recognition methods. Under such circumstances,
we show that incorporating object-level spatio-temporal relationships can lead to
significant improvements in inference of object category and state. Contextual
relationships are modeled by a dynamic Markov random field, in which recogni-
tion, localization and tracking are done simultaneously. Further, we define human
as the key object of the scene, which can be detected relatively robustly and there-
fore is used to guide the inference of other objects. Experiments are done on the
CHIL meeting video corpus. Performance is evaluated in terms of object detec-
tion and false alarm rates, object recognition confusion matrix and pixel-level
accuracy of object segmentation.
1 Introduction
Object recognition is a fundamental problem of computer vision. Its significance lies

not only in the static image domain but also in video understanding and analysis, e.g., is
the man typing on a laptop or writing on a pad of paper? What objects have been put on
the table and where are they? What is the motion of the passenger and his luggage if he
is carrying any? Answering questions of this kind requires the ability to recognize, lo-
calize and even track different categories of objects from videos captured with a camera
of usually broad view field.
There are a number of difficulties in this problem: background clutter, lack of
image detail, occlusion, multi-object co-occurrence and motion. To enhance purely
appearance-based approaches in the hope of overcoming these difficulties, we incor-
porate contextual information to aid object recognition and localization. There are three
key notions in our approach: 1) spatial relationships between different object categories
are utilized so that co-inference helps enhance accuracy; 2) temporal context is utilized
to accumulate object evidence and to track objects continuously; 3) we borrow tech-
niques from research efforts in single category object recognition to robustly detect key
objects (such as humans) and use them to reduce inference space for other objects.

410 Y. Li and R. Nevatia
Image observation and Inference with spatial-

key object detection temporal context
Spatial
relationship
Temporal
relationship
Result
Human (head-shoulder)
Table
Whiteboard
Computer
Projector
Paper
Fig. 1. Finding objects in spatio-temporal context
These concepts are modeled by a dynamic Markov random field (MRF). Figure 1
gives an illustration. Instead of letting each node represent a pixel or image blob in a
pre-defined grid, as is commonly done in segmentation, in our model a node represents
a hypothetical object in one frame, which enables integration of object-level informa-
tion during inference. Spatial and temporal relationships are modeled by intra-frame
and inter-frame edges respectively. Since objects are recognized on-the-fly and change
with time, the structure of the MRF is also dynamic. To avoid building an MRF with
excessive false hypothetical object nodes, key objects are detected first and provide con-
textual guidance for finding other objects. Inference over the resulting MRF gives an
estimate of the states of all objects through the sequence. We apply our approach to
meeting room scenes with humans as the key objects.
The rest of the paper is organized as follows: Section 2 summarizes related work by
categories; Section 3 gives the formulation of the model; Section 4 defines the potential
functions of the MRF and Section 5 describes the inference algorithm; Section 6 shows
the experimental results; Section 7 discusses about future work and concludes the paper.
2 Related Work
Our approach uses elements from both object recognition and detection. Object recog-
nition focuses on categorization of objects [1][2]; many approaches assume a close-up
view of a single object in the input image. Object detection focuses on single category
object classification and localization from the background [3][4][5]. Both have received
intense research interest recently, bringing forward a large body of literature. While our
approach assimilates several established ideas from the two, our emphasis is on inte-
gration of spatio-temporal context. We hereby focus on the recent growing effort in
tackling object-related problems based on contextual relationships.
Key Object Driven Multi-category Object Recognition, Localization and Tracking 411
Object in the scene. Modeling object-scene relationship enables the use of prior knowl-
edge regarding object category, position, scale and appearance. [6] learns a scene-
specific prior distribution of the reference position of each object class to improve
classification accuracy of image features. It assumes that a single reference position
explains all observed features. [7] proposes a framework for placing local object detec-
tion in the 3D scene geometry of an image. Some other work seeks to classify the scene
and objects at the same time [8][9]. [8] uses the recognized scene to provide strong prior
of object position and scale. Inter-object relationship is not considered. [9] proposes an
approach to recognize events and label semantic regions in images, but the focus is not
on localizing individual objects.
Object categorization and segmentation in context. Object segmentation and cat-

egorization are often combined to enhance each other. When multiple categories are
present, contextual knowledge fits in naturally [10][11][12]. [10] uses Conditional Ran-
dom Field (CRF) to combine appearance, shape and context. Shape filters are used to
classify each pixel, based on the appearance of a neighborhood; no object-level rela-
tionship is explicitly modeled. By counting the co-occurrence of every object pair, [11]
exploits object-level context to refine the category label after each image segment has
been categorized independently. While [11] does not model spatial relationship among
objects, [12] captures spatial relationship by laying a grid-structured MRF on the im-
age, with each node corresponding to the label of a rectangular image blob. Labeling
of one blob is dependent on the labels of its neighbors. However, such relationship is
constrained to adjacent image blobs.
Object and human action. There have been several attempts in collaborative recog-
nition of object category and human action [13][14][15]. [13] uses the hand motion to
improve the shape-based object classification from the top-down view of a desktop. In
[14], objects such as chair, keyboards are recognized from surveillance video of an of-
fice scene. Bayesian classification of regions is done completely based on human pose
and action signatures. Given estimated human upper body pose, [15] accomplishes hu-
man action segmentation and object recognition at the same time. All these approaches
require the ability of tracking human poses or recognizing action, which is not a trivial
task. But they have reflected the fact that many visual tasks are human centered. Namely
the objects of most interest for recognition are those interacting closely with humans.
This is also our motivation to choose human as the key object in our framework.
3 Model and Representation

In our approach, a dynamic MRF (Figure 2) is designed to integrate the relationship
between the object state and its observation, the spatial relationships between objects,
as well as the temporal relationships between the states of one object in successive
frames. The MRF has the structure of an undirected graph G, with a set of nodes V and
a set of edges E. Each node v ∈ V is associated with an unobserved state variable xv
and a observation yv . Since we are considering a temporal sequence, each node belongs
to exactly one time frame t.
t 1 t
\u t 1 ,u t
( xut 1 , xut )
xut1 xut \ u ,v ( xu , xv )
t t t t
xvt1 xvt
xwt1 xwt
\ v ( xv , yv )
t t t
yut1 yut
yvt 1 yvt
ywt 1 ywt
Fig. 2. The MRF defined in our problem (left) and an ideal graph structure for one input frame
(right). Section 5 explains how to build such a graph.
We use a node vt to represent a hypothetical object instance in frame t. Define

xvt = (cvt , pvt , svt ) (1)
as the state of the object, where cvt stands for the object’s category label, pvt for co-
ordinates of its centroid and svt for the logarithm of size1 . yvt is defined as the image
evidence of the object. There are two types of edges: intra-frame edges that represent
the spatial relationships between different objects, and inter-frame edges that represent
the temporal relationships between states of the same object in adjacent frames. Let the
potential functions be pairwise, in which case the distribution of the MRF factorizes as
1
p(x, y) = ψv,u (xv , xu ) ψv (xv , yv ), (2)
Z
(v,u)∈E v∈V
where x = {xv |v ∈ V} and y = {yv |v ∈ V}, ψv,u (xv , xu ) models the spatio-temporal
relationship, and ψv (xv , yv ) models the image observation likelihood. Given the struc-
ture of the MRF and the potential functions, the states of the objects can be inferred.
Note that rather than letting each node correspond to an image blob in a pre-defined
image grid or a pixel, as is commonly done in segmentation literature [12][10], we let
each node represent an object, which is similar to some tracking frameworks such as
the Markov chain in Particle Filtering and MRF in collaborative tracking proposed by
[16]. The reason is twofold: 1) object-based graph enables us to use object-level infor-
mation during inference, while pixel- or grid-based graph can only model inter-object
relationships locally along the boundary of objects; 2) object-based graph has fewer
nodes and therefore the complexity of inference is much lower. However, one draw-
back of object-based graph is that accurate segmentation cannot be directly obtained.
One new property of the object-based graph is that its structure is dynamic. In Section
5 we show that the nodes for new objects can be added to the graph online driven by
detected key objects. Before that we first give our models for the potential functions.
4 Potential Functions
There are three types of edges in our model, each associated with one kind of potential
function representing a specific semantic meaning.
1
Logarithm is used because scale is multiplicative.
4.1 Observation Potential ψv (xv , yv )

We use two sources of observation evidence. The first is a single-category object de-
tector for the key objects. For meeting room applications, we implement a patch-based
cascade human upper body detector following the method in [17]. Let c∗ stand for
the category label of key objects, for each xv = (cv , pv , sv ) with cv = c∗ , we de-
fine the observation potential to be the likelihood output of the detector: ψv (xv , yv ) =
p(c∗ |xv , yv ) = p(human|pv , sv ). Please refer to [18] for deriving probability of an
object class from a boosted classifier.
The second source of observation potential function targets all object categories of
interest. We build our object classifier based on the Bag of Features [1] approach and
combine it with image region. The motivation of our choice is the proven performance
of Bag of Feature and the suggestion in recent literature that classification based on
image segments provides better spatial coherence [2][11][14]. These ideas are tailored
to our needs. Specifically, interest points are detected with the DoG and Harris corner
detectors and at each interest point a 128d SIFT feature is extracted. During training,
these features are used to build a code book by clustering. Also every input image is
over-segmented by Mean Shift [19]; each segment is associated with interest points.
Base on both the point features and the statistics of pixel intensity of the segments, a
classifier is built to model p(c|ri ), defined as the likelihood of any given segment ri
belonging to category c. This could be done by standard Bag of Feature categorization,
or more sophisticated generative models such as [2]. We build a discriminative model
by using AdaBoost to select and weight features from the code book. Given p(c|ri ) for
any segment ri , the observation potential of object v is modeled as:

p(cv |ri )ζ(ri , xv )
ri ∈R(xv )
ψv (xv , yv ) = , (3)
ζ(ri , xv )
ri ∈R(xv )
where R(xv ) stands for the set of segments that is associated with the object v; ζ(ri , xv )
is a position weight for ri , which allows the use of object shape prior. In our implemen-
tation we let R(xv ) include all segments which has at least 50% of its area within v’s
bounding box, and ζ(ri , xv ) is defined as a Gaussian centered at pv . Figure 3 shows an
example for the category paper. We can see that it is hard to distinguish the paper from
a bright computer screen or the white board by appearance (feature point and region).
Fig. 3. An example of finding paper based on appearance. (a) Input image; (b) SIFT features
(green: feature with positive weight in the classifier, red: feature with negative weight); (c) Seg-
mentation; (d) observation likelihood p(paper |ri ) for each region ri (yellow: high likelihood).
Note that the observation potential here can be substituted by any object recognition
method, possibly with a more complicated model and higher accuracy such as [2]. Here
we do not elaborate on this since our emphasis is on the effect of introducing contextual
relationship.
4.2 Spatial Potential ψvt ,ut (xvt , xut )

Spatial potential function ψvt ,ut is defined on edges between nodes within one frame
but of different object categories. The purpose is to model inter-category correlation in
terms of position and scale, e.g., a person tends to sit on a chair beside a table and a
laptop is often near a person and on a table. Such correlation generalizes well in our
experience for the selected scenario.
When defining the form of the potential function, we want to avoid using very com-
plicated models which introduce risk of over-fitting. In practice, we find that a single
Gaussian function is sufficient for our amount of training data as well as the problem
itself. Denote N (μ, σ, x) as a Gaussian function with mean μ, variance σ and x as the
variable. Since nodes involved are from the same time frame, we suppress the subscript
t in this subsection. Define
ψv,u (xv , xu ) = N (μp (cu , cv ), σp (cu , cv ), pv − pu ) ·
N (μs (cu , cv ), σs (cu , cv ), sv − su ), (4)
where μp (cu , cv ), σp (cu , cv ), μs (cu , cv ) and σs (cu , cv ) are the model parameters that
describes the relative position and size of two object depending on their category labels
cu and cv . It is ideal to learn them by maximizing the sum of log likelihoods of all
training samples {x(i) }. However, this is difficult because x(i) s of different training
samples may have different dimensionalities (number of objects varies) and the graph
structures also differ. Therefore potential functions are learned independently for each
kind of edge in a piecewise manner [20]. The number of different spatial potential
functions is n(n−1)/2 for n categories. The parameters of the spatial potential function
between the categories c1 and c2 can be easily learned by maximizing

l= log ψv,u (x(j)
v , xu ),
(j)
(5)
j
(j) (j)
where {(xv , xu )} is the set of all pairs of objects that co-exist in a training sample
and satisfy cv = c1 and cu = c2 .
4.3 Temporal Potential ψvt−1 ,vt (xvt−1 , xvt )

To build the temporal potential function, feature points used in Section 4.1 are tracked
by optical flow through frames. Let the positions of feature points associated with object
(i)
v be {qt }m i=1 at frame t, xvt can be estimated from xvt −1 as:
1 (i)
m
(i)
Position : p̃vt = pvt−1 + (q − qt−1 ), (6)
m i=1 t
9 :
m (i)
i=1 Dist (qt , p̃vt )
Scale : s̃vt = svt−1 + log m (i)
, (7)
i=1 Dist(qt−1 , pvt−1 )
where Dist(·) is the distance between two points. The temporal potential is defined as
a Gaussian distribution centered at the estimated position and scale with fixed variance:
ψvt−1 ,vt (xvt−1 , xvt ) = N (p̃vt , σp , pvt )N (s̃vt , σs , svt ). (8)
5 Integration of Observation and Spatio-temporal Context

Given the graphical model defined above, there are two remaining issues in using it:
how to build such a graph on-the-fly and how to do inference. We solve them in a
unified manner by belief propagation (BP) [21][22]. Augmenting nodes are introduced
as nodes that do not correspond to any specific object but are responsible for generating
new object nodes by receiving belief messages from nodes of key objects. To distinguish
augmenting nodes from the others, we refer to other nodes as object nodes. BP is then
applied to compute the distribution p(xv |y) for all object nodes, from which the state of
every object can be estimated. Since message passing in BP is essential for augmenting
nodes, we first bring up the inference part and then introduce the augmenting nodes.
5.1 Inference
We choose BP as the inference method because of two main reasons. First, our graph has
cycles and the structure is not fixed (due to addition and removal of object nodes, also
inference is not done over the whole sequence but over a sliding window). Therefore it is
inconvenient to use methods that require rebuilding the graph (such as the junction tree
algorithm). While loopy BP is not guaranteed to converge to the true marginal, it has
proven excellent empirical performance. Second, BP is based on local message passing
and update, which is efficient and more importantly, gives us an explicit representation
of the interrelationship between nodes (especially useful for the augmenting nodes).
At each iteration of BP, the message passing and update process is as follows. Define
the neighborhood of a node u ∈ V as Γ (u) = {v|(u, v) ∈ E}, each node u send a
message to its neighbor v ∈ Γ (u):

mu,v (xv ) = α ψu,v (xu , xv )ψu (xu , yu ) mw,u (xu )dxu . (9)
xu w∈Γ (u)\v
The marginal distribution of each object v is estimated by

p(xv |y) = αψv (xv , yv ) mu,v (xv ). (10)
u∈Γ (v)
In our problem xv is a continuous variable whose distribution is non-Gaussian and

hard to represent in an analytical form; also, the observation potential function can
only be evaluated in a point-wise manner. Therefore we resort to nonparametric
version of the BP algorithm [22]. Messages are represented by a nonparametric kernel
density estimate. More details of this method can be found in [22]. As a result, a
weighted sample set is obtained to approximate the marginal distribution of each
(i) (i)
object node v: {xv , ωv }M i=1 ∼ p(xv |y). The sample set is generated by importance
(i) (i) (i)

sampling; namely sample {xv } ∼ u∈Γ (v) mu,v (xv ) and let ωv = ψv (xv , yv ).
We can then estimate the state of object v (except its category label) by MMSE: x̂v =
M (i) (i) M (i)
i=1 ωv xv / i=1 ωv .
5.2 Augmenting Nodes
Augmenting nodes find new objects by receiving “hints” (messages) from key object
nodes. It is reasonable because we are more interested in finding objects that are closely
related to key objects; by combining inter-category spatial relationships with detection
techniques specially developed for key objects, other objects can be detected and recog-
nized more robustly and efficiently.
Let the set of key objects in one frame be K, and consider finding new objects of
category c = c∗ . The ideal way is: for every subset K of K, make the hypothesis
that there is a new object a which is in context with K . Based on the NBP paradigm,
we estimate a’s state by p(xa |y) ∝ ψu (xa , ya ) v∈K mv,a (xa ). The number of such
hypotheses is exponential of |K|, so we simplify it by letting K contain only one key
object (it is reasonable because if a new object is hinted by a set of key objects it is at
least hinted by one in some extent). In this case K = {v}, the distribution of a’s state is
estimated as p(xa |y) ∝ ψa (xa , ya )mv,a (xa ). This is done for each v in K, each result
in a weighted sample set of a hypothetic new object’s state.
Further, if two hypotheses of the same category are close in position and scale,
they should be the same new object. So for each category, Agglomerative Clustering
is done on the union of the |K| sample sets to avoid creating duplicated nodes. For each
Fig. 4. Use of augmenting nodes to update graph structure. Augmenting nodes for each category
are shown as one (dotted circle). For weighted samples, red indicates the highest possible weight,
while blue indicates the lowest.
Table 1. Algorithm: inference over a sequence
Denote by Vt and Et the sets of nodes and edges in frame t respectively.

With the graph G over a L-frame sliding window containing frame (t − L) to (t − 1), proceed as follows
with the arrival of a new frame t:
– Output the estimated state x̂v for each object node v of frame (t − L). Remove sub-graph
(Vt−L , Et−L ) from G and move the sliding window one frame forward.
– Add new sub-graph (Vt , Et ) for frame t to G by algorithm in Table 2.
– Inference: perform the nonparametric BP algorithm over G. For each object node v a weighted sample
(i) (i)
set is obtained: {xv , ωv }Mi=1 ∼ p(xv |y). t M (i)
– Evaluate confidence of each object v by W = j=t−L+1
ω . If W < γ, remove node vj
i=1 vj
from frame j for each j = (t − L + 1) . . . t. γ is an empirical threshold.
Table 2. Algorithm: build the sub-graph for a new frame t
Build the sub-graph (Vt , Et ) for a new frame t as follows:
– For each object node vt−1 ∈ Vt−1 , let Vt ← Vt ∪ {vt }, E ← E ∪ {(vt−1 , vt )}. Pass message
forward along edge (vt−1 , vt ) to get an approximation of p(xvt |y) ∝ ψvt (xvt , yvt )mvt−1 ,vt (xvt ).
– Detect key object by applying p(c∗ |x) to all possible state x in the image. Cluster responses with
confidence higher than τc∗ . For each cluster non-overlapping with any existing node, create a new node
vt . Let the initial estimated state x̂vt be the cluster mean. Denote the set of all key object node as K.
– For each category c = c∗ :
- Create an augmenting node a for each key object node v ∈ K and add an edge (v, a) between
them.
(i) (i)
- For each such augmenting node and key object node pair {a, v}, sample {xa , ωa }M i=1 ∼
p(xa |y) ∝ ψa (xa , ya )mv,a (xa ).
(i) (i)
- Define the union of samples S = a {xa , ωa }M i=1 ; let S be the subset of S with samples
whose weight are higher than τc .
- Do clustering on S ; for each cluster non-overlapping with any existing node, create an object node
ut of category c. Let the initial estimated state x̂ut be the cluster mean.
- Vt ← Vt ∪ {ut }. Et ← Et ∪ {(ut , vt )|vt ∈ Vt , ψut ,vt (x̂ut , x̂vt ) > λ}.
- Remove augmenting nodes and corresponding edges.
high-weight cluster, a new object node is created. Figure 4 illustrates how to use aug-
menting nodes to update the graph.
More details of our overall algorithm and the algorithm of building sub-graph for
each new frame are shown in Table 1 and Table 2.
6 Experiments
Experiments are done on the CHIL meeting video corpus [23]. Eight categories of ob-
jects are of interest: human, table, chair, computer, projector, paper, cup and whiteboard
(or projection screen).
For testing we use 16 videos captured from three sites (IBM, AIT and UPC) and three
camera views for each site. Each sequence has about 400 frames. One frame out of every
60 is fully annotated for evaluation. For training the parameters of spatial potential func-
tion, we selected 200 images from two views of the IBM and UPC site (no intersection
between training images and test videos), and manually annotated the object size and
position. Observation models for objects are trained with object instances from images
of various meeting room and office scenes including a training part from CHIL.
We design our experiments to compare three methods with different levels of context:
1) no context, i.e. object observation model is directly applied to each frame; 2) spatial
context only, i.e. a MRF without the temporal edges is applied in a frame-by-frame
manner; 3) spatio-temporal context, i.e. the full model with both spatial and temporal
edges is applied to the sequence.
6.1 Quantitative Analysis
Quantitative analysis is performed with metrics focusing on three different aspects: ob-
ject detection and tracking, image segment categorization and pixel-level segmentation
accuracy.
Object-level detection and tracking. The overall object detection rate and false alarm
rate is shown in Figure 6(left). Two methods are compared: frame-based method with
only spatial context, and the spatio-temporal method using the complete model. For the
spatial-only method, an ROC curve is obtained by changing the threshold τc for creating
new object nodes. The result shows that integrating temporal information helps improve
detection rate and reduce false alarms, which is the effect of temporal smoothing and
evidence accumulation. In object-level evaluation we do not include the non-contextual
method, because the object observation model is based on classifying image segments,
and we find that applying exhaustive search using such a model does not give a mean-
ingful result. Some visual results of these methods can be found in Figure 5(a)(c)(d)
respectively.
Fig. 5. Comparison among observation with no context, inference using spatial relationship only
and inference using spatio-temporal relationship
0.8 0.8
No context
Pixel−level Segmentation Recall

0.7 0.7 Frame−based (spatial only)
Correct Detection No. / GT No.
Spatial−temoral
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
Frame−based (spatial only)
0.1 Spatial−temoral 0.1
0 0
0 0.1 0.2 0.3 0.4 0.2 0.4 0.6 0.8 1
False Alarm No. / GT No. Pixel−level Segmentation Precision
Fig. 6. Object detection rate and false alarm rate (left); pixel-level segmentation precision and
recall (right)
Table 3. Object tracking evaluation
Ground truth Mostly tracked Partially tracked

Category Fragments
trajectories trajectories (%GT) trajectories (%GT)
Human 64 46 (71.9%) 12 (18.8%) 12
Chair 30 10 (33.3%) 4 (13.3%) 2
Paper 40 21 (52.5%) 7 (17.5%) 0
Cup 11 2 (18.2%) 0 (0%) 0
Computer 24 10 (41.7%) 3 (12.5%) 2
Table 16 14 (87.5%) 0 (0%) 0
Screen 14 12 (85.7%) 1 (7.1%) 2
Projector 13 7 (53.8%) 2 (15.4%) 0
All 212 122 (57.5%) 29 (13.7%) 8
For the spatio-temporal method, we further evaluate its performance by the number
of objects that are consistently tracked through the sequence, as shown in Table 3. All
the numbers stand for trajectories, where mostly tracked is defined as at least 80% of the
trajectory is tracked, and partially tracked defined as at least 50% is tracked. When a
trajectory is broken into two, a fragment is counted. We can see that small objects such
as cups and computers are harder to detect and track. Paper has a high false alarm rate,
probably due to lack of distinct interior features (Figure 8(h) shows segments of hu-
man clothes detected as paper). Most fragments belong to human trajectories, because
humans exhibit much more motion than other objects.
Image segment-level categorization. To compare with the result of applying object

observation without contextual information, we compute the categorization accuracy
of all the image segments in the form of a confusion matrix (Figure 7). The matrix
shows that incorporating context helps reduce the confusion between different object
categories, such as paper versus whiteboard. It is also observed that many objects are
easily confused with table, mainly because they are often on top of or adjacent to the
table.
Pixel-level segmentation. We obtain segmentation of each object based on the like-

lihood p(c|ri ) of each segment ri classified as category c. Pixel-level precision and
recall rates of the three methods are shown in Figure 6(right). Similar to the previous
two evaluations, the spatio-temporal method gives the best result. The segmentation
und
und
und
a rd
ard
ard
r
r
pute
pute
pute
Bac or
tebo
kgro
kgro
kgro
tebo
tebo
ecto
ecto
an
an
an
ect
er
le
ir
er
le
er
le
ir
ir
Hum
Com
Hum
Hum
Com
Com
Whi
Cup
Proj
Whi
Whi
Cha
Cup
Proj
Cup
Proj
Ba c
Tab
Cha
Bac
Cha
Pa p
Tab
Tab
Pa p
Pa p
Human .38 .04 .50 .41 .07 .49 .51 .45
Chair .10 .21 .10 .57 .30 .09 .59 .30 .05 .64
Paper .12 .16 .03 .24 .12 .06 .23 .08 .21 .43 .03 .21 .04 .40 .19 .35
Cup .14 .10 .04 .03 .18 .19 .28 .09 .03 .38 .46 .17 .38 .40
Computer .15 .04 .31 .17 .30 .04 .27 .30 .38 .03 .44 .20 .40
Table .08 .03 .61 .24 .03 .03 .62 .27 .63 .32
Whiteboard .04 .03 .40 .51 .44 .56 .48 .51
Projector .03 .06 .04 .04 .35 .10 .36 .53 .18 .25 .06 .24 .53 .17
Background .08 .85 .03 .93 .93
(a) Without context (b) Spatial-only (c) Spatio-temporal
Fig. 7. Confusion matrix of image region categorization by different methods. The value at (i, j)
stands for the proportion of segments of category i classified as category j.
Fig. 8. Sample results of our method

accuracy is not high, since this is only a simple post-process. But from the sample re-
sults of the spatio-temporal method in Figure 8 we can see that most detected objects
are reasonably segmented when the object position and scale are correctly inferred.
6.2 Scenario Analysis

Figure 8 shows some sample results of our method on data from different meeting
room sites and views. Objects that are in close interaction with key objects (humans)
are detected more accurately. The method also has a tolerance to missed detection of
key objects, e.g., for the IBM site, although human detection rate is not high due to
complex background, most objects are reasonably detected (Figure 8(a)-(c)). However
such tolerance is to certain extent: Figure 8(m) shows a case when missed detections of
key objects cause failure in detecting other objects.
Partial occlusions are frequently encountered and handled, such as occlusions of
tables, whiteboards and laptops. But there is a bigger chance of failure when only a
small part of an object is visible, such as in Figure 8(a)-(d) the table is broken into two;
in Figure 8(n)(o) the table or part of it is missing from detection. This is also true for
small objects, e.g. in Figure 8(g)(h) the paper occluded by the hand is broken into two.
But in such case the result is still correct in the image segment level.
The bottleneck of performance is the observation model for objects other than key
objects. As in Figure 8(p) the computer and projector are missing simply because obser-
vation likelihood is low. Although contextual information improves the overall result,
the observation model in our current implementation is relatively simple compared with
the complexity of the object recognition problem.
7 Conclusion
In this paper we address the problem of recognizing, localizing and tracking multiple
categories of objects in a certain type of scenes. Specifically, we consider eight cate-
gories of common objects in meeting room videos. Given the difficulty of approaching
this problem by purely appearance-based methods, we propose the integration of spatio-
temporal context through a dynamic MRF, in which each node represents an object and
the edges represent inter-object relationships. New object hypotheses are proposed on-
line by adding augmenting nodes, which receive belief messages from the detected key
objects of the scene (humans in our case). Experimental results show that the perfor-
mance is greatly enhanced by incorporating contextual information.
There are many open problems and promising directions regarding the topic of ob-
ject analysis in video. First, a stronger object observation model is needed, and our
current training and testing sets are very limited. Second, we made no assumption of a
fixed camera, but it can be a strong cue for inference, e.g. the position and scale of the
stationary objects (such as tables) can be inferred from the activity area of the moving
objects (such as humans). Third, 3D geometry of the scene or depth information should
be useful for modeling occlusions. Last but not least, object recognition and tracking
can be combined with action recognition [14][15] so as to better understand the seman-
tics of human activities.
Acknowledgments. This research is supported, in part, by the U.S. Government VACE

program. Yuan Li is funded, in part, by a Provost’s Fellowship from USC.
References
1. Fergus, R., Perona, P., Zisserman, A.: Object class recognition by unsupervised scale-
invariant learning. In: CVPR (2003)
2. Cao, L., Fei-Fei, L.: Spatially coherent latent topic model for concurrent object segmentation
and classification. In: ICCV (2007)
3. Viola, P., Jones, M.: Rapid object detection using a boosted cascade of simple features. In:
CVPR (2001)
4. Dalal, N., Triggs, B.: Histograms of oriented gradients for human detection. In: CVPR (2005)
5. Wu, B., Nevatia, R.: Cluster boosted tree classifier for multi-view, multi-pose object detec-
tion. In: ICCV (2007)
6. Sudderth, E.B., Torralba, A., Freeman, W.T., Willsky, A.S.: Learning hierarchical models of
scenes, objects, and parts. In: ICCV (2005)
7. Hoiem, D., Efros, A.A., Hebert, M.: Putting objects in perspective. In: CVPR (2006)
8. Torralba, A., Murphy, K., Freeman, W., Rubin, M.: Context-based vision system for place
and object recognition. In: ICCV (2003)
9. Li, L.-J., Fei-Fei, L.: What, where and who? classifying events by scene and object recogni-
10. Shotton, J., Winn, J., Rother, C., Criminisi, A.: Textonboost: Joint appearance, shape and
context modeling for multi-class object recognition and segmentation. In: ECCV (2006)
11. Rabinovich, A., Vedaldi, A., Galleguillos, C., Wiewiora, E., Belongie, S.: Objects in context.
In: ICCV (2007)
12. Carbonetto, P., de Freitas, N., Barnard, K.: A statistical model for general contextual object
recognition. In: Pajdla, T., Matas, J(G.) (eds.) ECCV 2004. LNCS, vol. 3021, pp. 350–362.
13. Moore, D.J., Essa, I.A., Heyes, M.H.: Exploiting human actions and object context for recog-
nition tasks. In: ICCV (1999)
14. Peursum, P., West, G., Venkatesh, S.: Combining image regions and human activity for indi-
rect object recognition in indoor wide-angle views. In: ICCV (2005)
15. Gupta, A., Davis, L.S.: Objects in action: an approach for combining action understanding
and object perception. In: CVPR (2007)
16. Yu, T., Wu, Y.: Collaborative tracking of multiple targets. In: CVPR (2004)
17. Wu, B., Nevatia, R.: Tracking of multiple humans in meetings. In: V4HCI (2006)
18. Friedman, J., Hastie, T., Tibshirani, R.: Additive logistic regression: a statistical view of
boosting. Annals of Statistics 28(2), 337–407 (2000)
19. Comaniciu, D., Meer, P.: Mean shift: A robust approach toward feature space analysis. IEEE
Transaction on Pattern Analysis and Machine Intelligence 24(5), 603–619 (2002)
20. Sutton, C., McCallum, A.: Piecewise training for undirected models. In: Conference on Un-
certainty in Artificial Intelligence (2005)
21. Pearl, J.: Probabilistic Reasoning in Intelligent Systems. Morgan Kaufman, San Mateo
(1988)
22. Sudderth, E.B., Ihler, A.T., Freeman, W.T., Willsky, A.S.: Nonparametric belief propagation.
In: CVPR (2003)
23. CHIL: The chil project, http://chil.server.de/
A Pose-Invariant Descriptor for Human
Detection and Segmentation
Zhe Lin and Larry S. Davis
Institute of Advanced Computer Studies

University of Maryland, College Park, MD 20742
{zhelin,lsd}@umiacs.umd.edu
Abstract. We present a learning-based, sliding window-style approach

for the problem of detecting humans in still images. Instead of tra-
ditional concatenation-style image location-based feature encoding, a
global descriptor more invariant to pose variation is introduced. Specif-
ically, we propose a principled approach to learning and classifying
human/non-human image patterns by simultaneously segmenting human
shapes and poses, and extracting articulation-insensitive features. The
shapes and poses are segmented by an efficient, probabilistic hierarchi-
cal part-template matching algorithm, and the features are collected in
the context of poses by tracing around the estimated shape boundaries.
Histograms of oriented gradients are used as a source of low-level fea-
tures from which our pose-invariant descriptors are computed, and kernel
SVMs are adopted as the test classifiers. We evaluate our detection and
segmentation approach on two public pedestrian datasets.
1 Introduction
Human detection is a widely-studied problem in vision. It still remains challeng-
ing due to highly articulated body postures, viewpoint changes, varying illumi-
nation conditions, and background clutter. Combinations of these factors result
in large variability of human shapes and appearances in images. We present
an articulation-insensitive feature extraction method and apply it to machine
learning-based human detection. Our research goal is to robustly and efficiently
detect and segment humans under varying poses.
Numerous approaches have been developed for human detection in still im-
ages or videos. Most of them use shape information as the main discriminative
cue. These approaches can be roughly classified into two categories. The first
category models human shapes globally or densely over image locations, e.g.
shape template hierarchy in [1], an over-complete set of haar wavelet features
in [2], rectangular features in [3], histograms of oriented gradients (HOG) in [4]
or locally deformable Markov models in [5]. Global schemes such as [4, 6] are
designed to tolerate certain degrees of occlusions and shape articulations with a
large number of samples and have been demonstrated to achieve excellent per-
formance with well-aligned, more-or-less fully visible training data. The second
category of approaches uses local feature-based approaches to learn body part

424 Z. Lin and L.S. Davis
and/or full-body detectors based on sparse interest points and descriptors as

in [7, 8], from predefined pools of local curve segments [9, 10], k-adjacent seg-
ments [11], or edgelets [12]. In [13], several part detectors are trained separately
for each body part, and combined with a second-level classifier. Compared to
the global schemes, part (or local feature)-based approaches [12, 8, 14] are more
adept in handling partial occlusions, and flexible in dealing with shape artic-
ulations. Shape cues are also combined with motion cues for human detection
in [15, 16], simultaneous detection and segmentation in [17].
Dalal and Triggs [4] introduced a powerful image descriptor - HOG, and
provided an extensive experimental evaluation using linear and gaussian-kernel
SVMs as the test classifiers. Later, Zhu et al. [18] improved its computational
efficiency significantly by utilizing a boosted cascade of rejectors. Recently, Tuzel
et al. [6] reported better detection performance than [4] on the INRIA dataset.
They use covariant matrices as image descriptors and classify patterns on Rie-
mannian manifolds. Similarly, Maji et al. [19] also demonstrate promising results
using multi-level HOG descriptors and faster (histogram intersection) kernel
SVM classification. In [20], two-fold adaboost classifiers are adopted for simul-
taneous part selection and pedestrian classification. Ref. [21] combines different
features in a single classification framework.
Previous discriminative learning-based approaches mostly train a binary clas-
sifier on a large number of positive and negative samples where humans are
roughly center-aligned. These approaches represent appearances by concatenat-
ing information along 2D image coordinates for capturing spatially recurring
local shape events in training data. However, due to highly articulated human
poses and varying viewing angles, a very large number of (well-aligned) training
samples are required; moreover, the inclusion of information from whole images
inevitably make them sensitive to biases in training data (in the worst case, sig-
nificant negative effects can occur from arbitrary image regions), consequently
the generalization capability of the trained classifier can be compromised. Mo-
tivated by these limitations, we extract features adaptively in the local context
of poses, ı.e. we propose a pose-invariant feature extraction method for simul-
taneous human detection and segmentation. The intuition is that pose-adapted
features produce much better spatial repeatability and recurrence of local shape
events. Specifically, we segment human poses on both positive and negative sam-
ples1 and extract features adaptively in local neighborhoods of pose contours,
i.e. in the pose context. The set of all possible pose instances are mapped to a
canonical pose such that points on an arbitrary pose contour have one-to-one
correspondences to points in the canonical pose. This ensures that our extracted
feature descriptors correspond well to each other, and also invariant to varying
poses. Our main contributions are summarized as follows:
– An extended part-template tree model and an automatic learning algorithm

are introduced for simultaneous human detection and pose segmentation.
1
For negative samples, pose estimation is forced to proceed even though no person in
them.
A Pose-Invariant Descriptor for Human Detection and Segmentation 425
– A fast hierarchical part-template matching algorithm is used to estimate

human shapes and poses based on both gradient magnitude and orientation
matching. Human shapes and poses are represented by parametric models,
and the estimation problem is formulated and optimized in a probabilistic
framework.
– Estimated optimal poses are used to impose spatial priors (for possible hu-
mans) to encode pose-invariant features in nearby local pose contexts. One-
to-one correspondence is established between sets of contour points of an
arbitrary pose and a canonical pose.
The paper is organized as follows. Section 2 gives an overview; Section 3 de-
scribes the details of our pose-invariant feature extraction method; Section 4 in-
troduces our learning and classification schemes and demonstrates experiments
and evaluations; finally, Section 5 concludes the paper and discusses possible
future extensions.
2 Overview of the Approach

We illustrate and evaluate our approach mainly using the INRIA person
dataset2 [4] and the MIT-CBCL pedestrian dataset3 [2, 13]. In these datasets,
training and testing samples all consist of 128 × 64 image patches. Negative
samples are randomly selected from raw (person-free) images, positive samples
are cropped (from annotated images) such that persons are roughly aligned in
location and scale.
For each training or testing sample, we first compute a set of histograms
of (gradient magnitude-weighted) edge orientations for non-overlapping 8 × 8
rectangular regions (or cells) evenly distributed over images. Motivated by the
success of HOG descriptors [4] for object detection, we employ coarse-spatial
and fine-orientation quantization to encode the histograms, and normalization is
performed on groups of locally connected cells, i.e. blocks. Then, given the orien-
tation histograms, a probabilistic hierarchical part-template matching technique
is used to estimate shapes and poses based on an efficient part-based synthesis
approach under a probabilistic framework. A fast k-fold greedy search algorithm
is used for the likelihood optimization. The part-template tree model in [14]
used for the hierarchical matching is learned from a set of annotated silhouette
images. Given the pose and shape estimates, block features closest to each pose
contour point are collected; finally, the histograms of the collected blocks are
concatenated in the order of pose correspondence to form our feature descriptor.
As in [4], each block (consisting of 4 histograms) is normalized before collect-
ing features to reduce sensitivity to illumination changes. The one-to-one point
correspondence from an arbitrary pose model to the canonical one reduces sensi-
tivity of extracted descriptors to pose variations. Figure 1 shows an illustration
of our feature extraction process.
2
http://lear.inrialpes.fr/data
3
http://cbcl.mit.edu/software-datasets/PedestrianData.html
(a) (b) (c) (d) (e) (f)

Fig. 1. Overview of our feature extraction method. a) A training or testing image, b)
Part-template detections, c) Pose and shape segmentation, d) Cells overlaid onto pose
contours, e) Orientation histograms and cells overlapping with the pose boundary, f)
Block centers relevant to the descriptor.
3 Pose-Invariant Descriptors
3.1 Low-Level Feature Representation
For pedestrian detection, histograms of oriented gradients (HOG) [4] exhib-
ited superior performance in separating image patches into human/non-human.
These descriptors ignore spatial information locally, hence are very robust to
small alignment errors. We use a very similar representation as our low-level fea-
ture description, i.e. (gradient magnitude-weighted) edge orientation histograms.
Given an input image I, we calculate gradient magnitudes |GI | and edge
orientations OI using simple difference operators (−1, 0, 1) and (−1, 0, 1)t in
horizontal-x and vertical-y directions, respectively. We quantize the image re-
gion into local 8 × 8 non-overlapping cells, each represented by a histogram
of (unsigned) edge orientations (each surrounding pixel contributes a gradient
magnitude-weighted vote to the histogram bins). Edge orientations are quan-
tized into Nb = 9 orientation bins [k Nπb , (k + 1) Nπb ), where k = 0, 1...Nb − 1.
For reducing aliasing and discontinuity effects, we also use trilinear interpola-
tion as in [4] to vote for the gradient magnitudes in both spatial and orientation
dimensions. Additionally, each set of neighboring 2 × 2 cells form a block. This
results in overlapping blocks where each cell is contained in multiple blocks. For
reducing illumination sensitivity, we normalize the group of histograms in each
block using L2 normalization with a small regularization constant to avoid
dividing-by-zero. Figure 2 shows example visualizations of our low-level HOG
descriptors.
The above computation results in our low-level feature representation consist-
ing of a set of raw (cell) histograms (gradient magnitude-weighted) and a set of
normalized block descriptors indexed by image locations. As will be explained in
the following, both unnormalized cell histograms and block descriptors are used
for inferring poses and computing final features for detection.
3.2 Part-Template Tree Model

For highly articulated objects like humans, part-based detection approaches
(e.g. [7, 12]) have been shown to be capable of handling partial object/
Fig. 2. Examples of two training samples and visualization of corresponding (un-

normalized and L2 -normalized) edge orientation histograms
inter-occlusions and are flexible in modeling shape articulations. In contrast,

global shape template-based approaches are capable of simultaneously detect-
ing and segmenting human shapes, e.g. [1] and its generalization [22] using a
Bayesian inference. The merits of these two schemes are combined in a uni-
fied (top-down and bottom-up) optimization framework for simultaneous detec-
tion and segmentation in [14]. Specifically, it extends the hierarchical template
matching method in [1] by decomposing the global shape models into parts and
constructing a part-template tree for matching it to images hierarchically.
In order to more efficiently and reliably estimate human shapes and poses in
the image, we learn the part-template tree model [14] and extend the matching
algorithm in a probabilistic optimization framework. We train the part-template
tree on a set of annotated silhouette images to learn the distribution of part
models in each of the tree layers and to handle a wider range of articulations
of people. The structure of our learned part-template tree model is roughly
shown in Figure 3. The part-template tree was initially constructed using a
simple pose generator and body-part decomposer. Each part in the tree can
be viewed as a parametric model, where part location and sizes are the model
Fig. 3. An illustration of the extended part-template tree

parameters. As shown in the figure, the tree consists of 186 part-templates, i.e.
6 head-torso (ht) models, 18 upper-leg (ul) models, 162 lower-leg (ll) models,
and organized hierarchically based on the layout of human body parts in a top-
to-bottom manner. Due to the tree structure, a fast hierarchical shape (or pose)
matching scheme can be applied using the model. For example, using hierarchical
part-template matching (which will be explained later), we only need to match
24 part-templates to account for the complexity of matching 486 global shape
models using the method in [1], so it is extremely fast. For the details of the tree
model construction method, readers are referred to [14].
Learning the Part-template Tree. We learn the part-template tree model

based on a training set which consists of 404 (plus mirrored versions) manually
annotated binary silhouette images (white foreground and black background).
Those silhouettes are chosen from a subset of positive image patches of the IN-
RIA person database. Each of the training silhouette images is sent through the
tree from the root node to leaf nodes and the degree of coverage (both foreground
and background) consistency between each part template Tθj , j ∈ {ht, ul, ll} and
the observation is measured. Here, each part-template is considered to be covered
by a binary rectangular image patch M (see Figure 1(b) for an example). The
degree of coverage consistency ρ(θj |S) between a part-template Tθj and a silhou-
ette image S is defined as the pixel-wise similarity of the part-template coverage
image M (θj ) and the binary sub-silhouette S (corresponding to the same region
j
|S (x)−M(θj ,x)|
as the part-template), i.e. ρ(θj |S) = 1 − x i n , where n is the total
number of pixels in the rectangular part-template region. Then, we can estimate
the best set of part models θ∗ = {θj∗ } for the training silhouette S by maximum
likelihood estimation: θj∗ = arg maxθj ∈Θj ρ(θj |S), where Θj denotes the set of
all possible part template parameters. This process is repeated for all training
silhouettes and the ranges of part template models are estimated based on the
statistics of each part-template’s model parameters. The ranges of parameters
are evenly quantized to produce the final tree model (Figure 3). We also verified
that the average image of our learned global shape models (composition of parts)
is very similar to the mean of all training silhouettes (see Figure 4).
3.3 Probabilistic Hierarchical Part-Template Matching

We formulate the pose and shape estimation problem probabilistically as max-
imization of a global object likelihood L. In order to quickly evaluate the like-
lihood for a global pose (i.e. different parameter combinations of part models),
the object likelihood is simply modeled as a summation of matching scores of
part-template models in all tree layers. We can think of L as a log-likelihood and
the summation of the matching scores over different parts is equivalent to mul-
tiplication of probabilities. Given an image I (either training or testing sample)
and a candidate global pose model θ = {θj } (including part-template indices
and their locations and scales), in the simplest case, if we assume independence
between part-template models θj in different layers, the object likelihood can be
simply represented as follows:

L(θ|I) = L(θht , θul , θll |I) = L(θj |I). (1)
j∈{ht,ul,ll}
For the purpose of pose estimation, we should jointly consider different parts θj
for optimization of L. Hence, based on the layer structure of the tree in Figure 3,
the likelihood L is decomposed into conditional likelihoods as follows:
L(θ|I) = L(θht |I) + L(θul |θht , I) + L(θll |θht , θul , I)

= L(θht |I) + L(θul |θht , I) + L(θll |θul , I), (2)
where the decomposition is performed in a top-to-bottom order of the layers,

and independence is assumed between the two non-joining layers, ht and ll. We
used Eq. 2 as our optimization model.
Part-Template Likelihood. A part template Tθj (defined by model parame-

ters θj ) is characterized by its boundary curve segments (see Figure 3) and edge
orientations of points along the segment. We match individual part-templates
using a method similar to Chamfer matching [1]. Instead of using distance trans-
forms, we collect matching scores (magnitudes of corresponding orientation bins
in the map of edge orientation histograms) along the part-template contour.
The matching scores are measured using look-up tables for speed. Magnitudes
from neighboring histogram bins are weighted to reduce orientation biases and
to regularize the matching scores of each template point.
More formally, the likelihood L(θj (x, s)|I) of a part template-Tθj at location
x and scale s is modeled as follows:
1
L(θj (x, s)|I) = dI (x + st), (3)
|Tθj |
t∈Tθj
where |Tθj | denotes the length of the part-template, and t represents individual
contour points along the template. Suppose the edge orientation of contour point
t is O(t), its corresponding orientation bin index B(t) is computed as: B(t) =
[O(t)/(π/9)] ([x] denotes the maximum integer less-or-equal to x), and the un-
normalized (raw) orientation histogram at location (x + st) is H = {hi }. Then,

the individual matching score dI at contour point t is expressed as:

δ
dI (x + st) = w(b)hB(t)+b , (4)
b=−δ
where δ is a neighborhood range, and w(b) is a symmetric weight distribution4 .
Optimization. The structure of our part-template model and the form (sum-
mation) of the global object likelihood L suggest that the optimization problem
4
For simplicity, we use δ = 1, and w(1) = w(−1) = 0.25, w(0) = 0.5 in our
experiments.
can be solved by dynamic programming or belief propagation [23] to achieve glob-

ally optimal solutions. But, these algorithms are computationally too expensive
for dense scanning of all windows for detection. For efficiency, we perform the
optimization, i.e. the maximization of L, by a fast k-fold greedy search proce-
dure. Algorithm 1 illustrates the overall matching (optimization) process. We
keep scores for all nodes (k = 1, 2...K) in the second layer (i.e. the torso layer)
instead of estimating the best k in step 1 of the algorithm. In the following
steps, a greedy procedure is individually performed for each of those K nodes
(or threads).
Algorithm 1. Probabilistic Hierarchical Part-Template Matching

1) For a set of locations x and scales s, match all K head-torso part-
templates in layer L1 with the image and compute their part-template likeli-
hoods L(θhtk
(x, s)|I), k = 1, 2...K.
2) For k = 1...K, repeat the following steps (3)-(4), and select k = k ∗ and θ = θ∗
with the maximum L(θ|I).
3) According to the part-template model θht k
of Layer L1 , estimate the maxi-
∗
mum conditional-likelihood leg models θul |θht
k
in L2 and θll∗ |θul ∗
, θht
k
in L3 using
a greedy search algorithm along the tree.
4) Given the above part-template’s model estimates, compute the current global
object likelihood based on Eq. 2.
5) Return the global pose model estimates θ∗ = {θht k ∗
, θul , θll∗ }.
Pose model parameters estimated by the hierarchical part-template matching

algorithm are directly used for pose segmentation by part-synthesis (region con-
nection). Figure 4 shows the process of global pose (shape) segmentation by the
part-template synthesis.
Fig. 4. An illustration of pose segmentation. Top-Left: Mean image of training silhou-

ettes and our 486 learned global shape models; Top-Right: Best part-templates (three
images on the left side) in each layer are combined to produce final global shape and
pose estimates (two images on the right side); Bottom: example pose (shape) segmen-
tation on positive/negative samples.
3.4 Representation Using Pose-Invariant Descriptors
In our implementation, the global shape models (consisting of 3 part-template

types) are represented as a set of boundary points with corresponding edge ori-
entations. The range of the number of those model points are from 118 to 172. In
order to obtain a unified (constant dimensional) description of images with those
different dimensional pose models, and to establish a one-to-one correspondence
between contour points of different poses, we map the boundary points of any
pose model to those of a canonical pose model. The canonical pose model is
assumed to be occlusion-free, so that all contour points are visible. For human
upper bodies (heads and torso), the boundaries are uniformly sampled into 8 left
side and 8 right side locations; and the point correspondence is established be-
tween poses based on vertical y coordinates and side (left or right) information.
For lower bodies (legs), boundaries are uniformly sampled into 7 locations ver-
tically with 4 locations at each y value (inner leg sample points are sampled at
5 pixels apart from outer sample points in the horizontal direction). Figure 1(e)
shows an example of how the sampled locations are distributed).
Associated with each of those sample locations is a 36-dimensional feature
vector (L2 -normalized histogram of edge orientations of its closet 2 × 2 block in
the image). Hence, this mapping procedure generates a (8 × 2 + 7 × 4) × 36 =
1584 dimensional feature descriptor. Figure 1 illustrates the feature extraction
method. Note that only a subset of blocks are relevant for the descriptor, and
a block might be duplicated several times based on the frequency of contour
points lying inside the block.
4 Experiments
4.1 Datasets
We use both the INRIA person dataset and the MIT-CBCL pedestrian dataset
for detection and segmentation performance evaluation. The MIT-CBCL dataset
contains 924 front/back-view positive images (no negative images), and the INRIA
dataset contains 2416 positive training samples and 1218 negative training images
plus 1132 positive testing samples and 453 negative testing images. Comparing to
the MIT dataset, the INRIA dataset is much more challenging due to significant
pose articulations, occlusion, clutter, viewpoint and illumination changes.
4.2 Detection Performance
We evaluate our detection performance and compare it with other approaches

using Detection-Error-Tradeoff (DET) curves, plots of miss rates versus false
positives per window (FPPW).
Training. We first extract pose-invariant descriptors for the set of 2416 posi-
tive and 12180 negative samples and batch-train a discriminative classifier for
the initial training algorithm. We use the publically available LIBSVM tool [24]
for binary classification (RBF Kernel) with parameters tuned to C=8000,

gamma=0.04 (as the default classifier).
For improving performance, we perform one round of bootstrapping procedure
for retraining the initial detector. We densely scan 1218 (plus mirror versions)
person-free photos by 8-pixel strides in horizontal/vertical directions and 1.2
scale (down-sampling) factors (until the resized image does not contain any
detection window) to bootstrap false positive windows. This process generates
41667 ‘hard’ samples out of examined windows. These samples are normalized
to 128 × 64 and added to the original 12180 negative training samples and the
whole training process is performed again.
Testing. For evaluation on the MIT dataset, we chose its first 724 image patches
as positive training samples and 12180 training image images from the INRIA
dataset as negative training samples. The test set contains 200 positive samples
Detection Error Tradeoff (DET) curves Detection Error Tradeoff (DET) curves
0.5 0.5
Pose−inv descriptor Pose−inv descriptor, single scale
Class. on Riemannian Man. (1132 pos and 898016 neg windows)
Dalal&Triggs, Ker. HoG Pose−inv descriptor, multiple scales
Dalal&Triggs, Lin. HoG (1132 pos and 2156585 neg windows)
0.2 Zhu et al. Cascade of Rej. 0.2
0.1 0.1
Miss Rate
Miss Rate
0.05 0.05
0.02 0.02
0.01 −6 −5 −4 −3 −2 −1
0.01 −6 −5 −4 −3 −2 −1
10 10 10 10 10 10 10 10 10 10 10 10
False Positives Per Window (FPPW) False Positives Per Window (FPPW)
Confidence Distribution of Positive Test Samples Confidence Distribution of Negative Test Samples
7
1000 10
Positive test samples Single scale
900 Multiple scales
6
10
800
5
700 10
600
Frequency
Frequency
4
10
500
3
400 10
300
2
10
200
1
100 10
0
0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Positive probability (Confidence) Positive probability (Confidence)
Fig. 5. Detection performance evaluation on INRIA dataset. Top-Left: The proposed

approach (testing on single scale) is compared to Kernel HOG-SVM [4], Linear HOG-
SVM [4], Cascaded HOG [18], and Classification on Riemannian Manifold [6]. The
results of [4] are copied from the original paper (actual data might be slightly dif-
ferent), and the results of [18, 6] are obtained by running their original detectors on
the same test data. Top-Right: Performance comparison w.r.t. the number of negative
windows scanned. Bottom: Distribution of confidence values for positive and negative
test windows.
from the MIT dataset and 1200 negative samples from the INRIA dataset. As a
result, we achieve 1.0% true positive rate, and a 0.00% false positive rate even
without retraining. Direct comparisons on the MIT dataset are difficult since
there are no negative samples and no separation of training and testing samples
in this dataset. Indirect comparisons show that our result on this dataset are
similar to the performance achieved previously in [4].
For the INRIA dataset, we evaluated our detection performance on 1132 pos-
itive image patches and 453 negative images. Negative test images are scanned
exhaustively in the same way as in retraining. The detailed comparison of our
detector with current state of the art detectors on the INRIA dataset is plotted
using the DET curves as shown in Figure 5. The comparison shows that our
approach is comparable to state of the art human detectors. The dimensionality
of our features is less than half of that used in HOG-SVM [4], but we achieve
better performance. Another advantage of our approach is that it is capable
of not only detecting but also segmenting human shapes and poses. In this re-
gard, our approach can be further improved because our current pose model is
very simple and can be extended to cover a much wider range of articulations.
Fig. 6. Detection results. Top: Example detections on the INRIA test images, nearby
windows are merged based on distances; Bottom: Examples of false negatives (FNs)
and false positives (FPs) generated by our detector.
Fig. 7. Example results of pose segmentation
Figure 6 shows examples of detection on whole images and examples of false neg-
atives and false positives from our experiments. Note that FNs are mostly due to
unusual poses or illumination conditions, or significant occlusions; FPs mostly
appeared in highly-textured samples (such as trees) and structures resembling
human shapes.
4.3 Segmentation Performance

Figure 7 shows some qualitative results of our pose segmentation algorithm on
the INRIA dataset. Our pose model and probabilistic hierarchical part-template
matching algorithm give very accurate segmentations for most images in the MIT-
CBCL dataset and on over 80% of 3548 training/testing images in the INRIA
dataset. Significantly poor pose estimation and segmentation are observed in
about 10% of the images in the INRIA dataset, and most of those poor segmen-
tations were due to very difficult poses and significant misalignment of humans.
Our detection and segmentation system is implemented in C++ and the cur-
rent running time (on a machine with 2.2GHz CPU and 3GB memory) is as
follows. Both pose segmentation and feature extraction for 800 windows takes
less than 0.2 second; classifying 800 windows with the RBF-Kernel SVM classi-
fier takes less than 10 seconds; initial classifier training takes about 10 minutes
and retraining takes about two hours. The computational overhead is only due
to the kernel SVM classifier which can be replaced with a much faster boosted
cascade of classifiers [3] (which we have implemented recently and runs at 3
frames/second on a 320 × 240 image scanning 800 windows); this is comparable
to [6] (reported as less than 1 second scanning 3000 windows).
5 Conclusion
We proposed a pose-invariant (articulation-insensitive) image descriptor for

learning a discriminative human/non-human classifier. The descriptor is com-
puted adaptively based on human poses instead of concatenating features along
2D image locations as in previous approaches. Specifically, we estimate the poses
using a fast hierarchical matching algorithm based on a learned part-template
tree. Given the pose estimate, the descriptor is formed by concatenating lo-
cal features along the pose boundaries using a one-to-one point correspondence
between detected and canonical poses. Although our approach can handle the
majority of standing human poses, many of our misdetections are still due to
pose estimation failures. This suggests that the detection performance could be
further improved by extending the part-template tree model to handle more dif-
ficult poses and to cope with alignment errors in positive training images. We
are also investigating the addition of color and texture information to our local
contextual descriptor to improve the detection and segmentation performance.
Acknowledgement
This work was funded, in part, by Army Research Laboratory Robotics Col-
laborative Technology Alliance program (contract number: DAAD 19-012-0012
ARL-CTA-DJH). We would like to thank Fatih Porikli, Oncel Tuzel, and Mo-
hamed Hussein for providing results of their approaches for comparison.
References
1. Gavrila, D.M., Philomin, V.: Real-time object detection for smart vehicles. In:
ICCV (1999)
2. Papageorgiou, C., Evgeniou, T., Poggio, T.: A trainable pedestrian detection syste.
In: Proc. of Intelligent Vehicles (1998)
features. In: CVPR (2001)
4. Dalal, N., Triggs, B.: Histograms of oriented gradients for human detection. In:
CVPR (2005)
5. Wu, Y., Yu, T., Hua, G.: A statistical field model for pedestrian detection. In:
CVPR (2005)
6. Tuzel, O., Porikli, F., Meer, P.: Human detection via classification on riemannian
manifold. In: CVPR (2007)
7. Mikolajczyk, K., Schmid, C., Zisserman, A.: Human detection based on a proba-
bilistic assembly of robust part detectors. In: Pajdla, T., Matas, J(G.) (eds.) ECCV
8. Leibe, B., Seemann, E., Schiele, B.: Pedestrian detection in crowded scenes. In:
CVPR (2005)
9. Shotton, J., Blake, A., Cipolla, R.: Contour-based learning for object detection.
In: ICCV (2005)
10. Opelt, A., Pinz, A., Zisserman, A.: A boundary-fragment-model for object detec-
tion. In: Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3952,
11. Ferrari, V., Fevrier, L., Jurie, F., Schmid, C.: Groups of adjacent contour segments
for object detection. IEEE Trans. PAMI 30(1), 36–51 (2008)
12. Wu, B., Nevatia, R.: Detection of multiple, partially occluded humans in a single
image by bayesian combination of edgelet part detectors. In: ICCV (2005)
13. Mohan, A., Papageorgiou, C., Poggio, T.: Example-based object detection in im-
ages by components. IEEE Trans. PAMI 23(4), 349–361 (2001)
14. Lin, Z., Davis, L.S., Doermann, D., DeMenthon, D.: Hierarchical part-template
matching for human detection and segmentation. In: ICCV (2007)
15. Viola, P., Jones, M., Snow, D.: Detecting pedestrians using patterns of motion and
appearance. In: ICCV (2003)
16. Dalal, N., Triggs, B., Schmid, C.: Human detection using oriented histograms of
flow and appearance. In: Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006.
17. Sharma, V., Davis, J.W.: Integrating appearance and motion cues for simultaneous
detection and segmentation of pedestrians. In: ICCV (2007)
18. Zhu, Q., Avidan, S., Yeh, M.C., Cheng, K.T.: Fast human detection using a cascade
of histograms of oriented gradients. In: CVPR (2006)
19. Maji, S., Berg, A.C., Malik, J.: Classification using intersection kernel support
vector machines is efficient. In: CVPR (2008)
20. Sabzmeydani, P., Mori, G.: Detecting pedestrians by learning shapelet features. In:
CVPR (2007)
21. Wu, B., Nevatia, R.: Optimizing discrimination-efficientcy tradeoff in integrating
heterogeneous local features for object detection. In: CVPR (2008)
22. Gavrila, D.M.: A bayesian, exemplar-based approach to hierarchical shape match-
ing. IEEE Trans. PAMI 29(8), 1408–1421 (2007)
23. Felzenszwalb, P.F., Huttenlocher, D.P.: Pictorial Structures for Object Recogni-
tion. International Journal of Computer Vision 61(1), 55–79 (2005)
24. Chang, C.C., Lin, C.J.: LIBSVM: a library for support vector machines (2001),
http://www.csie.ntu.edu.tw/∼ cjlin/libsvm
Texture-Consistent Shadow Removal
Feng Liu and Michael Gleicher
Computer Sciences Department, University of Wisconsin-Madison

{fliu,gleicher}@cs.wisc.edu
Abstract. This paper presents an approach to shadow removal that

preserves texture consistency between the original shadow and lit area.
Illumination reduction in the shadow area not only darkens that area,
but also changes the texture characteristics there. We achieve texture-
consistent shadow removal by constructing a shadow-free and texture-
consistent gradient field. First, we estimate an illumination change
surface which causes the shadow and remove the gradients it induces. We
approximate the illumination change surface with illumination change
splines across the shadow boundary. We formulate estimating these
splines as an optimization problem which balances the smoothness be-
tween the neighboring splines and their fitness to the image data. Second,
we sample the shadow effect on the texture characteristics in the umbra
and lit area near the shadow boundary, and remove it by transforming
the gradients inside the shadow area to be compatible with the lit area.
Experiments on photos from Flickr demonstrate the effectiveness of our
method.
1 Introduction
Shadow removal is often required in digital photography as well as in many vision
applications. For clarity, we define the problem of shadow removal at the very
beginning. Following previous work [1,2,3], an image I can be represented as the
composition of the reflectance field R and the illumination field L as follows:
I(x, y) = R(x, y) · L(x, y)
A shadow image can be formulated by applying an illumination change surface

C(x, y) to the illumination field as follows:
Ĩ(x, y) = I(x, y) · C(x, y) (1)

I(x, y) = I(x, y) + C(x, y)
˜ (2)
where Equation 2 is Equation 1’ counterpart in the log domain. I, ˜ I and C are

the logarithms of Ĩ, I and C respectively. C is usually assumed to be 1 in the lit
area, a constant c(∈ (0, 1)) in the umbra area, and changing from c to 1 in the
penumbra area. This paper works in the log domain.
The normal goal of the shadow removal is to estimate the illumination change
surface C from Ĩ and recover the shadow free image I or I. It is important to

438 F. Liu and M. Gleicher
P U L
(a) original image (b) shadow boundary (c) zoom in (d) our result
Fig. 1. Given a rough shadow boundary ’P’ provided by users (b and c), our algorithm
removes the shadow (d). The red curve inside the brush stroke is the trajectory of the
brush center. Users do not need to provide a precise shadow boundary as shown in
(c) (Notice the eagle’s right wing.). The brush strokes divide the image into 3 areas:
definite umbra areas,’U’, definite lit areas, ’L’, and boundary, ’P’, which contains the
penumbra area as well as parts of the umbra and lit area.
examine how the illumination change surface C affects an image. Since an image
can be reconstructed from its gradient field with proper boundary conditions,
we focus on how C affects the gradient field in the log domain.
1. C will affect the gradients in the penumbra area where it is not uniform.
Ideally, C will not affect the gradients in the umbra and lit area since it is
uniform in these 2 areas, and is canceled off in calculating the gradients.
However, this is not often true in practice as explained in the following.
2. In practice, the imaging process suffers from noise and quantization errors.
Usually the signal to noise/quantization error ratio in the shadow area is
lower than in the lit area. In this way, C makes the effect of noise/quantization
error on the gradients in the shadow area more significant than in the lit area.
3. Normally, the poor lighting in shadow areas can weaken the texture, and
even diminish the details. However, this is not always true for many im-
ages containing highly specular surfaces. If the illumination is strong in the
scene, texture details in the lit area disappear; while in the shadow area, the
reduction of the illumination can keep the textures there.
4. If the surface response curve has a different shape in the shadow and lit area,
scaling up the shadow region to cancel C will change the texture character-
istics.
From the above observations, we can see that applying the illumination change
surface C not only affects the gradients in the penumbra area, it also affects
the characteristics of the gradient fields in the whole shadow area. We call the
former the shadow effect on the penumbra gradients and the latter the shadow
effect on the gradient characteristics in the shadow area.
1.1 Previous Work
This paper focuses on removing shadows from a single image. Many methods
have been presented to address this problem. Shadow removal is usually achieved
Texture-Consistent Shadow Removal 439
(a) original image (b) multiplying a constant (c) zeroing gradient
(d) texture preserving [3] (e) in-painting [4] (f) our result
Fig. 2. Motivating example. (b): multiplying constant to the image intensities inside
the shadow region. (c): zeroing gradients inside the shadow boundary. (d): texture-
preserving shadow removal [3]. (e): in-painting the shadow boundary region [4].
in two steps: shadow detection and image reconstruction. (Approaches to remov-

ing shadow using information from multiple images have also been presented(c.f.
[2,5]).)
Many methods have been presented to automatically detect shadow regions.
Finlayson et al. estimate an illumination invariant image based on an invariant
color model, and use this invariant image together with the original image to
locate the shadow region [6]. Similarly, Salvador et al. use invariant color features
to segment cast shadows [7]. Levine and Bhattacharyya [8] study properties
of color ratios across boundaries between regions in a segmented image, and
use a support vector machine to identify shadow regions based on these color
ratios. In digital photography, shadow boundaries are often specified through
user interactions [9,10]. Our algorithm relies on users to roughly specify the
shadow boundary, and refines it automatically.
Once shadow areas are located, they can be removed by multiplying a suitable
scalar to the shadow pixels to cancel the effect of the illumination change surface
C. This easy method can create a noticeable over-saturated band in the penumbra
area as illustrated in Fig. 2(b). To solve this problem, Baba et al. adapt the
multiplicative scalars C based on shadow densities [11]. Recently, Arbel and
Hel-Or estimate C by considering the surface geometry, and effectively remove
the shadow while preserving texture in both umbra and penumbra areas [3], as
illustrated in Fig. 2(d).
Alternatively, shadows can be removed by zeroing gradients in the penumbra
area and applying 2D integration to reconstruct the shadow-free image [2,6].
These methods usually work in the log image domain. As shown in Fig. 2(c),
zeroing gradients in the penumbra area nullifies the texture there, however. To
solve this problem, in-painting techniques are applied to fill in the missing tex-
ture [12,4]. However, in-painting sometimes introduces inconsistent textures as
illustrated in Fig. 2(e). Alternatively, Mohan et al. [10] estimate a soft shadow
model in the penumbra area, and remove shadow effect in the gradient domain
accordingly.
Although previous methods vary in estimating the illumination change sur-
face C, they share common ideas to reconstruct the shadow-free image in the
umbra area: multiplying a constant scalar to cancel the effect of C. Applying 2D
integration in the log domain with proper boundary conditions is equivalent to
multiplying a constant in the image domain. This scheme can effectively match
the overall illumination in the umbra area to that in the lit area. And using
proper scalar constants to the penumbra area can also cancel the shadow effect
on the penumbra area. However, these methods can not remove the shadow ef-
fect on the texture characteristics of the shadow area. Multiplying a constant
can magnify the noise and quantization error in the original shadow region. For
particular images with strong specular surface and strong lighting, the details
in the shadow area, which disappear in the lit area, will be enhanced. All these
lead to inconsistent texture between the shadow area and lit area. For example,
the texture in the shadow area in Fig. 2(c), (d) and (e) is not compatible with
that in the lit area.
1.2 Our Contribution

In this paper, we present a shadow removal method that preserves texture con-
sistency. Since textures manifest themselves by image gradients, our algorithm
works in the gradient domain. Specifically, we construct a new image gradient
field that removes the shadow effects on both the gradients in the penumbra
area and the characteristics on the gradients in the whole shadow area. From
this new image gradient field, we can reconstruct the shadow-free image by solv-
ing a Poisson equation.
Our major contribution is a method for constructing a shadow-free and
texture-consistent gradient field by removing the two-fold shadow effects on
the gradient field as mentioned previously. First, we simultaneously locate the
penumbra area and estimate the illumination change curves across the shadow
boundary by estimating and sampling the illumination change surface using
line segments. With the illumination change curves, we can cancel the effect of
shadow on the gradient field in the penumbra area. Second, we estimate the
shadow effect on the gradient distribution in the shadow area, and transform
the gradient field there to cancel the effect to be consistent with that in the lit
area. In this way, we obtain the texture-consistent gradient field. These two key
algorithms are detailed in § 2. Experiments on photos from Flickr demonstrate
the effectiveness of our algorithm as detailed in § 3.
2 Texture-Consistent Shadow Removal
In this paper, we provide a brush tool for users to mark the shadow boundary.
As illustrated in Fig. 1(c), users can select a brush with much larger size than
the boundary, and do not need to delineate the boundary precisely. The brush
strokes divide an image into three areas: definite umbra area, definite lit area,
and boundary, which consists of penumbra area as well as parts of the umbra
and lit area. Our algorithm precisely locates the penumbra area from the user
specified boundary, and removes the shadow seamlessly. A working example of
our algorithm is illustrated in Fig. 1.
This paper aims to remove shadow effects such that the resulting shadow-free
image has consistent texture between the shadow and lit area. We first construct
a new image gradient field that removes the gradients induced by the shadow
effect and has consistent gradient characteristics between the shadow and lit
area. Then we can reconstruct the shadow-free image from the new gradient
field through 2D integration by solving a Poisson equation similar to previous
work (c.f. [2,6,13]). The major challenge is to construct the new image gradient
field Gn given only the rough shadow boundary from users. In § 2.1, we de-
scribe a novel algorithm to estimate the illumination change curves across the
shadow boundary and cancel the effect of illumination change on the gradient
field in the penumbra area. In the § 2.2, we describe a method to estimate the
shadow effect on the texture characteristics in the shadow area and transform
the characteristics of gradients there to be compatible with that in the lit area.
2.1 Estimate Illumination Change in Penumbra Area
Properly handling the shadow boundary or the penumbra area is a challenge for
shadow removal. The ambiguity of the shadow boundary often makes automatic
shadow boundary detection methods fail. Relying on users to provide the pre-
cise shadow boundary casts a heavy burden on them. To relieve users’ burden,
Mohan et al. [10] presented a piece-wise model where users only need to specify
connected line segments to delineate the boundary. However, when dealing with
complex shadow boundaries like the eagle’s right wing in Fig. 1(c), their method
will still require users to specify a large number of key points. To further reduce
users’ burden, we only require a rough specification of the shadow boundary
from users using brush tools as illustrated in Fig. 1(c).
Given an inaccurate shadow boundary specification, our method simultane-
ously locates the shadow boundary precisely and estimates the illumination
change C(x, y) in Equation 2 in the penumbra area. The complex shape of
the shadow boundary makes devising a parametric model of C(x, y) difficult.
However, we observe that any line segment crossing the boundary has an easily
parameterizable illumination profile. Therefore, we model C(x, y) by sampling
line segments across the boundary and estimating a parametric model for each as
illustrated in Fig. 3(a). Since the user provided-boundary usually is not accurate
enough, unlike [3], we do not sample C(x, y) using line segments perpendicular
to the boundary. Instead, like [10], we use a vertical/horizontal sampling line per
0.5
Illumination change
0
-0.5
extent
extent
r r
-1.5
t1 t0 t2
Position on the sampling line
(a) vertical sampling lines (b) illumination change model
Fig. 3. Sampling illumination change surface using line segments. (a): vertical sampling
lines. (b): t0 and r are the brush center and brush radius. [t1 , t2 ] is the penumbra
area. extent is the range in the umbra and lit area, used to estimate the gradient
characteristics.
pixel along the boundary and use the estimated illumination change to cancel
the shadow effect on the gradient in Y/X direction. We estimate horizontal and
vertical illumination change sampling lines independently.
We model the illumination change along each line segment as the following
C 1 continuous piece-wise polynomial as illustrated in Fig. 3(b):
⎧
⎨ c, t < t1 ;
Cl (t) = f (t), t1 ≤ t ≤ t2 ; (3)
⎩
0, else.
This piece-wise polynomial model can be parameterized by 3 parameters, de-

noted as Ml (c, t1 , t2 ). Here t1 and t2 define the penumbra area along the sampling
line. (Without losing generality, we assume t < t1 lies in the umbra area and
t > t2 lies in the lit area.) c(≤ 0) is the reduction of the illumination in the umbra
area. f (t) is a cubic curve determined by the two boundary points, (t1 , c) and
(t2 , 0), and the derivatives at these two points, f (t1 ) = 0 and f (t2 ) = 0. This
illumination change model is determined by both the location of the penumbra
area and the characteristics how the illumination changes from c in the umbra
area to 0 in the lit area. Due to these combined properties, our method esti-
mates the penumbra area location and the illumination change simultaneously
by estimating the above piece-wise polynomial model.
Because we assume that the illumination change surface is smooth, neighbor-
ing illumination change models along the shadow boundary should be similar
to each other. So we solve for all these models simultaneously instead of fitting
each model separately. We formulate the problem of finding illumination change
models as an optimization problem, aiming to balance the fitness of the models
to the shadow image and the smoothness between neighboring models.

E= Ef it (Mli , I)
˜ +λ Esm (Mli , Mlj ) (4)
li li lj∈N (li)
where Ef it (Mli , I)
˜ measures the fitness error of the illumination change model
Mli to the original shadow image I, ˜ Esm (Mli , Mlj ) measures the similarity be-
tween Mli and Mlj , and N (li) denotes the neighborhood of sampling line li. λ
is a parameter, with a default value 10.
We measure Ef it (Mli , I),
˜ the fitness error of the model Mli to the shadow
image I,
˜ as how well the gradient in the penumbra area fits into its neighborhood
along the sampling line after shadow effect compensation according to Mli .
Ef it (Mli , I)
˜ = −Πt∈[t −r ,t +r ] ϕ(Ĝli (t), T tex )
i0 i i0 i li (5)
Ĝli (t) = G̃li (t) − Cli (t) (6)
where Cli is the illumination change curve of Mli as defined in Equation 3, Cli
is its first derivative, G̃li is the gradient along li, and Ĝli (t) is the gradient
after canceling the shadow effect. Tlitex is the texture distribution along li. ϕ(, )
measures the fitness of the gradient to the distribution Tlitex . We model the
texture distribution along li as a normal distribution N (μi , σi2 ) of the gradients,
which can be estimated explicitly from the umbra and lit extension along li as
illustrated in Fig. 3(b). Accordingly, we define the fitness measure as follows:
exp(−(Gli (t) − μi )2 /2σi2 )
ϕ(Gli (t), Tlitex ) = . (7)
2πσi2
We define Esm (Mli , Mlj ), the smoothness cost between neighboring illumination
change models as follows:
Esm (Mli , Mlj ) = γ(ci − cj )2 + (1 − γ)((t1i − t1j )2 + (t2i − t2j )2 )
where the first term measures the difference between the illumination steps from
the umbra to lit area, and the second term measures the difference between the
location of the penumbra area along sampling lines. We emphasize the fact that
the illumination change inside the umbra area is mostly uniform by weighting
the first term significantly. The default value for γ is 0.9.
Directly solving the minimization problem in Equation 4 is time-consuming.
We approximate the optimal solution in two steps:
1. For each sampling line li, we find an optimal illumination change model Mlio
which fits the shadow image most by minimizing the fitness error defined
in Equation 5. Since the extent of the penumbra area is small, we use a
brute-force search method.
2. With the optimal illumination change model Mlio of each sampling line, we
approximate the fitness error term in Equation 4 using the difference between
the illumination change model Mli and Mlio as follows:

E= Esm (Mli , Mlio ) + λ Esm (Mli , Mlj )
li li lj∈N (li)
The above energy minimization is a quadratic minimization problem. We

solve it using a Preconditioned Conjugate Gradient method [14].
(a) original image (b) after removing shadow (c) after texture transfer
8
−2
original −4
shadow removal −6
texture transfer
−8
20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180
(d) intensity along row 10 (e) gradient along row 10
Fig. 4. Reconstruct the gradient field for shadow removal. (a) shows the original image
and its gradient field along X direction. For the sake of illustration, we encode the
negative and positive gradient values using the GREEN and RED channels respectively.
From the original gradient field, we can see the shadow effect on the gradient field by
noticing the strong edges along the shadow boundary. By estimating the illumination
change across the penumbra area, the shadow effect on the gradient field is canceled
as illustrated in (b) and (d). However, as we can see in (b) and (e) right, the shadow
area is more contrasty than the lit area, causing inconsistent texture characteristics.
This inconsistency is removed after gradient transformation as shown in (c) and (e).
After obtaining the illumination change model along each sampling line, we
apply it to the gradient field to cancel the shadow effect according to Equation 6.
An example of canceling the shadow effect on the gradients in the penumbra area
is shown in Fig. 4(a) and (b).
2.2 Estimate Shadow Effect on Texture Characteristics

Canceling the shadow effect on the gradients in the penumbra area can effectively
match the illumination in the shadow area (including penumbra and umbra
area) to that in the lit area. However, as illustrated in Fig. 4(b) and (c), it
cannot guarantee the texture consistency between the shadow and lit area since
the shadow can also affect the texture characteristics in the whole shadow area
(§ 1). Our method estimates the shadow effect on the gradient characteristics
and transfers the shadow-effect free gradient characteristics to the shadow area
to make it compatible with the lit area.
Like transferring color between images [15], where the global color charac-
teristics of an image is parameterized using its sampling mean and deviation,
we model the texture characteristics using the sampling mean and deviation of
the gradient field. So if given the target mean and deviation, we transform the
gradient field in the shadow area as follows:
(Ĝs (x, y) − μ̂s ) ∗ σ̂ t

Gs (x, y) = μ̂t + (8)
σ̂ s
where Ĝs and Gs are the gradients in the shadow area before and after trans-
formation respectively, and μ̂s and σ̂ s are the mean and deviation of Ĝs . μ̂t and
σ̂ t are the target mean and deviation.
Like transferring color [15], using the characteristics parameters of the lit area
as the target parameters can achieve consistent texture characteristics between
the shadow and lit area. However, this scheme works well only if the texture
distribution is globally homogeneous in the image. Otherwise it can destroy local
textures in the shadow area. We calculate the target characteristics parameters
by estimating the shadow effect on the gradient distribution and canceling this
effect from the original gradient field. Assuming the gradient distribution around
the shadow boundary is homogenous and the shadow effect is independent of
the shadow-free image, we estimate the shadow effect parameters from gradients
around the boundary as follows:
5
μse = μsb − μlb
2 (9)
σse
2
= σbs 2 − σbl
where μse and σse are the mean and deviation of the shadow effect on gradients
in the shadow area. μsb and σbs are the mean and deviation of the gradients in
the umbra side along the shadow boundary(the extent parts as illustrated in
Fig. 3(b)) , and μlb and σbl are those in the lit area side. Accordingly, the target
mean and deviation can be calculated by canceling the shadow effect as follows:
t
μ̂ = μ̂
− μse
s
2 (10)
σ̂ t = σˆs − σse2
Fig. 4(b) and (c) shows that the gradient field transformation leads to consistent
texture characteristics between the shadow and lit area. Please refer to the whole
image in Fig. 6(a) to examine the consistency of the texture.
3 Results
We have experimented with our method on photos with shadows from Flickr.
These photos have different texture characteristics. We report some representa-
tive ones together with the results in Fig. 1, Fig. 2, Fig. 6, Fig. 7 and Fig. 8, as
well as comparison to many representative works [2,6,4,3,10]. (Please refer to
(a) original (b) zeroing-gradient
(c) result from [10] (d)our result
Fig. 5. Images in (a) and (c) are from [10]. (b) shadow removed by nullifying the
gradients in the boundary [2,6]. (c) shadow removed using the method from [10]. There,
not only the illuminance level in the lit area is changed, but also the shadow area is
not as contrasty as the lit area. Our method creates a texture-consistent result.
the electronic version of this paper to examine the results. Zooming

in on the images will be helpful for the examination.)
For all the experiments, users specify the shadow boundaries with a brush
tool. Users do not need to delineate the boundary precisely as shown in Fig. 1(c)
(notice the eagle’s right wing). They can pick a brush with much larger size than
the real shadow boundary area to cover the boundary as shown in the second
column of Fig. 6. Given the user specified shadow boundary, our system can
automatically perform shadow removal efficiently. The majority of the time is
spent on solving the Poisson equation, whose complexity is dependent on the
number of pixels in the shadow region. It takes about 3 seconds to remove a
shadow region with about 60,000 colored pixels on a 2.2GHz Athlon machine.
Fig. 2 and Fig. 5 compare our method to other representative methods.
Methods [2,6] cancel the shadow effect by zeroing the gradients in the boundary
area. In this way, the textures there are nullified as shown in Fig. 2(c). While
in-painting [4] can partially solve this problem, it sometimes destroys the
continuity of the texture as shown in Fig. 2(e). The recent method from [3] can
effectively remove shadow, however the texture in the original shadow area is
not consistent with that in the lit area as shown in Fig. 2(d). Our method can
not only remove the shadows, but also keep the texture consistency between the
shadow and lit area as shown in Fig. 2(f). Fig. 5 compares our method to the
recent work from [10]. While the illuminance between the lit and the original
shadow area is balanced in the result from [10], the illuminance level in the lit
area is changed. More overall, the lit and the original shadow area have different
(a) sandy beach
(b) pavement
(c) rock cliff

Fig. 6. Experiments results. Left: original images; Middle: shadow boundaries; Right:
our results.
(a) tree in hill (b) sandy beach
(c) desert sand dawn break (d) bridge over river
Fig. 7. Experiments results. Left: original images; Right: our results.
contrasty levels as shown in Fig. 5(c). Our method effectively removes the shadow
as well as keeps the consistent texture characteristics across the whole image as
shown in Fig. 5(d) and other examples. For instance, in the Fig. 7(b), the texture
of small shell grains in the shadow area and in the lit area is consistent. For the
desert example in Fig. 7(c), the highlights across the original shadow boundary
are consistent between the shadow and lit area. For the river surface example
in Fig. 7(d), the ripples in the shadow area are consistent with that in the lit
area. Particularly, the wavefront in the middle is continuous across the original
shadow boundaries. For the tree example in Fig. 7(a), the soil inside the shadow
region is consistent with the lit area surrounding it. The hill example in Fig. 8(a)
is similar.
(a) rock cliffs (b) mountain above clouds
(c) volcano above clouds (d) cast shadow of semi-transparent object
Fig. 8. Experiments results. Left: original images; Right: our results.

From the results in Fig. 6, 7 and 8, we can see that the proposed algorithm
can seamlessly remove shadows in images with various texture characteristics.
For example, the shadows are on the beach (Fig. 6(a)), on the road surfaces
(Fig. 6(b)), on the sands (Fig. 7(b)), on the desert (Fig. 7(c)), on the river
surface (Fig. 7(d)), on the hills (Fig. 7(a) and Fig. 8(a)), etc. Our method works
well on specular surfaces such as Fig. 6(a), as well as Lambertian surfaces, such
as examples in Fig. 7.
Examples in Fig. 8(b) and (c) are very interesting. Noticing the mountains
in these examples, shadow removal reveals the beautiful texture details in the
original dark shadow areas, which are concealed in the original shadow images.
What is particularly interesting is that shadow removal recovers the blue glacier
ice phenomenon1 in the Fig. 8(b) (Notice the blue-cyan area of the snow in the
left bottom.).
We found from the experiments that our method does not work well on some
images. Taking Fig. 8(d) as an example, the shadow area in the original im-
age looks more reddish than its surrounding lit area. This is because when the
lighting is blocked by the semi-transparent red leaf, its red component can still
pass through. For this kind of cast shadow, the general shadow model in Equa-
tion 2 used in previous work (including ours) does not hold. Noticing the original
shadow region in the resulting image, we can still sense the reddish component
there. In future, analyzing the caustics of shadow from its context may help solve
this problem. However, our current method is effective for many images.
4 Conclusion
In this paper, we presented a texture-consistent shadow removal method. Specif-

ically, we construct a shadow-effect free and texture-consistent gradient field
between the shadow and lit area and recover the shadow-free image from it
by solving a Poisson equation. The experiments on shadow images from Flickr
demonstrate the effectiveness of the proposed method.
Currently, our method provides users with a brush tool to specify the shadow
boundary. The brush tool is very popular in digital photography software. As
illustrated in the examples in previous sections, our method does not require
a precise shadow boundary. We envision our method a convenient tool for in-
teractive photo editing. Of course, integrating an automatic shadow detection
algorithm can make our method even easier to use.
We characterize texture characteristics using the sampling mean and devi-
ation of the gradient field. Based on our current experiments on photos from
Flickr, this global model works well. An important reason for its success is that
a global transformation on an image or its various representations usually pre-
serves important properties of the original image. In fact, similar models work
pretty well in other applications like color transfer [15] as well.
1
http://www.northstar.k12.ak.us/schools/joy/denali/OConnor/
colorblue.html
Acknowledgements. We would like to thank reviewers for their constructive

suggestions. The Flickr images are used under a Creative Commons license
from Flickr users: etamil, viktoria s, 82684220@N00, el chupacabrito, magnusvk,
30201239@N00, erikogan, 24342028@N00, mishox, gandhu, hamedmasoumi and
lexnger. This research was sponsored in part by NSF grant IIS-0416284.
References
1. Barrow, H., Tenenbaum, J.: Recovering intrinsic scene characteristics from images.
In: Computer Vision Systems. Academic Press, London (1978)
2. Weiss, Y.: Deriving intrinsic images from image sequences. In: IEEE ICCV, pp.
68–75 (2001)
3. Arbel, E., Hel-Or, H.: Texture-preserving shadow removal in color images contain-
ing curved surfaces. In: IEEE CVPR (2007)
4. Finlayson, G.D., Hordley, S.D., Lu, C., Drew, M.S.: On the removal of shadows
from images. IEEE Trans. Pattern Anal. Mach. Intell. 28(1), 59–68 (2006)
5. Liu, Z., Huang, K., Tan, T., Wang, L.: Cast shadow removal combining local and
global features. In: The 7th International Workshop on Visual Surveillance (2007)
6. Finlayson, G.D., Hordley, S.D., Drew, M.S.: Removing shadows from images. In:
7th European Conference on Computer Vision, pp. 823–836 (2002)
7. Salvador, E., Cavallaro, A., Ebrahimi, T.: Cast shadow segmentation using invari-
ant color features. Comput. Vis. Image Underst. 95(2), 238–259 (2004)
8. Levine, M.D., Bhattacharyya, J.: Removing shadows. Pattern Recognition Let-
ters 26(3), 251–265 (2005)
9. Wu, T.P., Tang, C.K., Brown, M.S., Shum, H.Y.: Natural shadow matting. ACM
Trans. Graph. 26(2), 8 (2007)
10. Mohan, A., Tumblin, J., Choudhury, P.: Editing soft shadows in a digital photo-
graph. IEEE Comput. Graph. Appl. 27(2), 23–31 (2007)
11. Baba, M., Mukunoki, M., Asada, N.: Shadow removal from a real image based on
shadow density. ACM SIGGRAPH 2004 Posters, 60 (2004)
12. Fredembach, C., Finlayson, G.D.: Hamiltonian path based shadow removal. In:
BMVC, pp. 970–980 (2005)
13. Pérez, P., Gangnet, M., Blake, A.: Poisson image editing. ACM Trans.
Graph. 22(3), 313–318 (2003)
14. Barrett, R., Berry, M., Chan, T.F., Demmel, J., Donato, J., Dongarra, J., Eijkhout,
V., Pozo, R., Romine, C., der Vorst, H.V.: Templates for the Solution of Linear
Systems: Building Blocks for Iterative Methods. SIAM, Philadelphia (1994)
15. Reinhard, E., Ashikhmin, M., Gooch, B., Shirley, P.: Color transfer between images.
IEEE Comput. Graph. Appl. 21(5), 34–41 (2001)
Scene Discovery by Matrix Factorization
Nicolas Loeff and Ali Farhadi
University of Illinois at Urbana-Champaign,

Urbana, IL, 61801
{loeff,afarhad2}@uiuc.edu
Abstract. What constitutes a scene? Defining a meaningful vocabulary for scene

discovery is a challenging problem that has important consequences for object
recognition. We consider scenes to depict correlated objects and present visual
similarity. We introduce a max-margin factorization model that finds a low di-
mensional subspace with high discriminative power for correlated annotations.
We postulate this space should allow us to discover a large number of scenes in
unsupervised data; we show scene discrimination results on par with supervised
approaches. This model also produces state of the art word prediction results in-
cluding good annotation completion.
1 Introduction
Classification of scenes has useful applications in content-based image indexing and re-
trieval and as an aid to object recognition (improving retrieval performance by removing
irrelevant images). Even though a significant amount of research has been devoted to
the topic, the questions of what constitutes a scene has not been addressed. The task
is ambiguous because of the diversity and variability of scenes but also mainly due to
the subjectivity of the task. Just like in other areas of computer vision such as activity
recognition, it is not simple to define the vocabulary to label scenes. Thus, most ap-
proaches have used the physical setting where the image was taken to define the scene
(e. g. beach, mountain, forest, etc.).
Previous work is focused on supervised approaches. It is common to use techniques

that do not share knowledge between scene types. For instance, In [12] Lazebnik pro-
poses a pyramid match kernel on top of SIFT features to measure image similarity and
applies it to classification of scenes using an SVM. Chapelle et al. [6] use global color
histograms and an SVM classifier.
Therefore other models build intermediate representations, usually as a bag of fea-
tures, in order to perform classification. Internal representations let classifiers share
features between scene classes. Quelhas and Odobez [19] propose a scene represen-
tation using mixtures of local features. Fei-Fei and Perona [13] use a modified Latent
Dirichlet Allocation model on bags of patches to create a topic representation of scenes.
Scenes are also directly labeled during training. Liu and Shah [14] use maximization
of mutual information between bags of features and intermediate concepts to create an
internal representation. These intermediate concepts are purely appearance based. On
top of it, they run a supervised SVM classifier. Bosch et al. [3] uses a pLSA model on

452 N. Loeff and A. Farhadi
top of bags of features to discover intermediate visual representations and a supervised

KNN classifier to identify scenes.
Other approaches first manually define a vocabulary for the internal representation
and then try to learn it. J. C. van Gemert et al. [22] describe scenes using “proto-
concepts” like vegetation, sky and water, and learning using image statistics and con-
text. Vogel and Schiele [24] manually label 9 different intermediate “concepts” (e. g.
water, sky, foliage) and learn a KNN classifier on top of this representation. Oliva and
Torralba [17] use global “gist” features and local spatial constraints, plus human la-
beled intermediate properties (such as “roughness” or “openness”) as an intermediate
representation.
We propose a different strategy. First, we aim to find scenes without supervision.
Second, we treat the building of the internal representation not as separate from a clas-
sification task, but as interdependent processes that must be learnt together.
What is a scene? In current methods, visual similarity is used to classify scenes into a
known set of types. We expect there are many types of scene, so that it will be hard to
write down a list of types in a straightforward way. We should like to build a vocabulary
of scene types from data. We believe that two images depict the same scene category if:
1. Objects that appear in one image could likely appear in the other
2. The images look similar under an appropriate metric.
This means one should be able to identify scenes by predicting the objects that are
likely to be in the image, or that tend to co-occur with objects that are in the image.
Thus, if we could estimate a list of all the annotations that could reasonably be attached
to the image, we could cluster using that list of annotations. The objects in this list of
annotations don’t actually have to be present – not all kitchens contain coffee makers –
but they need to be plausible hypotheses. We would like to predict hundreds of words
for each of thousands of images. To do so, we need stable features and it is useful to
exploit the fact that annotating words are correlated.
All this suggests a procedure akin to collaborative filtering. We should build a set of
classifiers, that, from a set of image features, can predict a set of word annotations that
are like the original annotations. For each image, the predicted annotations will include
words that annotators may have omitted, and we can cluster on the completed set of
annotations to obtain scenes. We show that, by exploiting natural regularization of this
problem, we obtain image features that are stable and good at word prediction. Clus-
tering with an appropriate metric in this space is equivalent to clustering on completed
annotations; and the clusters are scenes.
We will achieve this goal by using matrix factorization [21,1] to learn a word classi-
fier. Let Y be a matrix of word annotations per image, X the matrix of image features
per image, and W a linear classifier matrix, we will look for W to minimize
J(W ) = regularization(W ) + loss(Y, W t X) (1)
The regularization term will be constructed to minimize the rank of W , in order to im-
prove generalization by forcing word classifiers to share a low dimensional represen-
tation. As the name “matrix factorization” indicates, W is represented as the product
Scene Discovery by Matrix Factorization 453
Fig. 1. Matrix factorization for word prediction. Our proxy goal is to find a word classifier W
on image features X. W factorizes into the product W = F G. We regularize with the rank of
W ; this makes F t X a low-dimensional feature space that maximizes word predictive power.
In this space, where correlated words are mapped close, we learn the classifiers G.
between two matrices F G. This factorization learns a feature mapping (F ) with shared
characteristics between the different words. This latent representation should be a good
space to learn correlated word classifiers G (see figure 1).
Our problem is related to multi-task learning as clearly the problem of assigning one
word to an image is correlated with the other words. In a related approach [2] Ando
and Zhang learn multiple classifiers with a shared structure, alternating fixing the struc-
ture and learning SVM classifiers and fixing the classifiers to learn the structure using
SVD. Ando and Zhang propose an interesting insight into the problem: instead of do-
ing dimensionality reduction on the data space (like PCA), they do it in the classifier
space. This means the algorithm looks for low-dimensional structures with good pre-
dictive, rather than descriptive, power. This leads to an internal representation where the
tasks are easier to learn. This is a big conceptual difference with respect to approaches
like [14,3]. It is also different from the CRF framework of [20], where pairwise co-
occurrence frequencies are modeled.
Quattoni et al. [18] proposed a method for supervised classification of topics using
auxiliary tasks, following [2]. In contrast, our model we discover scenes without super-
vision. We also differ in that [18] first learns word classifiers, fixes them, and then finds
the space for the topic (scene) prediction. We learn both the internal structure and the
classifiers simultaneously, in a convex formulation. Thus our algorithm is able to use
correlation between words not only for the scene classification task but also for word
prediction. This results in improved word prediction performance. In section 4 we show
the model also produces better results than [18] for the scene task, even without having
the scene labels!
2 A Max-Margin Factorization Model

Consider a set of N images {xi }, each represented by a d-dimensional vector, and
M learning tasks which consist in predicting the word yim ∈ {−1, 1} for each im-
age using a linear classifier wm
t
xi . This can be represented as Y ∼ W t X for a matrix
Y ∈ {±1}M×N where each column is an image and each row a word, W ∈ Rd×M
is the classifier matrix and X ∈ Rd×N the observation matrix. We will initially con-
sider
that the words are decoupled (as in regular SVMs), and use the L2 regularization
2 2
m ||wm ||2 = ||W ||F (known as the Frobenius norm of W ). A suitable loss for a
max-margin formulation is the hinge function h(z) = max(0, 1 − z). The problem can
then be stated as
1 2
N M
min ||W ||F + C Δ(yim )h(yim · (wm
t
xi )) (2)
W 2 i=1 m=1
where C is the trade-off constant between data loss and regularization, and Δ is a slack
re-scaling term we introduce to penalize errors differently: false negatives Δ(1) = 1
and false positives Δ(−1) = < 1. The rationale is that missing word annotations are
much more common than wrong annotation for this problem.
Our word prediction formulation of the loss is different from [21] (a pure collabora-
tive filtering model) and [1] (a multi-class classifier), even though our tracenorm regu-
larization term is similar to theirs. Our formulation is, to the best of our knowledge, the
first application of the tracenorm regularization to a problem of these characteristics.
From [1] we took the optimization framework, although we are using different losses
and approximations and we are using BFGS to perform the minimization. Finally, we
introduce a unsupervised model on top of the internal representation this formulation
produces to discover scenes.
Matrix Factorization: In order to exploit correlations in the words, an alternative prob-

lem is to factor the matrix W = F G where F ∈ Rd×k can be interpreted as a mapping
of the features X into a k dimensional latent space and G ∈ Rk×M is a linear clas-
sifier on this space (i. e. Y ∼ Gt (F t X)). Regularization is provided by constraining
the dimensionality of the latent space (k) and penalizing the Frobenius norm of F and
G [21]. The minimization in F and G is unfortunately non-convex, and Rennie sug-
gested using the tracenorm (the minimum of the possible sum of Frobenius norms so
that W = F G)as an alternative regularization. As the tracenorm may also be written
as ||W ||Σ = l |γl | (where γl is the l−th singular value), tracenorm minimization
can be seen as minimizing the L1 norm of the singular values of W . This leads to a
low-rank solution, in which correlated words share features, while the Frobenius norm
of W (which minimizes the L2 norm of the singular values) assumes the words are
independent.
Minimization is now with respect to W directly, and the problem is convex. More-
over, the dimensionality k doesn’t have to be provided.
1 N M
min ||W ||Σ + C Δ(yim )h(yim · (wm
t
xi )) (3)
W 2 i=1 m=1
Rennie [21] showed (3) can be recast as a Semidefinite Program (SDP). Unfortunately,
SDPs don’t scale nicely with the number of dimensions of the problem, making any
decent size problem intractable. Instead, he proposed gradient descent optimization.
Fig. 2. Smooth approximations of the hinge function (left) and absolute value function (right),
used in the gradient descent optimization
2.1 Gradient Based Optimization

Equation 3 is not differentiable due to the hinge loss and the tracenorm, but the equation
can be approximated to arbitrary precision by a smoothed version. This allows to per-
form gradient based optimization. We will consider a smooth approximation hρ (z) of
the hinge loss h(z) that is exact for |1 − z| ≥ ρ, and is twice differentiable everywhere:
⎧
⎨ 4 z z>ρ
−z 3z 2
h(1 − z) ≈ hρ (1 − z) = 16ρ3 + 8ρ + 2 + 16 |z| ≤ ρ
3z 3σ
(4)
⎩
0 z < −ρ

For the tracenorm we use ||W ||Σ ≈ ||W ||S = l aσ (γl ), where the smoothed absolute
value is again exact for |x| ≥ σ and is twice differentiable everywhere,
5
|x| |x| > σ
aσ (x) = −z4 z2 (5)
8σ3 + 4σ + 8 |x| ≤ σ
3σ
In our experiments we use ρ = σ = 10−7 . Plots for both approximation are depicted in
figure 2.
We will then consider the smooth cost
J(W ; Y, X, σ, ρ) = JR (W ; σ) + C · JD (W ; Y, X, ρ) (6)
where the regularization cost is
JR (W, σ) = ||W ||S (7)
and the data loss term is

N
M
JD (W ; Y, X, ρ) = Δ(yim )hρ (yim · (wm
t
xi )) (8)
i=1 m=1
Using the SVD decomposition W = U DV t ,

∂JR
= U aσ (D)V t (9)
∂W
The gradient of the data loss term is

∂JD
= −(Δ(Y ) · hρ (Y · W t X))t (Y · X) (10)
∂W
where (A · B) is the Hadamard or element-wise product: (A · B)ij = aij bij . Exact
second order Newton methods cannot be used because of the size of the Hessian, so we
use limited-memory BFGS for minimization.
2.2 Kernelization
A interesting feature of problem 3 is that it admits a solution when high dimensional
features X are not available but instead the Gram matrix K = X t X is provided. Theo-
rem 1 in [1] can be applied with small modifications to prove that there exists a matrix
α ∈ RM×N so that the minimizer of (3) is W = Xα. But instead of solving the dual
Lagrangian problem we will use this representation of W to minimize the primal prob-
lem (actually, it’s smoothed version) using gradient descent. The derivatives in terms of
K and α only become
∂JR ∂ ||Xα||S X t ∂ ||Xα||S
= = = KαV D−1 aσ (D)V t (11)
∂α ∂α ∂Xα
using that D(V V t )D−1 = I, Xα = U DV t , and that K = X t X. The gradient of the
data loss term is
∂JD
= −K ∗ (Δ(Y ) · hρ (αt Kα) · Y ) (12)
∂W
3 Scene Discovery – Analysing the Latent Representation

Section 2.1 introduced a smooth approximation to the convex problem 3. After conver-
gence we obtain the classification matrix W . The solution does not provide the factor-
ization W = F G. Moreover, any decomposition W = F G is not unique as a full rank
transformation F̃ = F A, G̃ = A−1 G will produce the same W .
What is a good factorization then? As discussed in the section 1 clustering in the
latent space should be similar to clustering the word predictions. Since we define scenes
as having correlated words, a good factorization
of W should
maximally transfer the
correlation between the predicted words (W t X)t (W t X) to the correlation in the

latent space (At F t X)t (At F t X) . Identifying terms, A = (GGt )1/2 . In this space
(At F t X), images with correlated words (i. e. belonging to the same scene category)
should cluster naturally.
For the factorization of W we will use a truncated SVD decomposition and then we
will use this A. We will measure their similarity of images in this space using the cosine
distance.
4 Experiments
To demonstrate the performance of our scene discovery model we need a dataset with
multiple object labels per image. We chose the standard subset of the Corel image
collection [7] as our benchmark dataset. This subset has been extensively used and
consists of 5000 images grouped in 50 different sets (CDs). These images are separated
into 4500 training and 500 test images. The vocabulary size of this dataset is 374, out
of which 371 appear in train and 263 in test set. The annotation length varies from 1 to
5 words per image.
We employ features used in the PicSOM [23] image content analysis framework.
These features convey image information using 10 different, but not necessarily uncorre-
lated, feature extraction methods. Feature vector components include: DCT coefficients
of average color in 20x20 grid (analogous to MPEG-7 ColorLayout feature), CIE LAB
color coordinates of two dominant color clusters, 16 × 16 FFT of Sobel edge image,
MPEG-7 EdgeHistogram descriptor, Haar transform of quantised HSV color histogram,
three first central moments of color distribution in CIE LAB color space, average CIE
LAB color, co-occurence matrix of four Sobel edge directions, histogram of four Sobel
edge directions and texture feature based on relative brightness of neighboring pixels.
The final image descriptor is a 682 dimensonal vector. We append a constant value 1
to each vector to learn a threshold for our linear classifiers.
001 001 144 147 001 101 001
012 012 012 012 012 012 012
296 189 189 189 296 187 189
113 113 113 113 113 104 113
013 013 013 013 152 142 013
182 182 182 182 182 182 182
174 174 174 174 174 174 174
153 153 153 153 153 153 153
Fig. 3. Example clustering results on the Corel training set. Each row consists of the closest im-
ages to the centroid of a different cluster. The number on the right of each image is the Corel CD
label. The algorithm is able to discover scenes even when there is high visual variability in the
images (e. g. people cluster, swimmers, CD-174 cluster). Some of the scenes (e. g. sunsets, peo-
ple) clearly depict scenes, even if the images are come from different CDs. (For display purposes,
portrait images were resized)
Scene discovery. First, we explore the latent space described in section 3. As mentioned
there, the cosine distance is natural to represent dissimilarity in this space. To be able
to use it for clustering we will employ graph-based methods. We expect scene clusters
to be compact and thus use complete link clustering. We look initially for many more
clusters than scene categories, and then remove clusters with a small number of images
allocated to them. We reassign those images to the remaining clusters using the closest
5 nearest neighbors. This produced approximately 1.5 clusters per CD label. For the test
set we use again the 5 nearest neighbors to assign images to the train clusters. As shown
in figure 3, the algorithm found highly plausible scene clusters, even in the presence of
034 034 034 034 010 010 103
231 046 001 017 001 001 118
276 276 276 276 276 276 148
153 153 153 120 153 153 012
113 113 113 113 113 113 108
022 101 171 384 101 384 022
161 161 161 161 161 161 161
021 021 021 021 021 021 021
119 147 119 119 147 119 119
189 187 147 201 189
Fig. 4. Example results on the Corel test set. Each row consists of the closest 7 test images to
each centroid found on the training set. The number on the right of each image is the Corel CD
label. Rows correspond to scenes, which would be hard to discover with pure visual clustering.
Because our method is able to predict word annotations while clustering scenes, it is able to
discount large but irrelevant visual differences. Despite this, some of mistakes are due to visual
similarity (e. g. the bird in the last image of the plane cluster, or the skyscraper in the last image
of the mountain cluster). (For displaying purposes, portrait images were resized).
large visual variability. This is due to the fact that these images depict objects that
tend to appear together. The algorithm also generalizes well: when the clusters were
transfered to the test set it still produced a good output (see figure 4).
Word prediction. Our approach to scene discovery is based on the internal representa-
tion of the word classifier, so these promising results suggest a good word annotation
prediction performance. Table 1 shows the precision, recall and F1-measure of our word
prediction model is competitive with the best state-of-the-art methods using this dataset.
Changing the value of in equation 3 traces out the precision-recall curve; we show the
equal error rate (P = R) result. It is remarkable that the kernelized classifier does not
provide a substantial improvement over the linear classifier. The reason for this may lie
in the high dimensionality of the feature space, in which all points are roughly at the
same distance. In fact, using a standard RBF kernel produced significantly lower re-
sults; thus the sigmoid kernel, with a broarder support, performed much better. Because
to this and the higher computational complexity of the kernelized classifier, we will use
the linear classifier for the rest of the experiments.
The influence of the tracenorm regularization is clear when the results are com-
pared to independent linear SVMs on the same features (that corresponds to using the
Frobenius norm regularization, equation 2). The difference in performance indicates
Table 1. Comparison of the performance of our word annotation prediction method with that
of Co-occurance model (Co-occ), Translation Model (Trans), Cross-Media Relevance Model
(CMRM), Text space to image space (TSIS), Maximum Entropy model (MaxEnt), Continuous
Relevance Model (CRM), 3×3 grid of color and texture moments (CT-3×3), Inference Network
(InfNet), Multiple Bernoulli Relevance Models (MBRM), Mixture Hierarchies model (MixHier),
PicSOM with global features, and linear independent SVMs on the same features. The perfor-
mance of our model is provided for the linear and kernelized (sigmoid) classifiers.* Note: the
results of the PicSOM method are not directly comparable as they limit the annotation length to
be at most five (we do not place this limit as we aim to complete the annotations for each image).
Method P R F1 Ref
Co-occ 0.03 0.02 0.02 [16]
Trans 0.06 0.04 0.05 [7]
CMRM 0.10 0.09 0.10 [9]
TSIS 0.10 0.09 0.10 [5]
MaxEnt 0.09 0.12 0.10 [10]
CRM 0.16 0.19 0.17 [11]
CT-3×3 0.18 0.21 0.19 [25]
CRM-rect 0.22 0.23 0.23 [8]
InfNet 0.17 0.24 0.23 [15]
Independent SVMs 0.22 0.25 0.23
MBRM 0.24 0.25 0.25 [8]
MixHier 0.23 0.29 0.26 [4]
This work (Linear) 0.27 0.27 0.27
This work (Kernel) 0.29 0.29 0.29
PicSOM 0.35∗ 0.35∗ 0.35∗ [23]
VN\VXQFORXGVVHD
VHD WUHHELUGVVQRZIO\ VN\VXQMHWSODQH VN\ZDWHUEHDFK PRXQWDLQVN\ZDWHU
ZDYHVELUGVZDWHU SHRSOHVDQGVDLOERDWV FORXGVSDUN
EUDQFKOHDIELUGV VN\EXLOGLQJVVPRNH VQRZWUDLQWUDFNV WUHHSHRSOH

SHRSOHVKDGRZV VN\ZDWHUWUHH
QHVW WUDLQWUDFNVORFRPR ORFRPRWLYHUDLOURDG
UDLOURDG URDGVWRQHVWDWXH EULGJHVPRNHWUDLQ
WLYHUDLOURDG VFXOSWXUHSLOODU
SLOODU WUDFNVORFRPRWLYH
UDLOURDG
Fig. 5. Example word completion results. Correctly predicted words are below each image in
blue, predicted words not in the annotations (“False Positives”) are italic red, and words not
predicted but annotated (“False Negatives”) are in green. Missing annotations are not uncommon
in the Corel dataset. Our algorithm performs scene clustering by predicting all the words that
should be present on an image, as it learns correlated words (e. g. images with sun and plane
usually contain sky, and images with sand and water commonly depict beaches). Completed
word annotations are a good guide to scene categories while original annotations might not be;
this indicates visual information really matters.
the sharing of features among the word classifiers is beneficial. This is specially true for
words that are less common.
Annotation completion. The promising performance of the approach results from its
generalization ability; this in turn lets the algorithm predict words that are not anno-
tated in the training set but should have been. Figure 5 shows some examples of word
completion results. It should be noted that performance evaluation in the Corel dataset
is delicate, as missing words in the annotation are not uncommon.
Discriminative scene prediction. The Corel dataset is divided into sets (CDs) that do
not necessarily depict different scenes. As it can be observed in figure 3, some correctly
clustered scenes are spread among different CD labels (e. g. sunsets, people). In order
to evaluate our unsupervised scene discovery, we selected a subset of 10 out of the 50
CDs from the dataset so that the CD number can be used as a reliable proxy for scene
labels. The subset consists of CDs: 1 (sunsets), 21 (race cars), 34 (flying airplanes),
130 (african animals), 153 (swimming), 161 (egyptian ruins), 163 (birds and nests),
182 (trains), 276 (mountains and snow) and 384 (beaches). This subset has visually
very disimlar pictures with the same labels and visually similar images (but depicting
different objects) with different labels. The train/test split of [7] was preserved.
To evaluate the performance of the unsupervised scene discovery method, we label
each cluster with the most common CD label in the training set and then evaluate the
scene detection performance in the test set. We compare our results with the same clus-
tering thechnique on the image features directly. In this space the cosine distance losses
Table 2. Comparison of the performance of our scene discovery on the latent space with another
unsupervised method and four supervised methods on image features directly. Our model pro-
duced significantly better results that the unsupervised method on the image features, and is only
surpassed by the supervised kernelized SVM. For both unsupervised methods, clustering is done
on the train set and performance is measured on the test set (see text for details).
Method Accuracy
Unsupervised Latent space (this work) 0.848
Unsupervised Image features clustering 0.697
Supervised Image features KNN 0.848
Supervised Image features SVM (linear) 0.798
Supervised Image features SVM (kernel) 0.948
Supervised ”structural learning” [2,18] 0.818
its meaning and thus we use the euclidean distance. We also computed the performance
of two supervised approaches on the image features: k nearest neighbors (KNN), sup-
port vector machines (SVM), and “structural learning” (introduced in [2] and used in a
vision application -Reuters image classification- in [18]). We use a one-vs-all approach
for the SVMs. Table 2 show the the latent space is indeed a suitable space for scene de-
tection: it clearly outperforms clustering on the original space, and only the supervised
SVM using a kernel provides an improvement over the performance of our method.
The difference with [18] deserves further exploration. Their algorithm classifies top-
ics (in our case scenes) by first learning a classification of auxiliary tasks (in this case
words), based in the framework introduced in [2]. [18] starts by building independent
Fig. 6. Dendrogram for our clustering method. Our scene discovery model produces 1.5 proto-
scenes per scene. Clusters belonging to the same scene are among the first to be merged
Fig. 7. Future work includes unsupervised region annotation. Example images show promising
results for region labeling. Images are presegmented using normalized cuts (red lines), features
are computed in each region and fed to our classifier as if they were whole image features.
SVM classifiers on the auxiliary tasks/words. As we showed in table 1, this leads to

lower performance in word classification when compared to our correlated classifiers.
On top of this [18] runs an SVD to correlate the output of the classifiers. It is remarkable
that our algorithm provides a slight performance advantage despite the fact [18] is su-
pervised and learns the topic classifier directly, whereas our formulation is unsupervised
and does not use topic labels.
Figure 4 depicts a dendrogram of the complete-link clustering method applied to the
clusters found by our scene discovery algorithm. As expected clusters belonging to the
same scene are among the first to be merged together. The exception is a sunset cluster
that is merged with an airplane cluster before being merged with the rest of the sunset
clusters. The reason for this is that both cluster basically depict images where the sky
occupies most of the image. Is is pleasing that “scenery” clusters depicting mountains
and beaches are merged together with the train cluster (also depicts panoramic views);
the birds and animals clusters are also merged together.
5 Conclusions
Scene discovery and classification is an important and challenging task that has impor-
tant applications in object recognition. We have introduced a principled way of defining
a meaningful vocabulary of what constitutes a scene. We consider scenes to depict cor-
related objects and present visual similarity. We introduced a max-margin factorization
model to learn these correlations. The algorithm allows for scene discovery on par with
supervised approaches even without explicitly labeling scenes, producing highly plausi-
ble scene clusters. This model also produced state of the art word annotation prediction
results including good annotation completion.
Future work will include using our classifier for weakly supervised region annota-
tion/labeling. For a given image, we use normalized cuts to produce a segmentation.
Using our classifier, we know what words describe the image. We then restrict our clas-
sifier to these word subsets and to the features in each of the regions. Figure 7 depicts
examples of such annotations. These are promising preliminary results; since quantita-
tive evaluation of this procedure requires having a ground truth labels for each segment,
we only show qualitative results.
Acknowledgements
The authors would like to thank David Forsyth for helpful discussions.
This work was supported in part by the National Science Foundation under IIS -
0534837 and in part by the Office of Naval Research under N00014-01-1-0890 as part
of the MURI program. Any opinions, findings and conclusions or recommendations
expressed in this material are those of the author(s) and do not necessarily reflect those
of the National Science Foundation or the Office of Naval Research.
References
1. Amit, Y., Fink, M., Srebro, N., Ullman, S.: Uncovering shared structures in multiclass clas-
sification. In: ICML, pp. 17–24 (2007)
2. Ando, R.K., Zhang, T.: A high-performance semi-supervised learning method for text chunk-
ing. In: ACL (2005)
3. Bosch, A., Zisserman, A., Munoz, X.: Scene classification via plsa. In: Leonardis, A.,
Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3954, pp. 517–530. Springer, Hei-
delberg (2006)
4. Carneiro, G., Vasconcelos, N.: Formulating semantic image annotation as a supervised learn-
ing problem. In: CVPR, vol. 2, pp. 163–168 (2005)
5. Celebi, E., Alpkocak, A.: Combining textual and visual clusters for semantic image retrieval
and auto-annotation. In: 2nd European Workshop on the Integration of Knowledge, Seman-
tics and Digital Media Technology, 30 November - 1 December 2005, pp. 219–225 (2005)
6. Chapelle, O., Haffner, P., Vapnik, V.: SVMs for histogram-based image classification. IEEE
Transactions on Neural Networks, special issue on Support Vectors (1999)
7. Duygulu, P., Barnard, K., de Freitas, J.F.G., Forsyth, D.A.: Object recognition as machine
translation: Learning a lexicon for a fixed image vocabulary. In: Heyden, A., Sparr, G.,
Nielsen, M., Johansen, P. (eds.) ECCV 2002. LNCS, vol. 2353, pp. 97–112. Springer, Hei-
delberg (2002)
8. Feng, S.L., Manmatha, R., Lavrenko, V.: Multiple bernoulli relevance models for image and
video annotation. In: CVPR, vol. 02, pp. 1002–1009 (2004)
9. Jeon, J., Lavrenko, V., Manmatha, R.: Automatic image annotation and retrieval using cross-
media relevance models. In: SIGIR, pp. 119–126 (2003)
10. Jeon, J., Manmatha, R.: Using maximum entropy for automatic image annotation. In: Enser,
P.G.B., Kompatsiaris, Y., O’Connor, N.E., Smeaton, A.F., Smeulders, A.W.M. (eds.) CIVR
11. Lavrenko, V., Manmatha, R., Jeon, J.: A model for learning the semantics of pictures. In:
NIPS (2003)
12. Lazebnik, S., Schmid, C., Ponce, J.: Beyond bags of features: Spatial pyramid matching for
recognizing natural scene categories. In: CVPR, pp. 2169–2178 (2006)
13. Li, F.-F., Perona, P.: A bayesian hierarchical model for learning natural scene categories. In:
CVPR, vol. 2, pp. 524–531 (2005)
14. Liu, J., Shah, M.: Scene modeling using co-clustering. In: ICCV (2007)
15. Metzler, D., Manmatha, R.: An inference network approach to image retrieval. In: Enser,
P.G.B., Kompatsiaris, Y., O’Connor, N.E., Smeaton, A.F., Smeulders, A.W.M. (eds.) CIVR
16. Mori, Y., Takahashi, H., Oka, R.: Image-to-word transformation based on dividing and vector
quantizing images with words. In: Proc. of the First International Workshop on Multimedia
Intelligent Storage and Retrieval Management (1999)
17. Oliva, A., Torralba, A.B.: Modeling the shape of the scene: A holistic representation of the
spatial envelope. International Journal of Computer Vision 42(3), 145–175 (2001)
18. Quattoni, A., Collins, M., Darrell, T.: Learning visual representations using images with
captions. In: CVPR (2007)
19. Quelhas, P., Odobez, J.-M.: Natural scene image modeling using color and texture visterms.
Technical report, IDIAP (2006)
20. Rabinovich, A., Vedaldi, A., Galleguillos, C., Wiewiora, E., Belongie, S.: Objects in context.
In: ICCV (2007)
21. Rennie, J.D.M., Srebro, N.: Fast maximum margin matrix factorization for collaborative pre-
diction. In: ICML, pp. 713–719 (2005)
22. van Gemert, J.C., Geusebroek, J.-M., Veenman, C.J., Snoek, C.G.M., Smeulders, A.W.M.:
Robust scene categorization by learning image statistics in context. In: CVPRW Workshop
(2006)
23. Viitaniemi, V., Laaksonen, J.: Evaluating the performance in automatic image annotation:
Example case by adaptive fusion of global image features. Image Commun. 22(6), 557–568
(2007)
24. Vogel, J., Schiele, B.: Natural scene retrieval based on a semantic modeling step. In: CIVR,
pp. 207–215 (2004)
25. Yavlinsky, A., Schofield, E., Rger, S.: Automated image annotation using global features and
robust nonparametric density estimation. In: Leow, W.-K., Lew, M., Chua, T.-S., Ma, W.-
Y., Chaisorn, L., Bakker, E.M. (eds.) CIVR 2005. LNCS, vol. 3568, pp. 507–517. Springer,
Heidelberg (2005)
Simultaneous Detection and Registration for Ileo-Cecal
Valve Detection in 3D CT Colonography
Le Lu1,2 , Adrian Barbu1 , Matthias Wolf2 , Jianming Liang2 ,

Luca Bogoni2 , Marcos Salganicoff2, and Dorin Comaniciu1
1
Integrated Data Systems Dept., Siemens Corporate Research, Princeton, NJ 08540
2
Computer Aided Diagnosis Group, Siemens Medical Solutions USA, Malvern, PA 19355
Abstract. Object detection and recognition has achieved a significant progress

in recent years. However robust 3D object detection and segmentation in noisy
3D data volumes remains a challenging problem. Localizing an object generally
requires its spatial configuration (i.e., pose, size) being aligned with the trained
object model, while estimation of an object’s spatial configuration is only valid at
locations where the object appears. Detecting object while exhaustively searching
its spatial parameters, is computationally prohibitive due to the high dimension-
ality of 3D search space. In this paper, we circumvent this computational com-
plexity by proposing a novel framework capable of incrementally learning the
object parameters (IPL) of location, pose and scale. This method is based on a
sequence of binary encodings of the projected true positives from the original 3D
object annotations (i.e., the projections of the global optima from the global space
into the sections of subspaces). The training samples in each projected subspace
are labeled as positive or negative, according their spatial registration distances
towards annotations as ground-truth. Each encoding process can be considered
as a general binary classification problem and is implemented using probabilis-
tic boosting tree algorithm. We validate our approach with extensive experiments
and performance evaluations for Ileo-Cecal Valve (ICV) detection in both clean
and tagged 3D CT colonography scans. Our final ICV detection system also in-
cludes an optional prior learning procedure for IPL which further speeds up the
detection.
1 Introduction
Detecting and segmenting human anatomic structures in a 3D medical image volume

(e.g., CT, MRI) is very challenging. It demonstrates different aspects of difficulties as
2D counterparts of occlusion, illumination and camera configuration variations (for in-
stance, rotation-invariant, single-view or multi-view 2D face detection [9,15,4,6,10]).
Human anatomic structures are highly deformable by nature, which leads to large intra-
class shape, appearance and pose variation. However only a limited number of patient
image volumes are available for training. Another important issue is that the pose of
the anatomic structure for detection is generally unknown in advance. If we knew the
pose as a prior, the detection problem would be easier because we can train a model
for anatomic structures under a fixed pose specification and pre-align all testing data
(w.r.t. the known pose) to then evaluate their fitness values using the learned model.

466 L. Lu et al.
However we always face a chicken-and-egg problem in practice. When estimating the

pose configuration, the structure itself must be first detected and localized because pose
information is only meaningful in the area where the object exists. In this paper, our
goal is to localize and segment an anatomic structure using a bounding box under a full
3D spatial configuration (i.e., 3D translation, 3D scaling and 3D orientation).
Exhaustive search for 3D object detection and segmentation is infeasible, due to
the prohibitive computational time required in 9D space. Naturally one would consider
restricting the search space by concatenated subspaces. Since the global optima pro-
jections are not necessarily optima in the projected subspaces, such naı̈ve projection
strategies cannot guarantee to find the global optima. In this paper, we propose a novel
learning framework to tackle this problem. In training, we encode the projections of
“global optima” in the global parameter space to a sequence of subspaces as optima
for learning. Thus the obtained classifiers can direct the searching sequentially back to
“global optima” in testing.
Our encoding process is iterative. At each stage of encoding, we extract new train-
ing samples by scanning the object’s configuration parameters in the current projected
subspace, based on previously detected candidates/hypotheses from the preceding step.
The distances of these extracted samples w.r.t. their corresponding labeled object an-
notations are then utilized to separate these training samples into positive or negative
set. This ensures the projections of the global optima represented by positives in the
subspace for training, so that the global optima can be sequentially detected through
subspaces in testing. We repeat this process until the full object configuration parameter
spaces are explored. Each encoding process is a general binary classification problem,
and is specifically implemented using probabilistic boosting tree algorithm (PBT) [12].
We demonstrate the validity of our approach with the application on 3D object detec-
tion: fully automated Ileo-Cecal Valve1 (ICV) detection in 3D computed tomography
(CT) volumes. However our technique is generally applicable to other problems as 3D
object extraction in range-scanned data [3] or event detection in spatial-temporal video
volumes [7,1]. For event detection [7,1], only subvolumes with very pre-constrained
scales and locations in video are scanned for evaluation due to computational feasibil-
ity. Our 3D detection method allows full 9 degree-of-freedoms (DOF) of searching to
locate the object/event with optimal configurations (3D for translation, 3D for rotation
and 3D for scales).
Comparing with our previous empirical approach for cardiac heart segmentation
[19], this paper develops an explicit, formal mathematical formulation for the core ob-
ject detection and parameter learning algorithm (see section 2). It also presents a more
intuitive interpretation, theoretical insights and convergence analysis in section 4. The
task of ICV detection in 3D colonography is more challenging than the organ local-
ization in [19], without considering its boundary delineation. The rest of this paper is
organized as follows. We give the mathematical formulation of proposed incremental
parameter learning (IPL) algorithm in section 2 followed by the application on ICV
1
Ileo-Cecal Valve (ICV) is a small, deformable anatomic structure connecting the small and
large intestine in human body. In addition to its significant clinical value, automated detection
of ICV is of great practical value for automatic colon segmentation and automatic detection of
colonic cancer in CT colongraphy (CTC) [11,17,5].
Simultaneous Detection and Registration 467
detection in section 3 and its evaluation in section 4. We conclude the paper with dis-
cussion in section 5.
2 Incremental Parameter Learning
For noisy 3D medical data volumes, the scanning or navigation processes of finding in-
terested objects can be very ambiguous and time-consuming for human experts. When
the searched target is partially or fully coated by other types of noisy voxels (such as
colonic objects embedded within stool, or tagging materials in CT), 3D anatomic struc-
ture detection by human experts becomes extremely difficult and sometimes impossible.
These characteristics make it very necessary to solve the type of problems using com-
puter aided detection and diagnosis (CAD) system for clinic purpose. This is the main
motivation for our paper.
The diagram of our proposed incremental parameter learning (IPL) framework is
shown in figure 1, by taking a full 3D object detection problem as an illustrative exam-
ple. We define the detection task as finding a 3D bounding box including the object in
3D data volume as closely as possible. The object’s (or the box’s) spatial configuration
space Ω can be uniquely determined by its 3D (center) position (ΩT ), 3D size (ΩS ) and
Fig. 1. Algorithm framework of incremental parameter learning (IPL) by projections in a full

3D space including 3D translations, 3D rotations (poses) and 3D scales. The parameter box on
the top row represents the ground truth, or the global optimal solution in searching. In the second
row, left, center and right boxes show how the object spatial parameters are incrementally learned
from translation, scale, to rotation. means one-to-one corresponding parameter augmentation,
and × means Cartesian product in ΩT , ΩS , ΩR parameter spaces.
468 L. Lu et al.
3D pose (rotation angles ΩR ). However the prohibitive computational expense makes

it impossible for the direct searching (ie. scanning and verifying) strategy in this total
9D space2 . To address the computational feasibility, we decompose the 9D parame-
ter searching or learning process into three 3D steps: location finding (ΩT ), followed
by size adjustment (ΩS ) and orientation estimation (ΩR ). The general searching strat-
egy in sequentially decomposed subspaces can cause undesirable, sub-optimal solutions
because the global optima are not necessary to be optimal in the decomposed dimen-
sions as well. In this paper, we propose an incremental parameter learning framework
to tackle this problem with guaranteed training performance using ROC curves analysis
of multiple steps. In each step a “detection (using the detector from previous step)-
sampling-registration-training (the detector in the current step)” scheme is applied, as
explained later. In more detail, we formulate the following incremental parameter sub-
spaces
Ω1 : {ΩT } ⊂ Ω2 : {ΩT , ΩS } ⊂ Ω3 : {ΩT , ΩS , ΩR } (1)
where Ω3 = Ω, or
Ω1 ⊂ Ω2 ⊂ ... ⊂ Ωn = Ω (2)
more generally. In equation 1, the order of ΩS , ΩR is switchable, but ΩT needs to be
first learned. The object’s size and pose configurations can only be optimized where
object is found.
For training, a set of 3D objects are labeled with their bounding boxes {T, S, R}.
Without loss of generality, we assume that there is only one true object in each 3D
data volume. In the first step, we search into ΩT by scanning n samples {T1 , T2 , ...Tn }
around the true object positions {T } and set parameters ΩS , ΩR with the mean values
S ∗ , R∗ of {S} and {R} as priors. Prior learning itself is a general and important com-
puter vision problem. The mean-value (or median) prior setting is the simplest but not
necessary the only or optimal choice of formulation, which is selected for representa-
tion clarity in this section. For example, a more natural option is prior sampling from
the distribution formed by annotation parameters. In this paper, as an optional, more
problem-specific treatment, the prior configuration of ICV detection can be learned
from its informative orifice surface profiles and other side information using the same
training/detection strategy.
First, we compute the distances dist((Ti , S ∗ , R∗ ), (Tt , St , Rt )), i = 1, 2, ..., n be-
tween each of the sampled box candidates {(T1 ,S ∗, R∗ ); (T2 , S ∗, R∗ ); ...; (Tn , S ∗ , R∗ )}
and the annotated object bounding box (Tt , St , Rt ) as its corresponding ground truth
in the same volume. The translational distance metric dist((Ti , S ∗ , R∗ ), (Tt , St , Rt ))
is computed as the center-to-center Euclidean distance
dist((Ti , S ∗ , R∗ ), (Tt , St , Rt )) = Ci − Ct (3)
where Ci is the geometrical center of the sampling box (Ti , S ∗ , R∗ ) and Ct for the
ground truth box (Tt , St , Rt ). Then the box samples {(T1 , S ∗ , R∗ ); (T2 , S ∗ , R∗ ); ...;
(Tn , S ∗ , R∗ )} are divided into positive Φ+
T if
2
Assume that the searching step is M in each dimension, and the overall cost will be M 9 . If
M = 20, the searching cost will be 512 billion times! Our target gain is M 6 here.
dist((Ti , S ∗ , R∗ ), (Tt , St , Rt )) < θ1 (4)
or negative training set Φ−

T if
dist((Ti , S ∗ , R∗ ), (Tt , St , Rt )) > θ2 (5)

−
where θ2 > θ1 . Φ+ T and ΦT are learned using our implementation of a boosting based
probabilistic binary learner (PBT [12]). Steerable features [19] are computed from
each 3D bounding box and its including volume data for PBT training. After this,
the output classifier PT is able to distinguish sampled (in training) or scanned (in
testing) object boxes: higher positive-class probability values (close to 1) for boxes
which are close to their respective labeled object boxes, lower values (close to 0) for
boxes that are distant. For computational efficiency, only top M candidates are retained
as {(T1 , S ∗ , R∗ ); (T2 , S ∗ , R∗ ); ...; (Tm

, S ∗ , R∗ )} with highest output probabilities. If
there is only one existing object per volume (such as ICV) and the training function can
be perfectly learned by a classifier, M = 1 is sufficient to achieve the correct detec-
tion. In practice, we set M = 50 ∼ 100 for all intermediate detection steps to improve
robustness. It means that we maintain multiple detected hypotheses until the final result.
We then use these M intermediate detections as a basis to search in the next step.
Each candidate (Ti , S ∗ , R∗ ), i = 1, 2, ..., M is augmented as n samples: {(Ti , S1 , R∗ );
(Ti , S2 , R∗ ); ...; (Ti , Sn , R∗ )}. Overall M × n box candidates are obtained. Similarly,
they are divided into positive Φ+ S if
dist((Ti , Sj , R∗ ), (Tt , St , Rt )) < τ1 (6)
or negative training set Φ−

S if
dist((Ti , Sj , R∗ ), (Tt , St , Rt )) > τ2 (7)
for i = 1, 2, ..., M and j = 1, 2, ..., n. dist((Ti , Sj , R∗ ), (Tt , St , Rt )) is defined as a

box-to-box distance function which formulates 3D box differences in both ΩT and ΩS .
More generally,

dist(box1 , box2 ) = { v1i − v2i }/8 (8)
i=1,2,...,8
where v1i is one of the eight vertices of box1 and v2i is its according vertex of box2 .
v1i −v2i is the Euclidean distance between two 3D vectors v1i , v2i . Again PBT algorithm
and steerable features are used for training to get PS .
In the third step, PS is employed to evaluate the positive-class probabilities for M ×n
samples {Ti, Sj , R∗ )},i=1,2,...,M; j = 1, 2, ..., n, and keep a subset of M candidates
with the highest outputs. We denote them {(Ti , Si , R∗ )}, i = 1, 2, ..., M , which are
further expanded in ΩR as {(Ti , Si , Rj )}, i = 1, 2, ..., M ; j = 1, 2, ..., n. After this,
all the process is the same for training dataset construction and classifier training PR ,
as step 2. Box-to-box distance is employed and the two distance thresholds are denoted
as η1 and η2 . Finally we have {(Tk , Sk , Rk )}, k = 1, 2, ...M returned by our whole
algorithm as the object detection result of multiple hypotheses. In testing, there are
470 L. Lu et al.
three searching steps in ΩT , ΩS and ΩR , according to the training procedure. In each

step, we can scan and detect 3D object box candidates which are close to the global
optimum (i.e., the object’s true spatial configuration) in the current parameter subspace
(Ω1 → Ω2 → Ω3 ), using the learned classifier (PT , PS or PR ) respectively. The out-
put candidates are used as seeds of propagation in the next stage of incremental, more
accurate parameter optimization. The training samples at each step are expanded and
bootstrapped using the detection results at its previous step (and the global annotations
as reference). Note that we set smaller threshold margins,
(θ2 − θ1 ) > (τ2 − τ1 ) > (η2 − η1 ) (9)
for more desirable object detection/registration accuracy as steps of detection proceed.
The above incremental parameter learning process for 3D object detection is illus-
trated in figure 1. The parameter spaces (ΩT , ΩS and ΩR ) before search (prior), during
search (learning/optimizing) and after search (optimized) are displayed in red, yellow
and white shadows respectively. The mean parameter values T ∗ , S ∗ , R∗ estimated from
the labeled object annotations, are used as prior by default.
3 Ileo-Cecal Valve (ICV) Detection in 3D CT Colonography

Detecting Ileo-Cecal Valve (ICV) in 3D CT volumes is important for accurate colon
segmentationand colon polyp false positive reduction [11,17,5] that are required by
colon CAD system. Nevertheless, it is very challenging in terms of ICV’s huge varia-
tions in its internal shape/appearance and external spatial configurations: (X, Y, Z; Sx,
Sy , Sz ; ψ, φ, ω), or (ΩT ; ΩS ; ΩR ;). ICV is a relatively small-scaled (compared with
heart, liver, even kidney) and deformable human organ which opens and closes as a
valve. The ICV size is sensitive to the weight of patient and whether ICV is diseased. Its
position and orientation also vary of being a part of colon which is highly deformable.
To address these difficulties, we develop a two-staged approach that contains the prior
learning of IPL to prune ICV’s spatial configurations in position and orientation, fol-
lowed by the position, size and orientation estimation of incremental parameter learn-
ing. Figure 2 shows the diagram of our final system. To validate the proposed incremen-
tal parameter learning of ΩT ΩS ΩR , an ICV detection system without prior learning
is also experimentally evaluated.
3.1 Features
In the domain of 3D object detection, 3D Haar wavelet features [13] are designed to
capture region-based contrasts which is effective to classification. However 3D Haar
features are inefficient for object orientation estimation because they require a very
time-consuming process of rotating 3D volumes for integral volume computation. In
steerable features [19], only a sampling grid-pattern need to be translated, rotated and
re-scaled instead of data volumes. It allows fast 3D data evaluation and has shown to
be effective for object detection tasks [19]. It is composed by a number of sampling
grids/points where 71 local intensity, gradient and curvature based features are com-
puted at each grid. The whole sampling pattern models semi-local context. For details,
refer to [19].
Fig. 2. System diagram of Ileo-Cecal Valve detection. The upper block is prior learning and the
lower block is incremental parameter learning for ICV spatial parameter estimation. Examples of
the annotated ICV bounding boxes are shown in red.
Fig. 3. Steerable sampling grid patterns for (a) 3D point detector and (b) 3D box detector
In this paper, we design two specific steerable patterns for our ICV detection task as
shown in figure 3. In (a), we design an axis-based pattern for detecting ICV’s orifice.
Assume that the sampling pattern is placed with its center grid at a certain voxel v. It
contains three sampling axes as the gradient directions averaged in v’s neighborhoods
under three scales respectively. Along each axis, nine grids are evenly sampled. This
process is repeated for halfly and quarterly downsampled CT volumes as well. Alto-
gether we have M = 81 = 3 × 9 × 3 grid nodes which brings 71 × 81 = 5751 features.
In (b), we fit each box-based pattern with evenly 7 × 7 × 5 sampling grids. The total
feature number is 52185 by integrating features from three different scales. This type of
feature is used for all ΩT ΩS ΩR detection. The detector trained with axis pattern and
PBT is named 3D point detector; while the detector with box pattern and PBT is noted
as 3D box detector.
3.2 Prior Learning in ΩT and ΩR of IPL

If likely hypotheses ICV’s orifice can be found, its position in ΩT can be constrained,
then no explicitly exhaustive searching of position is needed. The ICV orifice has an
472 L. Lu et al.
(a) (b)
Fig. 4. (a) ICV orifice sampling pattern of three sampling axes and nine sampling grids along
each axis; (b) detected ICV voxel/orifice candidates shown in white
informative, but far from fully unique, surface profile that can possibly indicates ICV
location as multiple hypotheses. It also allows very efficient detection using a 3D point
detector which involves less feature computation (5751 vs. 52185 for training) than a
box detector. Further more, it is known that ICV orifice only lies on the colon surface
that is computed using a 3D version of Canny edge detection. Thus we can prune all
voxel locations inside the tissue or in the air for even faster scanning. An illustrative
example of the orifice sampling pattern and detection result is shown in figure 4. Note
that multiple clusters of detection may occur often in practice. From the annotated ICV
orifice positions in our training CT volume set, we generate the positive training sam-
ples for surface voxels within α1 voxel distance and negatives out of α2 voxel distance.
We set α2 > α1 , so the discriminative boosting training [12] will not focus on samples
with distances [α1 , α2 ] which are ambiguous for classifier training but not important for
target finding. The trained classifier PO is used to exhaustively scan all surface voxels,
prune the scanned ICV orifice candidates and only a few hypotheses (eg. N = 100)
are preserved. In summary, 3D point detector for ICV orifice detection is efficient and
suitable for exhaustive search as the first step.
Given any detected orifice hypothesis, we place ICV bounding boxes centering at
its location and with the mean size estimated from annotations. In the local 3D coor-
dinates of an ICV box, XY plane is assumed to be aligned with the gradient vector
of the orifice as its Z-axis. This is an important domain knowledge that we can use to
initially prune ICV’s orientation space ΩR in 2 degrees of freedom (DOF). Boxes are
then rotated around Z-axis with 10o interval to generate training samples. Based on
their box-to-box distances against the ground truth of ICV box3 and β1 , β2 threshold as
above, our routine process is: (1)generating positive/negative training sets by distance
thresholding; (2) training a PBT classifier PR using the box-level steerable features;
(3) evaluating the training examples using the trained classifier, and keeping top 100
hypotheses of probabilities (ρiR , i = 1, 2, ..., 100). In our experiments, we show results
with α1 = 4, α2 = 20 (normally out of the ICV scope), β1 = 6 and β2 = 30.
3
The ground truth annotations are normalized with the mean size to count only the translational
and orientational distances.
3.3 Incremental Parameter Learning in ΩT ΩS ΩR

In this section, we search for more accurate estimates of ICV position, scale and ori-
entation parameter configurations. Incremental parameter learning method described in
section 2 is implement. The box-level steerable features (as shown in figure 3(b)) and
PBT classifier are employed for all three steps. From section 3.2 we obtain 100 ICV
box hypotheses per volume with their positions and orientations pruned. Therefore we
select the order of incremental parameter learning as ΩT → ΩS → ΩR , where ΩT is
always the first step to locate itself and ΩS proceeds before aligned ΩR .
First, the position of each of the N hypotheses is shifted every one voxel in the range
of [−20, 20] of all X, Y and Z coordinates (ie. ΩT + ΔT ). This set of synthesized ICV
box samples is then splitted into the positive (< θ1 = 5 voxel distance) and negative
(> θ2 = 25 voxel distance) training sets for the PBT training of PT . Again the top 100
ICV box candidates in each CT volume (with the largest probability outputs ρiT using
PT ) are maintained. Next, the optimal estimates of ICV box scales are learned. We set
the size configuration of each survived hypotheses in ΩS , evenly with 2 voxel intervals
from the range of [23, 51] voxels in X, [15, 33] voxels in Y and [11, 31] voxels in Z
coordinates. The ranges are statistically calculated from the annotated ICV dataset.
In the same manner, we train the classifier PS and use it to obtain the top N candi-
dates of ρiS with more accurate estimates of ΩS . The distance thresholds are τ1 = 4 and
τ2 = 20 for positive/negative training respectively. Last, we adaptively add disturbances
from the previously aligned orientation estimates in prior learning (ie. ΩR + ΔR ). ΔR
varies with 0.05 intervals in [−0.3, 0.3] radians, 0.1 in ([−0.9, −0.3), (0.3, 0.9]) and 0.3
in ([−1.8, −0.9), (0.9, 1.8]). This strategy provides a finer scale of searching when closer
to the current orientation parameters (retained from PR in prior learning), to improve the
ΩR detection accuracy. PR is learned with the distance thresholds as η1 = 4 and η2 = 15.
After all steps of incremental parameter learning, the top one box candidate of the highest
probability value from PR is returned as the final ICV detection result by default.
Incremental parameter learning of ΩT , ΩS , ΩR is equivalent to exhaustive search in
ΩT ∪ΩS ∪ΩR if we can train mathematically perfect classifiers (100% recall at 0% false
positive rate) at all steps. This causes large positive within-class variations at early learn-
ing steps (e.g., detecting object location while tolerating unestimated poses and scales),
which decreases trainability in general. Classifiers with intrinsic “divide-and-conquer”
scheme as PBT [12] or cluster based tree [14] can be applied. In short, explicit ex-
haustive searching for parameter estimation is traded by implicit within-class variation
learning using data-driven clustering [12,14]. It also relaxes the requirement for training
accuracy by keeping multiple hypotheses during detection. In case of multiple object
detection, selecting top N candidates simply based on their class-conditional probabil-
ities can not guarantee to find all objects since a single target may cause many detec-
tions. Possible approaches are to exploit cluster based sampling [8] or Non-Maximum
Suppression by using the spatial locations of detected hypotheses.
4 Evaluation and Results

Convergence Analysis: The convergence analysis of incremental parameter learning
method is first based on the property of Receiver Operating Characteristic (ROC) curves
474 L. Lu et al.
during five stages of training. The training scale for our PBT classifier ranges over
10K ∼ 250K positives and 2M ∼ 20M negatives. The ROC curves are shown in
figure 5 (a). From the evidence of these plots, our training process are generally well-
performed and gradually improves for later steps. We then discuss the error distribution
curves between the top 100 ICV hypotheses maintained for all five stages of detection
and the ground truth, using five-fold cross-validation. The error curves, as shown in
figure 5 (b), also demonstrate that more accurate ICV spatial configurations can be ob-
tained as the detection process proceed through stages. This convergence is bounded by
the good training performance of ROC curves with positive-class distance boundaries
that are gradually more close to the global optima (or ground-truth) as 6, 5, 4, 4, and
(a) (b)
(c) (d)
Fig. 5. (a) Receiver operating characteristic curves of different stages of training in our Ileo-Cecal
Valve detection system. (b) Error ratio curves of top 100 ICV hypotheses of different stages of
detection. Each curve show the ratios of hypotheses (Y axis) under the particular error readings
(X-axis) against ground truth. All numbers are averaged over the testing sets of volumes, under
five-fold cross-validation of 116 total labeled ICV examples. (c) Overlap ratios between 114
detected ICV examples and their ground truth. (d) A typical example of 3D ICV detection in CT
Colonography, with overlap ratio of 79.8%. Its box-to-box distance as define in equation 8 is
3.43 voxels where the annotation box size is 29.0 × 18.0 × 12.0 voxels. Its orientational errors
are 7.68o , 7.77o , 2.52o with respect to three axes. The red box is the annotation; the green box is
the detection. This picture is better visualized in color.
decreasing distance margins between positive and negative classes (eg. β2 − β1 = 24;
θ2 − θ1 = 20; τ2 − τ1 = 16 and η2 − η1 = 11) over stages.
ICV Detection Evaluation: Our training set includes 116 ICV annotated volumes from
the dataset of clean colon CT volumes using both Siemens and GE scanners. With a
fixed threshold ρR > 0.5 for the final detection, 114 ICVs are found with the detection
rate of 98.3%, under five-fold cross-validation. After manual examination, we find that
the two missed ICVs have very abnormal shape from the general training pool which is
probably heavily diseased. The ICV detection accuracy is first measured by a symmetric
overlapping ratio between a detected box Boxd and its annotated ground truth Boxa
;
2 × V ol(Boxa Boxd )
γ(Boxa , Boxd ) = (10)
V ol(Boxa ) + V ol(Boxd )
where V ol() is the box-volume function (eg. the voxel number inside a box). The accu-
racy distribution over 114 detected ICV examples is shown in 5 (c). The mean overlap
ratio γ(Boxa , Boxd ) is 74.9%. This error measurement is directly relevant with our
end goal of removing polyp-like false findings in my CAD system. Addtionally the
mean and standard deviation of orientational detection errors are 5.89o , 6.87o, 6.25o ;
and 4.46o , 5.01o , 4.91o respectively for three axes. The distribution of absolute box-
box distances (ie. equation 8) has 4.31 voxels as its mean value, and 4.93 voxels for
the standard deviation. Two missed cases are further verified by clinician as heavily
diseased ICVs which are rare in nature. Our trained classifiers treat them as outliers.
Next we applied our detection system to other previously unseen clean and tagged
CT datasets. For clean data, 138 detections are found from 142 volumes. After manual
validation, 134 detections are true ICVs and 4 cases are Non-ICVs. This results a detec-
tion rate of 94.4%. We also detected 293 ICVs from 368 (both solid and liquid) tagged
colon CT volumes where 236 detections are real ICVs with 22 cases for Non-ICVs and
35 cases unclear (which are very difficult even for expert to make decision). Tagged CT
data are generally much more challenging than clean cases, under low-contrast imaging
and very high noise level of tagging materials. Some positive ICV detections are illus-
trated in figure 6. The processing time varies from 4 ∼ 10 seconds per volume on a P4
3.2G machine with 2GB memory.
Without prior learning for ICV detection, our system can achieve comparable detec-
tion performance as with prior learning. However it requires about 3.2 times more com-
putation time by applying a 3D box detector exhaustively on translational search, not
a cheaper 3D point detector as in prior learning. Note that prior learning is performed
in the exact same probabilistic manner as the incremental 3D translation, scale and ori-
entation parameter estimation. It is not a simple and deterministic task, and multiple
(e.g., 100) detection hypotheses are required to keep for desirable results.
Polyp False Positive (FP) Deduction: ICV contains many polyp-like local structures
which confuse colon CAD system [11,17,5]. By identifying a reasonably accurate
bound box for ICV, this type of ambiguous false positive polyp candidates can be re-
moved. For this purpose, we enhanced the ICV orifice detection stage by adding the
labeled polyp surface voxels into its negative training dataset. Other stages are conse-
quentially retained in the same way. Polyp FP deduction is tested on 802 unseen CT
476 L. Lu et al.
(a) (b) (c) (d)
(e) (f) (g) (h)

Fig. 6. (a,b) An example of ICV detection result from two viewpoints. The red box is the annota-
tion; the green box is the detection. (c,d,e,f,g,h) Examples of ICV detection results from unseen
clean colon CT volumes (c,d,e) and unseen solid (f) or liquid tagged (g,h) colon CT volumes. The
red box is the final detection result where no annotation available. Note that only a CT subvol-
ume surrounding the detected ICV box is visualized for clarity. This picture is better visualized
in color.
volumes: 407 clean volumes from 10 different hospital sites acquired on Siemens and
GE scanners; 395 tagged volumes, including iodine and barium preparations, from 2
sites acquired on Siemens and GE scanners. The ICV detection is implemented as post
filter for our existing colon CAD system and only applied on those candidates that are
labeled as “Polyp” in the preceding classification phases4 . In clean cases, ICV detec-
tion reduced the number of false positives (fp) from 3.92 fp/patient (2.04 fp/vol.) to
3.72 fp/patient (1.92 fp/vol.) without impacting the overall sensitivity of the CAD sys-
tem. It means that no true polyps were missed due to our ICV detection component
integrated. In tagged cases, ICV detection reduced the number of false marks from 6.2
fp/patient (3.15 fp/vol.) to 5.78 fp/patient (2.94 fp/vol.). One polyp out of 121 polyps
with a size range from 6 up to 25 mm was wrongly labeled as ICV, resulting in a sen-
sitivity drop of 0.8%. Another version implementation of using ICV detection as a soft
constraint, instead of a hard-decisioned post filter, avoids true polyp missing without
sacrificing FP Deduction. In summary our ICV system achieved 5.8% and 6.7% false
positive deduction rates for clean and tagged data respectively, which has significant
clinical importance.
Contextual K-Box ICV Model: To more precisely identify the 3D ICV region be-
sides detection, a contextual K-box model is experimented. The idea is using the final
ICV detection box B1 as an anchor to explore reliable expansions. For all other high
probability hypotheses {; B̂i } returned in the last step of detection, we sort them ac-
cording to V ol(B̂i − B1 B̂i ) while two constraints are satisfied: γ(B;1 , B̂i ) ≥ γ1 and
ρR (B̂i ) ≥ ρ1 . Then the box that gives the largest gain of V ol(B̂i −B1 B̂i ) is selected
4
Note that the use of ICV detection as post-process is dedicated to handle “difficult” polyp cases
which can not be correctly classified in preceding processes.
as the second box B2 . The two constraints guarantee that B2 is spatially correlated with
B1 (γ1 = 0.5) and is a highly likely ICV detection
hypothesis by itself ρ1 = 0.8. By
taking B1 and B2 as a union Boxd = B1 B2 , it is straightforward to expand the
model for K-box ICV model while K > 2. Our initial experimental results show that
2-box model improves the mean overlap ratio γ(Boxa , Boxd ) from 74.9% to 88.2%
and surprisingly removes 30.2% more Polyp FPs without losing true polyps.
Previous Work on ICV Detection: Our proposed approach is the first reported, fully
automatic Ileo-Cecal Valve detection system in 3D CT colonography, due to the dif-
ficulties discussed in sections 1 and 3. The closest previous work is by Summer et al.
[11] that is also considered as the state-of-art technique in medical imaging community.
We discuss and compare [11] and our work in two aspects. (1) For localization of ICV,
Summer et al. relies on a radiologist to interactively identify the ICV by clicking on
a voxel inside (approximately in the center of) the ICV. This is a requisite step for the
next classification process and takes minutes for an expert to finish. On the contrary, our
automatic system takes 4 ∼ 10 seconds for the whole detection procedure. (2) For clas-
sification, [11] primarily designs some heuristic rules discovered from dozens of cases
by clinicians. It depends on the performance of a volume segmentor [16] which fails
on 16% ∼ 38% ICV cases [11]. Their overall sensitivity of ICV detection is 49% and
50% based on the testing (70 ICVs) and training datasets (34 ICVs) [11], respectively.
This rule based classification method largely restricts its applicability and effectiveness
on recognizing varieties of ICV samples with their low detection rates reported in [11].
Our detection rate is 98.3% for training data and 94.4% for unseen data. The superiority
of our approach attributes to our effective and efficient incremental parameter learning
framework optimizing object spatial configuration in a full 3D parameter space, and the
discriminative feature selection algorithm (PBT + steerable features) exploring hun-
dreds of thousands volume features.
5 Conclusion and Discussion

In this paper, we present the incremental parameter learning framework to address gen-
eral 3D/2D object detection problem under high dimensional parameter spaces. The
challenges are not only the computational feasibility, but also how to obtain good so-
lutions in terms of the parameter searching complexity (essentially exponential to the
dimension). The effectiveness of our method is demonstrated using an application on
detecting Ileo-Cecal Valve (ICV) in 3D CT colonography with 9 DOF. To our best
knowledge, ICV detection is the first fully automatic system for localizing a small (ver-
sus the whole CT volume dimension), largely deformable, unconstrainedly posed and
possibly coated (by tagging material or stool in tagged volumes) 3D anatomic structure.
As a discussion, our proposed learning architecture is intuitively analogical to the
famous twenty questions games, where many highly complex information extraction
problems can be solved by using a flow of simpler, binary (yes/no), sequentially de-
pendent testings (question vs. answer). We leave explorations on more sophisticated
solution searching techniques [2,18] as future work.
478 L. Lu et al.
References
1. Blank, M., Gorelick, L., Shechtman, E., Irani, M., Basri, R.: Actions as Space-Time Shapes.
In: ICCV (2005)
2. Geman, D., Jedynak, B.: An Active Testing Model for Tracking Roads in Satellite Images.
IEEE Trans. Pattern Anal. Mach. Intell. 18(1), 1–14 (1996)
3. Han, F., Tu, Z., Zhu, S.C.: Range Image Segmentation by an Effective Jump-Diffusion
Method. IEEE Trans. PAMI 26(9) (2004)
4. Huang, C., Ai, H., Li, Y., Lao, S.: High-performance rotation invariant multiview face detec-
tion. IEEE Trans. PAMI 29(4), 671–686 (2007)
5. Jerebko, A., Lakare, S., Cathier, P., Periaswamy, S., Bogoni, L.: Symmetric Curvature Pat-
terns for Colonic Polyp Detection. In: Larsen, R., Nielsen, M., Sporring, J. (eds.) MICCAI
6. Jones, M., Viola, P.: Fast multi-view face detection. In: CVPR (2003)
7. Ke, Y., Sukthankar, R., Hebert, M.: Efficient Visual Event Detection using Volumetric Fea-
tures. In: ICCV (2005)
8. Lu, L., Hager, G.: Dynamic Background/Foreground Segmentation From Images and Videos
using Random Patches. In: NIPS (2006)
9. Rowley, H., Baluja, S., Kanade, T.: Neural Network-Based Face Detection. In: CVPR (1996)
10. Rowley, H., Baluja, S., Kanade, T.: Rotation Invariant Neural Network-Based Face Detec-
tion. In: CVPR (1998)
11. Summers, R., Yao, J., Johnson, C., Colonography, C.T.: with Computer-Aided Detection:
Automated Recognition of Ileocecal Valve to Reduce Number of False-Positive Detections.
Radiology 233, 266–272 (2004)
12. Tu, Z.: Probabilistic boosting-tree: Learning discriminative methods for classification, recog-
nition, and clustering. In: ICCV (2005)
13. Tu, Z., Zhou, X.S., Barbu, A., Bogoni, L., Comaniciu, D.: Probabilistic 3D polyp detection
in CT images: The role of sample alignment. In: CVPR (2006)
14. Wu, B., Nevatia, R.: Cluster Boosted Tree Classifier for Multi-View, Multi-Pose Object De-
tection. In: ICCV (2007)
15. Viola, P., Jones, M.: Rapid object detection using a boosted cascade of simple features. In:
CVPR, pp. 511–518 (2001)
16. Yao, J., Miller, M., Franaszek, M., Summers, R.: Colonic polyp segmentation in CT
Colongraphy-based on fuzzy clustering and deformable models. IEEE Trans. on Medical
Imaging (2004)
17. Yoshida, H., Dachman, A.H.: CAD techniques, challenges, and controversies in computed
tomographic colonography. Abdominal Imaging 30(1), 26–41 (2005)
18. Yuille, A.L., Coughlan, J.M.: Twenty Questions, Focus of Attention, and A*: A Theoretical
Comparison of Optimization Strategies. In: Pelillo, M., Hancock, E.R. (eds.) EMMCVPR
19. Zheng, Y., Barbu, A., Georgescu, B., Scheuering, M., Comaniciu, D.: Fast Automatic Heart
Chamber Segmentation from 3D CT Data Using Marginal space Learning and Steerable
Features. In: ICCV (2007)
Constructing Category Hierarchies
for Visual Recognition
Marcin Marszalek and Cordelia Schmid
INRIA Grenoble, LEAR, LJK

{marcin.marszalek,cordelia.schmid}@inrialpes.fr
Abstract. Class hierarchies are commonly used to reduce the complex-

ity of the classification problem. This is crucial when dealing with a large
number of categories. In this work, we evaluate class hierarchies cur-
rently constructed for visual recognition. We show that top-down as well
as bottom-up approaches, which are commonly used to automatically
construct hierarchies, incorporate assumptions about the separability of
classes. Those assumptions do not hold for visual recognition of a large
number of object categories. We therefore propose a modification which
is appropriate for most top-down approaches. It allows to construct class
hierarchies that postpone decisions in the presence of uncertainty and
thus provide higher recognition accuracy. We also compare our method
to a one-against-all approach and show how to control the speed-for-
accuracy trade-off with our method. For the experimental evaluation, we
use the Caltech-256 visual object classes dataset and compare to state-
of-the-art methods.
1 Introduction
Visual object classification is one of the basic computer vision problems. In spite
of significant research progress, the problem is still far from being solved and a
considerable effort is still being put into this research area [1].
In the last years, one could witness remarkable progress in the development of
robust image representations and also observe successful applications of sophis-
ticated machine learning techniques in computer vision. Developments in image
representation include research on interest point detectors [2,3], SIFT features [4]
and bag-of-features [5]. Support Vector Machines (SVMs) [6] were successfully
applied to vision with the design of specialized kernels [7,8]. Combining these
techniques allowed researchers to construct successful visual object recognition
systems [1]. We build on those works to construct our baseline.
Still, the typical problems that are tackled today by the state-of-the-art vi-
sual object class recognition systems, consist of only few object categories. Very
recently, datasets that include more than a hundred of categories, like the most
recent Caltech datasets [9,10], have been introduced. Furthermore, there is an
obvious need to further increase this number. In this paper we examine the
problem of classifying a large number of categories and use the Caltech-256 [10]
dataset for evaluation. Figure 1 shows a few sample images.

480 M. Marszalek and C. Schmid
rifle sword touring-bike mountn-bike zebra tower-pisa

(0%) (0%) (37%) (45%) (86%) (91%)
a) most difficult b) most confused c) easiest
Fig. 1. Sample Caltech-256 images for the most difficult (left), the most confused
(middle) and the easiest (right) classes are shown. In parentheses the per-class accuracy
of our method is given.
Multi-class classification problems are often handled by combining multiple

binary classifiers. Common setups for binary SVMs are based on competition
(one-against-rest), voting (one-against-one) or discarding subsequent hypotheses
(DAG-SVM). Unfortunately, this means linear (OAR, DAG) or square (OAO)
complexity in the number of classes and therefore does not scale well. In principle,
a direct multi-class approach is also possible for SVMs, but the optimization
problem becomes significantly more difficult and in spite of longer training times,
the accuracy is comparable to the one achieved with an OAR approach [6].
To deal with a large number of categories, many approaches combine binary
classifiers using class hierarchies. This usually results in logarithmic complexities.
The question of how to build such a hierarchy remains, however, open. Approaches
common in computer vision can be divided into two groups. First, the hierarchy
can be constructed top-down by recursive partitioning of the set of classes. To
find a decomposition, Chen et al. [11] used a sophisticated relaxed max-cut for-
mulation, while Liu et al. [12] simply employed k-means clustering. Second, the
hierarchy can be built bottom-up by agglomerative clustering. Zhigang et al. [13]
explored this approach. In principle, hierarchies could also be found by exhaus-
tive search or random sampling followed by cross-validation. Yuan et al. [14] com-
pared hierarchies found with exhaustive search with the ones constructed by a
k-means based top-down method. For a small number of categories, using a top-
down method resulted in performance comparable to employing an exhaustive
search. For a large number of categories, the exponential growth of possible
Constructing Category Hierarchies for Visual Recognition 481
solutions prohibits the use of exhaustive or random approaches, so we do not in-

clude them into our further consideration.
Class hierarchies can address the limitation of current systems to handle a
large number of object categories, but they can be used for visual recognition
in other ways as well. For example, Zweig and Weinshall [15] exploited class
hierarchies to combine models form different category levels, whereas He and
Zemel [16] used them to cope with missing and roughly-specified annotations. As
the need for class hierarchies increases, the purpose of this paper is to evaluate
the suitability of currently constructed hierarchies for visual recognition. We
observe that even the most recent methods tend to model class hierarchies with
trees [17]. As we will show, this imposes a hard constraint that leads to separation
problems when the number of categories increases. We propose a simple yet
powerful solution based on the relaxation of this constraint and the possibility of
postponing uncertain decisions until they can be reliably made. Furthermore, we
address the classification complexity in the number of classes by demonstrating
how one can control speed-for-accuracy trade-off with our method.
Hierarchical methods are also used at lower levels. Approaches like vocabulary
trees [18] that speed up feature matching are related to our work due to their hi-
erarchical nature. Similarly, kd-trees are hierarchic space partitioning structures
that can perform component-wise classification [19]. Note, however, that in this
work we focus on building high-level class hierarchies and look into the problem
of class-wise partitioning.
The rest of the paper is organized as follows. In Sect. 2 we evaluate existing
approaches for constructing class hierarchies. In Sect. 3 we propose a novel ap-
proach that avoids the separation problem present in existing methods. In Sect. 4
we experimentally confirm our findings and demonstrate the speed-for-accuracy
trade-off of our method. We conclude the paper in Sect. 5.
2 Existing Approaches
In this section we assume that some dissimilarity measure between classes is

given. Common approaches are to simply compute the distances between class
means [12,14] or to represent the classes in a high dimensional space using a
Gaussian kernel [13,20].
As discussed in the introduction, we can divide the existing methods for build-
ing class hierarchies into two main groups. In Subsect. 2.1 we consider some com-
monly used methods that construct the class hierarchy by top-down recursive
partitioning of the set of classes. In Subsect. 2.2 we discuss methods based on
bottom-up agglomerative clustering of the classes.
Given a class hierarchy, we can efficiently classify samples by descending the
resulting decision tree. In principle, any classifier could be used in the nodes of
the hierarchy to make the decision about the direction of descent. In practice,
Support Vector Machines are widely used for this task in computer vision.
Most often the hierarchies are represented as binary trees, which means that
at each node a binary decision is made on which of the two subtrees to choose.
Thus, a Support Vector Machine for each node of the tree can be trained. If
the tree is balanced, only &log2 N ' SVM runs are necessary to perform the N -
class classification. In the worst case (degenerated trees) the complexity is linear
in the number of classes. Therefore, in general, hierarchy-based classification
approaches scale well with the number of classes.
2.1 Top-Down Recursive Partitioning

K-means clustering. A set of classes can be clustered into k groups. This
determines the partitioning at a given level. When applied recursively, this simple
yet popular [12,14,20] method allows to construct a class hierarchy. K-means
clustering minimizes the distances to cluster centers, thus tries to find compact
clusters. This presumably leads to well separated clusters.
Normalized cuts. A dataset can be viewed as a fully connected undirected

graph GV = (V, EV ), where v ∈ V nodes correspond to the elements of the
dataset and edge weights k(v1 , v2 ) ∈ EV correspond to the similarity measure
between the elements. This is the starting point of many spectral clustering
algorithms. A graph cut partitions GV into GA and GB , where A ⊂ V and
B ⊂ V are two disjoint sets of nodes A B = V , i.e., A ∪ B = V and A ∩ B = ∅.
Shi and Malik [21] proposed the Normalized Cuts method to find a good cut
through such a graph. The idea is to minimize
assoc(A, B) assoc(A, B)
Ncut(A, B) = + (1)
assoc(A, V ) assoc(B, V )
where assoc(A, B) is the weight of all edges connecting the nodes between the
sets A and B, i.e.,
assoc(A, B) = k(a, b) . (2)
a∈A,b∈B
Note that assoc(A, B) is often denoted in the literature as cut(A, B). As the
distance measures used in spectral clustering are often positive definite, the
adjacency matrix EV is often denoted as K. The common choice is the RBF
kernel, which can be generalized to an extended Gaussian kernel
k(vi , vj ) = Kij = e− γ m(vi ,vj )

1
(3)
where m(vi , vj ) is the distance measure between the elements.
Finding the optimal normalized cut is NP-hard, therefore the following relax-
ation is commonly used:
wT D− 2 KD− 2 w
1 1
w∗ = arg max (4)

w wT w
such that
wT D 1 = 0 (5)

where D is a diagonal matrix with dii = j Kij and 1 is the vector of all ones.
The optimal w∗ can be found by computing the eigenvector corresponding to the
second largest eigenvalue of D− 2 KD− 2 . The sgn(wi∗ ) indicates whether vi ∈ A

1 1
or vi ∈ B. As it was shown by Rahimi and Recht [22], this relaxation can be

interpreted as finding a maximal hyperplanar gap.
By recursively partitioning the graph GV where V is a set of classes, one can
obtain a class hierarchy [11].
2.2 Bottom-Up Agglomerative Clustering
Given a distance measure between classes, agglomerative clustering can be used

to build a class hierarchy bottom up. Initially, all the classes belong to different
clusters. Then, the closest clusters are merged. It is assumed that merging close
elements will result in clusters that are better separated. The distances to the
new-formed class clusters are recomputed, such that the procedure can be applied
iteratively until all classes belong to one cluster. The merge order determines
the hierarchy. If during each step one looks for the two most similar clusters,
the hierarchy will have a form of a binary tree. Zhigang et al. [13] explored both
binary trees and k-trees.
2.3 Discussion
Most existing class hierarchy construction methods assume that at each level of
the hierarchy the feature-space can be partitioned into disjoint subspaces. We
predict an inevitable conflict between generalization and precision requirements.
Especially for the earliest decisions, where the boundary is supposed to split
very distinct categories of objects (natural vs. man-made objects for example),
a requirement is enforced to precisely trace the boundaries between tens or
hundreds of similar classes that fall at the explored decision boundary (a bear
vs. a teddy-bear and a fountain vs. waterfall for example). Note that a mistake
at a boundary of such a high-level decision is as costly as a mistake at lower
levels, where the classifier can tune to minor class differences without degrading
its generalization properties.
Given a few distinct visual object categories class separability can be good.
But this certainly cannot hold for hundreds or thousands of classes. Let us
motivate our hypothesis with some simplified examples before evaluating it ex-
perimentally.
Figure 2 presents some simplistic efforts to separate 2-dimensional multi-class
data with a linear boundary. A carefully crafted example (Fig. 2a) shows, that
even if any two of three classes can be easily separated with a hyperplane, it does
not assure good separation of all three classes. If there are few classes which are
well separated (Fig. 2b), a good recursive partitioning can be found. With the
growing number of classes, however, it will be increasingly difficult to find a
disjoint class-set partitioning (Fig. 2c).
As we show in Sect. 4, early enforcement of hard decisions can be costly in
the hierarchic setup and can significantly lower the classification performance.
Thus, we propose a novel approach for constructing top-down hierarchies, which
postpones final decisions in the presence of uncertainty.
a) crafted example b) few classes c) many classes
Fig. 2. Simple examples of separating 2-dimensional multi-class data with a linear de-
cision boundary. Difficulties to separate classes (left) might not arise for a few separated
classes (middle), but can emerge when the number of classes increases (right).
3 Our Approach
Our approach is based on the observation that finding a feature-space partition-
ing that reflects the class-set partitioning becomes more and more difficult with
a growing number of classes. Thus, we propose to avoid disjoint partitioning and
split the class-set into overlapping sets instead. This allows to postpone uncer-
tain classification decisions until the number of classes gets reduced and learning
good decision boundaries becomes tractable.
The proposed solution is to discover classes that lie on the partition boundary
and could introduce classification errors. Those classes should not be forced
into either of the partitions, but they should be included in both. With our
approach, a number of classes can still be separated with one decision. This
assures a computational gain compared to setups with linear complexity like
OAR. However, since disjoint partitioning is not enforced, the performance is
not degraded. As the resulting partitioning is relaxed, we call our hierarchy
Relaxed Hierarchy (RH).
Figure 3 demonstrates how our method applies to the problem sketched in
Subsect. 2.3. The boundary from Fig. 2a which separates members of a class
can be used if both subpartitions (Fig. 3a) contain this class. Moreover, the
subsequent splits are straightforward. Note that the resulting hierarchy (Fig. 3b)
is no longer a tree, but a rooted directed acyclic graph (DAG).
Our method can be applied to most top-down partitioning approaches. This
includes methods based on k-means clustering and normalized cuts. Here we
build on normalized cuts. Note that the kernel matrix constructed for SVMs
can be reused. Furthermore, only one eigenvector corresponding to the second
largest eigenvalue needs to be computed, so optimized algorithms can be used.
By partitioning the set of training samples S instead of the set of classes
C = {[s] : s ∈ S},1 a separating boundary between the samples can be found. A
disjoint bi-partitioning of samples S = A B leads to a disjoint tri-partitioning
1
[s] denotes the class assigned to sample s ∈ S.
a) solution to the example of Fig. 2a b) resulting hierarchy
Fig. 3. Illustration of our approach. Separation difficulties can be avoided by includ-

ing non-separable classes on both sides of the decision boundary. This can simplify
subsequent splits (left) and leads to a rooted DAG structure (right).
of classes C = A X B, where all classes in A have all samples in A, all classes

in B have all samples in B, and finally the rest of the classes X have samples in
both partitions. Our proposal is to split the set of classes C = L ∪ R so that the
classes in X belong to both sets, i.e., X = L ∩ R:
L = A ∪ X = {C : ∃s∈A [s] = C}
R = B ∪ X = {C : ∃s∈B [s] = C} . (6)
In practice, we can also slightly relax the requirement for A (B) to have all
samples in A (B). Given a partitioning p : S → {−1, 1} of the training set S, we
define a function q : C → [−1, 1] on the set of classes C:
1
q(C) = p(s) (7)
|C|
s∈C
where C ∈ C is a class.
This allows us to define a split:

L = q −1 [−1, 1−α)

R = q −1 (−1+α, 1] (8)
where q −1 denotes an inverse image

and α is a softening
parameter.
Note that in

this case A = q −1 [−1, −1+α] , B = q −1 [1−α, 1] and X = q −1 (−1+α, 1−α) ,
thus when α = 0 then the above simplifies to (6).
Figure 4 illustrates the split procedure. Let us consider a set of classes C
ordered according to q(C) values for each class C ∈ C. The L set consists of
classes L ∈ L with q(L) < 1 − α and the R set of classes R ∈ R with q(R) >
−1 + α. Increasing α reduces the split overlap. This produces more efficient
hierarchies, but can degrade performance.
We create our Relaxed Hierarchy by recursively splitting the class-sets Cn
until they contain only one class or a group of classes impossible to split,
Fig. 4. Illustration of the split procedure. Note how the value of α influences the
overlap.
i.e., until |Cn | = 1 or Ln = Rn . In the second case we use OAR on the sub-
set of classes that is too complex to split.
To train the hierarchy, for each node of the computed rooted DAG we train
an SVM using samples belonging to classes in Rn \ Ln as a positive set and to
classes in Ln \ Rn as a negative set. Note that samples belonging to classes in
Xn = Ln ∩Rn are not used for training. This does not matter, since classification
of a sample that belongs to a class in Xn is not relevant at this stage. This is the
key point of our method, since the decision for these classes could be erroneous
and is postponed till later.
For testing, the tree is descended until a leaf is reached. The decision is either
directly the class label (leaves containing only one class) or OAR classification is
performed on the remaining classes (complex leaves with more that one class).
4 Experiments
In Subsect. 4.1 we describe the implementation details of our image representa-

tion and the classifier used. Note, however, that different image representations
and classifiers can be combined with our Relaxed Hierarchy. Subsection 4.2 in-
troduces the dataset and the experimental setup. Results are presented and
discussed in Subsect. 4.3.
4.1 Image Representation and Image Dissimilarity Measure

Given an image, we use complementary Harris-Laplace [2] and Laplacian interest
point detectors [3] to find a sparse set of salient image regions. Both detectors
are invariant to scale transformations, they output circular regions at a charac-
teristic scale. Using the SIFT [4] descriptor, gradient orientation histograms are
computed over those regions.
To describe an image, we use the bag-of-features representation [5]. Given a
visual vocabulary, we represent the appearance of the image as a histogram of
vocabulary words occurring in the image. Each histogram entry hij ∈ Hi is the
proportion of all image features i assigned to a vocabulary word j with respect
to the total number of features in the image.
To compute the dissimilarity between the images, we use the χ2 distance
1 (hin − hjn )2
V
m(Hi , Hj ) = . (9)
2 n=1 hin + hjn
where V is the vocabulary size. We use k-means to construct the vocabulary and
V = 8000 in our experiments.
To use this distance measure in Support Vector Machines, we use the extended
Gaussian kernel, cf. (3). This results in a Mercer kernel [23]. The parameter γ is
set to the mean value of the distances between all training samples.
Using the above image representation with Support Vector Machines in the
OAR setup corresponds to the method of Zhang et al. [8]. This method has
shown an excellent performance on varying object class datasets, including 2005
and 2006 Pascal VOC challenges [8,24]. Extended with additional channels
and a separate optimization framework to combine them, this approach won the
Pascal VOC classification challenge in 2007 [1].
4.2 Caltech-256
We evaluate our approach on the Caltech-256 [10] dataset. It contains images of

256 object categories and an additional background class. Each category contains
at least 80 images of varying size and quality. The images were downloaded from
the web with the help of popular image search engines and then human-filtered.
We closely follow the experimental setup suggested by the dataset authors,
i.e., we use the first 250 categories of the dataset to measure the accuracy of
multi-class image classification. The first 15 images of each class are used for
training (this includes computing the visual vocabulary for the bag-of-features,
constructing the class hierarchy and training the SVMs) and all the remaining
images are used for testing. We report the average of the per-class classification
accuracy.
Figure 1 shows a few samples and the classification accuracies for the corre-
sponding classes. We can see that classes with very high intra-class variability
(like rifles and swords) are the most difficult. Our method also confuses two very
similar bike classes (34% touring bikes are classified as mountain bikes and 24%
mountain bikes as touring bikes). It performs very well on classes with discrim-
inative texture (like zebras) and those with low intra-class variability (like the
tower of Pisa).
4.3 Results
Figure 5 shows a class hierarchy constructed by our method for the Caltech-
256 dataset, displayed for a subset of 10 categories. The categories were chosen
to include animals, natural phenomena and man-made objects. They include
class pairs with apparent visual similarities that are semantically close (bear
and dog, top hat and cowboy hat) as well as those that have a secondary or
no semantic relationship at all (bear and teddy bear, top hat and Saturn). The
Fig. 5. Class hierarchy constructed by our method for the Caltech-256 dataset, dis-
played for a subset of 10 categories
hierarchy reveals many intuitive relationships and groupings. At the top node
man-made objects and natural phenomena (hats, lightning, rainbow, Saturn)
are separated from animals (octopus, starfish, bear). Classes at the partition
boundary (dog and teddy bear) are included in both partitions. Subsequent
splits further separate sea animals from land animals (with a teddy bear) and
hat-like objects (including Saturn) from natural phenomena and mascot-like
objects. Even though it is based on visual data only, the constructed hierarchy
turns out to be similar to hierarchies extracted from semantic networks [25].
Unlike the purely semantic hierarchies, however, it also groups classes that are
related by semantic links difficult to model (bear and teddy bear) or that feature
accidental similarity (top hat and Saturn).
Table 1 shows the average per-class classification accuracy on the Caltech-
256 dataset. The upper half of the table compares our approach, i.e., a Relaxed
Hierarchy (RH), to the OAR setup. We can see that the proposed hierarchy
does not lead to accuracy loss. The image representation is the one described in
Subsection 4.1. The lower half of the table shows a result for a different image
representation, i.e., based on a reimplementation of the method of Lazebnik at
al. [7]. This representation obtains better results for the Caltech-256 dataset,
as most objects are centered in the image and relatively small. Again, we can
Table 1. Average per-class classification accuracy on the Caltech-256 dataset
OAR (reimpl. of Zhang et al. [8]) 23.6%

Our RH (α = 0, sparse IPs) 23.4%
Griffin [10] (reimpl. of Lazebnik at al. [7]) 28%

Our RH (α = 0, dense/grid) 27.9%
observe that the results obtained with our RH and an OAR approach (see results
obtained by Griffin et al. [10]) are comparable.
As to be expected, our approach does not depend on the image representation.
Best results on Caltech-256 dataset in a similar setup (53% average accuracy
for 10 training images) where achieved by Varma [26] using a combination of
multiple channels. Our method could be combined with this multi-representation
approach. Note that it could even be applied to different data types, but this is
beyond the scope of this paper. In the following we use the image representation
described in Sect. 4.1 as it is fast to compute and does not impact the evaluation
of our class hierarchy construction.
Figure 6 compares the complexity in the number of categories. The complex-
ity in the OAR setup is linear (red squares). The complexity of our Relaxed
Hierarchy method is confirmed to be sublinear. The exact gain depends on the
parameter α, see the datapoints along the right edge. Note that α is expressed
here as r—the number of relaxed training samples per class, i.e., α = r/15. For
250 categories and a setting of α = 3/15 = 0.2 (blue diamonds) which corresponds
to minor performance loss, we observe a reduction of the computation time by
1/3. This ratio will further increase with the number of categories.
Figure 7 demonstrates the speed-for-accuracy trade-off (green circles) that
can be tuned with the α parameter. As shown in Sect. 3, with the increase of the
parameter value the set of classes is more willingly treated as separable. Greater
α values lead to better computational gain, but could degrade the classification
accuracy. Note that the complexity is sublinear independently of the parameter
setting (see Fig. 6), but for the smaller number of classes one may choose to
accept a small loss in accuracy for a significant gain in computation time. For
instance, for Caltech-256 we find the setting of α = 0.2 (r = 3) reasonable, as
the absolute loss in the accuracy is only about 2%, while the computational gain
250 1
OAR 10
2
Our RH (r=0) 3
200 Our RH (r=1) 0.8
SVM runs per test image
4
Our RH (r=3) 5
Relative accuracy
Our RH (r=5)
150 0.6
100 0.4
OAR
50 0.2 Our RH (r)
Std. top-down
Std. bottom-up
0 0
0 50 100 150 200 250 0 0.2 0.4 0.6 0.8 1
Number of classes Relative complexity
Fig. 6. Complexity in the number of Fig. 7. Speed-for-accuracy trade-off and

classes. The α relaxation parameter ex- comparison with existing approaches.
pressed in the number of per-class train- Next to the RH datapoints the α relax-
ing samples r (i.e., α = r/15) is given in ation parameter expressed in the number
parenthesis for our method. of samples r is shown (α = r/15).
of 1/3 is noticeable. Setting α = 0.33 (r = 5) leads to the computational gain of

3/5, but in exchange for another 2% of accuracy.
Figure 7 compares the results obtained with our class hierarchies for different
α values (green circles) to two existing methods for class hierarchy construc-
tion (triangles). The baseline top-down method follows the approach of Liu et
al. [12,14], but we use normalized cuts instead of k-means. This makes it more
comparable to our method and is also similar to the approach of Chen et al. [11].
The baseline bottom-up method follows the agglomerative clustering based ap-
proach of Zhigang et al. [13], but uses the same inter-class similarity measure
as the top-down approach. Note that the only difference between the compared
methods is the algorithm used for class hierarchy construction, i.e., we keep the
same image representation and settings of the Support Vector Machines. Still,
the magnitude of the difference is surprising. The standard bottom-up method
seems to fail completely. The standard top-down approach has low computa-
tional complexity, but the loss in terms of classification accuracy is enormous.
This confirms our claim, see Subsect. 2.3, that popular disjoint approaches for
construction of class hierarchies fail when dealing with a large number of visual
object categories. Note that the cited methods were evaluated on visual data and
performed well. However, the number of categories never exceeded 14 classes and
was usually kept below 10.
5 Summary
We have shown that existing approaches for constructing class hierarchies for
visual recognition do not scale well with the number of categories. Methods that
perform disjoint class-set partitioning assume good class separability and thus
fail to achieve good performance on visual data when the number of categories
becomes large. Thus, we have proposed a method that detects classes at the par-
titioning boundary and postpones uncertain decisions until the number of classes
becomes smaller. Experimental validation shows that our method is sublinear in
the number of classes and its classification accuracy is comparable to the OAR
setup. Furthermore, our approach allows to tune the speed-for-accuracy trade-off
and, therefore, allows to significantly reduce the computational costs.
Our method finds a reliable partitioning of the categories, but the hierarchy
may be far from optimal. Finding the optimal partitioning is a hard problem.
For the future work we plan use semantic information to drive the optimization.
Acknowledgments. M. Marszalek is supported by the European Community

under the Marie-Curie project Visitor. This work was partially funded by the
European research project Class.
References
1. Everingham, M., van Gool, L., Williams, C., Winn, J., Zisserman, A.: Overview and
results of classification challenge. In: The PASCAL VOC 2007 Challenge Workshop,
in conj. with ICCV (2007)
2. Mikolajczyk, K., Schmid, C.: Scale and affine invariant interest point detectors.
IJCV (2004)
3. Lindeberg, T.: Feature detection with automatic scale selection. IJCV (1998)
4. Lowe, D.: Distinctive image features form scale-invariant keypoints. IJCV (2004)
5. Willamowski, J., Arregui, D., Csurka, G., Dance, C.R., Fan, L.: Categorizing nine
visual classes using local appearance descriptors. In: IWLAVS (2004)
6. Schölkopf, B., Smola, A.: Learning with Kernels: Support Vector Machines, Regu-
larization, Optimization and Beyond (2002)
7. Lazebnik, S., Schmid, C., Ponce, J.: Beyond bags of features: Spatial pyramid
matching for recognizing natural scene categories. In: CVPR (2006)
8. Zhang, J., Marszalek, M., Lazebnik, S., Schmid, C.: Local features and kernels for
classification of texture and object categories: A comprehensive study. IJCV (2007)
9. Fei-Fei, L., Fergus, R., Perona, P.: One-shot learning of object categories. PAMI
(2007)
10. Griffin, G., Holub, A., Perona, P.: Caltech-256 object category dataset. Technical
report (2007)
11. Chen, Y., Crawford, M., Ghosh, J.: Integrating support vector machines in a hier-
archical output space decomposition framework. In: IGARSS (2004)
12. Liu, S., Yi, H., Chia, L.T., Deepu, R.: Adaptive hierarchical multi-class SVM clas-
sifier for texture-based image classification. In: ICME (2005)
13. Zhigang, L., Wenzhong, S., Qianqing, Q., Xiaowen, L., Donghui, X.: Hierarchical
support vector machines. In: IGARSS (2005)
14. Yuan, X., Lai, W., Mei, T., Hua, X., Wu, X., Li, S.: Automatic video genre cate-
gorization using hierarchical SVM. In: ICIP (2006)
15. Zweig, A., Weinshall, D.: Exploiting object hierarchy: Combining models from
different category levels. In: ICCV (2007)
16. He, X., Zemel, R.: Latent topic random fields: Learning using a taxonomy of labels.
In: CVPR (2008)
17. Griffin, G., Perona, P.: Learning and using taxonomies for fast visual category
recognition. In: CVPR (2008)
18. Nister, D., Stewenius, H.: Scalable recognition with a vocabulary tree. In: CVPR
(2006)
19. Philbin, J., Chum, O., Isard, M., Sivic, J., Zisserman, A.: Object retrieval with
large vocabularies and fast spatial matching. In: CVPR (2007)
20. Casasent, D., Wang, Y.C.: A hierarchical classifier using new support vector ma-
chines for automatic target recognition. Neural Networks (2005)
21. Shi, J., Malik, J.: Normalized cuts and image segmentation. PAMI (2000)
22. Rahimi, A., Recht, B.: Clustering with normalized cuts is clustering with a hyper-
plane. In: SLCV (2004)
23. Fowlkes, C., Belongie, S., Chung, F., Malik, J.: Spectral grouping using the
Nyström method. PAMI (2004)
24. Everingham, M., Zisserman, A., Williams, C., van Gool, L.: The PASCAL visual
object classes challenge 2006 (VOC 2006) results. Technical report (2006)
25. Marszalek, M., Schmid, C.: Semantic hierarchies for visual object recognition. In:
CVPR (2007)
26. Perona, P., Griffin, G., Spain, M.: The Caltech 256 Workshop. In: Conj. with ICCV
(2007)
Sample Sufficiency and PCA Dimension for Statistical
Shape Models
Lin Mei* , Michael Figl, Ara Darzi, Daniel Rueckert, and Philip Edwards∗
Dept. of Biosurgery and Surgical Technology Imperial College London, UK

{l.mei,eddie.edwards}@imperial.ac.uk
Abstract. Statistical shape modelling(SSM) is a popular technique in computer

vision applications, where the variation of shape of a given structure is modelled
by principal component analysis (PCA) on a set of training samples. The issue
of sample size sufficiency is not generally considered. In this paper, we propose
a framework to investigate the sources of SSM inaccuracy. Based on this frame-
work, we propose a procedure to determine sample size sufficiency by testing
whether the training data stabilises the SSM. Also, the number of principal modes
to retain (PCA dimension) is usually chosen using rules that aim to cover a per-
centage of the total variance or to limit the residual to a threshold. However, an
ideal rule should retain modes that correspond to real structural variation and dis-
card those that are dominated by noise. We show that these commonly used rules
are not reliable, and we propose a new rule that uses bootstrap stability analysis
on mode directions to determine the PCA dimension.
For validation we use synthetic 3D face datasets generated using a known
number of structural modes with added noise. A 4-way ANOVA is applied for
the model reconstruction accuracy on sample size, shape vector dimension, PCA
dimension, and the noise level. It shows that there is no universal sample size
guideline for SSM, nor is there a simple relationship to the shape vector dimen-
sion (with p-Value=0.2932). Validation of our rule for retaining structural modes
showed it detected the correct number of modes to retain where the conventional
methods failed. The methods were also tested on real 2D (22 points) and 3D
(500 points) face data, retaining 24 and 70 modes with sample sufficiency being
reached at approximately 50 and 150 samples respectively. We provide a foun-
dation for appropriate selection of PCA dimension and determination of sample
size sufficiency in statistical shape modelling.
1 Introduction
Statistical shape modelling (SSM) is a technique for analysing variation of shape and
generating or inferring unseen shapes. A set of sample shapes is collected and PCA is
performed to determine the principal modes of shape variation. These modes can be
optimised to fit the model to a new individual, which is the familiar active shape model
(ASM) [1,2,3]. Further information, such as texture, can be included to create an active
appearance model [4] or morphable model [5].
* We would like to thank Tyco Healthcare for funding Lin Mei’s PhD studentship. We are also
grateful to many other members of the Department of Computing and the Department of Bio-
surgery and Surgical Technology at Imperial College.
Sample Sufficiency and PCA Dimension for Statistical Shape Models 493
Despite its popularity, PCA-based SSMs are normally trained from datasets for which
the issue of sufficiency is not considered. The PCA dimension for an SSM is often
chosen by rules that assume either a given percentage or level of noise. As will be shown
later in this paper that these two methods are highly dependent on sample size.
In this paper, we review the discussions on sample size sufficiency for a closely
related field, common factor analysis (CFA), and design a mathematical framework
to investigate the source of PCA model error. This framework provides a theoretical
evaluation of the conventional rules for retaining PCA modes, and enables analysis
of sample size sufficiency for PCA. We then propose a rule for retaining only stable
PCA modes that uses a t-test between the bootstrap stability of mode directions from
the training data and those from pure Gaussian noise. The convergence of the PCA
dimension can then be used as an indication of sample sufficiency.
We verify our framework by a 4-way ANOVA for reconstruction accuracy is applied
to the models trained from synthetic datasets generated under different conditions. Our
PCA dimension rule and procedure for sample sufficiency determination are validated
on the synthetic datasets and demonstrated on real data.
2 Background
2.1 Minimum Sample Size for CFA
There is little literature on the issue of minimum sample size for PCA. In the related
field of CFA, however, this issue has been thoroughly discussed. CFA is commonly
used to test or discover common variation shared by different test datasets. Guidelines
for minimum sample size in CFA involve either a universal size regardless of the data di-
mension or a ratio to the data dimension. Recommendations for minimum size neglect-
ing the sample dimension and the number of expected factors vary from 100 to 500 [6].
Such rules are not supported by tests on real data. Doubts have been raised about a uni-
versal sample size guideline since it neglects the data dimension. Size-variable ratios
(SVR) may be more appropriate and values of between 2:1 to 20:1 have been sug-
gested [7]. There have been a number of tests using real data, but no correlation was
found between SVR and the mode stability [8], nor has any minimum value for SVR
emerged [9]. The minimum sample size needed in these real tests is not consistent ei-
ther, varying from 50 [8], to 78-100 [9], 144 [10], 400 [11] and 500 or more [12].
The inconsistency among these results shows that the minimum size depends on
some nature of the data other than its dimension. MacCallum et al. [13,14] proposed
a mathematical framework for relating the minimum sample size for CFA with its
communality and overdetermination level. They then designed an experiment using
4-way ANOVA to study the effects of communality, overdetermination level, model
error and sample size on the accuracy in recovering the genuine factors from synthetic
data. The results showed that communality had the dominant effect on the accuracy
regardless of the model error. The effect of overdetermination level was almost neg-
ligible when communality is high. In low communality tests, accuracy improves with
larger sample size and higher accuracy was found in tests with lower overdetermination
levels.
494 L. Mei et al.
There is no equivalent to communality and overdetermination level for PCA. Instead,

the factors we consider are the data dimension and the number of genuine structural
modes that are retained.
2.2 Number of Modes to Retain for SSM

Many rules choosing the number of modes to retain for SSM and PCA have been pro-
posed [6,15,16,17]. The most popular rule used in SSM for structural mode determina-
tion is simply to take the leading modes covering a percentage of the total variance in
the sample set. The percentage is arbitrarily set, which equivalent to simply assuming a
corresponding percentage of noise.
Another popular rule is to discard the least principal modes until the sum of total
variance, which is the model residual on the training data, reaches a certain threshold.
This threshold is normally set according to the error tolerance of the application.
Stability measurements for PCA have been proposed to determine the number of
modes. Given two shape models trained from different sample sets, Daudin et al [18]
used a sum of correlation coefficients between pairs of principal components; Besse
et al [19] used a loss function derived from an Euclidean distance between orthogonal
projectors; Babalola et al [20] used the Bhattacharya Metric to measure the similar-
ity of PCA models from different sample sets. Resampling techniques such as boot-
strapping [18] and jackknifing [19] can be used. The distribution of PCA modes across
the replicates reflects their distribution in the population, allowing stability analysis to
be performed. The selected principal modes span a subspace. Besse et al. proposed
a framework for choosing the number of modes based on their spanned-space stabil-
ity [21]. This method differentiates structural modes and noise-dominated modes when
the sample set is large. However, as will be shown in the section 4.3, this method can
only provide a estimation of the number of modes when the sample size is sufficient.
3 Theories
3.1 Sources of PCA Model Inaccuracy
We propose the following mathematical framework to examine the characteristics af-
fecting the sufficiency of a sample set drawn from a population with genuine modes of
< instead of
variation, listed in the column of A. Due to the presence of noise, we have X
X, and the PCA modes from X < are A.
< The model inaccuracy can be expressed as the
difference between the covariance matrices Δ = X <X< T − XX T .
Let X = AW and X < =A <W
= , we have:
<=A
X <W <W
= = AAT A = + (I − AAT )A
<W= (1)
Since A is orthonormal, (I − AAT ) is a diagonal matrix with only 1s and 0s. Hence
(I − AAT ) = N N T . Equation 1 becomes:
< = AAT A
X <W= + NNT A
<W= = AW
=A + N W
=N (2)
Applying equation 2 on the covariance matrix of X:
<X
X < T = AW = T AT + AW
=A W =T N T + N W
=A W = T AT + N W
=N W =T N T
=N W
A N A N

= AΣAA A + AΣAN N + N ΣN A A + N ΣN N N
T T T T
(3)
The model inaccuracy becomes:
Δ=X<X< T − XX T

= A(Σ AA − ΣAA )A + AΣAN N + N ΣN A A + N ΣN N N
T T T T
= (A(ΣEE )AT + AΣAN N T + N ΣN A AT ) + N ΣN N N T (4)
A PCA model error consists of two parts:

EN = N ΣN N N T , the error introduced by sampling noise modes that are orthogonal
to . EN depends only on the noise level introduced by human interaction or measure-
ment error during the process of building an SSM. Increasing sample size would cause
little reduction in this error if noise level remains the same.
El = A(ΣEE )AT + AΣAN N T + N ΣN A AT , the error along the subspace spanned
by structural modes, . This is due to noise affecting the sample coefficients and insuf-
ficient coverage of the dimensions in . Therefore, El increases with PCA dimension,
rank(A). It also affected by noise level because at high noise level some structural
modes with small variances may be swamped by noise.
Rather counter-intuitively, El is not dependent on the shape vector dimension, as will
be shown in the section 4.2. However, higher EN can result from higher shape vector
dimension, which therefore increases Δ.
3.2 Sample Size Requirement
According to the framework in section 3.1, the sample size requirement for PCA only
depends on two factors: number of structural modes in the dataset, and the level of
noise. Hence we propose the following procedure for sample sufficiency determination.
For a sample set, X, of n samples:
PCA Sample Size Sufficiency Test

1) Apply PCA on X, to get a set of modes B.
2) Starting with a reasonably small number, n∗ , construct a set Xj∗ of n∗ samples ran-
domly drawn, allowing repeats, from X.
3) Apply PCA to Xj∗ to get a set of modes Bj∗ and resolve mode correspondence with
respect to B.
4) Find the number of structural modes in Bj∗ , k.
5) Repeat 2-4 with an increased n∗ . If k converges before n∗ reaches n, we have suffi-
cient samples. Otherwise, further sample data is required.
Step 4 in this procedure requires determination of the number of structural modes,

which is a common question for PCA. These rules are sometimes is called stopping
rules, since they determine where to stop including the principal modes.
496 L. Mei et al.
3.3 PCA Dimension by the Stability of Mode Direction

It is generally assumed that most of the structural variation is modelled by the leading
principal modes. Noise affects the lower ranking modes and dominates those after a
cut-off point further down the rank. Going back to our framework, a cut-off point is
chosen for retaining principal modes in order to reduce EN . However, since genuine
structural variation may still be present in the least principal modes, discarding them
would increase El . There is trade-off between EN and El . Stopping rules should aim
at discarding only modes that are sufficiently dominated by noise.
Assuming noise is randomly spread across all the dimensions, mode instability can
be a good indication of the point where noise begins to dominate. There is a risk with
tests using the magnitude of the variance that stopping rules will be dominated by the
first few modes and fail to identify the correct cut-off point. Also, it is the mode direc-
tions that define the basis of a shape model for fitting or synthetic shape generation.
Therefore we propose a stopping rule based on the stability of the mode direction only.
Fig. 1. Comparison of Leading 8 Eigenmodes from two mutually exclusive sets of 50 samples
from our 3D face mesh database, aligned according to eigenvalue ranks. Darker texture implies
larger variation, showing many mismatched after the 4th mode.
3.3.1 Establishing Mode Correspondence

Examining individual modes requires mode correspondence. Normally, this is done by
matching those with the same eigenvalue ranks. Significant variation can be found be-
tween individual modes drawn from different sample sets with the same ranking, as
shown in figure 1. Although leading modes may correspond, mode 5 on the top seems
to correspond with mode 6 on the bottom, and modes after 6 on the top seems to corre-
spond to none at the bottom. However, the combined modes from different sample sets
may still span similar subspaces. Mode alignment can be achieved by minimising the
distance between these subspaces.
For the leading PCA modes {(ai , λi )|ai | = 1} of an n-dimensional distribution, we

define the principal spanned space (PSS) as the subspace k spanned by {ai }, where
the distance measure used by Besse et al.[19] can be applied:
d( k
, k ) = k − trace(AAT BBT ) (5)
where the columns of A and B are the modes spanning PSS k

and k .
For two sets of PCA modes, ai and bi , trained from different sample sets of a
common distribution, the following rule can be used to establish correspondence. The

first mode in ai corresponds to the mode of a replicate that minimises d( 1 , 1 ),
and we proceed iteratively. Assume we have already aligned k

, the PSS from the

first k modes in ai , to the spanned space from k modes in the replicate bi . The
k
mode in bi that corresponds to the k+1th mode in ai will be the one that minimises

d( k+1 , k+1 ).
3.3.2 Bootstrap Stability of PCA Modes

Bootstrap stability analysis can be used to analyse mode stability. We use the angles
between mode directions as the measurement of distance between corresponding modes
from different replicates. The instability, ξ, of mode ai is given by:
m
=i )·
j=1 arccos(ai j · α
ξ(ai ) = (6)
mπ
=i is the mean mode vector and m is the number of bootstrap replicates.
where α
3.3.3 Stopping Rule Based on a t-Test against Synthetic Gaussian Noise

Since noise-dominated modes should have higher instability than structural modes, a
threshold on ξ can be used to differentiate them from structural modes. However, the
choice for the threshold is arbitrary and is found to be sensitive to the size of replicates.
Instead, assuming the distribution of angles between corresponding modes is Gaussian,
a one-tailed t-test can be used to establish whether a mode is dominated by noise to a
given significance level.
We generate a pure Gaussian noise dataset to compare with the test dataset. All
conditions must be the same – the dimensionality, the number of samples in the dataset,
the number of replicates, and the number of samples in each replicate. Since we are
only interested in mode directions, the level of noise is not important. Let the angle for
the first pure noise mode to be α1 and the angle for the test samples to be ai , The null
hypothesis of the t-test is H0 : ξ(α1 ) > ξ(ai ). By rejecting H0 at a given confidence
level, one can safely conclude that a mode is not dominated by noise.
4 Experiments
We demonstrate the correctness of our theories with three sets of experiments. First a 4-
way ANOVA is performed on synthetic datasets to show how the PCA model accuracy
is affected by different features as it is discussed in section 3.1. Then we show that our
stopping rule is able to identify the correct number of modes in the synthetic samples
for which commonly used rules fail. This shows that our rule can be used to determine
PCA sample sufficiency, by following the procedure presented in section 3.2. This is
applied to two different sets of real samples.
4.1 Real Datasets

Two real shape datasets are used in the experiments. The first one comprises 150 sam-
ples of 3D faces with 5090 points each from University of Notre Dame [22] prepro-
cessed using Papatheodorou’s method [23], and the second one consists of 135 samples
498 L. Mei et al.
a b
Fig. 2. Examples from real 3D Face database (a) and landmarks of 2D AR Face database
a b c
Fig. 3. Examples of three synthetic faces generated with 70 modes with shape vector dimension
being 2100. Different noise levels are applied: 0.1mm (a), 0.25mm (b) and 0.5mm (c). Noise
starts to become visible in (b) and (c).
from the landmarks (22 points) [24] of 2D AR face database [25]. Examples from these
two datasets are shown in figure 2.
4.2 ANOVA Results

For validation of our framework, we generate a dataset consists of 8960 subsets, each
having different combinations of: sample sizes, numbers of modes to generate, levels
of Gaussian noise and decimated to different number of points. A list of choices for
different characteristics are shown as follows
Sample Sizes (SS): 50, 100, 150, 200, 250, 300, 350, 400
450, 500, 550, 600, 650, 700, 750, 800
Shape Vector Dimension: 300, 600, 900, 1200, 1500, 1800, 2100
Number of Genuine Modes: 10, 20, 30, 40, 50, 60, 70, 80
Gaussian Noise Levels (in mm): 0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5
Examples of faces generated under different conditions are given in figure 3.
PCA is applied to each of the 8960 subsets of the first synthetic dataset. Results are
compared to the original modes that used to generate the data. Measurement described
in equation 5 is used to calculate the error of the models trained from the subsets. A
4-way ANOVA was performed to find out which characteristics influence the model
accuracy. As shown in table 1, the results confirm the correctness of our framework
introduced in section 3.1. Sample size and number of genuine modes in the dataset act
as the major source of influence on the model accuracy. Noise also has a significant but
small influence. Also the result showed that the effect of sample dimension is negligible.
Table 1. Result of 4-Way ANOVA
Source Sum of Squares DoF Mean Squares F-Statistic p-Value

Sample Size 194.534 15 12.9689 11315.66 ¡0.03
Sample Dimension 0.008 6 0.0014 1.22 0.2932
Number of Genuine Modes 83.055 7 11.865 10352.48 ¡0.03
Gaussian Noise Level 0.513 9 0.057 49.74 ¡0.03
4.3 Number of Modes to Retain for SSM

We have validated previous stopping rules and our method using synthetic data gener-
ated with a known number of background modes. These shapes are generated using the
leading 80 modes of the model built from all the 150 3D Faces, decimated to 500 points
for faster experiments. Gaussian noise with 1mm standard deviation is added to each
element of the shape vector. Example faces from the synthetic set are shown in figure 4.
Stopping rules applied to this dataset should not retain more than 80 modes.
We validated the rule which retains 95% of the cumulative variance using synthetic
datasets sized from 100 to 600. Compactness plots are shown in figure 5(a). With in-
creasing sample size, the number of modes retained by this rule increases beyond 80,
where the noise dominates the variance. These noise modes contribute to an increasing
proportion of the total variance with increasing sample size, and the number of modes
covering 95% of the total variance increases accordingly. A similar trend was also found
Fig. 4. Synthetic faces generated using 80 modes, added 1mm Gaussian noise on to each element
of the shape vector with dimension 1500
500 L. Mei et al.
a b
Fig. 5. 95% thresholded compactness plots of synthetic 3D face datasets (a) with 100, 200, 400
and 600 samples and real 3D face datasets (b) with 30, 50, 100 and 150 samples. The number of
retained modes is clearly dependent on sample size.
Table 2. Number of modes to keep the point error below 1mm
Number of Samples 50 100 150 200 250 300 350 400 450 500
Number of Modes 32 60 95 108 120 140 169 186 204 219
Fig. 6. Instability of PSS for synthetic datasets for synthetic datasets sized from 200 to 2000
for the real data as shown in figure 5(b), which strongly suggests that this rule is un-
reliable and should not be used. A similar effect, as shown in table 2, was found for
the stopping rule that discards the least principal modes until the average error of each
point reaches 1mm.
Fig. 7. t-Test Based stopping rule on synthetic datasets
Fig. 8. Result of real datasets sufficiency test. Left: 2D faces; Right: 3D faces.
The method of Besse et al [21] was validated with synthetic datasets sized from
200 to 400. A plot of instability, measured as the distance between subspaces spanned
by different replicates, is shown in figure 6. Although this method provides a visible
indication of the correct number of modes to retain when the sample size is sufficiently
large, it cannot identify the lower number of modes that should be retained when the
sample size is insufficient.
Our method was validated with synthetic datasets sized from 100 to 2000. Figure 7
shows the number of modes to retain versus the sample size is also shown. Our stop-
ping rule does not have the tendency to go beyond 80 with large sample sizes. It also
identifies a lower number of stable modes to retain for smaller sample sizes. It appears
a sample size of around 500 is sufficient.
4.4 Sample Size Sufficiency Test for SSM
Figure 8 shows the results of sample size sufficiency tests on the three real datasets we
have. For the 2D dataset, the plot obviously converges at 24 modes with 50 samples.
502 L. Mei et al.
With the 3D faces, the graph appears close to convergence at around 70 modes for the
150 samples. These results suggest both face datasets are sufficient.
We propose a mathematical framework to analyse the sources of inaccuracy in PCA,

which suggests that only the number of genuine structural modes in the dataset and
the noise level will influence the minimum sample size required to stabilise the model.
There is no trivial relationship between the shape vector dimension and the required
sample size. These results are confirmed by a 4-way ANOVA on synthetic data.
We propose a stopping rule that retains modes according to the stability of their direc-
tions. We also establish mode correspondence from different training sets by minimis-
ing the distance between the principal spanned spaces rather than simply by the rank of
their eigenvalues. For a synthetic dataset generated with known structural modes plus
added noise, our method converges correctly where conventional methods did not.
The number of genuine structural modes is not known, but the number of modes to
use for a given dataset can be estimated using our stopping rule. The convergence of
this rule can then be used as an indicator of sample size sufficiency.
Resulting sample size sufficiency suggest 50 samples is sufficient for 2D face land-
marks(22 points), retaining 24 modes in total, and 150 samples is sufficient for the 3D
face meshes (500 points), where around 70 modes are retained. We believe this is the
first principled test for sample sufficiency and determination of the number of modes to
retain for SSM. It can also be applied to other applications of PCA and related fields.
References
1. Cootes, F., Hill, A., Taylor, C., Haslam, J.: The use of active shape models for locating
structures in medical images. In: Proc. IPMI, pp. 33–47 (1993)
2. Cootes, T., Taylor, C., Cooper, D., Graham, J.: Active shape models and their training and
application. Comput. Vis. Image Underst. 61(1), 38–59 (1995)
3. Sukno, F.M., Ordas, S., Butakoff, C., Cruz, S.: Active shape models with invariant optimal
features: Application to facial analysis. IEEE Trans. Pattern Anal. Mach. Intell. 29(7), 1105–
1117 (2007) (Senior Member-Alejandro F. Frangi)
4. Cootes, T., Edwards, G., Taylor, C.: Active appearance models. IEEE Transactions on Pattern
Analysis and Machine Intelligence 23(6), 681–685 (2001)
5. Blanz, V., Vetter, T.: Face recognition based on fitting a 3D morphable model. IEEE Trans-
actions On Pattern Analysis And Machine Intelligence 25, 1063–1074 (2003)
6. Osborne, J., Costello, A.: Sample size and subject to item ratio in principal components
analysis. Practical Assessment, Research and Evaluation 9(11) (2004)
7. Guadagnoli, E., Velicer, W.: Relation of sample size to the stability of component patterns.
Psychological Bulletin 103, 265–275 (1988)
8. Barrett, P., Kline, P.: The observation to variable ratio in factor analysis. Personality Study
and Group Behavior 1, 23–33 (1981)
9. Arrindell, W., van der Ende, J.: An empirical test of the utility of the observations-to-variables
ratio in factor and components analysis. Applied Psychological Measurement 9(2), 165–178
(1985)
10. Velicer, W., Peacock, A., Jackson, D.: A comparison of component and factor patterns: A
monte carlo approach. Multivariate Behavioral Research 17(3), 371–388 (1982)
11. Aleamoni, L.: Effects of size of sample on eigenvalues, observed communalities, and factor
loadings. Journal of Applied Psychology 58(2), 266–269 (1973)
12. Comfrey, A., Lee, H.: A First Course in Factor Analysis. Lawrence Erlbaum, Hillsdale (1992)
13. MacCallum, R., Widaman, K., Zhang, S., Hong, S.: Sample size in factor analysis. Psycho-
logical Methods 4, 84–99 (1999)
14. MacCallum, R., Widaman, K., Hong, K.P.S.: Sample size in factor analysis: The role of
model error. Multivariate Behavioral Research 36, 611–637 (2001)
15. Jackson, D.: Stopping rules in principal components analysis: a comparison of heuristical
and statistical approaches. Ecology 74, 2204–2214 (1993)
16. Jolliffe, I.: Principal Component Analysis, 2nd edn. Springer, Heidelberg (2002)
17. Sinha, A., Buchanan, B.: Assessing the stability of principal components using regression.
Psychometrika 60(3), 355–369 (2006)
18. Daudin, J., Duby, C., Trecourt, P.: Stability of principal component analysis studied by the
bootstrap method. Statistics 19, 341–358 (1988)
19. Besse, P.: PCA stability and choice of dimensionality. Statistics& Probability 13, 405–410
(1992)
20. Babalola, K., Cootes, T., Patenaude, B., Rao, A., Jenkinson, M.: Comparing the similarity of
statistical shape models using the bhattacharya metric. In: Larsen, R., Nielsen, M., Sporring,
J. (eds.) MICCAI 2006. LNCS, vol. 4190, pp. 142–150. Springer, Heidelberg (2006)
21. Besse, P., de Falguerolles, A.: Application of resampling methods to the choice of dimension
in PCA. In: Hardle, W., Simar, L. (eds.) Computer Intensive Methods in Statistics, pp. 167–
176. Physica-Verlag, Heidelberg (1993)
22. University of Notre Dame Computer Vision Research Laboratory: Biometrics database dis-
tribution (2007), http://www.nd.edu/∼cvrl/UNDBiometricsDatabase.html
23. Papatheodorou, T.: 3D Face Recognition Using Rigid and Non-Rigid Surface Registration.
PhD thesis, VIP Group, Department of Computing, Imperial College, London University
(2006)
24. Cootes, T.: The AR face database 22 point markup (N/A),
http://www.isbe.man.ac.uk/∼bim/data/tarfd markup/
tarfd markup.html
25. Martinez, A., Benavente, R.: The AR face database (2007),
http://cobweb.ecn.purdue.edu/∼aleix/aleix face DB.html
Locating Facial Features with an Extended
Active Shape Model
Stephen Milborrow and Fred Nicolls
Department of Electrical Engineering

University of Cape Town, South Africa
www.milbo.users.sonic.net
Abstract. We make some simple extensions to the Active Shape Model

of Cootes et al. [4], and use it to locate features in frontal views of
upright faces. We show on independent test data that with the extensions
the Active Shape Model compares favorably with more sophisticated
methods. The extensions are (i) fitting more landmarks than are actually
needed (ii) selectively using two- instead of one-dimensional landmark
templates (iii) adding noise to the training set (iv) relaxing the shape
model where advantageous (v) trimming covariance matrices by setting
most entries to zero, and (vi) stacking two Active Shape Models in series.
1 Introduction
Automatic and accurate location of facial features is difficult. The variety of
human faces, expressions, facial hair, glasses, poses, and lighting contribute to
the complexity of the problem.
This paper focuses on the specific application of locating features in unob-
structed frontal views of upright faces. We make some extensions to the Active
Shape Model (ASM) of Cootes et al. [4] and show that it can perform well in
this application.
2 Active Shape Models

This section describes Active Shape Models [8].
A landmark represents a distinguishable point present in most of the images
under consideration, for example, the location of the left eye pupil (Fig. 1). We
locate facial features by locating landmarks.
A set of landmarks forms a shape. Shapes are represented as vectors: all the
x- followed by all the y-coordinates of the points in the shape. We align one
shape to another with a similarity transform (allowing translation, scaling, and
rotation) that minimizes the average euclidean distance between shape points.
The mean shape is the mean of the aligned training shapes (which in our case
are manually landmarked faces).
The ASM starts the search for landmarks from the mean shape aligned to
the position and size of the face determined by a global face detector. It then

Locating Facial Features with an Extended Active Shape Model 505
Fig. 1. A face with correctly positioned landmarks. This image is from the BioID
set [15].
repeats the following two steps until convergence (i) suggest a tentative shape by
adjusting the locations of shape points by template matching of the image tex-
ture around each point (ii) conform the tentative shape to a global shape model.
The individual template matches are unreliable and the shape model pools the
results of the weak template matchers to form a stronger overall classifier. The
entire search is repeated at each level in an image pyramid, from coarse to fine
resolution.
It follows that two types of submodel make up the ASM: the profile model
and the shape model.
The profile models (one for each landmark at each pyramid level) are used
to locate the approximate position of each landmark by template matching.
Any template matcher can be used, but the classical ASM forms a fixed-length
normalized gradient vector (called the profile) by sampling the image along a
line (called the whisker ) orthogonal to the shape boundary at the landmark.
During training on manually landmarked faces, at each landmark we calculate
the mean profile vector ḡ and the profile covariance matrix Sg . During searching,
we displace the landmark along the whisker to the pixel whose profile g has lowest
Mahalanobis distance from the mean profile ḡ:
M ahalanobisDistance = (g − ḡ)T S−1

g (g − ḡ). (1)
The shape model specifies allowable constellations of landmarks. It gener-

ates a shape x̂ with
x̂ = x̄ + Φb (2)
where x̄ is the mean shape, b is a parameter vector, and Φ is a matrix of
selected eigenvectors of the covariance matrix Ss of the points of the aligned
training shapes. Using a standard principal components approach, we model as
much variation in the training set as we want by ordering the eigenvalues λi of
Ss and keeping an appropriate number of the corresponding eigenvectors in Φ.
We use a single shape model for the entire ASM but scale it for each pyramid
level.
506 S. Milborrow and F. Nicolls
We can generate various shapes with Equation 2 by varying the vector pa-
rameter b. By keeping the elements of b within limits (determined during model
building) we ensure that generated face shapes are lifelike.
Conversely, given a suggested shape x, we can calculate the parameter b that
allows Equation 2 to best approximate x with a model shape x̂. Cootes and
Taylor [8] describe an iterative algorithm that gives the b and T that minimizes
distance(x, T(x̄ + Φb)) (3)
where T is a similarity transform that maps the model space into the image
space.
3 Related Work
Active Shape Models belong to the class of models which after a shape is situated
near an image feature interact with the image to warp the shape to the feature.
They are deformable models likes snakes [16], but unlike snakes they use an
explicit shape model to place global constraints on the generated shape. ASMs
were first presented by Cootes et al. [3]. Cootes and his colleagues followed with
a succession of papers cumulating in the classical ASM described above [8] [4].
Many modifications to the classical ASM have been proposed. We mention
just a few. Cootes and Taylor [6] employ a shape model which is a mixture of mul-
tivariate gaussians, rather than assuming that the shapes come from the single
gaussian distribution implicit in the shape model of the classical ASM. Romdhani
et al. [22] use Kernel Principal Components Analysis [23] and a Support Vector
Machine. Their software trains on 2D images, but models non-linear changes to
face shapes as they are rotated in 3D. Rogers and Graham [21] robustify ASMs
by applying robust least-squares techniques to minimize the residuals between
the model shape and the suggested shape. Van Ginneken et al. [12] take the tack
of replacing the 1D normalized first derivative profiles of the classical ASM with
local texture descriptors calculated from “locally orderless images” [17]. Their
method automatically selects the optimum set of descriptors. They also replace
the classical ASM profile model search (using Mahalanobis distances) with a k-
nearest-neighbors classifier. Zhou et al. [25] estimate shape and pose parameters
using Bayesian inference after projecting the shapes into a tangent space. Li and
Ito [24] build texture models with AdaBoosted histogram classifiers. The Active
Appearance Model [5] merges the shape and profile model of the ASM into a
single model of appearance, and itself has many descendants. Cootes et al. [7]
report that landmark localization accuracy is better on the whole for ASMs than
AAMs, although this may have changed with subsequent developments to the
AAM.
4 Extensions to the ASM

We now look at some extensions to the classical ASM. Figure 3 (Sec. 5.1) shows
the increase in performance for each of these extensions.
2.5
point−to−point err relative to 68−point−model
2.0
1.5
1.0
0.5
0 10 20 30 40 50 60 70
number of landmarks
Fig. 2. Mean error versus number of landmarks
4.1 Number of Landmarks
A straightforward way to improve the mean fit is to increase the number of

landmarks in the model (Fig. 2). Fitting a landmark tends to help fitting other
landmarks, so results are improved by fitting more landmarks than are actually
needed. Search time increases roughly linearly with the number of landmarks.
Fig. 2 was constructed as follows from the XM2VTS [19] set of manually
landmarked faces. For a given number (from 3 to 68) of landmarks, that number
of landmarks was chosen randomly from the 68 in the XM2VTS test. With the
chosen landmarks, a model was built and tested to give one gray dot. This was
repeated ten times for each number of landmarks. The black line shows the mean
error for each number of landmarks.
4.2 Two Dimensional Profiles
The classical ASM uses a one-dimensional profile at each landmark, but using
two-dimensional “profiles” can give improved fits. Instead of sampling a one-
dimensional line of pixels along the whisker, we sample a square region around
the landmark. Intuitively, a 2D profile area captures more information around
the landmark and this information if used wisely should give better results.
During search we displace the sampling region in both the “x” and “y” direc-
tions, where x is orthogonal to the shape edge at the landmark and y is tangent to
the shape edge. We must rely on the face being approximately upright because 2D
profiles are aligned to the edges of the image. The profile covariance matrix Sg of a
set of 2D profiles is formed by treating each 2D profile matrix as a long vector (by
appending the rows end to end), and calculating the covariance of the vectors.
Any two dimensional template matching scheme can be used, but the au-
thors found that good results were obtained using gradients over a 13x13 square
around the landmark, after prescaling faces to a constant width of 180 pixels.
The values 13 and 180 were determined during model building by measurements
on a validation set, as were all parameter values in this paper (Sec. 5).
Gradients were calculated with a 3x3 convolution mask ((0,0,0),(0,-2,1),(0,1,0))
and normalized by dividing by the Frobenius norm of the gradient matrix. The
effect of outliers was reduced by applying a mild sigmoid transform to the elements
xi of the gradient matrix: xi = xi /(abs(xi ) + constant).
Good results were obtained using 2D profiles for the nose and eyes and sur-
rounding landmarks, with 1D profiles elsewhere.
4.3 Adding Noise during Training
The XM2VTS set used for training (Sec. 5) contains frontal images of mostly
caucasian working adults and is thus a rather limited representation of the variety
of human faces. A shape model built with noise added to the training shapes
helps the trained model generalize to a wider variety of faces. Good results can
be obtained with the following techniques:
1. Add gaussian noise with a standard deviation of 0.75 pixels to the x- and y-
positions of each training shape landmark. In effect, this increases variability
in the training set face shapes.
2. Randomly choose the left or the right side each face. Generate a stretching
factor for each face from a gaussian distribution with a standard deviation
of 0.08. Stretch or contract the chosen side of the face by multiplying the x
position (relative to the face center) of each landmark on that side by 1 + .
This is roughly equivalent to rotating the face slightly.
4.4 Loosening Up the Shape Model
In Equation 2, the constraints on the generated face shape are determined by the
number of eigenvectors neigs in Φ and the maximum allowed values of elements
in the parameter vector b. When conforming the shape suggested √ by the profile
models to the shape model, we clip each element bi of b to bmax λi where λi is
the corresponding eigenvalue The parameters neigs and bmax are global constants
determined during model building by parameter selection on a validation set.
See [8] for details.
The profile models are most unreliable when starting the search (for exam-
ple, a jaw landmark can snag on the collar), but become more reliable as the
search progresses. We can take advantage of this increase in reliability with two
modifications to the standard ASM procedure described above. The first mod-
ification sets neigs and bmax for the final pyramid level (at the original image
scale) to larger values. The second sets neigs and bmax for the final iteration at
each pyramid level to larger values. In both cases the landmarks at that stage of
the search tend to be already positioned fairly accurately, for the given pyramid
level. It is therefore less likely that the profile match at any landmark is grossly
mispositioned, allowing the shape constraints to be weakened.
These modifications are effective for 2D but not for 1D profiles. The 1D profile
matches are not reliable enough to allow the shape constraints to be weakened.
4.5 Trimming the Profile Covariance Matrices

For 2D profiles, calculation of the Mahalanobis distances dominates the overall
search time. We can reduce this time (with little or no effect on landmark location
accuracy) by “trimming” the covariance matrix.
The covariance between two pixels in a profile tends to be much higher for
pixels that are closer together. This means that we can ignore covariances for
pixels that are more than 3 pixels apart, or equivalently clear them to 0. Clear-
ing elements of a covariance matrix may result in a matrix that is no longer
positive definite (which is necessary for a meaningful Mahalanobis distance cal-
culation in Equation 1). We therefore adjust the trimmed matrix to a “nearby”
positive definite matrix. This can be done by iterating the following procedure a
few times: perform a spectral decomposition of the trimmed covariance matrix
A = QΛQT , set zero or negative eigenvalues in Λ to a small positive number,
reconstruct the matrix from the modified Λ, and re-trim. A suitable “small posi-
tive number” is iter nbr × abs(min(eig vals(A))). More rigorous ways of forcing
positive definiteness are presented in Gentle [11] and in Bates and Maechler [1].
Trimming the covariance matrices in conjunction with a sparse matrix multi-
plication routine roughly halves the overall search time.
4.6 Stacking Models

Accurate positioning of the start shape is crucial — it is unlikely that an ASM
search will recover completely from a bad start shape. One way of better posi-
tioning the start shape is to run two ASM searches in series, using the results
of the first search as the start shape for the second search. In practice is suffices
to use 1D profiles for the first model and to start the second model at pyramid
level 1, one level below full size. Stacking helps the worst fits, where the start
shape is often badly mis-positioned, but has little effect where the start shape
is already well positioned.
Before giving experimental results we briefly review model assessment in more
general terms [13]. The overall strategy for selecting parameters is
1. for each model parameter
2. for each parameter value
3. train on a set of faces
4. evaluate the model by using it to locate landmarks
5. select the value of the parameter that gives the best model
6. test the final model by using it to locate landmarks.
0.08
me17
(mean point to point me17
0.07
error / eye distance) time
for all BioID faces
found by the Viola
0.06
Jones detector
0.05
0.5
0.4
search time
0.04
0.3 including face

0.2 detection time
0.1 (secs on a
0.03
0 3GHz Pentium)
20 point model (1D)
68 point model (1D)
2D profiles
training noise
loosen shape model
trimmed
stacked
subsection: 4.1 4.1 4.2 4.3 4.4 4.5 4.6
Fig. 3. Relative performance of various models
Two processes are going on here: model selection which estimates the perfor-
mance of different models in order to choose one (steps 2-5 above), and model
assessment which estimates the final model’s performance on new data (step 6
above). We want to measure the generalization ability of the model, not its abil-
ity on the set it was trained on, and therefore need three independent datasets
(i) a training set for step 3 above (ii) a parameter selection or validation set for
step 4 above, and (iii) a test set for step 6 above.
For the training set we used the XM2VTS [19] set. We effectively doubled the
size of the training set by mirroring images, but excluded faces that were of poor
quality (eyes closed, blurred, etc.).
For the validation set we used the AR [18] set. So, for example, we used the
AR set for choosing the amount of noise discussed in section 4.3. We minimized
overfitting to the validation set by using a different subset of the AR data for
selecting each parameter. Subsets consisted of 200 randomly chosen images.
For the test set we used the BioID set [15]. More precisely, the test set is
those faces in the BioID set that were successfully found by the OpenCV [14]
implementation of the Viola-Jones face detector (1455 faces, which is 95.7% of
the total 1521 BioID faces).
We used manual landmarks for these three sets from the FGNET project [9].
Cross validation on a single data set is another popular approach. We did not
use cross validation because three datasets were available and because of the
many instances of near duplication of images within each dataset.
1.0
0.8
proportion
0.6
0.4
stacked model
0.2
CLM Fig 4(c)

0.0
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14

me17 (BioID with Viola Jones)
Fig. 4. Comparison to Constrained Local Model [10]
Following Cristinacce [10], we present results in terms of the me17 measure.

The me17 is calculated by taking the mean of the euclidean distances between
each of the 17 internal face points located by the search and the corresponding
manually landmarked point. This mean is normalized by dividing by the distance
between the manually landmarked eye pupils. We use only 17 of the 20 manually
landmarked BioID points because the 3 points near the sides of the face have a
high variability across human landmarkers.
5.1 Relative Performance

Fig. 3 summarizes and compares results from applying each of the modifica-
tions described in this paper. Each graph point represents the me17 averaged
over all faces in the test set, for the given model. Each model incorporates the
improvements of the models to its left but not to its right.
For example, the entry labeled 4.2 2D profiles shows results for the model
described in section 4.2. The model uses the 2D profiles described in that section
and incorporates the techniques prior to but not subsequent to section 4.2. The
graph shows that using 2D profiles decreases the me17 from 0.061 to 0.055 but
increases the search time from 110 to 410 ms.
The mean me17 of the final stacked model is 66% of the initial 20 point
model. The biggest single improvement comes from adding more points to the
model, followed by using 2D profiles, followed by stacking. A different test set or
different evaluation order would give somewhat different results, but the graph
is representative of the relative performance of the various modifications.
5.2 Comparison to Previously Published Results

Fig. 4 compares the best model in this paper, the stacked model (section 4.6), to
the Constrained Local Model presented in Cristinacce and Cootes [10]. Briefly,
the Constrained Local Model is similar to an Active Appearance Model [5], but
instead of modeling texture across the whole face it models a set of local feature
templates. During search, the feature templates are matched to the image using
an efficient shape constrained search. The model is more accurate and more
robust than the original Active Appearance Model.
The results in Cristinacce and Cootes’ paper appear to be the best previously
published facial landmark location results and are presented in terms of the me17
on the BioId set, which makes a direct comparison possible. The dotted curve in
Fig. 4 reproduces the curve in Fig. 4(c) in their paper. The figure shows that the
stacked model on independent data outperforms the Constrained Local Model.
The median me17 for the stacked model is 0.045 (2.4 pixels), the best me17 is
0.0235 (1.4 pixels), and the worst is 0.283 (14 pixels). The long right hand tail
of the error distribution is typical of ASMs.

This paper presented some modifications to the Active Shape Model which make
it competitive with more sophisticated methods of locating features in frontal
views of upright faces.
A few simple rules of thumb for improving ASMs became apparent. You can get
better fits by adding more landmarks. You can discard most elements of the co-
variance matrices for increased speed without loss of quality. You get better results
with a better start shape, and you can do this by running two models in series.
The techniques used in this paper are fairly standard. Perhaps the main contri-
bution of the paper is assembling them together in a sound fashion. Advantages
of the techniques are their simplicity and applicability for use in conjunction
with other methods. For example, extra landmarks and stacked models would
possibly improve the performance of the Constrained Local Model shown in
Fig. 4.
The results are still not as good as manual landmarks. Further work will inves-
tigate combining multiple profiling techniques at each landmark with a decision
tree [2] or related method. Here the training process would try different profiling
techniques at each landmark and build a decision tree (for each landmark) that
would select or combine techniques during searching.
Additional documentation and source code to reproduce the results in this
paper can be found at this project’s web site [20].
References
1. Bates, D., Maechler, M.: Matrix: A Matrix package for R. See the nearPD function
in this R package for methods of forcing positive definiteness (2008),
http://cran.r-project.org/web/packages/Matrix/index.html
2. Breiman, Friedman, Olshen, Stone: Classification and Regression Trees. Wadsworth
(1984)
3. Cootes, T.F., Cooper, D.H., Taylor, C.J., Graham, J.: A Trainable Method of
Parametric Shape Description. BMVC 2, 54–61 (1991)
4. Cootes, T.F., Taylor, C.J., Cooper, D.H., Graham, J.: Active Shape Models —
their Training and Application. CVIU 61, 38–59 (1995)
5. Cootes, T.F., Edwards, G.J., Taylor, C.J.: Active Appearance Models. In:
Burkhardt, H., Neumann, B. (eds.) ECCV 1998. LNCS, vol. 1407, pp. 484–498.
6. Cootes, T.F., Taylor, C.J.: A Mixture Model for Representing Shape Variation.
Image and Vision Computing 17(8), 567–574 (1999)
7. Cootes, T.F., Edwards, G.J., Taylor, C.J.: Comparing Active Shape Models with
Active Appearance Models. In: Pridmore, T., Elliman, D. (eds.) Proc. British Ma-
chine Vision Conference, vol. 1, pp. 173–182 (1999)
8. Cootes, T.F., Taylor, C.J.: Technical Report: Statistical Models of Appearance for
Computer Vision. The University of Manchester School of Medicine (2004),
www.isbe.man.ac.uk/∼ bim/refs.html
9. Cootes, T.F., et al.: FGNET manual annotation of face datasets (2002),
www-prima.inrialpes.fr/FGnet/html/benchmarks.html
10. Cristinacce, D., Cootes, T.: Feature Detection and Tracking with Constrained Local
Models. BMVC 17, 929–938 (2006)
11. Gentle, J.E.: Numerical Linear Algebra for Applications in Statistics. Springer,
Heidelberg (1998); See page 178 for methods of forcing positive definiteness
12. van Ginneken, B., Frangi, A.F., Stall, J.J., ter Haar Romeny, B.: Active Shape
Model Segmentation with Optimal Features. IEEE-TMI 21, 924–933 (2002)
13. Hastie, T., Tibshirani, R., Friedman, J.: The Elements of Statistical Learning: Data
Mining, Inference, and Prediction. Springer, Heidelberg (2003); See chapter 7 for
methods of model assessment
14. Intel: Open Source Computer Vision Library. Intel (2007)
15. Jesorsky, O., Kirchberg, K., Frischholz, R.: Robust Face Detection using the Haus-
dorff Distance. AVBPA 90–95 (2001)
16. Kass, M., Witkin, A., Terzopoulos, D.: Snakes: Active Contour Models. IJCV 1,
321–331 (1987)
17. Koenderink, J.J., van Doorn, A.J.: The Structure of Locally Orderless Images.
IJCV 31(2/3), 159–168 (1999)
18. Martinez, A.M., Benavente, R.: The AR Face Database: CVC Tech. Report 24
(1998)
19. Messer, K., Matas, J., Kittler, J., Luettin, J., Maitre, G.: XM2VTS: The Extended
M2VTS Database. AVBPA (1999)
20. Milborrow, S.: Stasm software library (2007),
http://www.milbo.users.sonic.net/stasm
21. Rogers, M., Graham, J.: Robust Active Shape Model Search. In: Heyden, A., Sparr,
G., Nielsen, M., Johansen, P. (eds.) ECCV 2002. LNCS, vol. 2353, pp. 517–530.
22. Romdhani, S., Gong, S., Psarrou, A.: A Multi-view Non-linear Active Shape Model
using Kernel PCA. BMVC 10, 483–492 (1999)
23. Scholkopf, S., Smola, A., Muller, K.: Nonlinear Component Analysis as a Kernel
Eigenvalue Problem. Neural Computation 10(5), 1299–1319 (1998)
24. Li, Y., Ito, W.: Shape Parameter Optimization for AdaBoosted Active Shape
Model. ICCV 1, 251–258 (2005)
25. Zhou, Y., Gu, L., Zhang, H.J.: Bayesian Tangent Shape Model: Estimating Shape
and Pose Parameters via Bayesian Inference. In: CVPR (2003)
Dynamic Integration of Generalized Cues for
Person Tracking
Kai Nickel and Rainer Stiefelhagen
Universität Karlsruhe (TH), InterACT

Am Fasanengarten 5, 76131 Karlsruhe, Germany
Abstract. We present an approach for the dynamic combination of mul-

tiple cues in a particle filter-based tracking framework. The proposed
algorithm is based on a combination of democratic integration and lay-
ered sampling. It is capable of dealing with deficiencies of single fea-
tures as well as partial occlusion using the very same dynamic fusion
mechanism. A set of simple but fast cues is defined, which allow us to
cope with limited computational resources. The system is capable of au-
tomatic track initialization by means of a dedicated attention tracker
permanently scanning the surroundings.
1 Introduction
Visual person tracking is a basic prerequisite for applications in fields like surveil-
lance, multimodal man-machine interaction or smart spaces. Our envisioned
scenario is that of an autonomous robot with limited computational resources
operating in a common space together with its users. The tracking range varies
from close distance, where the portrait of the user spans the entire camera im-
age, to far distance, where the entire body is embedded in the scene. In order to
tackle the problem, we present a multi-cue integration scheme within the frame-
work of particle filter-based tracking. It is capable of dealing with deficiencies of
single features as well as partial occlusion by means of the very same dynamic
fusion mechanism. A set of simple but fast cues is defined, allowing to cope with
limited on-board resources.
The choice of cues is a crucial design criterion for a tracking system. In real-
world applications, each single cue is likely to fail in certain situations such
as occlusion or background clutter. Thus, a dynamic integration mechanism is
needed to smooth over a temporary weakness of certain cues as long as there
are other cues that still support the track. In [1], Triesch and Von Der Malsburg
introduced the concept of democratic integration that weights the influence of
the cues according to their agreement with the joint hypothesis. The competing
cues in [1] were based on different feature types such as color, motion, and shape.
In this paper, we use the principle of democratic integration in a way that also
includes the competition between different regions of the target object. We show
that this allows us to deal with deficiencies of single feature types as well as with
partial occlusion using one joint integration mechanism.

Dynamic Integration of Generalized Cues for Person Tracking 515
The combination of democratic integration and particle filters has been ap-
proached before by Spengler and Schiele [2]. In their work, however, the inte-
gration weights were held constant, thus falling short behind the real power of
democratic integration. This has also been pointed out by Shen et al. [3], who
did provide a cue quality criterion for dynamic weight adaptation. This criterion
is formulated as the distance of the tracking hypothesis based on all cues and the
hypothesis based on the cue alone. The problem with this formulation is that,
due to resampling, the proposal distribution is generally strongly biased toward
the final hypothesis. Thus, even cues with uniformly mediocre scores tend to
agree well with the joint mean of the particle set. We therefore propose a new
quality criterion based on weighted MSE that prefers cues which actually focus
their probability mass around the joint hypothesis.
Democratic integration combines cues in the form of a weighted sum. In a
particle filter framework, this means that all cues have to be evaluated simulta-
neously for all particles. As pointed out by Pérez et al. [4], this can be alleviated
by layered sampling, if the cues are ordered from coarse to fine. In the proposed
algorithm, we therefore combine two-stage layered sampling with democratic in-
tegration on each stage to increase efficiency by reducing the required number
of particles.
For each object to be tracked, we employ one dedicated Condensation-like
tracker [5]. By using separate trackers instead of one single tracker running in
a joint state space, we accept the disadvantage of potentially not being able
to find the global optimum. On the other hand, however, we thereby avoid
the exponential increase in complexity that typically prevents the use of par-
ticle filters in high-dimensional state spaces. There are a number of approaches
dealing with this problem, such as Partitioned Sampling [6], Trans-dimensional
MCMC [7], or the Hybrid Joint-Separable formulation [8]. Although these ap-
proximations reduce the complexity of joint state space tracking significantly,
they still require noticeably more computational power than the separate tracker
approach.
The remainder of this paper is organized as follows: In section 2, we briefly
describe the concept of particle filters and layered sampling. In section 3 we
present our multi-cue integration scheme, which is the main contribution of this
paper. It is followed, in section 4, by the definition of the cues that we actually
use in the live tracking system. In section 5, the multi-person tracking logic
including automatic track initialization and termination is described. Finally,
section 6 shows the experiments and results.
2 Particle Filter-Based Tracking
Particle filters represent a generally unknown probability density function by a

(1..n) (1..n) (i)
set of random samples st and associated weights πt with πt = 1. In
one of the simplest cases, the Condensation algorithm [5], the evolution of the
particle set is a two-stage process which is guided by the observation and the
state evolution model:
516 K. Nickel and R. Stiefelhagen
1. The prediction step (including resampling): randomly draw n new particles

from the old set with a likelihood proportional to the particle weights. Prop-
agate the new particles by applying the state evolution model p(st |st−1 ).
2. The measurement step: adjust the weights of the new particles with respect
(i) (i)
to the current observation zt : πt ∝ p(zt |st ).
The final tracking hypothesis for the current time instance ŝt can be obtained
from the sample set as (i) (i)
ŝt = πt st (1)
i=0..n
2.1 Layered Sampling

Assuming that z is made up of M conditionally independent measurement
sources, i.e. different cues, the observation likelihood of a particle s can be fac-
torized as follows1 :
p(z|s) = p(zm |s) (2)
m=1..M
According to [4], the state evolution can then be decomposed into M successive
intermediate steps:

p(st |st−1 ) = pM (st |sM−1 ) · · · p1 (s1 |st−1 )ds1 · · · dsM−1 (3)
where s1 · · · sM−1 are auxiliary state vectors2 . In case of a Gaussian evolution

model, this corresponds to a fragmentation into M successive steps with lower
variances. Then, [4] make the approximation that the likelihood for the m-th
cue p(zm |s) can be incorporated after applying the m-th state evolution model
pm (sm |sm−1 ). This leads to a layered sampling strategy, where at the m-th stage
new samples are simulated from a Monte Carlo approximation of the distribu-
tion pm (sm |sm−1 )π m−1 with an associated importance weight π m ∝ p(zm |sm ).
As [4] point out, the benefit of layered sampling arises in cases where the cues
can be ordered from coarse to fine, e.g. the first cue produces a reliable but
rough estimation for the state, while the second cue produces a sharp and peaky
estimation. Then, the layered sampling approach will effectively guide the search
in the state space, with each stage refining the result from the previous stage.
We will apply layered sampling in section 5 in combination with the multi-cue
integration scheme described in the following.
3 Dynamic Multi-cue Integration

In the Bayesian tracking formulation used in this work, cues have the function of
scoring the match between a state vector s and the observation z. A joint score
combining the cues from the set of all cues C can be formulated as a weighted sum
1
The time index t is omitted for the sake of brevity wherever possible.
2
We omit the according formula for splitting the proposal distribution, because in
Condensation, the proposal distribution is identical to the evolution model.

p(z|s) = rc pc (z|s), (4)
c∈C
where pc (z|s) is the the
single-cue observation model, and rc is the mixture
weight for cue c, with c rc = 1.
Democratic integration [1] is a mechanism to dynamically adjust the mixture
weights rc , termed reliabilities, with respect to the agreement of the single cue c
with the joint result. For each cue, a quality measure qc is defined that quantifies
the agreement, with values close to zero indicating little agreement and values
close to one indicating good agreement. The reliabilities are updated after each
frame by a leaky integrator using the normalized qualities:
qc
rct+1 = (1 − τ )rct + τ (5)
c qc
with the parameter τ controlling the speed of adaptation.
3.1 Cue Quality Measure

In the original paper [1], tracking is implemented as an exhaustive search over a
support map, and the quality measure is defined over a single cue’s support
map. In [3], a different quality measure dedicated to particle filters is pro-
posed: Based on the current particle set s(1..n) and an auxiliary set of weights
(1..n)
πc ∝ pc (z|s(1..n) ), a tracking hypothesis ŝc is generated according to eq. 1 and
compared to the joint hypothesis ŝ. The L2 -norm distance |ŝc - ŝ|2 is normalized
by means of a sigmoid function and then taken as quality measure.
Although this formulation looks straightforward, there is a problem associated
with it: Imagine the common situation where a cue finds little or no support at
all, and therefore assigns uniform likelihood values to all of the particles. Let’s
assume further that the state of the target has not changed for a while, so that
in consequence, due to resampling, the particle distribution is equally spread
around the actual state. In this case, the cue-based hypothesis ŝc will be close
to ŝ resulting in a high quality value qc despite the fact that the cue is actually
not at all able to locate the target.
To address this problem, we need a quality measure that quantifies how well the
probability mass agglomerates around the joint hypothesis ŝ. The inverse mean-
(i)
square error ( i πc |s(i) − ŝ|22 )−1 of the particle set weighted with the respective
cue’s weights πc meets this requirement, but is dependent on the actual location of
the particles. We eliminate this dependency by relating the cue’s MSE to the MSE
of a hypothetical baseline cue which assigns uniform weights n1 to each particle.
Because a good cue is not only supposed to converge to the target location but
also to assign high values to the target, we multiply the term with the cue’s non-
normalized response at the joint hypothesis pc (z|ŝ). Thus, we come to the follow-
ing formulation for a universal cue quality measure in the context of particle-filter
based tracking:

|s − ŝ|λ
1 (i)
qc = i=1..n n(i) pc (z|ŝ) (6)
i=1..n πc |s
(i) − ŝ|λ
The exponent λ > 0 can be used to tweak the volatility of the quality measure:
high values of λ emphasize the quality difference between cues whereas low values
produce more similar qualities for all cues.
3.2 Generalized Cue Competition

In order to allow for a fruitful combination, the set of cues should be orthogonal
in the sense that different cues tend to fail under different circumstances. One
way to reduce the chances of co-occurrence of failure is to use different cue-
specific feature transformations F (z) like motion, color, or shape. Failure of one
feature can thus more likely be compensated by other features.
pc (z|s) = pc (F (z)|s) (7)
The other option to generate orthogonal cues is to use different state model
transformations A(s):
pc (z|s) = pc (z|A(s)) (8)
This is motivated by the fact that cues relying on certain aspects of the state
vector may still be used while other aspects of the state are not observable. In
our implementation, A(s) represents a certain projection from state space to
image space, i.e. a certain image sub-region of the target. This is useful in a
situation, where due to partial occlusion one region of the target object can be
observed, while another region cannot.
In this work, we aim at combining the advantages of both strategies, i.e.
dynamically combining cues that are based on different feature types as well as
dynamically weighting cues that focus on different regions of the target but are
based on the same feature type. Therefore, we use a generalized definition of the
cues c = (F , A) that comprises different feature types F (z) and different state
transformations A(s):
pc (z|s) = pF ,A (F (z)|A(s)), (9)
All cues in this unified set will then compete equally against each other, guided
by the very same integration mechanism. Thus, the self-organizing capabilities
of democratic integration can be used to automatically select the specific feature
types as well as the specific regions of the target that are most suitable in the
current situation.
3.3 Cue Model Adaptation

Certain cues, such as color models or templates, allow for online adaptation of
their internal parameters to better match the current target appearance. In [1],
this adaptation is described as a continuous update process with a fixed time
constant τc :
Pct+1 = (1 − τc )Pct + τc P̂c , (10)
with Pc being the internal parameters of cue c, and P̂c being new parameters
acquired from the image region given by the joint hypothesis ŝ.
One of the issues with adaptation is due to the fact that after an update
step, the cue is not guaranteed to perform better than before. Although the
update step always results in a higher score for the prototype region at ŝ, it can
happen that the updated model produces higher scores also for other regions
than the correct one. This actually reduces the cue’s discriminative power and,
in consequence, its reliability rc . We therefore propose the following test to be
carried out before accepting an update:
1. Calculate qc (eq. 6) using the new parameters P̂c
2. Perform the update step (eq. 10) only if qc > qc
4 Fast Cues for 3D Person Tracking

In the targeted application, one or more people are to be tracked in the vicinity
of an autonomous robot featuring a calibrated stereo camera. As the on-board
computational resources are strictly limited, cues have to be found that rely on
features that can be evaluated rapidly. Our proposed cues are based on the follow-
ing well-known feature types: difference image, color histogram back-projection,
Haar-feature cascades and stereo correlation.
As motivated in section 3.2, we use different transformations of the state vec-
tor in order to handle partial occlusion: some cues focus on the human head
region only, whereas other cues concentrate on the torso and legs region re-
spectively. These regions are determined using the ”3-box model” of the human
body depicted in Fig. 1. The real-world extensions of the 3 cuboids are geared
to model an average human being; their relative positions depend on the height
of the head above the ground plane.
By combining the feature types motion, color and stereo with the 3 different
body parts, and by using 4 different detectors, we obtain a total number of 13
cues that will be described in the following. Fig. 2 shows the different feature
types as a snapshot from a test sequence.
In the following, we will use F (z) to denote a feature map, i.e. an image in
which the intensity of a pixel is proportional to the presence of a feature, such
as color or motion. An image region corresponding to a state vector s will be
denoted as A(s) (see Fig. 1), |A(s)| is the size of the region, and A(s) F (z) is
the sum of pixel values of F (z) inside region A(s). All regions in our system are
rectilinear bounding boxes, so the sum can be calculated efficiently by means of
4 table lookups in the integral image [9].
4.1 Motion Cues

The difference image M(z) is generated by pixel-wise thresholding the absolute
difference of the current frame’s and the previous frame’s intensity images. For
a moving object, we can expect high values of M(z) in the region A(s) around
object’s current location s. The motion cue’s observation likelihood is given as:

A(s) M(z) A(s) M(z)
pM,A (z|s) = · (11)
|A(s)| M(z)
Fig. 1. The 3-box model of the human body: the state vector s is transformed into the
image space as the projection of a cuboid representing either the head, torso, or leg
region. The projection of the cuboid is approximated by a rectilinear bounding box.
camera image motion color person 1 color person 2 detectors
Fig. 2. Snapshot from a test sequence showing the different feature types. In this
visualization, the color support maps for head, torso and legs of the respective person
are merged into the RGB-channels of the image. The tracking result is superimposed.
The left factor seeks to maximize the amount of foreground within the region.
The right factor seeks to cover all foreground pixels in the image. It prevents the
motion cue from preferring tiny regions filled with motion, while ignoring the rest.
We employ 3 motion cues, termed m-h, m-t and m-l, dedicated to either
the head, torso or legs region as depicted in Fig. 1. We rely on the ability of
the integration mechanism (see section 3) to automatically cancel the influence
of the motion cues in case of camera motion. This is justified by the fact that
the agreement of the motion cues with the final tracking hypothesis will drop
whenever large portions of the image exceed the threshold.
4.2 Color Cues

We employ three adaptive color cues c-h, c-t, c-l for the three body regions.
For each of the cues, we use a 3-dimensional histogram with 16 bins per channel
in RGB color space that automatically adapts to the target region using the
mechanism described in section 3.3. A second histogram is built from the entire
image; it acts as a model for the background color distribution. The quotient his-
togram of the target histogram and the background histogram is back-projected
and forms the support map C(z) for a color cue. The observation likelihood is
given analogous to eq. 11 as:

A(s) C(z) A(s) C(z)
pC,A (z|s) = · (12)
|A(s)| C(z)
4.3 Detector Cues

For each particle, the head region A(s) is projected to the image plane, and the
bounding box of the projection is being classified with a single run of the detector
proposed by [9]. The detectors are organized stages that need to be passed one by
one in order to produce a positive response. The ratio m(A(s)) = ( stages passed ω
stages total )
can be interpreted as a confidence value for the detection, with the exponent ω
controlling the steepness of decay for each stage that is not being passed.
In order to smooth the scores of nearby particles, we define the score of a
particle s as the highest overlap between its region A(s) and all the positively
classified regions A ∈ {A(s(i) )|A(s(i) )is face}i=1..n by any of the other particles:
pD,A (z|s) = maxA m(A ) · d(A , A(s)), (13)
with d being a distance metric based on rectangle overlap.
We use four detector cues in total: one for frontal faces (d-f), one for left
(d-l) and one for right (d-r) profile faces, and one for upper bodies (d-u).
Implementation and training of the detectors is based on [10,11] as provided by
the OpenCV library.
4.4 Stereo Correlation Cues

In traditional stereo processing [12], a dense disparity map is generated by ex-
haustive area correlation followed by several post-filtering steps. Apart from the
computational effort of generating a dense disparity map, there is another, more
fundamental problem, namely the choice of the size of the area correlation win-
dow. If a windows is too large, it smoothes over fine details, if it is too small, it
tends to produce noisy results. In our approach, we can avoid these issues: we
use the entire target region A(s) as correlation window and search for optimal
correlation along the epipolar lines. The adaptive correlation window is thus as
large as possible and as small as necessary given the current size of the target.
The response of the stereo cue is given by the distance of the discovered
ˆ
disparity d(A(s)) and the hypothesized disparity d(A(s)):

−1
ˆ
pS,A (z|s) = 1 + |d(A(s)) − d(A(s))|κ , (14)
with κ being a parameter to control the volatility of the cue. The complexity
ˆ
of the local search for the disparity d(A(s)) is scale-invariant because it can
be implemented efficiently by means of integral images, as proposed by [13] for
dense disparity calculation. We employ 3 stereo cues, one for the head (s-h),
torso (s-t), and legs (s-l).
5 Multi-person Tracking Logic

As motivated in the introduction, we run one dedicated particle filter for each
person to be tracked. The state space consists of the location and velocity of the
person’s head centroid in 3-dimensional space: s(i) = (x, y, z, ẋ, ẏ, ż). The state
evolution p(st |st−1 ) is implemented as a 1st-order motion model with additive
Gaussian noise on the velocity components.
5.1 Democratic Integration and Layered Sampling

Multi-cue integration as described by eq. 4 is suitable for all kinds of cues that
are optional for the target, which means that the target may or may not have
the property implied by the cue at the moment. There are, however, cues that
are indispensable as track foundation and therefore must not be ruled out by
the fusion mechanism. In our application, this applies to the stereo cues: a track
should not be able to exist if it is not supported by at least one of the stereo
cues as these represent strict geometrical constraints. One way of ensuring this
would be to multiply the response of the stereo cues with the response of the
regular cues. A more efficient way is layered sampling as described in section 2.1.
We use it to evaluate the stereo cues CS ⊂ C before the regular cues CR ⊂ C,
as shown in Fig. 3. By evaluating the mandatory stereo cues first, followed by a
1,(1..n)
resampling step, the resulting particle set st clusters only in those regions
of the state space that are well supported by the stereo cues. The particles on
the second stage can now more efficiently evaluate the regular cues.
1st layer:
(1..n) (1..n)
– resample st−1 wrt. πt−1
– propagate with partial evolution model (cf. eq. 3)
1,(1..n) 1,(i) (i)
st ←− p1 (st |st−1 )
1,(i) 1,(i)
– evaluate stereo cues: πt ∝ c∈CS rc pc (z|st )
1,(i) 1,(i) 1,(i)
– apply collision penalty: πt ←− πt − v(st )
2nd layer:
1,(1..n) 1,(1..n)
– resample st wrt. πt
– propagate with partial evolution model (cf. eq. 3)
(1..n) (i) 1,(i)
st ←− p2 (st |st )
(i) (i)
– evaluate regular cues: πt ∝ c∈CR rc pc (z|st )
Dem. integration:
(i) (i)
– calculate track hypothesis ŝt = i πt st
– update reliabilities (cf. eqs. 5 and 6)
1,(1..n) 1,(1..n)
rc∈CS ←− ŝt , st , πt
(1..n) (1..n)
rc∈CR ←− ŝt , st , πt
Fig. 3. Two-stage layered sampling algorithm with democratic cue integration
Apart from the geometrical constraints implied by the stereo cues, there is
another strict constraint, namely the collision penalty, which is enforced in the
1st layer of the algorithm in Fig. 3. The function v(s) penalizes particles that
are close to those tracks with a higher track quality than the current track (see
following section). Thereby, we guarantee mutual exclusion of tracks.
5.2 Automatic Track Initialization
The question of when to spawn a new tracker and when to terminate a tracker
that has lost its target is of high importance, and can become more difficult than
the actual tracking problem. We define the quality measure for a tracker to be
the joint response from both stereo and regular cues at the tracker’s hypothesis ŝ:

Q(ŝ) = rc pc (z|ŝ) · rc pc (z|ŝ) (15)
c∈CS c∈CR
The final quality measure Q is a result of temporal filtering with a time

constant ν:
Qt+1 = (1 − ν)Qt + νQ(ŝ) (16)
Trackers falling below a certain threshold Q < Θ for a certain amount of time Γ
will be discarded.
In order to discover potential targets, we employ an additional tracker termed
attention tracker. The attention tracker permanently scans the state space,
searching for promising regions. It is, however, repelled by existing tracks by
means of the collision penalty v(s). Unlike regular trackers, 50% of the attention
tracker’s particles are not propagated by means of the state evolution model,
but are drawn randomly from the state space. This guarantees good coverage of
the state space and still allows some clustering around interesting regions. As
the attention tracker must remain general, its cues’ parameters are not allowed
to adapt. After each frame, the distribution of the attention tracker’s particles is
clustered with a k-means algorithm. If one of the clusters exceeds the threshold
Θ, a new regular tracker is initialized at that location.
6 Experiments
We evaluated the algorithm on 11 test sequences, some of them including camera

motion. The head’s bounding box was manually labeled in 3 of the 15 frames per
second to obtain the ground truth. In total, 2312 frames were labeled. From the
3D tracking output, a head-sized box was projected to the image and compared
to the manually labeled box. If there was no overlap between the boxes, the
frame was counted as a miss and a false positive. As the tracker was free to
output 0, 1 or more tracks, the number of misses and false positives do not need
to be identical.
Overall, the tracker showed solid performance throughout the experiments.
Critical situations for track loss – although it occurred rarely – were periods in
which the user rested virtually motionless either at far distance or in a turned-
away position, so that in consequence the detectors did not respond. Then, the
tracker had to rely solely on the automatically initialized color models, which
were not always significant enough. Another issue were phantom tracks that were
triggered by non-human motion or false detections. They were sometimes kept
alive by the color models which adapted to the false positive region. In most
Table 1. Tracking results on the evaluation set
misses false pos.

Fixed reliabilities (baseline) 10.2% 8.1%
Dynamic integration (Shen et al.) 11.1% 8.8%
Dynamic integration (equation 6) 4.6% 4.6%
Fig. 4. Evolution of cue reliabilities in an example sequence. The three stereo cues
constitute the first layer of the algorithm, their reliabilities sum up to 1. The remaining
ten cues are used in layer 2 and sum up to 1 likewise. In the beginning of the interval,
the subject approaches the camera. While he is walking (frame 250), the motion cues
for legs and torso (M-L,M-T) contribute significantly to the track. At around frame
300, the subject’s legs disappear, and in consequence the reliabilities of all leg-related
cues (M-L, C-L, S-L) drop automatically. While the subject is standing in front of the
camera (frames 300-500), the frontal face detection cue D-F and the the head color cue
C-H dominate the track. The influence of the head color cue C-H drops dramatically,
when the subject turns around (frame 520) and walks in front of the wooden pinboard,
which has a skin-color like appearance.
cases, however, this could be avoided by the adaptation control mechanisms

described in section 3.3.
Table 1 shows the results of the evaluation. The proposed algorithm was
compared to a baseline system with static reliabilities, and to a system using
the dynamic cue quality formulation by Shen et al. [3]. The proposed algorithm
clearly outperforms the two other systems both in the number of misses and
false positives. Figure 4 discusses the evolution of cue reliabilities for an example
sequence.
In the implementation, we made the following additions to the algorithm: The

color cue for the head region (c-h) is expected to converge to general skin color;
its model is therefore shared among all trackers. An new box-type for the upper
body detector was used; it comprises head and upper half of the torso. To avoid
dominance, we limited the range for a cue’s influence to 0.03 ≤ rc ≤ 0.6. We
found, however, that these situations rarely occur. Boxes that get projected
outside the visible range or that are clipped to less than 20% of their original
size, are scored with a minimum score of 0.001. The approximate runtime of
the algorithm was 30ms per frame for an empty scene, plus another 10ms per
person being tracked. These values are based on an image size of 320×240 pixels,
and a 2.4GHz Pentium CPU. The most important parameter values are given
in Table 2.
Table 2. Parameters of the algorithm
# of particles per tracker n = 150

Track threshold / timeout Θ = 0.25, Γ = 2s
Track quality time constant ν = 0.33
Cue reliability time constant τ = 0.25
Color update time constant τc = 0.01
Cue tweaking factors λ = 4, κ = 4, ω = 10
7 Conclusion
We have presented a new approach for dynamic cue combination in the frame-
work of particle filter-based tracking. It combines the concepts of democratic
integration and layered sampling and enables a generalized kind of competition
among cues. With this method, cues based on different feature types compete di-
rectly with cues based on different target regions. In this way, the self-organizing
capabilities of democratic integration can be fully exploited. In an experimental
validation, the proposed new cue quality measure has been shown to improve
the tracking performance significantly.
Acknowledgments
This work has been funded by the German Research Foundation (DFG) as part
of the Sonderforschungsbereich 588 ”Humanoid Robots”.
References
1. Triesch, J., Malsburg, C.V.D.: Democratic integration: Self-organized integration
of adaptive cues. Neural Comput. 13(9), 2049–2074 (2001)
2. Spengler, M., Schiele, B.: Towards robust multi-cue integration for visual tracking.
Machine Vision and Applications 14, 50–58 (2003)
3. Shen, C., Hengel, A., Dick, A.: Probabilistic multiple cue integration for particle
filter based tracking. In: International Conference on Digital Image Computing -
Techniques and Applications, pp. 309–408 (2003)
4. Pérez, P., Vermaak, J., Blake, A.: Data fusion for visual tracking with particles.
Proceedings of the IEEE 92(3), 495–513 (2004)
5. Isard, M., Blake, A.: Condensation–conditional density propagation for visual
tracking. International Journal of Computer Vision 29(1), 5–28 (1998)
6. MacCormick, J., Blake, A.: A probabilistic exclusion principle for tracking multiple
objects. International Journal of Computer Vision 39(1), 57–71 (2000)
7. Smith, K., Gatica-Perez, D., Odobez, J.M.: Using particles to track varying num-
bers of interacting people. In: IEEE Conf. on Computer Vision and Pattern Recog-
nition, Washington, DC, USA, pp. 962–969 (2005)
8. Lanz, O.: Approximate bayesian multibody tracking. IEEE Transactions on Pat-
tern Analysis and Machine Intelligence 28(9), 1436–1449 (2006)
9. Viola, P., Jones, M.: Robust real-time object detection. In: ICCV Workshop on
Statistical and Computation Theories of Vision (July 2001)
10. Lienhart, R., Maydt, J.: An extended set of haar-like features for rapid object
detection. In: ICIP, vol. 1, pp. 900–903 (September 2002)
11. Kruppa, H., Castrillon-Santana, M., Schiele, B.: Fast and robust face finding via
local context. In: IEEE Intl. Workshop on Visual Surveillance and Performance
Evaluation of Tracking and Surveillance (October 2003)
12. Scharstein, D., Szeliski, R.: A taxonomy and evaluation of dense two-frame stereo
correspondence algorithms. IJCV 47(1/2/3), 7–42 (2002)
13. Veksler, O.: Fast variable window for stereo correspondence using integral images.
In: IEEE Conf. on Computer Vision and Pattern Recognition, pp. 556–561 (2003)
Extracting Moving People from Internet Videos
Juan Carlos Niebles1,2 , Bohyung Han3 , Andras Ferencz3 , and Li Fei-Fei1

1
Princeton University, Princeton NJ, USA
2
Universidad del Norte, Colombia
3
Mobileye Vision Technologies, Princeton NJ, USA
Abstract. We propose a fully automatic framework to detect and ex-

tract arbitrary human motion volumes from real-world videos collected
from YouTube. Our system is composed of two stages. A person detector
is first applied to provide crude information about the possible locations
of humans. Then a constrained clustering algorithm groups the detections
and rejects false positives based on the appearance similarity and spatio-
temporal coherence. In the second stage, we apply a top-down pictorial
structure model to complete the extraction of the humans in arbitrary
motion. During this procedure, a density propagation technique based
on a mixture of Gaussians is employed to propagate temporal informa-
tion in a principled way. This method reduces greatly the search space
for the measurement in the inference stage. We demonstrate the initial
success of this framework both quantitatively and qualitatively by using
a number of YouTube videos.
1 Introduction
Human motion analysis is notoriously difficult because human bodies are highly
articulated and people tend to wear clothing with complex textures that obscure
the important features needed to distinguish poses. Uneven lighting, clutter,
occlusions, and camera motions cause significant variations and uncertainties.
Hence it is no surprise that the most reliable person detectors are built for
upright walking pedestrians seen in typically high quality images or videos.
Our goal in this work is to be able to automatically and efficiently carve out
spatio-temporal volumes of human motions from arbitrary videos. In particular,
we focus our attention on videos that are typically present on internet sites such
as YouTube. These videos are representative of the kind of real-world data that
is highly prevalent and important. As the problem is very challenging, we do
not assume that we can find every individual. Rather, our aim is to enlarge
the envelope of upright human detectors by tracking detections from typical to
atypical poses. Sufficient data of this sort will allow us in the future to learn
even more complex models that can reliably detect people in arbitrary poses.
Two example sequences and the system output are shown in Fig. 1.
Our first objective is to find moving humans automatically. In contrast to
much of the previous work in tracking and motion estimation, our framework
does not rely on manual initialization or a strong a priori assumption on the

528 J.C. Niebles et al.
0026 0038 0049 0054 0064
0061 0067 0077 0086 0108
Fig. 1. Two example outputs. Our input videos are clips downloaded from YouTube and
thus are often low resolution, captured by hand-held moving cameras, and contain a
wide range of human actions. In the top sequence, notice that although the boundary
extraction is somewhat less accurate in the middle of the jump, the system quickly
recovers once more limbs become visible.
number of people in the scene, the appearance of the person or the background,
the motion of the person or that of the camera. To achieve this, we improve a
number of existing techniques for person detection and pose estimation, leverag-
ing on temporal consistency to improve both the accuracy and speed of existing
techniques. We initialize our system using a state-of-the-art upright pedestrian
detection algorithm [1]. While this technique works well on average, it produces
many false positive windows and very often fails to detect. We improve this sit-
uation by building an appearance model and applying a two-pass constrained
clustering algorithm [2] to verify and extend the detections.
Once we have these basic detections, we build articulated models following
[3,4,5] to carve out arbitrary motions of moving humans into continuous spatio-
temporal volumes. The result can be viewed as a segmentation of the moving
person, but we are not aiming to achieve pixel-level accuracy for the extraction.
Instead, we offer a relatively efficient and accurate algorithm based on the prior
knowledge of the human body configuration. Specifically, we enhance the speed
and potential accuracy of [4,5] by leveraging temporal continuity to constrain
the search space and applying semi-parametric density propagation to speed up
evaluation.
The paper is organized as follows. After reviewing previous work in the area
of human motion analysis in Section 1.1, we describe the overall system archi-
tecture in Section 2. Two main parts of our system, person detection/clustering
and extraction of moving human boundaries, are presented in Sections 3 and 4,
respectively. Finally, implementation details and experimental results are de-
scribed in Section 5.
1.1 Related Work
Body Tracking. The most straightforward method to track humans is to

consider them as blobs and use generic object tracking methods such as [6,7].
More complex methods attempt to model the articulation of the body
Extracting Moving People from Internet Videos 529
[8,9,10,11,12,13,14,15]. Most of these methods rely on a manual initialization,

strong priors to encode the expected motion, a controlled or very simple envi-
ronment with good foreground/background separation, and/or seeing the motion
from multiple cameras.
Pedestrian Detection and Pose Estimation. Several fairly reliable pedes-

trian detection algorithms have been developed recently [1,16,17,18,19,20]. How-
ever, these methods typically deal with upright persons only, and the detection
accuracy is significantly reduced by even moderate pose variations. Furthermore,
these algorithms offer little segmentation of the human, providing only a bound-
ing box of the body.
To model body configurations, tree shaped graphical models have shown
promising results [3,4,5]. These generative models are often able to find an ac-
curate pose of the body and limbs. However, they are less adept at making a
discriminative decision: is there a person or not? They are typically also very
expensive computationally in both the measurement and inference steps.
We build on these models and address the discrimination problem by ini-
tializing detections with an upright person detector. To improve computational
efficiency, our algorithm exploits temporal information and uses more efficient
semi-parametric (Gaussian mixture) representations of the distributions.
Based on similar intuitions, [21] uses temporal information to reduce the
search space progressively in applying pictorial structures to videos. Ren et al.
[22] takes another approach to human pose estimation in videos by casting the
figure tracking task into a foreground/background segmentation problem using
multiple cues, though the algorithm seems to rely on objects having a high con-
trast with the background.
2 System Architecture
Our system consists of two main components. The first component generates
object-level hypotheses by coupling a human detector with a clustering algo-
rithm. In this part, the state of each person, including location, scale and trajec-
tory, is obtained and used to initialize the body configuration and appearance
models for limb-level analysis. Note that in this step two separate problems
– detection and data association – are handled simultaneously, based on the
spatio-temporal coherence and appearance similarity.
The second component extracts detailed human motion volumes from the
video. In this stage, we further analyze each person’s appearance and spatio-
temporal body configuration, resulting in a probability map for each body part.
We have found that we can improve both the robustness and efficiency of the
algorithm by limiting the search space of the measurement and inference around
the modes of the distribution. To do this, we model the density function as a
mixture of Gaussians in a sequential Bayesian filtering framework [23,24,25].
The entire system architecture is illustrated in Fig. 2. More details about each
step are described in the following two sections.
Fig. 2. Overall system
The focus of our work is to extract arbitrarily complex human motions from
YouTube videos that involve a large degree of variability. We face several difficult
challenges, including:
1. Compression artifacts and low quality of videos
2. Multiple shots in a video
3. Unknown number of people in each shot or sequence
4. Unknown human motion and poses
5. Unknown camera parameters and motion
6. Background clutter, motion and occlusions
We will refer back to these points in the rest of the paper as we describe how
the components try to overcome them.
3 People Detection and Clustering

As Fig. 2 shows, our system begins with a step to estimate location, scale, and
trajectories of moving persons. This step is composed of the following two parts.
3.1 Initial Hypothesis by Detection

We first employ an human detection algorithm [1] to generate a large num-
ber of hypotheses for persons in a video. This method, which trains a classifier
cascade using boosting of HOG features to detect upright standing or walking
people, has serious limitations. It only detects upright persons and cannot han-
dle arbitrary poses (challenge 4). The performance is degraded in the presence of
compression artifacts (challenge 1). Moreover, since it does not use any temporal
information, the detection is often inconsistent and noisy, especially in scale. It
is, therefore, difficult to reject false positives and recover miss-detections effec-
tively. The complexity increases dramatically when multiple people are involved
(challenge 3). This step, therefore, serves only as an initial hypotheses proposal
stage. Additional efforts are required to handle various exceptions.
3.2 People Clustering

The output of the person detector is a set of independent bounding boxes; there
are no links for the same individual between detections. The detections also
have significant noise, false alarms and miss-detections especially due to the
low quality of the video (challenge 1). In order to recover from these problems,
we incorporate a clustering algorithm based on the temporal and appearance
coherence of each person. The goal of clustering in our system is to organize all
correct detections into groups, where each corresponds to a single person in the
sequence (challenge 3), while throwing away false alarms. To achieve this, we
apply a constrained clustering paradigm [2] in two hierarchical stages, adding
both positive (should link) edges and negative (can not link) constraints between
the detections. See Fig. 3 for an example.
Stage 1. In the first stage, we focus on exploiting the temporal-coherence cue by

associating detections from multiple frames with the help of a low-level track-
ing algorithm [7]. When the first detection is observed, a low-level tracker is
initialized with the detected bounding box. A new detection in a consequent
frame is assigned to an existing track if it coherently overlaps with the tracker
predictions. In this case, we reinitialize the tracker with the associated detection
bounding box. When no existing track can explain the new detection, a new
track is created. Due to the complexity of the articulated human body, a low-
level tracker is susceptible to drift from the person. We thus limit the temporal
life of the tracker by counting the number of frames after the last detection and
terminating the track at the last detection if the maximum gap (e.g. 100 frames)
is surpassed. Very small clusters with few detections are discarded. The clusters
produced in this first stage are almost always correct but over-segmented tracks
(see Fig. 3 (b)). This is because the person detector often fails to detect a person
in the video for many frames in a row – especially when the person performs
some action that deviates from an upright pose.
Stage 2. The stage 2 agglomerative constrained clustering views the stage 1

clusters as atomic elements, and produces constraints between them with pos-
itive weights determined by appearance similarity and negative constraints de-
termined by temporal/positional incompatibility.
For the appearance similarity term, we select multiple high-scoring detection
windows for each stage 1 cluster, and generate probability maps for the head
and torso locations using a simple two-part pictorial structure [4]. We use these
results to (1) remove false detections by rejecting clusters that have unreliable
head/torso estimation results (e.g., high uncertainty in the estimated head and
torso locations), and (2) generate a weighted mask for computing color histogram
descriptors for both the head and the torso. The appearance of the person in
each cluster is then modeled with the color distributions of head and torso.
After the second pass of our hierarchical clustering, we obtain one cluster per
person in the sequence. Fig. 3 (c) illustrates the final clustering result, which shows
that three different persons and their trajectories are detected correctly, despite
the fact that the appearance of these individuals are very similar (Fig. 3 (d)).
300 300 300
250 250 250
200 200 200
150 150 150
100 100 100
50 50 50
0 0 0
0 200 0 200 0 200
(a) (b) (c) (d)
Fig. 3. Human detection and clustering result. From noisy detections, three tracks of
people are identified successfully by filling gaps and removing outliers. (In this figure,
the horizontal and vertical axis are the x locations and frame numbers, respectively.)
(a) Original detection (b) Initial clusters after step 1 (c) Final clusters (d) Example
images of three similar people that correctly clustered into different groups.
4 Extracting Spatio-temporal Human Motion Volume

We now have a cluster for each person, with a detection bounding box giving
the location, scale, and appearance in some subset of the frames. Our goal is to
find the body configuration for all the frames of the cluster (challenge 4), both
where we have detections and where we do not. In this section, we discuss how
to extract human body pose efficiently in every frame.
The existing algorithms for human motion analysis based on belief propaga-
tion such as [3,5] typically require exhaustive search of the input image because
minimal (or no) temporal information is employed for the inference. Our idea is
to propagate the current posterior to the next frame for the future measurement.
4.1 Overview
We summarize here the basic theory for the belief propagation and inference
in [3,4]. Suppose that each body part pi is represented with a 4D vector of
(xi , yi , si , θi ) – location, scale and orientation. The entire human body B is
composed of m parts, i.e. B = {p1 , p2 , . . . , pm }. Then, the log-likelihood given
the measurement from the current image I is

L(B|I) ∝ Ψ (pi − pj ) + Φ(pi ) (1)
(i,j)∈E i
where Ψ (pi − pj ) is the relationship between two body parts pi and pj , and Φ(pi )
is the observation for body part pi . E is a set of edges between directly connected
body parts. Based on the given objective function, the inference procedure by
message passing is characterized by

Mi (pj ) ∝ Ψ (pi − pj )O(pi ) (2)
pj

O(pi ) ∝ Φ(pi ) Mk (pi ) (3)
k∈Ci
where Mi (pj ) is the message from part pi to pj , O(pi ) is the measurement of

part pi , and Ci is a set of children of part pi . The top-down message from part
pj to pi for the inference is defined by

P (pi |I) ∝ Φ(pi ) Ψ (pi − pj )P (pj |I), (4)
pj
which generates the probability map of each body part in the 4D state.
Based on this framework, we propose a method to propagate the density
function in the temporal domain in order to reduce search space and temporally
consistent results. The rest of the section describes the details of our algorithm.
4.2 Initialization
The first step for human body extraction is to estimate an initial body configu-
ration and create a reliable appearance model. The initial location of the human
is given by the method presented in Section 3. Note that the bounding box pro-
duced by the detection algorithm does not need to be very accurate since most
of the background area will be removed by further processing. Once a potential
human region is found, we apply a pose estimation technique [4] based on the
same pictorial structure and obtain the probability map of the configuration of
each body part through the measurement and inference step. In other words, the
output of this algorithm is the probability map Pp (u, v, s, θ) for each body part
p, where (u, v) is location, s is scale and θ is orientation. A sample probability
map is presented in Fig. 4 (b)-(d). Although this method creates accurate proba-
bility maps for each human body part, it is too computationally expensive to be
used in video processing. Thus, we adopt this algorithm only for initialization.
4.3 Representation of Probability Map
The original probability map Pp is represented by a discrete distribution in 4D

space for each body part. There are several drawbacks of the discrete density
function. First of all, it requires a significant amount of memory space, which
is proportional to the image size and granularity of the orientations and scales,
even if most of the pixels in the image have negligible probabilities. Second,
the propagation of a smooth distribution is more desirable for the measurement
in the next step since a spiky discrete density function may lose a significant
number of potentially good candidates by sampling.
Instead of using the non-parametric and discrete probability map, we em-

ploy a parametric density function. However, finding a good parametric density
function is not straightforward, especially when the density function is highly
multi-modal as in human body. In our problem, we observe that the probability
map for each orientation is mostly uni-modal and close to a Gaussian distribu-
tion1 . We employ a mixture of N Gaussians for the initialization of human body
configuration, where N is the number of different orientations.
(k) (k)
Denote by xi and ωi (i = 1, . . . , n) the location and weight of each point in
the k-th orientation probability map. Let θ(k) be the orientation corresponding
the k-th orientation map. The mean (m(k) ), covariance (P(k) ) and weight (κ(k) )
of the Gaussian distribution for the k-th orientation map is then given by

(k) (k)

x(k) ωi x i
m(k) = = i (5)
θ(k) θ(k)
(k) (k)
− m(k) )(xi − m(k) ) 0
(k)
i ωi (xi
Vx 0
P(k) = = (6)
0 Vθ 0 Vθ
(k) (k)
κ(k) = xi / xi (7)
i k i
where Vx and Vθ are (co)variance matrices in spatial and angular domain, re-
spectively. The representation of the combined density function based on the
entire orientation maps is given by

1 2
N
1 κ(k)
fˆ(x) = exp − D x, x , P
(k) (k)
(8)
(2π)d/2 i=1 | P(k) |1/2 2

where D2 x, x(k) , P(k) is the Mahalanobis distance from x to x(k) with
covariance P(k) .
Although we simplify the density functions for each orientation as a Gaussian,
it is still difficult to manage them in an efficient way especially because the
number of components will increase exponentially when we propagate the density
to the next time step. We therefore adopt Kernel Density Approximation (KDA)
[26] to further simplify the density function with little sacrifice in accuracy. KDA
is a density approximation technique for a Gaussian mixture. The algorithm finds
the mode locations of the underlying density function by an iterative procedure,
such that a compact mixture of Gaussians based on the detected mode locations
is found.
Fig. 4 presents the original probability map and our approximation using a
mixture of Gaussians for each body part after the pose estimation. Note that
the approximated density function is very close to the original one and that the
multi-modality of the original density function is well preserved.
1
Arms occasionally have significant outliers due to their flexibility. A uni-modal
Gaussian fitting may result in more error here.
(a) (b) torso (c) upper-left arm (d) upper-left leg

Fig. 4. Comparison between the true probability map for the pose estimation (left in
each sub-figure) and its Gaussian mixture approximation (right) for each body part.
The approximated density functions are propagated for the measurement in the next
time step. Note that our approximation results look much wider since different scales
in the color palette are applied for better visualization.
4.4 Measurement, Inference and Density Propagation

Fast and accurate measurement and inference are critical in our algorithm. As
shown in Eq. (2) and (3), the bottom-up message is based on all the informa-
tion up to the current node as well as the relative configuration with the parent
node. Exhaustive search is good for generating the measurement information
at all possible locations. However, it is very slow and, more importantly, the
performance for the inference may be affected by spurious observations; noisy
measurement incurred by an object close to or moderately far from the real
person may corrupt the inference process. A desirable reduction of search space
not only decreases computation time, but also improves the accuracy. The search
space for measurement and inference is determined by a probability density func-
tion characterizing potential state of human body, where a mixture of Gaussians
are propagated in sequential Bayesian filtering framework [23,24,25].
In our method, we perform local search based on the spatio-temporal infor-
mation. We first diffuse the posterior density function from the previous frame,
which is done analytically thanks to the Gaussian mixture representation. Based
on the diffused density, locally dense samples are drawn to make measurements
and a discrete density function is constructed. Note that inference is performed
using the discrete density function. But a parametric representation of density
function is propagated to the next time step for the measurement. After the infer-
ence, the pose estimation density function is converted to a mixture of Gaussians
by the method described in Section 4.3. The posterior is given by the product
of the diffused density and the pose estimation density function in the current
frame. This step is conceptually similar to the integration of the measurement
and inference history (temporal smoothing). We denote by X and Z the state
and observation variable in the sequential Bayesian filtering framework, respec-
tively. The posterior at the time step t of the state is given by the product of
two Gaussian mixture as follows:
p(Xt |Z1:t ) ∝ p(Zt |Xt )p(Xt |Z1:t−1 ) (9)
9N :⎛N ⎞
1 2
= N (κi , xi , Pi ) ⎝ N (τj , yj , Qj )⎠ , (10)

i=1 j=1
Algorithm 1. Moving human body extraction

1: Apply human detection algorithm to a sequence
2: Apply clustering algorithm based on the detection. Create the initial body config-
uration and appearance at the first detection. Also, obtain the number of people
in the video.
3: Construct pose estimation density function for each body part based on a mixture
of Gaussians in the first frame, where it is also used as the posterior.
4: while not the end of sequence do
5: Go to the next frame
6: Diffuse the posterior of the previous frame
7: Perform the measurement and inference with the locally dense samples
8: Create a Gaussian mixture with the discrete pose estimation distribution
9: Compute the posterior by multiplying diffusion and pose estimation density
10: if there exists the detection of the same person then
11: Reinitialize the appearance and body configuration of the person (optional)
12: end if
13: end while
(a) (b) (c) (d)
Fig. 5. Density functions in one step of the human motion extraction. (a) Original
frame (cropped for visualization) (b) Diffused density function (c) Measurement and
inference results (d) Posterior (Note that the probability maps for all orientations are
shown in a single image by projection.)
where N (·) represents a Gaussian distribution with parameters of weight, mean,

and covariance. The first and second terms in the right hand side represent diffu-
sion and pose estimation density function, respectively. Note that the product of
two Gaussian mixtures is still a Gaussian mixture, but it causes the exponential
increase of the number of components. So KDA is required again to maintain a
compact representation of the density function.
The density propagation algorithm for inference is summarized in Algorithm 1,
and illustrated in Fig. 5.
5 Experiments
In order to evaluate our proposed approach, we have collected a dataset of 50
sequences containing moving humans downloaded from YouTube. The sequences
contain natural and complex human motions and various challenges mentioned
Table 1. Precision-Recall Table: Performance comparison
Detection only Detection & Clustering Full model

Prec Rec F Prec Rec F Prec Rec F
0.89 0.31 0.46 0.89 0.30 0.45 0.83 0.73 0.78
0.90 0.25 0.39 0.91 0.24 0.38 0.87 0.62 0.72
Rate 0.92 0.19 0.32 0.92 0.19 0.32 0.86 0.51 0.64
0.93 0.16 0.27 0.94 0.15 0.27 0.92 0.43 0.58
0.94 0.13 0.24 0.94 0.13 0.23 0.88 0.32 0.46
in Section 2. Many videos have multiple shots (challenge 2), so we divide the
original videos into several pieces based on the shot boundary detection, which
is performed by global color histogram comparison with threshold [27]. We deal
with each shot as a separate video. We have made this dataset public and it can
be found at http://vision.cs.princeton.edu/projects/extractingPeople.html.
Instead of 4D state space for human body configuration, 3D state space for
location and orientation is utilized and scale is determined based on the detection
size. Although person detector is not so accurate in scale estimate, the extraction
algorithm is robust enough to handle some variations of the scale. Also, the gaps
between detections are not generally long, and it is not often the case that we
observe significant change in scale between two detections.
The measurement is based on edge template and color histogram as in [4], but
search space for the measurement is significantly reduced. Fig. 5 (b) illustrates
the search space reduction, where low density areas are not sampled for the
observations.
We evaluate the retrieval performance of our system in terms of the precision-
recall measures. For each sequence, we have generated ground-truth by manually
labeling every human present in each frame with a bounding box. We compare
the precision-recall rates at three stages of our system: pedestrian detection only
[1], people detection and clustering, and the full model. For a fixed threshold of
the human detector, we obtain the three precision-recall pairs in each row of
Table 1. Our full system provides the highest performance in terms of the F-
measure2 . This reflects the fact that our system achieves much higher recall rates
by extracting non-upright people beyond the pedestrian detections.
We also evaluate the performance of our system in terms of the segmentation
of the moving people. We create ground-truth for the spatial support of the
moving people in the form of binary masks. We have labeled a random sample
of 122 people from our 50 sequences. The evaluation of the pose estimation is
performed at frames td , td +5 and td +10, where td is a frame containing a pedes-
trian detection, and no detections are available in [td + 1, td + 10]. The average
accuracies are 0.68, 0.68 and 0.63 respectively. Note that the accuracy decrease
in the extracted person mask is moderate, and the temporal error propagation
is small.
2
The F-measure is defined [28] as: 2 · (precision · recall)/(precision + recall).
0058 0335 0552 0567 0597
0013 0063 0136 0195 0204
0010 0024 0048 0057 0130
0000 0010 0145 0170 0366
0003 0024 0035 0061 0116
0001 0111 0131 0372 0830
0000 0008 0014 0022 0033
0000 0015 0046 0257 0278
0043 0097 0114 0212 0310
0328 0372 0381 0410 0451
Fig. 6. Experimental results for various sequences. Each row corresponds to a sep-
arate sequence and two failure examples are illustrated in the last two rows. Please
visit http://vision.cs.princeton.edu/projects/extractingPeople.html for more
sample videos.
The results for several YouTube videos are presented in Fig. 6. Various general
and complex human motions are extracted with reasonable accuracy, but there
are some failures that are typically caused by inaccurate measurements. In a
PC with a 2.33 GHz CPU, our algorithm requires around 10-20 seconds for the
measurement and inference of each person, one order of magnitude faster than
the full search method of [4].

We presented a method to automatically extract human motion volumes from
natural videos. Our system achieves promising results although many improve-
ments can still be made. Our future work is to make detection/tracking and
pose estimation module interact more closely to create positive feedback and
improve the quality of estimation. Currently, the measurement is based only on
the top-down pictorial structure, but we plan to incorporate bottom-up cues for
more robust and efficient processing. We also aim to build a large data set with
detailed labeling for human motion, which would be very helpful resource for
human motion analysis research [29,30,31].
References
1. Laptev, I.: Improvements of object detection using boosted histograms. In: BMVC,
Edinburgh, UK, vol. III, pp. 949–958 (2006)
2. Klein, D., Kamvar, S.D., Manning, C.D.: From instance-level constraints to space-
level constraints: Making the most of prior knowledge in data clustering. In: ICML
(2002)
3. Felzenszwalb, P., Huttenlocher, D.: Pictorial structures for object recognition.
IJCV 61, 55–79 (2005)
4. Ramanan, D.: Learning to parse images of articulated objects. In: NIPS, Vancouver,
Canada (2006)
5. Ramanan, D., Forsyth, D., Zisserman, A.: Tracking people by learning their ap-
pearance. PAMI 29, 65–81 (2007)
6. Lucas, B., Kanade, T.: An iterative image registration technique with an applica-
tion to stereo vision. IJCAI, 674–679 (1981)
7. Comaniciu, D., Ramesh, V., Meer, P.: Real-time tracking of non-rigid objects using
mean shift. In: CVPR, Hilton Head, SC, vol. II, pp. 142–149 (2000)
8. Cham, T., Rehg, J.: A multiple hypothesis approach to figure tracking. In: CVPR,
Fort Collins, CO, vol. II, pp. 219–239 (1999)
9. Deutscher, J., Blake, A., Reid, I.: Articulated body motion capture by annealed
particle filtering. In: CVPR, Hilton Head, SC (2000)
10. Han, T.X., Ning, H., Huang, T.S.: Efficient nonparametric belief propagation with
application to articulated body tracking. In: CVPR, New York, NY (2006)
11. Haritaoglu, I., Harwood, D., Davis, L.: W4: Who? When? Where? What? - A real
time system for detecting and tracking people. In: Proc. of Intl. Conf. on Automatic
Face and Gesture Recognition, Nara, Japan, pp. 222–227 (1998)
12. Lee, C.S., Elgammal, A.: Modeling view and posture manifolds for tracking. In:
ICCV, Rio de Janeiro, Brazil (2007)
13. Sigal, L., Bhatia, S., Roth, S., Black, M., Isard, M.: Tracking loose-limbed people.
In: CVPR, Washington DC, vol. I, pp. 421–428 (2004)
14. Sminchisescu, C., Triggs, B.: Covariance scaled sampling for monocular 3D body
tracking. In: CVPR, Kauai, Hawaii, vol. I, pp. 447–454 (2001)
15. Sminchisescu, C., Kanaujia, A., Li, Z., Metaxas, D.: Discriminative density prop-
agation for 3d human motion estimation. In: CVPR, San Diego, CA, vol. I, pp.
390–397 (2005)
16. Leibe, B., Seemann, E., Schiele, B.: Pedestrian detection in crowded scenes. In:
CVPR, San Diego, CA, vol. I, pp. 878–885 (2005)
CVPR, San Diego, CA, vol. I, pp. 886–893 (2005)
18. Tuzel, O., Porikli, F., Meer, P.: Human detection via classification on riemannian
manifolds. In: CVPR, Minneapolis, MN (2007)
19. Viola, P., Jones, M.J., Snow, D.: Detecting pedestrians using patterns of motion
and appearance. In: ICCV, Nice, France, pp. 734–741 (2003)
20. Wu, B., Nevatia, R.: Detection of multiple, partially occluded humans in a single
image by bayesian combination of edgelet part detectors. In: ICCV, Beijing, China,
vol. I, pp. 90–97 (2005)
21. Ferrari, V., Marin-Jimenez, M., Zisserman, A.: Progressive search space reduction
for human pose estimation. In: CVPR, Anchorage, AK (2008)
22. Ren, X., Malik, J.: Tracking as repeated figure/ground segmentation. In: CVPR,
Minneapolis, MN (2007)
23. Arulampalam, S., Maskell, S., Gordon, N., Clapp, T.: A tutorial on particle fil-
ters for on-line non-linear/non-gaussian bayesian tracking. IEEE Trans. Signal
Process. 50, 174–188 (2002)
24. Doucet, A., de Freitas, N., Gordon, N.: Sequential Monte Carlo Methods in Prac-
tice. Springer, Heidelberg (2001)
25. Han, B., Zhu, Y., Comaniciu, D., Davis, L.: Kernel-based bayesian filtering for
object tracking. In: CVPR, San Diego, CA, vol. I, pp. 227–234 (2005)
26. Han, B., Comaniciu, D., Zhu, Y., Davis, L.: Sequential kernel density approxima-
tion and its application to real-time visual tracking. PAMI 30, 1186–1197 (2008)
27. Lienhart, R.: Reliable transition detection in videos: A survey and practitioner’s
guide. International Journal of Image and Graphics 1, 469–486 (2001)
28. Van Rijsbergen, C.J.: Information Retreival. Butterworths, London (1979)
shapes. In: ICCV, Beijing, China, pp. 1395–1402 (2005)
30. Ke, Y., Sukthankar, R., Hebert, M.: Efficient visual event detection using volumet-
ric features. In: ICCV, Beijing, China, pp. 166–173 (2005)
31. Niebles, J.C., Wang, H., Fei-Fei, L.: Unsupervised learning of human action cate-
gories using spatial-temporal words. IJCV 79, 299–318 (2008)
Multiple Instance Boost Using Graph Embedding
Based Decision Stump for Pedestrian Detection
Junbiao Pang1,2,3, Qingming Huang1,2,3 , and Shuqiang Jiang2,3

1
Graduate university of Chinese Academy of Sciences, Beijing, 100190, China
2
Key Lab. of Intelligent Information Processing, Chinese Academy of Sciences (CAS)
3
Institute of Computing Technology, CAS, Beijing, 100190, China
{jbpang,qmhuang,sqjiang}@jdl.ac.cn
Abstract. Pedestrian detection in still image should handle the large

appearance and stance variations arising from the articulated structure,
various clothing of human as well as viewpoints. In this paper, we address
this problem from a view which utilizes multiple instances to represent
the variations in multiple instance learning (MIL) framework. Specifi-
cally, logistic multiple instance boost (LMIBoost) is advocated to learn
the pedestrian appearance model. To efficiently use the histogram fea-
ture, we propose the graph embedding based decision stump for the data
with non-Gaussian distribution. First the topology structure of the ex-
amples are carefully designed to keep between-class far and within-class
close. Second, K-means algorithm is adopted to fast locate the multiple
decision planes for the weak classifier. Experiments show the improved
accuracy of the proposed approach in comparison with existing pedes-
trian detection methods, on two public test sets: INRIA and VOC2006’s
person detection subtask [1].
1 Introduction
Pedestrian detection is a practical requirement of many today’s automated
surveillance, vehicle driver assistance systems and robot vision systems. How-
ever, the issue of large appearance and stance variations accompanied with dif-
ferent viewpoints makes pedestrian detection very difficult. The reasons can be
multifold, such as variable human clothing, articulated human structure and
illumination change, etc. The variations bring various challenges including miss-
alignment problem, which is often encountered in non-rigid object detection.
There exist a variety of pedestrian detection algorithms from the different per-
spectives, directly template matching [2], unsupervised model [3], traditional super-
vised model [4,5,6] and so on. Generally, these approaches cope with “mushroom”
shape – the torso is wider than the legs, which dominates the frontal pedestrian, and
deal with “scissor” shape – the legs are switching in walk, which dominates the lat-
eral pedestrian. However, for some uncommon stances, such as mounting on bike,
they incline to fail. In these conditions, the variations often impair the performance
of these conventional approaches. Fig. 1 shows some false negatives generated by
Dalal et al [4]. These false negatives are typically non-“mushroom” or non-“scissor”
shape, and have large variations between each other.

542 J. Pang, Q. Huang, and S. Jiang
Fig. 1. Some detection results in our method producing fewer false negatives than
Dalal et al do [4]
The key notion of our solution is that the variations are represented within
multiple instances, and the “well” aligned instances are automatically selected
to train a classifier via multiple instance learning (MIL) [7,8]. In MIL, a training
example is not singletons, but is represented as a “bag” where all of the instances
in a bag share the bag’s label. A positive bag means that at least one instance
in the bag is positive, while a negative bag means that all instances in the
bag are negative. To pedestrian detection, the standard scanning window is
considered as the “bag”, a set of sub-images in window are treated as instances.
If one instance is classified as pedestrian, the pedestrian is located in detection
stage. The logistic multiple instance boost (LMIBoost) [9] is utilized to learn the
pedestrian appearance, which assumes the average relationship between bag’s
label and instance’s label.
Considering the non-Gaussian distribution (which dominates the positive and
negative examples) and aims of detection (which are accurate and fast), a graph
embedding based weak classifier is proposed for histogram feature in boosting.
The graph embedding can effectively model the non-Gaussian distribution, and
maximally separate the pedestrians from negative examples in low dimension
space [10]. After feature is projected onto discriminative one dimension manifold,
K-means is utilized to fast locate the multiple decision planes for the decision
stump. The proposed weak classifier has the following advantages: 1) it handles
training examples with any distribution; and 2) it not only needs less computa-
tion cost, but also results in robust boosting classifier. The main contributions
of the proposed algorithm are summarized as following:
– The pose variations are handled by multiple instance learning. The variations
between examples are represented within the instances, and are automati-
cally reduced during learning stage.
– Considering the boost setting, graph embedding based decision stump is
proposed to handle training data with non-Gaussian distribution.
In the next section, related work is briefly summarized. Section 3 introduces

the LMIBoost for solving the variations. Section 4 first introduces the graph
embedding based discriminative analysis, and then presents the multi-channel
decision stump. In section 5, we describe the experimental settings for pedes-
trian detection. Finally the experiment and conclusion sections are provided,
respectively.
Multiple Instance Boost Using Graph Embedding Based Decision Stump 543
2 Related Work
Generally, the “mushroom” or “scissor” shape encourages the use of template
matching and traditional machine learning approach as discussed in section 1.
The contour templates are hierarchically matched via Chamfer matching [2]. A
polynomial support vector machine (SVM) is learned with Haar wavelets as hu-
man descriptor [5] (and variants are described in [11]). Similar to still images,
a real-time boosted cascade detector also uses Haar wavelets descriptor but ex-
tracted from space-time differences in video [6]. In [4], an excellent pedestrian
detector is described by training a linear SVM classifier using densely sampled
histogram of oriented gradients (HOG) feature (this is a variant of Lowe’s SIFT
descriptor [12]). In a similar approach [13], the near real-time detection perfor-
mance is achieved by training a cascade detector using SVM and HOG feature in
AdaBoost. However, their “fixed-template-style” detectors are sensitive to pose
variations. If the pose or appearance of the pedestrian has large change, the
“template”-like methods are doomed to fail. Therefore, more robust feature is
proposed to withstand translation and scale transformation [14].
Several existing publications have been aware of the pose variation problem,
and have handled it by “divide and conquer”– the parts based approach. In [15],
the body parts are explicitly represented by co-occurrences of local orientation
features. The separate detector is trained for each part using AdaBoost. Pedes-
trian location is determined by maximizing the joint likelihood of the part occur-
rences according to the geometric relations. Codebook approach avoids explicitly
modeling the body segments or the body parts, and instead uses unsupervised
methods to find part decompositions [16]. Recently, the body configuration esti-
mation is exploited to improve pedestrian detection via structure learning [17].
However, parts based approaches have two drawbacks. First, different part de-
tector has to be applied to the same image patch. This reduces the detection
speed. Second, labeling and aligning the local parts are tedious and time-costing
work in supervised learning. Therefore, the deformable part model supervised
learns the holistic classifier to coarsely locate the person, and then utilizes part
filters to refine body parts in unsupervised method [18].
The multiple instance learning(MIL) problem is first identified in [8], which
represents ambiguously labeled examples using axis-paralled hyperrectangles.
Previous applications of MIL in vision have focused on image retrieval [19]. The
seemingly most similar work to ours may be the upper-body detection [20].
Viola et al use Noisy-OR boost which assumes that only sparse instances are
upper-body in a positive bag. However, in our pedestrian detection setting, the
instances in a positive bag are all positive, and this facilitates to simply assume
that every instance in a bag contributes equally to the bag’s class label.
In pedestrian detection, the histogram feature (such as SIFT, HOG) is typi-
cally used. The histogram feature can be computed rapidly using an intermediate
data representation called “Integral Histogram” [21]. However, the efficient use
of the histogram feature is not well discussed. In [13], the linear SVM and HOG
feature is used as weak classifier. Kullback-Leibler (K-L) Boost uses the log-ratio
between the positive and negative projected histograms as weak classifier. The
projection function is optimized by maximizing the K-L divergence between the

positive and negative features [22]. SVM has high computational cost and hence
reduces the detection speed. Optimizing the projection function in K-L Boost is
also computationally costly and numerically unstable. Fisher linear discrimina-
tive analysis (FLDA) is used as weak classifier for histogram feature [23]. Despite
the success of FLDA for building weak classifier, it still has the following limita-
tions: it is optimal only in the case that the data for each class are approximate
Gaussian distribution with equal covariance matrix.
Although the histogram feature is projected into one dimension manifold using
the projection functions, the learned manifold does not directly supply classifi-
cation ability. The widely used decision stump is a kind of threshold-type weak
classifier, but a lot of discriminative information is lost [24]. Therefore, the single-
node, multi-channel split decision tree is introduce to exploit the discriminative
ability. In face detection [25], Huang et al use the histogram to approximate the
distributions of the real value feature by dividing the feature into many sub-
regions with equal width in RealBoost. Then a weak classifier based on a look
up table (LUT) function is built by computing the log-ratio on each sub-bins.
However, the equal regions unnecessarily waste decision stump in low discrimi-
native region. In [26], the unequal regions are obtained by exhaustively merging
or splitting the large number of histogram bins via Bayes decision rule. In this
paper, we avoid exhaustive searching and emphasize on fast designing the multi-
channel decision stump via K-means clustering.
3 Logistic Multiple Instance Boost

If pedestrian have uncommon stance, human-centering normalization often pro-
duces miss-aligned examples as illustrated in Fig. 1. Intuitively, some parts of
human can be aligned by shifting the normalization window. Therefore, we aug-
ment the training set by perturbing the training examples. The created instances
can take advantage of all information of the “omega” heads and the rectangle
bodies. Moreover, the augmented training set should cover the possible pose
variations for MIL. Fig. 2 illustrates the proposed approach.
Compared with traditionally supervised learning, an instance in MIL is indexed
with two indices: i which indexes the bag, and j which indexes the instance within
the bag. Given a bag xi , the conditional probability of the bag-level class yi is
1
ni
p(yi |xi ) = p(yij |xij ), (1)
ni j=1
where ni is the number of the instances in the i-th bag, yij is the instance-
level class label for the instance xij . Equation.(1) indicates that every instance
contributes equally to the bag’s label. This simple assumption is suitable for the
instances generated by perturbing around the person. Because the generated
every instance is positive pedestrian image.
The instance-level class probability is given as p(y|x) = 1/(1 + eβx ), where
β is the parameter to be estimated. Controlling the parameter β gives different
Fig. 2. Overview of the multiple instance learning process. The training example is first
converted into a bag of instances. Note that we only generate the instances spatially,
and the instances can also be generated at different scales. Therefore, the resulting
classifier will withstand the translation and scale transformation.
instance-level class probability, which gives different contribution to bag-level

probability. Ideally, the “well” aligned instances should be assigned higher prob-
ability than the non-aligned. Given a collection of N i.i.d bags x1 , . . . , xN , the pa-
rameter β can be estimated by maximizing the bag-level binomial log-likelihood
function
N
L= [yi logp(yi = 1|xi ) + (1 − yi )logp(yi = 0|xi )] . (2)
i
Equation.(2) can not be solved analytically. Xu el al [9] propose an boosting

method to maximize the log-likelihood function. We need to learn a bag-level
function F(x) = m cm fm (x) and the corresponding strong classifier H =
sign(F(x)), where weights c1 , . . . , cM ∈ R, the f is the bag-level weak classi-
fier. The expected empirical loss is
1
N
E[I(F(x) = y)] = − yi F(xi ), (3)
N i=1
where I(·) is the indicator function. We are interesting in wrapping the bag-level
weak classifier f with the instance-level weak classifier f . Using the Equation.(1),
Equation.(3) is converted into the instance-level’s exponential loss Ex Ey|x [e−yf ]
as e−yH ≥ I(H(x) = y), ∀M . One searches for the optimal update cm fm such
that minimizes
% &
Ex Ey|x e−yij Fm−1 (xij )−cm yij fm (xij ) = wi e[(2i −1)cm ] , (4)
i

where i = j 1fm (xij )=yij /ni , wi is the example’s weight. The error i describes
the discrepancy between the bag’s label and instance’s label. The instance in
positive bags with higher score f (xij ) gives higher confidence to the bag’s label,
even though there are some negative instances occurring in the positive bag.
Algorithm. 1 Graph embedding based decision stump

Input:
The training data {hi , yi }, i = 1, . . . , n
Training:
1. Learn the projection matrix P ∈ R1×D by Equation. (4), and project the data
into one dimension manifold {ĥi , yi }, ĥi = P hi .
2. Calculate the clustering center P c = {C1p , . . . , CN
p
p
} and N c = {C1n , . . . , CN
n
n
}
for the positive and negative data via K-means, where Np and Nn is the nu-
mber of clustering center.
3. Sort the clustering center C = {P c, N c} with ascendent order, and find the mid-
dle value rk = (Ck + Ck+1 )/2 as the rough decision plane.
4. Generate the histogram with the intervals σk = (rk , rk+1 ], and produce the class
label ωc for each interval via Bayesian decision rule.
5. Iteratively merge adjacent intervals with same decision label ωc to produce a set
of consistent intervals σ̂k .
Output:
A LUT function lup(k) on the merged intervals σ̂k , k = 1, . . . , K.
Therefore, the final classifier often classifies these bags as positive. The variations
problem in training examples will be reduced.
4 Graph Embedding Based Decision Stump

4.1 Supervised Graph Embedding
Let hi ∈ RD (i = 1, 2, . . . , n) be the D-dimensional histogram feature and yi ∈
{ωc }2c=1 be the associated class label. The feature is written as matrix form:
H = (h1 |h2 | . . . |hn ). Let G = {{hi }ni=1 , S} be an undirected weighted graph
with vertex set{hi }ni=1 and the similarity matrix S ∈ Rn×n . The element si,j of
matrix S measures the similarity of vertex pair i and j. The unsupervised graph
embedding is defined as the optimal low dimension vector representations for
the vertices of graph G

P∗ = min ||P hi − P hj ||2 si,j = min 2tr(P T HLH T P ), (5)
P T HMH T P =I P T HMH T P =I
i,j
where projection P ∈ Rd×D , (d < D) maps feature h from high dimension

space R
D
to low dimension space Rd . The elements in the diagonal matrix M is
mi,j = si,j , and the Laplacian matrix L is M − S.
i=j
The similarity si,j connects the relationship between high dimension and low
dimension space. If two vertexes hi and hj are close, si,j will be large, and vice
versa. To classification, the projection P should keep the between-class far and
within-class close. The similarity matrix S should reflect the separable ability. The
between-class similarity sbi,j and within-class similarity sw
i,j can be defined as
1
1
We refer the interested reader to [10] for more details.
positive samples
(a) clustering center
negative samples
(b)
clustering center
(c)
(d)
(e)
Fig. 3. A demonstration of generating the multi- Fig. 4. 4 type block feature

channel decision stump. (a)-(b) Cluster on positive and
negative examples, respectively. (d)Generate the deci-
sion stumps via histogram. (e)Merge the consistent de-
cision stumps.
5 5
1/n − 1/nc if yi = yj = ωc , 1/nc if yi = yj = ωc ,
sbi,j = sw = , (6)
1/n if yi = yj , i,j 0 if yi = yj ,
where nc is the cardinality of the ωc class. The pairwise sbi,j and sw i,j try to keep
within-class sample close (since sw
i,j is positive and s b
i,j is negative if yi = yj ) and
between-class sample pairs apart (since si,j is positive if yi = yj ). The projection
b
matrix P can be calculated by Fisher criterion
P∗ = max tr(P T H(Mb − Sb )H T P ). (7)

P T H(Mw −Sw )H T P =I
The projection matrix P = [p1 , p2 , . . . , pl ] are solved by generalized eigenvectors

corresponding to the l largest eigenvalues p in H(Mw − Sw )H T pl = λH(Mb −
Sb )H T pl .
4.2 Multi-channel Decision Stump
According to Bayesian decision theory, if class conditional probability p(ω1 |x) >
p(ω2 |x) we would naturally incline to decide that the true label of x is ω1 , and
vice versa. Using Bayes rule p(ω|x) = p(x|ω)p(ω), the optimal decision plane is
located at where p(x|ω1 ) = p(x|ω2 ) with p(ω1 ) = p(ω2 ). We obtain the Bayes
error p(error|x) = min[p(x|ω1 ), p(x|ω2 )]dx. However, the p(x|ωc ) is not directly
available. To accurately estimate the p(x|ωc ), histogram needs large numbers of
bins via uniform sampling in [25,26]. We avoid estimating the p(x|ωc ) with uniform
sampling or rejection sampling. As demonstrated in Fig. 3(c), we consider the local
region of feature space, and the location at the middle of two modal is a natural
decision plane. The decision plane would approximately minimize Bayes error, if
p(ω1 ) = p(ω2 ). Algorithm. 1 shows the graph embedding based decision stump.
Note that the number of decision planes is automatically decided.
5 Pedestrian Detection
To achieve the fast pedestrian detection, we adopt the cascade structure of de-
tector [6]. Each stage is designed to achieve high detection rate and modest false
positive rate. We combine K = 30 LMIBoost on HOG feature with rejection
cascade. To exploit the discriminative ability of HOG feature, we design 4 type
block feature as showed in Fig.4. In each cell, 9-bins HOG feature is extracted
and concatenated into a single histogram to represent the block feature. To ob-
tains a modicum of illumination invariance, the feature is normalized with L2
norm. The dimension of the 4 different type feature are 9, 18, 27 and 36, respec-
tively. The 453 × 4 number of block HOG feature can be computed from a single
detection window.
Assuming that the i-th cascade stage is trained, we classify all the possible
detection window on the negative training images with the cascade of the previ-
ous k-1 LMIBoost classifiers. The examples which are misclassified in scanning
window form the possible new negative training set. While, the positive training
samples do not change during bootstrap. Let Npi and Nni be cardinality of the
positive and negative training examples at i-th stage. Considering the influence
of asymmetric training data on the classifier and computer RAM limitations, we
constrain Npi and Nni to be approximately equal.
According to “There is no free lunch” theorem, it is very important to choose
suitable number of instances in a bag for training and detection. More instances
in a bag will represent more variations and improve the detection results, but
will also reduce the training and detection speed. We experimentally set 4 in-
stances for training and detection, respectively. Each level of cascade classifier
is optimized to correctly detect at least 99% of the positive bags, while reject at
least 40% of the negative bags.
6 Experiments
To test our method, we perform the experiments on two public dataset: INRIA
[4] and VOC2006 [1]. The INRIA dataset contains 1239 pedestrian images (2478
with their left-right reflections) and 1218 person-free images for training. In
the test set, there are 566 images containing pedestrians. The pedestrian images
provided by INRIA dataset have large variations (but most of them have standing
pose), different clothing and urban background. This dataset is very close to real-
life setting. The VOC2006’s person detection subtask supplies 319 images with
577 person as training set, and 347 images with 579 person as validation set.
675 images with 1153 person is supplied as test data. Note that the VOC2006’s
person detection dataset contains various human activities, different stances and
clothing. Some examples of the two different datasets are showed in Fig. 8.
6.1 Performance Comparisons on Multiple Datasets

We plot the detection error tradeoff curves on a log-log scale for INRIA dataset.
The y-axis corresponds to the miss rate, F alseN eg/(F alseN eg+T rueP os) and the
0.4
Dalal&Triggs−Ker. SVM + HOG [4]
Zhu et al−AdaBoost + HOG [13]
Dalal&Triggs−Linear SVM + HOG [4]
0.2 Tuzel et al−Boost + Covariance descriptor [14]
Our approach LMiBoost + HOG
miss rate
0.1
0.05
0.04
0.03
0.02
0.015
−5 −4 −3 −2
10 10 10 10
false postive per window (FPPW)
Fig. 5. Comparison results on INRIA dataset. Note that the curve of our detector is
generated by changing the number of cascade stage used.
x-axis corresponds to false positives per window(FPPW), F alseP os/(T rueN eg +

F alseP os). We compare ours results with [4,13,14] on INRIA dataset. Although
it has been noted that kernel SVM is computationally expensive, we consider
both the kernel and linear SVM method of [4]. Only the best performing result,
the L2-norm in HOG feature, is considered. Covariance descriptor [14] is also
compared. Fig. 5 shows that the performance of our method is comparable to
the state-of-art approaches. We achieve 4.3% miss rate at 10−4 FPPW. Notice
that all the results by other methods are quoted directly from the original papers,
since we perform the same separation of training-testing sets.
The Fig.7 shows the precision-recall cure on VOC2006 person detection sub-
task for comp3 [1]. The protocol of the comp3 is that the training data is com-
posed of the training set and validation set. The non-normalized examples are
first approximately aligned, and then be converted into a bag of instances. Some
truncated and difficult examples in training data are discarded. The standard
scanning window technique is adopted for detection, although the scanning win-
dow may be not suitable for VOC2006 detection subtask. The average preci-
sion scores is 0.23, which is better than the best results 0.164 reported by
INRIA Douze [1]. In Fig. 8, several detection results are showed for different
scenes with human having variable appearance and pose. Significantly overlap-
ping detection windows are averaged into a single window.
6.2 Analysis of the Weak Classifiers

For our next experiment, we conduct experiments to compare the performance
of different weak classifiers. A common set of parameters (such as, false positive
rate for every stage)are controlled equally for cascade training. Two detectors
are trained with different weak classifiers, including FLDA and graph embedding
based decision stump.
1
0.3
FLDA
0.9
0.2 Graph embeding based decision stump
0.8
0.7
miss rate
0.1
0.6
precision
0.5
0.05 0.4
0.04
0.3
0.03
0.2
0.02 0.1
0.015 0
−5 −4 −3
10 10 10 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
false postive per window (FPPW) recall
Fig. 6. Comparison on different weak Fig. 7. Performance on VOC2006 dataset

classifier
Fig. 8. Some detection samples on INRIA and VOC2006 datasets
The performance results on INRIA show that the detectors based on graph
embedding decision stump outperforms the detector based on FLDA in Fig. 6.
Unlike the other LUT weak classifier [26,25], the bins of decision stumps are
automatically decided by algorithm.
6.3 Analysis of the Detection Speed
There are 90% of the negative examples are rejected at first five stage. The speed
of the cascaded detector is directly related to the number of feature evaluated
per scanned sub-window. For INRIA dataset, on average our method requires to
evaluate 10.05 HOG feature per negative detection window. Densely scanning at
0.8 scale and 4 pixel step in a 320 × 240 image needs average 150ms under PC
with 2.8GHz CPU and 512RAM. While, 250ms for 320 × 240 image is reported
in Zhu et al’s detector [13].
We introduce the multiple instance learning into the pedestrian detection for
solving pose variations. The training example does not need to be well aligned,
but to be represented as a bag of instances. To efficiently utilizing histogram fea-
ture, a graph embedding based decision stump is proposed. The weak classifier
guarantees the fast detection and better discriminative ability. The promising
performances of the approach are shown on INRIA and VOC2006’s person de-
tection subtask.
Using multiple instance learning has enabled detector robust to the pose and
appearance variations. Theoretically, the more instances are supplied, the more
variations would be learned. Modeling the average relationship between the in-
stance’s label and bag’s label may be unsuitable when there are large numbers
of instances in a positive bag. In future, more experiments will be carried out to
compare the different way to model the relationship.
Acknowledgements
This work was supported in part by National Natural Science Foundation of

China under Grant 60773136 and 60702035, in part by National Hi-Tech De-
velopment Program (863 Program) of China under Grant 2006AA01Z117 and
2006AA010105. We would also thank the anonymous reviewers for their valuable
comments.
References
1. Everingham, M., Zisserman, A., Williams, C.K.I., Gool, L.V.: The PASCAL Visual
Object Classes Challenge (VOC 2006) Results (2006),
http://www.pascal-network.org/challenges/VOC/voc2006/results.pdf
2. Gavrila, D.M.: Pedestrian detection from a moving vehicle. In: Vernon, D. (ed.)
3. Bissacco, A., Yang, M., Soatto, S.: Detection human via their pose. In: Proc. NIPS
(2006)
Proc. CVPR, vol. I, pp. 886–893. IEEE, Los Alamitos (2005)
5. Papageorgiou, P., Poggio, T.: A trainable system for object detection. IJCV, 15–33
(2000)
6. Viola, P., Jones, M., Snow, D.: Detecing pedestrians using patterns of motion and
appearance. In: Proc. ICCV (2003)
7. Maron, O., Lozanno-Perez, T.: A framework for multiple-instance learning. In:
Proc. NIPS, pp. 570–576 (1998)
8. Dietterich, T., Lathrop, R., Lozano-Perez, T.: Solving the multiple instance prob-
lem with axis-parallel rectangles. Artifical intelligence, 31–71 (1997)
9. Xu, X., Frank, E.: Logistic regression and boosting for labeled bags of instances.
In: Dai, H., Srikant, R., Zhang, C. (eds.) PAKDD 2004. LNCS (LNAI), vol. 3056,
10. Sugiyama, M.: local fisher discriminat analysis for supervised dimensionality re-
duction. In: Proc. ICML (2006)
11. Monhan, A., Papageorgiou, C., Poggio, T.: Example-based object detection in im-
ages by componets. IEEE Trans. PAMI 23, 349–360 (2001)
12. Lowe, D.G.: Distinctive image features from scale-invariant keypoints, 91–110
(2004)
13. Zhu, Q., Avidan, S., Yeh, M.C., Cheng, K.T.: Fast human detection using a cascade
of histograms of oriented gradients. In: Proc. CVPR, vol. 2, pp. 1491–1498. IEEE,
Los Alamitos (2006)
14. Tuzel, O., Porikli, F., Meer, P.: Human detection via classification on riemannina
manifolds. In: Proc. CVPR. IEEE, Los Alamitos (2007)
15. Zisserman, A., Schmid, C., Mikolajczyk, K.: Human detection based on a proba-
bilistic assembly of robust part detectors. In: Pajdla, T., Matas, J(G.) (eds.) ECCV
16. Leibe, B., Seemann, E., Schiele, B.: Pedestrian detection in crowed scenes. In: Proc.
CVPR, pp. 878–885. IEEE, Los Alamitos (2005)
17. Tran, D., Forsyth, D.A.: Configuration estimates improve pedestrian finding. In:
Proc. NIPS (2007)
18. Felzenszwalb, P., Mcallester, D., Ramanan, D.: A discriminatively trained, multi-
scale, deformable part model. In: Proc. CVPR. IEEE, Los Alamitos (2008)
19. Maron, O., Ratan, A.: Multiple-instance learning for natural scene classification.
In: Proc. ICML (1998)
20. Viola, P., Platt, J.C., Zhang, C.: Multiple instance boosting for object detection.
In: Proc. NIPS (2006)
21. Porikli, F.M.: Integral histogram: a fast way to extract histogram in cartesian
space. In: Proc. CVPR, pp. 829–836. IEEE, Los Alamitos (2005)
22. Liu, C., Shum, H.Y.: Kullback-leibler boosting. In: Proc. CVPR, pp. 587–594
(2003)
23. Laptev, I.: Improvements of object detection using boosted histograms. In: BMVC
(2006)
features. In: Proc. CVPR. IEEE, Los Alamitos (2001)
25. Huang, C., Ai, H., Wu, B., Lao, S.: Boosting nested cascade detector for multi-view
face detection. In: Proc. ICPR. IEEE, Los Alamitos (2004)
26. Xiao, R., Zhu, H., Sun, H., Tang, X.: Dynamic cascades for face detection. In: Proc.
ICCV. IEEE, Los Alamitos (2007)
Object Detection from Large-Scale 3D Datasets
Using Bottom-Up and Top-Down Descriptors
Alexander Patterson IV, Philippos Mordohai, and Kostas Daniilidis
University of Pennsylvania
{aiv,mordohai,kostas}@seas.upenn.edu
Abstract. We propose an approach for detecting objects in large-scale

range datasets that combines bottom-up and top-down processes. In the
bottom-up stage, fast-to-compute local descriptors are used to detect
potential target objects. The object hypotheses are verified after align-
ment in a top-down stage using global descriptors that capture larger
scale structure information. We have found that the combination of spin
images and Extended Gaussian Images, as local and global descriptors
respectively, provides a good trade-off between efficiency and accuracy.
We present results on real outdoors scenes containing millions of scanned
points and hundreds of targets. Our results compare favorably to the
state of the art by being applicable to much larger scenes captured un-
der less controlled conditions, by being able to detect object classes and
not specific instances, and by being able to align the query with the best
matching model accurately, thus obtaining precise segmentation.
1 Introduction
Object detection and recognition in images or videos is typically done based on

color and texture properties. This paradigm is very effective for objects with
characteristic appearance, such as a stop sign or the wheel of a car. There are,
however, classes of objects for which 3D shape and not appearance is the most
salient feature. Cars are an object category, whose appearance varies a lot within
the class, as well as with viewpoint and illumination changes. Instead of repre-
senting these objects with a collection of appearance models, specific to each
viewpoint, several researchers have used range scanners and addressed object
recognition in 3D. Range as an input modality offers the advantages of using
the full dimensionality of an object and avoiding any scale ambiguity due to
projection. In addition, figure-ground segmentation is easier in 3D than in 2D
images since separation in depth provides powerful additional cues. On the other
hand, range sensors have significantly lower resolution compared to modern cam-
eras and alignment between the query and the database models still has to be
estimated. The challenges associated with object detection in 3D are due to
intra-class shape variations, different sampling patterns due to different sensors
or different distance and angle between the sensor and the object, targets that
are almost always partial due to self-occlusion and occlusion.

554 A. Patterson IV, P. Mordohai, and K. Daniilidis
Fig. 1. Cars detection results from real LIDAR data. Cars have false colors.
In this paper, we present an approach for detecting and recognizing objects

characterized by 3D shape from large-scale datasets. The input is a point cloud
acquired by range sensors mounted on moving vehicles. A part of the input is
used as training data to provide manually labeled exemplars of the objects of
interest, as well as negative exemplars where objects of interest are not present.
Our algorithm automatically detects potential locations for the target objects in
a bottom-up fashion. These locations are then processed by the top-down module
that verifies the hypothesized objects by aligning them with models from the
training dataset. We show results on a very large-scale dataset which consists of
hundreds of millions of points. To the best of our knowledge no results have been
published for datasets of this size. State-of-the-art 3D recognition systems on real
data [1,2,3,4,5] have shown very high recognition rates, but on high-resolution
scenes containing a few small objects captured in controlled environments.
We believe that our research is a first step towards fully automatic annotation
of large scenes. Recent advances in sensor technology have made the acquisition
and geo-registration of the data possible. Detailed 3D models can be generated
very efficiently to provide high-quality visualization [6], but their usefulness for
everyday applications is limited due to the absence of semantic annotation. Much
like image-based visualizations, such as the Google Street View, these represen-
tations cannot answer practical questions, such as “where is the nearest gas sta-
tion, mailbox or phonebooth”. Automatic methods for scene annotation would
dramatically increase the benefits users can derive from these large collections
of data. While our methods are not currently capable of addressing the problem
in its full extend, this paper introduces a framework for object detection from
range data that makes a step towards automatic scene annotation. Some results
on car detection can be seen in Fig. 1.
The main technical contribution of our work is the combination of a bottom-
up and a top-down process to efficiently detect and verify the objects of interest.
We use spin images [7] as local descriptors to differentiate between the target
Object Detection from Large-Scale 3D Datasets 555
objects and clutter and Extended Gaussian Images (EGIs) [8] to ascertain the
presence of a target at the hypothesized locations. This scheme enables us to
process very large datasets with high precision and recall. Training requires
little effort, since the user has to click one point in each target object, which
is then automatically segmented from the scene. The remaining points are used
as negative examples. Spin images are computed on both positive and negative
examples. EGIs only need to be computed for the positive exemplars of the
training set, since they are used to align the bottom-up detection with the model
database. Accurate alignment estimates between similar but not identical objects
enable us to segment the target objects from the clutter. The contributions of
our work can be summarized as follows:
– The combination of bottom-up and top-down processing to detect potential
targets efficiently and verify them accurately.
– The capability to perform training on instances that come from the same
object category as the queries, but are not necessarily identical to the queries.
– Minimal user efforts during training.
– Object detection for large-scale datasets captured in uncontrolled environ-
ments.
– Accurate segmentation of target objects from the background.
2 Related Work
In this section, we briefly overview related work on local and global 3D shape
descriptors and 3D object recognition focusing only on shape-based descriptors.
Research on appearance-based recognition has arguably been more active re-
cently, but is not directly applicable in our experimental setup.
Global shape descriptors include EGIs [8], superquadrics [9], complex EGIs
[10], spherical attribute images [11] and the COSMOS [12]. Global descriptors
are more discriminative since they encapsulate all available information. On the
other hand, they are applicable to single segmented objects and they are sensitive
to clutter and occlusion. A global representation in which occlusion is explicitly
handled is the spherical attribute image proposed by Hebert et al. [11].
A method to obtain invariance to rigid transformations was presented by Os-
ada et al. [13] who compute shape signatures for 3D objects in the form of shape
statistics, such as the distance between randomly sampled pairs of points. Liu
et al. [14] introduced the directional histogram model as a shape descriptor and
achieved orientation invariance by computing the spherical harmonic transform.
Kazhdan et al. [15] proposed a method to make several types of shape descrip-
tors rotationally invariant, via the use of spherical harmonics. Makadia et al. [16]
compute the rotational Fourier transform [17] to efficiently compute the corre-
lation between EGIs. They also propose the constellation EGI, which we use in
Section 5 to compute rotation hypotheses.
Descriptors with local support are more effective than global descriptors for
partial data and data corrupted by clutter. Stein and Medioni [18] combined sur-
face and contour descriptors, in the form of surface splashes and super-segments,
respectively. Spin images were introduced by Johnson and Hebert [7] and are
among the most popular such descriptors (See Section 4). Ashbrook et al. [19]
took a similar approach based on the pairwise relationships between triangles
of the input mesh. Frome et al. [20] extended the concept of shape contexts
to 3D. Their experiments show that 3D shape contexts are more robust to oc-
clusion and surface deformation than spin images but incur significantly higher
computational cost. Huber et al. [21] propose a technique to divide range scans
of vehicles into parts and perform recognition under large occlusions using spin
images as local shape signatures.
Local shape descriptors have been used for larger scale object recognition.
Johnson et al.[1] use PCA-compressed spin images and nearest neighbor search
to find the most similar spin images to the query. Alignment hypotheses are
estimated using these correspondences and a variant of the ICP algorithm [22] is
used for verification. Shan et al. [23] proposed the shapeme histogram projection
algorithm which can match partial object by projecting the descriptor of the
query onto the subspace of the model database. Matei et al. [3] find potential
matches for spin images using locality sensitive hashing. Geometric constraints
are then used to verify the match. Ruiz-Correa et all. [5] addressed deformable
shape recognition via a two-stage approach that computes numeric signatures
(spin images) to label components of the data and then computes symbolic
signatures on the labels. This scheme is very effective, but requires extensive
manual labeling of the training data. Funkhouser and Shilane [24] presented a
shape matching system that uses multi-scale, local descriptors and a priority
queue that generates the most likely hypotheses first.
In most of the above methods, processing is mostly bottom-up, followed in
some cases by a geometric verification step. A top-down approach was proposed
by Mian et al. [4] who represent objects by 3D occupancy grids which can be
matched using a 4D hash table. The algorithm removes recognized objects from
the scene and attempts to recognize the remaining data until no additional
library object can be found.
Our method can detect cars in real scenes in the presence of clutter and
sensor noise. Very few of the papers mentioned above ([1,2,3,4,5]) present results
on real data. Among the ones that do, Matei et al. [3] classified cars that had
been previously segmented. Johnson et al. [1], Carmichael et al. [2] and Mian
et al. [4] show object detection from real scenes containing multiple objects. It
should be noted, however, that the number of objects in the scene is small and
that all objects were presented to the algorithm during training. Ruiz-Correa
et. all [5] are able to handle intra-class variation, at the cost of large manual
labeling effort. The goal of our work is more ambitious than [1,2,3,4,20,23] in
order to make more practical applications possible. Our algorithm is not trained
on exemplars identical to the queries, but on other instances from the same
class. This enables us to deploy the system on very large-scale datasets with
moderate training efforts, since we only have to label a few instances from the
object categories we are interested in.
3 Algorithm Overview
Our algorithm operates on 3D point clouds and entails a bottom-up and a top-
down module. The steps for annotation and training are the following:
1. The user selects one point on each target object.
2. The selected target objects are automatically extracted from the background.
3. Compute surface normals for all points1 in both objects and background.
4. Compute spin images on a subset of the points for both objects and back-
ground and insert into spin image database DBSI (Section 4).
5. Compute an EGI for each object (not for the background). Compute con-
stellation EGI and density approximation. Insert into EGI database DBEGI
(Section 5).
Processing on test data is performed as follows:
1. Compute normals for all points and spin images on a subset of the points.
2. Classify spin images as positive (object) or negative (background) according
to their nearest neighbors in DBSI .
3. Extract connected components of neighboring positive spin images. Each
connected component is a query (object hypothesis).
4. Compute an EGI and the corresponding constellation EGI for each query.
5. For each query and model in DBEGI (Section 5):
(a) Compute rotation hypothesis using constellation EGIs.
(b) For each rotation hypothesis with low distance according to Section (5.3),
compute translation in frequency domain.
(c) Calculate the overlap between query and model.
6. If the overlap is above the threshold, declare positive detection (Section 5).
7. Label all points that overlap with each of the models of DBEGI after align-
ment as object points to obtain segmentation.
4 Bottom-Up Detection
The goal of the bottom-up module is to detect potential target locations in the
point cloud with a bias towards high recall to minimize missed detections. Since
detection has to be performed on very large point clouds, we need a represen-
tation that can be computed and compared efficiently. To this end we use spin
images [7]. A spin image is computed in a cylindrical coordinate system defined
by a reference point and its corresponding normal. All points within this region
are transformed by computing α, the distance from the reference normal ray and
β the height above the reference normal plane. Finally a 2D histogram of α and
β is computed and used as the descriptor. Due to integration around the normal
of the reference point, spin images are invariant to rotations about the normal.
This is not the case with 3D shape contexts [20] or EGIs (Section 5) for which
several rotation hypotheses have to be evaluated to determine a match. Since
1
During normal computation, we also estimate the reliability of the normals, which
is used to select reference points for the spin images.
10
12
14
2 4 6 8 10 12 14
Fig. 2. Left: spin image computation on real data. The blue circles delineate the cylin-
drical support region and the red vector is the normal at the reference point. Middle:
illustration of spin image computation. O is the reference point and n its normal. A
spin image is a histogram of points that fall into radial (α) and elevation (β) bins.
Right: the spin image computed for the point on the car.
spin image comparison is a simple distance between vectors, their comparisons

are computationally cheaper, but less discriminative.
Johnson and Hebert [7] computed spin images on meshes. This can compen-
sate for undesired effects due to varying sample density, since triangles contribute
to each bin of the histogram with their area. Triangulating the point cloud to
obtain a mesh is not trivial in our case, not only because of the computational
cost, but also due to noise, sparsity and sampling patterns of the data. Similar
to [20], we compute spin images directly from the point clouds and weigh the
contribution of each point by its inverse density to account for sampling differ-
ences. Local density is computed in balls centered at every point. Isolated points
are removed. Accounting for variations in point density is important for point
clouds captured by range sensors since the density of samples on a surface is a
function of sensor type, as well as distance and angle to the sensor.
Given a point cloud, regardless of whether it contains training or test data,
normals for all points are computed using tensor voting [25]. We then need to
select reference points for the spin images. Our experiments have shown that
spin images vary smoothly as long as the reference point is on the same surface
and the normal is accurate. Therefore, reference points need to be dense enough
to capture all surfaces of the object, but higher density is redundant. For cars,
a distance of 0.4m was found to offer a good trade-off between coverage and
computational efficiency. We obtain such a sampling by placing a 3D grid of the
desired resolution in the dataset and dropping vertices that have no scanned
points in their voxel. Since the reference points need to be among the points
sampled by the scanner, the retained vertices are moved to the median of the
nearest points to account for noisy samples. (The grid can be seen in the two
rightmost images of Fig. 3.) A spin image is computed for each of these points
unless the eigenvalues of the tensor after tensor voting indicate that the estimated
normal is unreliable [25]. Our spin images have 15 radial and 15 elevation bins
resulting in a 225-D descriptor.
For the training data, the user has to specify the targets, which are assumed to
be compact objects lying on the ground, by clicking one point on each. Then, an
Fig. 3. Left: input point cloud. Middle: Classification of spin images as target (blue)
and background (cyan). (Only the reference points are shown.) Right: target spin image
centers clustered into object hypotheses. Isolated target spin images are rejected.
automatic algorithm segments the object as a connected component protruding

from the ground. The ground can be reliably estimated in a small neighborhood
around the selected point as the lowest smooth surface that bounds the data.
Spin images computed for points on the targets are inserted into the spin image
database DBSI as positive exemplars, while spin images from the background
are inserted as negative exemplars. We have implemented the database using
the Approximate Nearest Neighbor (ANN) k-d tree [26].
During testing, query spin images are computed on reference points on a grid
placed on the test data as above. Each query spin image is classified according
to the nearest neighbor retrieved from DBSI . Some results on real data can be
seen in Fig. 3. Potential locations of the target objects can be hypothesized in
areas of high density of positive detections, while isolated false positives can be
easily pruned from the set of detections.
Object hypotheses (queries) are triggered by spin images that have been classi-
fied as positive (target). Target spin images are grouped into clusters by a simple
region growing algorithm that starts from a spin image reference point and con-
nects it to all neighboring target spin images within a small radius. When the
current cluster cannot be extended any further, the algorithm initializes a new
cluster. Raw points that are within a small distance from a cluster of spin images
are also added to it to form a query. Since neighboring spin images overlap, the
bottom-up portion of our algorithm is robust to some miss-classifications.
5 Top-Down Alignment and Verification
The second stage of processing operates on the queries (clustered points with
normals) proposed by the bottom-up stage and verifies whether targets exist at
those locations. Spin images without geometric constraints are not discriminative
enough to determine the presence of a target with high confidence. Spin image
classification is very efficient, but only provides local evidence for the presence
of a potential part of a target and not for a configuration of parts consistent
with a target. For instance a row of newspaper boxes can give rise to a number
of spin images that are also found in cars, but cannot support a configuration
of those spin images that is consistent with a car. The top-down stage enforces
these global configuration constraints by computing an alignment between the
query and the database models using EGI descriptors.
Early research has shown that there is a unique EGI representation for any
convex object [27], which can be obtained by computing the density function
of all surface normals on the unit sphere. If the object is not convex, its shape
cannot be completely recovered from the EGI, but the latter is still a powerful
shape descriptor. The EGI does not require a reference point since the relative
positions of the points are not captured in the representation. This property
makes EGIs effective descriptors for our data in which a reference point cannot
be selected with guaranteed repeatability due to occlusion, but the distribution
of normals is fairly stable for a class of objects.
5.1 Computing EGIs

EGIs are computed for the positive object examples in the training set. Objects
are segmented with assistance from the user, as described in Section 4. For the
test data, an EGI is computed for each object hypothesis extracted according
to the last paragraph of Section 4. Each EGI contains the normals of all input
points of the cluster, oriented so that they point outwards, towards the scanner.
These orientations can be computed since the trajectory of the sensor is available
to us. The majority of objects are scanned only from one side and, as a result,
the normals typically occupy at most a hemisphere of the EGI. This viewpoint
dependence occurs for both the queries and database objects and thus requires no
special treatment. If necessary, database models can be mirrored to increase the
size of the database without additional manual labeling since model symmetry
is modeled by the EGI.
5.2 Constellation EGIs

Unlike spin images, comparing two EGIs requires estimating a rotation that
aligns them before a distance can be computed. One can compute the rotational
Fourier transform [17] to efficiently compute all correlations between EGIs [16].
This technique is efficient if all rotations need to be computed, but it is sensitive
to clutter, missing parts and quantization. Our experiments have shown that
quantization can have adverse effects on rotation and distance computations
Fig. 4. Left: a database model of a car. Middle: illustration of an EGI in which points
are color-coded according to their density. Right: the corresponding constellation EGI.
using EGIs. We can use the constellation EGI to cue a more efficient distance
computation. Therefore, we avoid quantizing the orientations of the normals in
an EGI and do not treat it as an orientation histogram.
Instead of an exhaustive search using either spatial or Fourier methods, we use
a technique that generates discrete alignment hypotheses, which was originally
proposed in [16]. A constellation EGI records the locations of local maxima in the
distribution of normals in the EGI. We call these maxima stars, since they resem-
ble stars in the sky. An EGI and the corresponding constellation EGI for an object
can be seen in Fig. 4. Two constellation EGIs can be matched by sampling pairs of
stars that subtend the same angle on the sphere. Each sample generates matching
hypotheses with two stars of the other EGI. If the angles between each pair are
large enough and similar, a rotation hypothesis for the entire descriptor is gener-
ated. Note that a correspondence between two pairs of stars produces two possible
rotations. Similar rotations can be clustered to reduce the number of hypotheses
that need to be tested. The resulting set of rotations are evaluated based on the
distance between the entire EGIs and not just the stars.
5.3 Hypothesis Verification

Conceptually, the rotation hypothesis that achieves the best alignment of the
descriptors is the one that maximizes the cross-correlation between all normal
vectors of the first and the second EGI. This computation is exact, but computa-
tionally expensive since models and queries consist of thousands of points each.
To reduce the computational complexity, we select a smaller set of normals and
compute the weights of kernels which are centered on this set, thus closely ap-
proximating the original EGI via interpolation. This computation is performed
once per EGI and significantly reduces the cost of distance computation. Specif-
ically, to create an approximation of the EGI for a set of input normals, we
compute the density at all input normals on the sphere. We then select a subset
of samples by greedily choosing a predetermined number of points. Each choice
is made by computing the current interpolation via nearest neighbor, and then
adding the normal with the largest deviation between approximated and actual
values. Our method is similar to [28], but operates on the sphere. Once we have
a set of kernel centers Ns which is a subset of all normals N , the weights of the
kernels are computed as follows:
max(dmax − arccos(n̂Ti n̂j ), 0)

vij = , i ∈ N, j ∈ Ns
j (max(dmax − arccos(n̂i n̂j ), 0))
T
Dj = V † Di , (1)
Where Dj are the coefficients at the sparse set of normals Ns , and dmax is the
range of the kernel function. Using this new representation, we can compute
the distance between two EGIs, using a sparse set of samples, after applying a
rotation hypothesis. If the two shapes are identical, the density values should
be equal over the entire sphere. We measure the deviation from an ideal match
by predicting the density on the samples of one EGI using the interpolation
Fig. 5. Alignment of a database model (left car and left EGI) and a query (right car
and right EGI) that have been aligned. The car models are shown separately for clarity
of the visualization. Notice the accuracy of the rotation estimation. The query has been
segmented by the positive spin image clustering algorithm and the model by removing
the ground after the user specified one point.
function of the other EGI and comparing them with the original density values.
Specifically we use the l1 distance computed at the query points which we can
now interpolate once the normals Ns are rotated according to each hypothesized
rotation. The minimum distance provides an estimate of the best rotation to
align the two objects, but no estimate of translation and most importantly no
indication of whether the objects actually match. Typically, 1-5 rotations are
close enough to the minimum distance. For these, we estimate the translation
and compute the final distance in the following section.
5.4 Alignment and Distance Computation
Given the few best rotation hypotheses based in section 5.3, we compute the
translation that best aligns the two models in the frequency domain. We adopt
the translation estimation method of [16] in which translation is estimated us-
ing a Fourier transform in R3 . This is less sensitive to noise in the form of
missing parts or clutter than global alignment methods that estimate com-
plete rigid transformations in the Fourier domain. We begin by voxelizing the
model and the query to obtain binary occupancy functions in 3D. We then com-
pute their convolution efficiently using the fft and take the maximum as our
translation.
Finally, we need a measure of distance to characterize the quality of the align-
ment that is flexible enough to allow for deformation between the query and the
model. We experimented with the ICP distance [22], without performing ICP
iterations, but found the overlap between the query and model to be more ef-
fective because the quantization in the translation estimation caused large ICP
distance errors even though the models were similar. The overlap is computed as
the inlier ratio over all points of the model and query, where an inlier is a point
with a neighboring point from the other model that is closer than a threshold
distance and whose normal is similar to that of the point under consideration.
Figure 5 shows an alignment between a query and a database object and their
corresponding EGIs. Selecting the overlap points after alignment results in pre-
cise segmentation of the object from the background.
0.9
0.8
0.7
0.6
0.5
0.4
0.4 0.6 0.8 1
Fig. 6. Left: The precision-recall curve for car detection on 200 million points con-
taining 1221 cars. (Precision is the x-axis and recall the y-axis.) Right: Screenshot of
detected cars. Cars are in random colors and the background in original colors.
Fig. 7. Screenshots of detected cars, including views from above. (There is false nega-
tive at the bottom of the left image.) Best viewed in color.
We processed very large-scale point clouds captured by a moving vehicle equipped
with four range scanners and precise Geo-location sensors. The dataset consists of
about 200 million points, 2.2 million of which were used for training. The training
set included 17 cars which were selected as target objects. We compute 81,172 spin
images for the training set (of which 2657 are parts of cars) and 6.1 million for the
test set. Each spin image has a 15×15 resolution computed in a cylindrical support
region with height and radius both set to 2m. Reference points for the spin images
are selected as in Section 4 with an average distance between vertices of 0.4m. The
spin images of the training set are inserted into DBSI .
EGIs are computed for each target object in the training set, approximated
by picking a smaller set of 200 normals, that minimize the interpolation error on
all samples. The approximated EGIs are inserted into DBEGI , which is a simple
list with 17 entries. Since our method only requires very few representatives from
each class, we were able to perform the experiments using a few sedans, SUVs
and vans as models.
The query trouping threshold is set to 1m (Section 4). This groups points roughly
up to two grid positions away. The EGI matching thresholds (Section 5.2) are set
as follows: Each pair must of stars must subtend an angle of at least 30◦ and the
two angles must not differ by more than 5◦. Rotations that meet these requirements
are evaluated according to Section 5.3. For the best rotation hypotheses, the metric
used to make the final decision is computed: the percentage of inliers on both models
after alignment. For a point to be an inlier there has to be at least one other point
from the other model that is within 30cm and whose normal deviates by at most
35◦ from the normal of the current point. We have found the inlier fraction to be
more useful than other distance metrics.
Results on an test area comprising 220 million points and 1221 cars are shown
in Figs. 6 and 7. After bottom-up classification there were approximately 2200
detections of which about 1100 were correct. The top-down step removes about
1200 false positives and 200 true positives. The precision-recall curve as the inlier
threshold varies for the full system is shown in Fig. 6. For the point marked with
a star, there are 905 true positives, 74 false positives and 316 false negatives
(missed detections) for a precision of 92.4% and a recall of 74.1%.
7 Conclusion
We have presented an approach for object detection from 3D point clouds that
is applicable to very large datasets and requires limited training efforts. Its effec-
tiveness is due to the combination of bottom-up and top-down mechanisms to
hypothesize and test locations of potential target objects. An application of our
method on car detection has achieved very satisfactory precision and recall on an
area far larger than the test area of any previously published method. Moreover,
besides a high detection rate, we are able to accurately segment the objects of
interest from the background. We are not aware of any other methodology that
obtains comparable segmentation accuracy without being trained on the same
instances that are being segmented.
A limitation of our approach we intend to address is that search is linear in
the number of objects in the EGI database. We are able to achieve satisfactory
results with a small database, but sublinear search is a necessary enhancement
to our algorithm.
Acknowledgments
This work is partially supported by DARPA under the Urban Reasoning and
Geospatial ExploitatioN Technology program and is performed under National
Geospatial-Intelligence Agency (NGA) Contract Number HM1582-07-C-0018.
The ideas expressed herein are those of the authors, and are not necessarily
endorsed by either DARPA or NGA. This material is approved for public release;
distribution is unlimited.
The authors are also grateful to Ioannis Pavlidis for his help in labeling the
ground truth data.
References
1. Johnson, A., Carmichael, O., Huber, D., Hebert, M.: Toward a general 3-d matching
engine: Multiple models, complex scenes, and efficient data filtering. In: Image
Understanding Workshop, pp. 1097–1108 (1998)
2. Carmichael, O., Huber, D., Hebert, M.: Large data sets and confusing scenes in
3-d surface matching and recognition. In: 3DIM, pp. 358–367 (1999)
3. Matei, B., Shan, Y., Sawhney, H.S., Tan, Y., Kumar, R., Huber, D., Hebert, M.:
Rapid object indexing using locality sensitive hashing and joint 3d-signature space
estimation. IEEE Trans. on Pattern Analysis and Machine Intelligence 28(7), 1111–
1126 (2006)
4. Mian, A., Bennamoun, M., Owens, R.: Three-dimensional model-based object
recognition and segmentation in cluttered scenes. IEEE Trans. Pattern Analysis
and Machine Intelligence 28(10), 1584–1601 (2006)
5. Correa, S.R., Shapiro, L.G., Meila, M., Berson, G., Cunningham, M.L., Sze, R.W.:
Symbolic signatures for deformable shapes. IEEE Trans. on Pattern Analysis and
Machine Intelligence 28(1), 75–90 (2006)
6. Frueh, C., Jain, S., Zakhor, A.: Data processing algorithms for generating textured
3d building facade meshes from laser scans and camera images. IJCV 61(2), 159–
184 (2005)
7. Johnson, A.E., Hebert, M.: Using spin images for efficient object recognition in clut-
tered 3d scenes. IEEE Trans. on Pattern Analysis and Machine Intelligence 21(5),
433–449 (1999)
8. Horn, B.: Extended gaussian images. Proceedings of the IEEE 72(12), 1656–1678
(1984)
9. Solina, F., Bajcsy, R.: Recovery of parametric models from range images: The
case for superquadrics with global deformations. IEEE Transactions on Pattern
Analysis and Machine Intelligence 12(2), 131–147 (1990)
10. Kang, S., Ikeuchi, K.: The complex egi: A new representation for 3-d pose deter-
mination. IEEE Transactions on Pattern Analysis and Machine Intelligence 15(7),
707–721 (1993)
11. Hebert, M., Ikeuchi, K., Delingette, H.: A spherical representation for recognition
of free-form surfaces. IEEE Transactions on Pattern Analysis and Machine Intelli-
gence 17(7), 681–690 (1995)
12. Dorai, C., Jain, A.K.: Cosmos: A representation scheme for 3d free-form objects.
IEEE Trans. on Pattern Analysis and Machine Intelligence 19(10), 1115–1130
(1997)
13. Osada, R., Funkhouser, T., Chazelle, B., Dobkin, D.: Shape distributions. ACM
Transactions on Graphics 21(4) (2002)
14. Liu, X., Sun, R., Kang, S.B., Shum, H.Y.: Directional histogram model for three-
dimensional shape similarity. In: Int. Conf. on Computer Vision and Pattern Recog-
nition (2003)
15. Kazhdan, M., Funkhouser, T., Rusinkiewicz, S.: Rotation invariant spherical har-
monic representation of 3D shape descriptors. In: Symposium on Geometry Pro-
cessing (2003)
16. Makadia, A., Patterson, A.I., Daniilidis, K.: Fully automatic registration of 3d
point clouds. In: Int. Conf. on Computer Vision and Pattern Recognition, vol. I,
pp. 1297–1304 (2006)
17. Driscoll, J., Healy, D.: Computing fourier transforms and convolutions on the 2-
sphere. Advances in Applied Mathematics 15, 202–250 (1994)
18. Stein, F., Medioni, G.: Structural hashing: Efficient three dimensional object recog-
nition. IEEE Trans. on Pattern Analysis and Machine Intelligence 14(2), 125–145
(1992)
19. Ashbrook, A., Fisher, R., Robertson, C., Werghi, N.: Finding surface correspon-
dence for object recognition and registration using pairwise geometric histograms.
In: Burkhardt, H., Neumann, B. (eds.) ECCV 1998. LNCS, vol. 1407, pp. 674–686.
20. Frome, A., Huber, D., Kolluri, R., Bulow, T., Malik, J.: Recognizing objects in
range data using regional point descriptors. In: Pajdla, T., Matas, J(G.) (eds.)
21. Huber, D., Kapuria, A., Donamukkala, R., Hebert, M.: Parts-based 3d object clas-
sification. In: Int. Conf on Computer Vision and Pattern Recognition, vol. II, pp.
82–89 (2004)
22. Besl, P.J., McKay, N.D.: A method for registration of 3-d shapes. IEEE Trans. on
23. Shan, Y., Sawhney, H.S., Matei, B., Kumar, R.: Shapeme histogram projection
and matching for partial object recognition. IEEE Trans. on Pattern Analysis and
Machine Intelligence 28(4), 568–577 (2006)
24. Funkhouser, T., Shilane, P.: Partial matching of 3d shapes with priority-driven
search. In: Symposium on Geometry Processing (2006)
25. Medioni, G., Lee, M., Tang, C.: A Computational Framework for Segmentation
and Grouping. Elsevier, New York (2000)
26. Arya, S., Mount, D.M., Netanyahu, N.S., Silverman, R., Wu, A.Y.: An optimal
algorithm for approximate nearest neighbor searching. Journ. of the ACM 45, 891–
923 (1998)
27. Smith, D.A.: Using enhanced spherical images. Technical Report AIM-530. MIT
(1979)
28. Carr, J.C., Beatson, R.K., Cherrie, J.B., Mitchell, T.J., Fright, W.R., McCallum,
B.C., Evans, T.R.: Reconstruction and representation of 3d objects with radial
basis functions. In: SIGGRAPH, pp. 67–76. ACM, New York (2001)
Making Background Subtraction Robust to
Sudden Illumination Changes
Julien Pilet, Christoph Strecha, and Pascal Fua
École Polytechnique Fédérale de Lausanne, Switzerland

{julien.pilet,christoph.strecha,pascal.fua}@epfl.ch
http://cvlab.epfl.ch/
Abstract. Modern background subtraction techniques can handle grad-

ual illumination changes but can easily be confused by rapid ones. We
propose a technique that overcomes this limitation by relying on a sta-
tistical model, not of the pixel intensities, but of the illumination effects.
Because they tend to affect whole areas of the image as opposed to in-
dividual pixels, low-dimensional models are appropriate for this purpose
and make our method extremely robust to illumination changes, whether
slow or fast.
We will demonstrate its performance by comparing it to two repre-
sentative implementations of state-of-the-art methods, and by showing
its effectiveness for occlusion handling in a real-time Augmented Reality
context.
1 Introduction
Background subtraction is a critical component of many applications, ranging
from video surveillance to augmented reality. State-of-the-art algorithms can
handle progressive illumination changes but, as shown in Fig. 1, remain vulner-
able to sudden changes. Shadows cast by moving objects can easily be misinter-
preted as additional objects.
This is especially true of approaches [2,3,4,1] that rely on statistical back-
ground models that are progressively updated as time goes by. They can handle
both illumination effects and moving background elements, such as tree leaves or
flowing water. This is an obvious strength, but can result in mistakenly integrat-
ing foreground elements into the background model. This is a potentially serious
problem in surveillance applications: A forgotten luggage could accidentally be-
come part of the background. Furthermore, the model update is usually relatively
slow, making it difficult to rapidly adjust to sudden illumination changes and to
shadows cast by moving objects.
Here, we propose an approach that overcomes this problem by replacing the
statistical background model by a statistical illumination model. More specifi-
cally, we model the ratio of intensities between a stored background image and
an input image in all three channels as a Gaussian Mixture Model (GMM) that
accounts for the fact that different parts of the scene can be affected in differ-
ent ways. We incorporate this GMM in an efficient probabilistic framework that

568 J. Pilet, C. Strecha, and P. Fua
Fig. 1. Top row: Three very different input images and a model image of the same
scene. The changes are caused by lights being turned on one after the other and the
person moving about. Bottom row: Our algorithm successfully segments out the person
in all three input images. The rightmost image depicts the completely wrong output of
a state-of-the-art approach [1] applied on the third image.
accounts for texture, background illumination, and foreground colour clues. Its
parameters are computed by Expectation Maximization (EM) [5].
This approach reflects our key insight that, assuming that the background is
static, changes in intensity of non-occluded pixels are mainly caused by illumi-
nation effects that are relatively global: They are not the same in all parts of
the image but typically affect similarly whole portions of the image as opposed
to individual pixels. As a result, they can be modelled using GMMs with only
few components—2 in the experiments presented in this paper—which leads to
a very robust algorithm.
We will demonstrate that our algorithm outperforms state-of-the-art back-
ground subtraction techniques when illumination changes quickly. The key
difference between these techniques and ours is that they directly estimate dis-
tributions of pixel intensities as opposed to illumination effects as we do. We
will also show that our approach performs well in an Augmented Reality context
where a moving object is treated as the background from which occluders such
as the hands holding it must be segmented out.
2 Related Work
Many background subtraction algorithms try to update on-line a statistical back-

ground model. A pixel from a new image is then classified as background if it
fits the model. Wren et al. [2] represent the colour of each pixel by a three-
dimensional Gaussian, learned from colour observation of consecutive frames.
Since a single Gaussian is a poor approximation of the true probability density
function, GMMs were proposed instead [3,4]. These approaches have proved to
be effective at handling gradual illumination changes and repetitive dynamic
Making Background Subtraction Robust to Sudden Illumination Changes 569
backgrounds. Many improvements have been published since, such as a recent

method that dynamically selects the appropriate number of components for each
pixel [1]. We will use it as a benchmark against which we compare our approach
because it is representative of this whole class of techniques.
Introducing a GMM is not the only way to model a dynamic background.
Elgammal et al. proposed to model both background and foreground pixel in-
tensities by a nonparametric kernel density estimation [6]. In [7], Sheikh and
Shah proposed to model the full background with a single distribution, instead
of one distribution per pixel, and to include location into the model.
Because these methods do not decouple illumination from other causes of back-
ground changes, they are more sensitive to drastic light effects than our approach.
Shadows cast by moving objects cause illumination changes that follow them,
thereby hindering the integration of shadowed pixels into the background model.
This problem can be alleviated by explicitly detecting the shadows [8]. Most of
them consider them as binary [8], with the notable exception of [9] that also con-
siders penumbra by using the ratio between two images of a planar background.
Our approach also relies on image ratios, but treats shadows as a particular illu-
mination effect, a wider class that also include the possibility of switching lights
on.
Another way to handle illumination changes is by using illumination invariant
features, such as edges. Edge information alone is not sufficient, because some
part of the background might be uniform. Thus, Jabri et al. presented an ap-
proach to detect people fusing colour and edge information [10]. More recently,
Heikkilä and Pietikäinen modelled the background using histograms of local bi-
nary patterns [11]. The bilayer segmentation of live video presented in [12] fuses
colour and motion clues in a probabilistic framework. In particular, they observe
in a labeled training set the relation between the image features and their target
segmentation. We follow here a similar idea by training beforehand histograms
of correlation and amount of texture, allowing us to fuse illumination, colour
and texture clues.
3 Method
Our method can serve in two different contexts. For background subtraction,
where both the scene and the camera are static. For augmented reality applica-
tions, where an object is moving in the camera field and occlusions have to be
segmented for realistic augmentation.
Let us assume that we are given an unoccluded model image of a background
scene or an object. Our goal is to segment the pixels of an input image in two
parts, those that belong to the same object in both images and those that are
occluded. If we are dealing with a moving object, we first need to register the
input image and create an image that can be compared to the model image
pixelwise. In this work, we restrict ourselves to planar objects and use publicly
available software [13] for registration. If we are dealing with a static scene
and camera, that is, if we are performing standard background subtraction,
registration is not necessary. It is the only difference between both contexts, and
the rest of the method is common. In both cases, the intensity and colour of
individual pixels are affected mostly by illumination changes and the presence
of occluding objects.
Changes due to illumination effects are highly correlated across large portions
of the image and can therefore be represented by a low dimensional model that
accounts for variations across the whole image. In this work, we achieve this by
representing the ratio of intensities between the stored background image and an
input image in all three channels as a Gaussian Mixture Model (GMM) that has
very few components—2 in all the experiments shown in this paper. This is in
stark contrast with more traditional background subtraction methods [2,3,4,1]
that introduce a model for each pixel and do not explicitly account for the fact
that inter-pixel variations are correlated.
Following standard practice [14], we model the pixel colours of occluding ob-
jects, such as people walking in front of the camera, as a mixture of Gaussian
and uniform distributions.
To fuse these clues, we model the whole image — background, foreground and
shadows — with a single mixture of distributions. In our model, each pixel is
drawn from one of five distributions: Two Gaussian kernels account for illumi-
nation effects, and two more Gaussians, completed by a uniform distribution,
represent the foreground. An Expectation Maximization algorithm assigns pix-
els to one of the five distributions (E-step) and then optimizes the distributions
parameters (M-step).
Since illumination changes preserve texture whereas occluding objects radi-
cally change it, the correlation between image patches in the model and input
images provides a hint as to whether pixels are occluded or not in the latter,
especially where there is enough texture.
In order to lower the computational burden, we assume pixel independence.
Since this abusive assumption entails the loss of the relation between a pixel
and its neighbors, it makes it impossible to model texture. However, to circum-
vent this issue, we characterize each pixel of the input image by a five dimen-
sional feature vector: The usual red, green, and blue values plus the normalized
cross-correlation and texturedness values. Feature vectors are then assumed in-
dependent, allowing an efficient maximization of a global image likelihood, by
optimizing the parameters of our mixture. In the remainder of this section, we
introduce in more details the different components of our model.
3.1 Illumination Likelihood Model
First, we consider the background model, which is responsible for all pixels that
have a counterpart in the model image m. If a pixel ui of the input image u
shows the occlusion free target object, the luminance measured by the camera
depends on the light reaching the surface (the irradiance ei ) and on its albedo.
Irradiance ei is function of visible light sources and of the surface normal. Under
the lambertian assumption, the pixel value ui is: ui = ei ai , where ai is the
albedo of the target object at the location pointed by ui . Similarly, we can
write: mi = em ai ,with em assumed constant over the surface. This assumption
is correct if the model image m has been taken under uniform illumination, or
if a textured model free of illumination effects is available. Combining the above
equations yields:
ui ei
li = = ,
mi em
which does not depend on the surface albedo. It depends on the surface orienta-
tion and on the illumination environment. In the specific case of a planar surface
lit by distant light sources and without cast shadows, this ratio can be expected
to be constant for all i [9]. In the case of a 3 channel colour camera, we can write
the function li that computes a colour illumination ratio for each colour band:
% &
ui,r ui,g ui,b T
li = mi,r mi,g mi,b ,
where the additional indices r, g, b denotes the red, green and blue channel of
pixel ui , recpectively.
In our background illumination model we suppose that the whole scene can be
described by K different illumination ratios, that correspond to areas in ui with
different orientations and/or possible cast shadows. Each area is modelled by a
Gaussian distribution around the illumination ratio μk and with full covariance
Σk . Furthermore we introduce a set of binary latent variables xi,k that take the
value 1 iff pixel i belongs to Gaussian k and 0 otherwise. Then, the probability
of the ratio li is given by:

K
x xi,k
p(li | xi , μ, Σ) = πk i,k N (li ; μk , Σk ) , (1)
k=1
where μ, Σ denote all parameters of the K Gaussians. πk weights the relative

importance of the different mixture components. Even though the ratios li are
not directly observed, this model has much in common with a generative model
for illumination ratios.
So far we described the background model. The foreground model is respon-
sible for all pixels that do not correspond to the model image m. These pixels
are assumed to be generated by sampling the foreground distribution, which we
model as a mixture of K̄ Gaussians and a uniform distribution. By this choice,
we implicitely assume that the foreground object is composed of K̄ colours μk ,
handled by the normal distributions N (ui ; μk , Σk ), and some suspicious pixels
that occur with probability 1/2563. Again, as in the background model, the la-
tent variables are used to select a specific Gaussian or the uniform distribution.
The probability of observing a pixel value ui given the state of the latent variable
xi and the parameters μ, Σ is given by:

π xi,K+K̄+1 K̄
K+
K+K̄+1 x xi,k
p(ui | xi , μ, Σ) = πk i,k N (ui ; μk , Σk ) . (2)
2563
k=K+1
The overall model consist of the background (Eq. 1) and the foreground (Eq. 2)
model. Our latent variables xi select the one distribution among the total K+K̄+1
components which is active for pixel i. Consider figures 2(a) and 2(b) for example:
The background pixels could be explained by K = 2 illumination ratios, one
for the cast shadow and one for all other background pixels. The hand in the
foreground could be modelled by the skin colour and the black colour of the shirt
(K̄ = 2). The example in Fig. 2 shows clearly that the importance of the latent
variable components is not equal. In practice, there is often one Gausssian which
models a global illimination change, ı.e. most pixels are assigned to this model
by the latent variable component xi,k . To account for the possibly changing
importance, we have introduced πk that globally weight the contribution of all
Gaussian mixtures k = 1 . . . K̄ and the uniform distribution k = K̄ + 1.
A formal expression of our model requires combining the background pdf of
Eq. 1 and the foreground pdf of Eq. 2. However, one is defined over illumination,
whereas the other over pixel colour, making direct probabilities incompatible.
We therefore express the background model as a function of pixel colour instead
of illumination:
1
p(ui | xi , μ, Σ) = p(li | xi , μ, Σ) , (3)
| Ji |
where | Ji | is the determinant of the Jacobian of function li (ui ). Multiplying
this equation with Eq. 2 composes the complete colour pdf.
Some formulations define an appropriate prior model on the latent variables
x. Such a prior model would incorporate the prior belief that the model se-
lection x shows spatial [14] and spatio-temporal [12] correlations. These priors
on the latent variable x have shown to improve the performance of many vi-
sion algorithms [15]. However, they increase the complexity and slow down the
computation substantially. To circumvent this, we propose in the next section
a spatial likelihood model, which can be seen as a model to capture the spatial
nature of pixels and which allows real-time performance.
3.2 Spatial Likelihood Model

In this section, we present an image feature and a way to learn off-line its rela-
tionship with our target segmentation. Consider an extended image patch around
pixel i for which we extract a low dimensional features vector fi = [fi1 , fi2 ]. The
basic idea behind our spatial likelihood model is to capture texture while keep-
ing a pixel independence assumption. To achieve real-time performance we use
two features that can be computed very fast and model their distribution in-
dependently for the background and for the foreground, by histograms of the
discretized feature values. We use the normalized cross-correlation (NCC) be-
tween input and model image as one feature and a measure of the amount of
texture as the other feature. fi1 is given by:

j∈wi (uj − ūi ) (mj − m̄i )
fi =
1
,
2 2
j∈wi (uj − ūi ) j∈wi (mj − m̄i )

where wi denotes a window around pixel i, and ūi = |w1i | j∈wi uj is the average
over wi . The correlation is meaningful only in windows containing texture. Thus,
the texturedness of window i is quantified by:
(a) (b) (c)
(d) (e) (f)
Fig. 2. Elements of the approach. (a) Background image m. (b) Input image u. (c)
Textureness image f 2 . (d) Correlation image f 1 . (e) Probability of observing f on the
background, according to the histogram h(fi | vi ) (f) Probability of observing f on the
foreground, according to the histogram h̄(fi | v̄i ).
' '
2 2
fi2 = (uj − ūi ) + (mj − m̄i ) .
j∈wi j∈wi
We denote the background and foreground distributions by h(fi | vi ) and

h̄(fi | v̄i ), respectively. They are trained from a set of manually segmented image
pairs. Since joint correlation and amount of texture is modelled, the histograms
remain valid for new illumination conditions and for new backgrounds. There-
fore, the training is done only once, off-line. Once normalized, these histograms
model the probability of observing a feature fi on the backgound or on the fore-
ground. Fig. 3 depicts both distributions. One can see that both distributions
are dissociate, especially in highly textured areas.
Figure 2 shows a pair of model and input images, the corresponding texture
and correlation images fi2 and fi1 , and the results of applying the histograms
to f . It is obvious that the correlation measure is only meaningful in textured
areas. In uniform areas, because NCC is invariant to illumination, it can not
make the difference between a background with some uniform illumination or a
uniform foreground.
Both histograms are learnt in the two cases of background and foreground
which are related to the latent variable xi designing one of the distributions of
our model. Therefore, h can be used together with all background distributions
corresponding to {xi,1 , ..., xi,K } and h̄ with all foreground ones, corresponding
to {xi,K+1 , ..., xi,K+K̄+1 }.
Distribution over background pixels Distribution over foreground pixels
Probability Probability
0.07 0.07
0.06 0.06
0.05 0.05
0.04 0.04
0.03 0.03
0.02 0.02
0.01 120 0.01 120
0 100 0 100
80 80
1 60 1 60
0.5 40 0.5 40
0 20 Textureness 0 20 Textureness
Correlation -0.5 0 Correlation -0.5 0
-1 -1
Fig. 3. Joint correlation and texturedness distributions over background and fore-
ground pixels
3.3 Maximum Likelihood Estimation
Having defined the illumination and the spatial likelihood model we are now
in the position to describe the Maximum Likelihood (ML) estimation of the
combined model. Let θ = {μ, Σ, π} denote the vector of all unknowns. The ML
estimate θ̃ is given by:

θ̃ = arg max log p(u, f , x | θ) (4)
θ x
where p(u, f , x | θ) = p(u, x | θ)p(f , x | θ) represents the combined pdf of the

illumination and the spatial likelihood models given by the product of eqs. 3,
2 and the histrogram distributions h(fi | vi ), h̄(fi | v̄). Since the histogram dis-
tributions are computed over an image patch, the pixel contributions are not
independent. However, in order to reach the real-time constraints, we assume
the factorisation over all pixels i in Eq. 4 to be approximately true. We see this
problem as a trade-off between (i) a prior model on x, that models spatial inter-
actions [12,15] with a higher computational complexity and (ii) a more simple,
real time model for which the independence assumption is violated, in the hope
that the spatially dependent feature description f account for pixel dependence.
The pixel independence assumption simplifies the ML estimate to:

θ̃ = arg max log p(ui , li , fi , xi | θ) (5)
θ i xi
The expectation-maximization (EM) algorithm can maximize equation 5. It

alternates the computation between an expectation step (E-step), and a maxi-
mization step (M-step).
E-Step. On the (t + 1)th iteration the conditional expectation bt+1 of the log-
likelihood w.r.t. the posterior p(x | u, θ) is computed in the E-step. By construc-
tion, i.e. by the pixel independence, this leads to a closed-form solution for the
latent variable expectations bi , which are often called beliefs. Note, that in other
formulations, where the spatial correlation is modelled explicitly, the E-step re-
quires graph-cut optimisation [14] or other iterative approximations like mean
field [15]. The update equations for the expected values bi,k of xi,k are given by:
1 1
bt+1
i,k=1...K = πk N (li ; μtk , Σkt )h(fi | vi ) (6)
N | Ji |
1
bt+1 = πk N (ui ; μtk , Σkt )h̄(fi | v̄i ) (7)
i,k=K+1...K̄ N
1 1
bt+1 = πK+K̄+1 h̄(fi | v̄i ) ,
i,K̄+1 N 2563

where N = k bt+1i,k normalises the beliefs bi,k to one. The first line in Eq. 6
t+1
corresponds to the beliefs that the k th normal distribution of the illumination

background model is active for pixel i. Similarly, the other two lines (Eq. 7)
correspond to the beliefs w.r.t. for the foreground illumination model.
M-Step. Given the beliefs bt+1 i,k , the M-step maximises the log-likelihood by
replacing the binary latent variables xi,k by their expected value bt+1
i,k .

1 N
Nk i=1 bi,k li if k ≤ K ,
t+1
μt+1 = N (8)
i=1 bi,k ui otherwise
k 1 t+1
Nk
N
where Nk = i=1 bik .
t+1
Similarly, we obtain:
N t+1
Nk i=1 bi,k (li − μk ) (li − μk ) if k ≤ K
1 T
Σkt+1 = N (9)
i=1 bi,k (ui − μk )(ui − μk ) otherwise
1 t+1 T
Nk
Nk
πkt+1 = (10)
k Nk
Alternating E and M steps ensure convergence to a local minimum. After conver-

gence, we can compute the segmentation by summing the beliefs corresponding
to the foreground and the background model. The probability of a pixel beeing
described by the background model is therefore given by:

K
p(vi | θ̃, u) = bi,k . (11)
k=1
In the next section, we discuss implementation and performance issues.

Our algorithm can be used in two different manners. First, it can run on-line, with
a single E-M iteration at each frame, which allows fast computation. On very
abrupt illumination changes, convergence is reached after a few frames (rarely

more than 6). Second, the algorithm can run offline, with only two images as
input instead of a video history. In this case, several iterations, typically 5 to 10,
are necessary before convergence.
Local NCC can be computed efficiently with integral images, with a complex-
ity linear with respect to the number of pixels and constant with respect to the
window size. Thus, the complexity of the complete algorithm is also linear with
the number of pixels, and the full process of acquiring, segmenting, and display-
ing images is achieved at a rate of about 2.3 × 106 pixels per second, using a
single core of a 2.0GHz CPU. This is about 18 fps for half PAL (360x288), 12
FPS for 512x384, and 5-6 FPS for 720x576 images.
Correlation and texturedness images, as presented in section 3.2, are computed
from single channel images. We use the green channel only, because it is more
represented on a Bayer pattern. The correlation window is a square of 25 × 25
pixels, cropped at image borders.
For all experiments presented in the paper, K = 2 and K̄ = 2. The histograms
h and h̄ have been computed only once, from 9 pairs of images (about 2 × 106
training pixels). Training images do not contain any pattern or background used
in test experiments.
The function li as presented in previous section is sensitive to limited dynamic
range and to limited precision in low intensity values. Both following functions
assume the same role with more robustness and give good result:
! "T
ui,r ui,g ui,b
lia (ui ) = arctan arctan arctan
mi,r mi,g mi,b
! "T
ui,r + c ui,g + c ui,b + c
lic (ui ) =
mi,r + c mi,g + c mi,b + c
where c is an arbitrary positive constant. In our experiments, we use c = 64.
4 Results
In this section, we show results on individual frames of video sequences that
feature both sudden illumination changes and shadows cast by occluding ob-
jects. We also compare those results to those produced by state-of-the-art tech-
niques [1,11].
4.1 Robustness of Illumination Changes and Shadows

We begin by the sequence of Fig. 5 in which an arm is waved in front of a
cluttered wall. The arm casts a shadow, which affects the scene’s radiosity and
causes the camera to automatically adapt its luminosity settings. With default
parameters, the algorithm of [1] reacts to this by slowly adapting its background
model. However, this adaptation cannot cope with the rapidly moving shadow
and produces the poor result of Fig. 5(a). This can be prevented by increasing
(a) (b) (c)
(d) (e) (f)
Fig. 4. Segmenting the light switch test images from [16]. (a) Background model.
(b) Test image. (c) Manually segmented ground truth. (d) The output of Zivkovic’s
method [1]. (e) Result published in [11], using an approach based on local binary pat-
terns. (f) Our result, obtained solely by comparing (a) and (b). Unlike the other two
methods, we used no additional video frames.
the rate at which the background adapts, but, as shown in Fig. 5(b), it results
in the sleeve being lost. By contrast, by explicitly reevaluating the illumination
parameters at every frame, our algorithm copes much better with this situation,
as shown in Fig. 5(c). To compare these two methods independently of specific
parameter choices, we computed the ROC curve of Fig. 5(d). We take precision
to be the number of pixels correctly tagged as foreground divided by the total
number of pixels marked as foreground and recall to be the number of pixels
tagged as foreground divided by the number of foreground pixels in the ground
truth. The curve is obtained by binarizing using different thresholds for the prob-
ability of Eq. 11. We also represent different runs of [1] by crosses corresponding
to different choices of its learning rate and the decision threshold. As expected,
our method exhibits much better robustness towards illumination effects.
Fig. 1 depicts a sequence with even more drastic illumination changes that
occur when the subject turns on one light after the other. The GMM based-
method [1] immediately reacts by classifying most of the image as foreground. By
contrast, our algorithm correctly compares the new images with the background
image, taken to be the average of the first 25 frames of the sequence.
Fig. 4 shows the light switch benchmark of [16]. We again built the background
representation by averaging 25 consecutive frames showing the room with the
light switched off. We obtain good results when comparing it to an image where
the light is turned on even though, unlike the other algorithms [1,11], we use a
single frame instead of looking at the whole video. To foreground recall of 82%
that appears in [11] entails a precision of only 25%, whereas our method achieves
100
95
90
85
Precision
(a) (b) 80
75
70
65
60 Our method
Zivkovic method
55
100 95 90 85 80 75 70 65 60 55
Recall
(c) (d)
Fig. 5. Segmenting the hand of Fig. 2(b). (a) Result of [1] when the background model
adjusts too slowly to handle a quick illumination change. (b) When the background
model adjusts faster. (d) ROC curve for our method obtained by varying a threshold
on the probability of Eq. 11. The crosses represent results obtained by [1] for different
choices of learning rate and decision threshold.
49% for the same recall. With default parameters, the algorithm of [1] cannot
handle this abrupt light change and yields a precision of 13% for a recall of 70%.
Finally, as shown in Fig. 6, we ran our algorithm on one of the PETS 2006
video sequences that features an abandoned luggage to demonstrate that our
technique is indeed appropriate for surveillance applications because it does not
lose objects by unduly merging them in the background.
4.2 Augmented Reality

Because our approach is very robust to abrupt illumination changes, it is a
perfect candidate for occlusion segmentation in augmented reality. The task is
(a) (b) (c) (d)
Fig. 6. PETS 2006 Dataset. (a) Initial frame of the video, used as background model.
(b) Frame number 2800. (c) The background subtraction of [1]: The abandoned bag in
the middle of the scene has mistakenly been integrated into the background. (d) Our
method correctly segment the bag, the person who left after sitting on the bottom left
corner, and the chair that has been removed on the right.
(a) (b) (c) (d)
Fig. 7. Occlusion segmentation on a moving object. (a) Input frame in which the card
is tracked. (b) Traditional background subtraction provide unsatisfying results because
of the shadow cast by the hand, and because it learned the fingers hiding the bottom
left corner as part of the background. (c): Our method is far more robust and produces
a better segmentation. (d) We use its output as an alpha channel to convincingly draw
the virtual text and account for the occluding hand.
the following: A user holds an object that is detected and augmented. If the
detected pattern is occluded by a real object, the virtual object should also be
occluded. In order to augment only the pixels actually showing the pattern, a
visibility mask is required. Technically, any background subtraction technique
could produce it, by unwarping the input images in a reference frame, and by
rewarping the resulting segmentation back to the input frame.
The drastic illumination changes produced by quick rotation of the pattern
might hinder a background subtraction algorithm that has not been designed for
such conditions. That is why the Gaussian mixture based background subtraction
method of [1] has difficulties to handle our test sequence illustrated by figure 7.
On the other hand, the illumination modeling of our approach is able to handle
this situation well and, unsurprisingly, shows superior results. The quality of the
resulting segmentation we obtain allows convincing occluded augmented reality,
as illustrated by figure 7(d).
5 Conclusion
We presented a fast background subtraction algorithm that handles heavy illumi-
nation changes by relying on a statistical model, not of the pixel intensities, but
of the illumination effects. The optimized likelihood also fuses texture correlation
clues by exploiting histograms trained off-line.
We demonstrated the performance of our approach under drastic light changes
that state-of-the-art technique have trouble to handle.
Moreover, our technique can be used to segment the occluded parts of a mov-
ing planar object and therefore allows occlusion handling for augmented reality
applications.
Although we do not explicitely model spatial consistency, the learnt his-
tograms of correlation captures texture. Similarly, we could easily extend our
method by integrating temporal dependence using temporal features.
References
1. Zivkovic, Z., van der Heijden, F.: Efficient adaptive density estimation per image
pixel for the task of background subtraction. Pattern Recognition Letters 27(7),
773–780 (2006)
2. Wren, C., Azarbayejani, A., Darrell, T., Pentland, A.: Pfinder: Real-time tracking
of the human body. In: Photonics East, SPIE, vol. 2615 (1995)
3. Friedman, N., Russell, S.: Image segmentation in video sequences: A probabilistic
approach. In: Annual Conference on Uncertainty in Artificial Intelligence, pp. 175–
181 (1997)
4. Stauffer, C., Grimson, W.: Adaptive background mixture models for real-time
tracking. In: CVPR, pp. 246–252 (1999)
5. Bishop, C.: Pattern Recognition and Machine Learning. Springer, Heidelberg
(2006)
6. Elgammal, A., Duraiswami, R., Harwood, D., Davis, L.: Background and fore-
ground modeling using nonparametric kernel density for visual surveillance. Pro-
ceedings of the IEEE 90, 1151–1163 (2002)
7. Sheikh, Y., Shah, M.: Bayesian modeling of dynamic scenes for object detection.
PAMI 27, 1778–1792 (2005)
8. Prati, A., Mikic, I., Trivedi, M., Cucchiara, R.: Detecting moving shadows: Algo-
rithms and evaluation. PAMI 25, 918–923 (2003)
9. Stauder, J., Mech, R., Ostermann, J.: Detection of moving cast shadows for object
segmentation. IEEE Transactions on Multimedia 1(1), 65–76 (1999)
10. Jabri, S., Duric, Z., Wechsler, H., Rosenfeld, A.: Detection and location of people in
video images using adaptive fusion of color and edge information. In: International
Conference on Pattern Recognition, vol. 4, pp. 627–630 (2000)
11. Heikkila, M., Pietikainen, M.: A texture-based method for modeling the background
and detecting moving objects. PAMI 28(4), 657–662 (2006)
12. Criminisi, A., Cross, G., Blake, A., Kolmogorov, V.: Bilayer segmentation of live
video. In: CVPR, pp. 53–60 (2006)
13. Lepetit, V., Pilet, J., Geiger, A., Mazzoni, A., Oezuysal, M., Fua, P.: Bazar,
http://cvlab.epfl.ch/software/bazar
14. Rother, C., Kolmogorov, V., Blake, A.: Grabcut: Interactive foreground extraction
using iterated graph cuts. ACM SIGGRAPH (2004)
15. Fransens, R., Strecha, C., Van Gool, L.: A mean field EM-algorithm for coherent
occlusion handling in map-estimation problems. In: CVPR (2006)
16. Toyama, K., Krumm, J., Brumitt, B., Meyers, B.: Wallflower: principles and prac-
tice of background maintenance. In: International Conference on Computer Vision,
vol. 1, pp. 255–261 (1999)
Closed-Form Solution to Non-rigid 3D Surface
Registration
Mathieu Salzmann, Francesc Moreno-Noguer,

Vincent Lepetit, and Pascal Fua
EPFL - CVLab,
1015 Lausanne, Switzerland
Abstract. We present a closed-form solution to the problem of recov-

ering the 3D shape of a non-rigid inelastic surface from 3D-to-2D corre-
spondences. This lets us detect and reconstruct such a surface by matching
individual images against a reference configuration, which is in contrast
to all existing approaches that require initial shape estimates and track
deformations from image to image.
We represent the surface as a mesh, and write the constraints provided
by the correspondences as a linear system whose solution we express as
a weighted sum of eigenvectors. Obtaining the weights then amounts to
solving a set of quadratic equations accounting for inextensibility con-
straints between neighboring mesh vertices. Since available closed-form
solutions to quadratic systems fail when there are too many variables, we
reduce the number of unknowns by expressing the deformations as a lin-
ear combination of modes. The overall closed-form solution then becomes
tractable even for complex deformations that require many modes.
1 Introduction
3D shape recovery of deformable surfaces from individual images is known to be
highly ambiguous. The standard approach to overcoming this is to introduce a
deformation model and to recover the shape by optimizing an objective func-
tion [1,2,3,4,5,6,7,8] that measures the fit of the model to the data. However, in
practice, this objective function is either non-convex or involves temporal con-
sistency. Thus, to avoid being trapped in local minima, these methods require
initial estimates that must be relatively close to the true shape. As a result, they
have been shown to be effective for tracking, but not for registration without a
priori shape knowledge.
By contrast, we propose here a solution to detecting and reconstructing in-
elastic 3D surfaces from correspondences between an individual image and a
reference configuration, in closed-form, and without any initial shape estimate.
More specifically, we model flexible inelastic surfaces as triangulated meshes
whose edge lengths cannot change. Given an image of the surface in a known

This work was supported in part by the Swiss National Science Foundation and in
part by the European Commission under the IST-project 034307 DYVINE (Dynamic
Visual Networks).

582 M. Salzmann et al.
Fig. 1. 3D Reconstruction of non-rigid objects from and individual image and a refer-
ence configuration. Results were obtained in closed-form, without any initial estimate.
Top: Recovered mesh overlaid on the original image. Bottom: Re-textured side view of
the retrieved surface.
3D configuration, and correspondences between that model image and an input

image in which the shape is unknown, retrieving the mesh’s vertex coordinates
involves solving a rank-deficient linear system encoding the projection equations.
Taking our inspiration from our recent paper on rigid object pose estimation [9],
we express the solution of this linear system as a weighted sum of the correspond-
ing matrix’s eigenvectors associated with the smallest eigenvalues. We compute
these weights by using Extended Linearization [10] to solve a set of quadratic
constraints that preserve edge lengths. In its simplest form, this method is only
directly applicable to very small meshes because, for larger ones, the number of
unknowns after Extended Linearization grows fast, thus yielding an intractable
problem. We overcome this difficulty by expressing the surface deformations as
a linear combination of deformation modes. This preserves the linear formula-
tion of the correspondence problem, but dramatically reduces the size of the
corresponding linear system, while improving its conditioning. Therefore, the
quadratic constraints required to guarantee inextensibility are also expressed in
terms of a smaller number of variables, making Extended Linearization practical.
As a result, we can solve our problem in closed-form even when using enough
modes to model complex deformations such as those of Fig. 1, which yields a
3D reconstruction that jointly minimizes edge length variations and reprojects
correctly on the input image.
2 Related Work
3D reconstruction of non-rigid surfaces from images has attracted increasing
attention in recent years. It is a severely under-constrained problem and many
different kinds of prior models have been introduced to restrict the space of
possible shapes to a manageable size.
Closed-Form Solution to Non-rigid 3D Surface Registration 583
Most of the models currently in use trace their roots to the early physics-
based models that were introduced to delineate 2D shapes [11] and reconstruct
relatively simple 3D ones [12].
As far as 2D problems are concerned, their more recent incarnations have
proved effective for image registration [13,14] and non-rigid surface detection
[15,16]. Many variations of these models have also been proposed to address
3D problems, including superquadrics [1], triangulated surfaces [2], or thin-
plate splines [17]. Additionally, dimensionality reduction was introduced through
modal analysis [3,18], where shapes are represented as linear combinations of
deformation modes. Finally, a very recent work [19] proposes to set bounds on
distances between feature points, and use them in conjunction with a thin-plate
splines model to reconstruct inextensible surfaces.
One limitation of the physics-based models is that they rarely describe ac-
curately the non-linear physics of large deformations. In theory, this could be
remedied by introducing more sophisticated finite-element modeling. However,
in practice, this often leads to vastly increased complexity without a commensu-
rate gain in performance. As a result, in recent years, there has been increasing
interest in statistical learning techniques that build surface deformation models
from training data. Active Appearance Models [20] pioneered this approach by
learning low-dimensional linear models for 2D face tracking. They were quickly
followed by Active Shape Models [5] and Morphable Models [4] that extended it
to 3D. More recently, linear models have also been learned for structure-from-
motion applications [6,21] and tracking of smoothly deforming 3D surfaces [7].
There has also been a number of attempts at performing 3D surface recon-
struction without resorting to a deformation model. One approach has been
to use lighting information in addition to texture clues to constrain the recon-
struction process [8], which has only been demonstrated under very restrictive
assumptions on lighting conditions and is therefore not generally applicable.
Other approaches have proposed to use motion models over video sequences.
The reconstruction problem was then formulated either as solving a large lin-
ear system [22] or as a Second Order Cone Programming problem [23]. These
formulations, however, rely on tightly bounding the vertex displacements from
one frame to the next, which makes them applicable only in a tracking context
where the shape in the first frame of the sequence is known.
In all the above methods, shape recovery entails minimizing an objective func-
tion. In most cases, the function is non convex, and therefore, one can never be
sure to find its global minimum, especially if the initial estimate is far from the
correct answer. In the rare examples formulated as convex problems [23], the so-
lution involves temporal consistency, which again requires a good initialization.
By contrast, many closed-form solutions have been proposed for pose estima-
tion of rigid objects [24,25,26]. In fact, the inspiration for our method came from
our earlier work [9] in that field. However, reconstructing a deformable surface
involves many more variables than the 6 rigid motion degrees of freedom. In
the remainder of this paper, we show that this therefore requires a substantially
different approach.
3 Closed-Form 3D Reconstruction
In this section, we show that recovering the 3D shape of a flexible surface from
3D-to-2D correspondences can be achieved by solving a set of quadratic equations
accounting for inextensibility, which can be done in closed-form.
3.1 Notations and Assumptions
We represent our surface as a triangulated mesh made of nv vertices vi =

[xi , yi , zi ]T , 1 ≤ i ≤ nv connected by ne edges. Let X = [v1T , · · · , vnTv ]T be
the vector of coordinates obtained by concatenating the vi .
We assume that we are given a set of nc 3D-to-2D correspondences between
the surface and an image. Each correspondence relates a 3D point on the mesh,
expressed in terms of its barycentric coordinates in the facet to which it belongs,
and a 2D feature in the image.
Additionally, we assume the camera to be calibrated and, therefore, that its
matrix of intrinsic parameters A is known. To simplify our notations without
loss of generality, we express the vertex coordinates in the camera referential.
3.2 Linear Formulation of the Correspondence Problem
We first show that, given a set of 3D-to-2D correspondences, the vector of vertex
coordinates X can be found as the solution of a linear system.
Let x be a 3D point belonging to facet f with barycentric coordinates [a1 , a2 , a3 ].
3
Hence, we can write it as x = i=1 ai vf,i , where {vf,i }i=1,2,3 are the three ver-
tices of facet f . The fact that x projects to the 2D image location (u, v) can now
be expressed by the relation
⎡ ⎤
u
A (a1 vf,1 + a2 vf,2 + a3 vf,3 ) = k ⎣ v ⎦ , (1)
1
where k is a scalar accounting for depth. Since, from the last row of Eq. 1, k can
be expressed in terms of the vertex coordinates, we have
⎡ ⎤
! "
# $ vf,1 u
a1 B a2 B a3 B ⎣ ⎦
vf,2 = 0 , with B = A2×3 − A3 , (2)
v
vf,3
where A2×3 are the first two rows of A, and A3 is the third one. nc such corre-
spondences between 3D surface points and 2D image locations therefore provide
2nc linear constraints such as those of Eq. 2. They can be jointly expressed by
the linear system
MX = 0 , (3)
# $
where M is a 2nc × 3nv matrix obtained by concatenating the a1 B a2 B a3 B
matrices of Eq. 2.
Although solving this system yields a surface that reprojects correctly on

the image, there is no guarantee that its 3D shape corresponds to reality. This
stems from the fact that, for all practical purposes, M is rank deficient. More
specifically, even where there are many correspondences, one third, i.e. nv , of the
eigenvalues of MT M are very close to zero [22], as illustrated by Fig. 2(c). As
a result, even small amounts of noise produce large instability in the recovered
shape.
This suggests that additional constraints have to be added to guarantee a
unique and stable solution. In most state-of-the-art approaches, these constraints
are provided by deformation models and are enforced via an iterative method.
By contrast, we will argue that imposing inextensibility of the surface yields a
closed-form solution to the problem.
(a) (b)
5
6 6 x 10
x 10 x 10 5
3 3
4.5
2.5 2.5 4
3.5
2 2
3
1.5 1.5 2.5
2
1 1
1.5
1
0.5 0.5
0.5
0 0 0
50 100 150 200 5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
(c) (d) (e)
Fig. 2. (a,b) Original and side views of a surface used to generate a synthetic sequence.
The 3D shape was reconstructed by an optical motion capture system. (c,d) Eigenval-
ues of the linear system written from correspondences randomly established for the
synthetic shape of (a). (c) The system was written in terms of 243 vertex coordinates.
One third of the eigenvalues are close to zero. (d) The system was written in terms of
50 PCA modes. There are still a number of near zero eigenvalues. (e) First derivative
of the curve (d) (in reversed x-direction). We take the maximum value of nl to be the
one with maximum derivative, which corresponds to the jump in (d).
3.3 Inextensible Meshes
Following the idea introduced in [9], we write the solution of the linear system of
Eq. 3 as a weighted sum of the eigenvectors li , 1 ≤ i ≤ nv of MT M, which are
those associated with the eigenvalues that are almost zero. Therefore we write

nv
X= βi l i , (4)
i=1
since any such linear combination of li is in the kernel of MT M and produces

a mesh that projects correctly on the image. Our problem now becomes finding
appropriate values for the βi , which are the new unknowns.
We are now in a position to exploit the inextensibility of the surface by choos-
ing the βi so that edge lengths are preserved. Such βi can be expressed as the
solution of a set of quadratic equations of the form

nv
nv
βi lji − βi lki 2 = vjref − vkref 2 , (5)
i=1 i=1
where lji is the 3×1 sub-vector of li corresponding to the coordinates of vertex

vj , and vjref and vkref are two neighboring vertices in the reference configuration.
3.4 Extended Linearization

Typical closed-form approaches to solving systems of quadratic equations involve
linearizing the system and introducing new unknowns for the quadratic terms.
This results in a system of the form
Db = d , (6)
where b = [β1 β1 , · · · , β1 βnv , β2 β2 , · · · , β2 βnv , · · · , βnv βnv ]T is the vector of quad-
ratic terms, of size nv (nv + 1)/2. D is a ne × nv (nv + 1)/2 matrix built from the
known li , and d is the ne ×1 vector of edge lengths in the reference configuration.
Unfortunately, since, in hexagonal meshes, the number of edges grows as 3nv , the
number of quadratic unknown terms in the linearized system quickly becomes
larger than the number of equations.
In this paper, we solve this problem by using Extended Linearization [10], a
simple and powerful approach to creating new equations in a linearized system,
which performs better than Groebner bases and relinearization. The idea is to
multiply the original set of equations by the monomials, and linearize the re-
sulting system. In our particular case, we can, for example, multiply the existing
quadratic equations by each of the linear terms, thus creating new equations of
the form
9 n :
v nv

j
β1 βi l i − βi l i
k 2
= β1 vjref − vkref 2 ,
i=1 i=1
..
9 : .

nv
nv

βnv βi lji − βi lki 2 = βnv vjref − vkref 2 .
i=1 i=1
Let b = [β1 β1 β1 , · · · , β1 β1 βnv , β1 β2 β2 , · · · , β1 β2 βnv , β2 β2 β2 , · · · , βnv βnv βnv ]T ,

c
and bl = [β1 , · · · , βnv ]T . The resulting system can be written as

⎡ d1 ⎤
⎡ ⎤
0 · · · 0 D1,1 · · · 1,nv 2,2
· · · nv ,nv
· · · · · · · · · ! "
1 D1 D 1 D1 0 0
bl ⎢ ... ⎥
⎣ · · · · · · · · · · · · · · · · · · · · · · · · · · · · 1,1
−d 0 · · · 0 · · · · · · ··· ··· 0 D
·· ··· ··· ··· ···⎦
· · · Dnv ,nv 0 · · ·
b
c
=⎢ ⎥
⎣ 0 ⎦ , (7)
1 1 1 b
··· ··· ··· ··· ··· ··· ··· ··· ··· ··· ··· ··· ··· ··· .
.
.
where we only show the first line of the original system of Eq. 6 and its prod-
uct with β1 , and where Di,j 1 stands for the coefficient on the first line of D
corresponding to the product βi βj .
It can be shown that multiplying the inextensibility equations by all the βi
only yields a sufficient number of equations for very small meshes, i.e. less than
12 vertices for a hexagonal mesh. In theory, one could solve this problem by
applying Extended Linearization iteratively by re-multiplying the new equations
by the linear terms. However, in practice, the resulting system quickly becomes
so large that it is intractable, i.e. for a 10 × 10 mesh, the number of equations
only becomes larger than the number of unknowns when the size of the system
is of the order 1010 . In other words, Extended Linearization cannot deal with a
problem as large as ours and we are not aware of any other closed-form approach
to solving systems of quadratic equations that could. We address this issue in
the next section.
3.5 Linear Deformation Model

As discussed above, to solve the set of quadratic equations that express edge
length preservation, we need to reduce its size to the point where Extended
Linearization becomes a viable option. Furthermore, we need to do this in such
a way that the solution of the correspondence problem can still be expressed as
the solution of a system of linear equations, as discussed in Section 3.2. To this
end, we model the plausible deformations of the mesh as a linear combination
of nm deformation modes [6,7], much in the same spirit as those the morphable
models used to represent face deformations [4]. We write

nm
X = X0 + αi pi = X0 + Pα , (8)
i=1
where the pi are the deformation modes and the αi their associated weights.
In our implementation, modes were obtained by applying Principal Component
Analysis to a matrix of registered training meshes in deformed configurations,
from which the mean shape X0 was subtracted [7]. The pi therefore are the
eigenvectors of the data covariance matrix. Nonetheless, they could also have
been derived by modal analysis, which amounts to computing the eigenvectors
of a stiffness matrix, and is a standard approach in physics-based modeling [3].
In this formulation, recovering the shape amounts to computing the weights
α. Since the shape must satisfy Eq. 3, α must then satisfy
M(X0 + Pα) = 0 . (9)
When solving this system, to ensure that the recovered weights do not generate
shapes exceedingly far from our training data, we introduce a regularization term
by penalizing αi with the inverse of the corresponding eigenvalue σi of the data
covariance matrix. We therefore solve
! "! "
MP MX0 α
=0, (10)
wr S 0 1
where S is an nm × nm diagonal matrix whose elements are the σi−1 and wr is a

regularization weight that only depends on the maximum σi , and whose precise
value has only little influence on the results.
As shown in Fig. 2(d), we have considerably reduced the number of near-zero
eigenvalues. The system of Eq. 10 is therefore better conditioned than the one
of Eq. 3, but still does not yield a well-posed problem that would have a unique
solution. This is attributable to the fact that, because the solution is expressed
as a sum of deformation modes, inextensibility constraints, which are non linear,
are not enforced.
Nonetheless, we can follow the same procedure as in Sections 3.3 and 3.4.
We write the solution of the linear system of Eq. 10 as a weighted sum of the
eigenvectors l̃i , 1 ≤ i ≤ nl ) nm associated with the smallest eigenvalues of
its matrix, and find the weights β̃i as the solution of the linearized system of
quadratic equations
D̃b̃ = d̃ , (11)
where b̃ = [β̃1 , · · · , β̃nl , β̃1 β̃1 , · · · , β̃1 β̃nl , β̃2 β̃2 , · · · , β̃2 β̃nl , · · · , β̃nl β̃nl ]T now also
contains the linear terms arising in the quadratic equations from the mean shape
X0 . Furthermore, the system also encodes the additionnal linear equation that
constrains the β̃i l̃i,nm +1 to sum up to 1, where l̃i,nm +1 is the last element of l̃i .
Since in practice nl ) nm ) nv , the system is now much smaller. Therefore
a single iteration of Extended Linearization is sufficient to constrain its solution
while keeping it tractable, even for relatively large numbers of modes—in practice
up to 60—thus allowing complex deformations.
In this formulation, the number nl of eigenvectors strongly depends on the
number nm of modes used for the recovery. However, as shown in Fig. 2(e),
we can easily set the maximum number n̂l of eigenvectors to use by picking the
number corresponding to the maximum first derivative of the ordered eigenvalues
curve. We then simply test for all nl ≤ n̂l and pick the optimal value as the one
that, for a small enough reprojection error, gives the smallest mean edge length
variation. In practice, n̂l was typically about 25 when using 60 deformation
modes.
In this section we show that our method can be successfully applied to recon-
structing non-rigid shapes from individual images and a reference configuration.
We present results on both synthetic data and real images.
4.1 Synthetic Data

We first applied our method to images, such as those of Fig. 2(a), synthesized
by projecting known deformed shapes using a virtual camera. The deformed
shapes were obtained by recovering the 3D locations of reflective markers stuck
on a 200×200mm piece of cardboard with an optical motion capture system.
This allowed us to randomly create ncf perfect correspondences per facet to
σ = 0, r = 0% σ = 5, r = 0% σ = 10, r = 0%
g o g o g o
ncf = 5 ncf = 5 ncf = 5

Mean 3D distance [mm]

40 40 40
n =1 n =1 n =1
cf cf cf
ncf = 0.5 ncf = 0.5 ncf = 0.5
30 30 30
20 20 20
10 10 10
0 0 0
2 4 6 8 10 12 2 4 6 8 10 12 2 4 6 8 10 12
Mean curvature x 10
−4 Mean curvature x 10
−4
σ = 0, r = 5% σ = 5, r = 5% σ = 10, r = 5%
g o g o g o
ncf = 5 ncf = 5 ncf = 5


40 40 40
ncf = 1 ncf = 1 ncf = 1
n = 0.5 n = 0.5 n = 0.5
cf cf cf
30 30 30
20 20 20
10 10 10
0 0 0
2 4 6 8 10 12 2 4 6 8 10 12 2 4 6 8 10 12
Mean curvature x 10
−4
σ = 0, r = 10% σ = 5, r = 10% σ = 10, r = 10%

g o g o g o
ncf = 5 ncf = 5 ncf = 5

40
ncf = 1
40
ncf = 1 Mean 3D distance [mm] 40
ncf = 1
ncf = 0.5 ncf = 0.5 ncf = 0.5
30 30 30
20 20 20
10 10 10
0 0 0
2 4 6 8 10 12 2 4 6 8 10 12 2 4 6 8 10 12
Mean curvature x 10
−4
Fig. 3. Shape recovery of a 200×200mm synthetic mesh imaged by a virtual camera

placed 20cm away from it. Each plot shows the mean vertex-to-vertex 3D distance be-
tween the recovered surface and the ground-truth as a function of its mean curvature.
The three different curves in each graph correspond to a varying number of corre-
spondences per facet. Left to right, the gaussian noise added to the correspondences
increases. Top to bottom, the number of outliers grows. For each experiments, we plot
the average over 40 trials. The rightmost column shows in blue recovered shapes for
the ground-truth surface of Fig. 2(a,b), shown in red. The corresponding mean vertex-
to-vertex distances are 9mm, 19mm and 38mm. This highlights the fact that even for
distances around 40mm, the recovered shape remains meaningful.
which we added zero mean gaussian noise of variance σg . Finally, we simulated

outliers by setting the image coordinates of ro percents of the correspondences
to uniformly and randomly distributed values.
In Fig. 3, we show results as a function of the surface’s mean curvature,
the maximum one being that of Fig. 2(a). Each plot includes three curves cor-
responding to ncf = {5, 1, 1/2}, which depict the mean vertex-to-vertex 3D
distance between the recovered mesh and ground-truth. The plots are ordered
on a grid whose x-direction corresponds to σg = {0, 5, 10} and y-direction to
ro = {0%, 5%, 10%}. Each experiment was repeated 40 times, and we show the
average results. Note that the error grows with the mean curvature of the shape,
which is natural since the shape becomes more ambiguous when seen from the
viewpoint shown in Fig. 2(a). In the rightmost column, we display three shapes
reconstructed from the image of Fig. 2(a) with their corresponding ground-truth.
Note that even for average distances of 40mm between the true and recovered
σg = 5, ro = 0%, ncf = 5 σg = 5, ro = 0%, ncf = 5
Mean reprojection error [pixels]

15 10
closed−form closed−form

constrained opt. constrained opt.
8
10
6
4
5
2
0 0
2 4 6 8 10 12 2 4 6 8 10 12
Mean curvature x 10
−4
(a) (b)
Fig. 4. Comparison of our closed-form results against the results of constrained op-
timization. Optimization was performed on the vertex coordinates using Matlab’s
fmincon function, and starting from the flat position. (a) Mean vertex-to-vertex dis-
tance. (b) Reprojection error. Constrained optimization is both much slower and far
less accurate than our approach.
shape, the latter remains meaningful and could be used to initialize an iterative
algorithm.
In Fig. 4, we compare our results against results obtained with Matlab’s con-
strained optimization fmincon function. We use it to minimize the residual of
the linear system of Eq. 3 with respect to the vertex coordinates, under the
constraints that edge lengths must remain constant. We first tried to use the
similar representation in terms of modes. However, since the constraints could
never be truly satisfied, the algorithm would never converge towards an accept-
able solution. This forced us to directly use the vertex coordinates. To improve
convergence and prevent the surface from crumpling, we added a smoothness
term [11]. For all the frames, the initialization was set to the flat position. In
Fig. 4(a), we show the mean 3D vertex-to-vertex distance for the case where
σg = 5, ro = 0, and ncf = 5. The red curve corresponds to our closed-form solu-
tion and the blue one to constrained optimization. Note that our approach gives
much better results. Furthermore, it is also much faster, requiring only 1.5 min-
utes per frame as opposed to 1.5 hours for constrained optimization. Fig. 4(b)
shows the reprojection errors for the same cases.
4.2 Real Images

We tested our method on a folded bed-sheet, a piece of cloth and a t-shirt de-
forming in front of a 3-CCD DV-camera. In all these cases, we first established
SIFT [27] correspondences between the reference image and the input one. We
then detected the surface in 2D, which can be done in closed-form by sim-
ply solving the linear system built from SIFT matches, augmented with linear
smoothing equations [11]. For each facet, we then warped the reference image to
best match the input one based on the retrieved 2D shape, and finally established
dense correspondences by sampling the barycentric coordinates of the facet, and
matching small regions between the input image and the warped reference one
using normalized cross-correlation. Note that, even when we show results on
Fig. 5. 3D registration of a folded bed-sheet to an individual image given a reference

configuration. Top Row: Recovered mesh overlaid on the original image. Middle Row:
Synthesized textured view using the recovered shape. Bottom Row: Real side view
of the sheet from similar viewpoints. Despite lighting changes, the synthetic images
closely match the real ones.
Fig. 6. Shape recovery of a bed-sheet. Top Row: Recovered mesh overlaid on the orig-
inal image. Bottom Row: Mesh seen from a different viewpoint.
video sequences, nothing links one frame to the next, and no initialization is
required. Corresponding videos are given as supplementary material.
In the case of the sheet, we deformed it into several unrelated shapes, took
pictures from 2 different views for each deformation, and reconstructed the sur-
face from a single image and a reference configuration. In Fig. 5, we show the
results on four different cases. From our recovered shape, we generated synthetic
textured images roughly corresponding to the viewpoint of the second image.
As can be seen in the two bottom rows of Fig. 5, our synthetic images closely
match the real side views. Additionally, we also reconstructed the same sheet
Fig. 7. Shape recovery of a piece of cloth. From Top to Bottom: Mesh computed in
closed-form overlaid on the input image, side view of that mesh, refined mesh after 5
Gauss-Newton iterations.
Fig. 8. Shape recovery of the central part of a t-shirt. From Top to Bottom: Mesh
computed in closed-form overlaid on the input image, side view of that mesh, refined
mesh after 5 Gauss-Newton iterations.
from the images of a video sequence, and show the results in Fig. 6. Note that
no initialization was required, and that nothing links one frame to the next.
In Figs. 7 and 8, we show results for images of a piece of cloth and of a
t-shirt waved in front of the camera. Note that in both cases, the closed-form
solution closely follows what we observe in the videos. To further refine it, we
implemented a simple Gauss-Newton optimization technique, and minimize the
residual D̃b̃ − d̃ corresponding to Eq. 11 with respect to the β̃i . In the third
row of the figures, we show the refined mesh after 5 iterations this scheme.
This proved sufficient to recover finer details at a negligible increase in overall
computation time.
5 Conclusion
In this paper, we presented a closed-form solution to the problem of recover-

ing the shape of a non-rigid inelastic surface from an individual image and a
reference configuration. We showed that the reconstruction could be obtained
by solving a system of quadratic equations representing distance constraints be-
tween neighboring mesh vertices.
In future work, we intend to investigate what additional quadratic constraints
could be introduced to the current system of distance constraints. They could
come from additional sources of image information, such as lighting. Having
a larger number of quadratic equations would hopefully relieve the need for
Extended Linearization, and result in smaller, and therefore faster to solve, linear
systems.
References
1. Metaxas, D., Terzopoulos, D.: Constrained deformable superquadrics and nonrigid

motion tracking. PAMI 15, 580–591 (1993)
2. Cohen, L., Cohen, I.: Deformable models for 3-d medical images using finite ele-
ments and balloons. In: CVPR, pp. 592–598 (1992)
3. Pentland, A.: Automatic extraction of deformable part models. IJCV 4, 107–126
(1990)
4. Blanz, V., Vetter, T.: A Morphable Model for The Synthesis of 3–D Faces. ACM
SIGGRAPH, 187–194 (1999)
5. Matthews, I., Baker, S.: Active Appearance Models Revisited. IJCV 60, 135–164
(2004)
6. Torresani, L., Hertzmann, A., Bregler, C.: Learning non-rigid 3d shape from 2d
motion. In: NIPS (2003)
7. Salzmann, M., Pilet, J., Ilić, S., Fua, P.: Surface Deformation Models for Non-Rigid
3–D Shape Recovery. PAMI 29, 1481–1487 (2007)
8. White, R., Forsyth, D.: Combining cues: Shape from shading and texture. In:
CVPR (2006)
9. Moreno-Noguer, F., Lepetit, V., Fua, P.: Accurate Non-Iterative O(n) Solution to
the PnP Problem. In: ICCV (2007)
10. Courtois, N., Klimov, A., Patarin, J., Shamir, A.: Efficient algorithms for solving
overdefined systems of multivariate polynomial equations. In: Preneel, B. (ed.)
EUROCRYPT 2000. LNCS, vol. 1807. Springer, Heidelberg (2000)
11. Kass, M., Witkin, A., Terzopoulos, D.: Snakes: Active Contour Models. IJCV 1,
321–331 (1988)
12. Terzopoulos, D., Witkin, A., Kass, M.: Symmetry-seeking Models and 3D Object
Reconstruction. IJCV 1, 211–221 (1987)
13. Bartoli, A., Zisserman, A.: Direct Estimation of Non-Rigid Registration. In: BMVC
(2004)
14. Gay-Bellile, V., Bartoli, A., Sayd, P.: Direct estimation of non-rigid registrations
with image-base self-occlusion reasoning. In: ICCV (2007)
15. Pilet, J., Lepetit, V., Fua, P.: Real-Time Non-Rigid Surface Detection. In: CVPR
(2005)
16. Zhu, J., Lyu, M.R.: Progressive finit newton approach to real-time nonrigid surface
detection. In: ICCV (2007)
17. McInerney, T., Terzopoulos, D.: A Finite Element Model for 3D Shape Reconstruc-
tion and Nonrigid Motion Tracking. In: ICCV (1993)
18. Delingette, H., Hebert, M., Ikeuchi, K.: Deformable surfaces: A free-form shape
representation. Geometric Methods in Computer Vision (1991)
19. Perriollat, M., Hartley, R., Bartoli, A.: Monocular Template-based Reconstruction
of Inextensible Surfaces. In: BMVC (2008)
20. Cootes, T., Edwards, G., Taylor, C.: Active Appearance Models. In: Burkhardt,
H., Neumann, B. (eds.) ECCV 1998. LNCS, vol. 1407. Springer, Heidelberg (1998)
21. Llado, X., Bue, A.D., Agapito, L.: Non-rigid 3D Factorization for Projective Re-
construction. In: BMVC (2005)
22. Salzmann, M., Lepetit, V., Fua, P.: Deformable Surface Tracking Ambiguities. In:
CVPR (2007)
23. Salzmann, M., Hartley, R., Fua, P.: Convex Optimization for Deformable Surface
3–D Tracking. In: ICCV (2007)
24. Quan, L., Lan, Z.: Linear N-Point Camera Pose Determination. PAMI 21, 774–780
(1999)
25. Fiore, P.D.: Efficient linear solution of exterior orientation. PAMI 23, 140–148
(2001)
26. Ansar, A., Daniilidis, K.: Linear pose estimation from points or lines. PAMI 25,
578–589 (2003)
27. Lowe, D.: Distinctive Image Features from Scale-Invariant Keypoints. IJCV 20,
91–110 (2004)
Implementing Decision Trees and Forests on a
GPU
Toby Sharp
Microsoft Research, Cambridge, UK

toby.sharp@microsoft.com
Abstract. We describe a method for implementing the evaluation and

training of decision trees and forests entirely on a GPU, and show how
this method can be used in the context of object recognition.
Our strategy for evaluation involves mapping the data structure de-
scribing a decision forest to a 2D texture array. We navigate through the
forest for each point of the input data in parallel using an efficient, non-
branching pixel shader. For training, we compute the responses of the
training data to a set of candidate features, and scatter the responses
into a suitable histogram using a vertex shader. The histograms thus
computed can be used in conjunction with a broad range of tree learning
algorithms.
We demonstrate results for object recognition which are identical to
those obtained on a CPU, obtained in about 1% of the time.
To our knowledge, this is the first time a method has been proposed
which is capable of evaluating or training decision trees on a GPU. Our
method leverages the full parallelism of the GPU.
Although we use features common to computer vision to demonstrate
object recognition, our framework can accommodate other kinds of fea-
tures for more general utility within computer science.
1 Introduction
1.1 Previous Work
Since their introduction, randomized decision forests (or random forests) have
generated considerable interest in the machine learning community as new tools
for efficient discriminative classification [1,2]. Their introduction in the computer
vision community was mostly due to the work of Lepetit et al in [3,4]. This gave
rise to a number of papers using random forests for: object class recognition and
segmentation [5,6], bilayer video segmentation [7], image classification [8] and
person identification [9].
Random forests naturally enable a wide variety of visual cues (e.g. colour,
texture, shape, depth etc.). They yield a probabilistic output, and can be made
computationally efficient. Because of these benefits, random forests are being
established as efficient and general-purpose vision tools. Therefore an optimized
implementation of both their training and testing algorithms is desirable.

596 T. Sharp
This work is complementary to that of Shotton et al [6] in which the authors

demonstrate a fast recognition system using forests. Although they demonstrate
real-time CPU performance, they evaluate trees sparsely at 1% of pixels and
still achieve only 8 frames per second, whereas our evaluations are dense and
considerably quicker.
At the time of writing, the premium desktop CPU available is the Intel Core
2 Extreme QX9775 3.2 GHz quad-core. This chip has a theoretical peak perfor-
mance of 51.2 Gflops using SSE instructions (12.8 Gflops without SSE). With
DDR3 SDRAM at 200 MHz, system memory bandwidth peaks at 12.8 GB/s.
In contrast, the premium desktop GPU is the nVidia GeForce GTX 280. With
its 240 stream processors it has a theoretical peak of 933 Gflops and a memory
bandwidth of 141 GB/s.
In [10], the authors demonstrate a simple but effective method for performing
plane-sweep stereo on a GPU and achieve real-time performance. In [11], the
authors present belief propagation with a chequerboard schedule based on [12].
We follow in similar fashion, presenting no new theory but a method for realizing
the GPU’s computational power for decision trees and forests.
To our knowledge, this is the first time a method has been proposed which is
capable of evaluating or training decision trees on a GPU. In [13], the authors
explored the implementation of neural networks for machine learning on a GPU,
but did not explore decision trees.
1.2 Outline
Algorithm 1 describes how a binary decision tree is conceptually evaluated on
input data. In computer vision techniques, the input data typically correspond
to feature values at pixel locations. Each parent node in the tree stores a binary
function. For each data point, the binary function at the root node is evaluated
on the data. The function value determines which child node is visited next.
This continues until reaching a leaf node, which determines the output of the
procedure. A forest is a collection of trees that are evaluated independently.
In §2 we describe the features we use in our application which are useful for
object class recognition. In §3, we show how to map the evaluation of a decision
Fig. 1. Left: A 320 × 213 image from the Microsoft Research recognition database
[14] which consists of 23 labeled object classes. Centre: The mode of the pixelwise
distribution given by a forest of 8 trees, each with 256 leaf nodes, trained on a subset
of the database. This corresponds to the ArgMax output option (§3.3). This result was
generated in 7 ms. Right: The ground truth labelling for the same image.
Implementing Decision Trees and Forests on a GPU 597
Algorithm 1. Evaluate the binary decision tree with root node N on input x
1. while N has valid children do
2. if T estF eature(N, x) = true then
3. N ← N.RightChild
4. else
5. N ← N.Lef tChild
6. end if
7. end while
8. return data associated with N
forest to a GPU. The decision forest data structure is mapped to a forest texture
which can be stored in graphics memory. GPUs are highly data parallel machines
and their performance is sensitive to flow control operations. We show how to
evaluate trees with a non-branching pixel shader. Finally, the training of decision
trees involves the construction of histograms – a scatter operation that is not
possible in a pixel shader. In §4, we show how new GPU hardware features
allow these histograms to be computed with a combination of pixel shaders and
vertex shaders. In §5 we show results with speed gains of 100 times over a CPU
implementation.
Our framework allows clients to use any features which can be computed in a
pixel shader on multi-channel input. Our method is therefore applicable to more
general classification tasks within computer science, such as multi-dimensional
approximate nearest neighbour classification. We present no new theory but con-
centrate on the highly parallel implementation of decision forests. Our method
yields very significant performance increases over a standard CPU version, which
we present in §5.
We have chosen Microsoft’s Direct3D SDK and High Level Shader Language
(HLSL) to code our system, compiling for Shader Model 3.
2 Visual Features
2.1 Choice of Features
To demonstrate our method, we have adopted visual features that generalize
those used by many previous works for detection and recognition, including
[15,16,3,17,7]. Given a single-channel
input image I and a rectangle R, let σ
represent the sum σ(I, R) = x∈R I(x).
The features we use are differences of two such sums over rectangles R0 , R1
in channels c0 , c1 of the input data. The response of a multi-channel image I to
a feature F = {R0 , c0 , R1 , c1 } is then ρ(I, F ) = σ(I[c0 ], R0 ) − σ(I[c1 ], R1 ). The
Boolean test at a tree node is given by the threshold function θ0 ≤ ρ(I, F ) < θ1 .
This formulation generalizes the Haar-like features of [15], the summed rectan-
gular features of [16] and the pixel difference features of [3]. The generalization
of features is important because it allows us to execute the same code for all
the nodes in a decision tree, varying only the values of the parameters. This will
enable us to write a non-branching decision evaluation loop.
598 T. Sharp
(a) (b) (c)

Fig. 2. Processing images for feature computation. Each group of four rectangles rep-
resents a four-component (ARGB) texture, and each outline in the group represents a
single component (channel) of the texture. (a) An original sRGB image. (b) The image
is convolved with 16 filters to produce 16 data channels in 4 four-component textures
(§2.3). (c) The filtered textures are then integrated (§2.4).
The rectangular sums are computed by appropriately sampling an integral

image [15]. Thus the input data consists of multiple channels of integral images.
The integration is also performed on the GPU (§2.4).
2.2 Input Data Channels

Prior to integration, we pre-filter sRGB images by applying the bank of sepa-
rable 2D convolution filters introduced in [14] to produce a 16-channel result.
This over-complete representation incorporates local texture information at each
pixel. The convolution is also performed on the GPU (§2.3). The pipeline for
preparing input textures is shown in Figure 2.
2.3 Convolution with the Filter Bank

For object class recognition, we pre-filter images by convolving them with the
17-filter bank introduced in [14] to model local texture information. Whereas
the authors of that work apply their filters in the CIE Lab colour space, we
have found it sufficient to apply ours only to the non-linear R, G, B and Y
channels. The Gaussians are applied to the RGB channels, and the derivative
and Laplacian filters to the luma.
To perform the separable convolution on the GPU, we use the two-pass tech-
nique of [18].
Since the pixel shader operates on the four texture components in parallel,
up to four filters can be applied in one convolution operation. All 17 filters can
therefore be applied in 5 convolutions. In practice we prefer to omit the largest
scale Laplacian, applying 16 filters in 4 convolutions.
2.4 Image Integration

The sums over rectangular regions are computed using integral images [15].
Integral images are usually computed on the CPU using an intrinsically serial
method, but they can be computed on the GPU using prefix sums [19]. This
algorithm is also known as parallel scan or recursive doubling. For details on
how this can be implemented on the GPU, see [20].
bool TestFeature(sampler2D Input, float2 TexCoord, Parameters Params)

{ // Evaluate the given Boolean feature test for the current input pixel
float4 Sum1 = AreaSum(Input, TexCoord, Params.Rect1);
float4 Sum2 = AreaSum(Input, TexCoord, Params.Rect2);
float Response = dot(Sum1, Params.Channel1) - dot(Sum2, Params.Channel2);
return Params.Thresholds.x <= Response && Response < Params.Thresholds.y;
}
Fig. 3. HLSL code which represents the features used to demonstrate our system.
These features are suitable for a wide range of detection and recognition tasks.
2.5 Computation of Features
Figure 3 shows the HLSL code which is used to specify our choice of features
(§2.1). The variables for the feature are encoded in the Parameters structure.
The Boolean test for a given node and pixel is defined by the TestFeature
method, which will be called by the evaluation and training procedures as
necessary.
We would like to stress that, although we have adopted these features to
demonstrate our implementation and show results, there is nothing in our frame-
work which requires us to use a particular feature set. We could in practice use
any features that can be computed in a pixel shader independently at each input
data point, e.g. pixel differences, dot products for BSP trees or multi-level forests
as in [6].
3 Evaluation
Once the input data textures have been prepared, they can be supplied to a
pixel shader which performs the evaluation of the decision forest at each pixel
in parallel.
3.1 Forest Textures
Our strategy for the evaluation of a decision forest on the GPU is to transform
the forest’s data structure from a list of binary trees to a 2D texture (Figure 4).
We lay out the data associated with a tree in a four-component float texture,
with each node’s data on a separate row in breadth-first order.
In the first horizontal position of each row we store the texture coordinate of
the corresponding node’s left child. Note that we do not need to store the right
child’s position as it always occupies the row after the left child. We also store
all the feature parameters necessary to evaluate the Boolean test for the node.
For each leaf node, we store a unique index for the leaf and the required output
– a distribution over class labels learned during training.
To navigate through the tree during evaluation, we write a pixel shader that
uses a local 2D node coordinate variable in place of a pointer to the current node
(Figure 5). Starting with the first row (root node) we read the feature parameters
600 T. Sharp
Fig. 4. Left: A decision tree structure containing parameters used in a Boolean test at
each parent node, and output data at each leaf node. Right: A 7 × 5 forest texture built
from the tree. Empty spaces denote unused values.
float4 Evaluate(uniform sampler2D Forest, uniform sampler2D Input,

uniform float2 PixSize, in float2 TexCoord : TEXCOORD0) : COLOR0
{
float2 NodeCoord = PixSize * 0.5f;
// Iterate over the levels of the tree, from the root down...
[unroll] for (int nLevel = 1; nLevel < MAX DEPTH; nLevel++)
{
float LeftChild = tex2D(Forest, NodeCoord).x;
// Read the feature parameters for this node...
Parameters Params = ReadParams(Forest, NodeCoord, PixSize);
// Perform the user-supplied Boolean test for this node...
bool TestResult = TestFeature(Input, TexCoord, Params);
// Move the node coordinate according to the result of the test...
NodeCoord.y = LeftChild + TestResult * PixSize.y;
}
// Read the output distribution associated with this leaf node...
return Distribution(Forest, NodeCoord);
}
Fig. 5. An HLSL pixel shader which evaluates a decision tree on each input point in
parallel without branching. Here we have omitted evaluation on multiple and unbal-
anced trees for clarity.
and evaluate the Boolean test on the input data using texture-dependent reads.
We then update the vertical component of our node coordinate based on the
result of the test and the value stored in the child position field. This has the
effect of walking down the tree according to the computed features. We continue
this procedure until we reach a row that represents a leaf node in the tree, where
we return the output data associated with the leaf.
For a forest consisting of multiple trees, we tile the tree textures horizontally.
An outer loop then iterates over the trees in the forest; we use the horizontal
component of the node coordinate to address the correct tree, and the vertical
component to address the correct node within the tree. The output distribution
for the forest is the mean of the distributions for each tree.
This method allows our pixel shader to be non-branching (i.e. it does not con-
tain flow control statements) which is crucial for optimal execution performance.
3.2 Geometry and Texture Set-Up

In order to evaluate the forest at every data point, we render a rectangle that
covers the size of the input data. The first texture stage is assigned the forest
texture described above.
There is a slight difficulty with presenting the input data to the shader. Ac-
cording to our discussion of features (§2) we have 16-17 input channels and for
each tree node we wish to choose the correct channel on which to evaluate the
feature response. Unfortunately there is no legal way within the pixel shader
to index an array of textures based on a read operation. Neither is it possible
to create a 3D texture from the input channels and use the required channel to
address the third dimension. To overcome this difficulty, we tile each group of
four input channels into one large four-component texture, and bind the latter
to the second texture stage.
When we create the forest texture, we must modify appropriately the coor-
dinates of the feature rectangles so that they address the correct channel group
within the tiled texture as well as the correct spatial location. We also set unit
vectors (Channel0, Channel1 in Figure 3, 4) which are used in dot products to
select the correct scalar output.
3.3 Output Modes

To give clients sufficient control, we have implemented several different output
modes for forest evaluation:
– Distribution: Outputs the evaluated distribution over L labels into L out-
put channels
– ArgMax: Outputs the index of the label with the greatest probability in the
evaluated distribution
– ForestLeaves: Outputs the index of the leaf node reached in each of T trees
into T output channels
– TreeLeaf: Outputs the index of the leaf node reached in the first tree
Due to limitations of Direct3D 9, a maximum of 16 output channels can be used
for the Distribution and ForestLeaves modes. Where more than 16 label
posteriors are required, they can be computed efficiently in multiple passes from
one or more leaf images generated by the ForestLeaves option. If additionally
the number of trees exceeds 16, the forest can be split into groups of trees which
are evaluated in succession. Thus any number of trees and class labels can be
handled by our system.
4 Training
Training of randomized trees is achieved iteratively, growing a tree by one level
each training round. For each training round, a pool of candidate features is
602 T. Sharp
sampled, and these are then evaluated on all the training data to assess their
discriminative ability. Joint histograms over ground truth labels and feature
responses are created, and these histograms may be used in conjunction with
various learning algorithms, e.g. ID3 [21] or C4.5 [22], to choose features for new
tree nodes. Thus learning trees can be a highly compute-intensive task. We adopt
a general approach for efficient training on the GPU, suitable for any learning
algorithm.
A training database consists of training examples together with ground truth
class labels. Given a training database, a pool of candidate features and a decision
tree, we compute and return to our client a histogram that can be used to extend
the tree in accordance with a learning algorithm. For generality, our histogram is
4D and its four axes are: the leaf node index, ground truth class label, candidate
feature index and quantized feature response. Armed with this histogram, clients
can add two new children to each leaf of the current tree, selecting the most
discriminative candidate feature as the new test.
In one sweep over the training database we visit each labeled data point and
evaluate its response to each candidate feature. We also determine the active
leaf node in the tree and increment the appropriate histogram bin. Thus for each
training round we evaluate the discriminative ability of all candidate features at
all leaf nodes of the current tree.
4.1 Histogram Initialization

We store the 4D histogram in a single 2D texture by tiling 2D slices through
the histogram. The tiling strategy is chosen dynamically at runtime to avoid
exceeding the maximum texture width and height of 4096. Note that the size
of the histogram is independent of the size of the training database. Although
we have not found it necessary in practice, it would be possible to split very
large histograms across multiple textures. We use a 32-bit float texture for
convenience and to avoid overflow.
4.2 Training Data

We now iterate through the training database, evaluating all the feature re-
sponses and accumulating them into the histogram. To access the training data
as required we use a callback interface supplied by the client.
In order to evaluate the feature discrimination for every leaf node of the tree
simultaneously, we need to determine which leaf of the current tree would be
activated for each training data point. We achieve this by the method of §3
using the TreeLeaf output option to generate a leaf image that represents the
active leaf nodes of the current tree for the current training image.
4.3 Feature Evaluation

We request four training examples at a time from the database because this will
allow us to take full advantage of the GPU’s SIMD capabilities by operating on
four texture components in parallel (Figure 6a).
(a) (b) (c) (d)
Fig. 6. Processing training data. (a) A set of four training examples, already pre-
processed for feature computation. (b) The same data, rearranged so that each texture
contains corresponding channels from all four training examples. (c) The appropriate
textures are selected for a given feature and the box filters applied. (d) The final feature
response is the difference of the two box filtered textures.
Since we are pre-filtering our sRGB image data (§2), we can either perform
all the pre-processing to the training database in advance, or we can apply the
pre-processing as each image is fetched from the database. After the pre-filtering
(Figure 6a) we re-arrange the texture channels so that each texture contains one
filtered component from each of the four training examples (Figure 6b). The
input textures are thus prepared for evaluating our features efficiently.
We then iterate through the supplied set of candidate features, computing
the response of the current training examples to each feature. For each feature
we select two input textures according to the channels specified in the feature
(Figure 6c). We compute each box filter convolution on four training images
in parallel by passing the input texture to a pixel shader that performs the
necessary look-ups on the integral image. In a third pass, we subtract the two
box filter responses to recover the feature response (Figure 6d).
We ensure that our leaf image (§4.2) also comprises four components that
correspond to the four current training examples.
4.4 Histogram Accumulation

The computed feature responses are then accumulated into the 4D histogram,
using also the values of the leaf image and ground truth label at each pixel.
Histogram accumulation is a scatter operation rather than a gather operation,
so it cannot be implemented in a pixel shader. Instead we use a vertex shader,
inspired by [23], to perform the scattering.
As our vertex buffer input we supply a list of all the 2D texture coordinates
which cover the area of the feature response image. Our vertex shader (Figure 7)
uses this input coordinate to read the feature response value, ground truth label
and leaf index at one position from three supplied input textures. The other
value required for the histogram operation is the feature index which is passed
in as a constant. The shader then computes the 4D-to-2D mapping according to
the active tiling scheme (§4.1). Finally the 2D histogram coordinate is output
by the vertex shader.
604 T. Sharp
float4 Scatter(uniform sampler2D Textures[3], uniform float4 select,

in out float2 coord : TEXCOORD0, uniform float feature)
{
float4 address = {coord.x, coord.y, 0.0f, 0.0f};
float leaf = dot(tex2Dlod(Textures[0], address), select);
float label = dot(tex2Dlod(Textures[1], address), select);
float response = dot(tex2Dlod(Textures[2], address), select);
float2 pos = Map4DTo2D(leaf, label, response, feature);
return float4(pos.x, pos.y, 0.0f, 1.0f);
}
Fig. 7. An HLSL vertex shader that scatters feature response values to the appropriate
position within a histogram
A simple pixel shader emits a constant value of 1 and, with additive blending
enabled, the histogram values are incremented as desired.
We execute this pipeline four times for the four channels of data to be ag-
gregated into the histogram. A shader constant allows us to select the required
channel.
4.5 Quantization and Thresholds
In order to histogram the real-valued feature responses they must first be quan-
tized. We require that the client provides the number of quantization bins to
use for the training round. An interval of interest for response values is also pro-
vided for each feature in the set of candidates. In our Scatter vertex shader, we
then linearly map the response interval to the histogram bins, clamping to end
bins. We make the quantization explicit in this way because different learning
algorithms may have different criteria for choosing the parameters used for the
tree’s Boolean test.
One approach would be to use 20-30 quantization levels during a training
round and then to analyze the histogram, choosing a threshold value adaptively
to optimize the resulting child distributions. For example, the threshold could
be chosen to minimize the number of misclassified data points or to maximize
the KL-divergence. Although this method reduces training error, it may lead to
over-fitting. Another approach would be to use a very coarse quantization (only
2 or 3 bins) with randomized response intervals. This method is less prone to
over-fitting but may require more training rounds to become sufficiently discrim-
inative.
We have tested both of the above approaches and found them effective. We
currently favour the latter approach, which we use with the ID3 algorithm [21]
to select for every leaf node the feature with the best information gain. Thus we
double the number of leaf nodes in a tree after each training round.
We create flexibility by not requiring any particular learning algorithm. In-
stead, by focusing on computation of the histogram, we enable clients to adopt
their preferred learning algorithm efficiently.
5 Results
Our test system consists of a dual-core Intel Core 2 Duo 2.66 GHz and an nVidia
GeForce GTX 280. (Timings on a GeForce 8800 Ultra were similar.) We have
coded our GPU version using Direct3D 9 with HLSL shaders, and a CPU version
using C++ for comparison only. We have not developed an SSE version of the
CPU implementation which we believe may improve the CPU results somewhat
(except when performance is limited by memory bandwidth). Part of the appeal
of the GPU implementation is the ability to write shaders using HLSL which
greatly simplifies the adoption of vector instructions.
In all cases identical output is attained using both CPU and GPU versions.
Our contribution is a method of GPU implementation that yields a considerable
speed improvement, thereby enabling new real-time recognition applications.
We separate our results into timings for pre-processing, evaluation and train-
ing. All of the timings depends on the choice of features; we show timings for our
generalized recognition features. For reference, we give timings for our feature
pre-processing in Figure 8.
5.1 Tree Training
Training time can be prohibitively long for randomized trees, particularly with large
databases. This leads to pragmatic short-cuts such as sub-sampling training data,
which in turn has an impact on the discrimination performance of learned trees.
Our training procedure requires time linear in the number of training exam-
ples, the number of trees, the depth of the trees and the number of candidate
features evaluated.
Description Resolution CPU GPU Speed-up

(pixels) (ms) (ms) (×)
320 × 240 94 8.5 11.1
Convolution with 16 separable filters (§2.3)
640 × 480 381 15.5 24.6
320 × 240 9.2 7.0 1.31
Integration of 16 data channels (§2.4)
640 × 480 31 16.6 1.87
Fig. 8. Timings for feature pre-processing
Operation CPU GPU Speed-up

(s) (%) (s) (%) (×)
100× Leaf image computation (§4.2) 6.0 6 0.2 2 30
104 × Feature responses (§4.3) 39.5 41 0.2 2 198
10 × Histogram accumulations (§4.4) 52.1 53
4
11.8 96 4.4
Total 97.6 100 12.2 100 8.0
Fig. 9. Breakdown of time spent during one training round with 100 training examples
and a pool of 100 candidate features. Note the high proportion of time spent updating
the histogram.
606 T. Sharp
To measure training time, we took 100 images from the labeled object recog-
nition database of [14] with a resolution of 320×213. This data set has 23 labeled
classes. We used a pool of 100 candidate features for each training round. The
time taken for each training round was 12.3 seconds. With these parameters, a
balanced tree containing 256 leaf nodes takes 98 seconds to train. Here we have
used every pixel of every training image.
Training time is dominated by the cost of evaluating a set of candidate features
on a training image and aggregating the feature responses into a histogram.
Figure 9 shows a breakdown of these costs. These figures are interesting as they
reveal two important insights:
First, the aggregation of the histograms on the GPU is comparatively slow,
dominating the training time significantly. We experimented with various dif-
ferent method for accumulating the histograms, maintaining the histogram in
system memory and performing the incrementation on the CPU. Unfortunately,
this did not substantially reduce the time required for a training round. Most
recently, we have begun to experiment with using CUDA [24] for this task and
we anticipate a significant benefit over using Direct3D.
Second, the computation of the rectangular sum feature responses is extremely
fast. We timed this operation as able to compute over 10 million rectangular sums
per ms on the GPU. This computation time is insignificant next to the other
timings, and this leads us to suggest that we could afford to experiment with
more arithmetically complex features without harming training time.
5.2 Forest Evaluation

Our main contribution is the fast and parallel evaluation of decision forests on
the GPU. Figure 10 show timings for the dense evaluation of a decision forest,
with various different parameters.
Our method maps naturally to the GPU, exploiting its parallelism and cache,
and this is reflected in the considerable speed increase over a CPU version by
around two orders of magnitude.
Resolution Output Mode Trees Classes CPU GPU Speed-up

(pixels) (ms) (ms) (×)
320 × 240 TreeLeaf 1 N/A 70.5 0.75 94
320 × 240 ForestLeaves 4 N/A 288 3.0 96.1
320 × 240 Distribution 8 4 619 5.69 109
320 × 240 ArgMax 8 23 828 6.85 121
640 × 480 TreeLeaf 1 N/A 288 2.94 97.8
640 × 480 ForestLeaves 4 N/A 1145 12.1 95.0
640 × 480 Distribution 8 4 2495 23.1 108
640 × 480 ArgMax 8 23 3331 25.9 129
Fig. 10. Timings for evaluating a forest of decision trees. Our GPU implementation
evaluates the forest in about 1% of the time required by the CPU implementation.
(a) (b) (c) (d)
Fig. 11. (a)-(b) Object class recognition. A forest of 8 trees was trained on labelled
data for grass, sky and background labels. (a) An outdoor image which is not part of
the training set for this example. (b) Using the Distribution output option, the blue
channel represents the probability of sky and the green channel the probability of grass
at each pixel (5 ms). (c)-(d) Head tracking in video. A random forest was trained
using spatial and temporal derivative features instead of the texton filter bank. (c) A
typical webcam video frame with an overlay showing the detected head position. This
frame was not part of the training set for this example. (d) The probability that each
pixel is in the foreground (5 ms).
5.3 Conclusion
We have shown how it is possible to use GPUs for the training and evaluation
of general purpose decision trees and forests, yielding speed gains of around 100
times.
References
1. Amit, Y., Geman, D.: Shape quantization and recognition with randomized trees.
Neural Computation 9(7), 1545–1588 (1997)
2. Breiman, L.: Random forests. ML Journal 45(1), 5–32 (2001)
3. Lepetit, V., Fua, P.: Keypoint recognition using randomized trees. IEEE Trans.
Pattern Anal. Mach. Intell. 28(9), 1465–1479 (2006)
4. Ozuysal, M., Fua, P., Lepetit, V.: Fast keypoint recognition in ten lines of code.
In: IEEE CVPR (2007)
5. Winn, J., Criminisi, A.: Object class recognition at a glance. In: IEEE CVPR,
video track (2006)
6. Shotton, J., Johnson, M., Cipolla, R.: Semantic texton forests for image catego-
rization and segmentation. In: IEEE CVPR, Anchorage (2008)
7. Yin, P., Criminisi, A., Winn, J.M., Essa, I.A.: Tree-based classifiers for bilayer
video segmentation. In: CVPR (2007)
8. Bosh, A., Zisserman, A., Munoz, X.: Image classification using random forests and
ferns. In: IEEE ICCV (2007)
9. Apostolof, N., Zisserman, A.: Who are you? - real-time person identification. In:
BMVC (2007)
10. Yang, R., Pollefeys, M.: Multi-resolution real-time stereo on commodity graphics
hardware. In: CVPR, vol. (1), pp. 211–220 (2003)
11. Brunton, A., Shu, C., Roth, G.: Belief propagation on the gpu for stereo vision. In:
CRV, p. 76 (2006)
12. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient belief propagation for early vision.
International Journal of Computer Vision 70(1), 41–54 (2006)
608 T. Sharp
13. Steinkraus, D., Buck, I., Simard, P.: Using gpus for machine learning algorithms.
In: Proceedings of Eighth International Conference on Document Analysis and
Recognition, 2005, 29 August-1 September 2005, vol. 2, pp. 1115–1120 (2005)
14. Winn, J.M., Criminisi, A., Minka, T.P.: Object categorization by learned universal
visual dictionary. In: ICCV, pp. 1800–1807 (2005)
15. Viola, P.A., Jones, M.J.: Robust real-time face detection. International Journal of
Computer Vision 57(2), 137–154 (2004)
16. Shotton, J., Winn, J.M., Rother, C., Criminisi, A.: TextonBoost: Joint appearance,
shape and context modeling for multi-class object recognition and segmentation.
In: Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3951, pp.
17. Deselaers, T., Criminisi, A., Winn, J.M., Agarwal, A.: Incorporating on-demand
stereo for real time recognition. In: CVPR (2007)
18. James, G., O’Rorke, J.: Real-time glow. In: GPU Gems: Programming Techniques,
Tips and Tricks for Real-Time Graphics, pp. 343–362. Addison-Wesley, Reading
(2004)
19. Blelloch, G.E.: Prefix sums and their applications. Technical Report CMU-CS-90-
190, School of Computer Science, Carnegie Mellon University (November 1990)
20. Hensley, J., Scheuermann, T., Coombe, G., Singh, M., Lastra, A.: Fast summed-
area table generation and its applications. Comput. Graph. Forum 24(3), 547–555
(2005)
21. Quinlan, J.R.: Induction of decision trees. Machine Learning 1(1), 81–106 (1986)
22. Quinlan, J.: C4.5: Programs for Machine Learning. Morgan Kaufmann, California
(1992)
23. Scheuermann, T., Hensley, J.: Efficient histogram generation using scattering on
gpus. In: SI3D, pp. 33–37 (2007)
24. http://www.nvidia.com/cuda
General Imaging Geometry
for Central Catadioptric Cameras
Peter Sturm1 and João P. Barreto2

1
INRIA Rhône-Alpes and Laboratoire Jean Kuntzmann, Grenoble, France
2
ISR/DEEC, University of Coimbra, 3030 Coimbra, Portugal
Abstract. Catadioptric cameras are a popular type of omnidirectional

imaging system. Their imaging and multi-view geometry has been exten-
sively studied; epipolar geometry for instance, is geometrically speaking,
well understood. However, the existence of a bilinear matching constraint
and an associated fundamental matrix, has so far only been shown for
the special case of para-catadioptric cameras (consisting of a paraboloidal
mirror and an orthographic camera). The main goal of this work is to
obtain such results for all central catadioptric cameras. Our main result
is to show the existence of a general 15 × 15 fundamental matrix. This is
based on and completed by a number of other results, e.g. the formulation
of general catadioptric projection matrices and plane homographies.
1 Introduction and Previous Work

The geometry of single and multiple images has been extensively studied in com-
puter vision and photogrammetry [1]. The picture is rather complete for perspec-
tive cameras and many results have been obtained for other camera models too,
e.g. catadioptric [2,3,4,5], fisheyes [6,7], pushbroom [8], x-slit [9], oblique [10,11],
non-central mosaics [12,11]; this list is not intended to be exhaustive.
Besides perspective cameras, the most studied case is probably that of cata-
dioptric ones. Baker and Nayar have shown which catadioptric devices have a sin-
gle effective viewpoint, i.e. are central cameras [13]. Among those, the most useful
ones are the para-catadioptric and the hyper-catadioptric models, using a mirror
of paraboloidal/hyperboloidal shape, coupled with an orthographic/perspective
camera. The epipolar geometry of these devices has been studied by Svoboda
and Pajdla [2] who showed the existence of epipolar conics. Geyer and Daniilidis
have shown the existence of a fundamental matrix for para-catadioptric cameras
[4,14]; Sturm has extended this to fundamental matrices and trifocal tensors for
mixtures of para-catadioptric and perspective images [5]. Barreto showed that
the framework can also be extended to cameras with lens distortion due to the
similarities between the para-catadioptric and division models [15,16]
However, no such results have so far been obtained for the general catadiop-
tric camera model, i.e. including hyper-catadioptric cameras. In this paper, we
present a number of novel results concerning (bi-) linear formulations for single
and two-view geometry, valid for all central catadioptric cameras. First, we show
that the projection of a 3D point can be modeled using a projection matrix of

610 P. Sturm and J.P. Barreto
size 6 × 10 and how this may be used for calibrating catadioptric cameras with
a straightforward DLT approach, something which has not been possible up to
now. We then give analogous results for the backprojection of image points and
the projection of quadrics and conics. These are the basis for our main result,
the general fundamental matrix for catadioptric cameras. It is of size 15× 15 and
an explicit compact expression is provided. Finally, we also show the existence
of plane homographies, again of size 15 × 15, that relate sets of matching image
points that are the projections of coplanar scene points.
Our results, like those cited above for para-catadioptric cameras, are based on
the use of so-called lifted coordinates to represent geometric objects. For example,
2D points are usually represented by 3-vectors of homogeneous coordinates; their
lifted coordinates are 6-vectors containing all degree-2 monomials of the original
coordinates. Lifted coordinates have also been used to model linear pushbroom
cameras [8] and to perform multi-body structure from motion [17,18].
Organization. We describe our results in two ways, geometrically and alge-

braically. In the next section, we immediately describe all main results in a purely
geometrical way, which we find rather intuitive and which hopefully guides the
reader through the more technical later sections. In section 3, we introduce nota-
tions and background, mostly associated to Veronese maps which are extensively
used in this work. In sections 4 to 8, we develop algebraic formulations for the
projection of 3D points, backprojection of image points, projection of quadrics
and conics, epipolar geometry, and plane homographies.
2 Geometrical Description of Our Results

We use the sphere based model for catadioptric projection introduced by Geyer
and Daniilidis [19]. All central catadioptric cameras can be modeled by a unit
sphere and a perspective camera, such that the projection of 3D points can
be performed as follows in two steps. First, one projects the point onto the
sphere, to an intersection of the sphere and the line joining its center and the
3D point. There are two such intersection points; it is usually assumed that
the only physically feasible one can be singled out. That point is then projected
into the perspective camera. This model covers all central catadioptric cameras,
the type of which is encoded by the distance between the perspective camera
and the center of the sphere, e.g. 0 for perspective, 1 for para-catadioptric,
0 < ξ < 1 for hyper-catadioptric. Note that even for the para-catadioptric case,
where the true camera is an affine one, the camera in the sphere based model is
still perspective. An algebraic formulation of the model is given in section 3. We
now give geometrical and intuitive descriptions of our main results.
Projection of a 3D point. In the first step of the projection as described

above, the intersections of the sphere with the line spanned by its center and
the 3D point, are computed. There are two mathematical solutions which when
projected to the perspective camera, give the two mathematical image points, cf.
figure 1. For most real catadioptric cameras, only one of them can be observed;
General Imaging Geometry for Central Catadioptric Cameras 611
perspective
camera
Fig. 1. Left: Projection of a 3D point to two image points (in cyan). Middle: Backpro-
jection of an image point to two 3D lines. Right: Illustration of plane homography. The
image point in cyan in the right camera is backprojected to the scene plane, giving the
two red points. Each one of them is projected to the unit sphere of the second camera
on two blue points. The four image points in the second camera are shown in cyan.
the second intersection point with the sphere is hidden from the 3D point by the
mirror. An exception is the case of an elliptical mirror where a 3D point may
actually be seen twice in the image. Although for the most useful catadioptric
cameras, a 3D point is in reality visible in only one image point, it turns out
that in order to obtain multi-linear expressions for epipolar geometry, plane
homographies etc., both mathematical image points have to be considered, see
later. So, from now on we consider two image points per 3D point and similar
considerations will be done for backprojection etc. in the following. The algebraic
formulation of the projection of 3D points, in the form of a 6 × 10 general
catadioptric projection matrix, is given in section 4.
Projection of a 3D line. It is well known that a central catadioptric image of a

3D line, is a conic [19,20]. In this work, we do not explicitly require the projection
of 3D lines, but we still mention this result since it is helpful to understand the
epipolar geometry, see below.
Backprojection of an image point. First, the image point is backprojected

relative to the perspective camera, giving rise to a 3D line. Then, its two in-
tersection points with the sphere are computed. Finally, the two lines spanned
by the sphere center and these intersection points, are generated. These are the
backprojection lines of the image point, cf. figure 1 (middle). In section 5 we
show how to represent the union of two 3D lines by a single algebraic object,
leading to the formulation of a 6 × 6 backprojection matrix.
Two images of a plane. Consider q1 , the image of a 3D point on a scene

plane Π, in the first image. What are the possible matching points in the second
image? These can be determined as follows. Let us first backproject q1 to 3D; as
we have seen just before, this gives two 3D lines. Their intersections with Π are
the two points on that plane that may be observed in q1 . Let us project both of
them into the second image. This gives a total of four points in the second image,
all of which are mathematically plausible matches of q1 , cf. figure 1 (right). In

section 8 we give an algebraic representation of a catadioptric plane homography,
that maps one image point onto an entity representing its four possible matches.
Epipolar geometry. The basic question of epipolar geometry is: what is the
locus of points in the second image, that may be matches of a point q1 in the
first image? The answer follows from the insights explained so far. Let us first
backproject q1 . The two 3D lines we get can then be projected into the second
image, giving one conic each. Hence, the locus of matching points is the union of
two conics. This can be represented by a single geometric entity, a quartic curve
(note of course that not every quartic curve is the union of two conics).
Hence, if a multi-linear fundamental matrix exists that represents this epipo-
lar geometry, it must map an image point into some representation of a quartic
curve. The equation of a planar quartic curve depends on 15 coefficients (defined
up to scale), one per 4-th order monomial of a 2D point’s homogeneous coordi-
nates. Hence, we may expect the fundamental matrix to be of size 15 × · · · . Like
for perspective images, we may expect that the transpose of the fundamental
matrix gives the fundamental matrix going from the second to the first image.
The fundamental matrix for catadioptric images should thus intuitively be of
size 15 × 15. This is indeed the case, as is shown in section 7.
3 Background
Notations. We do not distinguish between a projective transformation and the

matrix representing it. Matrices are represented by symbols in sans serif font, e.g.
M and vectors by bold symbols, e.g. Q. Equality of matrices or vectors up to a
scalar factor is written as ∼. [a]× denotes the skew-symmetric matrix associated
with the cross product of a 3-vector a.
Camera model. As mentioned before, we use the sphere based model [19].
Without loss of generality, let the unit sphere be located at the origin and the
T
optical center of the perspective camera, at the point Cp = (0,0, −ξ) . The
perspective camera is modeled by the projection matrix P ∼ Ap Rp I −Cp . For
full generality, we include a rotation Rp ; this may encode an actual rotation
of the true camera looking at the mirror, but may also simply be a projective
change of coordinates in the image plane, like for para-catadioptric cameras,
where the true camera’s rotation is fixed, modulo rotation about the mirror axis.
Note that all parameters of the perspective camera, i.e. both its intrinsic and
extrinsic parameter sets, are intrinsic parameters for the catadioptric camera.
Hence, we replace Ap Rp by a generic projective transformation K from now on.
The intrinsic parameters of the catadioptric camera are thus ξ and K.
The projection of a 3D point Q goes as follows (cf. section 2). The two
intersection points of the sphere and the line joining its center and Q, are
. T
Q1 , Q2 , Q3 , ± Q21 + Q22 + Q23 . Their images in the perspective camera are
⎛ ⎞
Q1
q± ∼ Kr± ∼ K ⎝ . Q2
⎠
Q3 ± ξ Q1 + Q2 + Q3
2 2 2
In the following, we usually first work with the intermediate image points r± ∼
K−1 q± , before giving final results for the actual image points q± .
Plücker line coordinates. 3D lines may be represented by 6-vectors of so-
called Plücker coordinates. Let A and B be the non-homogeneous coordinates
of two generic 3D points. Let us define the line’s Plücker coordinates as the

T
6-vector L = AT − BT , (A × B)T .
All lines satisfy the Plücker constraint LT WL = 0 where W is

0I
W=
I0
Two lines L and L cut one another if and only if LT WL = 0. Consider a rigid
transformation for points
R t
0T 1
Lines are mapped accordingly using the transformation

R 03×3
T=
[t]× R R
Second order line complexes. A second order line complex is a set of 3D

lines that satisfy a quadratic equation in the Plücker coordinates [21]. It can be
represented by a symmetric 6 × 6 matrix C such that exactly the lines on the
complex satisfy LT CL = 0. Note that C is only defined up to adding multiples
of W. Henceforth we call second order line complexes shortly line complexes. In
this paper, we use line complexes to represent the union of two 3D lines. Rigid
displacements of line complexes are carried out as

−T −1 −1 RT 03×3
T CT with T = (1)
−RT [t]× RT
Lifted coordinates from symmetric matrix equations. The derivation

of (multi-) linear relations for catadioptric imagery requires the use of lifted
coordinates. The Veronese map Vn,d of degree d maps of P into points
points
n
n+d
of an m dimensional projective space P m , with m = − 1.
d
Consider the second order Veronese map V2,2 , that embeds the projective
plane into the 5D projective space, by lifting the coordinates of point q to
T
q̂ = q12 q1 q2 q22 q1 q3 q2 q3 q32
Vector q̂ and matrix qqT are composed by the same elements. The former can
be derived from the latter through a suitable re-arrangement of parameters.
Define v(U) as the vector obtained by stacking the columns of a generic matrix
U [22]. For the case of qqT , v(qqT ) has several repeated elements because of
matrix symmetry. By left multiplication with a suitable permutation matrix S
that adds the repeated elements, it follows that
⎛1 0 0 0 0 0 0 0 0⎞
0 1010000 0
q̂ = D−1 ⎝ 00 00 01 00 10 00 01 00 00 ⎠ v(qqT ), (2)
0 0000101 0
1 0000000 1

S
9
with D a diagonal matrix, Dii = j=1 Sij . This process of computing the lifted
representation of a point q can be extended to any second order Veronese map
Vn,2 independently of the dimensionality of the original space. It is also a mech-
anism that provides a compact representation for square symmetric matrices.
If U is symmetric, then it is uniquely represented by vsym (U), the column-wise
vectorization of its upper right triangular part:
T
vsym (U) = D−1 SU = (U11 , U12 , U22 , U13 , · · · , Unn )
Let us now discuss the lifting of linear transformations. Consider A such that
r = Aq. The relation rrT = A(qqT )AT can be written as a vector mapping
(rrT ) = (A ⊗ A)(qqT ),
with ⊗ denoting the Kronecker product [22]. Using the symmetric vectorization,
we have q̂ = vsym (qqT ) and r̂ = vsym (rrT ), thus:
r̂ = D−1 S(A ⊗ A)ST v̂

Â
We have just derived the expression for lifting linear transformations. A has a
lifted counterpart Â such that r = A q iff r̂ = Â q̂. For the case of a second
order Veronese map, the lifting of a 2D projective transformation A is Â of size
6 × 6. This lifting generalizes to any projective transformation, independently
of the dimensions of its original and target spaces, i.e. it is also applicable to
rectangular matrices. We summarize a few useful properties [22].
= = ÂB̂
AB >
A−1 = Â−1 =T = D−1 ÂT D
A (3)
Also, for symmetric matrices U and M, we have the following property:
U = AMAT ⇒ vsym (U) = Â vsym (M) (4)
Also, note that Â is non-singular iff A is non-singular. In this paper, we use the

following liftings: 3-vectors q to 6-vectors q̂, 4-vectors Q to 10-vectors Q̂, and
6-vectors u to 21-vectors û. Analogously, 3 × 3 matrices W are lifted to 6 × 6 ones
Ŵ and 3 × 4 matrices to 6 × 10 ones. We also use the fourth order Veronese map
ˆ containing the quartic monomials of
of P 2 , mapping 3-vectors q to 15-vectors q̂
q. We call this double lifting; it applies analogously to 3 × 3 matrices W, which

ˆ
are doubly lifted to 15 × 15 matrices Ŵ. Finally, note that applying two second
order Veronese maps in succession, is not equivalent to applying one fourth
order Veronese map: for a 3-vector q, q̂ ˆ is a 15-vector, whereas ŵ where w is
the 6-vector w = q̂, is of length 21. We thus denote successive application of two
second order liftings by ˚, e.g. for a 3 × 3 matrix E, we get a 21 × 21 matrix E̊.
4 Projection of 3D Points
As explained in section 2, a 3D point is mathematically projected to two image
points. How to represent two 2D points via a single geometric entity? One way
is to compute the degenerate dual conic generated by them, i.e. the dual conic
containing exactly the lines going through at least one of the two points. Let the
two image points be q+ and q− (see section 3). The dual conic is given by
⎛ 2 ⎞
Q1 Q1 Q2 Q1 Q3
Ω ∼ q+ qT T ⎝
− + q− q+ ∼ K Q1 Q2 Q2
2
Q2 Q3 ⎠ KT
Q1 Q3 Q2 Q3 Q3 − ξ (Q1 + Q2 + Q3 )
2 2 2 2 2
This can be written as a linear mapping of the 3D point’s lifted coordinates,

onto the vectorized matrix of the conic:
⎛ ⎞
1 0 0 00 0
⎜ 0 1 0 00 0 ⎟
⎜ ⎟
⎜ ⎟
< 6×6 ⎜ 0 0 1 0 0 0 ⎟ I6 06×4 Q̂
vsym (Ω) ∼ K ⎜ 0 0 0 10 0 ⎟
⎜ ⎟
⎝ 0 0 0 01 0 ⎠
−ξ 2 0 −ξ 2 0 0 1 − ξ 2

Xξ
So far, we have projected a 3D point given in the catadioptric camera’s local

coordinate system.
If we introduce extrinsic parameters of the camera, i.e. a pose
matrix T = R I −t then we can write the projection operation as

< 6×6 Xξ,6×6 R
vsym (Ω) ∼ K < 6×6 I6 T
4×6 Q̂10
where T depends only on t.
We have thus derived a 6 × 10 catadioptric projection matrix Pcata . Fur-
thermore, it can be decomposed, like the projection matrix of a perspective
camera, into two matrices containing either intrinsic or extrinsic parameters:

< ξR
Pcata = KX < 6×6 I6 T4×6

Acata Tcata
We find here a 6 × 6 catadioptric calibration matrix Acata . Note that the

restriction of the projection to points lying in a plane in 3D, leads straightfor-
wardly to a 6 × 6 homography, analogous to the plane-to-image homographies
used e.g. for perspective camera calibration.
Camera Calibration. Let us consider how to set up equations for calibrating

a camera. In the perspective case, a 2D–3D point correspondence allows to write
q ∼ PQ. One way to set up linear equations on P is to write [q]× PQ = 03 .
What is the analogous expression in the catadioptric case? Let again q and Q
be a 2D–3D point correspondence. Since each 3D point is projected to two 2D
points, one may not directly be able to compare q to the image of Q, unlike in
the perspective case. Instead, as mentioned above, the projection matrix maps Q
(rather, its lifted version) onto the coefficients of a degenerate dual conic Ω. The
point q must be one of the two generators of Ω. This implies that all lines through
q must lie on Ω. Hence: ∀p : p × q ∈ Ω, which gives ∀p : pT [q]× Ω[q]× p = 0.
Thus, [q]× Ω[q]× = 03×3 . This gives 6 constraints that can be written as

>×
[q] >× Pcata Q̂ = 06
vsym (Ω) = [q]
6×6
We thus find an expression that is very similar to that for perspective cameras
and that may be directly used for calibrating catadioptric cameras using e.g. a
standard DLT like approach. While a 3 × 3 skew symmetric matrix has rank
2, its lifted counterpart is rank 3. Therefore, each 3D-to-2D match provides 3
linear constraints on the 59 parameters of Pcata , and DLT calibration can be
done with a minimum of 20 matches.
5 Backprojection of Image Points
This is essential for deriving the proposed expression of the fundamental matrix.
Similarly to the case of projection, we want to express the backprojection func-
tion of a catadioptric camera as a linear mapping. Recall from section 2, that
the backprojection of an image point gives two 3D lines. How to represent two
3D lines via a single geometric entity? Several possibilities may exist; the one
that seems appropriate is to use a second order line complex: consider two 3D
lines L+ and L− . All lines that cut at least one of them, form a second order
line complex, represented by a 6 × 6 matrix C such that lines on the complex
satisfy LT CL = 0. The matrix C is given as (with W as defined in section 3)

C ∼ W L+ LT T
− + L− L+ W
The backprojection lines L± of an image point q are spanned by the origin

T T T T
(center of the sphere) and points (bT
± , 1) , thus L± ∼ b± 0 . Here,

T T
b± = (r r)Cp + ξr3 ± ξ r3 − (r r)(ξ − 1) r
2 2 2
with r ∼ K−1 q and Cp the center of the perspective camera (cf. section 3). The
line complex C generated by the two lines, is

0 0 0 0
C∼ ∼
0 b+ bT T
− + b− b+ 0 ξ 2 (rT r)e3 eT T T
3 − ξ r3 (e3 r + re3 ) + (ξ − 1)rr
2 2 T
where e3 = (0, 0, 1)T . C is by construction symmetric and of rank 2 and it has

9 non-zero coefficients. Let M be the lower right 3 × 3 submatrix of C and the
6-vector m its vectorized version: m = vsym (M). We have the following linear
backprojection equation:
m ∼ Bξ r̂ = Bξ K̂−1 q̂ (5)
with ⎛ 2 ⎞
ξ −1 0 0 0 0 0
⎜ 0 ξ2 − 1 0 0 0 0⎟
⎜ ⎟
⎜ 0 0 ξ −1 0 0 0 ⎟
2
Bξ = ⎜
⎜ 0
⎟
⎜ 0 0 −1 0 0 ⎟
⎟
⎝ 0 0 0 0 −1 0 ⎠
ξ2 0 ξ2 0 0 −1
We call Bcata = Bξ K̂−1 the backprojection matrix.
6 Projection of Line Complexes and Quadrics

To the best of our knowledge the projection of general quadric surfaces in cata-
dioptric cameras has never been studied. The existing literature concerns only
the projection of spheres, for calibration purposes [23]. The problem can be con-
veniently addressed by considering a line-based representation of quadrics, via
line complexes. The set of 3D lines tangent to a quadric form a line complex [21].
A conic on a 3D scene plane can also be represented by a line complex [24,25],
and the following results apply thus to the projection of both quadrics and con-
ics. As discussed in the previous section, a line complex can be represented by a
6 × 6 symmetric matrix C. Let C be split in 3 × 3 blocks:

U N
C∼
NT M
The image of C consists of all points q such that at least one of their backpro-
T T
jection rays lies on C. Let L± ∼ (bT± 0 ) be the Plücker coordinates of the two
backprojections of q (cf. section 5). Hence, q lies on the image of C iff
(LT T T T
+ CL+ )(L− CL− ) = (b+ Ub+ )(b− Ub− ) = 0
By replacing b± with its definition (section 5) and developing this equation, we

get the following constraint on the doubly lifted coordinates of q:
ˆ −T
ˆ T K̂
q̂ X û = 0
lc
Plc
where u = vsym (U), û is a 21-vector and the 15 × 21 matrix Xlc only depends
on the intrinsic parameter ξ and is highly sparse (not shown due to lack of
space). Since the coefficients of q̂ˆ are 4th order monomials of q, we conclude
that a central catadioptric image of any line complex, and thus of any quadric
or conic, is a quartic curve. We may call the 15×21 matrix Plc the line complex
projection matrix for catadioptric cameras. It maps the lifted coefficients of
the line complex to the 15 coefficients of the quartic curve in the image.
7 The General Catadioptric Fundamental Matrix

We are now ready to derive an analytical expression for the fundamental ma-
trix for any pair of catadioptric images. As suggested in section 2, we perform
the following steps: (i) Backproject a point q from the first image to its two
backprojection rays, represented by a line complex. (ii) Map the line complex
from the coordinate system of the first camera, to that of the second one, via
a rigid transformation (rotation and translation). (iii) Project the transformed
line complex into the second camera.
We already know from the previous sections that the result has to be a quartic
epipolar curve since it is the image of a line complex. In our case, the line complex
is degenerate (the “envelope” of just two lines – the backprojection rays – not a
full quadric). Hence, and as described in section 2, the quartic epipolar curve is
indeed the union of two conics.
Let us now derive the full expression of the catadioptric fundamental matrix.
The only remaining missing piece is that the backprojection of an image point
(step (i)) gives the coefficients of a line complex, but that the projection of the
line complex (step (iii)) requires its lifted coefficients. Hence, we need to insert
that lifting between steps (ii) and (iii).
Recall from section 5 that the backprojection line complex is obtained as:

0 0
C∼ with vsym (M) = Bcata q̂
0M
The rigid transformation of step (ii) gives (cf. equation (1) in section 3)
T T

−T −1 [t]× RMRT [t]× −[t]× RMRT
C ∼ T CT ∼
−RMRT[t]× RMRT
Recall from section 6 that the projection of a line complex, when expressed in
the local camera coordinate system, only involves its upper left 3 × 3 submatrix:
T
U = [t]× RMRT [t]× = EMET
where we encounter the well-known essential matrix E = [t]× R.
Since U and M are symmetric, property (4) from section 3 allows to write the
folllwing relation between their vectorized versions u = vsym (U), m = vsym (M):
u = Ê6×6 m
Finally, the required lifted coefficients of the line complex are obtained as:
û = E̊21×21 m̂
The last remaining detail is to express m̂ in terms of the image point q. From
m = Bcata q̂, we deduce
ˆ
m̂ = B̂cata,21×15 q̂
We can now introduce the catadioptric fundamental matrix:
Fcata,15×15 ∼ Plc E̊ B̂cata (6)
ˆ gives a quartic epipolar curve. The

We have already explained that Fcata q̂
epipolar constraint can thus be written as
ˆ T Fcata q̂
q̂ ˆ1 = 0
2
which has the familiar form known for perspective cameras. Fcata has rank 6 and
its left/right null space has dimension 9. While a perspective view has a single
epipole, in an omnidirectional view there are a pair of epipoles, e+ and e− ,
corresponding to the two antipodal intersections of the baseline with the sphere,
cf. section 4. The nullspace of Fcata comprises the doubly lifted coordinates
vectors of both epipoles. We conjecture that they are the only doubly lifted
3-vectors in the nullspace of Fcata , but this has to be proven in future work.
We have no space to discuss it, but for mixtures involving one hyper-catadiop-
tric and one other camera, the size of Fcata is smaller (15 × 6 for para–hyper and
6 × 6 for hyper–perspective). Other special cases are already known [14,5].
8 The General Catadioptric Plane Homography

We give an algebraic formulation of the different steps involved in the plane
homography operation, as described in section 2. This section omits many details
T
due to lack of space. Let Π = (nT , d) be the plane and q1 a point in the first
image. We start by backprojecting the point to the plane, giving two 3D points

dI3
Q± ∼ b±
−nT

Y4×3
with b± as in section 5. We project them to the second image using the projection
matrix given in section 4. To do so, we first have to lift the coordinates of these
3D points: Q̂± ∼ Ŷ10×6 b̂± . The projection then gives two dual conics in the
second image (cf. section 4), represented by 6-vectors ω± ∼ Pcata Ŷ b̂± .
Let us compute the following symmetric 6 × 6 matrix:

T T
Γ ∼ ω+ ω− + ω− ω+ ∼ Pcata Ŷ b̂+ b̂T
− + b̂− b̂T T T
+ Ŷ Pcata (7)

Z
What does Γ represent? It is the “union” of two degenerate dual conics, ω+

and ω− . Hence, it represents a dual quartic curve (this can also be proven more
formally). Further, each of the two dual conics represents two image points; it
contains exactly the lines going through at least one of them. Hence, Γ contains
exactly the lines going through at least one of the total of four considered image
points. These four points are nothing else than the four points explained in
section 2, i.e. the possible four matches of q1 .
The expression of the dual quartic curve represented by Γ can be written as
ˆT ˆ
l̂ γ = 0, where l̂ denotes the fourth order Veronese map of a generic line l, and
γ is a 15-vector containing sums of the 21 coefficients of Γ (this is analogous to

the reduction explained in equation (2)).
When developing the expression for Z in (7), it can be seen that its coefficients
ˆ 1 . The coefficients of γ may thus be computed via a linear mapping
are linear in q̂
ˆ
of q̂1 ; that mapping is given by a 15 × 15 matrix:
ˆ1
γ ∼ Hcata,15×15 q̂
The matrix Hcata is the catadioptric plane homography. Its explicit form is
omitted due to lack of space.
By the same approach as in section 4, we can derive the following constraint
equation:

ˆ
[q2 ]× Hcata q̂1 = 015
Of the 15 constraints contained in the above equation, only five are linearly
independent. Hence, in order to estimate the 152 = 225 coefficients of Hcata , we
need at least 45 matches.
In the special case of para-catadioptric cameras, the homography is of size 6×6
and each match gives 6 equations, 3 of which are linearly independent. Hence,
12 matches are needed to estimate the 36 coefficients of that plane homography.
9 Conclusions, Discussion, and Perspectives

Our motivation for this work is to get a complete picture of the imaging and
multi-view geometry of catadioptric (and other) cameras. We have shown that
the basic concepts – projection, backprojection, epipolar geometry, or plane ho-
mography – can all be written as (multi-) linear mappings. These results are first
of all of conceptual value, and we consider them as theoretical contributions.
Concerning potential practical applications, we note that a linear estimation
of the catadioptric fundamental matrix requires 224 matches. . . We thus do not
currently believe that Fcata will be of practical use.
However, the catadioptric projection matrix and the plane-to-image homog-
raphy described in section 4 may indeed prove useful for calibrating catadioptric
and possibly other omnidirectional cameras. We show this by an illustrative
experiment, cf. Figure 2. Corner extraction for calibration grids, despite being
trivial for perspective cameras, is still problematic for images with strong non-
linear distortions [26]. In the perspective case we typically indicate the area of
interest by manually clicking 4 corners; they enable the estimation of an ho-
mography and the projection of the grid into the image. The final position of
all corners is accurately determined by refining the initial estimate using im-
age processing techniques. Such a procedure has not been possible until now for
non-conventional imagery with non-perspective distortions.
From section 4 it follows that the homography from a plane to any catadioptric
image is represented by a 6 × 6 matrix H. We estimated it from the required mini-
mum of 12 manually selected matches by the DLT procedure suggested in section
4 (left side of Fig. 2). All corners were then projected into the image and refined
Fig. 2. Estimation of the homography mapping a planar grid into a catadioptric image.
It was determined from 12 clicked points (left side). Each corner is mapped into a pair
of image projections. The lines joining corresponding pairs form a pencil going through
the principal point which confirms the correctness of the estimation.
using a corner detector. From this initial step only 7 out of 91 points were missed.
The procedure was repeated a second time using all the good points and 6 more
points were correctly detected. The estimated homography maps each plane point
to a pair of antipodal image points (right side of Fig. 2). The shown result suggests
that the plane-to-image homography can be well estimated and that it is useful
for extracting and matching corners of a planar grid. Current work deals with cal-
ibrating the camera from such homographies, from multiple images of the grid.
There are several perspectives for our work. The shown results can be spe-
cialized to e.g. para-catadioptric cameras, leading to simpler expressions. It may
also be possible that due to the coordinate liftings used, some of the results hold
not only for catadioptric cameras, but also for other models, e.g. classical radial
distortion; this will be investigated. Current work is concerned with developing
practical DLT like calibration approaches for catadioptric cameras, using 3D or
planar calibration grids. Promising results have already been obtained.
Acknowledgements. João P. Barreto is grateful to the Portuguese Science

Foundation by generous funding through grant PTDC/EEA-ACR/68887/2006.
Peter Sturm acknowledges support by the French ANR project CAVIAR.
References
1. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision, 2nd edn.
Cambridge University Press, Cambridge (2004)
2. Svoboda, T., Pajdla, T.: Epipolar geometry for central catadioptric cameras.
IJCV 49, 23–37 (2002)
3. Kang, S.: Catadioptric self-calibration. In: CVPR, pp. 1201–1207 (2000)
4. Geyer, C., Daniilidis, K.: Structure and motion from uncalibrated catadioptric
views. In: CVPR, pp. 279–286 (2001)
5. Sturm, P.: Mixing catadioptric and perspective cameras. In: OMNIVIS, pp. 37–44
(2002)
6. Mičušı́k, B., Pajdla, T.: Structure from motion with wide circular field of view
cameras. PAMI 28(7), 1135–1149 (2006)
7. Claus, D., Fitzgibbon, A.: A rational function for fish-eye lens distortion. In: CVPR,
pp. 213–219 (2005)
8. Gupta, R., Hartley, R.: Linear pushbroom cameras. PAMI 19, 963–975 (1997)
9. Feldman, D., Pajdla, T., Weinshall, D.: On the epipolar geometry of the crossed-
slits projection. In: ICCV, pp. 988–995 (2003)
10. Pajdla, T.: Stereo with oblique cameras. IJCV 47, 161–170 (2002)
11. Seitz, S., Kim, J.: The space of all stereo images. IJCV 48, 21–38 (2002)
12. Menem, M., Pajdla, T.: Constraints on perspective images and circular panoramas.
In: BMVC (2004)
13. Baker, S., Nayar, S.: A theory of catadioptric image formation. In: ICCV, pp. 35–42
(1998)
14. Geyer, C., Daniilidis, K.: Properties of the catadioptric fundamental matrix. In:
Tistarelli, M., Bigun, J., Jain, A.K. (eds.) ECCV 2002. LNCS, vol. 2359, pp. 140–
15. Barreto, J.: A unifying geometric representation for central projection systems.
CVIU 103 (2006)
16. Barreto, J.P., Daniilidis, K.: Epipolar geometry of central projection systems using
veronese maps. In: CVPR, pp. 1258–1265 (2006)
17. Wolf, L., Shashua, A.: Two-body segmentation from two perspective views. In:
CVPR, pp. 263–270 (2001)
18. Vidal, R., Ma, Y., Soatto, S., Sastry, S.: Two-view multibody structure from mo-
tion. IJCV 68, 7–25 (2006)
19. Geyer, C., Daniilidis, K.: A unifying theory for central panoramic systems. In: Ver-
non, D. (ed.) ECCV 2000. LNCS, vol. 1843, pp. 445–461. Springer, Heidelberg (2000)
20. Barreto, J., Araujo, H.: Geometric properties of central catadioptric line images
and its application in calibration. PAMI 27, 1237–1333 (2005)
21. Semple, J., Kneebone, G.: Algebraic Projective Geometry. Claredon Press (1998)
22. Horn, R., Johnson, C.: Topics in Matrix Analysis. Cambridge University Press,
Cambridge (1991)
23. Ying, X., Zha, H.: Using sphere images for calibrating fisheye cameras under the
unified imaging model of the central catadioptric and fisheye cameras. In: ICPR,
pp. 539–542 (2006)
24. Ponce, J., McHenry, K., Papadopoulo, T., Teillaud, M., Triggs, B.: On the absolute
quadratic complex and its application to autocalibration. In: CVPR, pp. 780–787
(2005)
25. Valdés, A., Ronda, J., Gallego, G.: The absolute line quadric and camera autocal-
ibration. IJCV 66, 283–303 (2006)
26. Mei, C., Rives, P.: Single view point omnidirectional camera calibration from planar
grids. In: ICRA, pp. 3945–3950 (2007)
Estimating Radiometric Response Functions
from Image Noise Variance
Jun Takamatsu1,2 , Yasuyuki Matsushita3 , and Katsushi Ikeuchi1,4,

1
Microsoft Institute for Japanese Academic Research Collaboration (MS-IJARC)
2
Nara Institute of Science and Technology
j-taka@is.naist.jp
3
Microsoft Research Asia
yasumat@microsoft.com
4
Institute of Industrial Science, the University of Tokyo
ki@cvl.iis.u-tokyo.ac.jp
Abstract. We propose a method for estimating radiometric response

functions from observation of image noise variance, not profile of its
distribution. The relationship between radiance intensity and noise vari-
ance is affine, but due to the non-linearity of response functions, this
affinity is not maintained in the observation domain. We use the non-
affinity relationship between the observed intensity and noise variance to
estimate radiometric response functions. In addition, we theoretically de-
rive how the response function alters the intensity-variance relationship.
Since our method uses noise variance as input, it is fundamentally robust
against noise. Unlike prior approaches, our method does not require im-
ages taken with different and known exposures. Real-world experiments
demonstrate the effectiveness of our method.
1 Introduction
Many computer vision algorithms rely on the assumption that image intensity is
linearly related to scene radiance recorded at the camera sensor. However, this
assumption does not hold with most cameras; the linearity is not maintained in
the actual observation due to non-linear camera response functions. Linearization
of observed image intensity is important for many vision algorithms to work,
therefore the estimation of the response functions is needed.
Scene radiance intensity (input) I and observed image intensity (output) O
are related by the response function f as O = f (I). Assuming it is contin-
uous and monotonic, the response function can be inverted to obtain the in-
verse response function g (= f −1 ), and measured image intensities can be lin-
earized using I = g(O). Since only observed output intensities O are usually
available, most estimation methods attempt to estimate the inverse response
functions.

Institute of Industrial Science, University of Tokyo.

624 J. Takamatsu, Y. Matsushita, and K. Ikeuchi
1.1 Prior Work
Radiometric calibration methods assume known characteristics of radiance at

a camera to estimate unknown response functions. One class of these meth-
ods uses information about the ratio of input radiance intensities. The Mac-
beth color checker is used for estimating the response functions by several
research groups such as [1]. Nayar and Mitsunaga [2] use an optical filter
with spatially varying transmittance; the variation corresponds to the ratio.
Instead of using such special equipment, some methods use a set of input im-
ages taken with different exposures from a fixed viewpoint so that the ratio be-
comes available. These methods are divided into parametric and non-parametric
approaches.
In the parametric framework, Mann and Picard [3] propose a method that
assumes the response functions can be approximated by gamma correction func-
tions. Mitsunaga and Nayar use a polynomial function for the representation [4].
Grossberg and Nayar apply principal component analysis (PCA) to a database of
real-world response functions (DoRF) and show that the space of response func-
tions can be represented by a small number of basis functions [5]. Mann [6] uses
the comparametric equation that defines the relationship between two images
taken with different exposure times.
Debevec and Malik estimate response functions with a non-parametric repre-
sentation using a smoothness constraint [7]. Tsin et al . estimate non-parametric
forms of response functions using the statistical analysis [8]. Pal et al . propose
to use a Bayesian network consisting of probabilistic imaging models and prior
models of response functions [9]. While non-parametric approaches have greater
descriptive power, the large number of unknown variables often lead to compu-
tationally expensive or unstable algorithms.
A few authors have proposed more general estimation methods that allow
camera motion or scene motion between input images. Grossberg and Nayar [10]
estimate the function from images with scene motion using brightness his-
tograms, which avoid difficult pixelwise image registration. Kim and Polle-
feys [11] propose a method that allows free camera motion by finding the
correspondence among images. These methods can handle more general cases,
but still require images taken with multiple exposures.
Another class of the estimation methods is based on the physics of the imag-
ing process. Lin et al . estimate the function from the color blending around
edges in a single image [12]. For grayscale images, a 1D analogue of the 3D color
method is presented in [13]. Wilburn et al . propose to use the temporal blend-
ing of irradiance [14]. Matsushita and Lin propose an estimation method from
profiles of image noise distributions [15]; their method relies on the assumption
that profiles of noise distributions are symmetric. More recently, Takamatsu et
al . [16] propose to use a probabilistic intensity similarity measure [17] for the
estimation. Unlike these approaches, our method only uses noise variance, which
is considered a lower level of information.
Estimating Radiometric Response Functions from Image Noise Variance 625
Fig. 1. The noise variance in the input domain has an affine relationship with input
intensity level. Due to the non-linearity of the response function, the affine relationship
is lost in the output domain. The proposed method estimates the response functions
by recovering the affinity of the measured noise variances.
Response f Response f Response f

Output
Output
Output
Input Input Input

: Noise variance in input : Noise variance in output
Fig. 2. The relationship between response function and noise variances in input and
output domains. The magnitude of output noise variance (height of the filled region)
varies with the slope of the response function with a fixed input noise variance.
1.2 Proposed Method
In this paper, we propose a method for estimating camera response functions

from observation of image noise variance. Figure 1 illustrates the overview of
the proposed method. As shown in the figure, noise variance in the input do-
main has an affine relationship with input intensity. This relationship, however,
is not maintained in the output domain due to the non-linearity of camera re-
sponse functions. By estimating a function that projects the non-affine intensity-
variance relationship to an affine one, our method derives the response functions.
Figure 2 depicts the relationship between the response function and noise vari-
ances in input and output domains. The ratio of the input/output noise variance
is closely related to the slope of the response function.
Unlike Matsushita and Lin’s work [15], our method does not rely on profiles of
noise distributions which are more difficult to obtain in practice. The proposed
method has the following two advantages over their method. First, it uses the
more reasonable assumption; the affinity of intensity-variance relationship in the
input domain. We derive this relationship from the nature of photon (electron)
noise [18,19]. Second, the proposed method uses only noise variances rather than
noise profiles, therefore it only requires less amount of information about noise.
Practically, it makes the procedure of noise measurement easier.
This paper has two major contributions. It first theoretically derives how the
radiometric response function alters the intensity-variance relationship, and vice
versa. While it has been pointed out that these two quantities are related [8,20],
the exact alternation has not been explicitly described. Second, it introduces
a method that has a wide range of applicability even with noisy observations.
While many existing algorithms break down when the noise level is high, we
show that the proposed method is not sensitive to the noise level because it uses
the noise as information.
2 Noise Variance and Response Functions

This section provides a theoretical relationship between the response function
and noise variances in the input and output domains. We show that the levels
of noise variance in the input and output domains are related by the square of
the first derivative of the camera response function. This result is later used to
develop the estimation algorithm in Section 3.
2.1 Derivation of the Relationship between Noise Variance and

Response Functions
We treat the relationship between input or output intensity and noise variance as
the function similar to Liu et al . [20]. The noise variance function σO
2
(Õ) in the
output domain, where the noise-free output intensity is Õ, can be described as

σO
2
(Õ) = (O − μO )2 p(O|Õ)dO = (f (I) − μO )2 p(I|I)dI. ˜
The conditional density function (cdf) p(O|Õ) represents the noise distribution
in the output domain, i.e., the probability that the output intensity becomes O
when the noise-free intensity is Õ. Likewise, the cdf p(I|I) ˜ represents the noise
distribution in the input domain when the noise-free input intensity is I. ˜ The
function f is the response function, and Õ and I are related by f as Õ = f (I).
˜ ˜
μO (= μO (Õ)) is the expectation of the output intensity with the cdf p(O|Õ).
Using the Taylor series of f (I) around I˜ and assuming that the second- and
higher-degree terms of the series are negligible (discussed in Section 2.3), we
obtain
σO
2
(Õ) f 2 (I)σ
˜ 2 (I),
I
˜ (1)
where σI2 (I)
˜ is the function of noise variance in the input domain when the
noise-free input intensity is I.
˜ This equation shows the relationship between the
noise variance and the response functions. This relationship has been pointed out
in [8,20], however, the correctness of the equation was not thoroughly discussed.
To simplify the notation in derivation of Eq. (1), we define:

μI = μI (I) = Ip(I|I)dI,
˜ ˜ Id = μI − I,
˜

Mn = Mn (I) ˜ = (I − μI )n p(I|I)dI ˜ (n ≥ 0). (2)
μI is the expectation, or mean, of input intensity I with the cdf p(I|I).

˜ Id denotes
the difference between the expectation μI and the noise-free intensity I. ˜ Mn is
the n-th moment about the mean. Note that we do not put any assumptions
about the profile and model of the noise distributions.
The expectation of the output intensity μO where the noise-free output in-
tensity is Õ can be written as
∞
f (j) (I)
˜
μO = μO (Õ) = f (I)p(I|I)dI
˜ = Õ + Nj , (3)
j=1
j!
where Nj is the j-th moment about I˜ defined as follows.

j
j j−k
Nj = (I − I)
˜ j p(I|I)dI
˜ = ((I − μI ) + Id ) p(I|I)dI
j ˜ = I Mk . (4)
k d
k=0
Note that the response function f is represented using its Taylor series.
From the definition, the noise variance function σO 2
(Õ) where the noise-free
output intensity is Õ can be derived as
∞
9 :2 ∞ ∞
f (j) (I)
˜ ˜ f (k) (I)
f (j) (I) ˜
σO (Õ) =
2
Lj,j + 2 Lj,k , (5)
j=1
j! j=1
j! k!
k>j
where Lj,k is defined as

Lj,k = ((I − μI )j − Nj )((I − μI )k − Nk )p(I|I)dI
˜ = Nj+k − Nj Nk . (6)
As can be seen in the above equation, Lj,k is commutative (Lj,k = Lk,j ).

A detailed calculation gives us L1,1 = M2 = σI2 (I),
˜ L1,2 = 2Id σ 2 (I)
I
˜ + M3 , · · · .
Substituting these terms into Eq. (5), we obtain:
σO
2
(Õ) = f 2 (I)L
˜ 1,1 + f (I)f
˜ (I)L
˜ 1,2 + · · ·
= f 2 (I)σ ˜ + f (I)f
˜ I2 (I) ˜ (I)(2I
˜ d σI (I) + M3 ) + · · · .
2 ˜
(7)
Eq. (7) is the exact form of the relationship between the response function and
noise variances in the input and output domains. By discarding the second- and
higher-degree terms of Eq. (7), Eq. (1) is obtained. We discuss the validity of
this approximation in Section 2.3.
2.2 Noise Variance Function

Eq. (1) shows the relationship between the response function and noise variance
functions in the input and output domains. The input to our method is the
measured noise variance in the output domain. This section models the noise
variance function σI2 (I)
˜ in the input domain so that the estimation algorithm
for the inverse response function g can be developed.
Input intensity I with camera noise can be written as
I = aP + NDC + NS + NR , (8)
where a is a factor of photon-to-electron conversion efficiency with amplification,

and P is the number of photons. NDC , NS , and NR indicate dark current noise,
shot noise, and readout noise, respectively [21]1 . The noise-free input intensity I˜
equals to aP .
Now we consider the noise variance function in the input domain. We assume
the different noise sources are independent. From Eq. (8), the noise variance
function in the input domain, σI2 (I),
˜ can be written as
σI2 (I)
˜ = Iσ
˜ 2 + σ2 + σ2 ,
S DC R (9)
where σ∗2 denotes the variances of the different noise sources [8]. Eq. (9) can be
written in a simplified form as
σI2 (I)
˜ = AI˜ + B, (10)
where A = σS2 and B = σDC 2

+ σR2
. This equation clearly shows the affine
relationship between the noise-free input intensity I˜ and the noise variance σI2 (I).
˜
2.3 Validity of the Approximation

It is important to consider the validity of the approximation in Eq. (7). In
this section, we show it in the following steps. First, we show that Li,j becomes
exponentially smaller as i+j increases. Second, the second largest term in Eq. (7),
L1,2 , is small enough to be negligible compared with L1,1 through a detailed
calculation. Hereafter, we normalize the input and output intensity ranges from 0
to 1.
Relationship between Li,j and i + j. By assuming independence of different noise

sources, the second- and higher-order moments can be computed by summing
up the moments of noises from different sources. Three types of noise sources
must be considered: dark current noise, shot noise, and readout noise [21]. We
do not consider low-light conditions [22], so the effect of the dark current noise
becomes small.
The probability density of the readout noise can be considered to have a
normal distribution with a mean value equal to the noise-free input intensity I.
˜
The moment can be written as
5
0 (i is odd)
NRi = i/2 (11)
j=1 (2j − 1)σ 2
R (i is even).
In Eq. (11), (2j − 1) σR

2
) 1, so the i-th moment of the readout noise NRi about
the noise-free input intensity becomes exponentially smaller as i increases.
Shot noise is modeled as Poisson distribution [21]. From the theory of gener-
alized Poisson distribution [23], the moment MSi about the mean of the distri-
bution is defined as

MSi ai−2 σS2 Ĩ + O(σS4 Ĩ ) (i ≥ 2), (12)
1
The effect of the fixed-pattern noise is included in the term P .
since the minimum unit of the distribution equals to a (See Eq. (8)). σS2 Ĩ is the
variance of shot noise where the noise-free input intensity is I.
˜ By substituting
this into Eq. (4) yields
i
i i−j j−2 2
NSi Idi + Id a σSĨ . (13)
j=2
j

Even in the worst case where ji is overestimated as 2i , NSi becomes exponen-
tially smaller, since Eq. (13) is rewritten as

i
NSi ≤ Idi + (2Id )i−j (2a)j−2 (2σSĨ )2 , (14)
j=2
and we know that 2Id ) 1, 2a ) 1, and (2σSĨ )2 ) 1. This equation shows that
NSi exponentially decreases as i increases.
The term Li,j is defined as Li,j = Ni+j − Ni Nj in Eq. (6). Because the i-th
moment of image noise Ni can be computed as the sum of the readout and shot
noise as Ni = NRi + NSi , it also becomes exponentially smaller as i increases.
From these results, we see that the term Li,j becomes exponentially smaller as
i + j increases.
Ratio of L1,1 to L1,2 . Now we show that L1,2 is small enough to be negligible
compared with L1,1 . A detailed calculation gives us L1,2 = 2Id M2 + M3 . The
third moment of shot noise MS3 can be computed from Eq. (12). Also, the third
moment of readout noise can be obtained using Eq. (4) as
MR3 = NR3 − 3Id NR2 − Id3 = −3Id σR
2
− Id3 . (15)
From these results, the following equation is obtained:
L1,2 = 2Id M2 − 3Id σR
2
− Id3 + aσS2 Ĩ . (16)
Since M2 σR 2
+ σS2 Ĩ , a ) Id , if M2 ) Id2 , the order of L1,2 is roughly the
same as the order of Id M2 . Since Id is the difference between the noise-free
input intensity and the mean which can be naturally considered very small, it is
implausible to have cases where M2 ) Id2 .
From these results, the order of L1,2 is roughly equivalent to the order of
Id L1,1 , and Id is small because it is computed in the normalized input domain,
e.g., in the order of 10−2 ( 1/28) in 8-bit image case. Therefore, L1,2 is about
10−2 times smaller than L1,1 .
To summarize, L1,2 is sufficiently small compared with L1,1 , and Li,j decreases
exponentially as i + j increases. Also, because response functions are smooth,
f (I)
˜ ) f (I).
˜ Therefore, Eq. (7) can be well approximated by Eq. (1).
3 Estimation Algorithm
This section designs an evaluation function for estimating inverse response func-
tions g, using the result of the previous section.
3.1 Evaluation Function

From Eqs. (1) and (10), the noise variance σO
2
(Õ) of the output intensity O is
σO
2
(Õ) f 2 (I)σI
˜ + σ 2 = f 2 (I)(A
˜ 2 (I)
Q
˜ I˜ + B) + σ 2 .
Q (17)
σQ
2
is the variance of the quantization noise, which affects after applying the
response function. Using the inverse response function g, Eq. (17) can be
rewritten as
1
σO2
(Õ) = (Ag(Õ) + B) + σQ
2
. (18)
g (Õ)2
The variance of the quantization noise σQ 2
becomes σQ 2
= l2 /12, where l is the
quantization interval. Since its distribution is uniform, the following equation
holds: 2l l
1 2 2 l2
σQ =
2
x p(x)dx =
2
x dx = . (19)
− 2l l − 2l 12
In the following, we use σO 2
m
(Õ) to represent the measured noise variance to
discriminate from the analytic form of the noise variance σO 2
(Õ). Using Eq. (18)
and the measured noise variances σOm (Õ), our method estimates the inverse
2
response function g that minimizes the following evaluation function:

2
E1 (g; σO
2
m
(Õ)) = min σO2
(Õ) − σO
2
m
(Õ) . (20)
A,B
Õ
Eq. (20) involves the estimation of A and B, which can be simply solved by
linear least square fitting, given g.
To make the algorithm robust against the measuring errors, namely the er-
roneous component in the measured noise, we use weighting factors. Eq. (20) is
changed to
1 2 2
E2 (g; σO
2
(Õ)) = min w(Õ) σO (Õ) − σO
2
(Õ) , (21)
m
A,B w(Õ) m
Õ
where the weight function w(Õ) controls the reliability on the measured noise
variance σO2
m
(Õ) at the intensity level Õ. We use a Cauchy distribution
(Lorentzian function) for computing the weight function w(Õ):
1
w(Õ) = , (22)
e2 +ρ
where e is defined as e = σO 2
(Õ) − σO 2
m
(Õ). A damping factor ρ controls the
relationship between the difference e and weight w(Õ). As ρ becomes smaller,
the weight w(Õ) decreases more rapidly as the difference e increases.
We also add a smoothness constraint to the evaluation function, and the
evaluation function becomes
1 1
E3 (g; σO
2
(Õ)) = 2 E2 + λs g (Õ)2 , (23)
m
Õ σOm (Õ)
nÕ
Õ
where nÕ is the number of possible noise-free output intensity levels, e.g., 256 in
8-bit case. λs is
a regularization factor that controls the effect of the smoothness
constraint. 1/ Õ σO 2
m
(Õ) is a normalization factor that makes E2 independent
of the degree of noise level.
Our method estimates the inverse response function ĝ by minimizing Eq. (23)
given the measured noise variance σO 2
m
(Õ):

ĝ = argmin E3 g; σO 2
m
(Õ) . (24)
g
3.2 Representation of Inverse Response Functions

To reduce the computational cost, we represent the inverse response functions
using a parametric model proposed by Grossberg and Nayar [5]. In their method,
principal component analysis (PCA) is performed on the database of real-world
response functions (DoRF) to obtain a small number of eigenvectors that can
represent the space of the response functions. As done by Lin et al . [12,13],
we compute the principal components of the inverse response functions using
the DoRF. Using the principal components, we represent the inverse response
function g as g = g0 +Hc, where g0 is the mean vector of all the inverse response
functions, H is a matrix in which a column vector represents an eigenvector, and
c is a vector of PCA coefficients. Following Lin et al . [12,13], we use the first
five eigenvectors. Using this representation, the number of unknown variables is
significantly decreased, e.g., from 256 to 5 in the case of 8-bit images.
3.3 Implementation
In our implementation, we set the damping factor ρ to the variance of the dif-
ference e in Eq. (22). The regularization factor λs is set to 5 × 10−7 from our
empirical observation. Minimization is performed in an alternating manner. We
perform the following steps until convergence:
1. minimize the evaluation function in Eq. (23) with fixing the weight func-
tion w(Õ)
2. recompute the values of the weight function w(Õ) using the current estima-
tion result
We use the Nelder-Mead Simplex method [24] as the minimization algorithm
implemented in Matlab as a function fminsearch. The values of the weight
function w(Õ) are set to one for every Õ at the beginning. During the exper-
iments, we used five initial guesses for the inverse response function g as the
input to the algorithm. The converged result that minimizes the energy score is
finally taken as the global solution.
4 Experiments
We used two different setups to evaluate the performance of the proposed al-
gorithm; one is with multiple images taken by a fixed video camera, the other
is using a single image. The two setups differ in the means for collecting noise
variance information.
4.1 Multiple-Images Case

In this experiment, the measurements of noise variances are obtained by cap-
turing multiple shots of a static scene from a fixed viewpoint with fixed camera
parameters. From multiple images, a histogram of output intensities is created
for each pixel. From the histogram, the noise-free output intensity Õ is deter-
mined by taking the mode of the distribution, assuming that the noise-free in-
tensity should correspond to the most frequently observed signal. The pixelwise
histograms are then merged together to form the histogram h(O, Õ) for each
output intensity level Õ. Finally, the noise distribution p(O|Õ) is computed by
normalizing the histogram h(O, Õ) as
h(O, Õ)
p(O|Õ) = . (25)
O h(O, Õ)
Results. We used three different video cameras for this experiment: Sony
DCR-TRV9E (Camera A), Sony DCR-TRV900 NTSC (Camera B), and Sony
DSR-PD190P (Camera C). To obtain the ground truth of Camera C, we
used Mitsunaga and Nayar’s method [4], and the Macbeth color checker-based
method [1], and combined these results by taking the mean. For Camera A
and B, we used only the Macbeth color checker-based method [1] to obtain the
ground truth because the exposure setting was not available in these cameras.
The results obtained by the proposed method are compared with the ground
truth curves.
Figure 3 shows the results of our algorithm. The top row shows the plot of
the estimated inverse response functions with the corresponding ground truth
Fig. 3. Results of our estimation method. Top row: comparison of inverse response
functions. Bottom row: measured noise variance and fitting result.
curves. The bottom row shows the estimated and measured distributions of
noise variances; the horizontal axis is the normalized output, and the vertical
axis corresponds to the noise variance. Figure 4 shows the scenes used to obtain
these results.
Figure 3 (a) shows an estimation result using the blue channel of Camera A.
The maximum difference is 0.052 and the RMSE is 0.025 in terms of normalized
input. As shown in the bottom of (a), the noise variances in lower output levels
contain severe measured errors. Our algorithm is robust against such errors be-
cause of the use of adaptive weighting factors. Figure 3 (b) shows the result of
Camera B (green channel). The maximum difference is 0.037 and the RMSE is
0.022. Figure 3 (c) shows the estimation result of Camera C (red channel). The
input frames are obtained by setting the camera gain to 12 db which causes high
noise level. The maximum difference is 0.037 and the RMSE is 0.024.
Table 1 summarizes all the experimental results. For each camera, three differ-
ent scenes are used. The algorithm is applied to RGB-channels independently,
therefore 9 datasets for each camera are used. Disparity represents the mean
of maximum differences in normalized input. From these results, the proposed
method performs well even though the algorithm only uses the noise variance as
input.
4.2 Comparison with Another Noise-Based Estimation Method
Figure 5 shows the comparison between our method and Matsushita and Lin’s
method [15]. Unlike other estimation methods, these two methods take noise as
input. We use Camera B for the comparison.
As shown in the result, the estimation results are equivalent when the number
of images is relatively large. However, Matsushita and Lin’s method breaks down
when the number of samples becomes small, and our method shows significant
superiority. In statistics, it is known that variance of measured from samples’
variance is inversely proportional to the number of the samples. Therefore, the
measured variance becomes more stable than the profile of noise distribution
does, as the number of samples increases. In addition, Matsushita and Lin’s
symmetry criterion naturally requires large number of samples to make the noise
profiles smooth, while it does not hold in the lower number of samples in Figure 5.
These are why our method works well when the number of samples is relatively
small.
Table 1. Mean RMSE and disparity of

the estimated inverse response functions in
terms of normalized input. Three different
scenes were used for each camera.
Camera Mean RMSE Disparity

A. DCR-TRV9E 0.026 0.053
B. DCR-TRV900 0.024 0.040
Fig. 4. Recorded scenes corresponding to
C. DSR-PD190P 0.033 0.055
the results in Figure 3 (a-c)
Fig. 5. Comparison between our method and Matsushita and Lin’s method [15]. Our
method uses noise variance, but not profiles of noise distributions. Our method works
well even when the sampling number is relatively small.
4.3 Single-Image Case

We describe a single-image case where only one shot of the scene is available. In
this setup, the distribution of noise variances are collected from uniformly colored
image regions. However, the measured noise distribution is often insufficient to
determine the inverse response functions because the limited measurements do
not span the entire range of output levels. To better constrain the problem, we
use a prior model p(g) of the inverse response functions obtained from the DoRF
as done in [12] and [15].
Using the prior model p(g), the MAP (maximum a posteriori) estimation is
performed by maximizing the cdf p(g|σO 2
m
(Õ)) which represents the probability
of the inverse response function being g when the measured noise variances are
σO
2
m
(Õ) as

ĝ = argmax p(g|σO
2
m
(Õ)) = argmax log p(σO
2
m
(Õ)|g) + log p(g) . (26)
g g
The likelihood p(σO

2
m
(Õ)|g) is defined as
1
p(σO
2
(Õ)|g) = exp −λp E3 (g; σO
2
(Õ)) , (27)
m
Z m
where Z is the normalization factor, and λp is a regularization coefficient that

determines the weight on the evaluation function E3 . We empirically set λp to
2 × 104 in the experiments. The prior model p(g) is formed using a multivariate
Gaussian mixture model as

K
p(g) = αi N (g; μi , Σi ), (28)
i=1
where N represents a normal distribution with mean μ and covariance matrix Σ,

and αi is a weight factor. The prior model is obtained using the PCA coefficients
of the inverse response functions in the DoRF by applying the cross-entropy
method [25]. The number of normal distributions K is set to 5 in our experiments.
Fig. 6. Relationship between the noise level and mean RMSE of the estimates. Left
image shows one of the photographed scenes. Top row shows magnification of a part of
the image at different ISO levels. Bottom row shows the mean RMSE of RGB channels
at each ISO gain level, and demonstrates that our estimation method is independent
of noise levels.
Results. We used a Canon EOS-20D camera for the experiment. To obtain the
ground truth, we used Mitsunaga and Nayar’s method [4] using images taken
with different exposures. Since our focus is on estimating the inverse response
functions from the measured noise variances, we photographed a scene composed
of relatively flat and uniformly colored surfaces, so that the noise variances can
be easily obtained. The left image in Figure 6 shows one of two scenes used for
the experiment. We photographed them five times each at six different camera
gains (ISO 100 ∼ 3200). We manually selected 21 homogeneous image regions
to obtain the noise variances as input. In total, we ran our estimation algorithm
60 times (= 2 scenes × 5 shots × 6 ISO levels) for each RGB color channel.
Figure 6 summarizes the results of estimation at different ISO levels. The
noise level increases with the ISO gain level, as shown by the cropped images
on the top. The results indicate that the estimation is unaffected by the greater
noise level. The mean RMSE is almost constant across the different ISO levels,
which verifies that our method is not sensitive to the noise level.
5 Conclusions
In this paper, we have proposed the method for estimating a radiometric response
function using noise variance, not noise distribution, as input. The relationship
between the radiometric response function and noise variances in input and out-
put domains is explicitly derived, and this result is used to develop the estimation
algorithm. The experiments are performed for two different scenarios; one is with
multiple shots of the same scene, and the other is only from a single image. These
experiments quantitatively demonstrate the effectiveness of the proposed algo-
rithm, especially its robustness against noise. With our method, either special
equipment or images taken with multiple exposures are not necessary.
Limitations. It is better for our method that the measured noise variances cover
a wide range of intensity levels. Wider coverage provides more information to
the algorithm, so the problem becomes more constrained. This becomes an is-
sue, particularly in the single-image case. In the single-image case, we used a
simple method to collect the noise variances, but more sophisticated methods
such like [20] can be used to obtain more accurate measurements that could
potentially cover a wider range of intensity levels.
Acknowledgement
The authors would like to thank Dr. Bennett Wilburn for his useful feedback on
this research.
References
1. Chang, Y.C., Reid, J.F.: Rgb calibration for color image analysis in machine vision.
IEEE Trans. on Image Processing 5, 1414–1422 (1996)
2. Nayar, S.K., Mitsunaga, T.: High dynamic range imaging: Spatially varying pixel
exposures. In: Proc. of Comp. Vis. and Patt. Recog. (CVPR), pp. 472–479 (2000)
3. Mann, S., Picard, R.: Being ‘undigital’ with digital cameras: Extending dynamic
range by combining differently exposed pictures. In: Proc. of IS & T 48th Annual
Conf., pp. 422–428 (1995)
4. Mitsunaga, T., Nayar, S.K.: Radiometric self-calibration. In: Proc. of Comp. Vis.
and Patt. Recog. (CVPR), pp. 374–380 (1999)
5. Grossberg, M.D., Nayar, S.K.: What is the space of camera response functions? In:
Proc. of Comp. Vis. and Patt. Recog. (CVPR), pp. 602–609 (2003)
6. Mann, S.: Comparametric equations with practical applications in quantigraphic
image processing. IEEE Trans. on Image Processing 9, 1389–1406 (2000)
7. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from pho-
tographs. Proc. of ACM SIGGRAPH, 369–378 (1997)
8. Tsin, Y., Ramesh, V., Kanade, T.: Statistical calibration of ccd imaging process.
In: Proc. of Int’l Conf. on Comp. Vis. (ICCV), pp. 480–487 (2001)
9. Pal, C., Szeliski, R., Uyttendale, M., Jojic, N.: Probability models for high dynamic
range imaging. In: Proc. of Comp. Vis. and Patt. Recog. (CVPR), pp. 173–180
(2004)
10. Grossberg, M.D., Nayar, S.K.: What can be known about the radiometric response
function from images? In: Heyden, A., Sparr, G., Nielsen, M., Johansen, P. (eds.)
11. Kim, S.J., Pollefeys, M.: Radiometric alignment of image sequences. In: Proc. of
Comp. Vis. and Patt. Recog. (CVPR), pp. 645–651 (2004)
12. Lin, S., Gu, J., Yamazaki, S., Shum, H.Y.: Radiometric calibration from a single
image. In: Proc. of Comp. Vis. and Patt. Recog. (CVPR), pp. 938–945 (2004)
13. Lin, S., Zhang, L.: Determining the radiometric response function from a single
grayscale image. In: Proc. of Comp. Vis. and Patt. Recog. (CVPR), pp. 66–73
(2005)
14. Wilburn, B., Xu, H., Matsushita, Y.: Radiometric calibration using temporal irra-
diance mixtures. In: Proc. of Comp. Vis. and Patt. Recog. (CVPR) (2008)
15. Matsushita, Y., Lin, S.: Radiometric calibration from noise distributions. In: Proc.
of Comp. Vis. and Patt. Recog. (CVPR) (2007)
16. Takamatsu, J., Matsushita, Y., Ikeuchi, K.: Estimating camera response functions
using probabilistic intensity similarity. In: Proc. of Comp. Vis. and Patt. Recog.
(CVPR) (2008)
17. Matsushita, Y., Lin, S.: A probabilistic intensity similarity measure based on noise
distributions. In: Proc. of Comp. Vis. and Patt. Recog. (CVPR) (2007)
18. Janesick, J.R.: Photon Transfer. SPIE Press (2007)
19. Schechner, Y.Y., Nayar, S.K., Belhumeur, P.N.: Multiplexing for optimal lighting.
IEEE Trans. on Patt. Anal. and Mach. Intell. 29, 1339–1354 (2007)
20. Liu, C., Freeman, W.T., Szeliski, R., Kang, S.B.: Noise estimation from a single
image. In: Proc. of Comp. Vis. and Patt. Recog. (CVPR), pp. 901–908 (2006)
21. Healey, G.E., Kondepudy, R.: Radiometric ccd camera calibration and noise esti-
mation. IEEE Trans. on Patt. Anal. and Mach. Intell. 16, 267–276 (1994)
22. Alter, F., Matsushita, Y., Tang, X.: An intensity similarity measure in low-light
conditions. In: Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS,
23. Consul, P.C.: Generalized Poisson Distributions: Properties and Applications. Mar-
cel Dekker Inc., New York (1989)
24. Nelder, J.A., Mead, R.: A simplex method for function minimization. Computer
Journal 7, 308–312 (1965)
25. Botev, Z., Kroese, D.: Global likelihood optimization via the cross-entropy method
with an application to mixture models. In: Proc. of the 36th Conf. on Winter simul.,
pp. 529–535 (2004)
Solving Image Registration Problems Using
Interior Point Methods
Camillo Jose Taylor and Arvind Bhusnurmath
GRASP Laboratory, University of Pennsylvania
Abstract. This paper describes a novel approach to recovering a para-

metric deformation that optimally registers one image to another. The
method proceeds by constructing a global convex approximation to
the match function which can be optimized using interior point meth-
ods. The paper also describes how one can exploit the structure of the
resulting optimization problem to develop efficient and effective match-
ing algorithms. Results obtained by applying the proposed scheme to a
variety of images are presented.
1 Introduction
Image registration is a key problem in computer vision that shows up in a wide

variety of applications such as image mosaicing, medical image analysis, face
tracking, handwriting recognition, stereo matching and motion analysis. This
paper considers the problem of recovering the parameters of a deformation that
maps one image onto another. The main contribution is a novel approach to
this problem wherein the image matching problem is reformulated as a Linear
Program (LP) which can be solved using interior point methods. The paper
also describes how one can exploit the special structure of the resulting LP to
derive efficient implementations which can effectively solve problems involving
hundreds of thousands of pixels and constraints.
One of the principal differences between the proposed approach and other ap-
proaches that have been developed [1,2,3] is that the scheme seeks to construct a
global convex approximation to the matching function associated with the
registration problem as opposed to constructing a local convex model around the
current parameter estimate. The approach is intended for situations where the dis-
placements between frames are large enough that local matches at the pixel level
are likely to be ambiguous. For example, in the experiments we consider images
that are 320 pixels on side where individual pixels may be displaced by up to 40
pixels along each dimension. The approximation procedure is designed to capture
the uncertainties inherent in matching a given pixel to a wide swath of possible
correspondents.
One common approach to solving image matching problems proceeds by ex-
tracting feature points in the two images, establishing correspondences between
the frames, and then using a robust estimation procedure to recover the pa-
rameters of the transformation. This approach is exemplified by the work of

Solving Image Registration Problems Using Interior Point Methods 639
Mikolajczyk and Schmid [4] who proposed a very effective scheme for detecting
and matching interest points under severe affine deformations. This approach
works best when the interframe motion is close to affine since more complicated
deformation models can distort the feature points beyond recognition. Further,
it becomes increasingly difficult to apply robust estimation methods as the com-
plexity of the deformation model increases since an ever increasing number of
reliable point matches are required.
Belongie and Malik [5] proposed an elegant approach to matching shapes
based on information derived from an analysis of contour features. This approach
is similar to [4] in that it revolves around feature extraction and pointwise cor-
respondence. The method described in this work is very different from these in
that it avoids the notion of features altogether, instead it proceeds by construct-
ing a matching function based on low level correlation volumes and allows every
pixel in the image to constrain the match to the extent that it can.
Shekhovstov Kovtun and Hlavac [6] have developed a novel method for image
registration that uses Sequential Tree-Reweighted Message passing to solve a lin-
ear program that approximates a discrete Markov Random Field optimization
problem. Their work also seeks to construct a globally convex approximation
to the underlying image matching problem but the approach taken to formulat-
ing and solving the optimization problem differ substantially from the method
discussed in this paper.
Linear programming has been previously applied to motion estimation [7,8].
The work by Jiang et al.. [7] on matching feature points is similar to ours in that
the data term associated with each feature is approximated by a convex com-
bination of points on the lower convex hull of the match cost surface. However,
their approach is formulated as an optimization over the interpolating coeffi-
cients associated with these convex hull points which is quite different from the
approach described in this paper. Also their method uses the simplex method
for solving the LP while the approach described in this paper employs an inte-
rior point solver which allows us to exploit the structure of the problem more
effectively.
2 Image Registration Algorithm

The objective of the algorithm is to recover the deformation that maps a base
image onto a target image. This deformation is modeled in the usual manner
by introducing two scalar functions Dx (x, y, px ) and Dy (x, y, py ) which capture
the displacement of a pixel at location (x, y) along the horizontal and vertical
directions respectively [9,5,10]. Here px and py represent vectors of parameters
that are used to model the deformation. Consider for example an affine deforma-
tion where the horizontal displacements are given by Dx (x, y) = c1 + c2 x + c2 y,
then px = [c1 , c2 , c3 ] would capture the parameters of this transformation. In
the sequel we will restrict our consideration to models where the displacements
can be written as a linear function of the parameters. That is, if we let Dx and
640 C.J. Taylor and A. Bhusnurmath
Dy represent vectors obtained by concatenating the displacements at all of the

pixels then Dx = Cpx and Dy = Cpy for some matrix C. Here the columns of
the matrix C constitute the basis vectors of the displacement field [9].
2.1 Formulating Image Matching as an LP
The problem of recovering the deformation that maps a given base image onto a
given target image can be phrased as an optimization problem. For every pixel
in the target image one can construct an objective function, exy , which captures
how similar the target pixel is to its correspondent in the base image as a function
of the displacement applied at that pixel.
Figure 1(a) shows an example of one such function for a particular pixel in one
of the test images. This particular profile was constructed by computing the #2 dif-
ference between the RGB value of the target pixel and the RGB values of the pixels
in the base image for various displacements up to ±10 pixels in each direction.
Our goal then is to minimize an objective function E(px , py ) which models
how the discrepancy between the target and base images varies as a function of
the deformation parameters, px and py .

E(px , py ) = exy (Dx (x, y, px ), Dy (x, y, py )) (1)
x y
In general, since the component exy functions can have arbitrary form the land-
scape of the objective function E(px , py ) may contain multiple local minima
Fig. 1. (a) Error surface associated with particular pixel in the target image that
encodes how compatible that pixel is with various x, y displacements (b) Piecewise
planar convex approximation of the error surface
which can confound most standard optimization methods that proceed by con-
structing local approximations of the energy function.
The crux of the proposed approach is to introduce a convex approximation for
the individual objective functions exy . This leads directly to an approximation
of the global objective function E (px , py ) which is convex in the deformation
parameters. Once this has been done, one can recover estimates for the deforma-
tion parameters and, hence, the deformation by solving a convex optimization
problem which is guaranteed to have a unique minimum.
The core of the approximation step is shown in Figure 1(b), here the original
objective function is replaced by a convex lower bound which is constructed
by considering the convex hull of the points that define the error surface. This
convex lower hull is bounded below by a set of planar facets.
In order to capture this convex approximation in the objective function we
introduce one auxiliary variable z(x, y) for every pixel in the target image. There
are a set of linear constraints associated with each of these variables which reflect
the constraint that this value must lie above all of the planar facets that define
the convex lower bound.
z(x, y) ≥ aix (x, y)Dx (x, y, px ) + aiy (x, y)Dy (x, y, py ) − bi (x, y) ∀i (2)
Here the terms aix , aiy and bi denote the coefficients associated with each of the
facets in the approximation.
The problem of minimizing the objective function E (px , py ) can now be
rephrased as a linear program as follows:

minpx ,py ,z x y z(x, y) (3)
st z(x, y) ≥ ax (x, y)Dx (x, y, px ) + aiy Dy (x, y, py )
i
− b (x, y) ∀x, y, i
i
(4)
This can be written more compactly in matrix form as follows:

minpx ,py ,z 1T z (5)
st Ax Dx + Ay Dy − Iz z ≤ b
Dx = Cpx
Dy = Cpy
where Ax and Ay are Iz are sparse matrices obtained by concatenating the con-
straints associated with all of the planar facets and z and b are vectors obtained
by collecting the z(x, y) and bi (x, y) variables respectively.
Note that the Ax , Ay and Iz matrices all have the same fill pattern and are
structured as shown in equation 6, the non zero entries in the Iz matrix are all
1. In this equation M denotes the total number of pixels in the image and Si
refers to the number of planar facets associated with pixel i.
⎡ ⎤
a11 0 ··· ··· 0
⎢ a21 0 ··· ··· 0 ⎥
⎢ ⎥
⎢ .. ⎥
⎢ . 0 ··· ··· 0 ⎥
⎢ ⎥
⎢ aS 1 1 0 ··· ··· 0 ⎥
⎢ ⎥
⎢ 0 a12 0 ··· 0 ⎥
⎢ ⎥
⎢ 0 a22 0 ··· 0 ⎥
⎢ ⎥
A=⎢
⎢ 0
.. ⎥
⎥ (6)
⎢ . 0 ··· 0 ⎥
⎢ 0 aS 2 2 0 ··· 0 ⎥
⎢ ⎥
⎢ .. .. ⎥
⎢ 0 0 . . 0 ⎥
⎢ ⎥
⎢ 0 ··· ··· 0 a1M ⎥
⎢ ⎥
⎢ .. ⎥
⎣ 0 ··· ··· 0 . ⎦
0 · · · · · · 0 aS M M
The linear program shown in Equation 5 can be augmented to include constraints
on the displacement entries, Dx , Dy and the z values as shown in Equation 7.
Here the vectors blb and bub capture the concatenated lower and upper bound
constraints respectively. It would also be a simple matter to include bounding
constraints on the parameter values at this stage. Alternatively one could easily
add a convex regularization term to reflect a desire to minimize the bending
energy associated with the deformation.
minpx ,py ,z 1T z (7)
⎡ ⎤⎛ ⎞⎛ ⎞ ⎛ ⎞
Ax Ay −Iz C 0 0 px b
⎣ −I ⎦ ⎝ 0 C 0 ⎠ ⎝ py ⎠ ≤ ⎝ blb ⎠
I 0 0 I z bub
Note that the proposed approximation procedure increases the ambiguity asso-
ciated with matching any individual pixel since the convex approximation is a
lower bound which may significantly under estimate the cost associated with
assigning a particular displacement to a pixel. What each pixel ends up con-
tributing is a set of convex terms to the global objective function. The linear
program effectively integrates the convex constraints from tens of thousands of
pixels, constraints which are individually ambiguous but which collectively iden-
tify the optimal parameters. In this scheme each pixel contributes to constraining
the deformation parameters to the extent that it is able. Pixels in homogenous
regions may contribute very little to the global objective while well defined fea-
tures may provide more stringent guidance. There is no need to explicitly identify
distinguished features since local matching ambiguities are handled through the
approximation process.
2.2 Solving the Matching LP

Once the image registration problem has been reformulated as the linear program
given in equation 7 the barrier method [11] can be employed to solve the problem.
In this method, a convex optimization problem of the following form
min f0 (x)
st fi (x) ≤ 0, i = 1, . . . , m (8)

m
is solved by minimizing φ(x, t) = tf0 (x) − log(−fi (x)) for increasing values
i=1
of t until convergence. At each value of t a local step direction, the Newton step,
needs to be computed. This involves the solution of a system of linear equations
involving the Hessian and the gradient of φ(x, t). The Hessian can be computed
from the following expression H = [AT diag(s−2 )A] where s = b − Ax and s−2
denotes the vector formed by inverting and squaring the elements of s. Similarly
the gradient of the φ(x, t) can be computed from the following expression:
g = −tw − AT s−1 (9)

Then the Newton step is computed by solving
[AT diag(s−2 )A]δx = g (10)

For our matching problem, it can be shown that this Newton step system can
be written in the following form:
! "
Hp HzT δp gp
= (11)
Hz D 6 δz gz
where
! "
(C T D1 C) (C T D2 C)
Hp =
(C T D2 C) (C T D3 C)
# $
Hz = (D4 C) (D5 C) (12)
δp and δz denote proposed changes in the deformation parameters and the z

variables respectively and D1 , D2 , D3 , D4 , D5 , D6 are all diagonal matrices.
At this point we observe that since the matrix D6 is diagonal we can simplify
the linear system in Equation 11 via the Schur complement. More specifically
we can readily solve for δz in terms of δp as follows: δz = D6−1 (gz − Hz δp).
Substituting this expression back into the system yields the following expression
where all of the auxiliary z variables have been elided.
(Hp − HzT D6−1 Hz )δp = (gp − HzT D6−1 gz ) (13)

This can be written more concisely as follows:
Hp δp = gp (14)
In short, computing the Newton Step boils down to solving the linear system in
Equation 14. Note that the size of this system depends only on the dimension of
the parameter vector, p. For example if one were interested in fitting an affine
model which involves 6 parameters, 3 for px and 3 for py , one would only end
up solving a linear system with six degrees of freedom. Note that the compu-
tational complexity of this key step does not depend on the number of pixels
being considered or on the number of constraints that were used to construct the
convex approximation. This is extremely useful since typical matching problems
will involve hundreds of thousands of pixels and a similar number of constraint
equations. Even state of the art LP solvers like MOSEK and TOMLAB would
have difficulty solving problems of this size.
2.3 Deformation Models
Experiments were carried out with two classes of deformation models. In the first
class the displacements at each pixel are computed as a polynomial function of
the image coordinates. For example for a second order model:
Dx (x, y) = c1 + c2 x + c3 y + c4 xy + c5 x2 + c6 y 2 (15)
These deformations are parameterized by the coefficients of the polynomials. The
complexity of the model can be adjusted by varying the degree of the polynomial.
A number of interesting deformation models can be represented in this manner
include affine, bilinear, quadratic and bicubic.
Another class of models can be represented as a combination of an affine
deformation and a radial basis function. That is

Dx (x, y) = c1 + c2 x + c3 y + ki φ((x, y) − (xi , yi )) (16)
i
Once again the deformation model is parameterized by the coefficients c1 , c2 ,

c3 , ki and the function φ represents the interpolating kernel. Two different vari-
ants of this kernel were considered in the experiments, a Gaussian kernel, φ(r) =
exp(−(r/σ)2 ) and a thin plate spline kernel φ(r) = r2 log r. In the sequel we will
refer to the former as the Gaussian deformation model and the latter as the Thin
Plate Spline model.
In the experiments the coordinates of the kernel centers,(xi , yi ) were evenly
distributed in a grid over the the image. The complexity of the model can be
varied by varying the number of kernel centers employed. All of the experiments
that used this model employed 16 kernel centers arranged evenly over the image
in a four by four grid.
2.4 Coarse to Fine
It is often advantageous to employ image registration algorithms in a coarse

to fine manner [1]. In this mode of operation the base and target images are
downsampled to a lower resolution and then matched. The deformation recovered
from this stage is used to constrain the search for matches at finer scales. With
this scheme, gross deformations are captured at the coarser scales while the
finer scales fill in the details. It also serves to limit the computational effort
required since one can effectively constrain the range of displacements that must
be considered at the finer scales which limits the size of the correlation volumes
that must be constructed.
In the experiments described in section 3.1 the images are first downsampled
by a factor of 4 and then matched. The deformations computed at this scale in-
form the search for correspondences at the next finer scale which is downsampled
from the originals by a factor of 2.
Note that as the approach proceeds to finer scales, the convex approximation
is effectively being constructed over a smaller range of disparities which means
that it increasingly approaches the actual error surface.
Two different experiments were carried out to gauge the performance of the
registration scheme quantitatively. In the first experiment each of the images in
our data set was warped by a random deformation and the proposed scheme was
employed to recover the parameters of this warp. The recovered deformation was
compared to the known ground truth deformation to evaluate the accuracy of
the method.
In the second set of experiments the registration scheme was applied to por-
tions of the Middlebury stereo data set. The disparity results returned by the
method were then compared to the ground truth disparities that are provided
for these image pairs.
3.1 Synthetic Deformations

In these experiments the proposed scheme was applied to a number of different
images. In each case, a random deformation was constructed using a particular
motion model. The base image was warped by the deformation to produce the
target image and the registration algorithm was employed to recover this defor-
mation. In these experiments each of the base images was at most 320 pixels on
side. The deformations that were applied were allowed to displace the pixels in
the base image by up to ±12.5% of the image size. Hence for an image 320 pixels
on side each pixel in the image can be displaced by ±40 pixels along each dimen-
sion. The random deformations were specifically constructed to fully exercise the
range of displacements so the maximum allowed displacement values are achieved
in the applied warps. In order to recover such large deformations, the registration
scheme is applied in a coarse to fine manner as described in Section 2.4.
The underlying matching functions associated with each of the pixels in the
target image, exy , are constructed by simply comparing the pixel intensity in
the target image to the pixels in a corresponding range in the base image. This
is equivalent to conducting sum of squared difference (SSD) matching for each
pixel using a 1 × 1 matching window.
In order to provide a quantitative evaluation of the scheme, the recovered
deformation field, (Dx (x, y), Dy (x, y)) was compared to the known ground truth
deformation field (Dxt (x, y), Dyt (x, y)) and the mean, median and maximum dis-
crepancy between these two functions over the entire image was computed. The
Table 1. This table details the deformation applied to each of the images in the data
set and reports the discrepancy between the deformation field returned by the method
and the ground truth displacement field
error in pixels
Image Deformation Model no. of parameter mean median max
Football Gaussian 38 0.1524 0.1306 0.5737
Hurricane Gaussian 38 0.1573 0.1262 0.7404
Spine Affine 6 0.1468 0.1314 0.4736
Peppers Gaussian 38 0.1090 0.0882 0.7964
Cells Thin Plate Spine 38 0.1257 0.1119 0.8500
Brain Gaussian 38 0.1190 0.0920 0.8210
Kanji third degree polynomial 20 0.1714 0.0950 2.5799
Aerial bilinear 8 0.0693 0.0620 0.2000
Face1 Gaussian 38 0.1077 0.0788 0.6004
Face2 Gaussian 38 0.5487 0.3095 4.6354
results are tabulated in Table 1. This table also indicates what type of defor-
mation model was applied to each of the images along with the total number of
parameters required by that model.
Note that in every case the deformed result returned by the procedure is al-
most indistinguishable from the given target. More importantly, the deformation
fields returned by the procedure are consistently within a fraction of a pixel of
the ground truth values. The unoptimized Matlab implementation of the match-
ing procedure takes approximately 5 minutes to proceed through all three scales
and produce the final deformation field for a given image pair.
3.2 Stereo Data Set

The image registration scheme was applied to regions of the image pairs taken
from the Middlebury stereo data set. This data set was chosen because it included
ground truth data which allows us to quantitatively evaluate the deformation
results returned by the registration scheme. Here the vertical displacement be-
tween the two images is zero and the horizontal displacement field Dx (x, y) is
modeled as an affine function.
The correlation volume was computed using sum of squared difference match-
ing with a five by five correlation window. For the teddy image, the correlation
volume was constructed by considering displacements between 12 and 53 pixels
while for the venus image the displacement range was 3 to 20 pixels. In this
case, the convex lower bound approximations to the individual score functions
degenerates to a piecewise linear profile along the horizontal dimension.
In each of the images two rectangular regions were delineated manually and
an affine displacement model was fit to the pixels within those regions using the
proposed method.
Base Image Target Image Result
(a) Football
(b) Hurricane
(c) Spine
(d) Peppers
(e) Cells
Fig. 2. Results obtained by applying the proposed method to actual image pairs. The
first two columns correspond to the input base and target images respectively while
the last column corresponds to the result produced by the registration scheme.
The first column of Figure 4 shows the left image in the pair, the second
column shows what would be obtained if one used the raw SSD stereo results
and the final column shows the ground truth disparities.
Base Image Target Image Result
(f) Brain
(g) Kanji
(h) Aerial
(i) Face1
(j) Face2
Fig. 3. More Registration Results

Left Image SSD Disparity Solution Ground Truth Disparity
(a) Teddy
(b) Venus
Fig. 4. The proposed image registration scheme was applied to the delineated regions
in the Middlebury Stereo Data Set. The first column shows the left image, the second
column the raw results of the SSD correlation matching and the last column the ground
truth disparity.
Table 2. This table reports the discrepancy between the affine deformation field re-
turned by the method and the ground truth disparities within each region
error in pixels
Image Region mean median
teddy bird house roof 0.2558 0.2245
teddy foreground 0.9273 0.8059
venus left region 0.0317 0.0313
venus right region 0.0344 0.0317
The selected rectangles are overlaid on each of the images. These regions
were specifically chosen in areas where there was significant ambiguity in the
raw correlation scores to demonstrate that the method was capable of correctly
integrating ambiguous data. Table 2 summarizes the results of the fitting proce-
dure. The reconstructed disparity fields within the regions were compared to the
ground truth disparities and the mean and median discrepancy between these
two fields is computed over all of the pixels within the region.
4 Conclusion
This paper has presented a novel approach to tackling the image registration
problem wherein the original image matching objective function is approximated
by a linear program which can be solved using the interior point method. The
paper also describes how one can exploit the special structure of the resulting
linear program to develop efficient algorithms. In fact the key step in the resulting
resulting procedure only involves inverting a symmetric matrix whose dimension
reflects the complexity of the model being recovered.
While the convex approximation procedure typically increases the amount of
ambiguity associated with any individual pixels, the optimization procedure ef-
fectively aggregates information from hundreds of thousands of pixels so the net
result is a convex function that constrains the actual global solution. In a cer-
tain sense, the proposed approach is dual to traditional non-linear optimization
schemes which seek to construct a local convex approximation to the objective
function. The method described in this work proceeds by constructing a global
convex approximation over the specified range of displacements.
A significant advantage of the approach is that once the deformation model
and displacement bounds have been selected, the method is insensitive to ini-
tialization since the convex optimization procedure will converge to the same
solution regardless of the start point. This means that the method can be di-
rectly applied to situations where there is a significant deformation.
The method does not require any special feature detection or contour extrac-
tion procedure. In fact all of the correlation volumes used in the experiments
were computed using nothing more than pointwise pixel comparisons. Since the
method does not hinge on the details of the scoring function more sophisticated
variants could be employed as warranted. The results indicate the method pro-
duces accurate results on a wide range of image types and can recover fairly
large deformations.
References
1. Bajcsy, R., Kovacic, S.: Multiresolution elastic matching. Computer Vision, Graph-
ics and Image Processing 46(1), 1–21 (1989)
2. Cootes, T., Edwards, G., Taylor, C.: Active appearance models. IEEE Transactions
on Pattern Analysis and Machine Intelligence 23(6), 681–685 (2001)
3. Baker, S., Matthews, I.: Equivalence and efficiency of image alignment algorithms.
In: IEEE Conference on Computer Vision and Pattern Recognition, pp. 1090–1097
(2001)
4. Mikolajczyk, K., Schmid, C.: Scale and affine invariant interest point detectors.
5. Belongie, S., Malik, J.: Shape matching and object recognition using shape con-
texts. IEEE Transactions on Pattern Analysis and Machine Intelligence 24(24),
509 (2002)
6. Shekhovstov, A., Kovtun, I., Hlavac, V.: Efficient mrf deformation model for non-
rigid image matching. In: IEEE Conference on Computer Vision and Pattern
Recognition (2007)
7. Jiang, H., Drew, M., Li, Z.N.: Matching by linear programming and successive
convexification. PAMI 29(6) (2007)
8. Ben-Ezra, M., Peleg, S., Werman, M.: Real-time motion analysis with linear pro-
gramming. In: ICCV (1999)
9. Friston, K.J., Ashburner, J., Frith, C.D., Poline, J.B., Heather, J.D., Frackowiak,
R.S.J.: Spatial registration and normalization of images. Human Brain Mapping 2,
165–189 (1995)
10. Modersitzki, J.: Numerical Methods for Image Registration. Oxford University
Press, Oxford (2004)
11. Boyd, S., VandenBerghe, L.: Convex Optimization. Cambridge University Press,
Cambridge (2004)
3D Face Model Fitting for Recognition
Frank B. ter Haar and Remco C. Veltkamp
Department of Information and Computing Sciences, Utrecht University, the Netherlands
Abstract. This paper presents an automatic efficient method to fit a statistical

deformation model of the human face to 3D scan data. In a global to local fitting
scheme, the shape parameters of this model are optimized such that the produced
instance of the model accurately fits the 3D scan data of the input face. To in-
crease the expressiveness of the model and to produce a tighter fit of the model,
our method fits a set of predefined face components and blends these components
afterwards. Quantitative evaluation shows an improvement of the fitting results
when multiple components are used instead of one. Compared to existing meth-
ods, our fully automatic method achieves a higher accuracy of the fitting results.
The accurately generated face instances are manifold meshes without noise and
holes, and can be effectively used for 3D face recognition: We achieve 97.5%
correct identification for 876 queries in the UND face set with 3D faces. Our re-
sults show that contour curve based face matching outperforms landmark based
face matching.
1 Introduction
The use of 3D scan data for face recognition purposes has become a popular research
area. With high recognition rates reported for several large sets of 3D face scans, the 3D
shape information of the face proved to be a useful contribution to person identification.
The major advantage of 3D scan data over 2D color data, is that variations in scaling and
illumination have less influence on the appearance of the acquired face data. However,
scan data suffers from noise and missing data due to self-occlusion. To deal with these
problems, 3D face recognition methods should be invariant to noise and missing data,
or the noise has to be removed and the holes interpolated. Alternatively, data could be
captured from multiple sides, but this requires complex data acquisition. In this work
we propose a method that produces an accurate fit of a statistical 3D shape model of
the face to the scan data. The 3D geometry of the generated face instances, which are
without noise and holes, are effectively used for 3D face recognition.
Related work. The task to recognize 3D faces has been approached with many dif-
ferent techniques as described in surveys of Bowyer et al. [1] and Scheenstra et al. [2].
Several of these 3D face recognition techniques are based on 3D geodesic surface infor-
mation, such as the methods of Bronstein et al. [3] and Berretti et al. [4]. The geodesic
distance between two points on a surface is the length of the shortest path between two
points. To compute accurate 3D geodesic distances for face recognition purposes, a 3D
face without noise and without holes is desired. Since this is typically not the case with
laser range scans, the noise has to be removed and the holes in the 3D surface interpolated.
However, the success of basic noise removal techniques, such as Laplacian smoothing is
3D Face Model Fitting for Recognition 653
very much dependent on the resolution of the scan data. Straightforward techniques to
interpolate holes using curvature information or flat triangles often fail in case of com-
plex holes, as pointed out in [5]. The use of a deformation model to approximate new
scan data and interpolate missing data is a gentle way to regulate flaws in scan data.
A well known statistical deformation model specifically designed for surface meshes
of 3D faces, is the 3D morphable face model of Blanz and Vetter [6]. This statistical
model was built from 3D face scans with dense correspondences to which Principal
Component Analysis (PCA) was applied. In their early work, Blanz and Vetter [6] fit
this 3D morphable face model to 2D color images and cylindrical depth images from the
Cyberware T M scanner. In each iteration of their fitting procedure, the model parame-
ters are adjusted to obtain a new 3D face instance, which is projected to 2D cylindrical
image space allowing the comparison of its color values (or depth values) to the in-
put image. The parameters are optimized using a stochastic Newton algorithm. More
recently, Blanz et al. [7] proposed a method to fit their 3D morphable face model to
more common textured depth images. The fitting process is similar to their previous
algorithm, but now the cost function is minimized using both color and depth values
after the projection of the 3D model to 2D cylindrical image space. To initialize their
fitting process, they manually select seven corresponding face features on their model
and in the depth scan. A morphable model of expressions was proposed by Lu et al.
[8]. Starting from an existing neutral scan, they use their expression model to adjust the
vertices in a small region around the nose to obtain a better fit of the neutral scan to a
scan with a certain expression.
Non-statistical deformation models were proposed as well. Huang et al. [9] proposed
a global to local deformation framework to deform a shape with an arbitrary dimension
(2D, 3D or higher) to a new shape of the same class. They show their framework’s ap-
plicability to 3D faces, for which they deform an incomplete source face to a target face.
Kakadiaris et al. [10] deform an annotated face model to scan data. Their deformation
is driven by triangles of the scan data attracting the vertices of the model. The deforma-
tion is restrained by a stiffness, mass and damping matrix, which control the resistance,
velocity and acceleration of the model’s vertices. The advantage of such deformable
faces is that they are not limited to the statistical changes of the input shapes, so the de-
formation has less restrictions. However, this is also their disadvantage, because these
models cannot rely on statistics in case of noise and missing data.
Contribution. First, we propose a fully automatic algorithm to efficiently optimize
the parameters of the morphable face model, creating a new face instance that accurately
fits the 3D geometry of the scan data. Unlike other methods, ours needs no manual ini-
tialization, so that batch processing of large data sets has become feasible. Second, we
quantitatively evaluate our fitted face models and show that the use of multiple compo-
nents improves the fitting process. Thirdly, we show that our model fitting method is
more accurate than existing methods. Fourthly, we show that the accurately generated
face instances can be effectively used for 3D face recognition.
2 Morphable Face Model

In this work we fit the morphable face model of the USF Human ID 3D Database
[11] to 3D scan data to obtain a clean model of the face scan, that we use to identify
654 F.B. ter Haar and R.C. Veltkamp
3D faces. This statistical point distribution model (PDM) was built from 100 cylin-
drical 3D face scans with neutral expressions from which n=75,972 correspondences
were selected using an optic flow algorithm. Each face shape Si was described us-
ing the set of correspondences S = (x1 , y1 , z1 , ..., xn , yn , zn )T ∈ !3n and a mean
face S̄ was determined. PCA was applied to these 100 sets Si to obtain the m=99
most important eigenvectors of the PDM. The mean face S̄, the eigenvectors si =
(Δx1 , Δy1 , Δz1 , ..., Δxn , Δyn , Δzn )T , the eigenvalues λi (σi2 = λi ) and weights wi
m
are used to model new face instances according to Sinst = S̄ + i=1 wi σi si . Weight
wi represents the number of standard deviations a face instance morphs along eigen-
vector evi . Since the connectivity of the n correspondences in the PDM is known, each
instance is a triangular mesh with proper topology and without holes.
3 Face Scans
We fit the morphable face model to the 3D frontal face scans of the University of Notre
Dame (UND) Biometrics Database [12]. This set contains 953 range scans and a corre-
sponding 2D color texture from 277 different subjects. All except ten scans were used
in the Face Recognition Grand Challenge (FRGC v.1). Because the currently used mor-
phable model is based on faces with neutral expressions only, it makes no sense to use
collections containing many non-neutral scans such as the FRGC v.2. Nevertheless, our
proposed method performs well for the small expression variations of the UND set.
Throughout this work, we have only used the 3D scan data and neglected the available
2D color information.
We aim at 3D face recognition, so we need to segment the face from each scan. For
that, we employ our pose normalization method [13] that normalizes the pose of the face
and localizes the tip of the nose. Before pose normalization was applied to the UND
scan data, we applied a few basic preprocessing steps to the scan data: the 2D depth
images were converted to triangle meshes by connecting the adjacent depth samples
with triangles, slender triangles and singularities were removed, and only considerably
large components were retained.
The cleaned surface meshes were randomly sampled, such that every ≈2.0 mm2 of
the surface is approximately sampled once. The pose normalization method uses these
locations in combination with their surface normal as initial placements for a nose tip
template. To locations where this template fits well, a second template of global face
Fig. 1. Face segmentation. The depth image (left) is converted to a surface mesh (middle). The
surface mesh is cleaned, the tip of the nose is detected and the face segmented (right, in pink).
features is fitted to normalize the face’s pose and to select the tip of the nose. The face
is then segmented by removing the scan data with a Euclidean distance larger than 100
mm from the nose tip. These face segmentation steps are visualized in Fig. 1.
4 Face Model Fitting

In general, 3D range scans suffer from noise, outliers, and missing data and their resolu-
tion may vary. The problem with single face scans, such as the UND face scans, is that
large areas of the face are missing, which cannot be filled by simple hole filling tech-
niques. When the morphable face model is fitted to a 3D face scan, a model is obtained
that has no holes, has a proper topology, and has an assured resolution. By adjusting
the m=99 weights wi for the eigenvectors, the morphable model creates a new face
instance. To fit the morphable model to 3D scan data, we need to find the optimal set
of m weights wi . In this section, we describe a fully automatic method that efficiently
finds a proper model of the face scan in the m-dimensional space.
4.1 Distance Measure

To evaluate if an instance of the morphable face model is a good approximation of the
3D face scan, we use the Root Mean Square (RMS) distance of the instance’s vertices
to their closest points in the face scan. For each vertex point (p) from the instance (M1 ),
we find the vertex point (p ) in the scan data (M2 ) with the minimal Euclidean distance
emin (p, M2 ) = minp ∈M2 d(p, p ) , (1)
using a kD-tree. The RMS distance is then measured between M1 and M2 as:
?
@ n
@1
drms (M1 , M2 ) = A emin (pi , M2 )2 , (2)
n i=1
using n vertices from M1 . Closest point pairs (p,p ) for which p belongs to the bound-
ary of the face scan, are not used in the distance measure.
The morphable face model has n=75,972 vertices that cover the face, neck and ear
regions and its resolution in the upward direction is three times higher than in its side-
ways direction. Because the running time of our measure is dependent on the number of
vertices, we recreated the morphable face model such that it contains only the face (data
within 110 mm from the tip of the nose) and not the neck and ears. To obtain a more
uniform resolution of for the model, we reduced the upward resolution to one third of
the original model. The number of vertices of this adjusted morphable mean face is now
n=12,964 vertices, a sample every ≈2.6 mm2 of the face area.
4.2 Iterative Face Fitting

With the defined distance measure for an instance of our compressed morphable face
model, the m-dimensional space can be searched for the optimal instance. The fitting
is done by choosing a set of m weights wi , adjusting the position of the instance’s

vertices according to Sinst = S̄ + m i=1 wi σi si , measuring the RMS-distance of the
new instance to the scan data, selecting new weights and continue until the optimal
instance is found. Knowing that each instance is evaluated using a large number of
vertices, an exhaustive search for the optimal set of m weights is too computationally
expensive.
A common method to solve large combinatorial optimization problems is simulated
annealing (SA) [14]. In our case, random m-dimensional vectors could be generated
which represent different morphs for a current face instance. A morph that brings the
current instance closer to the scan data is accepted (downhill), and otherwise it is ei-
ther accepted (uphill to avoid local minima) or rejected with a certain probability. In
each iteration, the length of the m-dimensional morph vector can be reduced as imple-
mentation of the “temperature” scheme. The problem with such a naive SA approach
is that most random m-dimensional morph vectors are uphill. In particular close to the
optimal solution, a morph vector is often rejected, which makes it hard to produce an
accurate fit. Besides this inefficiency, it doesn’t take the eigensystem of the morphable
face model into account.
Instead, we propose an iterative downhill walk along the consecutive eigenvectors
from a current instance towards the optimal solution. Starting from the mean face S̄
(∀mi=1 wi = 0), try new values for w1 and keep the best fit, then try new values for w2
and keep the best fit, and continue until the face is morphed downhill along all m eigen-
vectors. Then iterate this process with a smaller search space for wi . The advantage
in computation costs of this method is twofold. First, the discrete number of morphs
in the selected search space directly defines the number of rejected morphs per itera-
tion. Second, optimizing one wi at a time means only a one (instead of m) dimensional
modification of the current face instance Snew = Sprev + (wnew − wprev )σi si .
Because the first eigenvectors induce the fitting of global face properties (e.g. face
height and width) and the last eigenvectors change local face properties (e.g. nose length
and width), each iteration follows a global to local fitting scheme (see Fig. 2). To avoid
local minima, two strategies are applied. (1) The selected wi in one iteration is not
evaluated in the next iteration, forcing a new (similar) path through the m-dimensional
space. (2) The vertices of the morphable face model are uniformly divided over three
Fig. 2. Face morphing along eigenvectors starting from the mean face (center column). Differ-
ent weights for the principal eigenvectors (e.g. i=1,2) changes the global face shape. For latter
eigenvectors the shape changes locally (e.g. i=50).
sets and in each iteration a different set is modified and evaluated. Only in the first
and last iteration all vertices are evaluated. Notice that this also reduces the number of
vertices to fit and thus the computation costs.
The fitting process starts with the mean face and morphs in place towards the scan
data, which means that the scan data should be well aligned to the mean face. To do so,
the segmented and pose normalized face is placed with its center of mass on the center
of mass of the mean face, and finely aligned using the Iterative Closest Point (ICP)
algorithm [15]. The ICP algorithm iteratively minimizes the RMS distance between
vertices. To further improve the effectiveness of the fitting process, our approach is
applied in a coarse fitting and a fine fitting step.
4.3 Coarse Fitting

The more the face scan differs from the mean face S̄, the less reliable the initial align-
ment of the scan data to the mean face is. Therefore, the mean face is coarsely fitted to
the scan data by adjusting the weights of the first ten principal eigenvectors (mmax =10)
in a single iteration (kmax =1) with 10 different values for wnew =[-1.35, -1.05, . . . , 1.05,
1.35] as in Algorithm ModelFitting(S̄,scan). Fitting the model by optimizing the
first ten eigenvectors results in the face instance Scoarse , with global face properties
similar to those of the scan data. After that, the alignment of the scan to Scoarse is
further improved with the ICP algorithm.
4.4 Fine Fitting

Starting with the improved alignment, we again fit the model to the scan data. This time
the model fitting algorithm is applied using all eigenvectors (mmax =m) and multiple
iterations (kmax =9). In the first iteration of Algorithm ModelFitting(S̄,scan), 10
new weight values wnew are tried for each eigenvector, to cover a large range of facial
variety. The best wnew for every sequential eigenvector is used to morph the instance
closer to the face scan. In the following kmax -1 iterations only four new weight values
wnew are tried around wi with a range wrange equal to wincr of the previous iteration.
By iteratively searching for a better wi in a smaller range, the weights are continuously
optimized. Local minima are avoided as described in Sect. 4.2. The range of the first
iteration and the number of new weights tried in each next iteration were empirically
selected as good settings.
4.5 Multiple Components

Knowing that the morphable model was generated from 100 3D face scans, an increase
of its expressiveness is most likely necessary to cover a large population. To increase
the expressiveness, also Blanz and Vetter [6] proposed to independently fit different
components of the face, namely the eyes, nose, mouth, and the surrounding region.
Because each component is defined by its own linear combination of shape parameters,
a larger variety of faces can be generated with the same model. The fine fitting scheme
from the previous section was developed to be applicable to either the morphable face
model as a whole, but also to individual components of this model.
Algorithm 1. ModelFitting(Sinst to scan)

1: wrange = 1.5, wincr = 0.3
2: for k ← 1 to kmax do
3: select vertices (uniform subset of component)
4: for i ← 1 to mmax do
5: wmin = wi − wrange + 12 wincr
6: wmax = wi + wrange − 12 wincr
7: for wnew ← wmin to wmax do
8: morph Sinst with wnew
9: drms (Sinst , scan) smaller → keep wnew
10: undo morph
11: wnew = wnew + wincr
12: morph Sinst with wi ← best wnew
13: wrange = wincr , wincr = 12 wincr
14: return Sinst
Component selection. All face instances generated with the morphable model are
assumed to be in correspondence, so a component is simply a subset of vertices in the
mean shape S̄ (or any other instance). We define seven components in our adjusted
morphable face model (see Fig. 3). Starting with the improved alignment, we can in-
dividually fit each of the components to the scan data using the fine fitting scheme,
obtaining a higher precision of the fitting process (as shown in Sect. 6.1). Individual
components for the left and right eyes and cheeks were selected, so that our method ap-
plies to non-symmetric faces as well. The use of multiple components has no influence
on the fitting time, because the total number of vertices remains the same and only the
selected vertices are modified and evaluated.
Component blending. A drawback of fitting each component separately is that in-
consistencies may appear at the borders of the components. During the fine fitting, the
border triangles of two components may start to intersect, move apart, or move across
(Fig. 3). The connectivity of the complete mesh remains the same, so two components
moving apart remain connected with elongated triangles at their borders. We solve these
inconsistencies by means of a post-processing step, as described in more detail below.
(a) (b) (c) (d) (e)
Fig. 3. Multiple components (a) may intersect (b1), move apart (b2), or move across (b3).
Simulating a cylindrical scan (c) and smoothing the new border vertices (d) solves these
problems (e).
Knowing that the morphable face model is created from cylindrical range scans and
that the position of the face instance doesn’t change, it is easy to synthetically rescan the
generated face instance. Each triangle of the generated face instance Sfine is assigned
to a component (Fig. 3a). A cylindrical scanner is simulated, obtaining a cylindrical
depth image d(θ, y) with a surface sample for angle θ, height y with radius distance
d from the y-axis through the center of mass of S̄ (Fig. 3c). Basically, each sample is
the intersection point of a horizontal ray with its closest triangle, so we still know to
which component it belongs. The cylindrical depth image is converted to a 3D triangle
mesh by connecting the adjacent samples and projecting the cylindrical coordinates to

3D. This new mesh Sfine has a guaranteed resolution depending on the step sizes of
θ and y, and the sampling solves the problem of intersecting and stretching triangles.
However, ridges may still appear at borders where components moved across. There-
fore, Laplacian smoothing is applied to the border vertices and their neighbors (Fig.
3d). Finally, data further then 110 mm from the tip of the nose is removed to have the
final model Sfinal (Fig. 3e) correspond to the segmented face. In Sect. 6.1, we evaluate
both the single and multiple component fits.
5 Face Recognition
Our model fitting algorithm provides a clean model of a 3D face scan. In this section,
we use this newly created 3D geometry as input for two 3D face matching methods.
One compares facial landmarks and the other compares extracted contour curves.
Landmarks. All vertices of two different instances of the morphable model are as-
sumed to have a one-to-one correspondence. Assuming that facial landmarks such as
the tip of the nose, corners of the eyes, etc. are morphed towards the correct position in
the scan data, we can use them to match two 3D faces. So, we assigned 15 anthropo-
morphic landmarks to the mean face and obtain their new locations by fitting the model
to the scan data. To match two faces A and B we use the sets of c=15 corresponding
landmark locations:

c
dcorr (A, B) = dp (ai , bi ) , (3)
i=1
where distance dp between two correspondences ai and bi is the squared difference in

Euclidean distance e to the nose tip landmark pnt :
dp (ai , bi ) = (e(ai , pnt ) − e(bi , pnt ))2 . (4)
Contour curves. Another approach is to fit the model to scans A and B and use the new
clean geometry as input for a more complex 3D face recognition method. To perform
3D face recognition, we extract from each fitted face instance three 3D facial contour
curves, and match only these curves to find similar faces. The three curves were ex-
tracted and matched as described by ter Haar and Veltkamp [13].
In more detail, after pose normalization and the alignment of the face scan to both S̄
and Scoarse , a correct pose of the face scan is assumed and thus a correct pose of the
final face instance Sfinal . Starting from the nose tip landmark pnt , 3D profile curves can
Fig. 4. The similarity of two 3D faces is determined using one-to-one correspondences, with on
the left 15 corresponding landmarks and on the right 135 corresponding contour samples. The
optimal XY-, C-, and G-contour curves (inner to outer) were extracted, for which the G-contour
uses the (colored) geodesic distances. The line shown in black is one of the Np profiles.
be extracted by walking the 3D surface in different directions (radii in the XY-plane).

Samples along a profile from one face should correspond to samples along the same
profile on another face. In case two faces are identical, these samples have the same
Euclidean distance to the tip of the nose pnt . For different faces, these samples cause
a dissimilarity. The 3D face matching algorithm extracts Np =45 profiles curves and
extracts from each profile curve:
.
– One XY-sample, the location where the distance (x2 + y 2 ) to pnt equals r.
– One C-sample, the location where the curve length of the profile to pnt equals r.
– One G-sample, the location on the profile where the length of the shortest geodesic
path over the entire surface to pnt equals r.
The shortest geodesic paths were computed using the fast marching method [16]. The
combination of Np =45 XY-samples at the same distance r builds a XY-contour, simi-
larly a C-contour and a G-contour are constructed. Based on a training set of morphable
face instances, the curves that were found most distinctive were selected, namely the
XY-contour at r=34 mm, the C-contour at r=68 mm, and the G-contour at r=77 mm.
The information of each 3D face instance is now reduced to a set of 135 (3 × Np ) 3D
sample points, with one-to-one correspondence to the same set of 135 samples in a dif-
ferent face instance. The similarity of faces A and B is again defined by dcorr , with
c=135 correspondences.
6 Results
The results described in this section are based on the UND face scans. For each of the
953 scans we applied our face segmentation method (Sect. 3). Our face segmentation
method correctly normalized the pose of all face scans and adequately extracted the tip
of the nose in each of them. The average distance and standard deviation of the 953
automatically selected nose tips to our manually selected nose tips was 2.3 ±1.2 mm.
Model fitting was applied to the segmented faces, once using only a single compo-
nent and once using multiple components. Both instances are quantitatively evaluated
in Sect. 6.1, and both instances were used for 3D face recognition in Sect. 6.2.
6.1 Face Model Fitting

In this section we evaluate the face model fitting as follows. Each segmented face was
aligned to S̄ and the coarse fitting method of Sect. 4.3 was applied. After the improved
alignment of the scan data to Scoarse , the fine fitting method of Sect. 4.4 was applied
to either the entire face (one component) or to each of the individual components (mul-
tiple components). For a fair comparison the same post-processing steps (Sect. 4.5)
were applied to both Sfine instances. Fig. 5 shows qualitative better fits when multi-
ple components are used instead of a single component. Globally, by looking at the
more frequent surface interpenetration of the fitted model and face scan, which means
a tighter fit. Locally, by looking at facial features, such as the nose, lips and eyes. Note
that our fitting method correctly neglects facial hair, which is often a problem for 3D
face recognition methods.
To quantitatively evaluate the produced fits, we determined the RMS distance (Eq.
2) for each of the fitted models to their face scan drms (Sfinal , scan) and for the scan
data to the face instance drms (scan, Sfinal ). Points paired with boundary points are not
included, so that results report merely the measurements in overlapping face regions.
Results are reported over all 953 scans in Table 1. They show that our face morphing
method provides accurate alignments of the morphable face model to the scan data for
both the single component and multiple components. All results are in favor of multiple
component morphs. Important to know is that the segmented UND faces have approx-
imately twice the number of vertices compared to the fitted face model. Therefore, the
closest point distances are higher for the scan to model case.
Fig. 5. Fitted face models Sfinal based on a single component (1st and 3rd column) and multiple
components (2nd and 4rd column) to scan data in blue. Results from the front and side view,
show a qualitative better fit of the multiple components to the scan data. The last two subjects on
the right were also used in [7].
Table 1. The quantitative evaluation (in mm) of our face fitting method
measure outliers M1 → M2 min max mean sd

drms yes 1 component → scan 0.478 3.479 0.776 0.176
drms yes 7 components → scan 0.374 2.076 0.608 0.123
drms yes scan → 1 component 0.787 7.236 1.115 0.561
drms yes scan → 7 components 0.696 6.269 0.935 0.503
davr .depth yes scan ↔ 1 component 0.393 4.704 0.692 0.290
davr .depth yes scan ↔ 7 components 0.254 2.542 0.444 0.197
davr .depth no scan ↔ 1 component 0.393 2.379 0.656 0.183
davr .depth no scan ↔ 7 components 0.254 1.818 0.423 0.120
Table 2. Recognition rates and mean average precisions based on landmarks and contour curves
for single and multiple component fits.
features model fit RR MAP

landmarks 1 component 85.8% 0.872
landmarks 7 components 85.2% 0.862
contours 1 component 96.3% 0.952
contours 7 components 97.5% 0.967
Comparison. Blanz et al. [7] reported the accuracy of their model fitting method
using the average depth error between the cylindrical depth images of the input scan and
the output model. The mean depth error over 300 FRGC v.1 scans was 1.02 mm when
they neglected outliers (distance > 10 mm) and 2.74 mm otherwise. To compare the
accuracy of our method with their accuracy, we produced cylindrical depth images (as in
Fig. 3c) for both the segmented face scan and the fitted model and computed the average
depth error |dscan (θ, y)−dfinal (θ, y)| without and with the outliers. For the fitted single
component these errors davr .depth are 0.656 mm and 0.692 mm, respectively. For the
fitted multiple components these errors are 0.423 mm and 0.444 mm, respectively. So
even our single component fits are more accurate then those of Blanz et al.
Our time to process a raw scan requires ≈3 seconds for the face segmentation, ≈1
second for the coarse fitting, and ≈30 seconds for the fine fitting on a Pentium IV 2.8
GHz. Blanz method reported ≈4 minutes on a 3.4 GHz Xeon processor, but includes
texture fitting as well. Huang et al. [9] report for their deformation model a matching
error of 1.2 mm after a processing time of 4.6 minutes.
6.2 Face Recognition

As described in Sect. 5, we can use the 953 morphed face instances to perform 3D face
recognition. For this experiment, we computed the 953 × 953 dissimilarity matrix and
generated for each of the 953 queries a ranked list of face models sorted in decreasing
similarity. From these ranked lists, we computed the recognition rate (RR) and the mean
average precision (MAP). A person is recognized (or identified) when the face retrieved
on top of the ranked list (excluding the query) belongs to the same subject as the query.
For 77 subjects only a single face instance is available which cannot be identified, so
the RR is based on the remaining 876 queries. The mean average precision (MAP) of
the ranked lists are reported, to elaborate on the retrieval of all relevant faces, i.e. all
faces from the same subject.
Four 3D face recognition experiments were conducted, namely face recognition
based on landmark locations from the fitted single component and the fitted multiple
components, and based on contour curves from the fitted single component and the
fitted multiple components. Results in Table 2 show that the automatically selected an-
thropomorphic landmarks are not reliable enough for effective 3D face recognition with
85.8% and 85.2% recognition rates (RR). Notice that the landmarks obtained from the
single component fit perform better than those from the multiple component fit. This
is probably caused by three landmarks (outer eye corners and Sellion) lying close to
component boundaries, where the fitting can be less reliable.
The fitted face model is an accurate representation of the 3D scan data. This accuracy
allows the contour based method to achieve high recognition rates (see Table 2). For the
single component fits, the contour matching achieves a RR of 96.3% and for multiple
component fits even 97.5%. For a high recognition rate, only one of the relevant faces
in the dataset is required on top of each ranked list. The reported MAPs show that most
of the other relevant faces are retrieved before the irrelevant ones. Some of the queries
that were not identified, have a non-neutral expression (happy, angry, biting lips, etc.)
while its relevant faces have a neutral expression. A face recognition method invariant
to facial expressions, will most likely increase the performance even further.
Comparison. Blanz et al. [7] achieved a 96% RR for 150 queries in a set of 150 faces
(from the FRGC v.1). To determine the similarity of two face instances, they computed
the scalar product of the 1000 obtained model coefficients. Using a set of facial depth
curves, Samir et al. [17] reported a 90.4% RR for 270 queries in a set of 470 UND
scans. Mian et al. [18] reported a 86.4% RR for 277 queries in a set of 277 UND
scans.
7 Concluding Remarks
Where other methods need manual initialization, we presented a fully automatic 3D face
morphing method that produces a fast and accurate fit for the morphable face model to
3D scan data. Based on a global to local fitting scheme the face model is coarsely
fitted to the automatically segmented 3D face scan. After the coarse fitting, the face
model is either finely fitted as a single component or as a set of individual components.
Inconsistencies at the borders are resolved using an easy to implement post-processing
method. Our results show that the use of multiple components produces a tighter fit of
the face model to the face scan, but assigned anthropomorphic landmarks may lose their
reliability for 3D face identification. Face matching using facial contours, shows higher
recognition rates based on the multiple component fits then for the single component
fits. This means that the obtained 3D geometry after fitting multiple components has
a higher accuracy. With a recognition rate of 97.5% for a large dataset of 3D faces,
our model fitting method proves to produce highly accurate fits usable for 3D face
recognition.
Acknowledgements
This research was supported by the FP6 IST Network of Excellence 506766 AIM@-
SHAPE and partially supported by FOCUS-K3D FP7-ICT-2007-214993. The authors
thank the University of South Florida for providing the USF Human ID 3D Database.
References
1. Bowyer, K.W., Chang, K., Flynn, P.: A survey of approaches and challenges in 3D and multi-
modal 3D + 2D face recognition. CVIU 101(1), 1–15 (2006)
2. Scheenstra, A., Ruifrok, A., Veltkamp, R.C.: A Survey of 3D Face Recognition Methods.
In: Kanade, T., Jain, A., Ratha, N.K. (eds.) AVBPA 2005. LNCS, vol. 3546, pp. 891–899.
3. Bronstein, A.M., Bronstein, M.M., Kimmel, R.: Three-dimensional face recognition.
IJCV 64(1), 5–30 (2005)
4. Berretti, S., Del Bimbo, A., Pala, P., Silva Mata, F.: Face Recognition by Matching 2D and
3D Geodesic Distances. In: Sebe, N., Liu, Y., Zhuang, Y.-t., Huang, T.S. (eds.) MCAM 2007.
5. Davis, J., Marschner, S.R., Garr, M., Levoy, M.: Filling holes in complex surfaces using
volumetric diffusion. 3DPVT, 428–861 (2002)
6. Blanz, V., Vetter, T.: A morphable model for the synthesis of 3D faces. SIGGRAPH, 187–194
(1999)
7. Blanz, V., Scherbaum, K., Seidel, H.P.: Fitting a Morphable Model to 3D Scans of Faces. In:
ICCV, pp. 1–8 (2007)
8. Lu, X., Jain, A.: Deformation Modeling for Robust 3D Face Matching. PAMI 30(8), 1346–
1356 (2008)
9. Huang, X., Paragios, N., Metaxas, D.N.: Shape Registration in Implicit Spaces Using Infor-
mation Theory and Free Form Deformations. PAMI 28(8), 1303–1318 (2006)
10. Kakadiaris, I., Passalis, G., Toderici, G., Murtuza, N., Theoharis, T.: 3D Face Recognition.
In: BMVC, pp. 869–878 (2006)
11. Sarkar.S.: USF HumanID 3D Face Database. University of South Florida
12. Chang, K.I., Bowyer, K.W., Flynn, P.J.: An Evaluation of Multimodal 2D+3D Face Biomet-
rics. PAMI 27(4), 619–624 (2005)
13. ter Haar, F.B., Veltkamp, R.C.: A 3D Face Matching Framework. In: Proc. Shape Modeling
International (SMI 2008), pp. 103–110 (2008)
14. Kirkpatrick, S., Gelatt, C.D., Vecchi, M.P.: Optimization by Simulated Annealing. Sci-
ence 220, 4598, 671–680 (1983)
15. Besl, P.J., McKay, N.D.: A method for registration of 3D shapes. PAMI 14(2), 239–256
(1992)
16. Kimmel, R., Sethian, J.: Computing geodesic paths on manifolds. Proc. of National Academy
of Sciences 95(15), 8431–8435 (1998)
17. Samir, C., Srivastava, A., Daoudi, M.: Three-Dimensional Face Recognition Using Shapes
of Facial Curves. PAMI 28(11), 1858–1863 (2006)
18. Mian, A.S., Bennamoun, M., Owens, R.: Matching Tensors for Pose Invariant Automatic 3D
Face Recognition. IEEE A3DISS (2005)
A Multi-scale Vector Spline Method for
Estimating the Fluids Motion on Satellite
Images
Till Isambert1,2 , Jean-Paul Berroir1,2, and Isabelle Herlin1,2

1
INRIA, Domaine de Voluceau, Rocquencourt, BP 105 - 78153 Le Chesnay Cedex,
France
2
CEREA, Joint Laboratory ENPC - EDF R&D, Université Paris-Est, France
Abstract. Satellite image sequences visualize important patterns of the

atmospheric and oceanographic circulation. Assessing motion from these
data thus has a strong potential for improving the performances of the
forecast models. Representing a vector field by a vector spline has been
proven efficient for fluid motion assessment: the vector spline formulation
makes it possible to initially select the locations where the conservation
equation has to be taken into account; it efficiently implements the 2nd
order div-curl regularity, advocated for turbulent fluids. The scientific
contribution of this article is to formulate vector splines in a multiscale
scheme, with the double objective of assessing motion even in the case of
large displacements and capturing the spectrum of spatial scales associ-
ated to turbulent flows. The proposed method only requires the inversion
of a band matrix, which is performed by an efficient numerical scheme
making the method tractable for large satellite image sequences.
1 Introduction
Meteorological and oceanographic forecasts are based on the numerical solving

of 3D circulation models that require an initial condition of their state variables,
among which velocity, i.e. winds and currents. The estimation of the initial
condition relies on real measurements ingested into models by means of data
assimilation [1]. Operational forecast models make use of in situ observations
for this purpose, provided by ground stations, balloons, drifting buoys, etc. One
challenge of environmental modelling is the ability to derive motion estimates
from dynamic satellite data in order to complement in situ data. Atmospheric
and oceanographic satellite images constitute a 2D visualization of a 3D fluid
flow. The link between the 2D motion and the 3D flow is non trivial: for instance,
on water vapor meteorological images, the pixel value results from contributions
of water particles in the vertical column, hence the 2D image motion is a com-
plex combination of real 3D motions. Correlation-based methods constitute the
current operational standard for assessing image motion from satellite imagery
(the so-called Atmospheric Motion Winds). Their applicability to environmen-
tal forecast is however limited owing to the lack of reliable tracers identifiable

666 T. Isambert, J.-P. Berroir, and I. Herlin
on images and to uncertainties affecting their height assignment. Despite these

limitations, image motion fields are considered to be an essential data source for
environmental forecast.
Assessing image motion requires defining an image processing model, consti-
tuted of a 2D conservation law and regularity properties. State-of-the-art models
for fluid flow estimation are characterized by the three following properties. (1)
The conservation equation applies either to the luminance [2] or to the mass [3,4].
(2) The regularity is modelled by the 2nd order div-curl constraint, which pro-
vides direct control of the Helmholtz decomposition of the motion field in terms
of divergence and vorticity [5]. A recent study [6] formulates the regularity with ef-
fective physical constraints: the retrieved flow is supposed steady and satifying the
incompressible Stokes equations. The applicability to non steady flows satisfying
the Navier-Stokes equations remains however open. (3) Multiscale coarse-to-fine
numerical schemes [7,8,9,10] are implemented in the case of large displacements,
that prevent the linearization of the conservation equation. The multiscale scheme
is further needed in the case of highly turbulent flows in which motion patterns
co-exist at different spatial scales.
Solving the image processing model is usually addressed by a gradient-based
minimization of an energy functional made up of two components: the confidence
in data and the regularity of the result. The data confidence is often the integral
over the image domain of the residual of the conservation equation. Applying
such an approach to satellite atmospheric or oceanographic sequences faces two
main difficulties. First, the conservation equation can be locally incorrect -3D
motions such as upwelling or convection constitute typical examples in which the
2D conservation equation is not valid- or unworkable -missing data, occlusions by
clouds, absence of contrast, motion parallel to contours. Accounting for the con-
servation equation in the whole image domain therefore leads to erroneous flow
estimates. Second, the iterative minimization of the 2nd order div-curl regularity
constraint leads to 4th order PDEs, hence to complex iterative minimization not
guaranteeing the convergence to a minimum.
An elegant solution to these two problems is provided by spline-based meth-
ods. First, they make it possible to easily formulate models in which data confi-
dence applies only at selected locations: the so-called control points. Second, they
do not require an iterative energy minimization. Vector-valued thin-plate splines
minimize the 2nd order div-curl regularity constraint and have been proved ef-
ficient to interpolate or approximate wind measurements provided by ground
stations [11]. The use of thin-plate vector splines for image motion estimation
has been further proposed by [12], who formulated a vector spline model in
which the 2D conservation law is satisfied at control points. Thin-plate splines
are, however, unsuitable for multiscale estimation: they are defined from a har-
monic radial basis function that increases with the distance to the control point
and hence does not provide a local information.
This paper presents an innovative multiscale formulation of the vector spline
approach for motion estimation. It keeps the nice properties of vector splines (use
of control points, div-curl regularity) but makes use of a parametric spline model
A Multi-scale Vector Spline Method 667
based on a compactly supported and rapidly decaying radial basis function, thus
adapted to multiscale representation. The solution is obtained by solving a sparse
and well-conditioned linear system. The motion is computed on a pyramidal
representation of images, as the sum of a coarse scale motion and increments from
one scale to the immediately finer one. Results are presented to demonstrate the
effectiveness of the characteristics of the multiscale vector spline: use of control
points, div-curl regularity and multiscale coarse-to-fine motion estimation.
This paper is organized as follows: the section 2 recalls the vector spline
theory applied to fluid motion estimation; the proposed multiscale vector spline
is presented in section 3. Results are analyzed in section 4, and conclusions and
prospects for future work are given in section 5.
2 Apparent Motion Estimation Using Vector Splines
Vector splines have been initially introduced [11] for the interpolation and ap-
proximation of vector observations. In this context, the vector spline model is
defined from: (1) a set of n control points xi in a spatial domain Ω; (2) a vec-
tor observation wi at each control point. The vector spline is solution of the
following minimization problem:
⎧ 1
⎨
min w2d
or: min (w(xi ) − wi ) + λ
2
wd2
⎩ w(x )Ω= w ∀i Ω (1)
i i i
Interpolation Approximation
In equation (1), the parameter λ of the approximating spline controls the com-
promise between regularity and confidence in data, wd denotes the 2nd order
div-curl semi-norm, defined as:
w2d = α∇div w2 + β∇curl w2 (2)
It is a semi-norm which zero-space is the set of affine vector fields. It has been
proven [11] that this minimization problem admits a unique solution: a thin-plate
spline based on the harmonic radial basis function φ:
−1
φ(x) = (128π) x4 log x (3)
The spline is defined by:

5
u = p(x) + ni=1 ai ( α1 φxx (x − xi ) + β1 φyy (x − xi )) + bi ( α1 − β1 )φxy (x − xi )
n
v = q(x) + i=1 ai ( α1 − β1 )φxy (x − xi ) + bi ( α1 φyy (x − xi ) + β1 φxx (x − xi ))
(4)
with p and q degree 1 polynomials. The coefficients of the spline (ai , bi and the
6 coefficients in p and q) are obtained by solving a linear system.
The use of vector splines for motion estimation from image data has been
proposed in [12] for the luminance conservation equation, and in [13] for mass
conservation. In contrast to the previous formulation, only an indirect obser-

vation is provided by the conservation equation. The latter is formulated as:
Lw + It = 0, L being a linear operator. If assuming luminance conservation,
we have Lw = ∇I.w, and for mass conservation: Lw = (∇I + I∇).w. For both
cases, at each control point xi , we have Li .w = −It , Li denoting the observation
operator at control point xi .
The vector spline model is then rewritten as:
⎧ 1
⎨
min w2d
or: min (Li w + It )2 + λ w2d (5)
⎩ L w +Ω I = 0 Ω
i t i
for the interpolation and approximation cases. It has been proven [12] that the
solution of (5) exists and is unique if the observation operators Li are linear and
non zero, and if the control points are non aligned. The solution is a thin-plate
spline, with the same basis function φ as in equation (3):

n
6
w= ci Li φ(x − xi ) + di pi (x) (6)
i=1 i=1
p = (pi ) being the canonical basis of degree 1 polynomials. The vectors of

coefficients c = (ci ) and d = (di ) are solution of:

Q + λId S c −It
= (7)
ST 0 d 0
Id being the identity matrix. The general term of Q (size n × n) is Li Lj φ(xi −

xj ) and S (size n × 6) is defined as: S = Li p.
The vector spline model used for image motion estimation shares the same
interesting properties as the model used for interpolating or approximating vec-
tor data: its solution is obtained without iterative minimization and it has few
parameters (i.e. α, β and λ). One will set α * β when processing images of
highly turbulent flows, thus favoring an estimation with large curl variability.
Reversely, β * α is appropriate for situations such as intense vertical activity
causing locally divergent 2D flows.
One main difference with vector data is that there are no predefined loca-
tions where observations are available. The control points selection process must
be defined prior to the estimation. The theoretical conditions guaranteeing the
existence and uniqueness of the solution can easily be met by selecting these
control points with a double thresholding: first on the magnitude of the spa-
tial gradient, discarding low contrast areas; second on the magnitude of the
motion index (ratio of temporal and spatial gradient), thus discarding visually
still areas. Depending on the applicative domain, further locations in specific
structures must be discarded. This is, for instance, the case of ejection filaments
in oceanography, as in these structures the motion is parallel to the contours
and the conservation equation is degenerated. The selection of control points
is therefore highly application-dependent; they should be as evenly distributed
as possible, but there is, to our knowledge, no criterion for defining an optimal
distribution of control points.
3 Multiscale Vector Splines

3.1 Parametric Spline Model
Thin-plate vector splines minimize the 2nd order div-curl regularity, but are
inappropriate for multiscale estimation as they are defined from a harmonic
basis function. A multiscale scheme actually requires using a basis function that
provides a local representation, hence locally supported or rapidly decaying. The
contribution of this paper is to formulate a multiscale model, based on a spline
parameterized by the scale value and on a pyramidal representation of images
at different scales.
We consider the spline approximation problem with the 2nd order div-curl
norm and either the luminance or the mass conservation equation, through the
observation operators Li assessed on the n control points xi :
n 1

min J(w) = (Li w − wi )2 + λ α∇div w2 + β∇curl w2 (8)
i=1 Ω
Rather than exactly solving equation (8), which would lead to the thin-plate
spline, the minimum is searched for among a set of spline functions suitable for
the multiscale formalism and satisfying the two following properties. (1) The
spline is defined from a unique bell-shaped radial basis function of unit support.
The choice of this function is not critical as long as it is positive, decreasing
and at least three times continuously differentiable in order to compute the 2nd
order div-curl semi-norm. We make use of the basis function ψ proposed by [14]
and defined as ψ(r) = (1 − r)6 (35r2 + 18r + 3) for |r| ≤ 1. (2) The spline is a
linear combination of translates of the basis function over a regular lattice of m
grid points, whose sampling defines the scale parameter h. These translates are
dilated by a factor γ proportional to h. The parameters defining the spline are
the m weights q = (qj ) (each weight qj being homogeneous to a motion vector
with u and v components) applied to the translates of the basis function. The
parametric expression of the vector spline is thus:
x − hvj
wq,h (x) = qj ψ( ) (9)
γ
vj ∈Z2 ,hvj ∈Ω
where vj spans a regular lattice of unit spacing in the image domain Ω.

A new expression of the functional J is defined by substituting, in equation (8),
w by its parametric form wq,h (9). Let us first consider the first term of J. If
the observation operator is based on the luminance conservation equation, its
new expression becomes:
Ix Ψ qu + Iy Ψ qv − It 2 = Al q − It 2 (10)

It being the n-dimensional vector of the temporal derivatives at the control

points; Ψ being the n × m matrix of general term ψ((xi − kj )/γ, (yi − lj )/γ) with
i indexing the n control points and j the m grid points (k, l); Ix and Iy are the
n × n diagonal matrices of the image spatial derivatives at the control points. In
the case of mass conservation, the first term of J becomes:
Ix Ψ qu + Iy Ψ qv + IDx Ψ + IDy Ψ − It 2 = Am q − It 2 (11)
where I is the n × n diagonal matrix formed by the image values at control

points, Dx Ψ and Dy Ψ are the matrices of the spatial derivatives of Ψ . Whatever
the conservation equation, the first term of J is then rewritten as a quadratic
function of q.
Let us now analyze the second term of J. By introducing the matrix of dif-
ferential operators Q(D):
√ √ √ √
α∂xx +
√ √ β∂yy √( α − √ β)∂xy
Q(D) = (12)
( α − β)∂xy α∂yy + β∂xx
J then factorizes as:

α ∇div w2 + β ∇curl w2 = Q(D)w2 (13)
The second term of J is finally rewritten as the quadratic expression Rq2 ,

with:
√ √ √ √
α∂
√ xx Ψ √
+ β∂yy Ψ √( α − √ β)∂xy Ψ
R= . (14)
( α − β)∂xy Ψ α∂yy + β∂xx Ψ
The substitution of w by the parametric expression wq,h allows J to be rewritten
as a quadratic function of q:
J(q) = Aq − It 2 + λRq2 (15)
with A being either Al or Am depending on the conservation equation chosen.

Finding the minimum of J with respect to q is now a linear optimization problem.
The matrices A and R, in (15), have a band structure since ψ has a compact
support of size γ. The width of the band depends on the ratio of γ to the scale
parameter h. If γ is smaller than h, the matrices A and R are diagonal and the
vector spline is zero everywhere except in the vicinity of the grid points. If γ
is large compared to h, the resulting vector spline can accurately approximate
the thin-plate spline, but the A and R matrices are dense and require a heavy
computational load. γ = 3h has been empirically chosen as a good compromise
between the computational speed and the accuracy of the spline. The band
structure allows an efficient numerical solving to be implemented.
3.2 Hierarchical Motion Estimation

A multiscale scheme is required for two main reasons. (1) The parametric spline
allows the image motion to be assessed, given a spatial scale parameter h, and
provided that the conservation equation can be computed. On satellite image

sequences, a too strong motion and/or a too coarse time sampling cause large
displacements between successive frames, preventing the linearization of the con-
servation equation. (2) Turbulent flows are associated with a large spectrum of
spatial and temporal scales. We therefore make use of a pyramidal scheme, in
which motion is hierachically computed from the coarsest to the finest scale.
Let I0 and I1 be two successive images of the sequence. Both are represented
using a pyramid, from the full resolution I0 (0) and I1 (0) to the coarsest scale
I0 (pmax ) and I1 (pmax ). To each index p corresponds a scale parameter h(p). The
motion is initially computed at the coarsest scale with the parametric spline at
scale h(pmax ), yielding the motion field w(pmax ). This initial coarse motion field
is then progresively refined at each scale h(p) by first compensating the image
I0 (p) with w(p + 1) and computing the motion increment δw(p) between the
compensated image and I1 (p). The finest scale motion (p = 0) is thus expressed
as the sum of the coarse scale motion w(pmax ) and of the increments describing
the finer resolutions:

0
w(0) = w(pmax ) + δw(p) (16)
p=pmax −1
The link between the scale parameter h(p) and the real spatial scale of the evolv-
ing image structures is not obvious: at one level of the pyramid, the motion is
computed using a scale parameter h(p) corresponding to a basis function of sup-
port γ = 3h(p). The basis function is thus able to represent motion patterns with
spatial size less than 3h(p); but there is no guarantee that all motion patterns of
that size will be represented: this will occur only if enough control points have
been selected in the existing patterns.
4 Results
The first result intends to demonstrate the efficiency of accounting for the con-
servation only at control points. For this purpose, the motion is computed using
Fig. 1. Extract of the test sequences: left, OPA; right: Meteosat

the multiscale vector spline and compared to the result of Corpetti’s method [15].
Both methods minimize the second order div-curl regularity constraint, make use
of either luminance or mass conservation and are solved in a multiscale scheme.
The two methods differ in the data confidence term of the minimized energy
(computed on control points selected by double thresholding for the multiscale
spline, on the whole image domain for Corpetti’s method) and in the numerical
minimization scheme (multiscale vector spline vs variational minimization). Two
comparisons are displayed. First, the motion is computed using the luminance
conservation equation on the synthetic ’OPA’ sequence (on the left in figure 1),
obtained by numerical simulation with the OPA ocean circulation model1 . The
2 2 0.3
1.5 1.5
0.2
1 1
0.5 0.5
0.1
0 0
0
−0.5 −0.5
−1 −1
−0.1
−1.5
−1.5
−2 −0.2
−2
Fig. 2. Motion fields estimated on the OPA sequence using luminance conservation.
Left to right: reference motion, multiscale spline, Corpetti and Mémin. Top to bottom:
motion field, streamlines, vorticity.
1
Thanks to Marina Levy, LOCEAN, IPSL, France.
Fig. 3. Motion fields estimated on the Meteosat sequence using mass conservation. Left:
multiscale spline, right: Corpetti and Mémin. Top to bottom: motion field, streamlines.
OPA sequence consists of simulated images of sea surface temperature, used for
computing motion. Additionaly the corresponding surface currents are available
and used as the reference field for validation purposes. The results are displayed
on figure 2. The mean angular error between the estimated and reference mo-
tion fields is 28 degrees for the multiscale spline and 42 degrees for Corpetti’s
method. The qualitative inspection of the motion field’s streamlines and vorticity
suggest that the motion of vortices is better assessed by the multiscale spline.
A similar comparison on a Meteosat-5 sequence2 acquired in the water vapor
band is displayed on figure 3. The mass conservation equation is used as the 2D
atmospheric flow can be considered as compressible to accomodate the effects
of vertical motion. For this sequence, a sole qualitative assessment of results is
possible. The multiscale spline is more accurate with respect to the location of
the central vortex. It furthermore succeeds in capturing a rotating motion in the
lower left part of the image, whereas Corpetti’s method incorrectly computes a
smooth laminar field.
The second comparison is intended to demonstrate that the 2nd order div-curl
regularity must be preferred to L2 regularity for fluid motion assessment. The lu-
minance conservation equation is considered and the motion is computed on the
OPA sequence by the multiscale spline and the Horn and Schunck method [2].
The results are displayed on figure 4. Three different results are presented corre-
sponding to different values of the λ coefficient assigned to the regularity compo-
nent, so that both methods are tested with low, medium and high regularization.
The angular errors for the multiscale spline are 30, 29 and 28 degrees (respec-
tively for low, medium and high regularity), for the Horn and Schunk method
43, 47 and 49 degrees. The spline method is much more efficient as far as the
detected location of eddies is concerned: only one vortex is detected by H&S
2
Copyright Eumetsat.
Fig. 4. Comparison of motion estimates on the OPA sequence, luminance conservation.

Top: multiscale spline with low, medium and high regularity; bottom: Horn and Schunk
with low, medium and high regularity.
method with low regularity, and none with medium and high regularity. This is
a consequence of the L2 regularization which favours laminar fields.
Figure 5 displays the motion fields estimated on the OPA sequence at three
different scales. At the coarsest scale, the main vortices appear in the upper part
of the image, and the large vortex in the bottom part is not detected at all. At
the intermediate scale, more vortices appear. At finest resolution the location
of vortices is improved and the large vortex in the bottom part of the image is
Fig. 5. Motion field estimated on the OPA sequence, from the coarsest to the finest
(full) resolution
even detected. This illustrates that the multiscale scheme actually links the size
of the spatial structure with the spatial scale of the spline, although this link is
not easy to interpret.

This paper proposes an innovative approach for assessing motion on atmospheric
and oceanographic satellite image sequences of fluid flows. Its characteristics are
the following. (1) The data confidence is only taken into account at specific
locations, the control points, where the conservation equation is workable, i.e.
valid in 2D, or not degenerated to 0=0. The comparison made with a method
using the same conservation equation and the same regularity constraint, but
computing the conservation equation over the whole image domain, proves that
using control points is essential for assessing the confidence in data. (2) The
regularity of the motion field is controlled by the 2nd order div-curl constraint.
This is especially adapted to fluid and turbulent flows, as it allows the variations
of divergence and vorticity of the retrieved motion field to be controlled. This
is highlighted by comparisons with methods based on L2 regularization, which
are unable to correctly assess the motion field in eddies. (3) A multiscale scheme
is used to to allow estimation even in the case of large displacements and to
provide a hierarchical representation of the motion field, distinguishing between
large and fine scale structures such as vortices.
Being based on control points and on the 2nd order div curl constraint, the
motion field should theoretically be expressed by a thin-plate spline. The theo-
retical contribution of this paper is to formulate a vector spline in a multiscale
scheme, using a parametric spline model. This approach has two main advan-
tages: (1) the minimum of the energy is obtained by solving a linear system; (2)
the matrix to invert has a band structure, allowing an efficient implementation
that makes the method affordable even for large sequences.
The approach requires further investigation on two issues. First, what is the
link between the spatial scale of basis functions and the extent of image struc-
tures such as vortices? There is no obvious link according to the results and
we believe that the control points must also be hierarchically organized for that
purpose. Second, it is well known that spatial and temporal scales of turbulent
flows are strongly linked, large vortices being stabler than smaller ones. It is thus
necessary to consider motion estimation not only from a pair of successive im-
ages, but from a longer sequence in order to establish a multiscale representation
both in space and time.
References
1. Korotaev, G., Huot, E., Le Dimet, F.X., Herlin, I., Stanichny, S., Solovyev, D.,
Wu, L.: Retrieving Ocean Surface Current by 4D Variational Assimilation of Sea
Surface Temperature Images. Remote Sensing of Environment (2007) (Special Issue
on Data Assimilation)
2. Horn, B., Schunck, B.: Determining optical flow. AI 17(1-3), 185–203 (1981)
3. Béréziat, D., Herlin, I., Younes, L.: A generalized optical flow constraint and its
physical interpretation. In: CVPR 2000, pp. 487–492 (2000)
4. Wildes, R., Amabile, M.: Physically based fluid flow recovery from image sequences.
In: CVPR 1997, Puerto Rico, pp. 969–975 (June 1997)
5. Gupta, S., Princ, J.: Stochastic models for div-curl optical flow methods. IEEE
Signal Processing Letter 3(2) (1996)
6. Ruhnau, P., Schnoerr, C.: Optical Stokes Flow Estimation: An Imaging-based Con-
trol Approach. Experiments in Fluids 42, 61–78 (2007)
7. Anandan, P.: A computational framework and an algorithm for the measurement
of visual motion. International Journal of Computer Vision 2, 283–310 (1989)
8. Bergen, J.R., Anandan, P., Hanna, K.J., Hingorani, R.: Hierarchical model-based
motion estimation. In: Sandini, G. (ed.) ECCV 1992. LNCS, vol. 588, pp. 237–252.
9. Enkelmann, W.: Investigation of multigrid algorithms for the estimation of optical
flow fields in image sequences. Computer Vision Graphics and Image Process-
ing 43(2), 150–177 (1988)
10. Moulin, P., Krishnamurthy, R., Woods, J.: Multiscale modeling and estimation of
motion fields for video coding (1997)
11. Amodei, L.: A vector spline approximation. Journal of approximation theory 67,
51–79 (1991)
12. Suter, D.: Motion estimation and vector splines. In: CVPR 1994 (1994)
13. Isambert, T., Herlin, I., Berroir, J., Huot, E.: Apparent motion estimation for
turbulent flows with vector spline interpolation. In: XVII IMACS, Scientific Com-
putation Applied Mathematics and Simulation, Paris, July 11-15 (2005)
14. Wendland, H.: Piecewise polynomial, positive definite and compactly supported ra-
dial basis functions of minimal degree. Advances in Computational Mathematics 4,
389–396 (1995)
15. Corpetti, T., Memin, E., Perez, P.: Dense estimation of fluid flows. PAMI 24(3),
365–380 (2002)
Continuous Energy Minimization
Via Repeated Binary Fusion
Werner Trobin1 , Thomas Pock1,2 , Daniel Cremers2 , and Horst Bischof1

1
Institute for Computer Graphics and Vision, Graz University of Technology
2
Department of Computer Science, University of Bonn
Abstract. Variational problems, which are commonly used to solve low-

level vision tasks, are typically minimized via a local, iterative optimiza-
tion strategy, e.g. gradient descent. Since every iteration is restricted to
a small, local improvement, the overall convergence can be slow and the
algorithm may get stuck in an undesirable local minimum. In this paper,
we propose to approximate the minimization by solving a series of bi-
nary subproblems to facilitate large optimization moves. The proposed
method can be interpreted as an extension of discrete graph-cut based
methods such as α-expansion or LogCut to a spatially continuous set-
ting. In order to demonstrate the viability of the approach, we evaluated
the novel optimization strategy in the context of optical flow estimation,
yielding excellent results on the Middlebury optical flow datasets.
1 Introduction
Several fundamental problems in computer vision can be classified as inverse, ill-
posed problems, where a direct solution is not possible (e.g. deblurring, stereo,
optical flow). In such cases, a prior model of the forward process can help to infer
physically meaningful solutions via a maximum a posteriori (MAP) estimation.
Such MAP formulations naturally lead to energy minimization problems [1],
where an energy term Eprior , representing the prior model, penalizes unlikely
solutions and a data consistency term Edata enforces a close fit to the observed
data:
min {Eprior (u) + λEdata (u)} . (1)
u
Since we are dealing with spatially (and radiometrically) discrete images, at some
point any optimization approach for (1) has to take the spatial discretization into
account – there are two predominant strategies to do that. One currently very
popular approach is to state the problem as a discrete, combinatorial optimiza-
tion problem on a Markov Random Field (MRF). Since MRFs are a powerful tool
for solving most low level vision tasks, a considerable research effort has been
dedicated to exploring minimization methods for MRF energies (cf. [2] for a com-
parison of state-of-the-art algorithms). Generally, the optimization approaches

This work was supported by the Austrian Science Fund under grant P18110-B15,
the Austrian Research Promotion Agency within the VM-GPU project (no. 813396),
and the Hausdorff Center for Mathematics.

678 W. Trobin et al.
are either based on message passing (e.g. loopy belief propagation by Pearl [3]
and sequential tree-reweighted message passing by Kolmogorov [4]) or on graph
cuts (α-β-swap and α-expansion, introduced by Boykov et al. [5], and the more
recent “range moves” by Veksler [6] and LogCut by Lempitsky et al. [7]). Re-
cently, Komodakis et al. proposed a fast optimization approach, based on the
duality theory of Linear Programming [8]. The second optimization strategy for
(1) uses the tools of the calculus of variations in a continuous context. Once
an optimality condition for the energy is derived, the differential operators are
discretized and a numerical scheme is used to minimize the energy. In contrast
to the aforementioned discrete MRF approach, in the variational approach the
discretization is postponed as long as possible.
This work will focus on limitations of the local optimization approaches used
in the variational context. In order to circumvent these limitations, we will in-
troduce a novel optimization strategy inspired by the discrete α-expansion and
LogCut algorithms. In contrast to local methods, such an optimization strategy
allows large moves and therefore is less likely to get stuck in bad local minima.
Unlike combinatorial optimization approaches, the solution space does not have
to be discretized and the algorithm does not induce systematic metrication er-
rors. The proposed variational technique also facilitates high-performance imple-
mentations on massively parallel GPUs and permits an extension to higher-order
“interactions” at low costs in time and memory.
After proposing the novel optimization strategy in Section 2, we evaluate this
technique for optical flow estimation in Section 3. Experiments in Section 4
illustrate state-of-the-art results.
2 Energy Minimization by Solving Binary Subproblems

In order to simplify the discussion, we will only consider continuous energy
functionals of the form
5 6

min Ψ Du(x), D2 u(x), . . . dx + λ ρ(u(x), x) dx , (2)
u Ω Ω
where Ω ⊂ IR and u : Ω → IR . The lefthand term, which corresponds

n m
to Eprior in (1), penalizes non-smooth solutions via a potentially robust error

function Ψ , while the righthand term enforces data consistency, and the free pa-
rameter λ allows balancing the influence of the two terms. Solving such equations
in a continuous setting entails deriving the Euler-Lagrange equations and find-
ing an iterative numerical scheme to calculate the solution. These schemes are
typically based on gradient descent and therefore every iteration will only result
in a local improvement of the current solution. Hence, these methods are prone
to get stuck in bad local minima. Moreover, the speed of convergence might be
quite slow. Combinatorial optimization approaches for MRFs, on the other hand,
are not restricted to small, local moves. For a lot of MRF energies, α-expansion
seems to offer a good trade-off between runtime and optimality properties. The
central idea of the α-expansion algorithm is to approximately minimize an (NP-
hard) multi-label problem, by repeatedly solving a series of binary subproblems.
Continuous Energy Minimization Via Repeated Binary Fusion 679
For each subproblem, the global optimum can efficiently be computed using a
max-flow/min-cut algorithm [5].
Inspired by [9], where Chambolle shows close links between the Total Variation
(TV) model and binary MRFs in the context of image denoising, and [10], where
Nikolova et al. show how to find globally optimal solutions for certain nonconvex
optimization problems by restating them as convex problems, we will pose the
minimization problem (2) as a sequence of binary subproblems. Each of these
subproblems can conceptually be understood as a continuous version of an α-
expansion move, i.e. the current solution is changed to a proposed alternative
solution, wherever this is energetically favorable. Repeatedly solving this binary
problem for varying proposed solutions, i.e. performing cycles, as it is called in
the α-expansion terminology, yields increasingly accurate results.
In order to formulate (2) as a binary problem, let α, β : Ω → IRm be two
arbitrary but fixed candidate solutions we will subsequently refer to as “proposed
solutions” or just as “proposals.” Moreover, let φ : Ω → {0, 1} be a binary
function to selectively combine α and β to a new solution
u(x) = (1 − φ(x)) α(x) + φ(x)β(x) . (3)
The function φ is free to vary across Ω, as long as the fused solution u fulfills the
regularity requirements posed by the considered energy. Plugging the combined
solution (3) into the model (2) yields
5

min Ψ Du(φ(x), x), D2 u(φ(x), x), . . . dx + (4)
φ∈F Ω
6
λ (1 − φ(x)) ρ(α(x), x) + φ(x)ρ(β(x), x) dx
Ω
with F = {φ : Ω → {0, 1}}, where we used the observation that
ρ [(1 − φ(x)) α(x) + φ(x)β(x), x] = (1 − φ(x)) ρ[α(x), x] + φ(x)ρ [β(x), x] .

(5)
Therefore, instead of directly estimating the solution u, the goal is now to find
the best possible fusion of the two fixed proposals α and β, by optimizing with
respect to φ. This construction is strongly related to the generalized α-expansion
presented in [7], which comprises the core of the LogCut algorithm.
Unfortunately, the binary problem is not convex and hence hard to solve in
a continuous context. Therefore, similar to [10], we first expand the range of φ
to the interval [0, 1] and then solve this relaxed problem in a continuous set-
ting. Despite the fact that such a relaxation clearly invalidates (5), the observed
continuous solutions are close to binary (see Fig. 3). This indicates that in the
vast majority of the domain there is a clear preference for either α or β, so we
argue that thresholding the continuous solution φ leads to a reasonable binary
fusion of α and β. To avoid a solution procedure that merely oscillates between
different proposals, we only accept a fused result, if the energy (4) decreases,
compared to using either α or β alone.
3 Application to Optical Flow Estimation

Estimating the optical flow between two consecutive images of a scene requires
finding corresponding points in the images. As mentioned above, this is an ill-
posed, inverse problem, so solely relying on the optical flow constraint, i.e. to
assume that the intensities remained constant, does not provide sufficient infor-
mation to infer meaningful flow fields. Horn and Schunck were the first to apply
the tools of the calculus of variations to impose a quadratic regularization on
the displacement field [11]. Black and Anandan [12] introduced methods from
robust statistics to avoid the quadratic penalization of changes in the flow field
and of violations of the optical flow constraint, thereby allowing for flow dis-
continuities and outliers in the data term, respectively. For a systematic survey
on variational formulations of the optical flow problem, we refer to Aubert et
al. [13].
While robust error functions have the desirable property to limit the influence
of outliers, most of them are not convex and therefore hard to optimize. Due to
the fact that gross outliers are rare in optical flow estimation, several authors
proposed a Total Variation (TV) regularization of the flow field and an L1 data
fidelity term [14,15,16]:
2 1

min |∇ud | dx + λ ρ(u(x), x) dx (6)
u Ω Ω
d=1
T
with u(x) = (u1 (x), u2 (x)) and ρ (u(x), x) = |I1 (x + u(x)) − I0 (x)|, where
I0 and I1 are the two input images. Since the data fidelity term ρ is nonlinear
in u, a local linearization is required at some point in the solution procedure,
which limits the approach to recovering small displacements. To circumvent this
restriction, the estimation procedure is typically performed on a scale pyramid
of the input images, cf. [14] for an in-depth discussion. Yet, such a scale pyramid
strategy often fails to recover the flow for small holes in foreground objects
(which allow to observe a distant background), or for fine structures in front of a
background which is moving due to ego motion. This effect can be seen around
the moving leafs in the “Schefflera” sequence of the Middlebury optical flow
evaluation dataset [17]. Figure 1(a) shows the input image 0 of the “Schefflera”
sequence; the corresponding ground truth flow can be seen as color-coded image1
in Fig. 1(b). Figure 1(c) shows a color-coded flow field, estimated using a scale
pyramid-based implementation of the TV-L1 optical flow model (6).
Before applying the new solution strategy, in a short technical detour, we
follow the approach of Aujol et al. [18] and introduce an auxiliary displacement
field v, yielding a strictly convex approximation of (6):
2 2 1
1 2
min |∇ud | dx + (ud − vd ) dx + λ ρ(v, x) dx , (7)
u,v Ω 2θ Ω Ω
d=1 d=1
1
The hue encodes the direction of the flow vector, while the saturation encodes its
magnitude. Regions of unknown flow (e.g. due to occlusions) are colored black.
(a) (b) (c) (d)
Fig. 1. (a) input image 0 of the Middlebury “Schefflera” sequence; (b) color-coded
ground truth flow (hue = direction, intensity = magnitude, black = unknown); color-
coded flows, estimated using: (c) a continuous TV-L1 flow model; (d) the proposed
optimization strategy (AAE = 2.91◦ )
where the small, positive constant θ ensures a tight coupling of u and v. In

contrast to the original energy (6), this convex approximation can be solved
using a dual approach. Our proposed optimization strategy can now easily be
applied by replacing v with a binary fusion of two flow proposals α and β:
2 2
1
min |∇ud | dx + (ud − [(1 − φ) αd + φβd ])2 dx
u,φ Ω 2θ Ω
d=1 d=1
6
+ λ (1 − φ) ρ(α, x) + φρ(β, x) dx . (8)
Ω
The relaxed version of (8) is a minimization problem in two variables, u and φ.

We therefore have to perform an alternating minimization procedure:
1. For φ fixed, solve for every ud :
5 6
1 2
min |∇ud | dx + (ud − [(1 − φ) αd + φβd ]) dx (9)
ud Ω 2θ Ω
2. For u fixed, solve for φ:

1
2
2
min (ud − [(1 − φ) αd + φβd ]) dx
φ 2θ Ω
d=1
6
+λ (1 − φ) ρ(α, x) + φρ(β, x) dx (10)
Ω
The subproblem (9) is the well understood image denoising model of Rudin,
Osher, and Fatemi [19]. For this model, Chambolle proposed an efficient and
globally convergent numerical scheme, based on a dual formulation [20]. In prac-
tice, a gradient descent/reprojection variant of this scheme performs better [9],
although there is no proof for convergence. In order to make this paper self-
contained, we reproduce the relevant results from [20,9]:
Proposition 1. The solution of (9) is given by
ud = [(1 − φ) αd + φβd ] − θ∇ · pd . (11)

The dual variable pd is obtained as the steady state of

τ
p˜d k+1 = pkd + ∇ θ∇ · pkd − [(1 − φ) αd + φβd ] ,
θ
p˜ k+1
pk+1
d = ,d - , (12)
max 1, |p˜d k+1 |
where k is the iteration number, pd0 = 0, and τ ≤ 1/4.

The subproblem (10) permits a direct solution.
Proposition 2. The solution of (10) is given by clamping
⎧ T
⎨ (u − α) (β − α) + λθ (ρ(β, x) − ρ(α, x))
where α = β
φ̃ = T
(β − α) (α − β) (13)
⎩
0 elsewhere
to the range [0, 1]: , , --

φ = max 0, min 1, φ̃ . (14)
Proof: Starting with the Euler-Lagrange equation of (10)

1 T
(u − [(1 − φ) α + φβ]) (α − β) + λ [ρ(β, x) − ρ(α, x)] , (15)
θ
we try to solve for φ, yielding
φ (β − α)T (α − β) = (u − α)T (β − α) + λθ (ρ(β, x) − ρ(α, x)) . (16)
Wherever α = β, ρ(α, x) = ρ(β, x), hence φ can arbitrarily be chosen in [0, 1].
Everywhere else, we can divide by (β − α)T (α − β), yielding (13).
Please note that the data residuals ρ(α, x) and ρ(β, x) are just constants. They
have to be calculated only once per fusion step, and the sole requirement is that
their range is IR+
0 , i.e. almost any cost function can be used. Once the relaxed
version of problem (8) is solved, a final thresholding of φ is required to obtain the
binary fusion of the two flow proposals α and β. Since the continuous solution of
φ is already close to binary, the threshold μ is not critical. Our heuristic solution
is to evaluate the energy of the original TV-L1 energy (8) for α, β, and a few
different thresholds μ ∈ (0, 1). Finally, we select the threshold yielding the flow
field with the lowest energy.
3.1 Extension to a Second-Order Prior

The presented approach is by no means limited to Total Variation regularization.
As an example, we will apply the proposed technique to an optical flow model
with a prior based on decorrelated second-order derivatives [21]. This second-
order prior has the intrinsic property to penalize only deviations from piecewise
affinity. Since spatial second-order derivatives are not orthogonal and the local
information of orientation and shape are entangled, a decorrelation is necessary.
In [22], Danielsson et al. used circular harmonic functions to map the the second-
order derivative operators into an orthogonal space. In two spatial dimensions,
the decorrelated operator is given by
B 2
1 ∂2 ∂2 √ ∂ ∂2 √ ∂2 T
♦= + , 2 − , 8 . (17)
3 ∂x2 ∂y2 ∂x2 ∂y2 ∂x ∂y
The magnitude of this operator, defined as the Euclidean vector norm

B ' 2 2 2 2 2
1 ∂2u ∂2u ∂ u ∂2u ∂ u
♦u = + + 2 − + 8 , (18)
3 ∂x2 ∂y
2 ∂x2 ∂y
2 ∂x ∂y
measures the local deviation of a function u from being affine. Adapting the TV-
L1 flow model (6) to the new prior is a matter of replacing the TV regularization
in (6–8) with the Euclidean norm of the new operator (18). Instead of minimizing
the ROF energy (9), step 1 of the alternate optimization of u and φ now amounts
to solving
5 6
1 2
min ♦ud dx + (ud − [(1 − φ) αd + φβd ]) dx (19)
ud Ω 2θ Ω
for every ud , while keeping φ fixed.

Proposition 3. The solution of (19) is given by
ud = [(1 − φ) αd + φβd ] − θ♦ · pd . (20)
The dual variable pd is obtained as the steady state of

τ # $
p˜d k+1 = pkd + ♦ [(1 − φ) αd + φβd ] − θ♦ · pkd ,
θ
p˜ k+1
pk+1
d = ,d - , (21)
max 1, |p˜d k+1 |
where k is the iteration number, pd0 = 0, and τ ≤ 3/112. For a proof and further
details please refer to [21].
Moreover, we employ the following standard finite differences approximation of
the ♦ operator:
⎛ ⎞
1
(ui,j−1 + ui,j+1 + ui−1,j + ui+1,j − 4ui,j )
⎜ 3 ⎟
⎜ ⎟
(♦u)i,j = ⎜ 2
(u i−1,j + u i+1,j − u i,j−1 − u i,j+1 ) ⎟ , (22)
⎝ 3 ⎠
3 (ui,j + ui+1,j+1 − ui,j+1 − ui+1,j )
8
where (i, j) denote the indices of the discrete image domain, enforcing Dirichlet
boundary conditions on ∂Ω. For details on the discretization of ♦ · p, please
consult [21].
In the continuous setting, such an extension requires minor adaptions of the
solution procedure and incurs only a small increase of time and memory require-
ments. Most combinatorial optimization approaches, however, are limited to unary
and pairwise clique potentials, hence such a second-order prior can not be used.
Extending combinatorial optimization algorithms to higher-order cliques (e.g. as
proposed in [5,23]) is either expensive in time and space or imposes restrictions on
the potentials, e.g. [23] restricts the potentials to the Potts model.
4 Experiments
In this section, we first restrict the optical flow model to a single dimension
(rectified stereo) in order to analyze its behavior in a simplified setting. In Section
4.2 we will use image sets from the Middlebury optical flow database [17] to
illustrate that the proposed algorithm yields state-of-the-art flow estimates.
Most of the algorithm has been implemented in C++, with the exception of the
numerical schemes of the solvers, which have been implemented using CUDA 1.0.
All subsequent experiments have been performed on an Intel Core 2 Quad CPU
at 2.66 GHz (the host code is single-threaded, so only one core was used) with
an NVidia GeForce 8800 GTX graphics card, running a 32 bit Linux operating
system and recent NVidia display drivers. Unless noted otherwise, in all our
experiments the parameters were set to λ = 50 and θ = 0.1.
4.1 Illustrative Stereo Experiment
A restriction of the optical flow model (8) to a single displacement u permits a

direct comparison of the estimated solutions to the global optimum (cf. [24] for
details on calculating the global optimum of this multi-label problem). Since the
approach presented in [24] is based on a discrete solution space, we further restrict
our method by using only constant disparity proposals in 0.5 pixel increments.
(a) (b) (c)
Fig. 2. (a) im2 of the Middlebury “Teddy” stereo pair; (b) the corresponding ground
truth disparity map; (c) a mask for the pixels, which are also visible in im6
Moreover, this (rectified) stereo setting simplifies discussing the effects caused by
the relaxation and the seemingly asymmetric formulation.
All experiments in this section use the grayscale version of the “Teddy” stereo
pair [25] and a set of constant disparity proposals in the range 0 to 59 pixels in
0.5 pixel increments. Figure 2(a) shows im2 of the stereo pair, Fig. 2(b) the corre-
sponding ground truth disparity, and Fig. 2(c) is the mask of non-occluded regions.
Relaxation and Thresholding. For every fusion step (4), the binary function φ :
Ω → {0, 1} has to be optimized. Since this is a non-convex problem, we proposed
to relax φ to the range [0, 1], solve the continuous problem, and finally threshold
φ. Obviously, this only leads to reasonable fusions of α and β, if the optimal φ is
close to binary. Figure 3 shows, how a 64-bin histogram of the relaxed function φ
evolves during a typical optimization procedure. Since φ is initialized with 0, in the
beginning proposal α is chosen over β in the whole image. However, the “traces”
in Fig. 3 illustrate that in several image regions the value of φ flips to 1, i.e. in
these regions proposal β is energetically favored. Once the algorithm converged,
the histogram of φ is close to binary, just as expected.
Symmetry. Since we initialize φ(x) = 0, the formulation appears to be asym-

metric with respect to the flow proposals α and β, but in practice the effects of
swapping the proposals is negligible. Figure 4(a) shows some intermediate solu-
tion of the “Teddy” disparity map, which is used as proposal α in the following
experiment. The energy of α is 325963 and there clearly is room for improvement.
The disparity map resulting from a fusion with the constant proposal β(x) = 20
has an energy of 284748 and can be seen in Fig. 4(b). To highlight the image
Fig. 3. The evolution of a 64-bin histogram of φ during the solution procedure. Starting
at φ(x) = 0, i.e. with proposal α, several image regions flip to the alternative proposal
β. It is clearly apparent that the converged histogram is close to binary. Please note
that a logarithmic scale is used for the “iteration” axis.
(a) (b) (c)
Fig. 4. Two intermediate disparity maps, before (a) and after (b) a binary fusion with
the proposal β(x) = 20. (c) shows the corresponding continuous optimum of φ.
regions that were switched to proposal β during this fusion step, Fig. 4(c) shows
the continuous optimum of φ (before thresholding). Please note that φ is mostly
binary, except for regions where neither α nor β are close to the true disparity.
Repeating the experiment with α and β switched leads to visually indistinguish-
able results and an energy of 284727, i.e. the order of the proposals does not
matter in the binary fusion step.
Disparity Estimation via Randomized Sweeping. In the discrete α-expansion al-

gorithm [5], the minimization of a multi-label problem is performed by “agglom-
erating” an increasingly accurate solution by repeatedly allowing the current
labeling to switch to an alternative label. One sweep through the label space
is called a cycle. In the following experiment on the “Teddy” stereo pair, we
2nd cycle
1e+06
3rd cycle
4th cycle
1st cycle
sequential
Energy
4e+05
global optimum
2e+05
1 10 100
Iteration
Fig. 5. Decrease of the energy of the flow field with every successful fusion. Results
of randomized runs are shown as thin, colored lines; the thick, black line shows the
progress for a sequential run. The dashed vertical lines delimit fusion cycles, the thick
horizontal line marks the global optimum for the TV-L1 flow model (6).
emulate this behavior by sweeping through a set of constant disparity proposals

in the range 0 to 59 in 0.5 pixel increments. Figure 5 depicts, how the energy
of the flow field decreases with every successful fusion step. Sweeping the “label
space” in consecutive order of the displacements results in the thick, black line
labeled “sequential”. The results for 25 distinct runs with a randomized order
of the proposals are illustrated using thin, colored lines. The dashed vertical
lines delimit fusion cycles, i.e. between two of those vertical lines every possible
disparity is tested exactly once. The thick horizontal line represents the energy
of the global optimum (E = 199891) for this model at λ = 50, see Fig. 6(c)
for a disparity map of this solution. It is clearly apparent that after three or
more cycles the particular sweeping order does not have a significant influence.
After eight cycles, the mean energy of all runs is 205114, with the best run being
roughly 2 % better than the worst. Two exemplary disparity maps are shown in
Figs. 6(a) and 6(b) – they differ from the global optimum (shown in Fig. 6(c))
mainly in occluded areas, i.e. areas, where Fig. 2(c) is black. The mean energy of
the disparity maps, estimated using this approximated minimization technique,
is 2.6% higher than the global optimum, which is consistent with empirical re-
sults reported for the discrete α-expansion algorithm.
Depending on the quality of the model, reducing the energy of a solution might
not result in a lower error on the true problem. Hence, in order to compare the
true errors of the global optimum and the approximated disparity fields, the
results have been evaluated on the Middlebury stereo vision benchmark [25]. At
an error threshold of 0.5, the global optimum mislabels 16.9 % of the pixels in
non-occluded regions; the estimated solution shown in Fig. 6(b) has a slightly
larger error of 17.2 %, which is consistent with the energy differences.
4.2 Optical Flow Estimation

For optical flow estimation, merely fusing a set of constant proposals, just like
we did in the stereo case, is not feasible. One of the main reasons is that the
expected range of motion is not known beforehand. Simply assuming very large
ranges to be on the safe side either results in a coarse sampling of the solu-
tion space or in a huge number of constant flow proposals. Such a brute force
method could be made tractable by using a scale pyramid, but this would
still be inelegant and slow. However, the presented optimization strategy is
not limited to fusing constant proposals – any flow field can be used as pro-
posed solution. Thus, one obvious solution to the problem is to estimate a set
of flow fields using a standard optical flow algorithm and fuse them. There-
fore, before starting any fusion experiments, a set of 27 TV-regularized flows
(λ ∈ {10, 25, 40, 65, 100, 150, 200, 500, 1000}, θ ∈ {0.05, 0.1, 0.15}) and a set of 24
second-order prior regularized flows (λ ∈ {10, 25, 40, 55, 90, 200, 500, 1000}, θ ∈
{0.05, 0.1, 0.15}) have been estimated using the algorithms described in [16,21].
Since the proposed optimization strategy is not limited to convex data terms,
we used a truncated data term ρ = min{1, 1 − r}, where r is the normalized
cross-correlation, calculated on 3 × 3 patches across all color channels.
(a) (b) (c)
(d)
(e) (f) (g)
(h) (i) (j)
Fig. 6. First row: (a), (b) disparity maps, estimated by repeatedly fusing constant
proposals using the proposed optimization strategy; (c) global optimum for the TV-
L1 model. Second row: average end-point error results for the Middlebury benchmark
dataset; the proposed method, labeled “CBF”, was ranked 2nd at the time of sub-
mission. Third row: (e–g) show color-coded flow fields for the sequences “Schefflera”,
“Grove”, and “Yosemite”. Last row: (h) shows a color-coded flow field for the Mid-
dlebury “RubberWhale” sequence, estimated using the second-order prior (AAE =
3.14◦ ); (i) and (j) show the color-coded flow fields for a TV-regularized fusion of TV and
second-order prior flows for the “RubberWhale” (AAE = 2.87◦ ) and the “Dimetrodon”
(AAE = 3.24◦ ) sequences.
A TV-regularized fusion of all 27 TV flows has been submitted to the Middle-

bury optical flow evaluation site, where at the time of submission this method
(labeled “CBF”) was ranked second or better for 14 out of 16 error measures.
Figure 6(d) shows the average end-point error of the six top-ranked algorithms,
and Figs. 6(e)–6(g) show some color-coded results. Please visit the evaluation
page at http://vision.middlebury.edu/flow/eval/ for other error measures
and further images. Due to the challenging diversity of the data sets, this was
the only experiment where we used λ = 90 to improve the results on “Schefflera”
and “Grove” at the cost of a rather noisy “Yosemite” result. On the “Urban”
sequence (640 × 480 pixels), the outlined flow estimation procedure took 218 s
(138 s for precalculating 27 TV-regularized flows and 80 s for 6 fusion cycles).
Comparing the Figs. 6(e) and 1(c) indicates a slight improvement of the es-
timated flow, but since we only fuse TV-L1 proposals, the correct solution is
never “offered.” For further improvements, we have to resort to the brute force
strategy of fusing a number of constant flow proposals, but since we already have
a good estimate of the flow, the solution space is small and it quickly converges.
The final result has an average angular error (AAE) of 2.91◦ (see Fig. 1(d)).
Furthermore, the second-order prior regularized optical flow algorithm (see
Section 3.1) was used to fuse the 24 precalculated second-order flows. Figure
6(h) shows a color-coded result for the “RubberWhale” sequence of the Mid-
dlebury training dataset (AAE = 3.14◦). Using all precalculated flows and a
TV-regularized fusion algorithm yields even better results: Figs. 6(i) and 6(j)
show the color-coded flow fields for the “RubberWhale” (AAE = 2.87◦ ) and
“Dimetrodon” (AAE = 3.24◦ ) sequences, respectively.
5 Conclusion
The presented optimization strategy permits large optimization moves in a vari-
ational context, by restating the minimization problem as a sequence of binary
subproblems. After verifying that the introduced approximations are reasonable,
we showed that typical solutions for a stereo problem are within a few percent of
the global optimum (in energy as well as in the true error measure). Finally, we
showed that applying this optimization strategy to optical flow estimation yields
state-of-the-art results on the challenging Middlebury optical flow dataset.
References
1. Mumford, D.: Bayesian rationale for energy functionals. In: Geometry-driven dif-
fusion in Computer Vision, pp. 141–153. Kluwer Academic Publishers, Dordrecht
(1994)
2. Szeliski, R., Zabih, R., Scharstein, D., Veksler, O., Kolmogorov, V., Agarwala, A.,
Tappen, M., Rother, C.: A comparative study of energy minimization methods for
Markov random fields with smoothness-based priors. IEEE Trans. Pattern Anal.
Mach. Intell. 30(6), 1068–1080 (2008)
3. Pearl, J.: Probabilistic Reasoning in Intelligent Systems: Networks of Plausible
Inference. Morgan Kaufmann, San Francisco (1988)
4. Kolmogorov, V.: Convergent tree-reweighted message passing for energy minimiza-

tion. IEEE Trans. Pattern Anal. Mach. Intell. 28(10), 1568–1583 (2006)
5. Boykov, Y., Veksler, O., Zabih, R.: Fast approximate energy minimization via
graph cuts. IEEE Trans. Pattern Anal. Mach. Intell. 23, 1222–1239 (2001)
6. Veksler, O.: Graph cut based optimization for MRFs with truncated convex priors.
In: Proc. of the CVPR (June 2007)
7. Lempitsky, V., Rother, C., Blake, A.: LogCut – efficient graph cut optimization
for Markov random fields. In: Proc. of the ICCV (October 2007)
8. Komodakis, N., Tziritas, G., Paragios, N.: Fast, approximately optimal solutions
for single and dynamic MRFs. In: Proc. of the CVPR (June 2007)
9. Chambolle, A.: Total variation minimization and a class of binary MRF models.
Energy Minimization Methods in Comp. Vision and Pattern Rec. 136–152 (2005)
10. Nikolova, M., Esedoglu, S., Chan, T.F.: Algorithms for finding global minimizers
of image segmentation and denoising models. SIAM J. on App. Math. 66 (2006)
11. Horn, B.K.P., Schunck, B.G.: Determining optical flow. Artificial Intelligence 17,
185–203 (1981)
12. Black, M.J., Anandan, P.: A framework for the robust estimation of optical flow.
In: Proc. of the ICCV, pp. 231–236 (May 1993)
13. Aubert, G., Deriche, R., Kornprobst, P.: Computing optical flow via variational
techniques. SIAM Journal on Applied Mathematics 60(1), 156–182 (2000)
14. Papenberg, N., Bruhn, A., Brox, T., Didas, S., Weickert, J.: Highly accurate op-
tic flow computation with theoretically justified warping. International Journal of
Computer Vision 67(2), 141–158 (2006)
15. Bruhn, A., Weickert, J., Kohlberger, T., Schnörr, C.: A multigrid platform for
real-time motion computation with discontinuity-preserving variational methods.
16. Zach, C., Pock, T., Bischof, H.: A duality based approach for realtime TV-L1
optical flow. In: Hamprecht, F.A., Schnörr, C., Jähne, B. (eds.) DAGM 2007. LNCS,
17. Baker, S., Scharstein, D., Lewis, J.P., Roth, S., Black, M., Szeliski, R.: A database
and evaluation methodology for optical flow. In: Proc. of the ICCV (2007)
18. Aujol, J.F., Gilboa, G., Chan, T.F., Osher, S.: Structure-texture image decompo-
sition – modeling, algorithms, and parameter selection. International Journal of
Computer Vision 67(1), 111–136 (2006)
19. Rudin, L., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal
algorithms. Physica D 60, 259–268 (1992)
20. Chambolle, A.: An algorithm for total variation minimization and applications.
Journal of Mathematical Imaging and Vision 20, 89–97 (2004)
21. Trobin, W., Pock, T., Cremers, D., Bischof, H.: An unbiased second-order prior for
high-accuracy motion estimation. In: Rigoll, G. (ed.) DAGM 2008. LNCS, vol. 5096,
22. Danielsson, P.E., Lin, Q.: Efficient detection of second-degree variations in 2D and
3D images. Journal of Visual Comm. and Image Representation 12, 255–305 (2001)
23. Kohli, P., Kumar, P., Torr, P.H.: P3 & beyond: Solving energies with higher order
cliques. In: Proc. of the CVPR (June 2007)
24. Pock, T., Schoenemann, T., Cremers, D., Bischof, H.: A convex formulation of
continuous multi-label problems. In: Forsyth, D., Torr, P., Zisserman, A. (eds.)
In: Proc. of the CVPR, vol. 1, pp. 195–202 (June 2003)
Unified Crowd Segmentation
Peter Tu, Thomas Sebastian, Gianfranco Doretto, Nils Krahnstoever,

Jens Rittscher, and Ting Yu
GE Global Research, Niskayuna, NY USA

tu@crd.ge.com
Abstract. This paper presents a unified approach to crowd segmenta-

tion. A global solution is generated using an Expectation Maximization
framework. Initially, a head and shoulder detector is used to nominate
an exhaustive set of person locations and these form the person hypothe-
ses. The image is then partitioned into a grid of small patches which are
each assigned to one of the person hypotheses. A key idea of this paper
is that while whole body monolithic person detectors can fail due to oc-
clusion, a partial response to such a detector can be used to evaluate the
likelihood of a single patch being assigned to a hypothesis. This captures
local appearance information without having to learn specific appearance
models. The likelihood of a pair of patches being assigned to a person
hypothesis is evaluated based on low level image features such as uni-
form motion fields and color constancy. During the E-step, the single and
pairwise likelihoods are used to compute a globally optimal set of assign-
ments of patches to hypotheses. In the M-step, parameters which enforce
global consistency of assignments are estimated. This can be viewed as
a form of occlusion reasoning. The final assignment of patches to hy-
potheses constitutes a segmentation of the crowd. The resulting system
provides a global solution that does not require background modeling
and is robust with respect to clutter and partial occlusion.
1 Introduction
The segmentation of crowds into individuals continues to be a challenging re-
search problem in computer vision [1, 2, 3, 4, 5]. The automation of video surveil-
lance systems in public venues such as airports, mass-transit stations and sports
stadiums requires the ability to detect and track individuals through complex
sites. We identify three challenges that make this problem particularly difficult:
(i) Partial occlusion. In many crowded scenes people can be partially occluded
by others. Monolithic detectors [2, 6, 7] that model the shape and appearance of
an entire person typically fail in such situations and hence cannot reliably detect
people in crowded environments. (ii) Dynamic backgrounds. When cameras are
fixed, statistical background models are commonly used to identify foreground
regions [8]. However, this approach fails when the background is dynamic. Fur-
ther, background modeling is not applicable for moving cameras, such as those
mounted on pan tilt devices or mobile platforms. (iii) Foreground clutter. The
presence of moving non-person objects such as luggage carts, shopping trolleys

692 P. Tu et al.
and cleaning equipment can clutter the foreground of the scene. A robust crowd
segmentation algorithm should be immune to foreground clutter without having
to explicitly model the appearance of every non-person object.
This paper presents a unified approach to crowd segmentation that effectively
addresses these three challenges. The proposed system combines bottom-up and
top-down approaches in a unified framework to create a robust crowd segmen-
tation algorithm. We first review a number of relevant approaches.
Low level feature grouping has been used to segment crowds [5, 9]. These ap-
proaches take advantage of the fact that the motion field for an individual is
relatively uniform and hence tracked corners with common trajectories can be
grouped together to form individuals. However, difficulties arise when multiple
individuals have similar trajectories. Monolithic classifiers capture the shape
and appearance space for the whole body using relatively simple learning meth-
ods [10,6,7]. The direct application of these classifiers to non-crowded scenes gen-
erates reasonable segmentations, however failure modes can occur when partial
occlusions are encountered. Part based constellation models [11, 12, 13] construct
boosted classifiers for specific body parts such as the head, the torso and the
legs, and each positive detection generates a Hough-like vote in a parametrized
person space. The detection of local maxima in this space constitutes a segmen-
tation. A similar approach [2] uses interest operators to nominate image patches
which are mapped to a learned code book. A drawback of these approaches is
that the identification of local maxima in the Hough space can be problematic
under crowded and cluttered environments - a global approach is required.
The previous approaches can be considered to be bottom-up methods where
local context is used. On the other hand, global approaches that rely on back-
ground segmentation has been proposed in [14, 4]. In [14], Markov Chain Monte
Carlo (MCMC) algorithms are used to nominate various crowd configurations
which are then compared with foreground silhouette images. However, this form
of random search can be computationally expensive. To address this issue an
Expectation Maximization (EM) based approach has been developed [4]. In this
framework, a hypothesis nomination scheme generates a set of possible person
locations. Image features are then extracted from foreground silhouettes and a
global search for the optimal assignment of features to hypotheses is performed.
The set of hypotheses that receive a significant number of assignments constitute
the final segmentation. Reliance on accurate foreground background segmenta-
tion is a weakness of both of these approaches.
1.1 Overview of the Unified Approach

In this paper we extend the global EM crowd segmentation framework [4] to use
appearance-based features that do not rely on background segmentation. A head
and shoulder classifier is used to generate an initial set of hypothesized person
locations, a grid of patches are then superimposed on the image. A globally op-
timal assignment of patches to hypotheses defines the final segmentation. The
likelihood of a single patch to hypothesis assignment is evaluated based on local
appearance. However, instead of learning an appearance and spatial distribution
Unified Crowd Segmentation 693
model for every possible patch location, we show that a partial response from a
monolithic whole body classifier operating solely on a given patch can discrimi-
nate between valid and invalid patch assignments. The framework also allows for
the inclusion of grouping terms based on low level image cues so that concepts
such as uniform motion and intra-garment color constancy can be leveraged.
During the E-step we estimate a globally optimal assignment of patches to per-
son hypotheses. The M-step ensures that globally consistent patch assignments
are chosen. This can be viewed as a form of occlusion reasoning.
2 Segmentation
This section provides a detailed overview of the proposed crowd segmentation
algorithm. Figure 1 depicts the various stages used to generate the final seg-
mentation of a crowded scene. We assume that the head and shoulders of all
detectable individuals can be observed. Hence, an initial set of hypothesized
person locations are nominated using a head and shoulders detector (see section
3 for details). These K nominated hypotheses are denoted by C := {ci }. The
parameters of this head and shoulders detector are chosen to minimize missed
detections, hence many false detections are also generated(see Figure 1a). The
scene is partitioned into a set of N rectangular patches Z = {zi }, as shown in
Figure 1b. The segmentation of the scene into individuals is achieved by a glob-
ally optimal assignment of these image patches to the initial hypotheses. The
potential assignment of an image patch to a person hypothesis is evaluated using
both direct affinity and pairwise affinity terms, as described below.
Let gk (zi ) denote the affinity associated with the direct assignment of patch
zi to hypothesis ck . One of the main thrusts of this paper is a novel method for
computing this affinity function based on local shape and appearance informa-
tion - this will be the topic of section 3. Figure 1c illustrates this step for the
patch shown in green. The width and color of the arrow connecting the patch
to a hypothesis indicates the strength of the affinity. Using camera calibration
information and a ground plane assumption, certain direct assignments can be
ruled out based on geometric reasoning (shown with black arrows).
Let gk (zi , zj ) denote the affinity associated with pairwise assignment of patch
zi and zj to hypothesis ck . In this application, pairwise assignment affinity is
computed based on the fact that a given individual’s clothing often exhibits
a certain amount of color and motion constancy. Hence, affinity is based on
a similarity measure sim(zi , zj ) of low-level image cues such as motion fields
and color distributions. In this paper we use the Bhattacharya distance measure
between the color histograms associated with each patch. Given such a measure
of similarity, we define
gk (zi , zj ) = sim(zi , zj )min(gk (zi ), gk (zj )). (1)
Figure 1d shows two pairwise assignments. The pair of pink patches have a large
degree of pairwise affinity while the pair of blue patches exhibit relatively small
pairwise affinity.
694 P. Tu et al.
C4 C4 C4 C4
C1 C1 C1 C1
C2 C3 C2 C3 C2 C3 C2 C3
C5 C5 C5 C5
C6 C6 C6 C6
zi
(a) (b) (c) (d)

Hypotheses Hypotheses
c1 c2 c3 c4 c5 c6 Null c1 c2 c3 c4 c5 c6 Null
z1 z1
z2 z2 C4 C4 C4
Patches
Patches
z3 z3 C1 C1 C1
M M C2 C3
C2 C3 C2 C3
zN zN
C5 C5 C5
Iteration 0 Iteration 1
Hypotheses Hypotheses
c1 c2 c3 c4 c5 c6 Null c1 c2 c3 c4 c5 c6 Null
C6 C6 C6
z1 z1
z2 z2
Patches
Patches
z3 z3
M M
zN zN
Iteration 10 Iteration final
(e) (f) (g) (h)
Fig. 1. This figure depicts the different steps of the proposed algorithm. a) An initial set
of person hypotheses, b) a partitioning of the scene into a grid of patches, c) an example
of the direct association affinity between the green patch and all the hypotheses where
the width of the arrow is commensurate with the assignment affinity, d) shows two
patches with strong pairwise affinity (pink) and two patches with weak pairwise affinity
(blue), e) depicts the soft assign process where patches are assigned to hypotheses, f)
shows the assignment of patches to hypotheses after the first E-step, g) shows the
result of the M-step consistency analysis where red patch assignments are deemed to
be inconsistent based on occlusion reasoning, h) the final segmentation after multiple
iterations of the EM algorithm.
Given an initial set of hypotheses, a patch partitioning and the direct/pairwise

affinity measures, a global segmentation can be generated based on an estimate
of the optimal set of patch to hypothesis assignments. To achieve this, an EM
approach similar to [4] is used. The patch to hypothesis assignments is defined
by an assignment vector V = {vi } of length N with vi ∈ [1, . . . , K], where
vi = k implies that patch zi is assigned to hypothesis ck . The focus of the EM
approach is to model the joint probability of an assignment vector V and the set
of patches Z i.e. p(V, Z; X) under the parametrization of a consistency model
X (see section 2.1). To this end the log likelihood of the patch assignment is
defined as:

K
N
K
N
L(V |Z; X) ∝ γ1 xik gk (zi ) δck (vi ) + γ2 xik xjk gk (zi , zj ) δck (vi )δck (vj ) ,
k=1 i=1 k=1 i,j=1
i =j
(2)
where δck (vi ) is an indicator function which is one when vi = k and zero oth-
erwise and xik is a consistency parameter that is computed during the M-step
(see section 2.1). During the E-step, the consistency parameters are fixed and
a distribution for V is computed such that the expectation V p(V )L(V |Z; X)
is maximized. It was shown in [4] that a mechanism similar to soft-assign [15]
can be used to efficiently perform the E-step search. Figure 1e illustrates this
iterative process where the distribution of V is parametrized by a matrix of di-
rect assignment probabilities. The element in the ith row and k th column of this
matrix is the probability of the assignment of the it h patch to the k th hypoth-
esis. The sum along each row must be equal to 1 and there can be no negative
values. At the first iteration all matrix values are set uniformly. During each
iteration of the soft assign process, the matrix probabilities are updated based
on the gradient of the expectation function. After a number of iterations, the
assignment probabilities are forced to take on binary values and this defines an
estimate of the most likely value of V . If a hypothesis receives no patches then
it is deemed to be a spurious detection. A null hypothesis is created to allow for
the potential assignment of patches to the background and clutter in the scene.
In this application the direct assignment affinity between a patch and the null
hypothesis is set to a nominal value. The reader is directed to [4] for more detail
regarding the E-step process.
2.1 Enforcing Consistency

Figure 1f depicts a set of patch to hypothesis assignments that result from one
pass of the E-step. Note that a certain number of patches are assigned to the
spurious hypothesis c6 . In this case, these assignments would seem to indicate
that the feet of c6 are visible but that the head is not. These types of incon-
sistencies are addressed via the consistency parameters X = {xik } which are
computed during the M-step. As can be seen from equation 2, when xik is low,
the likelihood of an assignment vector with vi = k is reduced. For this paper,
the primary concern is to enforce consistency based on occlusion reasoning. The
following axioms form the basis for such an approach:
– If the person is visible then their head should be visible

– The non-occluded region of a person should be contiguous.
Given the current estimate of V all the patches that are currently assigned to a
given hypothesis ck can be identified. For each patch zi that is assigned to ck , a
path between it and the head location specified by ck can be constructed such
that the number of patches encountered on the path that are not assigned to ck
is minimal. This process takes advantage of the inherent grid like structure of
the patches and can be computed efficiently using dynamic programming. The
696 P. Tu et al.
value of xik is set to 1 unless the minimum cost path has a cost that is greater
than a threshold, in which case xik is set to a low value. Prior to the first E-step,
all the values of X are set to 1. Using this simple process, inconsistencies such
as the legs are visible but the torso is not, can be identified and addressed before
the next iteration of the E-step. Figure 1g shows the result of an M-step analysis
where consistent patch assignments are shown in white and the inconsistent
patch assignments are shown in red. By reducing the value of the consistency
parameters for the red assignments, their inclusion in subsequent estimates of V
will be inhibited.
The EM algorithm operates by iterating between the E-step and the M-step
operations. The process terminates when the estimates of V have converged.
Figure 1h shows the final segmentation for this example. By employing a global
optimization scheme, the system need not rely solely on local information for
making segmentation decisions, which is not the case for many greedy approaches
to crowd segmentation. In the next section, the critical question of how to com-
pute the affinity of direct patch to hypothesis assignments will be addressed.
3 Classification
In the previous section a detailed description of the overall crowd segmentation
process was given. The focus of this section is to describe how the direct patch
to hypothesis affinity function gk (z) can be computed based on local shape and
appearance information. For this purpose we use a whole body monolithic person
classifier consisting of a set of weak classifiers selected by boosting. We will show
that for certain types of weak classifiers, the whole body classifier response can be
computed for a specific patch and that this response can be used to characterize
the patch to hypothesis affinity. The section begins with a discussion of the
basic whole body classifier followed by details regarding the generation of patch
specific partial responses.
3.1 Whole Body Classifier

A sample s is defined as a head location in the image (xh , yh ) and an associated
sample bounding box that covers the entire body region. Given a hypothesized
head location, the associated bounding box is computed based on camera cali-
bration information under the dominant ground plane assumption [14]. Given an
arbitrary sample, a whole body classifier produces either a positive or negative
response. As was shown in [16] a strong classifier (sc) can be constructed from
a linear combination of weak classifiers (wc). The weak classifier is chosen by
minimizing an expected error function that is based on labeled training data:

N
(1 − wc(si ) ∗ li )pi
E(wc) = , (3)
i
2
where si is the ith training sample, li is its label and pi is the probability as-
sociated with sample si . The sample probability distribution is modified in an
iterative fashion so as to emphasize samples that were mis-classified during pre-

vious iteration. In this application an iterative site-specific approach is used for
learning. Initial training data from the site of interest is manually labeled and
a classifier is constructed. It is then applied to additional imagery taken from
the site, and the resulting false positives are incorporated into the training data,
while correctly classified negative training samples are removed. A new classifier
is then constructed and this process is repeated until no new false positive re-
sponses are generated. In this manner a series of strong classifiers are constructed
which are combined to form a cascaded classifier. The type of weak classifiers
that are appropriate for this application is now considered.
3.2 Weak Classifiers
A particular type of weak classifier can be characterized as follows:
wc(s; R(s)), (4)
where R is a region of interest defined relative to the sample bounding box

associated with a sample s. If the average values for a set of image statistics
are above (or below) a given set of thresholds, then the weak classifier produces
a positive or negative response accordingly. Once the type of image statistics
have been selected, the weak classifier is essentially parametrized by the relative
location and dimensions of its region of interest R. In general the threshold
values for selected weak classifiers are determined during the learning phase of
the boosting process. If we restrict our hypothesis space to this type of weak
classifier, then a patch specific partial response for a whole body classifier can
be generated.
Based on the boosting algorithm a strong whole body classifier for the sample
s is defined as:

M
sc(s) = αi wc(s; Ri (s)) (5)
i=1
The basic idea for generating patch specific responses is that each weak classifier
will only collect statistics over the intersection of R(s) and the patch z. Since
average statics are used, the thresholds learned during boosting remain valid.
However, instead of having a 1/ − 1 response, each weak classifier will have its
response modulated by the ratio of the areas of R(s) ∩ z and R(s). Based on this
idea, the partial response for a strong classifier with respect to a given patch z
and sample s is defined as:

M
Ri (s)∩z
dx
sc(s, z) = αi wci (s, z) , (6)
i=1 Ri (s)
dx
where
wci (s, z) = wc(s; Ri (s) ∩ z). (7)
698 P. Tu et al.
Note that if the region of interest associated with a particular weak classifier
does not intersect with the patch z, then this weak classifier will have no effect
on the strong classifier decision.
For a given person hypothesis ck , a sample sk can be constructed so that
for a particular patch zi , the direct patch to hypothesis affinity measure can be
defined as:
gk (zi ) = sc(sk , zi ) (8)
Figure 2 shows a set of cascaded classifiers that were used to construct the whole
body classifier. In this application the image statistic used is the magnitude of
the edge responses for pixels that exhibited an orientation similar to the preferred
orientation of the weak classifier. Edge magnitude and orientation are calculated
using the Sobel operator. Given such a whole body classifier the next question
is to determine the appropriate patch size. If the patch is too large, then there
is risk of contamination by occlusion. On the other hand, if the patch is too
small the ability to discriminate between correct and incorrect patch assignments
diminishes. To understand this tradeoff a training set of positive and negative
whole body samples was collected. Patches with widths ranging from 0.25W to
1.0W (W = person width) were evaluated across the entire bounding box for
each training sample. For each relative patch location, the average number of
positively responding strong classifiers from the cascaded whole body classifier
was recorded. As shown in Figure 3, when the patch width was reduced below
0.5W the ability to discriminate between positive and negative samples was
reduced significantly. Thus for this application a nominal patch width of 0.5W
is chosen.
Fig. 2. This figure shows the six strong classifiers that were constructed for the whole
body classifier plus all six cascades shown together. Each pink box represents the
region of interest for a weak classifier. The line interior to each region of interest depict
the weak classifier’s preferred orientation. Green lines represent positive features (the
average statistic must be above its threshold) and red lines are for negative features
(average statistic must be below its threshold).
Positive Negative Patch Positive Negative Patch

Response Response Width Response Response Width
Avg. 2.7 Avg. 2.0 0.25 Avg. 3.7 Avg. 2.3 0.50
Avg. 4.4 Avg. 2.2 0.75 Avg. 4.9 Avg. 2.1 1.00
Fig. 3. This figure shows the effect of changing the patch size. Patch size is varied as a
function of W = person width. In each case the average number of positively responding
strong classifiers from the whole body cascaded classifier is shown as a function of patch
location for both positive (person) and negative (non-person) images. Note that when
the patch width is reduced below 0.5W the ability to discriminate between positive
and negative samples is significantly reduced.
3.3 Hypothesis Nomination

For this application, hypothesis nomination as described in section 2 is achieved
using a scanning window approach. For every possible sample, the partial re-
sponse for the whole body classifier is evaluated based on a patch covering the
hypothesized head and shoulder regions. The set of positive responses constitute
the initial hypotheses set C.
4 Experiments
Unrehearsed imagery acquired at a mass transit site serves as the source of test
imagery for this paper. A whole body classifier was trained for this site. We
first illustrate the intermediate steps of our approach on a few representative
frames (see Figure 4). The “Initial Hypothesis” column of figure 4 shows the
initial set of hypotheses generated by the head and shoulders classifier. Note that
while an appropriate hypothesis was generated for each person in each image,
several false hypotheses were also generated. The “Single Assignment” column of
figure 4 illustrates the direct affinity between each patch and each hypothesis as
computed using equation 8. Each patch is color coded based on the hypothesis
700 P. Tu et al.
Initial Single Grouping E-step M-step Final

Hypothesis Assignment
Fig. 4. Six stages of the crowd segmentation process are illustrated for four test images.
Overlapping patches of 0.5W are used. However, for clarity smaller patches are shown.
The initial hypotheses generated by the head and shoulder classifier are shown in
the first column. In the second column, the patches are shown color coded based on
their strongest direct assignment as calculated by the whole body classifier. The null
hypothesis is shown in black. In the third column, neighboring patches with strong
similarity measures based on color constancy are connected by green line segments.
The assignment of patches to hypotheses based on the first E-step is shown in the
fourth column. The assignment after multiple rounds of both the E and M steps are
shown in the fifth column. The final segmentation is shown in the last column.
for which it has the highest direct affinity. Patches that are black have the
greatest affinity for the null hypothesis. A significant number of patches have
the greatest direct affinity for their true hypothesis, however confusion occurs
when multiple hypotheses overlap. An example of this can be seen in row A
of the Single Assignment column. In addition, patches that are only associated
with false detections tend to have a greater affinity for the null hypothesis.
The “Grouping” column of figure 4 illustrates the effectiveness of the pairwise
assignment criteria. For purposes of clarity, only neighboring patches with high
similarity measures are shown to be linked in green (blue otherwise). Note that
Fig. 5. This figure shows an assortment of crowd segmentation results. Note that the
algorithm produces the correct segmentation in case of severe partial occlusion (right
column), and in presence of cleaning equipment (bottom left) and a variety of suitcases
and bags.
patches associated with the same article of clothing tend to be grouped together.
Also, the background often exhibits continuity in appearance and such patches
tend to be grouped together.
The “E-step” column of figure 4 shows the patch assignment after the first
iteration of the “E-step”. Most of the false hypotheses have received very few
patch assignments, while the true hypotheses have been assigned patches in
an appropriate manner. However, inconsistencies have also been generated. For
example, in row A of the “E-step” column, a number of patches have been
assigned to the bottom of the green hypothesis. As seen in the “M-step” column,
these inconsistent assignments have been correctly removed.
The “Final” column of figure 4 shows the final segmentation. Figure 5 shows
similar results from a variety of images. Note that the algorithm is successful
when confronted with partial occlusion and clutter such as the janitor’s equip-
ment and various suitcases.
The algorithm was also applied to a video sequence (see supplemental ma-
terial). To measure overall performance 117 frames were processed. The initial
hypothesis generator produced 480 true detections, 32 false detections and 79
missed detections. After application of the crowd segmentation algorithm, the
number of false detections were reduced by 72 percent at a cost of falsely reject-
ing 2 percent of the true detections. For purposes of comparison, we applied the
Histogram of Oriented Gradients (HOG) [6] to this dataset. Our implementation
uses camera calibration information for automatic scale selection. The perfor-
mance tabulated in Table 1 shows that our crowd segmentation outperformed
HOG, arguably this is due to partial occlusion.
702 P. Tu et al.
Table 1. Comparison of HOG [6] person detector to the proposed crowd segmentation
algorithm
True Detects Missed Detects False Alarms

Crowd Segmentation 470 89 9
HOG [6] 387 172 20
frame 1 frame 30 frame 80 frame 117
Fig. 6. Four example frames from tracking the results of the crowd segmentation
process
The purpose of crowd segmentation algorithms in general is to reliably track

the location of people over time. The final segmentation of the sequence previ-
ously described was processed by a general-purpose person tracking algorithm.
At every time step, locations and estimates of the location uncertainties are
projected into the scene ground-plane via an unscented transform. Our tracker
processes these detections in a scene ground-plane reference frame, where the
dynamical models are intuitively formulated. Our approach to tracking is simi-
lar to [17] and [18]. We follow an efficient detect and track approach [19] using
a JPDAF filter [20], which has excellent performance in high degrees of clutter
while being efficient in the presence of many targets. The tracking results in
Figure 6 show the trajectories of all people in the scene.
Fig. 7. This figure illustrates the effect of using motion fields in the pairwise patch
assignment. The example on the left shows a frame where the algorithm results in
both a false positive and a false negative. However, when the motion information from
dense optical flow is used the correct segmentation results, as shown on the right.
The results thus far used pairwise patch similarity function based on color
constancy as defined in Equation 1. However, this is not always enough as shown
in the left image of Figure 7 where the crowd segmentation algorithm resulted
in both a false and a missed detection. An experiment was performed where
the pairwise patch similarity measures were augmented by the use of a motion
consistency measure based on dense optical flow. As can be seen from the right
image in Figure 7, this results in a correct segmentation.
5 Discussion
The framework presented in this paper has incorporated many of the strengths of
previously proposed crowd segmentation methods into a single unified approach.
A novel aspect of this paper is that monolithic whole body classifiers were used
to analyze partially occluded regions by considering partial responses associated
with specific image patches. In this way appearance information is incorporated
into a global optimization process alleviating the need for foreground background
segmentation. The EM framework was also able to consider low level image cues
such as color histograms and thus take advantage of the potential color constancy
associated with clothing and the background. Parametrization of the likelihood
function allowed for the enforcement of global consistency of the segmentation.
It was shown that these parameters can be estimated during the M-step and
that this facilitates consistency based on occlusion reasoning.
In the course of experimentation it was found that at various times, different
aspects of the crowd segmentation system proved to be the difference between
success and failure. For example when confronted with clutter, the appearance
based classifiers provide the saliency required to overcome these challenges. How-
ever, when multiple people having similar clothing are encountered, the motion
field can become the discriminating factor. A robust system must be able to take
advantage of its multiple strengths and degrade gracefully when confronted by
their weaknesses.
References
1. Munder, S., Gavrila, D.: An experimental study on pedestrian classification. IEEE

Trans. on Pattern Analysis and Machine Intelligence 28(11), 1863–1868 (2006)
2. Leibe, B., Seemann, E., Schiele, B.: Pedestrian detection in crowded scenes. IEEE
Computer Vision and Pattern Recognition, 878–885 (2005)
3. Leibe, B., Cornelis, N., Cornelis, K., Gool, L.V.: Dynamic 3d scene analysis from
a moving vehicle. IEEE Computer Vision and Pattern Recognition, 1–8 (2007)
4. Rittscher, J., Tu, P.H., Krahnstoever, N.: Simultaneous estimation of segmentation
and shape. IEEE Computer Vision and Pattern Recognition 2, 486–493 (2005)
5. Brostow, G.J., Cipolla, R.: Unsupervised bayesian detection of independent motion
in crowds. IEEE Computer Vision and Pattern Recognition I, 594–601 (2006)
6. Dalal, N., Triggs, B.: Histograms of oriented gradients for human detection. IEEE
Computer Vision and Pattern Recognition, 886–893 (2005)
704 P. Tu et al.
7. Tuzel, O., Porikli, F., Meer, P.: Pedestrian detection via classification on rieman-
nian manifolds. IEEE Computer Vision and Pattern Recognition (2007)
8. Stauffer, C., Grimson, W.: Adaptive background mixture models for real-time
tracking. IEEE Computer Vision and Pattern Recognition 2, 246–252 (1998)
9. Rabaud, V., Belongie, S.: Counting crowded moving objects. IEEE Computer Vi-
sion and Pattern Recognition, 705–711 (2006)
10. Viola, P., Jones, M., Snow, D.: Detecting pedestrians using patterns of motion and
appearance. International Journal of Computer Vision 2, 734–741 (2003)
11. Fergus, R., Perona, P., Zisserman, A.: A visual category filter for Google images.
In: European Conference on Computer Vision, vol. 1, pp. 242–256 (2004)
12. Mikolajczyk, K., Schmid, C., Zisserman, A.: Human detection based on a proba-
bilistic assembly of robust part detectors. In: European Conference on Computer
Vision (2004)
13. Wu, B., Nevatia, R.: Detection and tracking of multiple partially occluded humans
by bayesian combination of edgelet based part detectors. International Journal of
Computer Vision 75(2), 247–266 (2007)
14. Zhao, T., Nevatia, R.R.: Bayesian human segmentation in crowded situations.
IEEE Computer Vision and Pattern Recognition 2, 459–466 (2003)
15. Chui, H., Rangarajan, A.: A new point matching algorithm for non-rigid registra-
tion. Computer Vision and Image Understanding 89(3), 114–141 (2003)
16. Viola, P., Jones, M.J.: Robust real-time face detection. International Journal of
Computer Vision 57(2), 137–154 (2004)
17. Krahnstoever, N., Tu, P., Sebastian, T., Perera, A., Collins, R.: Multi-view detec-
tion and tracking of travelers and luggage in mass transit environments. In: Proc.
Ninth IEEE International Workshop on Performance Evaluation of Tracking and
Surveillance (PETS) (2006)
18. Leibe, B., Schindler, K., Gool, L.V.: Coupled detection and trajectory estimation
for multi-object tracking. In: International Conference on Computer Vision (ICCV
2007), Rio de Janeiro, Brasil (October 2007)
19. Blackman, S., Popoli, R.: Design and Analysis of Modern Tracking Systems. Artech
House Publishers (1999)
20. Rasmussen, C., Hager, G.: Joint probabilistic techniques for tracking multi-part
objects. In: Proc. IEEE Conference on Computer Vision and Pattern Recognition,
pp. 16–21 (1998)
Quick Shift and Kernel Methods
for Mode Seeking
Andrea Vedaldi and Stefano Soatto
University of California, Los Angeles

Computer Science Department
{vedaldi,soatto}@ucla.edu
Abstract. We show that the complexity of the recently introduced

medoid-shift algorithm in clustering N points is O(N 2 ), with a small
constant, if the underlying distance is Euclidean. This makes medoid
shift considerably faster than mean shift, contrarily to what previously
believed. We then exploit kernel methods to extend both mean shift
and the improved medoid shift to a large family of distances, with com-
plexity bounded by the effective rank of the resulting kernel matrix, and
with explicit regularization constraints. Finally, we show that, under cer-
tain conditions, medoid shift fails to cluster data points belonging to the
same mode, resulting in over-fragmentation. We propose remedies for this
problem, by introducing a novel, simple and extremely efficient clustering
algorithm, called quick shift, that explicitly trades off under- and over-
fragmentation. Like medoid shift, quick shift operates in non-Euclidean
spaces in a straightforward manner. We also show that the accelerated
medoid shift can be used to initialize mean shift for increased efficiency.
We illustrate our algorithms to clustering data on manifolds, image seg-
mentation, and the automatic discovery of visual categories.
1 Introduction
Mean shift [9,3,5] is a popular non-parametric clustering algorithm based on the

idea of associating each data point to a mode of the underlying probability den-
sity function. This simple criterion has appealing advantages compared to other
traditional clustering techniques: The structure of the clusters may be rather
arbitrary and the number of clusters does not need to be known in advance.
Mean shift is not the only “mode seeking” clustering algorithm. Other ex-
amples include earlier graph-based methods [13] and, more recently, medoid
shift [20]. Unlike mean shift, medoid shift extends easily to general metric spaces
(i.e. spaces endowed with a distance). In fact, mean shift is essentially a gra-
dient ascent algorithm [3,5,24] and the gradient may not be defined unless the
data space has additional structure (e.g. Hilbert space or smooth manifold struc-
ture). While there have been recent efforts to generalize mean shift to non-linear
manifolds [21], medoid shift does not require any additional steps to be used on
curved spaces. Moreover, the algorithm is non-iterative and there is no need for a
stopping heuristic. Its biggest disadvantage is its computational complexity [20].

706 A. Vedaldi and S. Soatto
Fig. 1. Mode seeking algorithms. Comparison of different mode seeking algo-

rithms (Sect. 2) on a toy problem. The black dots represent (some of) the data points
xi ∈ X ⊂ R2 and the intensity of the image is proportional to the Parzen density esti-
mate P (x). Left. Mean shift moves the points uphill towards the mode approximately
following the gradient. Middle. Medoid shift approximates mean shift trajectories by
connecting data points. For reason explained in the text and in Fig. 2, medoid shifts
are constrained to connect points comprised in the red circles. This disconnects por-
tions of the space where the data is sparse, and can be alleviated (but not solved) by
iterating the procedure (Fig. 2). Right. Quick shift (Sect. 3) seeks the energy modes
by connecting nearest neighbors at higher energy levels, trading-off mode over- and
under-fragmentation.
Depending on the implementation, medoid shift requires between O(dN 2 + N 3 )

and O(dN 2 + N 2.38 ) operations to cluster N points, where d is the dimension-
ality of the data. On the other hand, mean shift is only O(dN 2 T ), where T
is the number of iterations of the algorithm, and clever implementations yield
dT ) N .
Contributions. In this paper we show that the computational complexity of

Euclidean medoid shift is only O(dN 2 ) (with a small constant), which makes it
faster (not slower!) than mean shift (Sect. 3). We then generalize this result to
a large family of non-Euclidean distances by using kernel methods [18], showing
that in this case the complexity is bounded by the effective dimensionality of
the kernel space (Sect. 3). Working with kernels has other advantages: First, it
extends to mean shift (Sect. 4); second, it gives an explicit interpretation of non-
Euclidean medoid shift; third, it suggests why such generalized mode seeking
algorithms skirt the curse of dimensionality, despite estimating a density in
complex spaces (Sect. 4). In summary, we show that kernels extend mode seeking
algorithms to non-Euclidean spaces in a simple, general and efficient way.
Can we conclude that medoid shift should replace mean shift? Unfortunately,
not. We show that the weak point of medoid shift is its inability to identify
consistently all the modes of the density (Sect. 2). This fact was addressed
implicitly by [20] who reiterate medoid shift on a simplified dataset (similar
to [2]). However, this compromises the non-iterative nature of medoid shift and
changes the underlying density function (which may be undesirable). Moreover,
we show that this fix does not always work (Fig. 2).
Quick Shift and Kernel Methods for Mode Seeking 707
We address this issue in two ways. First, we propose using medoid shift to sim-
plify the data and initialize the more accurate mean shift algorithm (Sect. 5.2 and
Sect. 5.3). Second, we propose an alternative mode seeking algorithm that can
trade off mode over- and under-fragmentation (Sect. 3). This algorithm, related
to [13], is particularly simple and fast, yields surprisingly good segmentations,
and returns a one parameter family of segmentations where model selection can
be applied.
We demonstrate these algorithms on three tasks (Sect. 5): Clustering on a
manifold (Sect. 5.1), image segmentation (Sect. 5.2), and clustering image sig-
natures for automatic object categorization (Sect. 5.3). The relative advantages
and disadvantages of the various algorithms are discussed.
2 Mode Seeking
Given N data points x1 , . . . , xN ∈ X = Rd , a mode seeking clustering algorithm

conceptually starts by computing the Parzen density estimate
1
N
P (x) = k(x − xi ), x ∈ Rd (1)
N i=1
where k(x) can be a Gaussian or other window.1 Then each point xi is moved
towards a mode of P (x) evolving the trajectory yi (t), t > 0 uphill, starting from
yi (0) = xi and following the gradient ∇P (yi (t)). All the points that converge to
the same mode form a cluster.
A mode seeking algorithm needs (i) a numerical scheme to evolve the trajec-
tories yi (t), (ii) a halting rule to decide when to stop the evolution and (iii) a
clustering rule to merge the trajectory end-points. Next, we discuss two algo-
rithms of this family.
Mean Shift. Mean shift [9,5] is based on an efficient rule to evolve the tra-
jectories yi (t) when the window k(x) can be written as ψ(x22 ) for a convex
function ψ(z) (for instance the Gaussian window has ψ(z) ∝ exp(−z)). The
idea is to bound the window from below by the quadric k(z ) ≥ k(z) + (z 22 −
z22)ψ̇(z22 ). Substituting in (1) yields
1
N

P (y ) ≥ P (y) + (y − xj 22 − y − xj 22 )ψ̇(y − xj 22 ), (2)
N j=1
and maximizing this lower bound at y = yi (t) yields the mean-shift update rule
N
1 j=1 ψ̇(yi (t) − xj 2 )xj
N 2
yi (t + 1) = argmax y − xj 2 ψ̇(yi (t) − xj 2 ) = N
2 2
. (3)
j=1 ψ̇(yi (t) − xj 2 )
y N j=1 2
1
The term “kernel” is also used in the literature. Here we use the term “window” to
avoid confusion with the kernels introduced in Sect. 3.
If the profile ψ(z) is monotonically decreasing, then P (yi (t)) < P (yi (t + 1)) at
each step and the algorithm converges in the limit (since P is bounded [5]). The
complexity is O(dN 2 T ), where d is the dimensionality of the data space and T is
the number of iterations. The behavior of the algorithm is illustrated in Fig. 1.
Medoid Shift. Medoid shift [20] is a modification of mean shift in which the
trajectories yi (t) are constrained to pass through the points xi , i = 1, . . . , N .
The advantage of medoid shift are: (i) only one step yi (1), i = 1, . . . , N has to
be computed for each point xi (because yi (t + 1) = yyi (t) (1)), (ii) there is no
need for a stopping/merging heuristic (as these conditions are met exactly), and
(iii) the data space X may be non-Euclidean (since to maximize (4) there is no
need to compute derivatives). Eventually, points are linked by steps into a forest,
with clusters corresponding to trees. The algorithm is illustrated in Fig. 1.
According to [20], the main drawback of medoid shift is speed. In fact, maxi-
mizing (3) restricted to the dataset amounts to calculating
1 2
N
yi (1) = argmax d (y, xj )φ̇(d2 (xj , xi )) (4)
y∈{x1 ,...,xN } N j=1
where d2 (x, y) = x − y22 in the Euclidean case. A basic implementation re-

quires O(N 3 + dN 2 ) operations, assuming O(d) operations to evaluate d2 (x, y).
However, by defining matrices Dkj = d2 (xk , xj ) and Fki = φ̇(Dik )/N , we can
rewrite (4) as

N
yi (1) = argmax Dkj Fji = argmax e
k DF ei (5)
k=1,...,N j=1 k=1,...,N
where ei denotes the i-th element of the canonical basis.2 As noted in [20],
O(N 2.38 ) operations are sufficient by using the fastest matrix multiplication
algorithm available. Unfortunately the hidden constant of this algorithm is too
large to be practical (see [12], pag. 501). Thus, a realistic estimate of the time
required is more pessimistic than what suggested by the asymptotic estimate
O(dN 2 + N 2.38 ).
Here we note that a more delicate issue with medoid shift is that it may fail
to properly identify the modes of the density P (x). This is illustrated in Fig. 2,
where medoid shift fails to cluster three real points −1, +1 and +1/2, finding
two modes −1 and +1 instead of one. To overcome this problem, [20] applies
medoid shift iteratively on the modes (in the example −1 and +1). However, this
solution is not completely satisfactory because (i) the underlying model P (x) is
changed (similarly to blurry mean shift [9,3]) and (ii) the strategy does not work
in all cases (for instance, in Fig. 2 points −1 and +1 still fail to converge to a
single mode).
Finally, consider the interpretation of medoid shift. When X is a Hilbert space,
medoid (and mean) shift follow approximately the gradient of the density P (x)
2
# $
For instance e2 = 0 1 0 . . . 0 .
0.2 0.2
0.1 0.1
0 0
−1 1 1.5 −1 1
Fig. 2. Medoid shift over-fragmentation. Left. We apply medoid shift to cluster

points −1, +1, +1/2 ∈ R using a Gaussian window of variance σ 2 = 1 (dashed green
lines). The density P (x) (red curve; Sect. 2) has a single mode, but medoid shift fails
to move the point −1 towards the mode (i.e. y−1 (1) = −1). The reason is that the
quadratic lower bound (2) (blue curve) is larger at −1 than it is at +1 or +1/2. Notice
that mean shift would have moved −1 towards the mode by a small, but finite amount,
eventually extracting the single mode. Right. The problem is not solved even if medoid
shift is reiterated [20] on the two modes −1 and +1 (where +1 has double mass), even
if the density P (x) does become blurrier [2,20].
(by maximizing the lower bound (3)). The gradient depends crucially on the
inner product and corresponding metric defined on X , which encodes the cost
of moving along each direction [22]. For general metric spaces X , the gradient
may not be defined, but the term d2 (x, y) in (4) has a similar direction-weighing
effect. In later sections we will make this connection more explicit.
3 Fast Clustering
Faster Euclidean Medoid Shift. We show that the complexity of Euclidean
medoid shift is only O(dN 2 ) (with a small# constant)$instead of O(dN 2 + N 2.38 )
(with a large constant) [20]. Let X = x1 . . . xN be the data matrix. Let
n = (X + X )1 be the vector of the squared norms of the data, where 1
denotes the vector of all ones and + the Hadamard (component wise) matrix
product. Then we have
D = 1n + n1 − 2X X, DF = n(1 F ) + 1(n F ) − 2X (XF ). (6)
The term 1(n F ) has constant columns and is irrelevant to the maximiza-
tion (5). Therefore, we need to compute
DF ∝ n(1 F ) − 2X (XF ), n = (X + X )1 = (I + X X)1 (7)
where I is the identity matrix.3 It is now easy to check that each matrix product
in (7) requires O(dN 2 ) operations only.
3
And we used the fact that (I AB)1 = (B
A)1.
Kernel Medoid Shift. An advantage of medoid shift is the possibility of com-

puting (4) for distances d2 (x, y) other than the Euclidean one [20]. The decom-
position (6) can still be carried out if the distance d2 (x, y) can be expressed as
K(x, x) + K(y, y) − 2K(x, y) for an appropriate positive definite (p.d.) kernel4
K [18]. Then we have D = 1n + n1 − 2K, and
DF ∝ n(1 F ) − 2KF, n = (I + K)1.
Unfortunately, the multiplication KF is still O(N 2.38 ). However, we can search
for a low-rank decomposition G G of K (we assume, without loss of generality,
that K is centered5 ). If G is a decomposition of rank d, then
DF ∝ n(1 F ) − 2G(G F ), n = (I + G G)1 = (G + G )1
can still be computed in O(dN 2 ) operations. The cost of decomposing K is
typically around O(d2 N ) [8,1]. See Fig. 3 for a basic implementation.
Quick Shift. In order to seek the mode of the density P (x), it is not necessary
to use the gradient or the quadratic lower bound (2). Here we propose quick
shift, which simply moves each point xi to the nearest neighbor for which there
is an increment of the density P (x). In formulas,
1
N
yi (1) = argmin Dij , Pi = φ(Dij ). (8)
j:Pj >Pi N j=1
Quick shift has four advantages: (i) simplicity; (ii) speed (O(dN 2 ) with a small
constant); (iii) generality (the nature of D is irrelevant); (iv) a tuning parameter
to trade off under- and over-fragmentation of the modes. The latter is obtained
because there is no a-priori upper bound on the length Dij of the shifts yi (0) →
yi (1). In fact, the algorithm connects all the points into a single tree. Modes
are then recovered by breaking the branches of the tree that are longer than a
threshold τ . Searching τ amounts to performing model selection and balances
under- and over-fragmentation of the modes. The algorithm is illustrated in
Fig. 1.
Quick shift is related to the classic algorithm from [13]. In fact, we can
rewrite (8) as
sign(Pj − Pi ) Pj − Pi
yi (1) = argmax , and compare it to yi (1) = argmax (9)
j=1,...,N Dij j:d(xj ,xi )<τ Dij
4
The kernel K should not be confused with the Parzen window k(z) appearing in (1).
In the literature, it is common to refer to the Parzen window as “kernel”, but in most
cases it has rather different mathematical properties than the kernel K we consider
here. An exception is when the window is Gaussian, in which cases k(d2 (x, y)) is a
p.d. kernel. In this case, we point out an interesting interpretation of mean shift as
a local optimization algorithm that, starting from each data point, searches for the
pre-image of the global data average computed in kernel space. This explains the
striking similarity of the mean shift update Eq. (3) and Eq. (18.22) of [19].
5
K is centered if K1 = 0. If this is not the case, we can replace K by K = HKH,
where H = I − 11
/N is the so-called centering matrix. This operation translates
the origin of the kernel space, but does not change the corresponding distance.
as given by [13]. Notice that (Pj − Pi )/Dij is a numerical approximation of the

gradient of P in the direction xj − xi . The crucial difference is that maximizing
the gradient approximation must be done in a neighborhood of each point defined
a-priori by the choice of the parameter τ . Thus, model selection in [13] requires
running the algorithm multiple times, one for each value of τ . In contrast, quick
shift returns at once the solutions for all possible values of τ , making model
selection much more efficient.
4 Cluster Refinement
In the previous section we introduced fast kernel medoid shift as an accelerated

version of non-Euclidean medoid shift. Since medoid shift may over-fragment
modes, quick shift was then proposed as a method to control under- and over-
fragmentation by the choice of a parameter τ . No algorithm, however, guarantees
the same accuracy of the slower mean shift.
It is then natural to ask whether mean shift could be extended to work in a
non-Euclidean setting. [20] cites the problem of defining the mean as the major
obstacle to this idea. [21] addresses this issue by defining mean shift vectors on
the tangent space of a non-linear manifold, but no proof of convergence is given,
and the applicability is limited by the fact that the data space needs to have a
manifold structure known analytically.
A simple solution to this problem is to extend kernel medoid to a correspond-
ing kernel mean shift procedure. Let K(·, ·) be a p.d. kernel on the data space
X . Then K(x, ·) is an element of the so called reproducing kernel Hilbert space
H [19], whose inner product is defined by letting K(x, ·), K(y, ·)H = K(x, y).
Points x ∈ Rd are then identified with elements K(x, ·) of the Hilbert space.
Given this identification, we can write ·, xH for ·, K(x, ·)H .
Kernel mean shift computes a “density6 ” on H
1
N
P (y) = k(d2H (y, xj )), y∈H (10)
N j=1
where d2H (xj , y) = y, yH + xj , xj H − 2y, xj H . Notice that y ∈ H, unlike
standard mean shift, does not belong necessarily to the data space X (up to
the identification x ≡ K(x, ·)). However, if k(z) is monotonically decreasing,
then maximizing w.r.t. y can be restricted to the linear subspace spanH X =
spanH {x1 , . . . , xn } ⊂ H (if not, the orthogonal projection of y onto that space
decreases simultaneously all terms d2H (xj , y)).
Therefore, we can express all calculations relative to spanH X. In particular,
if Kij = K(xi , xj ) is the kernel matrix, we have d2H (xj , y) = y Ky + e
j Kej −

2ej Ky where ej is the j-th vector of the canonical basis and y is a vector of N
coefficients. As in standard mean shift, the shifts are obtained by maximizing
the lower bound
6
The interpretation is discussed later.
(Kernel) Mean Shift (Kernel) Medoid Shift

function Z = meanshift(G, sigma) function map = medoidshift(G, sigma)
[d,N] = size(G) ; [d,N] = size(G) ;

oN = ones(N,1) ; oN = ones(N,1) ;
od = ones(d,1) ; od = ones(d,1) ;
n = (G’.*G’)*od ; n = (G’.*G’)*od ;
Z = G ; D = n*oN’ + oN*n’ - 2*(G’*G) ;

T = 100 ; F = - exp(- .5 * D’ / sigma^2) ;
for t=1:T Q = n * (oN’*F) - 2 * G’ * (G*F) ;
m = (Z’.*Z’)*od ;
D = m*oN’ + oN*n’ - 2*(Z’*G) ; [drop,map] = max(Q) ;
F = - exp(- .5 * D’ / sigma^2) ;
Y = F ./ (oN * (oN’*F)) ;
Z = G*Y ;
end
Fig. 3. Kernel mean and medoid shift algorithms. We show basic MATLAB
implementations of two of the proposed algorithms. Here K = G
G is a low-rank
decomposition G ∈ Rd×N of the (centered) kernel matrix and sigma is the (isotropic)
standard deviation of the Gaussian Parzen window. Both algorithms are O(dN 2 ) (for
a fixed number of iterations of mean shift), reduce to their Euclidean equivalents by
setting G ≡ X and Z ≡ Y , and can be easily modified to use the full kernel matrix K
rather than a decomposition G
G (but the complexity grows to O(N 3 )).

N
yi (t + 1) = argmax (y Ky + e
j Kej − 2ej Ky)φ̇(dH (xj , yi (t))).
2
y∈RN j=1
Deriving w.r.t. y and setting to zero yields the update equation

1
yi (t + 1) = (F ei ), Fji = φ̇(Dij ), Dij = d2H (yi (t), xj ). (11)
1 F ei
Low-rank approximation. Similarly to medoid shift, we can accelerate the

algorithm by using a low-rank decomposition K = G G of the (centered) ker-
nel matrix.
# It is $useful to switch to matrix notation for all the quantities. Let
Y = y1 , . . . yM be the trajectory matrix and define Z = GY the reduced
coordinates.7 The distance matrix D can be written compactly as
D = m1 + 1n − 2Y K = m1 + 1n − 2Z G;
where
m = (Y + Y K)1 = (Z + Z )1, n = (I + K)1 = (G + G )1.
At each iteration D is calculated in O(dN 2 ) operations. Then F = φ̇(D )/N

is evaluated component-wise. Finally the trajectories Y (or equivalently Z) are
updated by
Y ← F diag(F 1)−1 , Z ← GY.
7
Similarly, the data matrix X has reduced coordinates equal to G.
in O(N 2 ) operations. Notice that, by setting G ≡ X and Z ≡ Y in these equa-

tions, we obtain Euclidean mean shift back. See Fig. 3 for a basic implementation.
Interpretation, Regularization and Scaling. In Euclidean mean shift the

function P (x) is a non-parametric estimate of a probability density. Does the
same interpretation hold in kernel space? For any fixed data set of size N , we can
restrict our attention to the subspace space spanH X ⊂ H and interpret P (x) as
a probability density on this finite-dimensional space. Unfortunately, the num-
ber of dimensions of this space may be as large as the number of data points N ,
which makes the Parzen density estimate P (x) inconsistent (in the sense that
the variance does not converge to zero as N → ∞). So how do we make sense
of kernel mean shift? The idea is to use the fact that most of the dimensions
of spanH X are often unimportant. Formally, consider the eigen-decomposition
K = V ΣV = G G, G = Σ 2 V of the (centered) kernel matrix K. Assume
1
1
Σ 2 = N diag(σ1 , . . . , σN ), with σ1 ≥ σ2 ≥ · · · ≥ σN . According to this decompo-
sition, vectors x, y ∈ spanH X can be identified with their coordinates
# $g, z ∈ R
N
so that x, yH = g, z. Moreover the data matrix G = g1 . . . gn has null
mean8 and covariance GG /N = Σ/N = σ12 diag(λ21 , . . . , λ2N ). If λi decay fast,
the effective dimension of the data can be much smaller than N .
The simplest way to regularize the Parzen estimate is therefore to discard the
dimensions above some index d (which also improves efficiency). Another option
is to blur the coordinates z by adding a small Gaussian noise η of isotropic
standard deviation , obtaining a regularized variable z = z+η. The components
of z with smaller variance are “washed out” by the noise, and we canobtain a
consistent estimator of z by using the regularized Parzen estimate i=1 (g ∗
N
k)(zi ) (the same idea is implicitly used, for instance, in kernel Fisher discriminant
analysis [15], where the covariance matrix computed in kernel space is regularized
by the addition of 2 I). This suggests that using a kernel with sufficient isotropic
smoothing may be sufficient.
Finally, we note that, due to the different scalings λ1 , . . . , λN of the linear
dimensions, it might be preferable to use an adapted Parzen window, which
retains the same proportions [17]. This, combined . with the regularization ,
suggests us to scale each axis of the kernel by σ 2 λ2i + 2 .9
8
Because K is assumed to be centered, so that 1
G
(G1) = 1
K1 = 0.
9
So far we disregarded the normalization constant of the Parzen window k(x) as it
was irrelevant for our purposes. If, however, windows kσ (x) of variable width σ are
used [6], then the relative weights of the windows become important. Recall that in
the d dimensional Euclidean case one has kσ (0)/kσ (0) = (σ /σ)d . In kernel space
therefore one would have
?
@N
kσ (0) @ σ 2 λ2 + 2
=A i
2 λ2 + 2
.
kσ (0) i=1
σ i
Fig. 4. Clustering on a manifold. By using kernel ISOMAP we can apply kernel

mean and medoid shift to cluster points on a manifold. For the sake of illustration,
we reproduce an example from [20]. From left to right: Kernel mean shift (7.8s), non-
iterated kernel medoid shift (0.18s), iterated kernel medoid shift (0.48s), quick shift
(0.12s). We project the kernel space to three dimensions d = 3 as the residual dimen-
sions are irrelevant. All algorithms but non-iterated medoid shift segment the modes
successfully. Compared to [20], medoid shift has complexity O(dN 2 ), (with a small con-
stant and d = 3 N ) instead of O(N 3 ) (small constant) or O(N 2.38 ) (large constant).
5 Applications
5.1 Clustering on Manifolds
[20] applies medoid shift to cluster data on manifolds, based on the distance
matrix D calculated by ISOMAP. If the kernel matrix K = HDH /2, H =
I − N1 11 is p.d., we can apply directly kernel mean or medoid shift to the same
problem. If not, we can use the technique from [4] to regularize the estimate and
enforce this property. In Fig. 4 this idea is used to compare kernel mean shift,
kernel medoid shift and quick shift in a simple test case.
5.2 Image Segmentation
Image segmentation is a typical test case for mode seeking algorithms [5,16,20].
Usually mode seeking is applied to this task by clustering data {(p, f (p)), p ∈ Ω},
where p ∈ Ω are the image pixels and f (p) their color coordinates (we use the
same color space of [5]).
As in [5], we apply mean shift to segment the image into super-pixels (mean
shift variants can be used to obtain directly full segmentations [2,16,24]). We
compare the speed and segmentation quality obtained by using mean shift,
medoid shift, and quick shift (see Fig. 5 for further details).
Mean shift is equivalent to [5] and can be considered a reference to evaluate the
other segmentations. Non-iterative medoid shift (first column) over-fragments
significantly (see also Fig. 2), which in [20] is addressed by reiterating the algo-
rithm. However, since our implementation is only O(dN 2 ), medoid shift has at
least the advantage of being much faster than mean shift, and can be used to
speed up the latter. In Fig. 5 we compare the time required to run mean shift
from scratch and from the modes found by medoid shift. We report the speedup
(as the number of modes found by medoid shift over the number of pixels), the
Medoid Mean Quick 1 Quick 2

0.6s 6.5s (5.9s) 0.5s 0.5s
σ = 2 spd = 1.1 σ=2 σ = 2 τ = 6.7 σ = 2 τ = 11.3
3.5s 37.5s (34.0s) 2.7s 2.7s
σ = 5 spd = 2.5 σ=5 σ = 5 τ = 16.7 σ = 5 τ = 28.3
7.1s 33.4s (26.3s) 5.5s 5.5s
σ = 8 spd = 5.7 σ=8 σ = 8 τ = 26.7 σ = 8 τ = 45.3
Fig. 5. Image segmentation. We compare different mode seeking techniques for seg-
menting an image (for clarity we show only a detail). We report the computation time
in seconds (top-right corner of each figure). In order to better appreciate the intrin-
sic efficiency advantages of each method, we use comparable vanilla implementations
of the algorithms (in practice, one could use heuristics and advanced approximation
techniques [23] to significantly accelerate the computation). We use a Gaussian kernel
of isotropic standard deviation σ in the spatial domain and use only one optimization:
We approximate the support of the Gaussian window by a disk of radius 3σ (in the
spatial domain) which results in a sparse matrix F . Therefore the computational effort
increases with σ (top to bottom). The results are discussed in the text.
computation time of medoid+mean shift and, in brackets, the computation time

of the mean shift part only. Interestingly, the efficiency increases for larger σ, so
that the overall computation time actually decreases when σ is large enough.
Finally, we show the result of quick shift segmentation (last two columns)
for increasing values of the regularization parameter τ . Notice that quick shift
is run only once to get both segmentations (Sect. 3) and that the algorithm
is in practice much faster than the other two, while still producing reasonable
super-pixels.
5.3 Clustering Bag-of-Features
The interesting work [11] introduces a large family of positive definite kernels
for probability measures which includes many of the popular metrics: χ2 ker-
nel, Hellinger’s kernel, Kullback-Leibler kernel and l1 kernel. Leveraging on
these ideas, we can use kernel mean shift to cluster probability measures, and in
Algorithm Scal. Rate [%] Time [s]

mean no 93.9 29.9
mean yes 94.6 89.0
mean+medoid no 93.9 3.63
mean+medoid yes 94.5 6.28
quick no 96.3 0.62
quick yes 96.1 0.70
Fig. 6. Automatic visual categorization. We use kernel mean shift to cluster bag-
of-features image descriptors of 1600 images from Caltech-4 (four visual categories:
airplanes, motorbikes, faces, cars). Top. From left to right, iterations of kernel mean
shift on the bag-of-features signatures. We plot the first two dimensions of the rank-
reduced kernel space (z vectors) and color the points based on the ground truth labels.
In the rightmost panel the data converged to five points, but we artificially added
random jitter to visualize the composition of the clusters. Bottom. Samples from the
five clusters found (notice that airplane are divided in two categories). We also report
the clustering quality, as the percentage of correct labels compared to the ground truth
(we merge the two airplanes categories into one), and the execution time. We use basic
implementations of the algorithms, although several optimizations are possible.
particular histograms, such as the ones arising in bag-of-features [7] or similar

representations. In the rest of the section we experiment with the χ2 kernel

B
xb yb
Kχ2 (x, y) = 2
xb + yb
b=1
where x and y are histograms of B bins.

Inspired by [10], we attempt to automatically infer the object categories of
Caltech-4 in a completely unsupervised setting. We select at random 1600 im-
ages from the categories bike, airplanes, cars and faces. Instead of the more
sophisticated representation of [10], we compute a basic bag-of-feature image
representation as suggested by [25]: We extract multiscale Harris and DoG
interest points (of fixed orientation; see [25] and ref. therein) and calculate SIFT
descriptors [14], obtaining about 103 features per image. We then generate a vo-
cabulary of 400 visual words by clustering a random selection of such descriptors
by using k-means. For each image, we compute a bag-of-feature histogram x by
counting the number of occurrences of each visual word in that image. Finally,
we use the χ2 kernel to generate the kernel matrix, that we feed to our clustering
algorithms.
In Fig. 6 we compare kernel mean shift, kernel mean shift initialized by medoid
shift, and quick shift. The problem we solve is considerably harder than [10],
since in our case the number of clusters (categories) is unknown. All algorithms
discover five (rather than four) categories (Fig. 6), but the result is quite reason-
able since the category airplanes contains two distinct and visually quite different
populations (grounded and airborne airplanes). Moreover, compared to [10] we
do not try to explicitly separate an object from its background, but we use a
simple holistic representation of each image.
The execution time of the algorithms (Fig. 6) is very different. Mean shift
is relatively slow, at least in our simple implementation, and its speed greatly
improves when we use medoid shift to initialize it. However, consistently with
our image segmentation experiments, quick shift is much faster.
We also report the quality of the learned clusters (after manually merging the
two airplane subcategories) as the percentage of correct labels. Our algorithm
performs better than [10], that uses spectral clustering and reports 94% accuracy
on selected prototypes and as low as 85% when all the data are considered; our
accuracy in the latter case is at least 94%. We also study rescaling as proposed in
Sect. 4, showing that it (marginally) improves the results of mean/medoid shift,
but makes the convergence slower. Interestingly, however, the best performing
algorithm (not to mention the fastest) is quick shift.
6 Conclusions
In this paper we exploited kernels to extend mean shift and other mode seeking
algorithms to a non-Euclidean setting. This also clarifies issues of regularization
and data scaling when complex spaces are considered. In this context, we showed
how to derive a very efficient version of the recently introduced medoid shift
algorithm, whose complexity is lower than mean shift. Unfortunately, we also
showed that medoid shift often results in over-fragmented clusters. Therefore,
we proposed to use medoid shift to initialize mean shift, yielding a clustering
algorithm which is both efficient and accurate.
We also introduced quick shift, which can balance under- and over-fragmentation
of the clusters by the choice of a real parameter. We showed that, in practice, this
algorithm is very competitive, resulting in good (and sometimes better) segmenta-
tions compared to mean shift, at a fraction of the computation time.
Acknowledgment. Supported by AFOSR FA9550-06-1-0138 and ONR N00014-

08-1-0414.
References
1. Bach, F.R., Jordan, M.I.: Kernel independent componet analysis. Journal of Ma-
chine Learninig Research 3(1) (2002)
2. Carreira-Perpiñán, M.: Fast nonparametric clustering with gaussian blurring mean-
shift. In: Proc. ICML (2006)
3. Cheng, Y.: Mean shift, mode seeking, and clustering. PAMI 17(8) (1995)
4. Choi, H., Choi, S.: Robust kernel isomap. Pattern Recognition (2006)
5. Comaniciu, D., Meer, P.: Mean shift: A robust approach toward feature space
analysis. PAMI 24(5) (2002)
6. Comaniciu, D., Ramesh, V., Meer, P.: The variable bandwidth mean shift and
data-driven scale selection. In: Proc. ICCV (2001)
7. Csurka, G., Dance, C.R., Dan, L., Willamowski, J., Bray, C.: Visual categorization
with bags of keypoints. In: Proc. ECCV (2004)
8. Fine, S., Scheinberg, K.: Efficient SVM training using low-rank kernel representa-
tions. Journal of Machine Learninig Research (2001)
9. Fukunaga, K., Hostler, L.D.: The estimation of the gradient of a density function,
with applications in pattern recognition. IEEE Trans. 21(1) (1975)
10. Grauman, K., Darrell, T.: Unsupervised learning of categories from sets of partially
matching image features. In: Proc. CVPR (2006)
11. Hein, M., Bousquet, O.: Hilbertian metrics and positive definite kernels on proba-
bility measures. In: Proc. AISTAT (2005)
12. Knuth, D.: The Art of Computer Programming: Seminumerical Algorithms, 3rd
edn., vol. 2 (1998)
13. Koontz, W.L.G., Narendra, P., Fukunaga, K.: A graph-theoretic approach to non-
parametric cluster analyisis. IEEE Trans. on Computers c-25(9) (1976)
14. Lowe, D.: Implementation of the scale invariant feature transform (2007),
http://www.cs.ubc.ca/∼ lowe/keypoints/
15. Mika, S., Rätsch, G., Weston, J., Schölkopf, B., Müller, K.-R.: Fisher discrimi-
nant analysis with kernels. In: Proc. IEEE Neural Networks for Signal Processing
Workshop (1999)
16. Paris, S., Durand, F.: A topological approach to hierarchical segmentation using
mean shift. In: Proc. CVPR (2007)
17. Sain, S.R.: Multivariate locally adaptive density estimation. Comp. Stat. and Data
Analysis, 39 (2002)
18. Schölkopf, B.: The kernel trick for distances. In: Proc. NIPS (2001)
19. Schölkopf, B., Smola, A.J.: Learning with Kernels. MIT Press, Cambridge (2002)
20. Sheikh, Y.A., Khan, E.A., Kanade, T.: Mode-seeking by medoidshifts. In: Proc.
CVPR (2007)
21. Subbarao, R., Meer, P.: Nonlinear mean shift for clustering over analytic manifolds.
In: Proc. CVPR (2006)
22. Sundaramoorthy, G., Yezzi, A., Mennucci, A.: Sobolev active contours. Int. J. Com-
put. Vision 73(3) (2007)
23. Yang, C., Duraiswami, R., Gumerov, N.A., Davis, L.: Improved fast Gauss trans-
form and efficient kernel density estimation. In: Proc. ICCV (2003)
24. Yuan, X., Li, S.Z.: Half quadric analysis for mean shift: with extension to a se-
quential data mode-seeking method. In: Proc. CVPR (2007)
25. Zhang, J., Marszalek, M., Lazebnik, S., Schmid, C.: Local features and kernels for
classification of texture and object categories: A comprehensive study IJCV (2006)
A Fast Algorithm for Creating a Compact and
Discriminative Visual Codebook
Lei Wang1 , Luping Zhou1 , and Chunhua Shen2

1
RSISE, The Australian National University, Canberra ACT 0200, Australia
2
National ICT Australia (NICTA) , Canberra ACT 2601, Australia
Abstract. In patch-based object recognition, using a compact visual

codebook can boost computational efficiency and reduce memory cost.
Nevertheless, compared with a large-sized codebook, it also risks the loss
of discriminative power. Moreover, creating a compact visual codebook
can be very time-consuming, especially when the number of initial visual
words is large. In this paper, to minimize its loss of discriminative power,
we propose an approach to build a compact visual codebook by maxi-
mally preserving the separability of the object classes. Furthermore, a
fast algorithm is designed to accomplish this task effortlessly, which can
hierarchically merge 10,000 visual words down to 2 in ninety seconds.
Experimental study shows that the compact visual codebook created in
this way can achieve excellent classification performance even after a
considerable reduction in size.
1 Introduction
Recently, patch-based object recognition has attracted particular attention and
demonstrated promising recognition performance [1,2,3,4]. Typically, a visual
codebook is created as follows. After extracting a large number of local patch
descriptors from a set of training images, k-means or hierarchical clustering is
often used to group these descriptors into n clusters, where n is a predefined
number. The center of each cluster is called “visual word”, and a list of them
forms a “visual codebook”. By labelling each descriptor of an image with the
most similar visual word, this image is characterized by an n-dimensional his-
togram counting the number of occurrences of each word. The visual codebook
can have critical impact on recognition performance. In the literature, the size
of a codebook can be up to 103 or 104 , resulting in a very high-dimensional
histogram.
A compact visual codebook has advantages in both computational efficiency
and memory usage. For example, when linear or nonlinear SVMs are used, the
complexity of computing the kernel matrix, testing a new image, or storing the
support vectors is all proportional to the codebook size, n. Also, many algo-
rithms working well in a low dimensional space will encounter difficulties such

National ICT Australia is funded by the Australian Government’s Backing Aus-
tralia’s Ability initiative, in part through the Australian Research Council. The au-
thors thank Richard I. Hartley for many insightful discussions.

720 L. Wang, L. Zhou, and C. Shen
as singularity or unreliable parameter estimate when the dimensions increase.

This is often called the “curse of dimensionality”. A compact visual codebook
provides a lower-dimensional representation and can effectively avoid these dif-
ficulties. Moreover, in patch-based object recognition, the histogram used to
represent an image is essentially a discrete approximation of the distribution
of visual words in that image. A large-sized visual codebook may overfit this
distribution, as pointed out in [5]. Pioneering work of creating a compact and
discriminative visual codebook has been seen recently in [4], which hierarchically
merges the visual words in a large-sized initial codebook. To minimize the loss of
discriminative ability, the work in [4] requires the new histograms to maximize
the conditional probability of the true labels of training images (or image regions
in their work). This is a rigorous but complicated criterion that involves non-
trivial computation after each merging operation. Moreover, at each level of the
hierarchy, the optimal pair of words to be merged are sought by an exhaustive
search. These lead to a heavy computational load when dealing with large-sized
initial codebooks.
Creating a compact codebook is essentially a dimensionality reduction prob-
lem. To preserve the discriminative power, any classification performance re-
lated criterion may be adopted, for example, the rigorous Bayes error rate, error
bounds or distances, class separability measure, or that used in [4]. We pay
particular interest to the class separability measure because of its simplicity
and efficiency. By using this measure, we build a compact visual codebook that
maximally preserves the separability of the object classes. More importantly, we
propose a fast algorithm to accomplish this task effortlessly. By this algorithm,
the class separability measure can be immediately evaluated once two visual
words are merged. Also, searching for the optimal pair of words to be merged is
cast as a 2D geometry problem and testing a small number of pairs is sufficient
to find the optimal pair. Given an initial codebook of 10,000 visual words, the
proposed fast algorithm can hierarchically merge them down to 2 words in ninety
seconds. As experimentally demonstrated, our algorithm can produce a compact
codebook which is comparable to or even better than that obtained by [4], but
our algorithm needs much less computational overhead, especially when the size
of the initial codebook is large.
2 The Scatter-Matrix Based Class Separability Measure

This measure involves the Within-class scatter matrix (H), the Between-class
scatter matrix (B), and the Total scatter matrix (T). Let (x, y) ∈ (Rn × Y)
denote a training sample, where Rn stands for an n-dimensional input space,
and Y = {1, 2, · · · , c} is the set of c class labels. The number of samples in the
i-th class is denoted by li . Let mi be the mean vector of the i-th class and m
be the mean vector of all classes. The scatter matrices are defined as
c % &
H= li
j=1 (xij − mi )(xij − mi )
i=1
B= c
l%i (mi − m)(mi − m)
& (1)
i=1 li
T = ci=1 j=1 (xij − m)(xij − m) =H+B .

A Fast Algorithm 721
A large class separability means small within-class scattering but large between-
class scattering. A combination of two of them can be used as a measure, for
example, tr(B)/tr(T) or |B|/|H|, where tr(·) and | · | denote the trace and de-
terminant of a matrix, respectively. In these measures the scattering of data
is evaluated through the mean and variance, which implicitly assumes a Gaus-
sian distribution for each class. This drawback is overcome by incorporating
the kernel trick and it makes the scatter-matrix based measure quite useful, as
demonstrated in Kernel based Fisher Discriminate Analysis (KFDA) [6].
3 The Formulation of Our Problem

Given an initial codebook of n visual words, we aim to obtain a codebook con-
sisting of m (m ) n) visual words in the sense that when represented with these
m visual words, the c object classes can have maximal separability.
Recall that with a set of visual words, a training image can be represented
by a histogram which contains the number of occurrences of each word in this
image. Let xn (xn ∈ Rn ) and xm (xm ∈ Rm ) denote the histograms when n
and m visual words are used, respectively. In the following, we first discuss an
ideal way of solving our problem, and show that such a way is impractical for
patch-based object recognition. This motivates us to propose the fast algorithm
in this paper.
Inferring m visual words from the n initial ones is essentially a dimensionality
reduction problem. It can be represented by a linear transform as
xm = W xn (2)
where W (W ∈ Rn×m ) is an n × m matrix. Let Bn and Tn denote the between-
class and total -class scatter matrices when the training images are represented
by xn . The optimal linear transform, W , can be expressed as
tr(W Bn W)
W = arg max . (3)
W∈R n×m tr(W Tn W)
Note that the determinant-based measure is not adopted because n is often
much larger than the number of training images, making |Bn | and |Tn | zero.
The problem in (3) has been studied in [7] recently1 . The optimal W is located
by solving a series of Semi-Definite Programming (SDP) problems. Nevertheless,
this SDP-based approach quickly becomes intractable when n exceeds 100, which
is far less than the number encountered in practical object recognition. Moreover,
the W in patch-based object recognition may have the following constraints:
1. Wij ∈ {0, 1} if requiring the m new visual words to have meaningful and
determined content;2
1
Note that this problem is not simply the Fisher Discriminant Analysis problem.
Please see [7] for the details.
2
For example, when discriminating motorbikes from airplanes, the content of a visual
word will be “handle bar” and/or “windows” rather than 31% handle bar, 27%
windows, and 42% something else.
m
2. j=1 Wij = 1 if requiring that each of the n visual words only be assigned
to one of the m visual words.
3. If no words are to be discarded, the constraint of ni=1 Wij ≥ 1 will be
imposed because each of the n visual words must be assigned to one of the
m visual words;
This results in a large-scale integer programming problem. Efficiently and opti-
mally solving it may be difficult for the state-of-the-art optimization techniques.
In this paper, we adopt a suboptimal approach that hierarchically merges two
words while maximally maintaining the class separability at each level.
4 A Fast Algorithm of Hierarchically Merging Visual

Words
To make the hierarchical merging approach efficient, we need: i) Once two visual
words are merged, the resulting class separability can be quickly evaluated; ii)
In searching for the best pair of words to merge, the search scope has to be as
small as possible. In the following, we show how these requirements are achieved
with the scatter-matrix based class separability measure.
4.1 Fast Evaluation of Class Separability

Let xti = [xti1 , · · · , xtit ] (i = 1, · · · , l) be the i-th training image when t visual
words are used, where t (t = n, n − 1, · · · , m) indicates the current level in the
hierarchy. Let Kt be the Gram matrix defined by {Kt }ij = xti , xtj . Let Kt−1 rs
be the resulting Gram matrix after merging the r-th and s-th words at level t.
Their relationship is derived as
t−1
rs }ij = xi
{Kt−1 t−1
, xt−1 = k=1 xt−1 ik xjk
t−1
t j
= k=1 xtik xtjk − xtir xtjr − xtis xtjs + (xtir + xtis )(xtjr + xtjs )
t (4)
= k=1 xtik xtjk + xtir xtjs + xtis xtjr
= {Kt }ij + {Atrs }ij + {Atrs }ji
where Atrs is a matrix defined as Atrs = Xtr (Xts ) , where Xtr is [xt1r , · · · , xtlr ].
Hence, it can be obtained that
t
rs = K + Ars + (Ars ) .
Kt−1 t t
(5)
A similar relationship exists between the class separability measures at t and
t − 1 levels. Let Bt−1 and Tt−1 be the matrices B and T computed with xt−1 .
It can be proven (the proof is omitted) that for a c-class problem,
c
1
Kt−1
rs,i 1 1
Kt−1
rs 1 1
Kt−1
rs 1
tr(Bt−1
rs ) = − ; rs ) = tr(Krs ) −
tr(Tt−1 t−1
(6)
i=1
l i l l
rs,i is computed by the training images from class i. It can be verified

where Kt−1

rs,i = Ki + Ars,i + (Ars,i ) . The li is the number of training images from
that Kt−1 t t t
class i, and l is the total number. Note that 1 Atrs 1 = 1 (Atrs ) 1, where 1 is
a vector consisting of “1”. By combining (5) and (6), we obtain that

1 Kti 1 1 Kt 1
c 1 Atrs,i 1 1 Atrs 1
tr(Bt−1
rs ) =
c
i=1 li
− l
+2 i=1 li
− l

c 1 Atrs,i 1 1 Atrs 1 (7)
= tr(Bt ) + 2 i=1 li
− l
tr(Bt ) + f (Xtr , Xts ),
where f (Xtr , Xts ) denotes the second term in the previous step. Similarly,

1 Kt 1 1 Atrs 1
rs ) = tr(K ) −
tr(Tt−1 t
+ 2 tr(Atrs ) +

l
l
1 Atrs 1 (8)
= tr(Tt ) + 2 tr(Atrs ) − l
tr(Tt ) + g(Xtr , Xts ) .
Since both tr(Bt ) and tr(Tt ) have been computed at level t before any merging
operation, the above results indicate that to evaluate the class separability after
merging two words, only f (Xtr , Xts ) and g(Xtr , Xts ) need to be calculated.
In the following, we further show that at any level t (m ≤ t < n), f (Xtr , Xts )
and g(Xtr , Xts ) can be worked out with little computation. Three cases are dis-
cussed in turn.
i) Neither the r-th nor the s-th visual word is newly generated at level t.
This means that both of them are directly inherited from level t+1. Assuming
that they are numbered as p and q at level t + 1, it can be known that
f (Xtr , Xts ) = f (Xt+1

p , Xq );
t+1
(9)
ii) Just one of the r-th and the s-th visual words is newly generated at level t.
Assume that the r-th visual word is newly generated by merging the u-th
and the v-th words at level t + 1, that is, Xtr = Xt+1
u + Xt+1
v . Furthermore,
assume that Xts is numbered as q at level t + 1. It can be shown that
Atrs = Xtr (Xts ) = (Xt+1

u + Xt+1 t+1
v )(Xq )
t+1 t+1 t+1 t+1
= Xu (Xq ) + Xv (Xq ) (10)
uq + Avq .
= At+1 t+1
In this way, it can be obtained that

c 1 Atrs,i 1 1 Atrs 1
f (Xtr , Xts ) = 2 i=1 li
− l

c 1 At+1 1 1 t+1
A 1 1 At+1
vq,i 1 1 At+1
vq 1 (11)
=2 i=1
uq,i
li
− uq
l
+2 c
i=1 li
− l
= f (Xt+1 t+1
u , Xq ) + f (Xt+1 t+1
v , Xq );
iii) Both the r-th and the s-th visual words are newly generated at level t.
This case does not exist because only one visual word can be newly generated
at each level of a hierarchical clustering.
The above analysis shows that f (Xtr , Xts ) can be obtained either by directly
copying from level t + 1 or by a single addition operation. All of the analysis
applies to g(Xtr , Xts ). Hence, once the r-th and the s-th visual words are merged,
the class separability measure, tr(Bt−1 t−1
rs )/tr(Trs ), can be immediately obtained
by two addition and one division operations.
Computational complexity. The time complexity of calculating f (Xni , Xnj ) or

g(Xni , Xnj ) is analyzed. There are n(n − 1)/2 values to be computed in total,
each of which involves computing the matrix Anij which needs l2 multiplications.
Both terms of 1 Anij,k 1 (k = 1, 2, · · · , c) and 1 Anij 1 can be obtained by l2
c
additions. Finally, i=1 ( l1i )1 Anij,k 1 + (− 1l )1 Anij 1 can be worked out in c + 1
multiplications and c additions. Hence, computing all f (Xni , Xnj ) or g(Xni , Xnj )
needs
n(n − 1) # 2 $
(l + c + 1) multiplications + (l2 + c) additions ,
2
resulting in the complexity of O(n2 l2 ). In practice, the load of computing Anij
can be lower because the histogram xn is often sparse. Also, f (Xni , Xnj ) and
g(Xni , Xnj ) share the same Anij . The memory cost for storing all of the f (Xni , Xnj )
and g(Xni , Xnj ) in double precision format is n(n − 1) × 8 Bytes, leading to space
complexity of O(n2 ). When n equals 10, 000 (this is believed to be a reasonably
large size for an initial visual codebook used in patch-based object recognition),
the memory cost will be about 800 MByte, which is bearable for a desktop
computer today. Moreover, the memory cost decreases quadratically with respect
to the level because the total number of f or g is t(t − 1)/2 at a given level t.
4.2 Fast Search for the Optimal Pair of Words to Merge

Although the class separability can now be quickly evaluated once a pair of
words are merged, there are t(t−1)
2 possible pairs at level t from which we need
to find the optimal pair to merge. If an exhaustive search is used to identify
this optimal pair, the total number of pairs that are tested in the hierarchical

merging process will be nt=m+1 t(t−1)2 . For n = 10, 000 and m = 2, this number
is as large as 1.67 × 1011 . Using an exhaustive search will significantly prolong
the merging process. In the following, we propose a more efficient search strategy
by making use of the properties of the scatter-matrix based class separability
measure, which allows us to convert the search problem to a simple 2D geometry
problem. Denote f (Xtr , Xts ) and g(Xtr , Xts ) by f t and g t in short, respectively.
Recall that the class separability measure after merging two visual words is
tr(Bt−1 ) tr(Bt ) + f t f t − (−tr(Bt ))
J = = = t
t−1
tr(T ) tr(T ) + g
t t g − (−tr(Tt ))
As illustrated in Fig. 1, geometrically, the value of J equals the slope of the line
AB through A(−tr(Tt ), −tr(Bt )) and B(g t , f t ).
The coordinates of A and B are restricted by the following properties of the
scatter matrices:
i) From the definition in (1), it is known that
tr(Ht ) ≥ 0; tr(Bt ) ≥ 0; tr(Tt ) = tr(Ht ) + tr(Bt ) ≥ tr(Bt )
As a result, the point A must lie within the third quadrant of the Cartesian
coordinate system gOf and above the line of f − g = 0. The domain of A is
marked as a hatched region in Fig. 1.
ii) The coordinator of B(g t , f t ) must satisfy the following constraints:
tr(Bt−1 ) ≥ 0 =⇒ tr(Bt ) + f t ≥ 0 =⇒ f t ≥ −tr(Bt )

tr(Tt−1 ) ≥ 0 =⇒ tr(Tt ) + g t ≥ 0 =⇒ g t ≥ −tr(Tt )
tr(Tt−1 ) ≥ tr(Bt−1 ) =⇒ tr(Tt ) + g t ≥ tr(Bt ) + f t
=⇒ f t − g t − (tr(Tt ) − tr(Bt )) ≤ 0
They define three half-planes in the coordinate system gOf and the point
B(g t , f t ) must lie within the intersection, the blue-colored region in Fig. 1.
Therefore, finding the optimal pair of words whose combination produces the
largest class separability becomes finding the optimal point B which maximizes
the slope of the line AB, where the coordinate of A is fixed at a given level t.
Fig. 1. Illustration of the region where A(−tr(Tt ), −tr(Bt )) and B(g t , f t ) reside
Indexing structure. To realize the fast search, a polar coordinate based indexing
structure is used to index the t(t − 1)/2 points of B(g, f ) at level t, as illustrated
in Fig. 2. Each point B is assigned into a bin (i,j) according to its distance from
the origin and its polar angle, where i = 1, · · · , K and j = 1, · · · , S. The K is
the number of bins with respect to the distance from the origin, whereas S is the
number of bins with respect to the polar angle. In Fig. 2, this indexing structure
is illustrated by K concentric circles, each of which is further divided into S
segments. The total number of bins is KS. Through this indexing structure, we
can know which points B reside in a given bin. In this paper, the number of
circles K is set as 40, and their radius are arranged as ri = ri+1 /2. The S is set
as 36, which evenly divides [0, 2π) into 36 bins.
Fig. 2. The point A is fixed when searching for B which makes the line AB have the
largest slope. The line AD is tangent to the second largest circle CK−1 at D, and it
divides the largest circle CK into two parts, region I and II. Clearly, a point B in region
I always gives AB a larger slope than any point in region II. Therefore, if the region I
is not empty, the best point B must reside there and searching region I is sufficient.
Search strategy. As shown in Fig. 2, let D denote the point where the line AD is
tangent to the second largest circle, CK−1 . The line AD divides the largest circle
CK into two parts. When connected with A, a point B lying above AD (denoted
by region I) always gives a larger slope than any point below it (denoted by
region II). Therefore, if the region I is not empty, all points in the region II
can be safely ignored. The search is merely to find the best point B from the
region I which gives AB the largest slope. To carry out this search, we have to
know which points reside in the region I. Instead of exhaustively checking each of
the t(t−1)
2 points against AD, this information is conveniently obtained via the
above indexing structure. Let θE and θF be the polar angles of E and F where
the line AD and CK intersect. Denote the bins (with respect to the polar angle)
into which they fall by S1 and S2 , respectively. Thus, searching the region I can
be accomplished by searching the bin (i, j) with i = K and j = S1 , · · · , S2 .3
Clearly, the area of the searched region is much smaller than the area of CK
for moderate K and S. Therefore, the number of points B(g, f ) to be tested
can be significantly reduced, especially when the point B distributes sparsely in
the areas away from the origin. If the region I is empty, move the line AD to
be tangent to the next circle, CK−2 , and repeat the above steps. After finding
the optimal pair of words and merging them, all points B(g, f ) related to the
two merged words will be removed. Meanwhile, new points related to the newly
3
The region that is actually searched is slightly larger than the region I. Hence, the
found best point B will be rejected if it is below the line AD. This also means that
the region I is actually empty.
generated word will be added and indexed. This process is conveniently realized
in our algorithm by letting one word “absord” the other. Then, we finish the
operation at level t and move to level t−1. Our algorithm is described in Table 1.
Before ending this section, it is worth noting that this search problem may
be tackled by the dynamic convex hull [8] in computational geometry. Given
the point A, the best point B must be a vertex of the convex hull of the
points B(g, f ). At each level t, part of points B(g, f ) are updated, resulting in a
dynamically changing convex hull. The technique of dynamic convex hull can be
used to update the vertex set accordingly. This will be explored in future work.
Table 1. The fast algorithm for hierarchically merging visual words
Input: The l training images represented as {(xi , yi )}li=1 (xi ∈ Rn , yi ∈ {1, · · · , c}).
The n is the size of an initial visual codebook and yi is the class label of xi
m: the size of the target visual codebook.
Output: The n − m level merging hierarchy
Initialization:
compute f (Xn i , Xj ) and g(Xi , Xj ) (1 ≤ i < j ≤ n) and store them in memory
n n n
n(n−1)
Index the 2
points of B(g, f ) with a polar coordinate quantized into bins
Compute A(−tr(Tn ), −tr(Bn ))
Merging operation:
for t = n, n − 1, · · · , m
(1) fast search for the point B(g , f ) that gives the line AB the largest
slope, where f = f (Xtr , Xts ) and g = g (Xtr , Xts )
(2) compute tr(Bt−1 ) and tr(Tt−1 ) and update the point A:
tr(Bt−1 ) = tr(Bt ) + f (Xtr , Xts ); tr(Tt−1 ) = tr(Tt ) + g (Xtr , Xts )
(3) update f (Xtr , Xti ) and g(Xtr , Xti )
f (Xtr , Xti )=f (Xtr , Xti ) + f (Xts , Xti ); g(Xtr , Xti )=g(Xtr , Xti ) + g(Xts , Xti )
remove f (Xts , Xti ) and g(Xts , Xti )
(4) re-index f (Xtr , Xti ) and g(Xtr , Xti )
end
5 Experimental Result
The proposed class separability measure based fast algorithm is tested on four
classes of the Caltech-101 object database [9], including Motorbikes (798 images),
Airplanes (800), Faces easy (435), and BACKGROUND Google (520), as shown
in Fig. 3. A Harris-Affine detector [10] is used to locate interest regions, which
are then represented by the SIFT descriptor [11]. Other region detectors [12] and
descriptors [13] can certainly be used because our algorithm has no restriction
on this. The number of local descriptors extracted from the images of the four
classes are about 134K, 84K, 57K, and 293K, respectively. Our algorithm is
applicable to both binary and multi-class problems. This experiment focuses

on the binary case, including both object categorization and object detection
problems. To accumulate statistics, the images of the two object classes to be
classified are randomly split as 10 pairs of training/test subsets. Restricted to
the images in a training subset (those in a test subset are only used for test),
their local descriptors are clustered to form the n initial visual words by using
k-means clustering. Each image is then represented by a histogram containing
the number of occurrences of each visual word.
Fig. 3. Example images of Motorbikes, Airplanes, Faces easy, and BACK-

GROUND Google in [9] used in this experiment
Three algorithms are compared in creating a compact visual codebook, includ-

ing k-means clustering (KMS in short), the algorithm proposed in [4] (PRO in
short), and our class separability measure (CSM in short) based fast algorithm.
In this experiment, the k-means clustering is used to cluster the local descrip-
tors of the training images by gradually decreasing the value of k. Its result is
used as a baseline. The CSM and PRO are applied to the initial n-dimensional
histograms to hierarchically merge the visual words (or equally, the bins). For
each algorithm, the obtained lower-dimensional histograms are used by a clas-
sifier to separate the two object classes. Linear and nonlinear SVM classifiers
with a Gaussian RBF kernel are used. Their hyper-parameters are tuned via
k-fold cross-validation. The three algorithms are compared in terms of: i) the
time and memory cost with respect to the number of initial visual words; ii) the
recognition performance achieved by the obtained compact visual codebooks.
We aim to show that our proposed CSM-based fast algorithm can achieve the
recognition performance comparable to or even better than the PRO algorithm
but it is much faster in creating a compact codebook.
5.1 Result on Time and Memory Cost

The time and memory cost is independently evaluated on a synthetic data set.
Fixing the number of training images at 100, the size of the initial visual code-
book varies between 10 and 10,000. The number of occurrences of each visual
Comparison of time cost Comparison of memory cost

150 1500
Our CSM algorithm with the proposed fast search Our CSM algorithm with the proposed fast search
Our CSM algorithm with an exhaustive search

Our CSM algorithm with an exhaustive search
The PRO algorithm in [4]
Time cost (in minutes)
Memory cost (in MB)

100 1000
50 500
0 0
2 2.5 3 3.5 4 2 2.5 3 3.5 4

Logarithm (base 10) of the size of an initial codebook Logarithm (base 10) of the size of an initial codebook
(a) (b)
Fig. 4. Time and peak memory cost Comparison of our CSM algorithm (using
the proposed fast search or an exhaustive search) and the PRO algorithm in [4]. The
horizontal axis is the size (in logarithm) of an initial visual codebook, while the vertical
axes are time and peak memory cost in (a) and (b), respectively. As shown, the CSM
algorithm with the fast search significantly reduces the time cost for a large-sized visual
codebook with acceptable memory usage.
word used in a histogram is randomly sampled from {0, 1, 2, · · · , 99}. In this ex-
periment, the CSM-based fast algorithm is compared with the PRO algorithm
which uses an exhaustive search to find the optimal pair of words to merge. We
implement the PRO algorithm according to [4], including a trick suggested to
speed up the algorithm by only updating the terms related to the two words to
be merged. Meanwhile, to explicitly show the efficiency of the fast search part in
our algorithm, we purposely replace the fast search in the CSM-based algorithm
with an exhaustive search to demonstrate the quick increase on time cost. A ma-
chine with 2.80GHz CPU and 4.0GB memory is used. The result is in Fig. 4. As
seen in sub-figure(a), the time cost of the PRO algorithm goes up quickly with
the increasing codebook size. It takes 1, 624 seconds to hierarchically cluster 1000
visual words to 2, whereas the CSM algorithm with an exhaustive search only
uses 9 seconds to accomplish this. The less time cost is attributed to the simplic-
ity of the CSM criterion and the fast evaluation method proposed in Section 4.1.
The CSM algorithm with the fast search achieves the highest computational ef-
ficiency. It only takes 1.55 minutes to hierarchically merge 10,000 visual words
to 2, and the time cost increases to 141.1 minutes when an exhaustive search
is used. As shown in sub-figure(b), the price is that the fast search needs more
memory (1.45GB for 10,000 visual words) to store the indexing structure. We
believe that such memory usage is acceptable for a personal computer today.
In the following experiments, the discriminative power of the obtained compact
visual codebooks is investigated.
5.2 Motorbikes vs. Airplanes

This experiment discriminates the images of a motorbike from those containing
an airplane. In each of the 10 pairs of training/test subsets, there are 959 training
Object recognition: Motorbikes vs. Airplanes Object recognition: Motorbikes vs. Airplanes
0.14 0.12
k−means clustering (KMS) k−means clustering (KMS)
Classification error (Nonlinear SVM)

Classification error (Linear SVM)
Our CSM algorithm Our CSM algorithm

0.12 The PRO algorithm in [4] 0.1
0.1 0.08
0.08 0.06
0.06 0.04
0.04 0.02
0.02 0
900 700 500 300 100 80 60 40 20 0 900 700 500 300 100 80 60 40 20
The size of the obtained compact visual codebook The size of the obtained compact visual codebook
(a) (b)
Fig. 5. Motorbikes vs. Airplanes Comparison of classification performance of the

compact visual codebooks generated by k-means clustering (KMS), the PRO algorithm
in [4], and our class separability measure (CSM) algorithm. Linear and nonlinear SVM
classifiers are used in (a) and (b), respectively. The CSM-based algorithm still gives the
excellent classification result when the codebook size has been considerably reduced.
images and 639 test images. An initial visual codebook of size 1, 000 is created
by using k-means clustering. The CSM algorithm with the fast search hierarchi-
cally clusters them into 2 words in 6 seconds, whereas the PRO algorithm takes
6, 164 seconds to finish this. Based on the obtained compact visual codebook, a
new histogram is created to represent each image. With the new histograms, a
classifier is trained on a training subset and evaluated on the corresponding test
subset. The average classification error rate is plotted in Fig. 5. The sub-figure
(a) shows the result when a linear SVM classifier is used. As seen, the compact
codebook generated by k-means clustering has poor discriminative power. Its
classification error rate goes up with the decreasing size of the compact code-
book. This is because k-means clustering uses the Euclidean distance between
clusters as the merging criterion, which is not related to the classification perfor-
mance. In contrast, the CSM and PRO algorithms achieve better classification
performance, indicating that they well preserve the discriminative power in the
obtained compact codebooks. For example, when the codebook size is reduced
from 1000 to 20, these two algorithms still maintain excellent classification per-
formance, with an increase of error rate less than 1%. Though the classification
error rate of our CSM algorithm is a little bit higher (about 1.5%) at the initial
stage, it soon drops to a level comparable to the error rate given by the PRO
algorithm with the decreasing codebook size. Similar results can be observed
from Fig. 5(b) where a nonlinear SVM classifier is employed.
5.3 Faces Easy vs. Background Google
This experiment aims to separate the images containing a face from the back-
ground images randomly collected from the Internet. In each training/test split,
there are 100 training images and 1, 498 test images. The number of initial visual
Object detection: Faces−easy vs. BACKGROUND_Google Object detection: Faces−easy vs. BACKGROUND_Google
0.35 0.26
k−means clustering (KMS) k−means clustering (KMS)
Classification error (Nonlinear SVM)

0.24
Classification error (Linear SVM)
Our CSM algorithm

Our CSM algorithm
0.3 The PRO algorithm in [4] 0.22
0.2
0.25
0.18
0.16
0.2
0.14
0.12
0.15
0.1
0.1 0.08
900 700 500 300 100 80 60 40 20 900 700 500 300 100 80 60 40 20
The size of the obtained compact viusal codebook The size of the obtained compact visual codebook
(a) (b)
Fig. 6. Face-easy vs. Background Google Comparison of classification perfor-

mance of the small-sized visual codebooks generated by k-means clustering (KMS),
the PRO algorithm in [4], and our proposed class separability measure (CSM). Lin-
ear and nonlinear SVM classifiers are used in (a) and (b), respectively. As shown, the
CSM-based algorithm gives the best compact and discriminative codebooks.
words is 1, 000. They are hierarchically clustered into two words in 6 seconds by
our CSM algorithm with the fast search and in 1, 038 seconds by the PRO al-
gorithm. Again, with the newly obtained histograms, a classifier is trained and
evaluated. The averaged classification error rates are presented in Fig. 6. In this
experiment, the classification performance of the PRO algorithm is not as good
as before. This might be caused by the hyper-parameters used in the PRO al-
gorithm. Their values are preset according to [4] but may be task-dependent. In
contrast, our CSM algorithm achieves the best classification performance. The
small-sized compact codebooks consistently produce the error rate comparable to
that of the initial visual codebook. This indicates that our algorithm effectively
makes the compact codebooks preserve the discriminative power of the initial
codebook. An additional advantage of our algorithm is that the CSM criterion
is free of parameter setting. Meanwhile, a short “transition period” is observed
on the CSM algorithm in Fig. 6, where the classification error rate goes up and
then drops at the early stage. This interesting phenomenon will be looked into
in future work.
6 Conclusion
To obtain a compact and discriminative visual codebook, this paper proposes
using the separability of object classes to guide the hierarchical clustering of
initial visual words. Moreover, a fast algorithm is designed to avoid a lengthy
exhaustive search. As shown by the experimental study, our algorithm not only
ensures the discriminative power of a compact codebook, but also makes the
creation of a compact codebook very fast. This delivers an efficient tool for patch-
based object recognition. In future work, more theoretical and experimental
study will be conducted to analyze its performance.
References
1. Agarwal, S., Awan, A.: Learning to detect objects in images via a sparse, part-
based representation. IEEE Transactions on Pattern Analysis and Machine Intel-
ligence 26(11), 1475–1490 (2004)
2. Csurka, G., Dance, C.R., Fan, L., Willamowski, J., Bray, C.: Visual categoriza-
tion with bags of keypoints. In: Proceedings of ECCV International Workshop on
Statistical Learning in Computer Vision, pp. 1–22 (2004)
3. Jurie, F., Triggs, B.: Creating efficient codebooks for visual recognition. In: Pro-
ceedings of the Tenth IEEE International Conference on Computer Vision, vol. 1,
pp. 604–610 (2005)
4. Winn, J., Criminisi, A., Minka, T.: Object categorization by learned universal
visual dictionary. In: Proceedings of the Tenth IEEE International Conference on
Computer Vision, vol. 2, pp. 1800–1807 (2005)
5. Varma, M., Zisserman, A.: A statistical approach to texture classification from
single images. International Journal of Computer Vision 62(1-2), 61–81 (2005)
6. Mika, S., Rätsch, G., Weston, J., Schölkopf, B., Müller, K.R.: Fisher discriminant
analysis with kernels. In: Hu, Y.H., Larsen, J., Wilson, E., Douglas, S. (eds.) Neural
Networks for Signal Processing IX, pp. 41–48. IEEE, Los Alamitos (1999)
7. Shen, C., Li, H., Brooks, M.J.: A convex programming approach to the trace quo-
tient problem. In: Yagi, Y., Kang, S.B., Kweon, I.S., Zha, H. (eds.) ACCV 2007,
Part II. LNCS, vol. 4844, pp. 227–235. Springer, Heidelberg (2007)
8. Overmars, M.H., van Leeuwen, J.: Maintenance of configurations in the plane.
Journal of Computer and System Sciences 23(2), 166–204 (1981)
9. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few
training examples: an incremental bayesian approach tested on 101 object cate-
gories. In: Conference on Computer Vision and Pattern Recognition Workshop,
vol. 12, pp. 178–178 (2004)
10. Mikolajczyk, K., Schmid, C.: Scale & affine invariant interest point detectors. In-
ternational Journal of Computer Vision 60(1), 63–86 (2004)
11. Lowe, D.G.: Object recognition from local scale-invariant features. In: Proceedings
of the Seventh IEEE International Conference on Computer Vision, vol. 2, pp.
1150–1157 (1999)
itzky, F., Kadir, T., Gool, L.V.: A comparison of affine region detectors. Interna-
tional Journal of Computer Vision 65(1-2), 43–72 (2005)
13. Mikolajczyk, K., Schmid, C.: A performance evaluation of local descriptors. IEEE
Transactions on Pattern Analysis and Machine Intelligence 27(10), 1615–1630
(2005)
A Dynamic Conditional Random Field Model
for Joint Labeling of Object and Scene Classes
Christian Wojek and Bernt Schiele
Computer Science Department

TU Darmstadt
{wojek,schiele}@cs.tu-darmstadt.de
Abstract. Object detection and pixel-wise scene labeling have both

been active research areas in recent years and impressive results have
been reported for both tasks separately. The integration of these differ-
ent types of approaches should boost performance for both tasks as ob-
ject detection can profit from powerful scene labeling and also pixel-wise
scene labeling can profit from powerful object detection. Consequently,
first approaches have been proposed that aim to integrate both object
detection and scene labeling in one framework. This paper proposes a
novel approach based on conditional random field (CRF) models that ex-
tends existing work by 1) formulating the integration as a joint labeling
problem of object and scene classes and 2) by systematically integrating
dynamic information for the object detection task as well as for the scene
labeling task. As a result, the approach is applicable to highly dynamic
scenes including both fast camera and object movements. Experiments
show the applicability of the novel approach to challenging real-world
video sequences and systematically analyze the contribution of different
system components to the overall performance.
1 Introduction
Today, object class detection methods are capable of achieving impressive re-
sults on challenging datasets (e.g. PASCAL challenges [1]). Often these methods
combine powerful feature vectors such as SIFT or HOG with the power of dis-
criminant classifiers such as SVMs and AdaBoost. At the same time several
authors have argued that global scene context [2,3] is a valuable cue for ob-
ject detection and therefore should be used to support object detection. This
context-related work however has nearly exclusively dealt with static scenes. As
this paper specifically deals with highly dynamic scenes we will also model object
motion as an additional and important cue for detection.
Pixel-wise scene labeling has also been an active field of research recently. A
common approach is to use Markov or conditional random field (CRF) models to
improve performance by modeling neighborhood dependencies. Several authors
have introduced the implicit notion of objects into CRF-models [4,5,6,7]. The
interactions between object nodes and scene labels however are often limited to
uni-directional information flow and therefore these models have not yet shown

734 C. Wojek and B. Schiele
the full potential of simultaneously reasoning about objects and scene. By for-
mulating the problem as a joint labeling problem for object and scene classes,
this paper introduces a more general notion of object-scene interaction enabling
bidirectional information flow. Furthermore, as we are interested in dynamic
scenes, we make use of the notion of dynamic CRFs [8], which we extend to deal
with both moving camera and moving objects.
Therefore we propose a novel approach to jointly label objects and scene
classes in highly dynamic scenes for which we introduce a new real-world dataset
with pixel-wise annotations. Highly dynamic scenes are not only a scientific chal-
lenge but also an important problem, e.g. for applications such as autonomous
driving or video indexing where both the camera and the objects are moving
independently. Formulating the problem as a joint labeling problem allows 1) to
model the dynamics of the scene and the objects separately which is of particular
importance for the scenario of independently moving objects and camera, and 2)
to enable bi-directional information flow between object and scene class labels.
The remainder of this paper is structured as follows. Section 2 reviews related
work from the area of scene labeling and scene analysis in conjunction with
object detection. Section 3 introduces our approach and discusses how object
detection and scene labeling can be integrated as a joint labeling problem in
a dynamic CRF formulation. Section 4 introduces the employed features, gives
details on the experiments and shows experimental results. Finally, section 5
draws conclusions.
2 Related Work
In recent years, conditional random fields (CRFs) [9] have become a popular
framework for image labeling and scene understanding. However, to the best
of our knowledge, there is no work which explicitly models object entities in
dynamic scenes. Here, we propose to model objects and scenes in a joint label-
ing approach on two different layers with different information granularity and
different labels in a dynamic CRF [8].
Related work can roughly be divided into two parts. First, there is related
work on CRF models for scene understanding, and second there are approaches
aiming to integrate object detection with scene understanding.
In [10] Kumar&Hebert detect man-made structures in natural scenes using a
single-layered CRF. Later they extend this work to handle multiple classes in a
two-layered framework [5]. Kumar&Hebert also investigated object-context in-
teraction and combined a simple boosted object detector for side-view cars with
scene context of road and buildings on a single-scale database of static images.
In particular, they are running inference separately on their two layers and each
detector hypothesis is only modeled in a neighborhood relation with an entire
region on the second layer. On the contrary, we integrate multi-scale objects in
a CRF framework where inference is conducted jointly for objects and context.
Additionally, we propose to model edge potentials in a consistent layout by ex-
ploiting the scale given by a state-of-the-art object detector [11]. Torralba et al.
[7] use boosted classifiers to model unary and interaction potentials in order
A Dynamic CRF Model for Joint Labeling of Object and Scene Classes 735
to jointly label object and scene classes. Both are represented by a dictionary
of patches. However, the authors do not employ an object detector for entire
objects. In our work we found a separate object detector to be essential for im-
proved performance. Also Torralba et al. use separate layers for each object and
scene class and thus inference is costly due to the high graph connectivity, and
furthermore they also work on a single-scale database of static images. We intro-
duce a sparse layer to represent object hypotheses and work on dynamic image
sequences containing objects of multiple scales. Further work on simultaneous
object recognition and scene labeling has been conducted by Shotton et al. [6].
Their confusion matrix shows, that in particular object classes where color and
texture cues do not provide sufficient discriminative power on static images –
such as boat, chair, bird, cow, sheep, dog – achieve poor results. While their
Texton feature can exploit context information even from image pixels with a
larger distance, the mentioned object classes remain problematic due to the un-
known object scale. Furthermore, He et al. [4] present a multi-scale CRF which
contains multiple layers relying on features of different scales. However, they
do not model the explicit notion of objects and their higher level nodes rather
serve as switches to different context and object co-occurrences. Similarly, Ver-
beek&Triggs [12] add information about class co-occurrences by means of a topic
model. Finally, several authors proposed to adopt the CRF framework for ob-
ject recognition as a standalone task [13,14,15] without any reasoning about the
context and only report results on static single-scale image databases.
Dynamic CRFs are exploited by Wang&Ji [16] for the task of image segmenta-
tion with intensity and motion cues in mostly static image sequences. Similarly,
Yin&Collins [17] propose a MRF with temporal neighborhoods for motion seg-
mentation with a moving camera.
The second part of related work deals with scene understanding approaches
from the observation of objects. Leibe et al. [18] employ a stereo camera system
together with a structure-from-motion approach to detect pedestrians and cars in
urban environments. However, they do not explicitly label the background classes
which are still necessary for many applications even if all objects in the scene
are known. Hoiem et al. [3] exploit the detected scales of pedestrians and cars
together with a rough background labeling to infer the camera’s viewpoint which
in turn improves the object detections in a directed Bayesian network. Contrary
to our work, object detections are refined by the background context but not
the other way round. Also, only still images are handled while the presence of
objects is assumed. Similarly, Torralba [2] exploits filter bank responses to obtain
a scene prior for object detection.
3 Conditional Random Field Models

The following section successively introduces our model. It is divided into three
parts: the first reviews single layer CRFs, the second additionally models objects
in a separate layer and the last adds the scene’s and objects’ dynamics.
We denote the input image at time t with xt , the corresponding class labels
at the grid cell level with yt and the object labels with ot .
3.1 Plain CRF: Single Layer CRF Model for Scene-Class Labeling
In general a CRF models the conditional probability of all class labels yt given an
input image xt . Similar to others, we model the set of neighborhood relationships
N1 up to pairwise cliques to keep inference computationally tractable. Thus, we
model

log(PpCRF (yt |xt , N1 , Θ)) = Φ(yit , xt ; ΘΦ )+ Ψ (yit , yjt , xt ; ΘΨ )− log(Z t )
i (i,j)∈N1
(1)
Z t denotes the so called partition function, which is used for normalization.
N1 is the set of all spatial pairwise neighborhoods. We refer to this model as
plain CRF.
Unary Potentials. Our unary potentials model local features for all classes C
including scene as well as object classes. We employ the joint boosting framework
[19] to build a strong classifier H(c, f(xti ); ΘΦ ) = M m=1 hm (c, f(xi ); ΘΦ ). Here,
t
f(xi ) denotes the features extracted from the input image for grid point i. M is
t
the number of boosting rounds and c are the class labels. hm are weak learners
with parameters ΘΦ and are shared among the classes for this approach. In
order to interpret the boosting confidence as a probability we apply a softmax
transform [5]. Thus, the potential becomes:
exp H(k, f(xti ); ΘΦ )
Φ(yit = k, xt ; ΘΦ ) = log (2)
c exp H(c, f(xi ); ΘΦ )
t
Edge Potentials. The edge potentials model the interaction between class
labels at two neighboring sites yit and yjt in a regular lattice. The interaction
strength is modeled by a linear discriminative classifier with parameters ΘΨ =
wT and depends on the difference of the node features dtij := |f(xti ) − f(xtj )|.

1
Ψ (yi , yj , x ; ΘΨ ) =
t t t
w T
δ(yit = k)δ(yjt = l) (3)
dtij
(k,l)∈C
3.2 Object CRF: Two Layer Object CRF for Joint Object and
Scene Labeling
Information that can be extracted from an image patch locally is rather limited
and pairwise edge potentials are too weak to model long range interactions.
Ideally, a complete dense layer of hidden variables would be added to encode
possible locations and scales of objects, but since inference for such a model
is computationally expensive we propose to inject single hidden variables ot =
{ot1 , . . . , otD } (D being the number of detections) as depicted in figure 1(a). To
instantiate those nodes any multi-scale object detector can be employed.
The additional nodes draw object appearance from a strong spatial model
and are connected to the set of all corresponding hidden variables {yt }otn whose
κt
ot ot+1
Λ Λ
t
Δ
Ω yt y t+1 Ω
Φ Φ
xt xt+1
(a) (b)
Fig. 1. (a) Graphical model for the object CRF ; note that different edge colors denote
different potentials; (b) Graphical model for our full dynamic CRF ; observed nodes are
grey, hidden variables are white, for the sake of readability we omit the spatial layout
of yt with the corresponding edge potential Ψ
evidence {xt }otn support the object hypotheses. The new nodes’ labels in this
work are comprised of O = {object, background}; but the extension to multiple
object classes is straight forward. Thus, we introduce two new potentials into
the CRF model given in equation (1) and yield the object CRF :
log(PoCRF (yt , ot |xt , Θ)) = log(PpCRF (yt |xt , N2 , Θ)) + (4)

Ω(on , x ; ΘΩ ) +
t t
Λ(yi , yj , on , x ; ΘΨ )
t t t t
n (i,j,n)∈N3
Note that N2 ⊂ N1 denotes all neighborhoods where no object is present in the

scene, whereas N3 are all inter-layer neighborhoods with hypothesized object
locations. Ω is the new unary object potential, whereas Λ is the inter-layer edge
potential.
Unary Object Potentials. To define object potentials we use a state-of-the-

art object detector. More specifically, we use a sliding window based multi-scale
approach [11] where a window’s features are defined by g({xt }otn ) and classified
with a linear SVM, the weights being v and b being the hyperplane’s bias. To
get a probabilistic interpretation for the classification margin, we adopt Platt’s
method [20] and fit a sigmoid with parameters s1 and s2 using cross validation.
1
Ω(otn , xt ; ΘΩ ) = log (5)
1 + exp(s1 · (vT · g({xt }otn ) + b) + s2 )
Consequently, the parameters are determined as ΘΩ = {v, b, s1 , s2 }.
Inter-Layer Edge Potentials. For the inter-layer edge potentials we model

the neighborhood relations in cliques consisting of two underlying first layer
nodes yit , yjt and the object hypothesis node otn . Similar to the pairwise edge
potentials on the lower layer, the node’s interaction strength is modeled by a

linear classifier with weights ΘΛ = u.

1
Λ(yit , yjt , otn , xt ; ΘΛ ) = T
u δ(yit = k)δ(yjt = l)δ(otn = m) (6)
dtij
(k,l)∈C;m∈O
It is important to note, that the inter-layer interactions are anisotropic and scale-
dependent. We exploit the scale given by the object detector to train different
weights for different scales and thus can achieve real multi-scale modeling in the
CRF framework. Furthermore, we use different sets of weights for different parts
of the detected object enforcing an object and context consistent layout [15].
3.3 Dynamic CRF: Dynamic Two Layer CRF for Object and Scene
Class Labeling
While the additional information from an object detector already improves the
classification accuracy, temporal information is a further important cue. We
propose two temporal extensions to the framework introduced so far. For highly
dynamic scenes – such as the image sequences taken by a driving car, which we
will use as an example application to our model, it is important to note that
objects and the remaining scene have different dynamics and thus should be
modeled differently. For objects we estimate their motion and track them with
a temporal filter in 3D space. The dynamics for the remaining scene is mainly
caused by the camera motion in our example scenario. Therefore, we use an
estimate of the camera’s ego motion to propagate the inferred scene labels at
time t as a prior to time step t + 1.
Since both – object and scene dynamics – transfer information forward to fu-
ture time steps, we employ directed links in the corresponding graphical model
as depicted in figure 1(b). It would have also been possible to introduce undi-
rected links, but those are computationally more demanding. Moreover, those
might not be desirable from an application point of view, due to the backward
flow of information in time when online processing is required.
Object Dynamics Model. In order to model the object dynamics we employ

multiple extended Kalman filters [21] – one for each object. For the dynamic
scenes dataset which we will use for the experimental section the camera cali-
bration is known and the sequences are recorded from a driving car. Additionally,
we assume the objects to reside on the ground plane. Consequently, Kalman fil-
ters are able to model the object position in 3D coordinates. Additionally, the
state vector contains the objects’ width and speed on the ground plane as well
as the camera’s tilt and all state variables’ first derivative with respect to time.
For the motion model we employ linear motion dynamics with the acceleration
being modeled as system noise which proved sufficient for the image sequences
used below. The tracks’ confidences are given by the last associated detection’s
score. Hence, we obtain the following integrated model:
log(PtCRF (yt , ot |xt , Θ)) = log(PpCRF (yt |xt , N2 , Θ)) + (7)

κt (otn , ot−1 , xt ; Θκ ) + Λ(yit , yjt , otn , xt ; ΘΛ )
n (i,j,n)∈N3
where κt models the probability of an object hypothesis otn at time t given

the history of input images. It replaces the previously introduced potentials
for objects Ω. The parameter vector consists of the detector’s parameters and
additionally of the Kalman filter’s dynamics {A, W } and measurement model
{Ht , Vt } and thus Θκ = ΘΩ ∪ {A, W, Ht , Vt }.
Scene Dynamic Model. In the spirit of recursive Bayesian state estimation

under the Markovian assumption, the posterior distribution of yt−1 is used as a
prior to time step t. However, for dynamic scenes the image content needs to be
transformed to associate the grid points with the right posterior distributions.
In this work we estimate the projection Q from yt to yt+1 given the camera’s
translation and calibration (ΘΔt ). Thus, we obtain an additional unary potential
for yt .
Δt (yit , yt−1 ; ΘΔt ) = log(PtCRF (yQ−1 (i) |Θ))
t−1
(8)
The complete dynamic CRF model including both object and scene dynamics
as depicted in figure 1(b) then is
log(PdCRF (yt , ot , xt |yt−1 , ot−1 , Θ)) = log(PtCRF (yt , ot |xt , Θ)) +

Δt (yit , yt−1 ; ΘΔt ) (9)
i
3.4 Inference and Parameter Estimation

For inference in the undirected graphical models we employ sum-product loopy
belief propagation with a parallel message update schedule. For parameter esti-
mation we take a piecewise learning approach [22] by assuming the parameters
of unary potentials to be conditionally independent of the edge potentials’ pa-
rameters. While this no longer guarantees to find the optimal parameter setting
for Θ, we can learn the model much faster as discussed by [22].
Thus, prior to learning the edge potential models we train parameters ΘΦ ,
ΘΩ for the unary potentials. The parameter set Θκ for the Kalman filter is set
to reasonable values by hand.
Finally, the edge potentials’ parameter sets ΘΨ and ΘΛ are learned jointly in
a maximum likelihood setting with stochastic meta descent [23]. As proposed by
Vishwanathan et al. we assume a Gaussian prior with meta parameter σ on the
linear weights to avoid overfitting.
4 Experiments
To evaluate our model’s performance we conducted several experiments on two
datasets. First, we describe our features which are used for texture and location
based classification of scene labels on the scene label CRF layer. Then we in-
troduce features employed for object detection on the object label CRF layer.
Next, we briefly discuss the results obtained on the Sowerby database and fi-
nally we present results on image sequences on a new dynamic scenes dataset,
which consist of car traffic image sequences recorded from a driving vehicle under
challenging real-world conditions.
4.1 Features for Scene Labeling
Texture and Location Features. For the unary potential Φ at the lower
level as well as for the edge potentials Ψ and inter-layer potentials Λ we employ
texture and location features. The texture features are computed from the 16 first
coefficients of the Walsh-Hadamard transform. This transformation is a discrete
approximation of the cosine transform and can be computed efficiently [24,25] –
even in real-time (e.g. on modern graphics hardware). The features are extracted
at multiple scales from all channels of the input image in CIE Lab color space.
As a preprocessing step, a and b channels are normalized by means of a gray
world assumption to cope with varying color appearance. The L channel is mean-
variance normalized to fit a Gaussian distribution with a fixed mean to cope with
global lighting variations. We also found that normalizing the transformation’s
coefficients according to Varma&Zisserman [26] is beneficial. They propose to
L1 -normalize each filter response first and then locally normalize the responses
at each image pixel. Finally, we take the mean and variance of the normalized
responses as feature for each node in the regular CRF lattice. Additionally, we
use the grid point’s coordinates within the image as a location cue. Therefore,
we concatenate the pixel coordinates to the feature vector.
HOG. In the experiments described below we employ a HOG (Histogram of Ori-

ented Gradients) detector [11] to generate object hypotheses. HOG is a sliding
window approach where features are computed on a dense grid. First, histograms
of gradient orientation are computed in cells performing interpolation with re-
spect to the gradient’s location and with respect to the magnitude. Next, sets of
neighboring cells are grouped into overlapping blocks, which are normalized to
achieve invariance to different illumination conditions. Our front and rear view
car detector has a window size of 20 × 20 pixels. It is trained on a separate
dataset of front and rear car views containing 1492 positive instances from the
LabelMe database [27] and 178 negative images.
4.2 Results
Sowerby Dataset. The Sowerby dataset is a widely used benchmark for CRFs,
which contains 7 outdoor rural landscape classes. The dataset comprises 104
images at a resolution of 96×64 pixels. Following the protocol of [5] we randomly
selected 60 images for training and 44 images for testing. Some example images
with inferred labels are shown in figure 2. However, this dataset does neither
contain image sequences nor cars that can be detected with an object detector
Table 1. Comparison to previously reported results on the Sowerby dataset
Pixel-wise accuracy
Unary classification plain CRF model
He et al. [4] 82.4% 89.5%
Kumar&Hebert [5] 85.4% 89.3%
Shotton et al. [6] 85.6% 88.6%
This paper 84.5% 91.1%
and thus we can only compare our plain CRF model (equation 1) with previous
work on this set.
The experiments show that our features and CRF parameter estimation is
competitive to other state-of-the-art methods. Table 1 gives an overview of pre-
viously published results and how those compare to our model (see figure 3).
While the more sophisticated Textons features [6] do better for unary classifica-
tion, our CRF model can outperform those since our edge potentials are learned
from training data. For this dataset we use a grid with one node for each input
pixel, while the Gaussian prior σ was set to 1.25. The Walsh-Hadamard trans-
form was run on the input images at the aperture size of 2, 4, 8 and 16 pixels.
Moreover, we used a global set of weights for the isotropic linear classifiers of
the edge potentials, but distinguish between north-south neighborhood relations
and east-west neighborhood relations.
Unary plain CRF Unary plain CRF

Input image Input image
classification result classification result
Sky Street object Road surface Building

Vegetation Car Road marking
Fig. 2. Sowerby dataset example results
Dynamic Scenes Dataset. To evaluate our object and dynamic CRF we set
up a new dynamic scenes dataset with image sequences consisting of overall 1936
images1 . The images are taken from a camera inside a driving car and mainly
show rural roads with high dynamics of driving vehicles at an image resolution
of 752 × 480 pixels. Cars appear at all scales from as small as 15 pixels up to 200
pixels. The database consists of 176 sequences with 11 successive images each.
It is split into equal size training and test sets of 968 images.
1
The dataset is available at http://www.mis.informatik.tu-darmstadt.de.
To evaluate pixel level labeling accuracy the last frame of each sequence is
labeled pixel-wise, while the remainder only contains bounding box annotations
for the frontal and rear view car object class. Overall, the dataset contains the
eight labels void, sky, road, lane marking, building, trees & bushes, grass and
car. Figure 3 shows some sample scenes. For the following experiments we used
8 × 8 pixels for each CRF grid node and texture features were extracted at the
aperture sizes of 8, 16 and 32 pixels.
We start with an evaluation of the unary classifier performance on the scene
class layer. Table 2 lists the pixel-wise classification accuracy for different varia-
tions of the feature. As expected location is a valuable cue, since there is a huge
variation in appearance due to different lighting conditions. Those range from
bright and sunny illumination with cast shadows to overcast. Additionally, mo-
tion blur and weak contrast complicate the pure appearance-based classification.
Further, we observe that normalization [26] as well as multi-scale features are
helpful to improve the classification results.
Table 2. Evaluation of texture location features based on overall pixel-wise accuracy;

Multi-scale includes feature scales of 8, 16 and 32 pixels, Single-scale is a feature scale
of 8 pixels; note that these number do not include the CRF model – adding the plain
CRF to the best configuration yields an overall accuracy of 88.3%
Normalization
on off
multi-scale single-scale multi-scale single-scale
on 82.2% 81.1% 79.7% 79.7%
Location
off 69.1% 64.1% 62.3% 62.3%
Next, we analyze the performance of the different proposed CRF models. On the
one hand we report the overall pixel-wise accuracy. On the other hand the pixel-
wise labeling performance on the car object class is of particular interest. Overall,
car pixels cover 1.3% of the overall observed pixels. Yet, those are an important
fraction for many applications and thus we also report those for our evaluation.
For the experiments we used anisotropic linear edge potential classifiers with
16 parameter sets, arranged in four rows and four columns. Moreover, we dis-
tinguish between north-south and east-west neighborhoods. For the inter-layer
edge potentials we trained different weight sets depending on detection scale
(discretized in 6 bins) and depending on the neighborhood location with respect
to the object’s center.
Table 3 shows recall and precision for the proposed models. Firstly, the em-
ployed detector has an equal error rate of 78.8% when the car detections are eval-
uated in terms of precision and recall. When evaluated on a pixel-wise basis the
performance corresponds to 60.2% recall. The missing 39.8% are mostly due to
the challenging dataset. It contains cars with weak contrast, cars at small scales
and partially visible cars leaving the field of view. Precision for the detector eval-
uated on pixels is 37.7%. Wrongly classified pixels are mainly around the objects
and on structured background on which the detector obtains false detections.
Table 3. Pixel-wise recall and precision for the pixels labeled as Car and overall
accuracy on all classes
With object Including object

No objects
layer dynamics
Recall Precision Acc. Recall Precision Acc. Recall Precision Acc.
CRF 50.1% 57.7% 88.3% 62.9% 52.3% 88.6% 70.4% 57.8% 88.7%
dyn. CRF 25.5% 44.8% 86.5% 75.7% 50.8% 87.1% 78.0% 51.0% 88.1%
Let us now turn to the performance of the different CRF models. Without
higher level information from an object detector plain CRFs in combination
with texture-location features achieve a recall of 50.1% with a precision of 57.7%.
The recognition of cars in this setup is problematic since CRFs optimize a global
energy function, while the car class only constitutes a minor fraction of the data.
Thus, the result is mainly dominated by classes which occupy the largest regions
such as sky, road and trees.
With higher level object information (object CRF ) recall can be improved up
to 62.9% with slightly lower precision resulting from the detector’s false positive
detections. However, when objects are additionally tracked with a Kalman filter,
we achieve a recall of 70.4% with a precision of 57.8%. This proves that the
object labeling for the car object class leverages from the object detector and
additionally from the dynamic modeling by a Kalman filter.
Additionally, we observe an improvement of the overall labeling accuracy.
While plain CRFs obtain an accuracy of 88.3%, the object CRF achieves 88.6%
while also including object dynamics further improves the overall labeling accu-
racy to 88.7%. The relative number of 0.4% might appear low, but considering
that the database overall only has 1.3% of car pixels, this is worth noting. Thus,
we conclude that not only the labeling on the car class is improved but also the
overall scene labeling quality.
When the scene dynamics are modeled additionally and posteriors are prop-
agated over time (dynamic CRF ), we again observe an improvement of the
achieved recall from 25.5% to 75.7% with the additional object nodes. And also
the objects’ dynamic model can further improve the recall to 78.0% correctly
labeled pixels. Thus, again we can conclude that the CRF model exploits both
the information given by the object detector as well as the additional object
dynamic to improve the labeling quality.
Finally, when the overall accuracy is analyzed while the scene dynamic is
modeled we observe a minor drop compared to the static modeling. However, we
again consistently observe that the object information and their dynamics allow
to improve from 86.5% without object information to 87.1% with object CRFs
and to 88.1% with the full model.
The consistently slightly worse precision and overall accuracy for the dynamic
scene models need to be explained. Non-car pixels wrongly labeled as car are
mainly located at the object boundary, which are mainly due to artifacts of the
Ground Unary Dynamic

Input image plain CRF Object CRF
truth classification CRF
(a)
(b)
(c)
(d)
(e)
(f)
Void Sky Road Lane marking

Trees & bushes Grass Building Car
(a) (b) (c)
(d) (e) (f)
Fig. 3. Dynamic scenes dataset example result scene labels and corresponding detec-
tions in left-right order (best viewed in color); note that detections can be overruled
by the texture location potentials and vice versa
scene label forward propagation. Those are introduced by the inaccuracies of the
speedometer and due to the inaccuracies of the projection estimation.
A confusion matrix for all classes of the dynamic scenes database can be
found in table 4. Figure 3 shows sample detections and scene labelings for the
different CRF models to illustrate the impact of the different models and their
improvements. In example (d) for instance the car which is leaving the field of
view is mostly smoothed out by a plain CRF and object CRF, while the dynamic
CRF is able to classify almost the entire area correctly. Additionally, the smaller
cars which get smoothed out by a plain CRF are classified correctly by the object
and dynamic CRF. Also note that false object detections as in example (c) do
not result in a wrong labeling of the scene.
Table 4. Confusion matrix in percent for the dynamic scenes dataset; entries are
row-normalized
Building
marking
Inferred
Trees &
bushes
Grass
Road
Lane
Void
Car
Sky
True class Fraction
Sky 10.4% 91.0 0.0 0.0 7.7 0.5 0.4 0.3 0.1
Road 42.1% 0.0 95.7 1.0 0.3 1.1 0.1 0.5 1.3
Lane marking 1.9% 0.0 36.3 56.4 0.8 2.9 0.2 1.8 1.6
Trees & bushes 29.2% 1.5 0.2 0.0 91.5 5.0 0.2 1.1 0.4
Grass 12.1% 0.4 5.7 0.5 13.4 75.3 0.3 3.5 0.9
Building 0.3% 1.6 0.2 0.1 37.8 4.4 48.4 6.3 1.2
Void 2.7% 6.4 15.9 4.1 27.7 29.1 1.4 10.6 4.8
Car 1.3% 0.3 3.9 0.2 8.2 4.9 2.1 2.4 78.0
5 Conclusions
In this work we have presented a unifying model for joint scene and object class
labeling. While CRFs greatly improve unary pixel-wise classification of scenes
they tend to smooth out smaller regions and objects such as cars in landscape
scenes. This is particularly true when objects only comprise a minor part of the
amount of overall pixels. We showed that adding higher level information from
a state-of-the-art HOG object detector ameliorates this shortcoming. Further
improvement – especially when objects are only partially visible – is achieved
when object dynamics are properly modeled and when scene labeling information
is propagated over time. The improvement obtained is bidirectional, on the one
hand the labeling of object classes is improved, but on the other hand also the
remaining scene classes benefit from the additional source of information.
For future work we would like to investigate how relations between different
objects such as partial occlusion can be modeled when multiple object classes
are detected. Additionally, we seek to improve the ego-motion estimation of the
camera to further improve the performance. This will also allow us to employ
motion features in the future. Finally, we assume that the integration of different
sensors such as radar allow for a further improvement of the results.
Acknowledgements. This work has been funded, in part, by Continental Teves

AG. Further we thank Joris Mooij for publically releasing libDAI and BAE
Systems for the Sowerby Dataset.
References
1. Everingham, M., Zisserman, A., Williams, C., van Gool, L.: The pascal visual
object classes challenge results. Technical report, PASCAL Network (2006)
2. Torralba, A.: Contextual priming for object detection. IJCV, 169–191 (2003)
3. Hoiem, D., Efros, A.A., Hebert, M.: Putting objects in perspective. In: CVPR
(2006)
4. He, X., Zemel, R.S., Carreira-Perpiñán, M.Á.: Multiscale conditional random fields
for image labeling. In: CVPR (2004)
5. Kumar, S., Hebert, M.: A hierarchical field framework for unified context-based
classification. In: ICCV (2005)
6. Shotton, J., Winn, J., Rother, C., Criminisi, A.: Textonboost: Joint appearance,
shape and context modeling for multi-class object recognition and segmentation. In:
Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3951. Springer,
Heidelberg (2006)
7. Torralba, A., Murphy, K.P., Freeman, W.T.: Contextual models for object detec-
tion using boosted random fields. In: NIPS (2004)
8. McCallum, A., Rohanimanesh, K., Sutton, C.: Dynamic conditional random fields
for jointly labeling multiple sequences. In: NIPS Workshop on Syntax, Semantics
(2003)
9. Lafferty, J.D., McCallum, A., Pereira, F.C.N.: Conditional random fields: Proba-
bilistic models for segmenting and labeling sequence data. In: ICML (2001)
10. Kumar, S., Hebert, M.: Discriminative random fields: A discriminative framework
for contextual interaction in classification. In: ICCV (2003)
CVPR (2005)
12. Verbeek, J., Triggs, B.: Region classification with markov field aspect models. In:
CVPR (2007)
13. Quattoni, A., Collins, M., Darrell, T.: Conditional random fields for object recog-
nition. In: NIPS (2004)
14. Kapoor, A., Winn, J.: Located hidden random fields: Learning discriminative parts
for object detection. In: Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006.
LNCS, vol. 3954. Springer, Heidelberg (2006)
15. Winn, J., Shotton, J.: The layout consistent random field for recognizing and seg-
menting partially occluded objects. In: CVPR (2006)
16. Wang, Y., Ji, Q.: A dynamic conditional random field model for object segmenta-
tion in image sequences. In: CVPR (2005)
17. Yin, Z., Collins, R.: Belief propagation in a 3D spatio-temporal MRF for moving
object detection. In: CVPR (2007)
18. Leibe, B., Cornelis, N., Cornelis, K., Van Gool, L.: Dynamic 3D scene analysis
from a moving vehicle. In: CVPR (2007)
19. Torralba, A., Murphy, K.P., Freeman, W.T.: Sharing features: Efficient boosting
procedures for multiclass object detection. In: CVPR (2004)
20. Platt, J.: Probabilistic outputs for support vector machines and comparison to
regularized likelihood methods. In: Smola, A.J., Bartlett, P., Schoelkopf, B., Schu-
urmans, D. (eds.) Advances in Large Margin Classifiers, pp. 61–74 (2000)
21. Kalman, R.E.: A new approach to linear filtering and prediction problems. Trans-
actions of the ASME–Journal of Basic Engineering 82, 35–45 (1960)
22. Sutton, C., McCallum, A.: Piecewise training for undirected models. In: 21th An-
nual Conference on Uncertainty in Artificial Intelligence (UAI 2005) (2005)
23. Vishwanathan, S.V.N., Schraudolph, N.N., Schmidt, M.W., Murphy, K.P.: Accel-
erated training of conditional random fields with stochastic gradient methods. In:
ICML (2006)
24. Hel-Or, Y., Hel-Or, H.: Real-time pattern matching using projection kernels.
PAMI 27, 1430–1445 (2005)
25. Alon, Y., Ferencz, A., Shashua, A.: Off-road path following using region classifica-
tion and geometric projection constraints. In: CVPR (2006)
26. Varma, M., Zisserman, A.: Classifying images of materials: Achieving viewpoint
and illumination independence. In: Tistarelli, M., Bigun, J., Jain, A.K. (eds.)
ECCV 2002. LNCS, vol. 2359. Springer, Heidelberg (2002)
27. Russell, B.C., Torralba, A., Murphy, K.P., Freeman, W.T.: LabelMe: A database
and web-based tool for image annotation. IJCV 77, 157–173 (2008)
Local Regularization for Multiclass Classification
Facing Significant Intraclass Variations
Lior Wolf and Yoni Donner
The School of Computer Science

Tel Aviv Univerisy
Tel Aviv, Israel
Abstract. We propose a new local learning scheme that is based on the

principle of decisiveness: the learned classifier is expected to exhibit large
variability in the direction of the test example. We show how this prin-
ciple leads to optimization functions in which the regularization term is
modified, rather than the empirical loss term as in most local learning
schemes. We combine this local learning method with a Canonical Corre-
lation Analysis based classification method, which is shown to be similar
to multiclass LDA. Finally, we show that the classification function can
be computed efficiently by reusing the results of previous computations.
In a variety of experiments on new and existing data sets, we demon-
strate the effectiveness of the CCA based classification method compared
to SVM and Nearest Neighbor classifiers, and show that the newly pro-
posed local learning method improves it even further, and outperforms
conventional local learning schemes.
1 Introduction
Object recognition systems, viewed as learning systems, face three major chal-
lenges: First, they are often required to discern between many objects; second,
images taken under uncontrolled settings display large intraclass variation; and
third, the number of training images provided is often small.
Previous attempts to overcome these challenges use prior generic knowledge
on variations within objects classes [1], employ large amounts of unlabeled data
(e.g., [2]), or reuse previously learned visual features [3]. Here, we propose a more
generic solution, that does not assume nor benefit from the existence of prior
learning stages or of an additional set of training images.
To deal with the challenge of multiple classes, we propose a Canonical Cor-
relation Analysis (CCA) based classifier, which is a regularized version of a
recently proposed method [4], and is highly related to Fisher Discriminant Anal-
ysis (LDA/FDA). We treat the other two challenges as one since large intraclass
variations and limited training data both result in a training set that does not
capture well the distribution of the input space. To overcome this, we propose a
new local learning scheme which is based on the principle of decisiveness.
In local learning schemes, some of the training is deferred to the prediction
phase, and a new classifier is trained for each new (test) example. Such schemes

Local Regularization for Multiclass Classification 749
have been introduced by [5] and were recently advanced and shown to be ef-
fective for modern object recognition applications [6] (see references therein for
additional references to local learning methods). One key difference between our
method and the previous contribution in the field is that we do not select or di-
rectly weigh the training examples by their proximity to the test point. Instead,
we modify the objective function of the learning algorithm to reward components
in the resulting classifier that are parallel to the test example. Thus, we encour-
age the classification function (before thresholding takes place) to be separated
from zero.
Runtime is a major concern for local learning schemes, since a new clas-
sifier needs to be trained or adjusted for every new test example. We show
how the proposed classifier can be efficiently computed by several rank-one up-
dates to precomputed eigenvectors and eigenvalues of constant matrices, with
the resulting time complexity being significantly lower than that of a full eigen-
decomposition. We conclude by showing the proposed methods to be effective
on four varied datasets which exhibit large intraclass variations.
2 Multiclass Classification Via CCA

We examine the multiclass classification problem with k classes, where the goal
nn training samples (xi , yi ), with xi ∈ R and
m
is to construct a classifier given
yi ∈ {1, 2, . . . , k}. We assume i=1 xi = 0 (otherwise we center the data). Our
approach is to find a transformation T : Rm → Rl and class vectors vj ∈ Rl such
that the transformed inputs T (xi ) would be close to the class vector vyi corre-
sponding to their class. Limiting the discussion at first to linear transformations,
we represent T by a m × l matrix A such that T (x) = A x. The formulation of
the learning problem is therefore:

n
min A xi − vyi 2 (1)
A,{vj }k
j=1 i=1
Define V to be the k × l matrix with vj as its j’th row, so vj = V ej . Also

define zi = eyi where ej is the j’th column of the identity k × k matrix Ik . Using
these definitions, vyi = V zi and Equation 1 becomes:

n
min A xi − V zi 2 (2)
A,V
i=1
This expression can be further simplified by defining the matrices X ∈ Rm×n ,

Z ∈ Rk×n : X = (x1 , x2 , . . . , xn ), Z = (z1 , z2 , . . . , zn ). Equation 2 then becomes:
min tr(A XX A) + tr(V ZZ V ) − 2 tr(A XZ V ) (3)

A,V
This expression is not invariant to arbitrary scaling of A and Z. Furthermore,

we require the l components of the transformed vectors A xi and V zi to be
750 L. Wolf and Y. Donner
pairwise uncorrelated since there is nothing to be gained by correlations between

them. Therefore, we add the constraints A XX A = V ZZ V = Il , leading
to the final problem formulation:
max tr(A XZ V )
A,V
subject to A XX A = V ZZ V = I (4)
This problem is solved through Canonical Correlation Analysis (CCA) [7]. A
simple solution involves writing the corresponding Lagrangian and setting the
partial derivatives to zero, yielding the following generalized eigenproblem:

0 XZ ai XX 0 ai
= λ (5)
ZX 0 vi i
0 ZZ vi
where λi , i = 1..l are the leading generalized eigenvalues, ai are the columns of
A, and vi are, as defined above, the columns of V . To classifying a new sample
x, it is first transformed to A x, and then compared to the k class vectors, i.e.,
the predicted class is given by arg min1≤j≤k ||A x − vj ||.
This classification scheme is readily extendable to non-linear functions that
satisfy Mercer’s conditions by using Kernel CCA [8,9]. Kernel CCA is also equiv-
alent to solving a generalized eigenproblem of the form of Equation 5, so although
we refer directly to linear CCA throughout this paper, our conclusions are equally
valid for Kernel CCA.
In Kernel CCA, or in the linear case when m > n, and in many other com-
mon scenarios, the problem is ill-conditioned and regularization techniques are
required [10]. For linear regression, ridge regularization is often used, as is its
equivalent in CCA and Kernel CCA [8]. This involves replacing XX and ZZ
in Equation 5 with XX + ηX I and ZZ + ηZ I, where ηX and ηZ are regular-
ization parameters. In the CCA case presented here, for multiclass classification,
since the number of training examples n is not smaller than the number of
classes k, regularization need not be used for Z and we set ηZ = 0. Also, since
the X regularization is relative to the scale of the matrix XX , we scale the
regularization parameter ηX as a fraction of the largest eigenvalue of XX .
The multiclass classification scheme via CCA presented here is equivalent to
Fisher Discriminant Analysis (LDA). We provide a brief proof of this equivalence.
A previous lemma was proven by Yamada et al [4] for the unregularized case.
Lemma 1. The multiclass CCA classification method learns the same linear
transformation as multiclass LDA.
Proof. The generalized eigenvalue problem in Equation 5, with added ridge reg-
ularization, can be represented by the following two coupled equations:
(XX + ηIm )−1 XZ v = λa (6)
(ZZ )−1 ZX a = λv (7)
Any solution (a, v, λ) to the above system satisfies:
(XX + ηIm )−1 XZ (ZZ )−1 ZX a = (XX + ηIm )−1 XZ λv = λ2 a (8)
(ZZ )−1 ZX (XX + ηIm )−1 XZ v = (ZZ )−1 ZX λa = λ2 v (9)
Thus the columns of the matrix A are the eigenvectors corresponding to the
largest eigenvalues
of (XX + ηIm )−1 XZ (ZZ )−1 ZX . Examine the prod-
uct ZZ = i=n eyi e

yi . It is a k × k diagonal matrix with the number of train-
ing samples in each class (denoted Ni ) along its diagonal. Therefore, (ZZ )−1 =
n
diag( N11 , N12 , . . . , N1k ). Now examine XZ : (XZ )i,j = Xi,s Zj,s = Xi,s .
s=1 s:ys =j
Hence, the j’th column is the sum of all training samples of the class j. Denote by
X̄j the mean of the training samples belonging to the class j, then the j’th column
of XZ is Nj X̄j . It follows that

k
Nj2
k
XZ (ZZ )−1 ZX = X̄j X̄j = Nj X̄j X̄j = SB (10)
j=1
Nj j=1
Where SB is the between-class scatter matrix defined in LDA [11]. Let ST =

XX be the total scatter matrix ST . ST = SW + SB (where SW is LDA’s
within-class scatter matrix), and using ST in LDA is equivalent to using SW .
Hence, the multiclass CCA formulation is equivalent to the eigen-decomposition
of (SW + ηI)−1 SB , which is the formulation of regularized multiclass LDA.
Our analysis below uses the CCA formulation; the LDA case is equivalent, with
some minor modifications to the way the classification is done after the linear
transformation is applied.
3 Local Learning Via Regularization
The above formulation of the multiclass classification problem is independent of

the test vector to be classified x. It may be the case that the learned classifier is
“indifferent” to x, transforming it to a vector A x which has a low norm. Note
that by the constraint V ZZ V = I, the norm of the class vectors vj is Nj−0.5
which is roughly constant for balanced data sets. This possible mismatch between
the norm of the transformed example and the class vectors may significantly
decrease the ability to accurately classify x. Furthermore, when the norm of
A x is small, it is more sensitive to additive noise.
In local learning, the classifier may be different for each test sample and
depends on it. In this work, we discourage classifiers that are indifferent to x,
and have low A x2 . Hence, to discourage indifference (increase decisiveness),
we add a new term to the CCA problem:
max tr(A XZ V ) + ᾱ tr(A xx A)

A,V
subject to A XX A = V ZZ V = I (11)
tr(A xx A) = A x2 , and the added term reflects the principle of decisive-
ness. ᾱ is a parameter corresponding to the trade-off between the correlation
term and the decisiveness term. Adding ridge regularization as before to the
solution of Equation 11, and setting α = ᾱλ−1 gives the following generalized
eigenproblem:

0 XZ a XX + ηI − αxx 0 a
= λ (12)
ZX 0 v 0 ZZ v
Note that this form if similar to the CCA based multiclass classifier presented
in Section 2 above, except that the ridge regularization matrix ηI is replaced by
the local regularization matrix ηI −αxx . We proceed to analyze the significance
of this form of local regularization. In ridge regression, the influence of all eigen-
vectors is weakened uniformly by adding η to all eigenvalues before computation
of the inverse. This form of regularization encourages smoothness in the learned
transformation. In our version of local regularization, smoothness is still achieved
by the addition of η to all eigenvalues. The smoothing effect is weakened, how-
ever, by α, in the component parallel to x. This can be seen by the representation
xx = Ux λx Ux , for Ux Ux = Ux Ux = I, with λx = diag(x2 , 0, . . . , 0). Now
ηI − αxx = Ux (ηI − αλx )Ux , and the eigenvalues of the regularization ma-
trix are (η − α, η, η, . . . , η). Hence, the component parallel to x is multiplied by
η − α while all others are multiplied by η. Therefore, encouraging decisiveness by
adding the term αA x2 to the maximization goal is a form of regularization
where the component parallel to x is smoothed less than the other components.
4 Efficient Implementation
In this section we analyze the computational complexity of our method, and
propose an efficient update algorithm that allows it to be performed in time
comparable to standard CCA with ridge regularization. Our algorithm avoids
fully retraining the classifier for each testing example by training it once using
standard CCA with uniform ridge regularization, and reusing the results in the
computation of the local classifiers.
Efficient training of a uniformly regularized multiclass CCA classifier.
In the non-local case, training a multiclass CCA classifier consists of solving
Equations 6 and 7, or, equivalently, Equations 8 and 9. Let r = min(m, k), and
note that we assume m ≤ n, since the rank of the data matrix is at most n,
and if m > n we can change basis to a more compact representation. To solve
Equations 8 and 9, it is enough to find the eigenvalues and eigenvectors of a
r × r square matrix. Inverting (XX + ηIm )−1 and (ZZ )−1 and reconstructing
the full classifier (A and V ) given the eigenvalues and eigenvectors of the r × r
matrix above can be done in O(m3 + k 3 ). While this may be a reasonable effort
if done once, it may become prohibitive if done repeatedly for each new test
example. This, however, as we show below, is not necessary.
Representing the local learning problem as a rank-one modification.
We first show the problem to be equivalent to the Singular Value Decomposition
(SVD) of a (non-symmetric) matrix, which is in turn equivalent to the eigen-
decomposition of two symmetric matrices. We then prove that one of these two
matrices can be represented explicitly as a rank-one update to a constant (with

regards to the new test example) matrix whose eigen-decomposition is computed
only once. Finally, we show how to efficiently compute the eigen-decomposition
of the modified matrix, how to derive the full solution using this decomposition
and how to classify the new example in time complexity much lower than that
of a full SVD.
Begin with a change of variables. Let Ā = (XX + ηIm − αxx ) 2 A and
1
V̄ = (ZZ ) 2 V . By the constraints (Equation 11, with added ridge and local
1
regularizations), Ā and V̄ satisfy Ā Ā = A (XX + ηIm − αxx )A = I and

V̄ V̄ = V ZZ V = I. Hence, the new variables are orthonormal and the CCA
problem formulation (Equation 4) with added ridge regularization becomes:
max tr(Ā (XX + ηIm − αxx )− 2 XZ (ZZ )− 2 V̄ )

1 1
Ā,V̄
subject to Ā Ā = V̄ V̄ = I (13)
Define:
M0 = (XX + ηIm )− 2 XZ (ZZ )− 2 = U0 Σ0 R0

1 1
(14)
− 12 − 12
M = (XX + ηIm − αxx ) XZ (ZZ ) = U ΣR (15)
where U ΣR is the Singular Value Decomposition (SVD) of M and similarly

U0 Σ0 R0 for M0 . Then the maximization term of Equation 13 is Ā U ΣR V̄ ,
which under the orthonormality constraints of Equation 13, and since we seek
only l components, is maximized by Ā = U0|l and V̄ = R0|l , which are the l left
and right singular vectors of M corresponding to the l largest singular values.
Since M M = RΣ 2 R , the right singular vectors can be found by the eigen-
decomposition of the symmetric M M . We proceed to show how M M can be
represented explicitly as a rank-one update to M0 M0 . Define JX = (XX +
ηIm )−1 , then JX is symmetric as the inverse of a symmetric matrix, and by the
Sherman-Morrison formula [12],
JX αxx JX
(XX + ηX Im − αxx )−1 = (JX − αxx )−1 = JX +
1 − αx JX x
α
= JX + (JX x) (JX x) = (XX + ηX Im )−1 + βbb (16)
1 − αx JX x
where β = 1−αxα JX x and b = JX x. β and b can both be computed using O(m2 )

operations, since JX is known after being computed once. Now,
M M = (ZZ )− 2 ZX (XX + ηIm − αxx )−1 XZ (ZZ )− 2

1 1

= (ZZ )− 2 ZX (XX + ηIm )−1 + βbb XZ (ZZ )− 2
1 1
= M0 M0 + β(ZZ )− 2 ZX bb XZ (ZZ )− 2
1 1
= M0 M0 + βcc (17)
(18)
where c = (ZZ )− 2 ZX b, and again c is easily computed from b in O(km)

1
operations. Now let w = R0 c

c
(so w = 1) and γ = βc2 to arrive at the
representation
M M = R0 (Σ02 + γww )R0 (19)
It is left to show how to efficiently compute the eigen-decomposition of a rank-
one update to a symmetric matrix, whose eigen-decomposition is known. This
problem has been investigated by Golub [13] and Bunch et al. [14]. We propose
a simple and efficient algorithm that expands on their work. We briefly state
their main results, without proofs, which can be found in the original papers.
The first stage in the algorithm described in Bunch et al. [14] is deflation,
transforming the problem to equivalent (and no larger) problems S + ρzz satis-
fying that all elements of z are nonzero, and all elements of S are distinct. Then,
under the conditions guaranteed by the deflation stage, the new eigenvalues can
be found. The eigenvalues of S+ρzz satisfying that all elements of z are nonzero

s
zi2
and all elements of S are distinct are the roots of f (λ) = 1+ρ di −λ , where s is
i=1
the size of the deflated problem, zi are the elements of z and di are the elements
of the diagonal of S. [14] show an iterative algorithm with a quadratic rate of
convergence, so all eigenvalues can be found using O(s2 ) operations, with a very
small constant as shown in their experiments. Since the deflated problem is no
larger than k, this stage requires O(k 2 ) operations at most. Once the eigenvalues
have been found, the eigenvectors of Σ02 + γww can be computed by
(S − λi I)−1 z
ξi = (20)
(S − λi I)−1 z
using O(k) operations for each eigenvector, and O(k 2 ) in total to arrive at the
representation
M M = R0 R1 Σ1 R1 R0 (21)
Explicit evaluation of Equation 21 to find V̂ requires multiplying k × k, which
should be avoided to keep the complexity O(m2 + k 2 ). The key observation is
that we do not need to find V explicitly but only A x − vi for i = 1, 2, . . . , k,
with vi being the i’th class vector (Equation 1). The distances we seek are:
A x − vi 2 = A x2 + vi 2 − 2vi A x (22)
with vi 2 = Ni (see Section 3). Hence, finding all exact distances can be done
by computation of x AA x−V A x, since vi is the i’th row of V . Transforming
back from V̄ to V gives V = (ZZ )− 2 V̄ , where (ZZ )− 2 needs to be computed
1 1
only once. From Equations 6 and 21,
A x = Σ1−1 V ZX (XX + ηIm − αxx )−1 x

= Σ1−1 R1 R0 (ZZ )− 2 ZX (XX + ηIm )−1 + βbb x
1
(23)
All the matrices in Equation 23 are known after the first O(k 3 + m3 ) compu-
tation and O(k 2 + m2 ) additional operations per test example, as we have shown
above. Hence, A x can be computed by a sequence of matrix-vector multiplica-

tions in time O(k 2 + m2 ), and similarly for
V A x = (ZZ )− 2 R0 R1 A x
1
(24)
Thus, the distances of the transformed test vector x from all class vectors can
be computed in time O(m2 + k 2 ), which is far quicker than O(m3 + k 3 ) which
is required by training the classifier from scratch, using a full SVD. Note that
the transformation of a new vector without local regularization requires O(ml)
operations, and the classification itself O(kl) operations. The difference between
the classification times of a new test vector using local regularization, therefore,
is O(m2 + k 2 ) compared to O ((m + k)l) using uniform regularization.
5 Experiments
We report results on 3 data sets: a new Dog Breed data set, the CalPhotos
Mammals collection [15], and the “Labeled Faces in the Wild” face recognition
data set [16]. These data sets exhibit a large amount of intraclass variation.
The experiments in all cases are similar and consist of multiclass classifica-
tion. We compare the following algorithms: Nearest Neighbor, Linear All-Vs-All
SVM (a.k.a “pairwise”, ”All-Pairs”), Multiclass CCA (the method of Section 2),
and Local Multiclass CCA (Section 3). The choice of using All-Vs-All SVM is
based on its simplicity and relative efficiency. A partial set of experiments ver-
ified that One-Vs-All SVM classifiers perform similarly. It is well established in
the literature that the performance of other multiclass SVM schemes is largely
similar [6,17]. Similar to other work in object recognition we found Gaussian-
kernel SVM to be ineffective, and to perform worse than Linear SVM for every
kernel parameter we tried. Evaluating the performance of non-linear versions of
Multiclass CCA and Local Multiclass CCA is left for future work.
We also compare the conventional local learning scheme [5], which was devel-
oped further in [6]. In this scheme the k nearest neighbors of each test point are
used to train a classifier. In our experiments we have scanned over a large range
possible neighborhood sizes k to verify that this scheme does not outperform
our local learning method regardless of k. Due to the computational demands of
such tests, they were only performed on two out of the four data sets.
Each of the described experiments was repeated 20 times. In each repetition a
new split to training and testing examples was randomized, and the same splits
were used for all algorithms. Note that due to the large intraclass variation,
the standard deviation of the result is typically large. Therefore, we use paired
t-tests to verify that the reported results are statistically significant.
Parameter selection. The regularization parameter of the linear SVM algo-
rithm was selected by a 5-fold cross-validation. Performance, however, is pretty
stable with respect to this parameter. The regularization parameter of Multiclass
CCA and Local Multiclass CCA η was fixed at 0.1 times the leading eigenvalue
of XX , a value which seems to be robust in a large variety of synthetic and real
Dog Breed data set CalPhoto Mammals
Bullmastiff Chow Chow Black Rhinoceros Prairie Dog
Fig. 1. Sample images from the Dog Breed and CalPhoto Mammal data sets
data sets. The local regularization parameter β was set at 0.5η in all experiments,
except for the ones done to evaluate its effect on performance.
Image representation. The visual descriptors of the images in the Dog Breed
and CalPhotos Mammels data sets are computed by the Bag-of-SIFT implemen-
tation of Andrea Vendaldi [18]. This implementation uses hierarchical K-means
[19] for partitioning the descriptor space. Keypoints are selected at random lo-
cations [20]. Note that the dictionary for this representation was recomputed at
each run in order to avoid the use of testing data during training. Using the
default parameters, this representation results in vectors of length 11, 111.
The images in the face data set are represented using the Local Binary Pat-
tern [21] image descriptor, which were adopted to face identification by [22]. An
LBP is created at a particular pixel location by thresholding the 3 × 3 neighbor-
hood surrounding the pixel with the central pixels intensity value, and treating
the subsequent pattern as a binary number. Following [22], we set a radius of 2
and sample at the boundaries of 5 pixel blocks, and bin all patterns for which
there are more than 2 transition from 0 to 1 in just one bin. LBP representations
for a given image are generated by dividing an image into several windows and
creating histograms of the LBPs within each window.
5.1 Results on Individual Data Sets

Dog Breed images. The Dog Breed data set contains images of 34 dog species,
with 4–7 photographs each, a total of 177 images. The images were collected from
the internet, and as can be seen in Figure 1 are quite diverse.
Table 1 compares the classification results for a varying number of train-
ing/testing examples per breed. The results demonstrate that Local Multiclass
CCA performs better than Multiclass CCA, which in turn performs better than
Nearest Neighbor and SVM. Since the images vary significantly, the results ex-
hibit a large variance. Still, all differences in the table are significant (p < 0.01),
except for the difference between Multiclass CCA and SVM in the case of 3
training images per breed.
Table 1. Mean (± standard deviation) recognition rates (in percents) for the Dog
Breed data set. Each column is for a different number of training and testing examples
per breed for the 34 dog breeds.
Algorithm 1 training / 3 test 2 training / 2 test 3 training / 1 test

Nearest Neighbor 11.03 ± 1.71 14.85 ± 3.96 18.68 ± 6.35
All-Pairs Linear SVM 11.03 ± 1.71 17.50 ± 4.37 23.82 ± 6.32
Multiclass CCA 13.43 ± 3.56 19.63 ± 4.99 24.12 ± 6.92
Local Multiclass CCA 15.78 ± 3.63 21.25 ± 4.56 26.18 ± 6.39
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 10 20 30 40 50 60 70 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(a) (b)
√
Fig. 2. Mean performance and standard deviation (normalized by 20) for additional
experiments on the Dog Breed data set. (a) k-nearest neighbors based local learning.
The x axis depicts k, the size of the neighberhood. Top line – the performance of
the Multiclass CCA classifier, Bottom dashed line – the performance of SVM. (b)
Performance for various values of the local regularization parameter. The x axis depicts
the ratio of β and η.
To further understand the nature of the local learning method we performed

two additional more experiments. Figure 2(a) demonstrates that the conventional
local learning scheme, based on k-nearest neighbors does not seem to improve
performance for any values of k. Figure 2(b) demonstrates that the performance
of the Local CCA method is stable with respect to the additional parameter α.
CalPhoto Mammals. The mammal collection of the CalPhoto image reposi-
tory [15] contains thousands of images. After filtering out all images for which
the Latin species name does not appear and species for which there are less than
4 images, 3, 740 images of 256 species remain. For each species, the images vary
considerably, as can be seen in Figure 1.
In each experiment 10, 20 or 40 random species are selected. Each contributes
2 random training images and 2 test ones. Table 2 compares the classification
results. Once again, Local Multiclass CCA outperforms the uniform Multiclass
CCA, followed by SVM and NN. All performance differences in the table are
statistically significant, except for SVM and Multiclass CCA for 40 classes.
Table 2. Mean (± standard deviation) recognition rates (percents) for the Mammals
data set. Each column is for a different number of random classes per experiment. Each
experiment was repeated 20 times.
Algorithm 10 classes 20 classes 40 classes

Nearest Neighbor 25.50 ± 8.57 20.25 ± 7.86 14.13 ± 3.89
All-Pairs Linear SVM 28.75 ± 10.87 25.38 ± 9.22 17.13 ± 4.20
Multiclass CCA 33.00 ± 11.63 28.75 ± 9.78 18.88 ± 4.81
Local Multiclass CCA 36.00 ± 11.19 31.87 ± 10.06 21.00 ± 5.48
Table 3. Mean (± STD) recognition rates (percents) for “Labeled Faces in the Wild”.
Columns differ in the number of random persons per experiment.
Algorithm 10 persons 20 persons 50 persons 100 persons

Nearest Neighbor 36.00 ± 12.73 25.25 ± 7.20 18.10 ± 3.77 15.27 ± 1.90
All-Pairs Linear SVM 35.00 ± 13.67 24.37 ± 5.55 18.55 ± 3.91 14.10 ± 2.39
Multiclass CCA 40.50 ± 14.68 29.25 ± 6.93 24.15 ± 5.51 20.55 ± 2.99
Local Multiclass CCA 41.25 ± 14.77 31.25 ± 6.46 25.70 ± 5.07 21.40 ± 3.02
Labeled Faces in the Wild. From the Labeled Faces in the Wild dataset [16],
we filtered out all persons which have less than four images. 610 persons and a
total of 6, 733 images remain. The images are partly aligned via funneling [23],
and all images are 256 × 256 pixels. We only use the center 100 × 100 sub-
image, and represent it by LBP features of a grid of non-overlapping 16 pixels
blocks.
The number of persons per experiment vary from 10 to 100. For each run, 10,
20, 50 or 100 random persons and 4 random images per person are selected. 2 are
used for training and 2 for testing. Table 3 compares the classification results.
While the differences may seem small, they are significant (p < 0.01) and Local
Multiclass CCA leads the performance table followed by Multiclass CCA and
either NN or SVM. Additional experiments conducted for the 50 persons split
show that k-nearest neighbors based local learning hurts performance for all
values of k, for both SVM and Multiclass CCA.
Acknowledgments
This research is supported by the Israel Science Foundation (grants No. 1440/06,
1214/06), the Colton Foundation, and a Raymond and Beverly Sackler Career
Development Chair.
References
1. Fei-Fei, L., Fergus, R., Perona, P.: A bayesian approach to unsupervised one-shot
learning of object categories. In: ICCV, Nice, France, pp. 1134–1141 (2003)
2. Belkin, M., Niyogi, P.: Semi-supervised learning on riemannian manifolds. Machine
Learning 56, 209–239 (2004)
3. Bart, E., Ullman, S.: Cross-generalization: learning novel classes from a single ex-
ample by feature replacement. In: CVPR (2005)
4. Yamada, M., Pezeshki, A., Azimi-Sadjadi, M.: Relation between kernel cca and
kernel fda. In: IEEE International Joint Conference on Neural Networks (2005)
5. Bottou, L., Vapnik, V.: Local learning algorithms. Neural Computation 4 (1992)
6. Zhang, H., Berg, A.C., Maire, M., Malik, J.: Svm-knn: Discriminative nearest
neighbor classification for visual category recognition. In: CVPR (2006)
7. Hotelling, H.: Relations between two sets of variates. Biometrika 28, 321–377 (1936)
8. Akaho, S.: A kernel method for canonical correlation analysis. In: International
Meeting of Psychometric Society (2001)
9. Wolf, L., Shashua, A.: Learning over sets using kernel principal angles. J. Mach.
Learn. Res. 4, 913–931 (2003)
10. Neumaier, A.: Solving ill-conditioned and singular linear systems: A tutorial on
regularization (1998)
11. Hastie, T., Tibshirani, R., Friedman, J.: The elements of statistical learning: data
mining, inference and prediction. Springer, Heidelberg (2001)
12. Sherman, J., Morrison, W.J.: Adjustment of an inverse matrix corresponding to
changes in the elements of a given column or a given row of the original matrix.
Annals of Mathematical Statistics 20, 621 (1949)
13. Golub, G.: Some modified eigenvalue problems. Technical report, Stanford (1971)
14. Bunch, J.R., Nielsen, C.P., Sorensen, D.C.: Rank-one modification of the symmetric
eigenproblem. Numerische Mathematik 31, 31–48 (1978)
15. CalPhotos: A database of photos of plants, animals, habitats and other natural
history subjects [web application], animal–mammals collection. bscit, University
of California, Berkeley,
http://calphotos.berkeley.edu/cgi/
img query?query src=photos index&where-lifeform=Animal--Mammal
16. Huang, G.B., Ramesh, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild:
A database for studying face recognition in unconstrained environments. University
of Massachusetts, Amherst, Technical Report 07-49 (2007)
17. Rifkin, R., Klautau, A.: In defense of one-vs-all classification. Journal of Machine
Learning Research 5 (2004)
18. Vedaldi, A.: Bag of features: A simple bag of features classifier (2007),
http://vision.ucla.edu/∼ vedaldi/
19. Nister, D., Stewenius, H.: Scalable recognition with a vocabulary tree. In: CVPR
(2006)
20. Nowak, E., Jurie, F., Triggs, B.: Sampling strategies for bag-of-features image classi-
fication. In: European Conference on Computer Vision. Springer, Heidelberg (2006)
21. Ojala, T., Pietikainen, M., Harwood, D.: A comparative-study of texture measures
with classification based on feature distributions. Pattern Recognition 29 (1996)
22. Ahonen, T., Hadid, A., Pietikainen, M.: Face recognition with local binary pat-
terns. In: Pajdla, T., Matas, J(G.) (eds.) ECCV 2004. LNCS, vol. 3024. Springer,
Heidelberg (2004)
23. Huang, G.B., Jain, V., Learned-Miller, E.: Unsupervised joint alignment of complex
images. ICCV (2007)
Saliency Based Opportunistic Search for Object Part
Extraction and Labeling
Yang Wu1,2 , Qihui Zhu2 , Jianbo Shi2 , and Nanning Zheng1

1
Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University
ywu@aiar.xjtu.edu.cn, nnzheng@mail.xjtu.edu.cn
2
Department of Computer and Information Science, University of Pennsylvania
wuyang@seas.upenn.edu, qihuizhu@seas.upenn.edu,
jshi@cis.upenn.edu
Abstract. We study the task of object part extraction and labeling, which seeks
to understand objects beyond simply identifiying their bounding boxes. We start
from bottom-up segmentation of images and search for correspondences between
object parts in a few shape models and segments in images. Segments comprising
different object parts in the image are usually not equally salient due to uneven
contrast, illumination conditions, clutter, occlusion and pose changes. Moreover,
object parts may have different scales and some parts are only distinctive and
recognizable in a large scale. Therefore, we utilize a multi-scale shape repre-
sentation of objects and their parts, figural contextual information of the whole
object and semantic contextual information for parts. Instead of searching over a
large segmentation space, we present a saliency based opportunistic search frame-
work to explore bottom-up segmentation by gradually expanding and bounding
the search domain. We tested our approach on a challenging statue face dataset
and 3 human face datasets. Results show that our approach significantly outper-
forms Active Shape Models using far fewer exemplars. Our framework can be
applied to other object categories.
1 Introduction
We are interested in the problem of object detection with object part extraction and la-
beling. Accurately detecting objects and labeling their parts requires going inside the
object’s bounding box to reason about object part configurations. Extracting object parts
with the right configuration is very helpful for recognizing object details. For example,
extracting facial parts helps with recognizing faces and facial expressions, while under-
standing human activities requires knowing the pose of a person.
A common approach to solve this problem is to learn specific features for object
parts [1][2]. We choose a different path which starts with bottom-up segmentation and
aligns shape models to segments in test images. Our observation is that starting from
salient segments, it is unlikely to accidentally align object parts to background edges.
Therefore, we can search efficiently and avoid accidental alignment.
Our approach includes three key components: correpondence, contextual informa-
tion and saliency of segments. There exist algorithms incorporating correspondence
and contextual information such as pictorial structures [3] and contour context selec-
tion [4], both showing good performance on some object categories. The disadvantage

Saliency Based Opportunistic Search for Object Part Extraction and Labeling 761
Image Contour(TH=0.4) Contour 45Seggments 60Segments ZoomͲinSegmentation
Fig. 1. Saliency of contours and segments. The second image is a group of salient contours from
contour grouping [5] by setting a lower threshold to the average edge strength, while the third one
contains all the contours from contour grouping. It shows that by thresholding the saliency of con-
tour segments, we either get some foreground contours missing (under-segmented) or have a lot
of clutter come in (over-segmented). The same thing happens to image segmentation. Segments
comprising object parts pop out in different segmentation levels, representing different saliencies
(cut costs). The last three images show such a case.
is that these methods ignore image saliency. Therefore, they cannot tell accidental align-
ment of faint segments in the background from salient object part segments. However, it
is not easy to incorporate saliency. A naive way of using saliency is to find salient parts
first, and search for less salient ones depending on these salient ones. The drawback is
that a hard decision has to be made in the first step of labeling salient parts, and mistakes
arising from this step cannot be recovered later. Moreover, object parts are not equally
hard to find. Segments belonging to different object parts may pop out at different seg-
mentation levels (with different numbers of segments), as shown in Figure 1. One could
start with over-segmentation to cover all different levels. Unfortunately, by introducing
many small segments at the same time, segment saliency will be lost, which defeats the
purpose of image segmentation. Fake segmentation boundaries will also cause many
false positives of accidentally aligned object parts.
We build two-level contexts and shape representations for objects and their parts,
with the goal of high distinctiveness and efficiency. Some large object parts (e.g. facial
silhouettes) are only recognizable as a whole in a large scale, rather than as a sum of
the pieces comprising them. Moreover, hierarchical representation is more efficient for
modeling contextual relationships among model parts than a single level representation
which requires a large clique potential and long range connections. Two different levels
of contextual information is explored: figural context and semantic context. The former
captures the overall shape of the whole object, and the latter is formed by semantic
object parts.
In this paper, we propose a novel approach called Saliency Based Opportunistic
Search for object part extraction and labeling, with the following key contributions:
1. Different levels of context including both figural and semantic context are used.
2. Bottom-up image saliency is incorporated into the cost function.
3. We introduce an effective and efficient method of searching over different segmen-
tation levels to extract object parts.
762 Y. Wu et al.
2 Related Work
It has been shown that humans recognize objects by their components [6] or parts [7].
The main idea is that object parts should be extracted and represented together with
the relationships among them for matching to a model. This idea has been widely used
for the task of recogize objects and their parts [8,9,3]. Figural and semantic contextual
information play an important role in solving this problem. Approaches that take ad-
vantage of figural context include PCA and some template matching algorithms such as
Active Shape Models (ASM) [10] and Active Appearance Models (AAM) [11]. Tem-
plate matching methods like ASM usually use local features (points or key points) as
searching cues, and constrain the search by local smoothness or acceptable variations
of the whole shape. However, these methods require good initialization. They are sen-
sitive to clutter and can be trapped in local minima. Another group of approaches are
part-based models, which focus on semantic context. A typical case is pictorial struc-
ture [3]. Its cost function combines both the individual part matching cost and pair-wise
inconsistency penalties. The drawback of this approach is that it has no figural context
measured by the whole object. It may end up with many “OK” part matches without a
global verification, especially when there are many faint object edges and occlusions in
the image. Recently, a multiscale deformable part model was proposed to detect objects
based on deformable parts [1], which is an example that uses both types of contextual
information. However, it focuses on training deformable local gradient-based features
for detecting objects, but not extracting object parts out of the images.
3 Saliency Based Opportunistic Search
Problem definition. The problem we are trying to solve is to extract and label ob-
ject parts based on contextual information, given an image and its segmentations, as
shown in Figure 1. Multiple models are used to cover some variations of the object (see
Figure 2 for the models we have used on faces). Extracting and labeling object parts
requires finding the best matched model. The problem can be formulated as follows:
Input
– Model: M = {M1 , M2 , . . . , Mm }; each model Mk has a set of labeled parts

{pk1 , pk2 , . . . , pkn }. They are all shape models made of contours and line segments.
– Image: S = {s1 , s2 , . . . , sl } is a set of region segments and contour segments
coming from different segmentation levels from the image. For region segments,
only boundaries are used for shape matching.
Output
– Best matched model Mk .

– Object part labels L(S). L(si ) = j, if si belongs to part pkj , or else L(si ) = 0.
This can be formulated as a shape matching problem, which aims to find sets of
segments whose shapes match to part models. However, the segments comprising the
object parts are not equally hard to extract from the image, and grouping them to objects
Fig. 2. Different models for faces. They are hand designed models obtained from 7 real images,
each of them representing one pose. Facial features are labeled in different colors.
and object parts also requires checking the consistency among them. We call these ef-
forts “grouping cost”, which is not measured by shape but can be helpful to differentiate
segments belonging to object parts from those belonging to the background. Therefore,
we combine these two into such a cost function:
C labeling = C shape + C grouping (1)
C shape measures the shape matching cost between shape models and labeled segments
in the image, which relays much on correspondence and context. C grouping is the
grouping cost, which can be measured in different ways, but in this paper it is mainly
about the bottom-up saliency based editing cost.
The cost function above is based on the following three key issues.
1. Correspondence (u). A way to measure the dissimilarity between a shape model

and a test image. The correspondence is defined on control points. Features com-
puted on these control points represent the shape information and then the corre-
spondences are used to measure the dissimilarity. Let U M = {a1 , a2 , . . . , aNa }
be a set of control points on the model, and U I = {b1 , b2 , . . . , bNb } be the set on
the image. We use uij to denote the correspondence between control points ai and
bj where uij = 1 indicates they are matched, otherwise uij = 0. Note that this
correspondence is different from the one between object parts and image segments.
2. Context (x and y). The idea of using the context is to choose the correct context
on both model and test image sides for shape matching invariant to clutter and
occlusion. x and y are used here to denote the context selection of either segments
or parts on the model and the image, respectively.
3. Saliency. A property of bottom-up segments which represents how difficult it is
to separate the segment from the background. Coarse-level segmentation tends to
produce salient segments, while finer-level segmentation extracts less salient ones,
but at the same time introduces background clutter. Local editing on the salient gap
between two salient segments can help to get good segments out without bringing
in clutter, but it needs contextual guidence.
Saliency based editing. Segmentation has problems when the image segments have
different saliencies. Under-segmentation could end up with unexpected leakages, while
over-segmentation may introduce clutter. A solution for this problem is to do some
local editings. For example, adding a small virtual edge at the leakage place can make
the segmentation much better without increasing the number of segments. Zoom-in in
a small area is also a type of editing that can be effective and efficient, as presented in
764 Y. Wu et al.
Figure 1. Small costs for editing can result in big improvement on shape matching
cost. This is based the shape integrity and the non-additive distance between shapes.
However, editings need the contextual information from the model.
Suppose there are a set of possible editing operations z which might lead to better
segmentation. zk = 1 means that editing k is chosen, otherwise zk = 0. Note that
usually it is very hard to discover and precompute all the editings beforehand. There-
fore, this editing index vector z is dynamic, and it appends on the fly. After doing some
editings, some new segments/(part hypotheses) will come out, meanwhile we can still
keep the original segments/parts. Therefore, a new variable yedit = yedit (y, z) is used
to denote all the available elements which includes both the original ones in y and the
new ones induced by editing z. Let Ckedit be the edit cost for editing k.
Our cost function (1) of object part labeling and extraction can be written as follows:
min C labeling (x, y, z, u) = C shape (x, y, z, u) + C grouping (z) =

x,y,z,u

Na
Nb
M↔I
[β· uij Cij (x, yedit ) + CiF ↔M (x, u)] + Ckedit zk (2)
i=1 j=1 k

s.t. j uij ≤ 1, i = 1, ..., Na
x: selection indicator of model segments/parts.

y: selection indicator of image segments/parts.
z: selection vector of editing operations.
u: correspondence of control points between the image and model.
yedit (y, z): selection indicator of image segments/parts edited by z.
The three summations in equation (2) correspond to three different types of cost: mis-
match cost C M↔I (x, yedit , u), miss cost C F ↔M (x, u) and edit cost C edit (z). The
M↔I
mismatch cost, Cij (x, yedit ) = fi (x) − fj (yedit ) denotes the feature dissim-
ilarity between two corresponding control points. To prevent the cost function from
biasing to fewer matches, we add the miss cost CiF ↔M (x) = fif ull − ( j uij )fi (x)
to denote how much of the model has not been matched by the image. It encourages
M↔I
more parts to be matched on the model side. There is a trade-off between Cij
F ↔M
and Ci , where β ≥ 0 is a controlling factor. Note that · can be any norm
function1.
The rest of this section focuses on the two parts of our cost function. Shape matching
will be performed on two levels of contexts and saliency based editing will result in the
opportunistic search approach.
3.1 Two-Level Context Based Shape Matching
We extend the shape matching method called contour context selection in [4] to two
different contextual levels: “figural context selection” and “semantic context selection”.
1
In our shape matching we used L1 norm.
Figural context selection. Figural context selection matches a segment-based holistic

shape model to an object hypothesis represented by segments, which may have clutter
and missing segments. We optimize the following cost function:
min C f igural (x, y, u) =

x,y,u

Na
Nb

[β· uij SCiM (x) − SCjI (y)+SCiF − ( j uij ) · SCiM (x))]
i=1 j=1

M↔I
Cij (x,yedit ) CiF↔M (x,u)
geo
s.t. uij ui j Ci,j,i ,j ≤ Ctol (3)
i,j,i ,j
where SCiM (x) and SCjI (y) is defined as the Shape Context centered at model control
geo
point ai and image control point bj . Ci,j,i ,j is the geometric inconsistent cost of cor-
respondences u. Ctol is the maximum tolerance of the geometric inconsistency. We use

Shape Context [12] as our feature descriptor. Note that the size of Shape Context his-
togram is large enough to cover the whole object model, and this is a set-to-set matching
problem. Details for this algorithm can be found in [4].
Semantic context selection. Similarly we explore semantic context to select consistent

object part hypotheses. We first generate part hypotheses using almost the same context
selection algorithm as the one presented above. The selection operates on parts instead
of the whole object. Figure 3 shows an example of generating a part hypothesis.
In semantic context selection, we reason about semantic object parts. Hence we ab-
stract each part (on either model or test image) as a point located at its center with its
part label. We place control points on each one of the part centers.
part
γC
Suppose Cjpart is the matching cost of part hypothesis j. We use wjP = e eγ ∈
j
[ eγ , 1], γ ∈ [0, 1] as its weight. Then the cost function for semantic context
1
selection is:
Fig. 3. Semantic context selection. Left: Part hypothesizing. a) A local part region around the eye
in the image, with segments and control points. c) A model template of the eye with control points.
Selection result on the image is shown in b). Right: Consistent part grouping. Semantic-level
shape context centered on the left eye captures semantic contextual information of the image. A
subset of those parts form a mutually consistent context and we group them by matching with the
semantic-level shape context on the model shown in the middle.
766 Y. Wu et al.
min C semantic (x, y, u) =

x,y,u

Na
Nb

[β· uij wjP SCiM (x) − SCjI (y)+SCiF − ( j uij ) · SCiM (x))]
i=1 j=1

M↔I
Cij (x,yedit ) CiF↔M (x,u)
(4)
The variable definitions are similar to figural context selection, except for two differ-
ences: 1) selection variables depend on the correspondences and 2) Shape Context no
longer counts edge points, but object part labels.
The desired output of labeling L(S) is implicitly given in the optimization variables.
During part hypothesis generation, we put labels of candidate parts onto the segments.
Then after semantic context selection, we confirm some labels and discard the others
using the correspondence uij between part candidates and object part models.
3.2 Opportunistic Search
Labeling object parts using saliency based editing potentially requires searching over a
very large state space. Matching object shape and its part configuration requires com-
puting correspondences and non-local context. Both of them have exponentially many
choices. On top of that, we need to find a sequence of editings, such that the resulting
segments and parts produced by these editings are good enough for matching.
The key intuition of our saliency based opportunistic search is that we start from
coarse segmentations which produce salient segments and parts to guarantee low
saliency cost. We iteratively match configuration of salient parts to give a sequence
of bounds to the search zone of the space which needs to be explored. The possible
spatial extent of the missing parts is bounded by their shape matching cost and the
edit cost (equally, saliency cost). Once the search space has been narrowed down, we
“zoom-in” to the finer scale segmentation to rediscover missing parts (hence with lower
saliency). Then we “zoom-out” to do semantic context selection on all the part hy-
potheses. Adding these new parts improves the bound on the possible spatial extent
and might suggest new search zones. This opportunistic search allows both high effi-
ciency and high accuracy of object part labeling. We avoid extensive computation by
narrowing down the search zone. Furthermore, we only explore less salient parts if there
exist salient ones supporting them, which avoids producing many false positives from
non-salient parts.
Search Zone. In each step t of the search, given (x(t−1) , y(t−1) , z(t−1) , u(t−1) ), we use
ΔC M↔I (x, yedit ) to denote the increment of C M↔I (x, yedit ) (the first summation in
equation (2)). ΔC F ↔M (x, u) and ΔC edit (z) are similarly defined. By finding missing
parts, we seek to decrease the cost (2). Therefore, we introduce the following criterion
for finding missing parts:
βΔC M↔I (x, y, z) + ΔC F ↔M (x, u) + ΔC edit (z) ≤ 0 (5)
We write C M↔I (x, y, z)=C M↔I (x, yedit ) since yedit depends on editing vector z.
Algorithm 1. Saliency Based Opportunistic Search

1: Initialize using figural context selection. For each part k, compute Z(k) based on u from
figural context selection. Set (x(0) , y(0) , z(0) , u(0) ) to zeros. Set t = 1.
2: Compute search zones for all the missing parts. Find all missing parts by thresholding the
solution x(t−1) .
for each missing part pk
If Z(k) = ∅, compute search zone set Z(k) by equation (9) and (10).
end
3: Zoom-in search zone. Update editing set z.
(t−1)
for each xk where Z(k) = ∅
Perform Ncut segmentation for each zoom-in window indexed by elements in Z(k).
Generate part hypotheses. Set Z(k) = ∅.
If no candidates can be found, go to the next missing part.
Update z from part hypotheses.
end
4: Evaluate configurations with re-discovered parts.
Terminate if z does not change.
Update (x(t) , y(t) , z(t) , u(t) ) with the rediscovered parts using equation (4).
Terminate if C semantic (x, y, u) does not improve.
t = t + 1. Go to step 2.
The estimation of bounds is based on the intuition that if all the missing parts can be
found, then no miss cost is needed to pay any more. Therefore, according to equation (4):

ΔC F ↔M (x) ≥ − CiF ↔M (x, u). (6)
i
This is the upper bound for the increment of either one of the other two items in equation
(5) when any new object part is matched.
(t)
Suppose a new editing zα = 1|z(t−1) =0 matches a new object part ak to a part
α
(t) (t−1)
hypothesis in the image b . Let k ↔ # indicate uk = 1 and j ukj = 0. Then this
editing at least has to pay the cost of matching ak to b (we do not know whether others
will also match or not):
C|k↔ = βΔC M↔I (x, y, z)|k↔ + Cαedit . (7)
The first item on the right of equation (7) is the increment of mismatch ΔC M↔I
(x, yedit ) when a new object part ak get matched to b . It can be computed based
on the last state of the variables (x(t−1) , y(t−1) , z(t−1) , u(t−1) ). According to above
equations, we get

βΔC M↔I (x, y, z)|k↔ + Cαedit − CiF ↔M (x(t−1) , u(t−1) ) ≤ 0 (8)
i
Since we use Shape Context for representation and matching, the mismatch is non-
decreasing. And also the editing cost is nonnegative, so we abtain the bounds for the
(t)
new editing zα = 1|z(t−1) =0 . Let Z(k) denote the search zone for object part k. Then
α
we can compute two bounds for Z(k):
768 Y. Wu et al.
1 F ↔M (t−1) (t−1)
(Supremum) Z sup (k) = {zα |ΔC M↔I (x, y, z)|k↔ ≤ C (x ,u )}
β i i
(9)

(Infimium) Z inf
(k) = {zα |Cαedit ≤ CiF ↔M (x(t−1) , u(t−1) )} (10)
i
where Z sup gives the supremum of the search zone, i.e. upper bound of zoom-in win-
dow size, and Z inf gives the infimum of the search zone, i.e. lower bound of zoom-in
window size. When the number of segments is fixed, the saliency of the segments de-
creases as the window size becomes smaller. Z sup depends on mismatch and Z inf
depends on the edit cost (i.e. saliency). In practice, one can sample the space of the
search zone, and check which ones fall into these two bounds.
Our opportunistic search is summarized in Algorithm 1.
4 Implementation
4.1 A Typical Example
We present more details on the opportunistic search using faces as an example in
Figure 4. We found that usually the whole shape of the face is more salient than individ-
ual facial parts. Therefore, the procedure starts with figural context and then switchs to
semantic context. We concretize our algorithm for this problem in the following steps.
The same procedure can be applied to similar objects.
1. Initialization: Object Detection. Any object detection method can be used, but it
is not a necessary step2 . We used shape context voting [13] to do this task, which
can handle different poses using a small set of positive training examples.
2. Context Based Alignment. First, use C f igural in equation (3) to select the best
matched model Mk and generate the correspondences uf igural for rough align-
ment3 . When the loop comes back again, update the alignment based on usemantic .
Estimate locations for other still missing parts.
3. Part Hypotheses Generation. Zoom in on these potential part locations by crop-
ping the regions and do Ncut segmentation to get finer scale segmentation. Then
match them to some predefined part models. The resulting matching score is used
to prune out unlikely part hypotheses, according to the bound of the cost function.
4. Part Hypotheses Grouping. Optimize C semantic in equation (4). Note that the best
scoring group may consist of only a subset of the actual object parts.
5. Termination Checking. If no better results can be obtained, then we go to the next
step. Or else we update semantic context and go back to step 2.
6. Extracting Facial Contours. This is a special step for faces only. With the final set
of facial parts, we optimize C f igural again to extract the segments that correspond
to the face silhouette, which can be viewed as a special part of the face.
2
Figural context selection can also be used to do that [4].
3
In practice, we kept best two model hypotheses.
Fig. 4. Saliency based opportunistic search, using faces as an example. Top: the flowchart. Bot-
tom: results of each step for 3 different examples. Typically the iteration converges after only one
or two rounds. Rectangles with different colors indicate the zoom-in search zones for different
parts. Note that when zoom-in is performed for the first time, two adjacent parts can be searched
together for efficiency. This figure is best viewed in color.
4.2 Two-Level Context Selection

For simplification, we do not consider any editing in figural context selection. Then
equation (3) is an integer programming problem, we relaxed the variables to solve it
with LP. Details of this context selection algorithm can be found in [4].
For semantic context selection, we need to search for correspondences and part se-
lection variables simultaneously because they are highly dependent, unlike the situation
in figural context selection. Therefore, we introduce a correspondence context vector
PijM = uij x to expand the selection space for model parts:
M
PijM ∈ {0, 1}|U |
: PijM (i ) ⇔ uij = 1 ∧ x(i ) = 1 (11)
Similarly, we define the correspondence context vector for image parts,
I
PijI ∈ {0, 1}|U |
: PijI (j ) ⇔ uij = 1 ∧ y(j ) = 1 (12)
In addition to the cost in equation (4), constraints on context correspondence vec-
tor P M , P I are enforced such that the semantic context viewed by different parts are
770 Y. Wu et al.
Table 1. Constraints on context correspondence vector P M , P I . For example, Context complete-

ness requires that contexts must include all the matched parts. If both i and i are
P matched parts,
the context viewed from i must include i , i.e. (y(i) = 1) ∧ (y(i ) = 1) ⇒ j PijM (i ) = 1,
which is relaxed as the constraint in row 4. Other constraints are constructed in a similar way.
P P
Self consistency j PijM (i) = y(i), I
i Pij (j) = x(j)
P M
P M
P (i ) ≤ y(i ), P (i ) ≤ y(i )
One-to-one matching Pi ijI Pj ijI
i Pij (j ) ≤ x(j ), j Pij (j ) ≤ x(j )
Context reflexitivity PijM (i ) ≤ PijM (i), PijI (j ) ≤ PijI (j)

P P
Context completeness y(i) − j PijM (i ) ≤ 1 − y(i ), x(j) − i PijI (j ) ≤ 1 − x(j )
P M P M P I P I
Mutual context support j Pij (i ) = j Pi j (i), i Pij (j ) = i Pi j (j)
Fig. 5. Left: averaged models for ASM1. Right: averaged model for ASM3.
Table 2. Comparison of experimental details on Emperor-I dataset
Method No. of Poses Silhouette No. of Training No. of Test Average point error
ASM1 7 w 138 86 0.2814
ASM2 5 w/o 127 81 0.2906
ASM3 3 w/o 102 70 0.3208
Ours 7 w 7+16 86 0.1503
Table 3. Average error, normalized by distance between eyes for ASM vs. our method
Method Global Eyebrows Eyes Nose Mouth Silhouette

ASM1 0.3042 0.2923 0.2951 0.2715 0.2524 0.3126
Ours 0.1547 0.2015 0.1142 0.1546 0.1243 0.1353
consistent with each other. These constraints are summarized by the table 1. The cost
function and constraints are linear. We relaxed the variables and solved it with LP.
5 Experiments and Results
Datasets. We tested our approach on both statue faces from the Emperor-I dataset [14]
and real faces from various widely used face databases (UMIST, Yale, and Caltech
Faces). Quantitative comparison was done on the Emperor-I dataset and we also show
some qualitative results on a sample set of all these datasets. The statue face dataset has
All Parts Eyes

1.5 1.5
ASM1 ASM1
ASM2 Ours
ASM3
Average Point Error
Average Point Error

Ours
1 1
0.5 0.5
0 0
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Image Number Image Number
Fig. 6. Average point error vs. image number. All the values are normalized by the estimated
distance of two eyes in each image. The vertical dot-dash lines separate images of different poses.
some difficulties that normal faces do not have: lack of color cue, low contrast, inner
clutter, and great intra-subject variation.
Comparison measurement. The comparison is between Active Shape Models [10] and
our approach. Since we extract facial parts by selecting contours, our desired result is
that the extracted contours are all in the right places and correctly labeled. However,
ASM generates point-wise alignment between the image and a holistic model. Due to
the differences, we chose to use “normalized average point alignment error” measure-
ment for alignment comparison.
Since our results are just labeled contours, we do not have point correspondences for
computing the point alignment error. Therefore, we relaxed the measurement to the dis-
tance between each ground truth key point and its closest point on the contours belong
to the same part. To make the comparison fair, we have exactly the same measurement
for ASM by using spline interpolation to generate “contours” for its facial parts. We use
0.35 times the maximum height of the ground truth key points as an approximation of
the distance between two eyes invariant to pose changes as the our normalizing factor.
Experiments. There are two aspects of our Emperor-I dataset that may introduce dif-
ficulties for ASM: few training examples with various poses and dramatic face silhou-
ette changes. Therefore, we designed three variants of ASM to compensate for these
challenges, denoted in our plots as “ASM1”,“ASM2”,“ASM3”. Table 2 shows the dif-
ferences. Basically, ASM2 and ASM3 disregard face silhouette and work on fewer
poses that may have relatively more exemplars. Note that ASM3 even combined the
training data of the three near-frontal poses as a whole. We used “leave-one-out” cross-
validation for ASM. For our method, we picked up 7 images for different poses (one for
each pose), labeled them and extracted the contours out to work as our holistic models.
Moreover, we chose facial part models (usually combined by 2 or 3 contours) from a
total of 23 images which also contained these 7 images. Our holistic models are shown
in Figure 2 and Figure 5 shows those averaged ones for ASM.
In Figure 6, we show the alignment errors for all the facial parts together and also
those only for the eyes. Other facial parts have similar results so we leave them out. In-
stead, we provide a summary in Table 3 and a comparison in the last column of Table 2,
where each entry is the mean error across the test set or test set fold, as applicable. We
772 Y. Wu et al.
Rough Alignment
Contour Grouping
ASM
2 1 2 1
42 1 3
3 2 1 7 4 3 8 5 7
Our Result
2 13 4 7
4
3
5
7 6 2
4 7
8 5 5 2 13 5
5 4 7
6 6 8 6 8
5 6
6
6
Final Alignment
Rough Alignment
Contour Grouping
ASM
Our Result
Final Alignment
Fig. 7. A subset of the results. Upper group is on the Emperor-I dataset and the lower is for
real faces from various face databases (1-2 from UMIST, 3-4 from Yale, and 5-7 from Caltech).
Matched models, control points and labeled segments are superimposed on the images.
can see that our method performs significantly better than ASM on all facial parts with
significantly fewer training examples. We provide a qualitative evaluation of the results
in Figure 7, where we compare the result of ASM and our method on a variety of im-
ages containing both statue faces and real faces. These images show great variations,
especially of those statue faces. Note that the models are only trained on statue faces.
6 Conclusion
We proposed an object part extraction and labeling framework which incorporates two-
level contexts and saliency based opportunistic search. The combination of figural con-
text on the whole object shape and semantic context on parts enables robustly search
matching of object parts and image segments in cluttered images. Saliency further im-
proves this search by gradually exploring salient bottom-up segmentations and bound-
ing it via shape matching cost. Experimental results on several challenging face datasets
demonstrate that our approach can accurately label object parts such as facial features
and resist to accidental alignment.
Acknowledgment. This research was supported by China Scholarship Council, Na-

tional Science Foundation (Grant NSF-IIS-04-47953(CAREER) and NSF-IIS-03-33036
(IDLP)), National Basic Research Program of China (Grant No. 2006CB708303 and No.
2007CB311005), and National High-Tech Research and Development Plan of China
(Grant No. 2006AA01Z192). We would like to acknowledge the help from Praveen
Srinivasan, and the discussions and technical help from Liming Wang. Special thanks
are given to Geng Zhang for experimental help on ASM and the ground truth labeling.
References
1. Felzenszwalb, P., McAllester, D., Ramanan, D.: A discriminatively trained, multiscale, de-
formable part model. In: CVPR (2008)
2. Ferrari, V., Jurie, F., Schmid, C.: Accurate object detection with deformable shape models
learnt from images. In: CVPR, pp. 1–8 (2007)
3. Felzenszwalb, P.F., Huttenlocher, D.P.: Pictorial structures for object recognition.
IJCV 61(1), 55–79 (2005)
4. Zhu, Q., Wang, L., Wu, Y., Shi, J.: Contour context selection for object detection: A single
exemplar suffices. In: ECCV (2008)
5. Zhu, Q., Shi, J.: Untangling cycles for contour grouping. In: ICCV (2007)
6. Biederman, I.: Recognition by components: A theory of human image understanding. Psy-
chR 94(2), 115–147 (1987)
7. Pentland, A.: Recognition by parts. In: ICCV, pp. 612–620 (1987)
8. Amit, Y., Trouve, A.: Pop: Patchwork of parts models for object recognition. IJCV 75(2),
267–282 (2007)
9. Sudderth, E., Torralba, A., Freeman, W., Willsky, A.: Learning hierarchical models of scenes,
objects, and parts. In: ICCV, pp. 1331–1338 (2005)
10. Cootes, T., Taylor, C., Cooper, D., Graham, J.: Active shape models: Their training and ap-
plication. CVIU 61(1), 38–59 (1995)
11. Cootes, T., Edwards, G., Taylor, C.: Active appearance models. PAMI 23(6), 681–685 (2001)
774 Y. Wu et al.
12. Belongie, S., Malik, J., Puzicha, J.: Shape matching and object recognition using shape con-
texts. IEEE Trans. Pattern Anal. Mach. Intell. (2002)
13. Wang, L., Shi, J., Song, G., fan Shen, I.: Object detection combining recognition and seg-
mentation. In: ACCV (1), pp. 189–199 (2007)
14. Chen, C.: The First Emperor of China. Voyager Company (1994)
Stereo Matching: An Outlier Confidence Approach
Li Xu and Jiaya Jia
Department of Computer Science and Engineering

The Chinese University of Hong Kong
{xuli,leojia}@cse.cuhk.edu.hk
Abstract. One of the major challenges in stereo matching is to handle partial

occlusions. In this paper, we introduce the Outlier Confidence (OC) which dy-
namically measures how likely one pixel is occluded. Then the occlusion infor-
mation is softly incorporated into our model. A global optimization is applied to
robustly estimating the disparities for both the occluded and non-occluded pix-
els. Compared to color segmentation with plane fitting which globally partitions
the image, our OC model locally infers the possible disparity values for the out-
lier pixels using a reliable color sample refinement scheme. Experiments on the
Middlebury dataset show that the proposed two-frame stereo matching method
performs satisfactorily on the stereo images.
1 Introduction
One useful technique to reduce the matching ambiguity for stereo images is to incor-
porate the color segmentation into optimization [1,2,3,4,5,6]. Global segmentations im-
prove the disparity estimation in textureless regions; but most of them do not necessarily
preserve accurate boundaries. We have experimented that, when taking the ground
truth occlusion information into optimization, very accurate disparity estimation can
be achieved. This shows that partial occlusion is one major source of matching errors.
The main challenge of solving the stereo problems now is the appropriate outlier detec-
tion and handling.
In this paper, we propose a new stereo matching algorithm aiming to improve the
disparity estimation. Our algorithm does not assign each pixel a binary visibility value
indicating whether this pixel is partially occluded or not [7,4,8], but rather introduces
soft Outlier Confidence (OC) values to reflect how confident we regard one pixel as an
outlier. The OC values, in our method, are used as weights balancing two ways to infer
the disparities. The final energy function is globally optimized using Belief Propagation
(BP). Without directly labeling each pixel as “occlusion” or “non-occlusion”, our model
has considerable tolerance of errors produced in the occlusion detection process.
Another main contribution of our algorithm is the local disparity inference for out-
lier pixels, complementary to the global segmentation. Our method defines the disparity
similarity according to the color distance between pixels and naturally transforms color
sample selection to a general foreground or background color inference problem using
image matting. It effectively reduces errors caused by inaccurate global color segmen-
tation and gives rise to a reliable inference of the unknown disparity of the occluded
pixels.

776 L. Xu and J. Jia
We also enforce the inter-frame disparity consistency and use BP to simultaneously

estimate the disparities of two views. Experimental results on the Middlebury dataset [9]
show that our OC model effectively reduces the erroneous disparity estimate due to
outliers.
2 Related Work
A comprehensive survey of the dense two-frame stereo matching algorithms was given
in [10]. Evaluations of almost all stereo matching algorithms can be found in [9]. Here
we review previous work dealing with outliers because, essentially, the difficulty of
stereo matching is to handle the ambiguities.
Efforts of dealing with outliers are usually put in three stages in stereo matching –
that is, the cost aggregation, the disparity optimization, and the disparity refinement.
Most approaches use outlier truncation or other robust functions for cost computation
in order to reduce the influence of outliers [2,11].
Window-based methods aggregate matching cost by summing the color differences
over a support region. These methods [12,13] prevent depth estimation from aggre-
gating information across different depth layers using the color information. Yoon and
Kweon [14] adjusted the support-weight of a pixel in a given window based on the
CIELab color similarity and its spatial distance to the center of the support window.
Zitnick et al. [12] partitioned the input image and grouped the matching cost in each
color segment. Lei et al. [15] used segmentation to form small regions in a region-tree
for further optimization.
In disparity optimization, outliers are handled in two ways in general. One is to
explicitly detect occlusions and model visibility [7,4,8]. Sun et al. [4] introduced the
visibility constraint by penalizing the occlusions and breaking the smoothness between
the occluded and non-occluded regions. In [8], Strecha et al. modeled the occlusion as
a random outlier process and iteratively estimated the depth and visibility in an EM
framework in multi-view stereo. Another kind of methods suppresses outliers using
extra information, such as pixel colors, in optimization. In [16,6], a color weighted
smoothness term was used to control the message passing in BP. Hirschmuller [17] took
color difference as the weight to penalize large disparity differences and optimized the
disparities using a semi-global approach.
Post-process was also introduced to handle the remaining outliers after the global or
local optimization. Occluded pixels can be detected using a consistency check, which
validates the disparity correspondences in two views [10,4,17,6]. Disparity interpola-
tion [18] infers the disparities for the occluded pixels from the non-occluded ones by
setting the disparities of the mis-matched pixels to that of the background. In [1,3,4,5,6],
color segmentation was employed to partition images into segments, each of which is
refined by fitting a 3D disparity plane. Optimization such as BP can be further applied
after plane fitting [4,5,6] to reduce the possible errors.
Several disparity refinement schemes have been proposed for novel-view synthe-
sis. Sub-pixel refinement [19] enhances details for synthesizing a new view. In [12]
and [20], boundary matting for producing seamless view interpolation was introduced.
These methods only aim to synthesize natural and seamless novel-views, and cannot be
directly used in stereo matching to detect or suppress outliers.
Stereo Matching: An Outlier Confidence Approach 777
3 Our Model
Denoting the input stereo images as Il and Ir , and the corresponding disparity maps as
Dl and Dr respectively, we define the matching energy as
E(Dl , Dr ; Il , Ir ) = Ed (Dl ; Il , Ir ) + Ed (Dr ; Il , Ir ) + Es (Dl , Dr ), (1)
where Ed (Dl ; Il , Ir ) + Ed (Dr ; Il , Ir ) is the data term and Es (Dl , Dr ) defines the
smoothness term that is constructed on the disparity maps. In our algorithm, we not
only consider the spatial smoothness within one disparity map, but also model the con-
sistency of disparities between frames.
As the occluded pixels influence the disparity estimation, they should not be used in
stereo matching. In our algorithm, we do not distinguish between occlusion and image
noise, but rather treat all problematic pixels as outliers. Outlier Confidences (OCs) are
computed on these pixels, indicating how confident we regard one pixel as an outlier.
The outlier confidence maps Ul and Ur are constructed on the input image pair. The
confidence Ul (x) or Ur (x) on pixel x is a continuous variable with value between 0 and
1. Larger value indicates higher confidence that one pixel is an outlier, and vice versa.
Our model combines an initial disparity map and an OC map for two views. In the
following, we first introduce our data and smoothness terms. The construction of the
OC map will be described in Section 4.2.
3.1 Data Term
In the stereo configuration, pixel x in Il corresponds to pixel x − dl in Ir by disparity

dl . Similarly, x in Ir corresponds to x + dr in Il . All possible disparity values for dl
and dr are uniformly denoted as set Ψ , containing integers between 0 and N , where N
is the maximum positive disparity value. The color of pixel x in Il (or Ir ) is denoted as
Il (x) (or Ir (x)). We define the data term Ed (Dl ; Il , Ir ) on the left image as
f0 (x, dl ; Il , Ir ) f1 (x, dl ; Il )
Ed (Dl ; Il , Ir ) = [(1 − Ul (x))( ) + Ul (x)( )], (2)
x
α β
where α and β are weights. f0 (x, d; Il , Ir ) denotes the color dissimilarity cost between
two views. f1 (x, d; Il ) is the term defined as the local color and disparity discontinuity
cost in one view. Ed (Dr ; Il , Ir ) on the right image can be defined in a similar way.
The above two terms, balanced by the outlier confidence Ul (x), model respectively
two types of processes in disparity computation. Compared to setting Ul (x) as a bi-
nary value and assigning pixels to either outliers or inliers, our cost terms are softly
combined, tolerating possible errors in pixel classification.
For result comparison, we give two definitions of f0 (x, dl ; Il , Ir ) respectively cor-
responding to whether the segmentation is incorporated or not. The first is to use the
color and distance weighted local window [14,6,19] to aggregate color difference be-
tween conjugate pixels:
(1)
f0 (x, dl ; Il , Ir ) = min(g(Il (x) − Ir (x − dl )1 ), ϕ), (3)
where g(·) is the aggregate function defined similarly to Equation (2) in [6]. We use the
default parameter values (local window size 33 × 33, βcw = 10 for normalizing color
differences, γcw = 21 for normalizing spatial distances). ϕ determines the maximum
cost for each pixel, whose value is set as the average intensity of pixels in the correlation
volume.
The second definition is given by incorporating the segmentation information.
Specifically, we use the Mean-shift color segmentation [21] with default parameters
(spatial bandwidth 7, color bandwidth 6.5, minimum region size 20) to generate color
segments. A plane fitting algorithm using RANSAC (similar to that in [6]) is then ap-
plied to producing the regularized disparity map dpf . We define
(2) (1)
f0 (x, dl ; Il , Ir ) = (1 − κ)f0 (x, dl ) + κα|d − dpf |, (4)
where κ is a weight balancing two terms.

f1 (x, dl ; Il ) is defined as the cost of assigning local disparity when one pixel has
chance to be an outlier.

f1 (x, dl ; Il ) = ωi (x; Il )δ(dl − i), (5)
i∈Ψ
where δ(·) is the Dirac function, Ψ denotes the set of all disparity values between 0 and
N and ωi (x; Il ) is a weight function for measuring how disparity dl is likely to be i.
We omit subscript l in the following discussion of ωi (x; Il ) since both the left and right
views can use the similar definitions.
For ease of explanation, we first give a general definition of weight ωi (x; I), which,
in the following descriptions, will be slightly modified to handle two extreme situations
with values 0 and 1. We define
L(I(x), Ii (Wx ))
ωi (x; I) = 1 − , (6)
L(I(x), Ii (Wx )) + L(I(x), I=i (Wx ))
where I(x) denotes the color of pixel x and Wx is a window centered at x. Suppose
after initialization, we have collected a set of pixels x detected as inliers within each
Wx (i.e., U (x ) = 0), and have computed disparities for these inliers. We denote by Ii
the set of inliers whose disparity values are computed as i. Similarly, I=i are the inliers
with the corresponding disparity values not equal to i. L is a metric measuring the color
difference between I(x) and its neighboring pixels Ii (Wx ) and I=i (Wx ). One example
is shown in Figure 1(a) where a window Wx is centered at an outlier pixel x. Within
Wx , inlier pixels are clustered into I1 and I=1 . ω1 (x; I) is computed according to the
color similarity between x and other pixels in the two clusters.
(6) is a function to assign an outlier pixel x a disparity value, constrained by the color
similarity between x and the clustered neighboring pixels. By and large, if the color
distance between x and its inlier neighbors with disparity i is small enough compared
to the color distance to other inliers, ωi (x; I) should have a large value, indicating high
chance to let dl = i in (5).
Now the problem is on how to compute a metric L that appropriately measures
the color distance between pixels. In our method, we abstract color sets Ii (Wx ) and
Wx R
I(x)
I =i(∗)
I1
I(x) G
I1 I i(∗)
B
(a) (b)
R R
I(x)
I(x) I =i(∗)
I =i(∗) I i(∗)
G G
I i(∗)
B B
(c ) (d)
Fig. 1. Computing disparity weight ω . (a) Within a neighborhood window Wx , inlier pixels are
clustered into I1 and I =1 . (b)-(d) illustrate the color projection. (b) The projection of I(x) on
vector I i(∗) − I =i(∗) is between two ends. (c-d) The projections of I(x) are out of range, thereby
are considered as extreme situations.
I=i (Wx ) by two representatives I i(∗) and I =i(∗) respectively. Then L is simplified
to a color metric between pixels. We adopt the color projection distance along vector
I i(∗) − I =i(∗) and define
L(I(x), c) = I(x) − c, I i(∗) − I =i(∗) , (7)
where ·, · denotes the inner product of two color vectors and c can be either I i(∗) or
I =i(∗) . We regard I i(∗) − I =i(∗) as a projection vector because it measures the absolute
difference between two representative colors, or, equivalently, the distance between sets
Ii (Wx ) and I=i (Wx ).
Projecting I(x) to vector I i(∗) − I =i(∗) also makes the assignment of two extreme
values 0 and 1 to ωi (x; I) easy. Taking Figure 1 as an example, if the projection of I(x)
on vector I i(∗) − I =i(∗) is between two ends, its value is obviously between 0 and 1, as
shown in Figure 1 (b). If the projection of I(x) is out of one end point, its value should
be 0 if it is close to I i(∗) or 1 otherwise (Figure 1 (c) and (d)). To handle the extreme
cases, we define the final ωi (x; I) as
⎧
⎨0 if I − I =i(∗) , I i(∗) − I =i(∗) < 0
ωi (x; I) = 1 if I i(∗) − I, I i(∗) − I =i(∗) < 0
⎩
ωi (x; I) Otherwise
which is further expressed as

(I − I =i(∗) )T (I i(∗) − I =i(∗) )
ωi = T , (8)
I i(∗) − I =i(∗) 22
where ⎧
⎨0 x < 0
T (x) = 1 x > 1 (9)
⎩
x otherwise
(I−I =i(∗) )T (I i(∗) −I =i(∗) )
Note that term I i(∗) −I =i(∗) 22
defined in (8) is quite similar to an alpha matte
model used in image matting [22,23] where the representative colors I i(∗) and I =i(∗)
are analogous to the unknown foreground and background colors. The image matting
problem is solved by color sample collection and optimization. In our problem, the
color samples are those clustered neighboring pixels Ii (Wx ) and I=i (Wx ).
With the above analysis, computing the weight ωi is naturally transformed to an
image matting problem where the representative color selection is handled by applying
an optimization algorithm. In our method, we employ the robust matting with optimal
color sample selection approach [23]. In principle, I i(∗) and I =i(∗) are respectively
selected from Ii (Wx ) and I=i (Wx ) based on a sample confidence measure combining
two criteria. First, either I i(∗) or I =i(∗) should be similar to the color of the outlier pixel
I, which makes weight ωi approach either 0 or 1 and the weight distribution hardly
uniform. Second, I is also expected to be a linear combination of I i(∗) and I =i(∗) .
This is useful for modeling color blending since outlier pixels have chance to be the
interpolation of color samples, especially for those on the region boundary.
Using the sample confidence definition, we get two weights and a neighborhood
term, similar to those in [23]. Then we apply the Random Walk method [24] to com-
pute weight ωi . This process is repeated for all ωi ’s, where i = 0, · · · , N . The main
benefit that we employ this matting method is that it provides an optimal way to select
representative colors while maintaining spatial smoothness.
3.2 Smoothness Term

Term Es (Dl , Dr ) contains two parts, representing intra-frame disparity smoothness and
inter-frame disparity consistency:
f3 (x, x , dl , dr ) f2 (x, x , dl )
Es (Dl , Dr ) = [ ( )+ ( )+
λ γ
x x ∈N1 (x) x ∈N2 (x)
f3 (x, x , dr , dl ) f2 (x, x , dr )
( )+ ( )], (10)
λ γ
x ∈N1 (x) x ∈N2 (x)
where N1 (x) represents the N possible corresponding pixels of x in the other view and
N2 (x) denotes the 4-neighborhood of x in the image space. f2 is defined as
f2 (x, x , di ) = min(|di (x) − di (x ))|, τ ), i ∈ {l, r}, (11)
where τ is a threshold set as 2. To define (11), we have also experimented with using
color weighted smoothness and observed that the results are not improved.
We define f3 (·) as the disparity correlations between two views:
f3 (x, x , dl , dr ) = min(|dl (x) − dr (x )|, ζ) and
f3 (x, x , dr , dl ) = min(|dr (x) − dl (x )|, ζ) , (12)
where ζ is a truncation threshold with value 1. We do not define a unique x correspond-

ing to x because x is unknown in the beginning. The other reason is that both f2 and
f3 are the costs for disparity smoothness. In f2 , all neighboring pixels are encoded in
N2 though di (x) is not necessarily similar to all di (x ). So we introduce f3 with the
similar thought for reducing the disparity noise in global optimization considering the
inter-frame consistency.
4 Implementation
The overview of our framework is given in Algorithm 1, which consists of an initial-
ization step and a global optimization step. In the first step, we initialize the disparity
maps by minimizing an energy with the simplified data and smoothness terms. Then we
compute the Outlier Confidence (OC) maps. In the second step, we globally refine the
disparities by incorporating the OC maps.
Algorithm 1. Overview of our approach

1. Initialization:
1.1 Initialize disparity map D by setting U = 0 for all pixels.
1.2 Estimate Outlier Confidence map U .
2. Global Optimization:
2.1 Compute data terms using the estimated outlier confidence maps.
2.2 Global optimization using BP.
4.1 Disparity Initialization

To initialize disparities, we simply set all values in Ul and Ur to zeros and optimize the
objective function combining (2) and (10):
f0 (x, dl ) + f0 (x, dr )
( ) + Es (Dl , Dr ). (13)
x
α
Because of introducing the inter-frame disparity consistency in (12), our Markov Ran-
dom Field (MRF) based on the defined energy is slightly different from the regular-grid
MRFs proposed in other stereo approaches [2,25]. In our two-frame configuration, the
MRF is built on two images with (4 + N ) neighboring sites for each node. N is the
total number of the disparity levels. One illustration is given in Figure 2 where a pixel
x in Il not only connects to its 4 neighbors in the image space, but also connects to all
possible corresponding pixels in Ir .
We minimize the energy defined in (13) using Belief Propagation. The inter-frame
consistency constraint makes the estimated disparity maps contain less noise in two
frames. We show in Figure 3(a) the initialized disparity result using the standard 4-
connected MRF without defining f3 in (10). (b) shows the result using our (4 + N )-
connected MRF. The background disparity noise is reduced.
Il Ir
Fig. 2. In our dual view configuration, x (marked with the cross) is not only connected to 4
neighbors in one image, but also related to N possible corresponding pixels in the other image.
The total number of neighbors of x is 4 + N .
(1) (2)
Depending on using f0 in (3) or f0 in (4) in the data term definition, we obtain
two sets of initializations using and without using global color segmentation. We shall
compare in the results how applying our OC models in the following global optimiza-
tion improves both of the disparity maps.
4.2 Outlier Confidence Estimation
We estimate the outlier confidence map U on the initial disparity maps. Our following
discussion focuses on estimating Ul on the left view. The right view can be handled in
a similar way. The outlier confidences, in our algorithm, are defined as
⎧
⎨1 |dl (x) − dr (x − dl (x))| ≥ 1
bx (d∗ )−bmin
Ul (x) = T ( bo −bmin ) bx (d∗ ) > t ∧ |dl (x) − dr (x − dl (x))| = 0 (14)
⎩
0 Otherwise
considering 2 cases.
Case 1: Our MRF enforces the disparity consistency between two views. After dis-
parity initialization, the remaining pixels with inconsistent disparities are likely to be
occlusions. So we first set the outlier confidence Ul (x) = 1 for pixel x if the inter-frame
consistency is violated, i.e., |dl (x) − dr (x − dl (x))| ≥ 1.
Case 2: Besides the disparity inconsistency, pixel matching with large matching cost
is also unreliable. In our method, since we use BP to initialize the disparity maps, the
matching cost is embedded in the output disparity belief bx (d) for each pixel x. Here,
we introduce some simple operations to manipulate it. First, we extract bx (d∗ ), i.e.,
the smallest belief, for each pixel x. If bx (d∗ ) < t, where t is a threshold, the pixel
should be regarded as an inlier given the small matching cost. Second, a variable bo is
computed as the average of the minimal beliefs regarding all occluded pixels detected
in Case 1, i.e., bo = Ul (x)=1 bx (d∗ )/K where K is the total number of the occluded
pixels. Finally, we compute bmin as the average of top n% minimal beliefs among all
pixels. n is set to 10 in our experiments.
Using the computed bx (d∗ ), bo , and bmin , we estimate Ul (C x) for pixels neither de-
tected as occlusions nor treated as inliers by setting

bxC (d∗ ) − bmin
x) = T
Ul (C , (15)
bo − bmin
(a) (b) (c) (d)
Fig. 3. Intermediate results for the “Tsukuba” example. (a) and (b) show our initial disparity maps
by the 4-connected and (4 + N )-connected MRFs respectively without using segmentation. The
disparity noise in (b) is reduced for the background. (c) Our estimated OC map. (d) A disparity
map constructed by combining the inlier and outlier information. The disparities for the outlier
pixels are set as the maximum weight ωi . The inlier pixels are with initially computed disparity
values.
where T is the function defined in (9), making the confidence value in range [0, 1]. (15)
indicates if the smallest belief bx (d∗ ) of pixel x is equal to or larger than the average
smallest belief of the occluded pixels detected in Case 1, the outlier confidence of x
will be high, and vice versa.
Figure 3(c) shows the estimated outlier coefficient map for the “tsukuba” example.
The pure black pixels represent inliers where Ul (x) = 0. Generally, the region consist-
ing of pixels with Ul (x) > 0 is wider than the ground truth occluded region. This is
allowed in our algorithm because Ul (x) is only a weight balancing pixel matching and
color smoothness. Even if pixel x is mistakenly labeled as an outlier, the disparity esti-
mation in our algorithm will not be largely influenced because large Ul (x) only makes
the disparity estimation of x rely more on neighboring pixel information, by which d(x)
still has a large chance to be correctly inferred.
To illustrate the efficacy of our OC scheme, we show in Figure 3(d) a disparity map
directly constructed with the following setting. Each inlier pixel is with initially com-
puted disparity value and each outlier pixel is with the disparity i corresponding to the
maximum weight ωi among all ωj ’s, where j = 0, · · · , N . It can be observed that
even without any further global optimization, this simple maximum-weight disparity
calculation already makes the object boundary smooth and natural.
4.3 Global Optimization
With the estimated OC maps, we are ready to use global optimization to compute the
final disparity maps combining costs (2) and (10) in (1). Two forms of f0 (·) ((3) and
(4)) are independently applied in our experiments for result comparison.
The computation of f1 (x, d; I) in (5) is based on the estimated OC maps and the
initial disparities for the inlier pixels, which are obtained in the aforementioned steps.
To compute ωi for outlier pixel x with Ul (x) > 0, robust matting [23] is performed
as described in Section 3.1 for each disparity level. The involved color sampling is
performed in each local window with size 60 × 60. Finally, the smoothness terms are
embedded in the message passing of BP. An acceleration using distance transform [25]
is adopted to construct the messages.
5 Experiments
In experiments, we compare the results using and without using the Outlier Confidence
maps. The performance is evaluated using the Middlebury dataset [10]. All parameters
used in implementation are listed in Table 1 where α, β and κ are the weights defined
in the data term. γ and λ are for intra-frame smoothness and inter-frame consistency
respectively. ϕ, τ , and ζ are the truncation thresholds for different energy terms. t is
the threshold for selecting possible outliers. As we normalize the messages after each
message passing iteration by subtracting the mean of the messages, the belief bmin is
negative, making t = 0.9bmin > bmin .
A comparison of the state-of-the-art stereo matching algorithms is shown in
Table 2 extracted from the Middlebury website [9]. In the following, we give detailed
explanations.
Table 1. The parameter values used in our experiments. N is the number of the disparity levels.
c is the average of the correlation volume. bmin is introduced in (15).
Parameters α β κ γ λ ϕ τ ζ t
value ϕ 0.8 0.3 5.0 5N c 2.0 1.0 0.9bmin
Table 2. Algorithm evaluation on the Midellbury data set. Our method achieves overall rank 2 at
the time of data submission.
Avg. Tsukuba Venus Teddy Cones

Algorithm Rank nonocc all disc nonocc all disc nonocc all disc nonocc all disc
Adap.BP [5] 2.3 1.11 1.37 5.79 0.10 0.21 1.44 4.22 7.06 11.8 2.48 7.92 7.32
Our method 3.6 0.88 1.43 4.74 0.18 0.26 2.40 5.01 9.12 12.8 2.78 8.57 6.99
DoubleBP [6] 3.7 0.88 1.29 4.76 0.14 0.60 2.00 3.55 8.71 9.70 2.90 9.24 7.80
SPDou.BP [19] 4.6 1.24 1.76 5.98 0.12 0.46 1.74 3.45 8.38 10.0 2.93 8.73 7.91
SymBP+occ [4] 8.8 0.97 1.75 5.09 0.16 0.33 2.19 6.47 10.7 17.0 4.79 10.7 10.9
Table 3. Result comparison on the Middlebury dataset using (1st and 3rd rows) and without using
(2nd and 4th rows) OC Maps. The segmentation information has been incorporated for the last
two rows.
Overall Tsukuba Venus Teddy Cones

Algorithm Rank nonocc all disc nonocc all disc nonocc all disc nonocc all disc
COLOR 16 1.12 3.29 5.92 0.49 1.48 6.78 10.5 16.9 21.1 3.42 12.1 8.26
COLOR+OC 5 0.83 1.41 4.45 0.25 0.31 3.22 10.1 14.6 19.9 3.22 9.82 7.40
SEG 4 0.97 1.75 5.23 0.30 0.70 3.98 5.56 9.99 13.6 3.04 8.90 7.60
SEG+OC 2 0.88 1.43 4.74 0.18 0.26 2.40 5.01 9.12 12.8 2.78 8.57 6.99
5.1 Results without Using Segmentation

In the first part of our experiments, we do not use the segmentation information. So data
(1)
term f0 defined in (3) is used in our depth estimation.
(a)
(b)
Fig. 4. Disparity result comparison. (a) Disparity results of “SEG” (b) Our final disparity results
using the Outlier Confidence model (“SEG+OC”).
We show in the first row of Table 3 (denoted as “COLOR”) the statistics of the
initial disparities. The algorithm is detailed in Section 4.1. We set U (x) = 0 for all x’s
and minimize the energy defined in (13). Then we estimate the OC maps based on the
initial disparities and minimize the energy defined in (1). We denote the final results as
“COLOR+OC” in the second row of Table 3.
Comparing the two sets of results, one can observe that incorporating the outlier in-
formation significantly improves the quality of the estimated disparity maps. The over-
all rank jumps from initial No. 16 to No. 5, which is the highest position for all results
produced by the stereo matching algorithms without incorporating segmentation.
In analysis, for the “Teddy” example, however, our final disparity estimate does not
gain large improvement over the initial one. It is because that the remaining errors are
mostly caused by matching large textureless regions, which can be addressed by color
segmentation.
5.2 Results Using Segmentation
In this part of the experiments, we incorporate the segmentation information by using

(2)
the data term f0 defined in (4). Our initial disparities are denoted as “SEG”. Our final
results obtained by applying the global optimization incorporating the Outlier Confi-
dences are denoted as “SEG+OC”. We show in the third and forth rows of Table 3 the
error statistics of the initial disparity maps and our refined results. The average rank
rises from 6.9 to 3.6 and the overall rank jumps from No. 4 to No. 2. The improve-
ment validates the effectiveness of our approach in handling outliers and its nature of
complementarity to color segmentation.
The computed disparity maps are shown in Figure 4, where (a) and (b) respectively
show the results of “SEG” and “SEG+OC”. A comparison of disparity errors is demon-
strated in Figure 5 using the “Cones” example. The magnified patches extracted from
the error maps are shown in (b). The comparison shows that our approach can primarily
improve the disparity estimation for outlier pixels.
(a)
(b)
Fig. 5. Error comparison on the “Cones” example. (a) shows the disparity error maps for “SEG”
and “SEG+OC” respectively. (b) Comparison of three magnified patches extracted from (a). The
“SEG+OC” results are shown on the right of each patch pair.
Finally, the framework of our algorithm is general. Many other existing stereo match-
ing methods can be incorporated into the outlier confidence scheme by changing f0 to
other energy functions.
6 Conclusion
In this paper, we have proposed an Outlier-Confidence-based stereo matching algo-
rithm. In this algorithm, the Outlier Confidence is introduced to measure how likely
that one pixel is an outlier. A model using the local color information is proposed for
inferring the disparities of possible outliers and is softly combined with other data terms
to dynamically adjust the disparity estimate. Complementary to global color segmenta-
tion, our algorithm locally gathers color samples and optimizes them using the matting
techniques in order to reliably measure how one outlier pixel can be assigned a disparity
value. Experimental results on the Middlebury data set show that our proposed method
is rather effective in disparity estimation.
Acknowledgements
This work was fully supported by a grant from the Research Grants Council of Hong
Kong (Project No. 412708) and is affiliated with the Microsoft–CUHK Joint Laboratory.
References
1. Tao, H., Sawhney, H.S., Kumar, R.: A global matching framework for stereo computation.
In: ICCV, pp. 532–539 (2001)
2. Sun, J., Zheng, N.N., Shum, H.Y.: Stereo matching using belief propagation. IEEE Trans.
Pattern Anal. Mach. Intell. 25(7), 787–800 (2003)
3. Hong, L., Chen, G.: Segment-based stereo matching using graph cuts. In: CVPR (1), pp.
74–81 (2004)
4. Sun, J., Li, Y., Kang, S.B.: Symmetric stereo matching for occlusion handling. In: CVPR (2),
pp. 399–406 (2005)
5. Klaus, A., Sormann, M., Karner, K.F.: Segment-based stereo matching using belief propaga-
tion and a self-adapting dissimilarity measure. In: ICPR (3), pp. 15–18 (2006)
6. Yang, Q., Wang, L., Yang, R., Stewénius, H., Nistér, D.: Stereo matching with color-weighted
correlation, hierarchical belief propagation and occlusion handling. In: CVPR (2), pp. 2347–
2354 (2006)
7. Kang, S.B., Szeliski, R.: Extracting view-dependent depth maps from a collection of images.
8. Strecha, C., Fransens, R., Van Gool, L.J.: Combined depth and outlier estimation in multi-
view stereo. In: CVPR (2), pp. 2394–2401 (2006)
9. Scharstein, D., Szeliski, R.: http://vision.middlebury.edu/stereo/eval/
10. Scharstein, D., Szeliski, R.: A taxonomy and evaluation of dense two-frame stereo corre-
spondence algorithms. International Journal of Computer Vision 47(1-3), 7–42 (2002)
11. Zhang, L., Seitz, S.M.: Parameter estimation for mrf stereo. In: CVPR (2), pp. 288–295
(2005)
12. Zitnick, C.L., Kang, S.B., Uyttendaele, M., Winder, S.A.J., Szeliski, R.: High-quality video
view interpolation using a layered representation. ACM Trans. Graph. 23(3), 600–608 (2004)
13. Yoon, K.J., Kweon, I.S.: Stereo matching with the distinctive similarity measure. In: ICCV
(2007)
14. Yoon, K.J., Kweon, I.S.: Adaptive support-weight approach for correspondence search. IEEE
Trans. Pattern Anal. Mach. Intell. 28(4), 650–656 (2006)
15. Lei, C., Selzer, J.M., Yang, Y.H.: Region-tree based stereo using dynamic programming op-
timization. In: CVPR (2), pp. 2378–2385 (2006)
16. Strecha, C., Fransens, R., Gool, L.J.V.: Wide-baseline stereo from multiple views: A proba-
bilistic account. In: CVPR (1), pp. 552–559 (2004)
17. Hirschmüller, H.: Accurate and efficient stereo processing by semi-global matching and mu-
tual information. In: CVPR (2), pp. 807–814 (2005)
18. Hirschmüller, H., Scharstein, D.: Evaluation of cost functions for stereo matching. In: CVPR
(2007)
19. Yang, Q., Yang, R., Davis, J., Nistér, D.: Spatial-depth super resolution for range images. In:
CVPR (2007)
20. Hasinoff, S.W., Kang, S.B., Szeliski, R.: Boundary matting for view synthesis. Computer
Vision and Image Understanding 103(1), 22–32 (2006)
21. Comaniciu, D., Meer, P.: Mean shift: A robust approach toward feature space analysis. IEEE
Trans. Pattern Anal. Mach. Intell. 24(5), 603–619 (2002)
22. Chuang, Y.Y., Curless, B., Salesin, D., Szeliski, R.: A bayesian approach to digital matting.
In: CVPR (2), pp. 264–271 (2001)
23. Wang, J., Cohen, M.F.: Optimized color sampling for robust matting. In: CVPR (2007)
24. Grady, L.: Random walks for image segmentation. IEEE Trans. Pattern Anal. Mach. In-
tell. 28(11), 1768–1783 (2006)
25. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient belief propagation for early vision. In:
CVPR (1), pp. 261–268 (2004)
Improving Shape Retrieval by Learning Graph
Transduction
Xingwei Yang1 , Xiang Bai2,3 , Longin Jan Latecki1 , and Zhuowen Tu3
1
Dept. of Computer and Information Sciences, Temple University, Philadelphia
{xingwei,latecki}@temple.edu
2
Dept. of Electronics and Information Engineering, Huazhong University of Science
and Technology, P.R. China
xiang.bai@gmail.com
3
Lab of Neuro Imaging, University of California, Los Angeles
zhuowen.tu@loni.ucla.edu
Abstract. Shape retrieval/matching is a very important topic in com-

puter vision. The recent progress in this domain has been mostly driven
by designing smart features for providing better similarity measure be-
tween pairs of shapes. In this paper, we provide a new perspective to
this problem by considering the existing shapes as a group, and study
their similarity measures to the query shape in a graph structure. Our
method is general and can be built on top of any existing shape match-
ing algorithms. It learns a better metric through graph transduction by
propagating the model through existing shapes, in a way similar to com-
puting geodesics in shape manifold. However, the proposed method does
not require learning the shape manifold explicitly and it does not require
knowing any class labels of existing shapes. The presented experimen-
tal results demonstrate that the proposed approach yields significant
improvements over the state-of-art shape matching algorithms. We ob-
tained a retrieval rate of 91% on the MPEG-7 data set, which is the
highest ever reported in the literature.
1 Introduction
Shape matching/retrieval is a very critical problem in computer vision. There

are many different kinds of shape matching methods, and the progress in in-
creasing the matching rate has been substantial in recent years. However, all of
these approaches are focused on the nature of shape similarity. It seems to be
an obvious statement that the more similar two shapes are, the smaller is their
difference, which is measured by some distance function. Yet, this statement
ignores the fact that some differences are relevant while other differences are ir-
relevant for shape similarity. It is not yet clear how the biological vision systems
perform shape matching; it is clear that shape matching involves the high-level
understanding of shapes. In particular, shapes in the same class can differ signif-
icantly because of distortion or non-rigid transformation. In other words, even
if two shapes belong to the same class, the distance between them may be very

Improving Shape Retrieval by Learning Graph Transduction 789
Fig. 1. Existing shape similarity methods incorrectly rank shape (b) as more similar
to (a) than (c)
Fig. 2. A key idea of the proposed distance learning is to replace the original shape
distance between (a) and (e) with a geodesic path in the manifold of know shapes,
which is the path (a)-(e) in this figure
large if the distance measure cannot capture the intrinsic property of the shape.
It appears to us that all published shape distance measures [1,2,3,4,5,6,7] are
unable to address this issue. For example, based on the inner distance shape
context (IDSC) [3], the shape in Fig. 1(a) is more similar to (b) than to (c),
but it is obvious that shape (a) and (c) belong to the same class. This incorrect
result is due to the fact that the inner distance is unaware that the missing tail
and one front leg are irrelevant for this shape similarity judgment. On the other
hand, much smaller shape details like the dog’s ear and the shape of the head
are of high relevance here. No matter how good a shape matching algorithm is,
the problem of relevant and irrelevant shape differences must be addressed if we
want to obtain human-like performance. This requires having a model to capture
the essence of a shape class instead of viewing each shape as a set of points or a
parameterized function.
In this paper, we propose to use a graph-based transductive learning algo-
rithm to tackle this problem, and it has the following properties: (1) Instead
of focusing on computing the distance (similarity) for a pair of shapes, we take
advantage of the manifold formed by the existing shapes. (2) However, we do not
explicitly learn the manifold nor compute the geodesics [8], which are time con-
suming to calculate. A better metric is learned by collectively propagating the
similarity measures to the query shape and between the existing shapes through
graph transduction. (3) Unlike the label propagation [9] approach, which is semi-
supervised, we treat shape retrieval as an unsupervised problem and do not re-
quire knowing any shape labels. (4) We can build our algorithm on top of any
existing shape matching algorithm and a significant gain in retrieval rates can
be observed on well-known shape datasets.
Given a database of shapes, a query shape, and a shape distance function,
which does not need to be a metric, we learn a new distance function that is
790 X. Yang et al.
expressed by shortest paths on the manifold formed by the know shapes and the
query shape. We can do this without explicitly learning this manifold. As we
will demonstrate in our experimental results, the new learned distance function
is able to incorporate the knowledge of relevant and irrelevant shape differences.
It is learned in an unsupervised setting in the context of known shapes. For
example, if the database of known shapes contains shapes (a)-(e) in Fig. 2, then
the new learned distance function will rank correctly the shape in Fig. 1(a) as
more similar to (c) than to (b). The reason is that the new distance function
will replace the original distance (a) to (c) in Fig.1 with a distance induced by
the shortest path between in (a) and (e) in Fig.2.
In more general terms, even if the difference between shape A and shape C
is large, but there is a shape B which has small difference to both of them, we
still claim that shape A and shape C are similar to each other. This situation is
possible for most shape distances, since they do not obey the triangle inequality,
i.e., it is not true that d(A, C) ≤ d(A, B) + d(B, C) for all shapes A, B, C [10].
We propose a learning method to modify the original shape distance d(A, C).
If we have the situation that d(A, C) > d(A, B) + d(B, C) for some shapes
A, B, C, then the proposed method is able to learn a new distance d (A, C) such
that d (A, C) ≤ d(A, B) + d(B, C). Further, if there is a path in the distance
space such that d(A, C) > d(A, B1 ) + . . . + d(Bk , C), then our method learns
a new d (A, C) such that d (A, C) ≤ d(A, B1 ) + . . . + d(Bk , C). Since this path
represents a minimal distortion morphing of shape A to shape C, we are able to
ignore irrelevant shape differences, and consequently, we can focus on relevant
shape differences with the new distance d .
Our experimental results clearly demonstrate that the proposed method can
improve the retrieval results of the existing shape matching methods. We ob-
tained the retrieval rate of 91% on part B of the MPEG-7 Core Experiment
CE-Shape-1 data set [11], which is the highest ever bull’s eye score reported in
the literature. As the input to our method we used the IDSC, which has the
retrieval rate of 85.40% on the MPEG-7 data set [3]. Fig. 3 illustrates the ben-
efits of the proposed distance learning method. The first row shows the query
shape followed by the first 10 shapes retrieved using IDSC only. Only two flies
are retrieved among the first 10 shapes. The results of the learned distance for
the same query are shown in the second row. All of the top 10 retrieval results
Fig. 3. The first column shows the query shape. The remaining 10 columns show the
most similar shapes retrieved from the MPEG-7 data set. The first row shows the
results of IDSC [3]. The second row shows the results of the proposed learned distance.
are correct. The proposed method was able to learn that the shape differences
in the number of fly legs and their shapes are irrelevant. The remainder of this
paper is organized as follows. In Section 2, we briefly review some well-known
shape matching methods and the semi-supervised learning algorithms. Section 3
describes the proposed approach to learning shape distances. Section 4 relates
the proposed approach to the class of machine learning approaches called label
propagation. The problem of the construction of the affinity matrix is addressed
in Section 5. Section 6 gives the experimental results to show the advantage of
the proposed approach. Conclusion and discussion are given in Section 7.
2 Related Work
The semi-supervised learning problem has attracted an increasing amount of in-

terest recently, and several novel approaches have been proposed. The existing
approaches could be divided into several types, multiview learning [12], gener-
ative model [13], Transductive Support Vector Machine (TSVM) [14]. Recently
there have been some promising graph based transductive learning approaches
proposed, such as label propagation [9], Gaussian fields and harmonic functions
(GFHF) [15], local and global consistency (LGC) [16], and the Linear Neigh-
borhood Propagation (LNP) [17]. Zhou et al. [18] modified the LGC for the
information retrieval. The semi-supervised learning problem is related to mani-
fold learning approaches, e.g., [19].
The proposed method is inspired by the label propagation. The reason we
choose the framework of label propagation is it allows the clamping of labels.
Since the query shape is the only labeled shape in the retrieval process, the label
propagation allows us to enforce its label during each iteration, which naturally
fits in the framework of shape retrieval. Usually, GFHF is used instead of label
propagation, as both methods can achieve the same results[9]. However, in the
shape retrieval, we can use only the label propagation, the reason is explained
in detail in Section 4.
Since a large number of shape similarity methods have been proposed in the
literature, we focus our attention on methods that reported retrieval results
on the MPEG-7 shape data set (part B of the MPEG-7 Core Experiment CE-
Shape-1). This allows us to clearly demonstrate the retrieval rate improvements
obtained by the proposed method. Belongie et al. [1] introduced a novel lo-
cal representation of shapes called shape context. Ling and Jacobs [3] modified
the shape context by considering the geodesic distance of contour instead of
the Euclidean distance, which improved the classification of articulated shapes.
Latecki and Lakaemper [4] used visual parts for shape matching. In order to
avoid problems associated with purely global or local methods, Felzenszwalb
and Schwartz [5] also described a hierarchical matching method. Other hierar-
chical methods include the hierarchical graphical models in [20] and hierarchical
procrustes matching [6].
There is a significant body of work on distance learning [21]. Xing et al.
[22] propose estimating the matrix W of a Mahalanobis distance by solving a
792 X. Yang et al.
convex optimization problem. Bar-Hillel et al. [23] also use a weight matrix W to
estimate the distance by relevant component analysis (RCA). Athitsos et al. [24]
proposed a method called BoostMap to estimate a distance that approximates a
certain distance. Hertz’s work [25] uses AdaBoost to estimate a distance function
in a product space, whereas the weak classifier minimizes an error in the original
feature space. All these methods’ focus is a selection of suitable distance from
a given set of distance measures. Our method aims at improving the retrieval
performance of a given distance measure.
3 Learning New Distance Measures

We first describe the classical setting of similarity retrieval. It applies to many
retrieval scenarios like image, document, key word, and shape retrieval. Given is
a set of objects X = {x1 , . . . , xn } and a similarity function sim: X × X → R+
that assigns a similarity value (a positive integer) to each pair of objects.
We assume that x1 is a query object(e.g., a query shape), {x2 , . . . , xn } is a
set of known database objects (or a training set). Then by sorting the values
sim(x1 , xi ) in decreasing order for i = 2, . . . , n we obtain a ranking of database
objects according to their similarity to the query, i.e., the most similar database
object has the highest value and is listed first. Sometimes a distance measure is
used in place of the similarity measure, in which case the ranking is obtained
by sorting the database objects in the increasing order, i.e., the object with the
smallest value is listed first. Usually, the first N ) n objects are returned as the
most similar to the query x1 .
As discussed above, the problem is that the similarity function sim is not
perfect so that for many pairs of objects it returns wrong results, although it
may return correct scores for most pairs. We introduce now a method to learn
a new similarity function simT that drastically improves the retrieval results of
sim for the given query x1 .
Let wi,j = sim(xi , xj ), for i, j = 1, . . . , n, be a similarity matrix, which is also
called an affinity matrix. We define a sequence of labeling functions ft : X →
[0, 1] with f0 (x1 ) = 1 and f0 (xi ) = 0 for i = 2, . . . , n. We use the following
recursive update of function ft :
n
j=1 wij ft (xj )
ft+1 (xi ) = n (1)
j=1 wij
for i = 2, . . . , n and we set

ft+1 (x1 ) = 1. (2)
We have only one class that contains only one labeled element being the query
x1 . We define a sequence of new learned similarity functions restricted to x1 as
simt (x1 , xi ) = ft (xi ). (3)
Thus, we interpret ft as a set of normalized similarity values to the query x1 .

Observe that sim1 (x1 , xi ) = w1,i = sim(x1 , xi ).
We iterate steps (1) and (2) until the step t = T for which the change is
below a small threshold. We then rank the similarity to the query x1 with simT .
Our experimental results in Section 6 demonstrate that the replacement of the
original similarity measure sim with simT results in a significant increase in the
retrieval rate.
The steps (1) and (2) are used in label propagation, which is described in
Section 4. However, our goal and our setting are different. Although label prop-
agation is an instance of semi-supervised learning, we stress that we remain in
the unsupervised learning setting. In particular, we deal with the case of only
one known class, which is the class of the query object. This means, in particular,
that label propagation has a trivial solution in our case limt→∞ ft (xi ) = 1 for all
i = 1, . . . , n, i.e., all objects will be assigned the class label of the query shape.
Since our goal is ranking of the database objects according to their similarity to
the query, we stop the computation after a suitable number of iterations t = T .
As is the usual practice with iterative processes that are guaranteed to converge,
the computation is halted if the difference ||ft+1 − ft || becomes very slow, see
Section 6 for details.
If the database of known objects is large, the computation with all n objects
may become impractical. Therefore, in practice, we construct the matrix w using
only the first M < n most similar objects to the query x1 sorted according to
the original distance function sim.
4 Relation to Label Propagation

Label propagation is formulated as a form of propagation on a graph, where
node’s label propagates to neighboring nodes according to their proximity. In
our approach we only have one labeled node, which is the query shape. The key
idea is that its label propagates ”faster” along a geodesic path on the manifold
spanned by the set of known shapes than by direct connections. While following
a geodesic path, the obtained new similarity measure learns to ignore irrelevant
shape differences. Therefore, when learning is complete, it is able to focus on
relevant shape differences. We review now the key steps of label propagation
and relate them to the proposed method introduced in Section 3.
Let {(x1 , y1 ) . . . (xl , yl )} be the labeled data, y ∈ {1 . . . C}, and {xl+1 . . . xl+u }
the unlabeled data, usually l ) u. Let n = l + u. We will often use L and U to
denote labeled and unlabeled data respectively. The Label propagation supposes
the number of classes C is known, and all classes are present in the labeled
data[9]. A graph is created where the nodes are all the data points, the edge
between nodes i, j represents their similarity wi,j . Larger edge weights allow
labels to travel through more easily. We define a n × n probabilistic transition
matrix P as a row-wise normalized matrix w.
wij
Pij = n (4)
k=1 wik
where Pij is the probability of transit from node i to node j. Also define a l × C
label matrix YL , whose ith row is an indicator vector for yi , i ∈ L: Yic = δ(yi,c ).
794 X. Yang et al.
The label propagation computes soft labels f for nodes, where f is a n×C matrix
whose rows can be interpreted as the probability distributions over labels. The
initialization of f is not important. The label propagation algorithm is as follows:
1. Initially, set f (xi ) = yi for i = 1, . . . , l and f (xj ) arbitrarily (e.g., 0) for
xj ∈ Xu n
wij f (xj )
2. Repeat until convergence: Set f (xi ) = n
j=1
wij , ∀xi ∈ Xu and set
j=1
f (xi ) = yi for i = 1, . . . , l (the labeled objects should be fixed).
In step 1, all nodes propagate their labels to their neighbors for one step. Step 2 is
critical, since it ensures persistent label sources from labeled data. Hence instead
of letting the initial labels fade way, we fix the labeled data. This constant push
from labeled nodes, helps to push the class boundaries through high density
regions so that they can settle in low density gaps. If this structure of data fits
the classification goal, then the algorithm can use unlabeled data to improve
learning.
f
Let f = ( L ). Since fL is fixed to YL , we are solely interested in fU . The
fU
matrix P is split into labeled and unlabeled sub-matrices
! "
PLL PLU
P = (5)
PUL PUU
As proven in [9] the label propagation converges, and the solution can be com-
puted in closed form using matrix algebra:
fU = (I − PUU )−1 PUL YL (6)
However, as the label propagation requires all classes be present in the labeled
data, it is not suitable for shape retrieval. As mentioned in Section 3, for shape
retrieval, the query shape is considered as the only labeled data and all other
shapes are the unlabeled data. Moreover, the graph among all of the shapes is
fully connected, which means the label could be propagated on the whole graph.
If we iterate the label propagation infinite times, all of the data will have the
same label, which is not our goal. Therefore, we stop the computation after a
suitable number of iterations t = T .
5 The Affinity Matrix

In this section, we address the problem of the construction of the affinity matrix
W . There are some methods that address this issue, such as local scaling [26],
local liner approximation [17], and adaptive kernel size selection [27].
However, in the case of shape similarity retrieval, a distance function is usually
defined, e.g., [1,3,4,5]. Let D = (Dij ) be a distance matrix computed by some shape
distance function. Our goal is to convert it to a similarity measure in order to con-
struct an affinity matrix W . Usually, this can be done by using a Gaussian kernel:
Dij
2
wij = exp(− ) (7)
σij
2
Previous research has shown that the propagation results highly depend on the
kernel size σij selection [17]. In [15], a method to learn the proper σij for the kernel
is introduced, which has excellent performance. However, it is not learnable in
the case of few labeled data. In shape retrieval, since only the query shape has
the label, the learning of σij is not applicable. In our experiment, we use use an
adaptive kernel size based on the mean distance to K-nearest neighborhoods [28]:
σij = C · mean({knnd(xi ), knnd(xj )}) (8)
where mean({knnd(xi ), knnd(xj )}) represents the mean distance of the K-nearest
neighbor distance of the sample xi , xj and C is an extra parameter. Both K and
C are determined empirically.
In this section, we show that the proposed approach can significantly improve
retrieval rates of existing shape similarity methods.
6.1 Improving Inner Distance Shape Context

The IDSC [3] significantly improved the performance of shape context [1] by
replacing the Euclidean distance with shortest paths inside the shapes, and ob-
tained the retrieval rate of 85.40% on the MPEG-7 data set. The proposed
distance learning method is able to improve the IDSC retrieval rate to 91.00%.
For reference, Table 1 lists some of the reported results on the MPEG-7 data
set. The MPEG-7 data set consists of 1400 silhouette images grouped into 70
classes. Each class has 20 different shapes. The retrieval rate is measured by
the so-called bull’s eye score. Every shape in the database is compared to all
other shapes, and the number of shapes from the same class among the 40 most
similar shapes is reported. The bull’s eye retrieval rate is the ratio of the total
number of shapes from the same class to the highest possible number (which is
20 × 1400). Thus, the best possible rate is 100%.
In order to visualize the gain in retrieval rates by our method as compared to
IDSC, we plot the percentage of correct results among the first k most similar
shapes in Fig. 4(a), i.e., we plot the percentage of the shapes from the same class
among the first k-nearest neighbors for k = 1, . . . , 40. Recall that each class has
20 shapes, which is why the curve increases for k > 20. We observe that the
proposed method not only increases the bull’s eye score, but also the ranking of
the shapes for all k = 1, . . . , 40.
We use the following parameters to construct the affinity matrix: C = 0.25
and the neighborhood size is K = 10. As stated in Section 3, in order to increase
computational efficiency, it is possible to construct the affinity matrix for only
part of the database of known shapes. Hence, for each query shape, we first
retrieve 300 the most similar shapes, and construct the affinity matrix W for
only those shapes, i.e., W is of size 300 × 300 as opposed to a 1400 × 1400 matrix
if we consider all MPEG-7 shapes. Then we calculate the new similarity measure
796 X. Yang et al.
1 1 1
0.9 0.9 0.9
0.8 0.8 0.8

percentage of correct results

0.7 0.7 0.7
0.6 0.6 0.6
0.5 0.5 0.5
0.4 0.4 0.4
0.3 0.3 0.3
0.2 0.2 0.2
0.1 0.1 0.1
0 0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
number of most similar shapes number of most similar shape number of most similar shape
(a) (b) (c)
Fig. 4. (a) A comparison of retrieval rates between IDSC [3] (blue circles) and the
proposed method (red stars) for MPEG-7. (b) A comparison of retrieval rates between
visual parts in [4] (blue circles) and the proposed method (red stars) for MPEG-7. (c)
Retrieval accuracy of DTW (blue circles) and the proposed method (red stars) for the
Face (all) dataset.
Table 1. Retrieval rates (bull’s eye) of different methods on the MPEG-7 data set
Alg. CSS Vis. Parts SC IDSC Hierarchical Shape Tree IDSC+DP

+TPS +DP Procrustes + our
[29] [4] [1] [3] [6] [5] method
Score 75.44% 76.45% 76.51% 85.40% 86.35% 87.70% 91.00%
simT for only those 300 shapes. Here we assume that all relevant shapes will be
among the 300 most similar shapes. Thus, by using a larger affinity matrix we
can improve the retrieval rate but at the cost of computational efficiency.
In addition to the statistics presented in Fig. 4, Fig. 5 illustrates also that
the proposed approach improves the performance of IDSC. A very interesting
case is shown in the first row, where for IDSC only one result is correct for the
query octopus. It instead retrieves nine apples as the most similar shapes. Since
the query shape of the octopus is occluded, IDSC ranks it as more similar to an
apple than to the octopus. In addition, since IDSC is invariant to rotation, it
confuses the tentacles with the apple stem. Even in the case of only one correct
shape, the proposed method learns that the difference between the apple stem is
relevant, although the tentacles of the octopuses exhibit a significant variation
in shape. We restate that this is possible because the new learned distances are
induced by geodesic paths in the shape manifold spanned by the known shapes.
Consequently, the learned distances retrieve nine correct shapes. The only wrong
results is the elephant, where the nose and legs are similar to the tentacles of
the octopus.
As shown in the third row, six of the top ten IDSC retrieval results of lizard are
wrong. since IDSC cannot ignore the irrelevant differences between lizards and
sea snakes. All retrieval results are correct for the new learned distances, since the
proposed method is able to learn the irrelevant differences between lizards and
the relevant differences between lizards and sea snakes. For the results of deer
(fifth row), three of the top ten retrieval results of IDSC are horses. Compared
Fig. 5. The first column shows the query shape. The remaining 10 columns show the
most similar shapes retrieved by IDSC (odd row numbers) and by our method (even
row numbers).
to it, the proposed method (sixth row) eliminates all of the wrong results so that
only deers are in the top ten results. It appears to us that our new method learned
to ignore the irrelevant small shape details of the antlers. Therefore, the presence
of the antlers became a relevant shape feature here. The situation is similar for
the bird and hat, with three and four wrong retrieval results respectively for
IDSC, which are eliminated by the proposed method.
An additional explanation of the learning mechanism of the proposed method
is provided by examining the count of the number of violations of the triangle
inequality that involve the query shape and the database shapes. In Fig. 6(a),
the curve shows the number of triangle inequality violations after each iteration
of our distance learning algorithm. The number of violations is reduced signif-
icantly after the first few hundred iterations. We cannot expect the number of
violations to be reduced to zero, since cognitively motivated shape similarity may
sometimes require triangle inequality violations [10]. Observe that the curve in
Fig. 6(a) correlates with the plot of differences ||ft+1 − ft || as a function of t
shown in (b). In particular, both curves decrease very slow after about 1000
798 X. Yang et al.
3000 0.7
0.6
2500
0.5
2000
0.4
1500
0.3
1000
0.2
500
0.1
0 0
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000
(a) (b)
Fig. 6. (a) The number of triangle inequality violations per iteration. (b) Plot of dif-
ferences ||ft+1 − ft || as a function of t.
Table 2. Retrieval results on Kimia Data Set [30]
Algorithm 1st 2nd 3rd 4th 5th 6th 7th 8th 9th 10th
SC [30] 97 91 88 85 84 77 75 66 56 37
Shock Edit [30] 99 99 99 98 98 97 96 95 93 82
IDSC+DP [3] 99 99 99 98 98 97 97 98 94 79
Shape Tree [5] 99 99 99 99 99 99 99 97 93 86
our method 99 99 99 99 99 99 99 99 97 99
iterations, and at 5000 iterations they are nearly constant. Therefore, we se-
lected T = 5000 as our stop condition. Since the situation is very similar in all
our experiments, we always stop after T = 5000 iterations.
Besides MPEG-7, We also present experimental results on the Kimia Data
Set [30]. The database contains 99 shapes grouped into nine classes. As the
database only contains 99 shapes, we calculate the affinity matrix based on all
of the shape in the database. The parameters used to calculate the affinity matrix
are: C = 0.25 and the neighborhood size is K = 4. We changed the neighborhood
size, since the data set is much smaller than the MPEG-7 data set. The retrieval
results are summarized as the number of shapes from the same class among the
first top 1 to 10 shapes (the best possible result for each of them is 99). Table 2
lists the numbers of correct matches of several methods. Again we observe that
our approach could improve IDSC significantly, and it yields a nearly perfect
retrieval rate.
6.2 Improving Visual Part Shape Matching
Besides the inner distance shape context [3], we also demonstrate that the pro-
posed approach can improve the performance of visual parts shape similarity [4].
We select this method since it is based on very different approach than IDSC.
In [4], in order to compute the similarity between shapes, first the best possible
correspondence of visual parts is established (without explicitly computing the
visual parts). Then, the similarity between corresponding parts is calculated and
aggregated. The settings and parameters of our experiment are the same as for
IDSC as reported in the previous section except we set C = 0.4. The accuracy
of this method has been increased from 76.45% to 86.69% on the MPEG-7 data
set, which is more than 10%. This makes the improved visual part method one
of the top scoring methods in Table 1. A detailed comparison of the retrieval
accuracy is given in Fig. 4(b).
6.3 Improving Face Retrieval

We used a face data set from [31], where it is called Face (all). It addresses a
face recognition problem based on the shape of head profiles. It contains several
head profiles extracted from side view photos of 14 subjects. There exist large
variations in the shape of the face profile of each subject, which is the main reason
why we select this data set. Each subject is making different face expressions,
e.g., talking, yawning, smiling, frowning, laughing, etc. When the pictures of
subjects were taken, they were also encouraged to look a little to the left or
right, randomly. At least two subjects had glasses that they put on for half of
their samples.
The head profiles are converted to sequences of curvature values, and nor-
malized to the length of 131 points, starting from the neck area. The data set
has two parts, training with 560 profiles and testing with 1690 profiles. The
training set contains 40 profiles for each of the 14 classes. As reported on [31],
we calculated the retrieval accuracy by matching the 1690 test shapes to the
560 training shapes. We used a dynamic time warping (DTW) algorithm with
warping window [32] to generate the distance matrix, and obtained the 1NN
retrieval accuracy of 88.9% By applying our distance learning method we in-
creased the 1NN retrieval accuracy to 95.04%. The best reported result on [31]
has the first nearest neighbor (1NN) retrieval accuracy of 80.8%. The retrieval
rate, which represents the percentage of the shapes from the same class (profiles
of the same subject) among the first k-nearest neighbors, is shown in Fig. 4(c).
The accuracy of the proposed approach is stable, although the accuracy of DTW
decreases significantly when k increases. In particular, our retrieval rate for k=40
remains high, 88.20%, while the DTW rate dropped to 60.18%. Thus, the learned
distance allowed us to increase the retrieval rate by nearly 30%. Similar to the
above experiments, the parameters for the affinity matrix is C = 0.4 and K = 5.

In this work, we adapted a graph transductive learning framework to learn new
distances with the application to shape retrieval. The key idea is to replace the
distances in the original distance space with distances induces by geodesic paths
in the shape manifold. The merits of the proposed technique have been vali-
dated by significant performance gains over the experimental results. However,
like semi-supervised learning, if there are too many outlier shapes in the shape
database, the proposed approach cannot improve the results. Our future work
800 X. Yang et al.
will focus on addressing this problem. We also observe that our method is not
limited to 2D shape similarity but can also be applied to 3D shape retrieval,
which will also be part of our future work.
Acknowledgements
We would like to thank Eamonn Keogh for providing us the Face (all) dataset.
This work was support in part by the NSF Grant No. IIS-0534929 and by the
DOE Grant No. DE-FG52-06NA27508.
References
1. Belongie, S., Malik, J., Puzicha, J.: Shape matching and object recognition using
shape contexts. IEEE Trans. PAMI 24, 705–522 (2002)
2. Tu, Z., Yuille, A.L.: Shape matching and recognition - using generative models
and informative features. In: Pajdla, T., Matas, J(G.) (eds.) ECCV 2004. LNCS,
3. Ling, H., Jacobs, D.: Shape classification using the inner-distance. IEEE Trans.
PAMI 29, 286–299 (2007)
4. Latecki, L.J., Lakämper, R.: Shape similarity measure based on correspondence of
visual parts. IEEE Trans. PAMI 22(10), 1185–1190 (2000)
5. Felzenszwalb, P.F., Schwartz, J.: Hierarchical matching of deformable shapes. In:
CVPR (2007)
6. McNeill, G., Vijayakumar, S.: Hierarchical procrustes matching for shape retrieval.
In: Proc. CVPR (2006)
7. Bai, X., Latecki, L.J.: Path similarity skeleton graph matching. IEEE Trans.
PAMI 30, 1282–1292 (2008)
8. Srivastava, A., Joshi, S.H., Mio, W., Liu, X.: Statistic shape analysis: clustering,
learning, and testing. IEEE Trans. PAMI 27, 590–602 (2005)
9. Zhu, X.: Semi-supervised learning with graphs. In: Doctoral Dissertation. Carnegie
Mellon University, CMU–LTI–05–192 (2005)
10. Vleugels, J., Veltkamp, R.: Efficient image retrieval through vantage objects. Pat-
tern Recognition 35(1), 69–80 (2002)
11. Latecki, L.J., Lakämper, R., Eckhardt, U.: Shape descriptors for non-rigid shapes
with a single closed contour. In: CVPR, pp. 424–429 (2000)
12. Brefeld, U., Buscher, C., Scheffer, T.: Multiview dicriminative sequential learning.
In: Gama, J., Camacho, R., Brazdil, P.B., Jorge, A.M., Torgo, L. (eds.) ECML
2005. LNCS (LNAI), vol. 3720. Springer, Heidelberg (2005)
13. Lawrence, N.D., Jordan, M.I.: Semi-supervised learning via gaussian processes. In:
NIPS (2004)
14. Joachims, T.: Transductive inference for text classification using support vector
machines. In: ICML, pp. 200–209 (1999)
15. Zhu, X., Ghahramani, Z., Lafferty., J.: Semi-supervised learning using gaussian
fields and harmonic functions. In: ICML (2003)
16. Zhou, D., Bousquet, O., Lal, T.N., Weston, J., Scholkopf., B.: Learning with local
and global consistency. In: NIPS (2003)
17. Wang, F., Wang, J., Zhang, C., Shen., H.: Semi-supervised classification using
linear neighborhood propagation. In: CVPR (2006)
18. Zhou, D., Weston, J.: Ranking on data manifolds. In: NIPS (2003)
19. Roweis, S.T., Saul, L.K.: Nonlinear dimensionality reduction by locally linear em-
bedding. Science 290, 2323–2326 (2000)
20. Fan, X., Qi, C., Liang, D., Huang, H.: Probabilistic contour extraction using hier-
archical shape representation. In: Proc. ICCV, pp. 302–308 (2005)
21. Yu, J., Amores, J., Sebe, N., Radeva, P., Tian, Q.: Distance learning for similarity
estimation. IEEE Trans. PAMI 30, 451–462 (2008)
22. Xing, E., Ng, A., Jordanand, M., Russell, S.: Distance metric learning with appli-
cation to clustering with side-information. In: NIPS, pp. 505–512 (2003)
23. Bar-Hillel, A., Hertz, T., Shental, N., Weinshall, D.: Learning distance functions
using equivalence relations. In: ICML, pp. 11–18 (2003)
24. Athitsos, V., Alon, J., Sclaroff, S., Kollios, G.: Bootmap: A method for efficient
approximate similarity rankings. In: CVPR (2004)
25. Hertz, T., Bar-Hillel, A., Weinshall, D.: Learning distance functions for image
retrieval. In: CVPR, pp. 570–577 (2004)
26. Zelnik-Manor, L., Perona, P.: Self-tuning spectral clustering. In: NIPS (2004)
27. Hein, M., Maier, M.: Manifold denoising. In: NIPS (2006)
28. Wang, J., Chang, S.F., Zhou, X., Wong, T.C.S.: Active microscopic cellular image
annotation by superposable graph transduction with imbalanced labels. In: CVPR
(2008)
29. Mokhtarian, F., Abbasi, F., Kittler, J.: Efficient and robust retrieval by shape
content through curvature scale space. In: Smeulders, A.W.M., Jain, R. (eds.)
Image Databases and Multi-Media Search, pp. 51–58 (1997)
30. Sebastian, T.B., Klein, P.N., Kimia, B.: Recognition of shapes by editing their
shock graphs. IEEE Trans. PAMI 25, 116–125 (2004)
31. Keogh, E.: UCR time series classification/clustering page,
http://www.cs.ucr.edu/∼ eamonn/time series data/
32. Ratanamahatana, C.A., Keogh, E.: Three myths about dynamic time warping. In:
SDM, pp. 506–510 (2005)
Cat Head Detection - How to Effectively Exploit Shape
and Texture Features
Weiwei Zhang1, Jian Sun1 , and Xiaoou Tang2

1
Microsoft Research Asia, Beijing, China
{weiweiz,jiansun}@microsoft.com
2
Dept. of Information Engineering, The Chinese University of Hong Kong, Hong Kong
xtang@ie.cuhk.edu.hk
Abstract. In this paper, we focus on the problem of detecting the head of cat-like
animals, adopting cat as a test case. We show that the performance depends cru-
cially on how to effectively utilize the shape and texture features jointly. Specifi-
cally, we propose a two step approach for the cat head detection. In the first step,
we train two individual detectors on two training sets. One training set is normal-
ized to emphasize the shape features and the other is normalized to underscore
the texture features. In the second step, we train a joint shape and texture fusion
classifier to make the final decision. We demonstrate that a significant improve-
ment can be obtained by our two step approach. In addition, we also propose a set
of novel features based on oriented gradients, which outperforms existing leading
features, e. g., Haar, HoG, and EoH. We evaluate our approach on a well labeled
cat head data set with 10,000 images and PASCAL 2007 cat data.
1 Introduction
Automatic detection of all generic objects in a general scene is a long term goal in im-
age understanding and remains to be an extremely challenging problem duo to large
intra-class variation, varying pose, illumination change, partial occlusion, and cluttered
background. However, researchers have recently made significant progresses on a par-
ticularly interesting subset of object detection problems, face [14,18] and human detec-
tion [1], achieving near 90% detection rate on the frontal face in real-time [18] using
a boosting based approach. This inspires us to consider whether the approach can be
extended to a broader set of object detection applications.
Obviously it is difficult to use the face detection approach on generic object detection
such as tree, mountain, building, and sky detection, since they do not have a relatively
fixed intra-class structure like human faces. To go one step at a time, we need to limit
the objects to the ones that share somewhat similar properties as human face. If we can
succeed on such objects, we can then consider to go beyond. Naturally, the closest thing
to human face on this planet is animal head. Unfortunately, even for animal head, given
the huge diversity of animal types, it is still too difficult to try on all animal heads. This
is probably why we have seen few works on this attempt.
In this paper, we choose to be conservative and limit our endeavor to only one type
of animal head detection, cat head detection. This is of course not a random selection.

Cat Head Detection - How to Effectively Exploit Shape and Texture Features 803
(a) cat-like animal (b) cats
Fig. 1. Head images of animals of the cat family and cats
Our motivations are as follows. First, cat can represent a large category of cat-like an-
imals, as shown in Figure 1 (a). These animals share similar face geometry and head
shape; Second, people love cats. A large amount of cat images have been uploaded and
shared on the web. For example, 2,594,329 cat images had been manually annotated
in flickr.com by users. Cat photos are among the most popular animal photos on the
internet. Also, cat as a popular pet often appears in family photos. So cat detection can
find applications in both online image search and offline family photo annotation, two
important research topics in pattern recognition. Third, given the popularity of cat pho-
tos, it is easy for us to get training data. The research community does need large and
challenging data set to evaluate the advances of the object detection algorithm. In this
paper, we provide 10,000, well labeled cat images. Finally and most importantly, the cat
head detection poses new challenges for object detection algorithm. Although it shares
some similar property with human face so we can utilize some existing techniques, the
cat head do have much larger intra-class variation than the human face, as shown in
Figure 1 (b), thus is more difficult to detect.
Directly applying the existing face detection approaches to detect the cat head has
apparent difficulties. First, the cat face has larger appearance variations compared with
the human face. The textures on the cat face are more complicated than those on the
human face. It requires more discriminative features to capture the texture information.
Second, the cat head has a globally similar, but locally variant shape or silhouette. How
to effectively make use of both texture and shape information is a new challenging issue.
It requires a different detection strategy.
To deal with the new challenges, we propose a joint shape and texture detection ap-
proach and a set of new features based on oriented gradients. Our approach is a two step
approach. In the first step, we individually train a shape detector and a texture detector
to exploit the shape and appearance information respectively. Figure 2 illustrates our
basic idea. Figure 2 (a) and Figure 2 (c) are two mean cat head images over all training
images: one aligned by ears to make the shape distinct; the other is aligned to reveal the
texture structures. Correspondingly, the shape and texture detectors are trained on two
differently normalized training sets. Each detector can make full use of most discrimi-
native shape or texture features separately. Based on a detailed study of previous image
and gradient features, e.g., Haar [18], HoG [1], EOH [7], we show that a new set of
804 W. Zhang, J. Sun, and X. Tang
Shape Texture
(a) (b) (c)
Fig. 2. Mean cat head images on all training data. (a) aligned by ears. More shape information is
kept. (b) aligned by both eyes and ears using an optimal rotation+scale transformation. (c) aligned
by eyes. More texture information is kept.
carefully designed Haar-like features on oriented gradients give the best performance
in both shape and texture detectors.
In the second step, we train a joint shape and texture detector to fuse the outputs
of the above two detectors. We experimentally demonstrate that the cat head detection
performance can be substantially improved by carefully separating shape and texture
information in the first step, and jointly training a fusion classifier in the second step.
1.1 Related Work

Since a comprehensive review of the related works on object detection is beyond the
scope of the paper, we only review the most related works here.
Sliding window detection vs. parts based detection. To detect all possible objects
in the image, two different searching strategies have been developed. The sliding win-
dow detection [14,12,18,1,17,15,20] sequentially scans all possible sub-windows in the
image and makes a binary classification on each sub-window. Viola and Jones [18] pre-
sented the first highly accurate as well as real-time frontal face detector, where a cascade
classifier is trained by AdaBoost algorithm on a set of Haar wavelet features. Dalal and
Triggs [1] described an excellent human detection system through training a SVM clas-
sifier using HOG features. On the contrary, the parts based detection [5,13,9,6,3] detects
multiple parts of the object and assembles the parts according to geometric constrains.
For example, the human can be modeled as assemblies of parts [9,10] and the face can
be detected using component detection [5].
In our work, we use two sliding windows to detect the “shape” part and “texture”
part of the cat head. A fusion classifier is trained to produce the final decision.
Image features vs. gradient features. Low level features play a crucial role in the
object detection. The image features are directly extracted from the image, such as
intensity values [14], image patch [6], PCA coefficients [11], and wavelet coefficients
[12,16,18]. Henry et al.[14] trained a neural network for human face detection using
the image intensities in 20 × 20 sub-window. Haar wavelet features have become very
popular since Viola and Jones [18] presented their real-time face detection system. The
image features are suitable for small window and usually require a good photometric
normalization. Contrarily, the gradient features are more robust to illumination changes.
The gradient features are extracted from the edge map [4,3] or oriented gradients, which
mainly include SIFT [8], EOH [7], HOG [1], covariance matrix[17], shapelet [15], and
edgelet [19]. Tuzel et al. [17] demonstrated very good results on human detection using
the covariance matrix of pixel’s 1st and 2nd derivatives and pixel position as features.
Shapelet [15] feature is a weighted combination of weak classifiers in a local region. It
is trained specifically to distinguish between the two classes based on oriented gradients
from the sub-window. We will give a detailed comparison of our proposed features with
HOG and EOH features in Section 3.1.
2 Our Approach – Joint Shape and Texture Detection

The accuracy of a detector can be dramatically improved by first transforming the object
into a canonical pose to reduce the variability. In face detection, all training samples are
normalized by a rotation+scale transformation. The face is detected by scanning all sub-
windows with different orientations and scales. Unfortunately, unlike the human face,
the cat head cannot be well normalized by a rotation+scale transformation duo to the
large intra-class variation.
In Figure 2, we show three mean cat head images over 5,000 training images by
three normalization methods. In Figure 2 (a), we rotate and scale the cat head so that
both eyes appear on a horizontal line and the distance between two ears is 36 pixels.
As we can see, the shape or silhouette of the ears is visually distinct but the textures in
the face region are blurred. In a similar way, we compute the mean image aligned by
eyes, as shown in Figure 2 (c). The textures in the face region are visible but the shape
of the head is blurred. In Figure 2 (b), we take a compromised method to compute an
optimal rotation+scale transformation for both ears and eyes over the training data, in a
least square sense. As expected, both ears and eyes are somewhat blurred.
Intuitively, using the optimal rotation+scale transformation may produce the best re-
sult because the image normalized by this method contains two kinds of information.
However, the detector trained in this way does not show superior performance in our ex-
periments. Both shape and texture information are lost to a certain degree. The discrim-
inative power of shape features or texture features is hurt by this kind of compromised
normalization.
2.1 Joint Shape and Texture Detection

In this paper, we propose a joint shape and texture detection approach to effectively
exploit the shape and texture features. In the training phase, we train two individual
detectors and a fusion classifier:
1. Train a shape detector, using the aligned training images by mainly keeping the
shape information, as shown in Figure 2 (a); train a texture detector, using the
aligned training image by mainly preserving the texture information, as shown in
Figure 2 (c). Thus, each detector can capture most discriminative shape or texture
features respectively.
2. Train a joint shape and texture fusion classifier to fuse the output of the shape and
texture detectors.
In the detection phase, we first run the shape and texture detectors independently.
Then, we apply the joint shape and texture fusion classifier to make the final decision.
Specifically, we denote {cs , ct } as output scores or confidences of the two detectors,
and {fs , ft } as extracted features in two detected sub-windows. The fusion classifier is
trained on the concatenated features {cs , ct , fs , ft }.
Using two detectors, there are three kinds of detection results: both detectors re-
port positive at roughly the same location, rotation, and scale; only the shape detector
reports positive; and only the texture detector reports positive. For the first case, we
directly construct the features {cs , ct , fs , ft } for the joint fusion classifier. In the sec-
ond case, we do not have {ct , ft }. To handle this problem, we scan the surrounding
locations to pick a sub-window with the highest scores by the texture detector, as il-
lustrated in Figure 3. Specifically, we denote the sub-window reported by the detector
as [x, y, w, h, s, θ], where (x, y) is window’s center, w, h are width and height, and s, θ
are scale and rotation level. We search sub-windows for the texture/shape detector in
the range [x ± w/4] × [y ± h/4] × [s ± 1] × [θ ± 1]. Note that we use real value score of
the texture detector and do not make 0-1 decision. The score and features of the picked
sub-window are used for the features {ct , ft }. For the last case, we compute {cs , fs } in
a similar way.
To train the fusion classifier, 2,000 cat head images in the validation set are used as
the positive samples, and 4,000 negative samples are bootstrapped from 10,000 non-cat
images. The positive samples are constructed as usual. The key is the construction of the
negative samples which consist of all incorrectly detected samples by either the shape
detector or the texture detector in the non-cat images. The co-occurrence relationship
of the shape features and texture features are learned by this kind of joint training. The
learned fusion classifier is able to effectively reject many false alarms by using both
shape and texture information. We use support vector machine (SVM) as our fusion
classifier and HOG descriptors as the representations of the features fs and ft .
The novelty of our approach is the discovery that we need to separate the shape
and texture features and how to effectively separate them. The latter experimental re-
sults clearly validate the superiority of our joint shape and texture detection. Although
the fusion method might be simple at a glance, this is exactly the strength of our ap-
proach: a simple fusion method already worked far better than previous non-fusion
approaches.
{cs , fs } {ct , ft } {ct , ft } {cs , fs }
(a) (b)
Fig. 3. Feature extraction for fusion. (a) given a detected sub-window (left) by the shape detector,
we search a sub-window (right, solid line) with highest score by the texture detector in sur-
rounding region (right, dashed line). The score and features {ct , ft } are extracted for the fusion
classifier. (b) similarly, we extract the score and features {cs , fs } for the fusion.
3 Haar of Oriented Gradients
To effectively capture both shape and texture information, we propose a set of new
features based on oriented gradients.
3.1 Oriented Gradients Features
Given the image I, the image gradient −

→
g (x) = {gh , gv } for the pixel x is computed as:
gh (x) = Gh ⊗ I(x), gv (x) = Gv ⊗ I(x), (1)
where Gh and Gv are horizontal and vertical filters, and ⊗ is convolution operator. A
bank of oriented gradients {gok }K →
−
k=1 are constructed by quantifying the gradient g (x)
on a number of K orientation bins:

|−
→g (x)| θ(x) ∈ bink
go (x) =
k
, (2)
0 otherwise
where θ(x) is the orientation of the gradient − →g (x). We call the image gok oriented
gradients channel. Figure 4 shows the oriented gradients on a cat head image. In this
example, we quantify the orientation into four directions. We also denote the sum of
oriented gradients of a given rectangular region R as:

S k (R) = gok (x). (3)
x∈R
It can be very efficiently computed in a constant time using integral image technique [18].
Since the gradient information at an individual pixel is limited and sensitive to noise,
most of previous works aggregate the gradient information in a rectangular region to
form more informative, mid-level features. Here, we review two most successful fea-
tures: HOG and EOH.
Fig. 4. Oriented gradients channels in four directions

HOG-cell. The basis unit in the HOG descriptor is the weighted orientation histogram
of a “cell” which is a small spatial region, e.g., 8 × 8 pixels. It can be represented as:
HOG-cell(R) = [S 1 (R), ..., S k (R), ..., S K (R)]. (4)
The overlapped cells (e.g., 4 × 4) are grouped and normalized to form a larger spatial
region called “block”. The concatenated histograms form the HOG descriptor.
In Dalal and Triggs’s human detection system [1], a linear SVM is used to classify
a 64 × 128 detection window consisting of multiple overlapped 16 × 16 blocks. To
achieve near real-time performance, Zhu et al. [21] used HOGs of variable-size blocks
in the boosting framework .
EOH. Levi and Weiss [7] proposed three kinds of features on the oriented gradients:
EOH1 (R, k1, k2) = (S k1 (R) + )/(S k2 (R) + ),

EOH2 (R, k) = (S k (R) + )/( j (S j (R) + )),
EOH3 (R, R, k) = (S k (R) − S k (R))/sizeof (R),
where R is the symmetric region of R with respect to the vertical center of the detection
window, and is a small value for smoothing. The first two features capture whether
one direction is dominative or not, and the last feature is used to find symmetry or the
absence of symmetry. Note that using EOH features only may be insufficient. In [7],
good results are achieved by combining EOH features with Haar features on image
intensity.
Fig. 5. Haar of Oriented Gradients. Left: in-channel features. Right: orthogonal features.
3.2 Our Features - Haar of Oriented Gradients

In face detection, the Haar features demonstrated their great ability to discover local
patterns - intensity difference between two subregions. But it is difficult to find dis-
criminative local patterns on the cat head which has more complex and subtle fine scale
textures. On the contrary, the above oriented gradients features mainly consider the
marginal statistics of gradients in a single region. It effectively captures fine scale tex-
ture orientation distribution by pixel level edge detection operator. However, it fails to
capture local spatial patterns like the Haar feature. The relative gradient strength be-
tween neighboring regions is not captured either.
To capture both the fine scale texture and the local patterns, we need to develop a set
of new features combining the advantage of both Haar and gradient features. Taking a
close look at Figure 4, we may notice many local patterns in each oriented gradients
channel which is sparser and clearer than the original image. We may consider that
the gradient filter separates different orientation textures and pattern edges into several
channels thus greatly simplified the pattern structure in each channel. Therefore, it is
possible to extract Haar features from each channel to capture the local patterns. For
example, in the horizontal gradient map in Figure 4, we see that the vertical textures
between the two eyes are effectively filtered out so we can easily capture the two eye
pattern using Haar features. Of course, in addition to capturing local patterns within a
channel, we can also capture more local patterns across two different channels using
Haar like operation. In this paper, we propose two kinds of features as follows:
In-channel features
S k (R1 ) − S k (R2 )
HOOG1 (R1 , R2 , k) = . (5)
S k (R1 ) + S k (R2 )
These features measure the relative gradient strength between two regions R1 and R2
in the same orientation channel. The denominator plays a normalization role since we
do not normalize S k (R).
Orthogonal-channel features
∗
S k (R1 ) − S k (R2 )
HOOG2 (R1, R2, k, k ∗) = , (6)
S k (R1 ) + S k∗ (R2 )
where k ∗ is the orthogonal orientation with respect to k, i.e., k ∗ = k + K/2. These fea-
tures are similar to the in-channel features but operate on two orthogonal channels. In
theory, we can define these features on any two orientations. But we decide to compute
only the orthogonal-channel features based on two considerations: 1) orthogonal chan-
nels usually contain most complementary information. The information in two channels
with similar orientations is mostly redundant; 2) we want to keep the size of feature pool
small. The AbaBoost is a sequential, “greedy” algorithm for the feature selection. If the
feature pool contains too many uninformative features, the overall performance may
be hurt. In practice, all features have to be loaded into the main memory for efficient
training. We must be very careful about enlarging the size of features.
Considering all combinations of R1 and R2 will be intractable. Based on the success
of Haar features, we use Haar patterns for R1 and R2 , as shown in Figure 5. We call the
features defined in (5) and (6), Haar of Oriented Gradients (HOOG).
4.1 Data Set and Evaluation Methodology
Our evaluation data set includes two parts, the first part is our own data, which includes
10,000 cat images mainly obtained from flickr.com; the second part is from PASCAL
2007 cat data, which includes 679 cat images. Most of our own cat data are near frontal
view. Each cat head is manually labeled with 9 points, two for eyes, one for mouth,
and six for ears, as shown in Figure 6. We randomly divide our own cat face images
Fig. 6. The cat head image is manually labeled by 9 points
into three sets: 5,000 for training, 2000 for validation, and 3,000 for testing.We follow
the PASCAL 2007 original separations of training, validation and testing set on the cat
data. Our cat images can be downloaded from http://mmlab.ie.cuhk.edu.hk/ for research
purposes.
We use the evaluation methodology similar to PASCAL challenge for object detec-
tion. Suppose the ground truth rectangle and the detected rectangle are rg and rd , and
the area of those rectangles are Ag and Ad . We say we correctly detect a cat head only
when the overlap of rg and rd is larger than 50%:
(A ∩A )
1 if (Agg ∪Add ) > 50% ,
D(rg , rd ) = , (7)
0 otherwise
where D(rg , rd ) is a function used to calculate detection rate and false alarm rate.

Training samples. To train the shape detector, we align all cat head image with respect
to ears. We rotate and scale the image so that two tips of ears appear on a horizontal line
and the distance between two tips is 36 pixel. Then, we extract a 48 × 48 pixel region,
centered 20 pixels below two tips. For the texture detector, a 32 × 32 pixel region is
extracted. The distance between two eyes is 20 pixel. The region is centered 6 pixel
below two eyes.
Features. We use 6 unsigned orientations to compute the oriented gradients features.
We find the improvement is marginal when finer orientations are used. The horizontal
and vertical filters are [−1, 0, 1] and [−1, 0, 1]T . No thresholding is applied on the com-
puted gradients. For both shape and texture detector, we construct feature pools with
200,000 features by quantifying the size and location of the Haar templates.
4.3 Comparison of Features

First of all, we compare the proposed HOOG features with Haar, Haar + EOH, and
HOG features on both shape detector and texture detector using our Flickr cat data set.
For the Haar features, we use all four kinds of Haar templates. For the EOH features,
we use default parameters suggested in [7]. For the HOG features, we use 4 × 4 cell
size which produces the best results in our experiments.
0.9
0.9
0.8 0.8
0.7 0.7
Recall
Recall
0.6 0.6
0.5 0.5
Haar Haar
Haar+EOH 0.4 Haar+EOH
0.4
HOG HOG
our feature our feature
0.3 0.3
0 50 100 150 200 250 300 0 50 100 150 200 250 300 350 400 450 500
False Alarm Count False Alarm Count
(a) shape detector (b) texture detector
Fig. 7. Comparison of Haar, Haar+EOH, HOG, and our features
Figure 7 shows the performances of the four kinds of features. The Haar feature on
intensity gives the poorest performance because of large shape and texture variations
of the cat head. With the help of oriented gradient features, Haar + EOH improves the
performance. As one can expect, the HOG features perform better on the shape detector
than on the texture detector. Using both in-channel and orthogonal-channel information,
the detectors based on our features produce the best results.
shape detector texture detector
(a) (b) 0o (d) (e) 0o
(c) 60o, 150o (f) 30o, 120o
Fig. 8. Best features leaned by the AdaBoost. Left (shape detector): (a) best Haar feature on image
intensity. (b) best in-channel feature. (c) best orthogonal feature on orientations 60o and 150o .
Right (texture detector): (d) best Haar feature on image intensity. (e) best in-channel feature. (f)
best orthogonal-channel feature on orientations 30o and 120o .
In Figure 8, we show the best in-channel features in (b) and (e), and the best
orthogonal-channel features in (c) and (f), learned by two detectors. We also show the
best Haar features on image intensity in Figure 8 (a) and (d). In both detectors, the best
in-channel features capture the strength differences between a region with strongest
horizontal gradients and its neighboring region. The best orthogonal-channel features
capture the strength differences in two orthogonal orientations.
In the next experiment we investigate the role of in-channel features and orthogonal-
channel features. Figure 9 shows the performances of the detector using in-channel
features only, orthogonal-channel features only, and both kinds of features. Not surpris-
ingly, both features are important and complementary.
1 1
0.95 0.95
Precision
Precision
0.9 in-channel 0.9 in-channel
orthogonal-channel orthogonal-channel
0.85 0.85
in-channel + in-channel +
orthogonal-channel orthogonal-channel
0.8 0.8
0.5 0.6 Recall 0.7 0.8 0.5 0.6 Recall 0.7 0.8
(a) shape detector (b) texture detector
Fig. 9. The importance of in-channel features and orthogonal-channel features
4.4 Joint Shape and Texture Detection

In this sub-section, we evaluate the performance of the joint fusion on the Flickr cat
data. To demonstrate the importance of decomposing shape and texture features, we also
train a cat head detector using training samples aligned by an optimal rotation+scale
transformation for the comparison. Figure 10 shows four ROC curves: a shape detec-
tor, a texture detector, a head detector using optimal transformation, and a joint shape
and texture fusion detector. Several important observations can be obtained: 1) the per-
formance of fusion detector is substantially improved! For a given total false alarm
count 100, the recall is improved from 0.74/0.75/0.78 to 0.92. Or the total false alarm
is reduced from 130/115/90 to 20, for a fixed recall 0.76. In image retrieval and search
applications, it is a very nice property since high precision is preferred; 2) the head
detector using optimal transformation does not show superior performance. The dis-
criminative abilities of both shape and texture features are decreased by the optimal
transformation; 3) the maximal recall value of the fusion detector (0.92) is larger than
the maximal recall values of three individual detectors(0.77/0.82/0.85). This shows the
complementary abilities of two detectors - one detector can find many cat heads which
is difficult to the other detector; 4) note that the curve of fusion detector is very steep in
the low false alarm region, which means the fusion detector can effectively improve the
recall while maintain a very low false alarm rate.
The superior performance of our approach verifies a basic idea in object detection–
context helps! The fusion detector finds surrounding evidence to verify the detection re-
sult. In our cat head detection, when the shape detector reports a cat, the fusion detector
checks the surrounding shape information. If the texture detector says it may be a cat,
we increase the probability to accept this cat. Otherwise, we decrease the probability to
reject this cat.
1
0.95
0.9
0.85
0.8
Recall
0.75
0.7
0.65
Shape
0.6
Texture
0.55 Optimal Align
Shape+Texture
0.5
0 100 200 300 400 500 600
False Alalm Count
Fig. 10. Joint shape and texture detection
1 1
PASCA2007 Best
Haar
0.9 Our approach
0.9 Haar+EOH
0.8 0.8 HOG
0.7
our approach
0.7
Precision
Precision
0.6 0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
0 0. 1 0.2 0.3 0 .4 0.5 0 .6 0. 7 0.8 0.9 1
0 0. 1 0.2 0.3 0 .4 0. 5 0.6 0 .7 0 .8 0. 9 1
Recall Recall
(a) Competition 3 (b) Competition 4
Fig. 11. Experiments on PASCAL 2007 cat data. (a) our approach and best reported method on
Competition 3 (specified training data). (b) four detectors on Competition 4 (arbitrary training
data).
Figure 12 gives some detection examples having variable appearance, head shape,
illumination, and pose.
4.5 Experiment on the PASCAL 2007 Cat Data

We also evaluate the proposed approach on the PASCAL 2007 cat data [2]. There are
two kinds of competitions for the detection task: 1) Competition 3 - using both training
and testing data from PASCAL 2007; 2) Competition 4 - using arbitrary training data.
Figure 11 (a) shows the precision-recall curves of our approach and the best reported
method [2] on Competition 3. We compute the Average Precision (AP) as in [2] for a
convenient comparison. The APs of our approach and the best reported method is 0.364
and 0.24, respectively. Figure 11(b) shows the precision-recall curves on Competition
4. Since there is no reported result on Competition 4, we compare our approach with
the detectors using Haar, EOH, and HoG respectively. All detectors are trained on the
Fig. 12. Detection results. The bottom row shows some detected cats in PASCAL 2007 data.
same training data. The APs of four detectors (ours, HOG, Haar+EOH, Harr) are 0.632,
0.427, 0.401, and 0.357. Using larger training data, the detection performance is signif-
icantly improved. For example, the precision is improved from 0.40 to 0.91 for a fixed
recall 0.4. Note that the PASCAL 2007 cat data treat the whole cat body as the object
and only small fraction of the data contain near frontal cat face. However, our approach
still achieves reasonable good results (AP=0.632) on this very challenging data (the best
reported method’s AP=0.24).

In this paper, we have presented a cat head detection system. We achieved excellent
results by decomposing texture and shape features firstly and fusing detection results
secondly. The texture and shape detectors also greatly benefit from a set of new oriented
gradient features. Although we focus on the cat head detection problem in this paper,
our approach can be extended to detect other categories of animals. In the future, we
are planing to extend our approach to multi-view cat head detection and more animal
categories. We are also interest in exploiting other contextual information, such as the
presence of animal body, to further improve the performance.
References
1. Dalal, N., Triggs, B.: Histograms of oriented gradients for human detection. In: CVPR, vol. 1,
pp. 886–893 (2005)
2. Everingham, M., van Gool, L., Williams, C., Winn, J., Zisserman, A.: The PASCAL Visual
Object Classes Challenge (VOC 2007) Results (2007),
http://www.pascal-network.org/challenges/VOC/voc2007/
workshop/index.html
3. Felzenszwalb, P.F.: Learning models for object recognition. In: CVPR, vol. 1, pp. 1056–1062
(2001)
4. Gavrila, D.M., Philomin, V.: Real-time object detection for smart vehicles. In: CVPR, vol. 1,
pp. 87–93 (1999)
5. Heisele, B., Serre, T., Pontil, M., Poggio, T.: Component-based face detection. In: CVPR,
vol. 1, pp. 657–662 (2001)
6. Leibe, B., Seemann, E., Schiele, B.: Pedestrian detection in crowded scenes. In: CVPR,
vol. 1, pp. 878–885 (2005)
7. Levi, K., Weiss, Y.: Learning object detection from a small number of examples: the impor-
tance of good features. In: CVPR, vol. 2, pp. 53–60 (2004)
8. Lowe, D.G.: Object recognition from local scale-invariant features. In: ICCV, vol. 2, pp.
1150–1157 (1999)
9. Mikolajczyk, K., Schmid, C., Zisserman, A.: Human detection based on a probabilistic
assembly of robust part detectors. In: Pajdla, T., Matas, J(G.) (eds.) ECCV 2004. LNCS,
10. Mohan, A., Papageorgiou, C., Poggio, T.: Example-based object detection in images by com-
ponents. IEEE Trans. Pattern Anal. Machine Intell. 23(4), 349–361 (2001)
11. Munder, S., Gavrila, D.M.: An experimental study on pedestrian classification. IEEE Trans.
Pattern Anal. Machine Intell. 28(11), 1863–1868 (2006)
12. Papageorgiou, C., Poggio, T.: A trainable system for object detection. Intl. Journal of Com-
puter Vision 38(1), 15–33 (2000)
13. Ronfard, R., Schmid, C., Triggs, B.: Learning to parse pictures of people. In: ECCV, vol. 4,
pp. 700–714 (2004)
14. Rowley, H.A., Baluja, S., Kanade, T.: Neural network-based face detection. IEEE Trans.
Pattern Anal. Machine Intell. 20(1), 23–38 (1998)
15. Sabzmeydani, P., Mori, G.: Detecting pedestrians by learning shapelet features. In: CVPR
(2007)
16. Schneiderman, H., Kanade, T.: A statistical method for 3d object detection applied to faces
and cars. In: CVPR, vol. 1, pp. 746–751 (2000)
17. Tuzel, O., Porikli, F., Meer, P.: Human detection via classification on riemannian manifolds.
In: CVPR (2007)
18. Viola, P., Jones, M.J.: Robust real-time face detection. Intl. Journal of Computer Vision 57(2),
137–154 (2004)
19. Wu, B., Nevatia, R.: Detection of multiple, partially occluded humans in a single image by
bayesian combination of edgelet part detectors. In: ICCV, vol. 1, pp. 90–97 (2005)
20. Xiao, R., Zhu, H., Sun, H., Tang, X.: Dynamic cascades for face detection. In: ICCV, vol. 1,
pp. 1–8 (2007)
21. Zhu, Q., Avidan, S., Yeh, M.-C., Cheng, K.-T.: Fast human detection using a cascade of
histograms of oriented gradients. In: CVPR, vol. 2, pp. 1491–1498 (2006)
Motion Context: A New Representation for
Human Action Recognition
Ziming Zhang, Yiqun Hu, Syin Chan, and Liang-Tien Chia
Center for Multimedia and Network Technology, School of Computer Engineering,

Nanyang Technological University, Singapore 639798
{zhan0154,yqhu,asschan,asltchia}@ntu.edu.sg
Abstract. One of the key challenges in human action recognition from

video sequences is how to model an action sufficiently. Therefore, in this
paper we propose a novel motion-based representation called Motion
Context (MC), which is insensitive to the scale and direction of an ac-
tion, by employing image representation techniques. A MC captures the
distribution of the motion words (MWs) over relative locations in a lo-
cal region of the motion image (MI) around a reference point and thus
summarizes the local motion information in a rich 3D MC descriptor.
In this way, any human action can be represented as a 3D descriptor
by summing up all the MC descriptors of this action. For action recog-
nition, we propose 4 different recognition configurations: MW+pLSA,
MW+SVM, MC+w3 -pLSA (a new direct graphical model by extending
pLSA), and MC+SVM. We test our approach on two human action video
datasets from KTH and Weizmann Institute of Science (WIS) and our
performances are quite promising. For the KTH dataset, the proposed
MC representation achieves the highest performance using the proposed
w3 -pLSA. For the WIS dataset, the best performance of the proposed
MC is comparable to the state of the art.
1 Introduction
With the development of advanced security systems, human action recognition

in video sequences has become an important research topic in computer vision,
whose aim is to make machines recognize human actions using different types of
information, especially the motion information, in the video sequences.
The basic process for this problem can be divided into three issues: First,
how to detect the existence of human actions? Second, how to represent human
actions? Lastly, how to recognize these actions? Many research works have been
done to address these issues (e.g. [1], [2], [3], [4], [5], [6]). In this paper, we mainly
focus on the second issue, that is, how to represent human actions after having
detected their existence. In our approach, we model each video sequence as a
collection of so-called motion images (MIs), and to model the action in each
MI, we propose a novel motion-based representation called motion context
(MC), which is insensitive to the scale and direction of an action, to capture
the distribution of the motion words (MWs) over relative locations in a local

818 Z. Zhang et al.
hand- hand-
boxing jogging running walking
clapping waving
Frame Group
Motion Context Motion Image
Fig. 1. Illustrations of the frame groups, motion images, and our motion context rep-
resentations on the KTH dataset. This figure is best viewed in color.
region around a reference point and thus summarize the local motion informa-
tion in a rich, local 3D MC descriptor. Fig.1 illustrates some MIs and their
corresponding MC representations using the video clips in the KTH dataset.
To describe an action, only one 3D descriptor is generated by summing up all
the MC descriptors of this action in the MIs. For action recognition, we employ
3 different approaches: pLSA [7], w3 -pLSA (a new direct graphical model by
extending pLSA) and SVM [8]. Our approach is tested on two human action
video datasets from KTH [2] and Weizmann Institute of Science [9], and the
performances are quite promising.
The rest of this paper is organized as follows: Section 2 reviews some related
works in human action recognition. Section 3 presents the details of our MC rep-
resentation. Section 4 introduces the 3 recognition approaches. Our experimental
results are shown in Section 5, and finally Section 6 concludes the paper.
2 Related Work
Each video sequence can be considered as a collection of consecutive images

(frames), which makes it possible to model human actions using some image
representation techniques. One influential model is the Bag-of-Words (BOW)
model (e.g. [4], [6], [10], [11]). This model represents each human action as a
collection of independent codewords in a pre-defined codebook generated from
the training data. However, videos contain temporal information while images
do not. So how to exploit this temporal information becomes a key issue for
human action representation.
Based on image representation techniques, many research works have shown
that temporal information can be integrated with the interesting point detectors
and descriptors to locate and describe the interesting points in the videos. Laptev
et al. [1] proposed a 3D interesting point detector where they added the temporal
constraint to the Harris interesting point detector to detect local structures in
the space-time dimensions. Efros et al. [12] proposed a motion descriptor using
Motion Context: A New Representation for Human Action Recognition 819
the optical flow from different frames to represent human actions. Recently,
Scovanner et al. [4] applied sub-histograms to encode local temporal and spatial
information to generate a 3D version of SIFT [13] (3D SIFT), and Savarese et al.
[14] proposed so-called “spatial-temporal correlograms” to encode flexible long
range temporal information into the spatial-temporal motion features.
However, a common issue behind these interesting point detectors is that the
detected points sometimes are too few to sufficiently characterize the human
action behavior, and hence reduce the recognition performance. This issue has
been avoided in [6] by employing the separable linear filter method [3], rather
than such space-time interesting point detectors, to obtain the motion features
using a quadrature pair of 1D Gabor filters temporally.
Another way of using temporal information is to divide a video into smaller
groups of consecutive frames as the basic units and represent a human action as
a collection of the features extracted from these units. In [15], [5], every three
consecutive frames in each video were grouped together and integrated with
their graphical models as a node to learn the spatial-temporal relations among
these nodes. Also in [16], the authors took the average of a sequence of binary
silhouette images of a human action to create the “Average Motion Energy”
representation. Similarly, [17] proposed a concept of “Motion History Volumes”,
an extension of “Motion History Images” [18], to capture the motion information
from a sequence of video frames.
After the human action representations have been generated, both discrim-
inative approaches (e.g. kernel approaches [2]) and generative approaches (e.g.
pLSA [19], MRF [15], [5], semi-LDA [10], hierarchical graphical models [6]) can
be employed to recognize them.
3 Motion Context Representation
A motion context representation is generated based on the motion words which

are extracted from the motion images.
3.1 Motion Image
We believe that effective utilization of the temporal information is crucial for

human action recognition. In our approach, we adopt the strategy in [17], that
is, to group the consecutive frames of each video sequence according to their
temporal information.
More specifically, to generate a motion image (MI), first U ∗ V frames of
a video sequence are extracted, converted into gray scale and divided into non-
overlapping U groups, each with V consecutive frames. Then we calculate the
standard deviation (stdev) among the frames within a group pixel by pixel to
detect the motion information. Finally, putting the stdev values into the cor-
responding pixel positions, a MI is generated for each frame group. Fig.2 illus-
trates the MI generation process for a frame group. Motions usually cause strong
changes in the pixel intensity values at the corresponding positions among the
820 Z. Zhang et al.
Group of Frames Motion Image
Calucate
Standard Deviation
Fig. 2. Illustration of the MI generation process for a frame group. The black dots
denote the pixel intensity values.
consecutive frames. Since stdev can measure the variances of the pixel intensity
values, it can definitely detect motions.
We would like to mention that the length of each group, V , should be long
enough to capture the motion information sufficiently but not too long. Fig.3
illustrates the effects of different V on the MIs of human running and walking.
If V = 5, the difference between the two actions is quite clear. With V increased
to 60, the motion information of both actions spreads in the MIs, making it
difficult to distinguish them. A further investigation of V will be essential in our
MC representation.
5 frames 60 frames 5 frames 60 frames

running
walking
Fig. 3. Illustration of effects of different lengths of frame groups on the MIs using
human running and walking
3.2 Motion Word

The concept of motion words (MWs) refers to that of visual words in the BOW
model. After generating the MIs, some image interesting point detectors are first
applied to locate the important patches in the MIs. Then image descriptors are
employed to map these patches into a high dimensional feature space to generate
local feature vectors for them. Next, using clustering approaches such as K-
means, these local feature vectors in the training data are clustered to generate
a so-called motion word dictionary where the centers of the clusters are treated
as the MWs.
3.3 Motion Context

For each MW, there is one important affiliated attribute, its location in the corre-
sponding MI. For human action recognition, the relative movements of different
parts of the body are quite useful. To capture the structures of these relative
movements, we introduce the concept of motion context (MC). This concept
M o tio n
w o rd
Normalized
distance
gl ve
an at i
e
el
R
Fig. 4. Illustration of our MC representation (left) and its 3D descriptor (right). On
the left, P denotes a MW at an interesting point, O denotes the reference point, Θ and
S denote the relative angle and normalized distance between P and O in the support
region (the black rectangle), respectively, and the shaded sector (blue) denotes the
orientation of the whole representation. On the right, each MW is quantized into a
point to generate a 3D MC descriptor. This figure is best viewed in color.
is inspired by Shape Context (SC) [20], which has been widely used in object
recognition. The basic idea of SC is to locate the distribution of other shape
points over relative positions in a region around a pre-defined reference point.
Subsequently, 1D descriptors are generated to represent the shapes of objects.
In our representation, we utilize the polar coordinate system to capture the
relative angles and distances between the MWs and the reference point (the
pole of the polar coordinate system) for each action in the MIs, similar to SC.
This reference point is defined as the geometric center of the human motion, and
the relative distances are normalized by the maximum distance in the support
region, which makes the MC insensitive to changes in scale of the action. Here,
the support region is defined as the area which covers the human action in the
MI. Fig.4 (left) illustrates our MC representation. Suppose that the angular
coordinate is divided into M equal bins, the radial coordinate is divided into
N equal bins and there are K MWs in the dictionary, then each MW can be
put into one of the M *N bins to generate a 3D MC descriptor for each MC
representation, as illustrated in Fig.4 (right). To represent a human action in
each video sequence, we sum up all the MC descriptors of this action to generate
one 3D descriptor with the same dimensions.
When generating MC representations, another factor should also be consid-
ered, that is, the direction of the action, because the same action may occur in
different directions. E.g. a person may be running in one direction or the oppo-
site direction. In such cases, the distributions of the interesting points in the two
corresponding MIs should be roughly symmetric about the y-axis. Combining
the two distributions for the same action will reduce the discriminability of our
representation. To avoid this, we define the orientation of each MC representa-
tion as the sector where most interesting points are detected, e.g. the shaded
one (blue) in Fig. 4 (left). This sector can be considered to represent the main
characteristics of the motion in one direction. For the same action but in the
822 Z. Zhang et al.
original Motion Context

processed Motion Context
representation of an action
representation of the action
Pre-defined aligning side of all the

in the opposite direction
orientations
Flipping
Orientation Orientation
y-axis y-axis
Fig. 5. Illustration of aligning an inconsistent MC representation of an action in the

opposite direction. The pre-defined orientation of the actions is the left side of y-axis.
opposite direction, we then align all the orientations to the pre-defined side by
flipping the MC representations horizontally around the y-axis. Thus our repre-
sentation is symmetry-invariant. Fig.5 illustrates this process. Notice that this
process is done automatically without the need to know the action direction.
The entire process of modeling human actions using the MC representation
is summarized in Table 1.
Table 1. The main steps of modeling the human actions using the MC representation
Step 1 Obtain the MIs from the video sequences.

Step 2 Generate the MC representation for each human action in the MIs.
Step 3 Generate the 3D MC descriptor for each MC representation.
Step 4 Sum up all the 3D MC descriptors of an action to generate one 3D descriptor
to represent this action.
4 Action Recognition Approaches

We apply 3 different approaches to recognize the human actions based on the
MWs or the 3D MC descriptors: pLSA, w3 -pLSA and SVM.
4.1 pLSA
pLSA aims to introduce an aspect model, which builds an association between
documents and words through the latent aspects by probability. Here, we follow
the terminology of text classification where pLSA was used first. The graphical
model of pLSA is illustrated in Fig.6 (a).
Suppose D = {d1 , . . . , dI }, W = {w1 , . . . , wJ } and Z = {z1 , . . . , zK } denote a
document set, a word set and a latent topic set, respectively. pLSA models the
joint probability of documents and words as:

P (di , wj ) = P (di , wj , zk ) = P (wj |zk )P (zk |di )P (di ) (1)
k k
w.
d z w d z
s
(a) pLSA (b) w 3-pLSA
Fig. 6. Graphical models of pLSA (a) and our w3 -pLSA (b)
where P (di , wj , zk ) denotes the joint probability of document di , topic zk and

word wj , P (wj |zk ) denotes the probability of wj occurring in zk , P (zk |di ) denotes
the probability of di classified into zk , and P (di ) denotes the prior probability
of di modeled as a multinomial distribution.
Furthermore, pLSA tries to maximize the L function below:

L= n(di , wj ) log P (di , wj ) (2)
i j
where n(di , wj ) denotes the document-word co-occurrence table, where the num-
ber of co-occurrences of di and wj is recorded in each cell.
To learn the probability distributions involved, pLSA employs the Expecta-
tion Maximization (EM) algorithm shown in Table 2 and records P (wj |zk ) for
recognition, which is learned from the training data.
Table 2. The EM algorithm for pLSA
E-step:
P (zk |di , wj ) ∝ P (wj |zk )P (zk |di )P (di )
M-step:
P (wj |zk ) ∝ i n(di , wj )P (zk |di , wj )
P (zk |di ) ∝ j n(di , wj )P (zk |di , wj )

P (di ) ∝ j n(di , wj )
4.2 w3 -pLSA
To bridge the gap between the human actions and our MC descriptors, we extend
pLSA to develop a new graphical model, called w3 -pLSA. See Fig.6 (b), where d
denotes human actions, z denotes latent topics, w, θ and s denote motion words,
and the indexes in the angular and radial coordinates in the polar coordinate
system, respectively.
Referring to pLSA, we model the joint probability of human actions, motion
words and their corresponding indices in the angular and radial coordinates as

P (di , wj , θm , sr ) = P (di , wj , θm , sr , zk ) = P (di )P (zk |di )P (wj , θm , sr |zk )
k k
(3)
824 Z. Zhang et al.
and maximize the L< function below.

L< = n(di , wj , θm , sr ) log P (di , wj , θm , sr ) (4)
i j m r
Similarly, to learn the probability distributions involved, w3 -pLSA employs

the Expectation Maximization (EM) algorithm shown in Table 3 and records
P (wj , θm , sr |zk ) for recognition, which is learned from the training data.
Table 3. The EM algorithm for w3 -pLSA
E-step:
P (zk |di , wj , θm , sr ) ∝ P (wj , θm , sr |zk )P (zk |di )P (di )
M-step:
P (wj , θm , sr |zk ) ∝ i n(di , wj , θm , sr )P (zk |di , wj , θm , sr )
P (zk |di ) ∝ j,m,r n(di , wj , θm , sr )P (zk |di , wj , θm , sr )

P (di ) ∝ j,m,r n(di , wj , θm , sr )
4.3 Support Vector Machine

A support vector machine (SVM) [8] is a powerful tool for binary classification
tasks. First it maps the input vectors into a higher dimensional feature space,
then it conducts a separating hyperplane to separate the input data, finally on
each side of this hyperplane two parallel hyperplanes are conducted. SVM tries
to find the separating hyperplane which maximizes the distance between the
two parallel hyperplanes. Notice that in a SVM, there is an assumption that
the larger the distance between the two parallel hyperplanes the smaller the
generalization error of the classifier will be.
Specifically, suppose the input data is {(x1 , y1 ), (x2 , y2 ), · · · , (xn , yn )} where
xi (i = 1, 2, · · · , n) denotes the input vector and the corresponding yi (i = 1, 2,
· · · , n) denotes the class label (positive “1” and negative “-1”). Then the sepa-
rating hyperplane is defined as w · x + b = 0 and the two corresponding parallel
hyperplanes are w · x + b = 1 for the positive class and w · x + b = −1 for the
negative class, where w is the vector perpendicular to the separating hyperplane
and b is a scalar. If a test vector xt satisfies w · xt + b > 0, it will be classified as
a positive instance. Otherwise, if it satisfies w · xt + b < 0, it will be classified as
a negative instance. A SVM tries to find the optimal w and b to maximize the
distance between the two parallel hyperplanes.
5 Experiments
Our approach has been tested on two human action video datasets from KTH
[2] and Weizmann Institute of Science (WIS) [9]. The KTH dataset is one of
the largest datasets for human action recognition containing six types of human
actions: boxing, handclapping, handwaving, jogging, running, and walking. For
each type, there are 99 or 100 video sequences of 25 different persons in 4 differ-
ent scenarios: outdoors (S1), outdoors with scale variation (S2), outdoors with
different clothes (S3) and indoors (S4), as illustrated in Fig.7 (left). In the WIS
dataset, there are altogether 10 types of human actions: walk, run, jump, gallop
sideways, bend, one-hand wave, two-hands wave, jump in place, jumping jack,
and skip. For each type, there are 9 or 10 video sequences of 9 different persons
with the similar background, as shown in Fig.7 (right).
S1 S2 S3 S4
boxing
bend
side
clappin g
hand-
skip
jack
waving
hand-
jump
walk
joggin g
w ave1
pjum p
runn ing
w ave2
walking
run
Fig. 7. Some sample frames from the KTH dataset (left) and the WIS dataset (right)
5.1 Implementation
To generate MC representations for human actions, we need to locate the ref-
erence points and the support regions first. Some techniques in body tracking
(e.g. [21]) can be applied to locate the areas and the geometric centers of the
human bodies in each frame group of a video sequence. The integration of the
areas of a person can be defined as its support region and the mean of its centers
can be defined as the reference point for this action in the MI. However, this
issue is beyond the purpose of this paper. So considering that in our datasets
each video sequence only contains one person, we simply assume that in each MI
the support region of each human action covers the whole MI, and we adopted a
simple method to roughly locate the reference points. First, we generated one MI
from every 5-frame group of each video sequence empirically. Then a Gaussian
filter was applied to denoise these MIs so that the motion information from the
background was suppressed. Next, we used the Canny edge detector to locate
the edges in each MI, and finally took the geometric center of the edge points as
the reference point for the action.
After locating the reference points, we followed the steps in Table 1 to generate
the MC representations for human actions. The detector and descriptor involved
in Step 2 are the Harris-Hessian-Laplace detector [22] and the SIFT descriptor
826 Z. Zhang et al.
Table 4. Comparison (%) between our approach and others on the KTH dataset
Rec.Con. Tra.Str. boxing hand-c hand-w jogging running walking average

MW+pLSA SDE 85.2 91.9 91.7 71.2 73.6 82.1 82.62
LOO 82.0 90.9 91.0 82.0 79.0 83.0 84.65
MW+SVM SDE 90.4 84.8 82.8 65.1 76.1 82.0 80.20
LOO 85.0 82.8 82.0 62.0 70.0 87.0 78.14
MC+w3 -pLSA SDE 98.4 90.8 93.9 79.3 77.9 91.7 88.67
LOO 95.0 97.0 93.0 88.0 84.0 91.0 91.33
MC+SVM SDE 91.7 91.6 88.1 78.0 84.7 90.4 87.42
LOO 88.0 93.9 91.0 77.0 85.0 90.0 87.49
Savarese et al. [14] LOO 97.0 91.0 93.0 64.0 83.0 93.0 86.83
Wang et al. [10] LOO 96.0 97.0 100.0 54.0 64.0 99.0 85.00
Niebles et al. [19] LOO 100.0 77.0 93.0 52.0 88.0 79.0 81.50
Dollár et al. [3] LOO 93.0 77.0 85.0 57.0 85.0 90.0 81.17
Schuldt et al. [2] SDE 97.9 59.7 73.6 60.4 54.9 83.8 71.72
Ke et al. [24] SDE 69.4 55.6 91.7 36.1 44.4 80.6 62.96
Wong et al. [25] SDE 96.0 92.0 83.0 79.0 54.0 100.0 84.00
[13], and the clustering method used here is K-means. Then based on the MWs
and the MC descriptors of the training data, we trained pLSA, w3 -pLSA and
SVM for each type of actions separately, and a test video sequence was classified
to the type of actions with the maximum likelihood.
5.2 Experimental Results

To show the efficiency of our MC representation and the discriminability of the
MWs, we designed 4 different recognition configurations: MW+pLSA, MW+SVM,
MC+w3 -pLSA, and MC+SVM. Here we used libsvm [23] with the linear kernel. To
utilize the MWs, we employed the BOW model to represent each human action as
a histogram of the MWs without the M *N spatial bins.
First, we tested our approach on the KTH dataset. We adopted two different
training strategies: split-data-equally (SDE) and leave-one-out (LOO). The SDE
strategy means that the video collection is divided into two equal sets randomly:
one as the training data (50 video sequences) and the other as the test data for
each type of actions, and we repeated this experiment for 15 times. In the LOO
strategy, for each type of actions, only the video sequences of one person are
selected as the test data and the rest as the training data, and when applying
this strategy to the KTH dataset, for each run we randomly selected one person
for each type of actions as the test data and repeated this experiment for 15
times. Empirically, in our model, the number of MWs is 100, and the numbers
of the quantization bins in the angular and radial dimensions are 10 and 2,
respectively. The number of latent topics in both graphical models is 40.
Table 4 shows our average recognition rate for each type of actions and the
comparison with others on the KTH dataset under different training strate-
gies and recognition configurations. From this table, we can draw the following
Table 5. Comparison (%) between our approach and others on the WIS dataset. Notice
that “✕” denotes that this type of actions was not involved in their experiments.
Rec.Con. bend jack jump pjump run side skip walk wave1 wave2 ave.
MW+pLSA 77.8 100.0 88.9 88.9 70.0 100.0 60.0 100.0 66.7 88.9 84.1
MW+SVM 100.0 100.0 100.0 77.8 30.0 77.8 40.0 100.0 100.0 100.0 81.44
MC+w3 -pLSA 66.7 100.0 77.8 66.7 80.0 88.9 100.0 100.0 100.0 100.0 88.0
MC+SVM 100.0 100.0 100.0 88.9 80.0 100.0 80.0 80.0 100.0 100.0 92.89
Wang et al. [16] 100.0 100.0 89.0 100.0 100.0 100.0 89.0 100.0 89.0 100.0 96.7
Ali et al. [26] 100.0 100.0 55.6 100.0 88.9 88.9 100.0 100.0 100.0 92.6
Scovanner [4] 100.0 100.0 67.0 100.0 80.0 100.0 50.0 89.0 78.0 78.0 84.2
Niebles et al. [6] 100.0 100.0 100.0 44.0 67.0 78.0 56.0 56.0 56.0 72.8
conclusions: (1) MWs without any spatial information are not discriminative
enough to recognize the actions. MW+pLSA returns the best performance
(84.65%) using MWs, which is lower than the state of the art. (2) MC repre-
sentation usually achieves better performances than MWs, which demonstrates
that the distributions of the MWs are quite important for action recognition.
MC+w3 -pLSA returns the best performance (91.33%) among all the approaches.
Unlike the KTH dataset, the WIS dataset only has 9 or 10 videos for each
type of human actions, which may result in underfit when training the graphical
models. To utilize this dataset sufficiently, we only used the LOO training strat-
egy to learn the models for human actions and tested on all the video sequences.
We compare our average recognition rates with others in Table 5. The experi-
mental configuration of the MC representation is kept the same as that used on
the KTH dataset, while the number of MWs used in the BOW model is modified
empirically to 300. The number of latent topics is unchanged. From this table,
we can see that MC+SVM still returns the best performance (92.89%) among
the different configurations, which is comparable to other approaches and higher
than the best performance (84.1%) using MW. These results demonstrate that
our MC presentation can model the human actions properly with the distribu-
tions of the MWs.
6 Conclusion
We have demonstrated that our Motion Context (MC) representation, which

is insensitive to changes in the scales and directions of the human actions, can
model the human actions in the motion images (MIs) effectively by capturing the
distribution of the motion words (MWs) over relative locations in a local region
around the reference point and thus summarize the local motion information
in a rich 3D descriptor. To evaluate this novel representation, we adopt two
training strategies (split-data-equally (SDE) and leave-one-out (LOO)), design
4 different recognition configurations (MW+pLSA, MW+SVM, MC+w3 -pLSA,
and MC+SVM) and test them on two human action video datasets from KTH
828 Z. Zhang et al.
and Weizmann Institute of Science (WIS). The performances are promising. For
the KTH dataset, all configurations using MC outperform existing approaches
where the best performances are obtained using w3 -pLSA (88.67% for SDE and
91.33% for LOO). For the WIS dataset, our MC+SVM returns the comparable
performance (92.89%) using the LOO strategy.
References
1. Laptev, I., Lindeberg, T.: Space-time interest points. In: ICCV (2003)
2. Schuldt, C., Laptev, I., Caputo, B.: Recognizing human actions: a local svm ap-
proach. In: ICPR 2004, vol. III, pp. 32–36 (2004)
3. Dollár, P., Rabaud, V., Cottrell, G., Belongie, S.: Behavior recognition via sparse
spatio-temporal features. In: VS-PETS (October 2005)
4. Scovanner, P., Ali, S., Shah, M.: A 3-dimensional sift descriptor and its application
to action recognition. ACM Multimedia, 357–360 (2007)
5. Wang, Y., Loe, K.F., Tan, T.L., Wu, J.K.: Spatiotemporal video segmentation
based on graphical models. Trans. IP 14, 937–947 (2005)
6. Niebles, J., Fei Fei, L.: A hierarchical model of shape and appearance for human
action classification. In: CVPR 2007, pp. 1–8 (2007)
7. Hofmann, T.: Unsupervised learning by probabilistic latent semantic analysis. In:
Mach. Learn., Hingham, MA, USA, vol. 42, pp. 177–196. Kluwer Academic Pub-
lishers, Dordrecht (2001)
8. Burges, C.J.C.: A tutorial on support vector machines for pattern recognition. In:
Data Mining and Knowledge Discovery, vol. 2, pp. 121–167 (1998)
shapes. In: ICCV 2005, vol. II, pp. 1395–1402 (2005)
10. Wang, Y., Sabzmeydani, P., Mori, G.: Semi-latent dirichlet allocation: A hierar-
chical model for human action recognition. In: HUMO 2007, pp. 240–254 (2007)
11. Ikizler, N., Duygulu, P.: Human action recognition using distribution of oriented
rectangular patches. In: HUMO 2007, pp. 271–284 (2007)
12. Efros, A., Berg, A., Mori, G., Malik, J.: Recognizing action at a distance. In: ICCV
2003, pp. 726–733 (2003)
13. Lowe, D.: Distinctive image features from scale-invariant keypoints. International
14. Savarese, S., Sel Pozo, A., Fei-Fei, J.N.L.: Spatial-temporal correlations for unsu-
pervised action classification. In: IEEE Workshop on Motion and Video Comput-
ing, Copper Mountain, Colorado (2008)
15. Wang, Y., Tan, T., Loe, K.: Video segmentation based on graphical models. In:
CVPR 2003, vol. II, pp. 335–342 (2003)
16. Wang, L., Suter, D.: Informative shape representations for human action recogni-
tion. In: ICPR 2006, vol. II, pp. 1266–1269 (2006)
17. Weinland, D., Ronfard, R., Boyer, E.: Free viewpoint action recognition using mo-
tion history volumes. Computer Vision and Image Understanding 104 (Novem-
ber/December 2006)
18. Bobick, A., Davis, J.: The recognition of human movement using temporal tem-
plates. PAMI 23(3), 257–267 (2001)
19. Niebles, J., Wang, H., Wang, H., Fei Fei, L.: Unsupervised learning of human action
categories using spatial-temporal words. In: BMVC 2006, vol. III, p. 1249 (2006)
20. Belongie, S., Malik, J., Puzicha, J.: Shape context: A new descriptor for shape
matching and object recognition. In: NIPS, pp. 831–837 (2000)
21. Bissacco, A., Yang, M.H., Soatto, S.: Fast human pose estimation using appearance
and motion via multi-dimensional boosting regression. In: CVPR (2007)
22. Mikolajczyk, K., Schmid, C.: A performance evaluation of local descriptors. IEEE
Transactions on Pattern Analysis & Machine Intelligence 27, 1615–1630 (2005)
23. Chang, C., Lin, C.: Libsvm: a library for support vector machines, Online (2001)
24. Ke, Y., Sukthankar, R., Hebert, M.: Efficient visual event detection using volu-
metric features. In: International Conference on Computer Vision, vol. 1, p. 166
(October 2005)
25. Wong, S., Kim, T., Cipolla, R.: Learning motion categories using both semantic
and structural information. In: CVPR 2007, pp. 1–6 (2007)
26. Ali, S., Basharat, A., Shah, M.: Chaotic invariants for human action recognition.
In: ICCV 2007, pp. 1–8 (2007)
Temporal Dithering of Illumination
for Fast Active Vision
Srinivasa G. Narasimhan1 , Sanjeev J. Koppal1 , and Shuntaro Yamazaki2

1
The Robotics Institute, Carnegie Mellon University, USA
2
National Institute of Advanced Industrial Science and Technology, Japan
Abstract. Active vision techniques use programmable light sources,

such as projectors, whose intensities can be controlled over space and
time. We present a broad framework for fast active vision using Dig-
ital Light Processing (DLP) projectors. The digital micromirror array
(DMD) in a DLP projector is capable of switching mirrors “on” and
“off” at high speeds (106 /s). An off-the-shelf DLP projector, however,
effectively operates at much lower rates (30-60Hz) by emitting smaller
intensities that are integrated over time by a sensor (eye or camera)
to produce the desired brightness value. Our key idea is to exploit this
“temporal dithering” of illumination, as observed by a high-speed cam-
era. The dithering encodes each brightness value uniquely and may be
used in conjunction with virtually any active vision technique. We ap-
ply our approach to five well-known problems: (a) structured light-based
range finding, (b) photometric stereo, (c) illumination de-multiplexing,
(d) high frequency preserving motion-blur and (e) separation of direct
and global scene components, achieving significant speedups in perfor-
mance. In all our methods, the projector receives a single image as input
whereas the camera acquires a sequence of frames.
1 Introduction
Projectors are commonly used as programmable light sources for a variety of

active vision techniques including structured light range finding [1,2,3,4,5,6],
photometry-based reconstruction [7,8], relighting [9], light transport analysis
[10,11] and depth from defocus [12]. The intensity and color of the scene’s illu-
mination can be controlled over space and time depending on the task at hand.
For instance, projecting a set of colored striped patterns onto a scene alleviates
the problem of finding correspondences for 3D reconstruction [2].
Recently, Digital Light Processing (DLP) technology (http://www.dlp.com)
has enabled mass production of low cost projectors with high quality. The key
component of a DLP projector is the Digital Micromirror Device (DMD). Each
mirror in the DMD is 14 × 14 microns and can switch between two orientations,
+12o and −12o [13]. In one orientation, incident light is reflected by the mirror
toward the outside scene and in the other, light is reflected onto a black surface
within the projector. These mirrors can switch between orientations in a few
microseconds, enabling high precision control of illumination. As a result, the

Temporal Dithering of Illumination for Fast Active Vision 831
DMD device has found applications in areas ranging widely from microscopy to
chemistry to holographic displays [13].
The operating principle of the DMD device has also been exploited in com-
puter vision and graphics. Nayar et al. [14] re-engineer a DLP projector into
a DMD-camera and demonstrate the notion of programmable imaging for ap-
plications including adaptive dynamic range and optical filtering and matching.
Based on the theory of compressive sampling, a single pixel camera has been
implemented where the DMD device used to compute optical projections of
scene radiance [15]. Jones et al. [16] modify a DLP projector using custom made
FPGA-based circuitry to obtain 1-bit projection at 4800Hz. Using this, they
generate high speed stereoscopic light field displays. McDowall and Bolas [17]
use a specially re-programmed high speed projector based on Multiuse Light
Engine (MULE) technology to achieve range finding at kilohertz rates.
1.1 Temporal Dithering in a DLP Projector
In order to project a desired intensity value, the DLP projector emits a series of
light pulses of different time intervals [13]. A sensor aggregates the pulses of light
over the duration of its integration time (say, 1/30s in a video camera) to capture
the final gray-valued brightness. This Pulse-Width modulation (PWM) by the
projector is unique for every input intensity and can be termed as “temporal
dithering” of the illumination. As we shall show, this dithering allows us to
encode scene illumination in novel ways to achieve significant speedup in the
performance of virtually any active vision technique.
But how do we capture this high speed dithering? The exposure time (1/30s)
of a video camera is too long to observe the temporal illumination dithering
clearly. One possibility is to precisely synchronize the camera with a DLP pro-
jector and to expose the camera only for the duration of a single projected light
pulse (a few microseconds). Raskar et al [18] and Cotting et al [19] use this
technique to embed illumination patterns in the scene that cannot be observed
with the naked eye. The focus of these works is on intelligent office applications
with 30-60Hz performance requirements.
In contrast, our work focuses on exploiting the temporal dithering for fast ac-
tive vision. For this, we use a novel combination of a high speed camera and an
off-the-shelf DLP projector. Figure 1 illustrates the dithering of an 8-bit InFocus
IN38 DLP projector as observed by a Photron PCI-1024 high speed camera. A cal-
ibration image composed of 5 × 5 pixel blocks each with a different intensity value
from 0 to 255 is input to the projector. Each intensity at a pixel C in this calibra-
tion image is projected onto a flat screen using a unique temporal dithering DC (t),
over discrete time frames t. The high speed camera observes the projected im-
ages at 10 kHz. Notice the significant variation in the images recorded. The plot in
Figure 1(d) shows the patterns emitted by the projector for 4 input brightnesses
(165, 187, 215, 255), as measured over 100 camera frames. The temporal ditherings
corresponding to all the 256 input intensities in the calibration image are collated
into a photograph for better visualization of this principle. The temporal dithering
is stable and repeatable but varies for each projector-camera system.
832 S.G. Narasimhan, S.J. Koppal, and S. Yamazaki
Fig. 1. Reverse engineering a DLP Projector: (a) A DLP projector converts the input
intensity received into a stream of light pulses that is then projected onto a screen. A
high speed camera viewing the screen aggregates the brightness over the duration of
its integration time. (b) A calibration image composed of 5 × 5 pixel blocks each with
a different intensity from 0 to 255 is input to the projector. (c) The camera records the
projector output at 10 kHz. In (d) we show gray-valued intensities measured over time
by the high speed camera for 4 example intensities input to the projector. Notice the
significant variations in the plots. In (e), the temporal dithering for all 256 projector
input intensities is collated into an image. This temporal dithering is repeatable and
can be used to encode illumination in a novel way, enabling fast active vision.
1.2 Implications of Temporal Dithering

The high speed illumination modulation of a DLP projector can be exploited to
speed up a series of well-known active vision problems, making them applicable
to dynamic scenes. For each of these problems, we select a simple existing al-
gorithm to demonstrate our technique, although more sophisticated algorithms
may be used to achieve further speed up:
(a) The unique encoding of intensities allows us to obtain camera-projector
pixel correspondences allowing 3D reconstruction at high speeds.
(b) By multiplexing illumination from three projectors, we compute the surface
normals using photometric stereo [8] at high speeds.
(c) We de-multiplex illumination [20] from multiple projectors to capture the
appearances of a dynamic scene from different lighting directions.
(d) We demonstrate the ability to project high frequency complementary pat-
terns to separate the direct and global components [10] in a dynamic scene.
(e) We discuss motion blurring of an object illuminated by a DLP projector
and captured by a low frame rate camera (30-60 Hz). The temporal dithering
Fig. 2. Illumination and acquisition setup for structured light based 3D reconstruction:
The Photron high speed camera is placed vertically above the Infocus DLP projector.
A vertical plane is placed behind the scene (statue) for calibration.
preserves higher frequencies in the motion-blurred image. This is similar to

the work of Raskar et al [21] who demonstrate that fast camera shutter
modulation during image acquisition preserves higher frequencies.
In methods (a)-(d), the projector receives a single image as input via a com-
puter, whereas the high speed camera acquires a sequence of frames. The effective
speedup achieved depends on the task at hand and the quality of the result desired
given the signal-to-noise ratio in the captured images. In addition, the intensity
variation due to dithering can be observed reliably even with camera frame rates
as low as 300 fps enabling applications with slower performance requirements. Un-
like previous work, our techniques do not require any projector-camera synchro-
nization, hardware modification or re-programming of the DMD device, or the
knowledge of proprietary dithering coding schemes. Thus, we believe this work to
be widely applicable. Better visualizations of all our results are available through
our website (http://graphics.cs.cmu.edu/projects/dlp-dithering).
2 Projector-Camera Correspondence for Reconstruction
Structured light-based triangulation has commonly been used for 3D reconstruc-

tion [1]. A known set of spatio-temporally varying intensity patterns is projected
onto a scene and the reflected images are used to find the corresponding pixels
between the projector and the camera. The corresponding pixels are then tri-
angulated spatially (or by temporal analysis [4]) to obtain 3D structure. It is
assumed that the scene motion is negligible while the patterns are projected.
Since projectors have been operated at 30-60Hz, most implementations achieve
slower than real-time performances. Fast illumination dithering in a DLP pro-
jector enables high speed reconstruction.
Our goal is to obtain correspondences between the projector and camera pixels
at high speeds. Consider a high speed camera viewing a dynamic scene that is
illuminated by the DLP projector. A single image composed of a set of horizontal
lines of randomly chosen colors and intensities is input to the projector via a
Fig. 3. Results of 3D reconstruction using the DLP projector for a moving statue:
(a) Three frames captured by the high speed camera illustrate the fast modulation of
illumination incident on the scene. 20 continuous frames are used to match the inten-
sity variation observed on the scene point against the normalized intensity variation
observed on the vertical plane behind the object. (b) The best match finds correspon-
dences between projector and camera pixels. The error map is shown in (c). The (d)
disparity and (e) recovered shape of the object is shown from different viewpoints.
laptop. Let I(t) be the vector of intensities observed, over a set of frames, at a
scene point P . The normalized correlation between I(t) and temporal dithering
function DC (t) for each C (Section 1.1) is computed to obtain the projector pixel
C corresponding to the image pixel P . But how do we synchronize the frames
from the projector and the camera? One approach is to include a small planar
patch in the scene where correspondence between the corners of the patch can be
easily established (say, manually). This correspondence allows us to synchronize
the measured intensity vector with the temporal dithering.
We performed two experiments with a rotating statue and with a cloth waved
quickly in front of the high speed camera. For convenience, the camera and the
Fig. 4. 3D reconstruction of a cloth waved: (a) Twenty continuous images acquired

are used to reconstruct the 3D structure of a dynamically deforming cloth object. We
show three sample input images here taken at different times. (b) The reconstructed
cloth in different frames shows complex structure such as folds and creases that can
only be captured at high speeds.
projector are placed in a fronto-parallel configuration with a vertical baseline

(see Figure 2). The high speed camera is operated at 3kHz and the projector
is reverse engineered at this speed as described in Section 1.1. A homogeneous
vertical plane is used as the background behind the scene. The dithering DC (t)
can be captured from the pixels on this plane and simultaneously with the object.
Hence, in this setting, we simply correlate the normalized brightness vector I(t)
of a point on the object with the dithering vector DC (t) at every pixel C on the
background plane and no additional synchronization is needed.
Here, twenty frames were used to obtain correspondences, taking 20/3000 =
1/150s . In general, the number of frames necessary depends on the desired
matching accuracy and the SNR of the acquired images. By sliding the 20 frame
window across the full image sequence, 3D reconstruction can be obtained at
rates as high as 3 kHz (speed of camera). However, in practice, the rate of
reconstruction is lower considering the speed of the object’s motion. Figures 3
and 4 show the correspondences, the disparity and the reconstructions obtained
for the two scenes.
Note that the dithering can also be observed at lower frame rates and hence
a lower cost camera may be used for slower reconstructions. We repeated our
experiments for the same scene at four lower frame rates — 1500Hz, 600Hz,
Fig. 5. Reconstructions obtained using videos captured at reduced frame rates. Even
at 300Hz, the quality of the reconstruction obtained remains acceptable indicating that
temporal dithering can be exploited at this frame rate.
300Hz and 120Hz. Figure 5 shows the reconstructions obtained. The frame rate
of 120Hz is too low to capture the required intensity variation and hence, the
projector-camera pixel correspondences are unreliable. However, at 300Hz, the
reconstruction quality is still acceptable indicating that the temporal dithering
can be exploited even at this frame rate.
3 Illumination De-multiplexing for Appearance Capture

Acquiring scene appearance from multiple lighting directions is necessary for
image-based relighting and photometric recovery of scene properties (BRDF and
3D structure). In most works, the scene is assumed to be static and the acqui-
sition can take any where from a few seconds to several minutes. Using DLP
illumination, we capture the appearances of a dynamic scene from multiple light-
ing directions. For this, we draw upon the idea of illumination de-multiplexing
[20], where the images of the scene are simultaneously captured from multiple
source directions and de-multiplexed in software to obtain the desired images
under each lighting direction. This technique increases the signal-to-noise of the
captured images while keeping the number of captured images unchanged.
The difference between Schechner et al. [20] and our technique is in the cod-
ing: they use binary Hadamard codes, whereas we rely on the temporal dithering
of DLP illumination. The acquisition setup consists of three DLP projectors (In-
focus IN38 and LP120, and Sony XGA DataProjector) that simultaneously illu-
minate the scene from different directions. Since we wish to illuminate the scene
uniformly, a single constant brightness image is input to each of the projectors.
The three projectors differ in their brightness and contrast ratings and dither-
ing behaviors. The captured intensity at time instant t is written as a sum of
irradiances due to the illuminations from all projectors (k = 1 . . . 3):

3
I(t) = Dk (t)Ek (t) (1)
k=1
where, Dk (t) is the dithering intensity of the projector k at time t and Ek (t) is
the irradiance due to the scene as if illuminated only from projector k but with
Fig. 6. Demultiplexing illumination from three projectors to create appearances under

each lighting direction: The scene consists of a wiry polyhedral ball falling vertically.
Notice the mixing of the shadows in the multiplexed images in (a). A mirror sphere is
placed in the scene to measure the dithering intensities Dk from the three projectors.
(b) The results of demultiplexing for two frames in the video sequence. A total of 10
frames are used to demultiplex. Some ghosting artifacts are due to the noisy estimation
of source intensities Dk . The variation in the contrast and brightness of the three de-
multiplexed images are due to the quality of the three projectors. Projector 2 (InFocus
IN38) is the most recent one and has the highest quality.
unit intensity. The intensities Dk (t) can be obtained by observing a stationary

mirror sphere placed in the scene. The observed irradiances I(t) over time form
a linear system which is solved to obtain the appearances Ek (t) of the scene
from each individual lighting direction. In practice, since the projectors are not
synchronized when they illuminate the scene, the dithering intensities Dk vary
significantly over time, and hence the linear system is well-conditioned.
Figure 6 shows the results of applying the above approach to a scene with a
falling wiry ball. Notice the 3 shadows of the ball and the mirror sphere that
appear mixed in the multiplexed image I(t). For robustness, we use 10 frames
to solve the above linear system. Notice separation of the shadows in the de-
multiplexed images. As before, the effective rate of demultiplexing depends on
the SNR in the high speed camera. We have thus far ignored color information,
however, when the three DLP projectors emit intensities in different spectral
bands, the de-multiplexing algorithm can be used to colorize the acquired high
speed gray-scale video.
4 Illumination Multiplexing for Photometric Stereo
Photometric stereo is a widely used method to recover the surface normals and
albedos of objects that are photographed under different lighting directions.
There are many variants of this approach and we chose the one by Hertzmann
and Seitz [8] for its simplicity. In their work, the appearance of the scene under
varying lighting is matched with that of an example sphere made of the same
material (same BRDF) as the scene. The point on the sphere that produces the
best match is the normal of the scene point. We will extend this approach for
fast moving scenes that are simultaneously illuminated from different directions.
The scene in our experiments consists of a sphere and a falling pear both
painted in the same manner (Figure 7) and illuminated by three DLP projec-
tors simultaneously from different directions. The projectors and camera are far
enough away from the scene to assume orthographic viewing and distant lighting.
Since each projector must uniformly illuminate the scene, we provide a single
constant brightness image as input to each projector (with different brightness
values). The high speed camera records images at 3 kHz.
The projectors are de-synchronized and hence, the “multiplexed illumination”
results in significant variation in the observed intensities. The normalized inten-
sities at a scene point are compared to those observed on the sphere. The surface
normal of the scene point is that of the point on the sphere which produced the
best match. A matching length of 10 frames achieved robust results. A sliding
window of 10 frames can be used to generate the normals up to a rate of 3 kHz.
As before, the speed of the object determines the effective performance rate.
Figure 7 shows the normals of the pear as it falls and bounces on a table.
5 Complementary Patterns for Direct-Global Separation
The radiance of a scene point can be divided into two components - (a) the direct
component Ld , due to the direct illumination from the light source and (b) the
global component Lg due to the illumination indirectly reaching the scene point
from other locations in the scene [10]. The global component Lg includes effects
like interreflections, subsurface and volumetric scattering and translucency. Na-
yar et al [10] demonstrated that using high frequency illumination, it is possible
to separate the two components and obtain novel visualizations of the compo-
nents for the first time. A particular choice for high frequency illumination is
Fig. 7. Photometric stereo by example: The scene consists of a fast moving pear and a
sphere that are both painted similarly. Three DLP projectors simultaneously illuminate
the scene and the camera operates at 3000Hz. The projectors and camera are far
enough away from the scene to assume orthographic viewing and distant lighting. The
surface normal at a point on the falling pear is computed by matching the normalized
observed intensities to those at the points on the sphere. Since the projectors are
not synchronized, the variation in multiplexed illumination from the 3 projectors is
significant enough to obtain good matches for surface normals. A matching length of
10 frames achieved robust results.
a checker board pattern and its complement (with alternate bright and dark
squares), both of which are projected sequentially for separation.
We exploit illumination dithering to obtain separation at video rates. However,
in our setup, it is possible to input only one image to the DLP projector in 1/60s
and we have no control over the temporal dithering. So, how do we project comple-
mentary patterns much faster than 1/60s? We selected two specific input bright-
nesses 113 and 116 whose dithered patterns are shown in the plot of Figure 8.
Notice how the two patterns “flip” from bright to dark and vice versa over time.
Hence, a checker pattern with these two brightnesses are input to the projector.
The dithering ensures that the two complementary patterns occur at high speeds.
Let the observed temporally dithered values for input values 113 and 116 be a and
b, respectively, and the fraction of pixels that correspond to the value a be α (0.5
in our experiments). The two captured images are [10]:
L+ (x, y) = aLd + [(1 − α)b + αa]Lg

L− (x, y) = bLd + [(1 − α)a + αb]Lg . (2)
To solve the above equations, we need to know a and b in every frame. For this,
we place a white planar diffuse surface behind the scene of interest. For points on
this plane, Lg = 0 and Ld is a constant. This allows us to estimate a and b up to a
Fig. 8. Direct-Global Separation using DLP Dithering: (a) The DLP projector and
the camera are co-located using a beam splitter. A single checker pattern with two
intensities 113 and 116 are input to the projector. The plot shows how the input
intensities are dithered by the projector over time. Notice that at certain time instants,
the patterns flip between bright and dark. Thus, the projector emits complementary
checker patterns as in (b) onto the scene that are used to separate the direct and global
components (c). The flip occurs once in 1/100s.
single scale factor. Then, the above linear system can be solved at every pixel to
obtain the separation. There is one additional complication in our setup beyond
the method in [10]: it is hard to find out whether a scene point receives intensity
a or intensity b from just the observed appearance of the scene. To address this
problem, we co-locate the projector and the camera using a beam-splitter as
shown in Figure 8. The pixels of the projector are automatically corresponded
with those of the camera.
The scene in our experiment consists of a set of white ping-pong balls dropped
from a hand. The ping-pong balls are mostly diffuse. Notice that the direct
Fig. 9. Motion blurring under DLP illumination and fluorescent illumination: The
scene consists of a heavy brick falling rapidly and an image is captured with exposures
1/60s (a) and 1/125s (b). Under fluorescent illumination, the motion blur appears
as a smear across the image losing high frequencies. The temporal dithering in DLP
projectors acts as a high frequency modulator that convolves with the moving object.
The motion-blurred image still preserves some of the high spatial frequencies. Six copies
of the text “ECCV08” in (a) and 2 copies in (b) are clearly visible.
component for each ball looks like the shading on a sphere (with dark edges) and
the indirect component includes the interreflections between the balls (notice the
bright edges). For the hand, the direct component is only due to reflection by the
oils near the skin surface and is dark. The indirect component includes the effect
of subsurface scattering and dominates the intensity. The checker pattern “flips”
once in approximately 1/100s and hence we achieve separation at 100Hz. Due
to finite resolution of the camera and the narrow depth of field of the projector,
a 1-pixel blur is seen at the edges of the checker pattern. This results in the grid
artifacts seen in the results.
6 Flutter Flash: Motion-Blur under DLP Illumination
Motion-blur occurs when the scene moves more than a pixel within the inte-
gration time of a camera. The blur is computed as the convolution of the scene
motion with a box filter of width equal to the camera integration time. Thus,
images captured of fast moving objects cause a smear across the pixels losing
significant high frequencies. Deblurring images is a challenging task that many
works have addressed with limited success. A recent approach by Raskar et al.
[21] uses an electronically controlled shutter in front of the camera to modulate
the incoming irradiance at speeds far greater than the motion of the object. In
other words, the box filter is replaced by a series of short pulses of different
widths. The new convolution between the object motion and the series of short
pulses results in images that preserve more high frequencies as compared to the
box filter. This “Flutter Shutter” approach helps in making the problem better
conditioned. Our approach is similar in spirit to [21] with one difference: the fast
shutter is simulated by the temporal dithering of the DLP illumination. Note that
the DLP illumination dithering is significantly faster than mechanical shutters1 .
Figure 9 shows the images captured by with 1/60s exposure. The scene consists
of a brick with the writing “ECCV08” falling vertically. When illuminated by
a fluorescent source, the resulting motion-blur appears like a smear across the
image. On the other hand, when the scene is illuminated using a DLP projector,
we see 6 distinct copies of the text that are translated downward. A Canny edge
detector is applied to the captured image to illustrate the copies. If we knew
the extent of motion in the image, the locations of strong edges can be used as
a train of delta signals that can be used for deblurring the image. In 9(b), we
show an example of deblurring the image captured with 1/125s exposure. As
in the deblurred images obtained the flutter shutter case, the DLP illumination
preserves more high frequencies in the motion-blurred image.
7 Discussion
Speed vs. accuracy trade-off. One limitation of our approach is the require-
ment of a high speed camera. The acquisition speed of the camera and the
effective speed of performance achieved depend on the task at hand and the
signal-to-noise ratio of the captured images. For instance, the decision to use
10 frames for demultiplexing illumination or photometric stereo, or to use 20
frames for structured light, was mainly influenced by the noise characteristics of
the camera. A more scientific exploration of this trade-off is required to better
understand the benefits of our approach to each technique. A future avenue of
research is to design 2D spatial intensity patterns that create temporal dithering
codes that are optimal for the task at hand.
Issues in reverse engineering. The images shown in Figure 1 are dark for the
input brightness range of 0 to 90. Despite the claim from manufacturers that the
projector displays 8-bits of information, only about 160 patterns are usable for
our experiments. To compensate for this, the projector performs spatial dithering
in addition to temporal dithering in a few pixel blocks. This is an almost random
effect that is not possible to reverse engineer without proprietary information
from the manufacturers. We simply average a small neighborhood or discard
such neighborhoods from our processing.
Other active vision techniques and illumination modulations. We be-
lieve that the temporal illumination dithering can be applied to a broader range
of methods including pixel-wise optical flow estimation and tracking, projector
defocus compensation and depth from defocus [12] and spectral de-multiplexing.
While we exploit the temporal dithering already built-in to the projector, we do
not have a way of controlling it explicitly. Better control is obtained by using
a more expensive and special high speed MULE projector [17]. Finally, strobe
lighting, fast LED [22] and flash modulation are also effective in temporally
varying (not dithering) the illumination.
1
Faster shutters can be realized by electronically triggering the camera.
Acknowledgements
This research was supported in parts by ONR grants N00014-08-1-0330 and
DURIP N00014-06-1-0762, and NSF CAREER award IIS-0643628. The authors
thank the anonymous reviewers for their useful comments.
References
1. Will, P.M., Pennington, K.S.: Grid coding: A preprocessing technique for robot
and machine vision. AI 2 (1971)
2. Zhang, L., Curless, B., Seitz, S.M.: Rapid shape acquisition using color structured
light and multi-pass dynamic programming. 3DPVT (2002)
3. Davis, J., Nehab, D., Ramamoothi, R., Rusinkiewicz, S.: Spacetime stereo: A uni-
fying framework for depth from triangulation. In: IEEE CVPR (2003)
4. Curless, B., Levoy, M.: Better optical triangulation through spacetime analysis. In:
ICCV (1995)
5. Young, M., Beeson, E., Davis, J., Rusinkiewicz, S., Ramamoorthi, R.: Viewpoint-
coded structured light. In: IEEE CVPR (2007)
In: CVPR (2003)
7. Zickler, T., Belhumeur, P., Kriegman, D.J.: Helmholtz stereopsis: Exploiting reci-
procity for surface reconstruction. In: Heyden, A., Sparr, G., Nielsen, M., Johansen,
P. (eds.) ECCV 2002. LNCS, vol. 2352, pp. 869–884. Springer, Heidelberg (2002)
8. Hertzmann, A., Seitz, S.M.: Shape and materials by example: A photometric stereo
approach. In: IEEE CVPR (2003)
9. Wenger, A., Gardner, A., Tchou, C., Unger, J., Hawkins, T., Debevec, P.: Perfor-
mance relighting and reflectance transformation with time-multiplexed illumina-
tion. ACM SIGGRAPH (2005)
10. Nayar, S.K., Krishnan, G., Grossberg, M.D., Raskar, R.: Fast separation of di-
rect and global components of a scene using high frequency illumination. ACM
SIGGRAPH (2006)
11. Sen, P., Chen, B., Garg, G., Marschner, S.R., Horowitz, M., Levoy, M., Lensch,
H.P.A.: Dual photography. ACM SIGGRAPH (2005)
12. Zhang, L., Nayar, S.K.: Projection defocus analysis for scene capture and image
display. ACM SIGGRAPH (2006)
13. Dudley, D., Duncan, W., Slaughter, J.: Emerging digital micromirror device (dmd)
applications. In: Proc. of SPIE, vol. 4985 (2003)
14. Nayar, S.K., Branzoi, V., Boult, T.: Programmable imaging using a digital mi-
cromirror array. In: IEEE CVPR (2004)
15. Takhar, D., Laska, J., Wakin, M., Duarte, M., Baron, D., Sarvotham, S., Kelly,
K., Baraniuk, R.: A new compressive imaging camera architecture using optical-
domain compression. Computational Imaging IV at SPIE Electronic Imaging
(2006)
16. Jones, A., McDowall, I., Yamada, H., Bolas, M., Debevec, P.: Rendering for an
interactive 360 degree light field display. ACM SIGGRAPH (2007)
17. McDowall, I., Bolas, M.: Fast light for display, sensing and control applications.
In: IEEE VR Workshop on Emerging Display Technologies (2005)
18. Raskar, R., Welch, G., Cutts, M., Lake, A., Stesin, L., Fuchs, H.: The office of
the future: A unified approach to image-based modeling and spatially immersive
displays. ACM SIGGRAPH (1998)
19. Cotting, D., Naef, M., Gross, M., Fuchs, H.: Embedding imperceptible patterns
into projected images for simultaneous acquisition and display. In: ISMAR (2004)
20. Schechner, Y.Y., Nayar, S.K., Belhumeur, P.N.: A theory of multiplexed illumina-
21. Raskar, R., Agrawal, A., Tumblin, J.: Coded exposure photography: Motion de-
blurring using fluttered shutter. ACM SIGGRAPH (2006)
22. Nii, H., Sugimoto, M., Inami, M.: Smart light-ultra high speed projector for spatial
multiplexing optical transmission. In: IEEE PROCAMS (2005)
Compressive Structured Light for Recovering
Inhomogeneous Participating Media
Jinwei Gu, Shree Nayar, Eitan Grinspun,

Peter Belhumeur, and Ravi Ramamoorthi
Columbia University, New York, NY 10027, USA

jwgu@cs.columbia.edu
Abstract. We propose a new method named compressive structured

light for recovering inhomogeneous participating media. Whereas con-
ventional structured light methods emit coded light patterns onto the
surface of an opaque object to establish correspondence for triangulation,
compressive structured light projects patterns into a volume of partici-
pating medium to produce images which are integral measurements of the
volume density along the line of sight. For a typical participating medium
encountered in the real world, the integral nature of the acquired images
enables the use of compressive sensing techniques that can recover the
entire volume density from only a few measurements. This makes the
acquisition process more efficient and enables reconstruction of dynamic
volumetric phenomena. Moreover, our method requires the projection of
multiplexed coded illumination, which has the added advantage of in-
creasing the signal-to-noise ratio of the acquisition. Finally, we propose
an iterative algorithm to correct for the attenuation of the participating
medium during the reconstruction process. We show the effectiveness of
our method with simulations as well as experiments on the volumetric
recovery of multiple translucent layers, 3D point clouds etched in glass,
and the dynamic process of milk drops dissolving in water.
1 Introduction
Structured light has a long history in the computer vision community [1]. It has
matured into a robust and efficient method for recovering the surfaces of objects.
By projecting coded light patterns on the scene, and observing it using a camera,
correspondences are established and the 3D structure of the scene is recovered
by triangulation. Over the years, researchers have developed various types of
coding strategies, such as binary codes, phase shifting, spatial neighborhood
coding, etc. All structured light range finding approaches are based on a common
assumption: Each point in the camera image receives light reflected from a single
surface point in the scene.
However, many real-world phenomena can only be described by volume den-
sities rather than boundary surfaces. Such phenomena are often referred to as
participating media. Examples include translucent objects, smoke, clouds, mix-
ing fluids, and biological tissues. Consider an image acquired by photographing

846 J. Gu et al.
a volume of a participating medium. Unlike in the case of an opaque object,

here each pixel receives scattered light from all points along the line of sight
within the volume. Narasimhan et al. [2] have addressed the problem of recov-
ering opaque surfaces immersed in a participating medium (rather than a clear
medium) using structured light range finding. The goal in this case is to make
existing structured light range finding methods robust to light scattering by the
medium, rather than recover the medium itself.
The problem of recovering the volume density of a participating medium (in
particular, smoke), was addressed by Hawkins et al. [3]. They used a high-powered
laser sheet and a high-speed camera (5000fps) to measure thin slices of a smoke
density field via scanning, which is similar to a technique termed laser-induced
fluorescence (LIF) in the fluid imaging community [4]. Fuchs et al. [5] proposed
the idea of shooting a set of static laser rays into the volume and using spatial in-
terpolation to reconstruct the volume. However, the measurements are inherently
sparse in this case and hence the recovered density is low in resolution.
In this paper, we show that by using coded light patterns, one can make the
measurement of a participating medium highly efficient in terms of acquisition
time as well as illumination power. In particular, we exploit the fact that the
brightness measurements made at image pixels correspond to true line-integrals
through the medium (see Fig. 1a), and then solve for its volumetric density. We
consider both spatially- and temporally-coded light patterns. Because the pat-
terns are predetermined, measurement and reconstruction time are decoupled.
We target low-density inhomogeneous media, for which the density function is
sparse in an appropriately-chosen basis1 ; this allows us to harness compressive
sensing techniques [6,7] that accurately reconstruct a signal from only a few
measurements. We refer to our approach as compressive structured light.
We show that compressive structured light is more economical than a straight-
forward sequential scanning of a volume. Whereas the sampling rate of the latter
is limited by the desired resolution, the sampling rate of the former is restricted
by the sparsity of the data–a considerably more relaxed constraint for low-density
phenomena. Since our approach requires fewer measurements, it naturally en-
ables the recovery of dynamic participating media. An added advantage of com-
pressive structured light, is that it requires the projection of multiplexed coded
illumination which results in measurements with higher signal-to-noise ratio [8].
An important practical consequence is that light sources of significantly lower
power than in the case of sequential scanning can be used.
We have implemented our approach using a digital projector and a camera as
shown in Fig. 1c. The projector and the camera are synchronized and both oper-
ate at 360fps. Using 24 coded light patterns, we are able to recover a 1283 volume
at 15fps. Using this system, we have recovered various types of inhomogeneous
participating media, as shown in §7.
1
“sparse” does not necessarily imply that the volume density must be sparsely dis-
tributed in space. It means that the density can be represented with a few non-zero
coefficients in an appropriately-chosen basis, such as, wavelets, gradients, principal
components, etc.
Compressive Structured Light 847
Participating
Medium
m
y
I y,z ρ ( x, y, z) Projector
Projector
I(y,z) L(x,y)
Camera
x
Milk
Camera Drops
n
z L x,y
(a) (b) (c)
Fig. 1. (a) Compressive structured light for recovering inhomogeneous participating

media. Coded light is emitted along the z-axis to the volume while the camera ac-
quires images as line-integrated measurements of the volume density along the x-axis.
The light is coded in either the spatial domain or temporal domain with a predeter-
mined sequence. We reconstruct the volume density from the measurements by using
compressive sensing techniques. (b) Image formation model for participating medium
under single scattering. The image irradiance at one pixel, I(y, z), depends on the inte-
gral along the x-axis of the projector’s light, L(x, y), and the medium density, ρ(x, y, z),
along a ray through the camera center; refer to (1)–(2). (c) Experiment setup.
2 Related Work
Compressive Sensing. Compressive sensing [6,7] is a nascent field of applied
mathematics with a variety of successful applications including imaging [9], med-
ical visualization [10], and face recognition [11]. It offers a theoretical framework
to reconstruct “sparse” signals from far fewer samples than required by the con-
ventional Shannon sampling theorem. Our work builds on the basic formulation
of compressive sensing, which we augment with auxiliary terms specific to the
reconstruction of volume density.
Reconstruction of Volumetric Phenomena. There are several recent works in

reconstruction of volumetric phenomena from multiple views. Hasinoff et al. [12]
used two views to reconstruct flames by assuming flames are surfaces in the
volume. Based on tomographic algorithms, Ihrke et al. [13,14] and Trifonov et
al. [15] used eight views and 72–360 views, respectively, for recovering flames (and
smoke) as well as transparent objects. We mentioned light-based methods [3,4,5]
earlier in the previous section. For a comprehensive survey of works in this area,
see Ihrke et al. [16].
Multiplexed Illumination. Our work is also related to multiplexed illumination [8]

in that both use coded light as illumination. However, there is a fundamental
difference: Whereas the conventional multiplexing aims at increasing signal-to-
noise ratio of the measurements, our work aims at increasing the efficiency of
the acquisition process, i.e., to reconstruct high dimensional signals from a few
measurements. In summary, both the coding strategies and the reconstruction
algorithms are different.
848 J. Gu et al.
3 Background on Compressive Sensing

In its simplest form, compressive sensing seeks a solution of the underdetermined
linear system Ax = b, where x ∈ Rn is a sparse signal, A is an m × n matrix
(“measurement ensemble”), and b is the vector of m measurements, with m < n.
Compressive sensing theory asserts that one can recover the signal from far
fewer measurements than the dimension of the signal, if the signal is sparse—it is
represented with few non-zero coefficients in a suitable basis—and the measure-
ments are uncorrelated, in the sense that each measurement is an inner product
of the signal with a test function that has a necessarily dense representation in
the chosen basis. Given a measurement ensemble matrix A, compressive sam-
pling theory predicts that x is the minimizer of ||x||1 , subject to Ax = b.
The above reconstruction strategy has been shown to work well for sparse
signal estimation, even from a noisy measurement [17]. In our work, we augment
the basic problem above with auxiliary terms that enforce the nonnegative con-
straint for the signal, and that exploit the sparsity not only of the signal value
but also its gradient. Indeed, our evaluation (§6, §7) indicates that using the
sparsity of the gradient is important for accurate and efficient reconstruction.
4 Image Formation Model

In this section, we derive the relationship between the volume density ρ and
the image irradiance I of the camera under our camera/projector setting. We
focus on non-emissive participating media with relatively low density in which
multiple scattering is assumed to be negligible.
As shown in Fig. 1b, each camera pixel receives light scattered from a row of
voxels along the line of sight in the volume (i.e., the red line in Fig. 1b). Consider
one such voxel. Before the light scattered by this voxel reaches the camera, it is
first attenuated as it travels from the projector to the voxel, scattered at the voxel,
and then attenuated as it travels from the voxel to the camera. Under the assump-
tion of single scattering, the final radiance sensed by the camera from this partic-
ular voxel is [18]: L(x, y) · exp(−τ1 ) · σs · ρ(x, y, z) · p(θ) · exp(−τ2 ), where ρ(x, y, z)
is the participating medium’s volume density at the voxel, p(θ) is the phase func-
tion, and τ1 = σt s1 ρds1 and τ2 = σt s2 ρds2 are the “optical distance” from the
projector to the voxel and from the voxel to the camera, respectively; σs and σt
are the scattering coefficient and the attenuation coefficient of the participating
medium [18]. Since p(θ) is the same for all voxels under orthographic projection
for both the camera and the projector, the above formula can be simplified to
(up to a scale related to p(θ) and σs ): L(x, y) · exp (−(τ1 + τ2 )) · ρ(x, y, z). The
image irradiance, I(y, z), which is the integral of the scattered light from all the
voxels along the line, is therefore

I(y, z) = L(x, y) · exp (−(τ1 + τ2 )) · ρ(x, y, z)dx . (1)
x
In the low density case, or when σt is relatively small compared with the scat-
tering, the effect of attenuation usually can be ignored [3,5], i.e., the exponential
term in the above equation is equal to 1. Equation (1) thus can be reduced to a
linear projection of the light and the volume density,

I(y, z) = ρ(x, y, z) · L(x, y) dx . (2)
x
For media where the attenuation cannot be ignored, we present a simple, iterative
method based on iterative relinearization (see §5.3).
5 Compressive Structured Light
In this section, we explain the idea of compressive structured light for recovering
inhomogeneous participating media. For participating media, each camera pixel
receives light from all points along the line of sight within the volume. Thus
each camera pixel is an integral measurement of one row of the volume density.
Whereas conventional structured light range finding methods seek to triangulate
the position of a single point, compressed structured light seeks to reconstruct
the 1D density “signal” from a few measured integrals of this signal.
This is clearly a more difficult problem. One way to avoid this problem is
to break the integrals into pieces which can be measured directly. The price,
however, is the deterioration of either spatial resolution or temporal resolution
of the acquisition. Existing methods either illuminate a single slice at a time and
scan the volume (see Fig. 2a and [4,3]), thus sacrificing temporal resolution, or
they illuminate a single pixel per row and use interpolation to reconstruct the
volume (e.g., Fig. 2b and [5]), sacrificing spatial resolution.
In contrast, the proposed compressive structured light method uses the light
much more efficiently, projecting coded light patterns that yield “signatures,” or
integral measurements, of the unknown volume density function.
The didactic illustration in Fig. 1a depicts a simple lighting/viewpoint geom-
etry under orthographic projection, with the camera viewpoint along the x-axis,
and the projector emitting along the z-axis. Consider various coding strategies
y
e
e
tim
tim
Camera
z (a) x (b) (c) (d)
Fig. 2. Different coding strategies of the light L(x, y) at time t for recovering inho-
mogeneous participating media: (a) scan (one stripe turned on) [4,3]; (b) laser-lines
interpolation (one pixel turned on per one row) [5]; (c) Spatial coding of compressive
structured light (all pixels are turned on with random values per time frame); (d) Tem-
poral coding of compressive structured light (random binary stripes are turned on per
time frame). Compressive structured light, shown in (c) and (d), recovers the volume
by reconstructing the 1D signal along x-axis from a few integral measurements.
850 J. Gu et al.
of the 3D light function L(x, y, t): Spatial codes (Fig. 2c) recover the volume
from a single image by trading spatial resolution along one dimension; Tempo-
ral codes (Fig. 2d) trade temporal resolution by emitting a sequence of vertical
binary stripes (with no coding along y-axis), so that full spatial resolution is
retained.2
We will see that these compressive structured light codes yield high efficiency
both in acquisition time and illumination power; this comes at the cost of a more
sophisticated reconstruction process, to which we now turn our attention.
5.1 Formulation
Consider first the case of spatial coding. Suppose we want to reconstruct a volume
at the resolution n×n×n (e.g., n = 100). The camera and the projector have the
resolution of M ×M pixels (e.g., M = 1024). Therefore, one row of voxels along
the x-axis (refer to the red line in Fig. 1a) will receive light from m = M/n (e.g.,
m = 1024/100 ≈ 10) rows of the projector’s pixels. The light scattered by these
voxels in the viewing direction will then be measured, at each z-coordinate,
by a vertical column of m camera pixels. Thus, using the fact that we have
greater spatial projector/camera resolution than voxel resolution, we can have
m measurements for each n unknowns. Similarly, we can also acquire these m
measurements using temporal coding, i.e., changing the project light patterns at
each of the m time frames.
Without loss of generality, we use l1 = L(x, 1),· · · ,lm = L(x, m) to denote the
m rows of pixels from the projector, and b1 = I(1, z),· · · ,bm = I(m, z) to denote
the image irradiance of the m pixels in the camera image. Let x = [ρ1 ,· · · ,ρn ]T
be the vector of the voxel densities along the row. Assuming no attenuation, the
image irradiance for each of these m pixels is a linear projection of the light and
the voxels’ density from (2): bi = lTix, i = 1,· · · ,m. Rewriting these m equations
in matrix form, we have: Ax = b, where A = [l1 ,· · · ,lm ]T is a m × n matrix,
T
b = [b1 ,· · · ,bm ] is a m×1 vector.
Thus, if attenuation is not considered, the problem of recovering the volume is
formulated as the problem of reconstructing the 1D signal x given the constraints
Ax = b. To retain high spatial and temporal resolution, we often can only afford
far fewer measurements than the number of unknowns, i.e., m < n, which means
the above equation is an underdetermined linear system and optimization is
required to solve for the best x according to certain priors.
One benefit of this optimization-based reconstruction is high efficiency in
acquisition, which we quantify using the measurement cost, m/n, where m is
the number of the measurements and n is the number of unknowns (i.e., the
dimension of the signal). For example, the measurement cost of the scanning
method [4,3] is one. We show that by exploiting the sparsity of the signal, we
can reconstruct the volume with much lower measurement cost (about 18 to 14 ).
2
All of the 4 methods shown in Fig. 2 can be equally improved using color channels.
Table 1. Different norms used for reconstruction
Method Optimization Functional Constraints

Least Square (LS) ||Ax − b||2
Nonnegative Least Square (NLS) ||Ax − b||2 x≥0
CS-Value ||x||1 Ax = b, x ≥ 0
CS-Gradient ||x ||1 Ax = b, x ≥ 0
CS-Both ||x||1 + ||x ||1 Ax = b, x ≥ 0
5.2 Reconstruction Via Optimization

Formulation. Solving the underdetermined linear system requires some prior
(assumed) knowledge of the unknown signal, which can be represented as opti-
mization functionals or constraints on the data. We consider several alternatives,
as listed in Table 1. Besides the commonly-used Least Square (LS) and Nonneg-
ative Least Square (NLS), we consider functionals using #1 -norms, as these bias
toward sparse representations:3
First, we observe that for many natural volumetric phenomena, often only
a small portion of the entire volume is occupied by the participating media.
For example, consider the beautiful ribbon patterns generated by smoke; simi-
larly, sparsity was implicitly used to reconstruct (surface-like) flames [12]). This
suggests the use of the #1 -norm of the signal value (CS-Value).
Furthermore, the sparsity of gradients of natural images is well studied [20,21].
Related work in image restoration [22] uses nonlinear optimization to minimize
“total variation,” i.e., the sum of #2 -norm of image gradient. In this vein, we
consider the use of #1 -norm on the signal’s gradient (CS-Gradient).
Finally, consider a dynamic process, such as milk dissolving in water: here
diffusion decreases the signal value’s sparsity over time, but it increases the
gradient sparsity. Motivated by this observation, we consider the sum of #1 -
norms of both the value and the gradient (CS-Both), so that the algorithm has
the ability to “adapt” for the sparsity.
Analysis. Comparison of these reconstruction methods is first performed on 1D

synthetic signals. These signals are randomly sampled rows from the volume
density of smoke acquired in Hawkins et al. [3]. We restrict the measurement
cost, m/n, to be 1/4. The measurement ensemble, A, is generated in a way
that each element is drawn independently from a normal distribution and each
column is normalized to 1, which is effectively a white noise matrix and is known
to be good for compressive sensing [7]. NRMSE (normalized root mean squared
error) is used as the measure of error.
The reconstruction results are shown in Fig. 3. The commonly-used LS per-
forms the worst, since it merely minimizes the errors without using any prior on
the data. With the nonnegative constraint added, NLS has better performance.
CS-Value and CS-Gradient are better than NLS given that both use one more
3
LS and NLS are solved with SVD and Levenberg-Marquardt, respectively. The other
functionals are formulated as Linear Programming (LP) and solved with GLPK [19].
852 J. Gu et al.
Ground Truth LS NLS CS-Value CS-Gradient CS-Both
0.18 0.18 0.18 0.18 0.18 0.18
0.16
0.14
ρ 0.16
0.14
Reconstruction 0.16
0.14
0.16
0.14
0.16
0.14
0.16
0.14
0.12 0.12 0.12 0.12 0.12 0.12
0.1 0.1 0.1 0.1 0.1 0.1
0.08 0.08 0.08 0.08 0.08 0.08
0.06 0.06 0.06 0.06 0.06 0.06
0.04 0.04 0.04 0.04 0.04 0.04
0.02 0.02 0.02 0.02 0.02 0.02
0 0 0 0 0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
x NRMSE = 0.330 0.177 0.026 0.007 0.001

0.7 0.7 0.7 0.7 0.7 0.7
0.6 0.6 0.6 0.6 0.6 0.6
0.5 0.5 0.5 0.5 0.5 0.5
0.4 0.4 0.4 0.4 0.4 0.4
0.3 0.3 0.3 0.3 0.3 0.3
0.2 0.2 0.2 0.2 0.2 0.2
0.1 0.1 0.1 0.1 0.1 0.1
0 0 0 0 0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
0.272 0.076 0.052 0.014 0.005

0.35 0.35 0.35 0.35 0.35 0.35
0.3 0.3 0.3 0.3 0.3 0.3
0.25 0.25 0.25 0.25 0.25 0.25
0.2 0.2 0.2 0.2 0.2 0.2
0.15 0.15 0.15 0.15 0.15 0.15
0.1 0.1 0.1 0.1 0.1 0.1
0.05 0.05 0.05 0.05 0.05 0.05
0 0 0 0 0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
0.266 0.146 0.053 0.024 0.021
Fig. 3. Comparison of different reconstruction methods. The first column is the orig-
inal signal. The remaining columns show reconstruction results (red dashed lines) for
different methods, given the measurement cost, m/n, is equal to 1/4. The value below
each plot is the NRMSE(normalized root mean squared error) of reconstruction.
prior—the sparsity on the signal value or on the signal gradient. The fact that
CS-Gradient is better than CS-Value indicates that the sparsity on the signal
gradient holds better than the sparsity on the signal value. Finally, as expected,
CS-Both outperforms other methods due to its adaptive ability. In our trials,
the favorable performance of CS-Both was not sensitive to changes of the rela-
tive weighting of the value and gradient terms. These observations carry over to
the 3D setting (see Fig. 4), where we reconstruct a 1283 volume; note that this
requires 128 × 128 independent 1D reconstructions.
5.3 Iterative Attenuation Correction

Until now, we have not considered the attenuation in the image formation model
in (1) yet. To take into account attenuation, we use a simple iterative relineariza-
tion algorithm as follows:
1. Assume no attenuation, solve the optimization problem with techniques from
§5.2 to get the initial reconstruction of the volume density ρ(0) .
2. At iteration k, assuming σt is known4 , compute the attenuated light as:
L(k) (x, y, z) = exp (− (τ1 + τ2 ))·L(x, y), where τ1 and τ2 are computed using
ρ(k−1) as shown in §4.
3. With the attenuated light L(k) (x, y, z), (1) becomes a linear equation. We
solve for ρ(k) and go to next iteration until it converges.5
Since our overall framework accommodates the scanning method [4,3] and the
interpolation method [5] as special cases, the iterative algorithm could be directly
applied to these prior methods as well.
4
The attenuation coefficient, σt , of the participating medium can be obtained from
literature, specified by a user, or be measured by a second camera taking the shad-
owgram of the volume.
5
In practice, we found that the algorithm usually converges within 3-4 iterations.
6 Validation Via Simulation
To further validate our method, we perform simulations on a synthetic volume.

The volume is generated from a triangular mesh of a horse and it is discretized
into 1283 voxels. For each voxel, if it is inside the mesh, the density is designed
to be proportional to the distance from the center of the voxel to the center
of the mesh, otherwise the density is 0. Fig. 4a shows the volume where blue
corresponds to the lowest density while yellow corresponds to the highest density.
A slice of the volume is shown in Fig. 4b.
Both spatial coding and temporal coding of compressive structured light are
tested. The measurement cost, m/n, is fixed to 1/4. For spatial coding, we use a
random color image with resolution of 1280×1280 as the coded light from the
projector. This gives us m = 1280/128×3 = 30 measurements to recover densities
of 128 voxels on one row of the volume. Based on (1), a single image (shown
in Fig. 4c) is generated from the camera view and used for reconstruction. For
temporal coding, we use random binary stripes as illumination and generate 32
images for reconstruction. One of these images is shown in Fig. 4g. CS-Both is
used to reconstruct the volume for both cases. As shown in Fig. 4, both meth-
ods accurately reconstruct the volume. Moreover, Fig. 4(right) shows the recon-
struction errors and reconstructed slices at different iterations of attenuation
correction, which demonstrates the effectiveness of the iterative algorithm.
We also evaluate different reconstruction methods at various measurement
costs from 1/16 to 1. The results are shown as a table in Fig. 5. Conclusions
similar to the ones from the previous 1D signal simulation (Fig. 3) can be drawn
from these results: (1) As expected, all methods have improvements as the mea-
surement cost increases. (2) Without using any prior of the data, LS is the worst
for reconstruction with insufficient measurements. (3) CS-Gradient and CS-Both
0.03
Error (NRMSE)
Spatial Coding of Compressive Structured Light
y
0.02
0.01
z
x
0 1 4 7 10 13 16
(a) Ground Truth (c) Coded Image (d) Reconstructed Slice (e) Reconstructed Volume at 2 Views (a) Iterations
Temporal Coding of Compressive Structured Light
Ground Truth Iteration 1 Iteration 2 Iteration 3

(b) Sample Slice (g) Coded Image (h) Reconstructed Slice (i) Reconstructed Volume at 2 Views
(b)
Fig. 4. Simulation results of volume reconstruction using compressive structured light.

LEFT: (a) The original volume where blue means the lowest density and yellow means
the highest density. (b) A slice of the volume. The top and the bottom row on the right
shows the reconstruction results for spatial coding and temporal coding, respectively.
For each row, from left to right are the coded image acquired by the camera, the
reconstruction of the slice, and the reconstructed volume under two different views.
RIGHT: (a) Reconstruction errors and (b) slices with iterative attenuation correction.
854 J. Gu et al.
m/n 1/16 1/8 1/4 1/2 1
LS
NLS
CS-Value
CS-Gradient
CS-Both
Fig. 5. Comparison of different reconstruction methods at different measurement costs,

m/n. CS-Both outperforms other methods.
largely outperform other methods, especially for low measurement cost, which
indicating strong sparsity in the signal’s gradient. (4) CS-Both is better than
CS-Gradient, especially at low measurement cost (e.g., as shown in Fig. 5 at
m/n = 1/16). Based on these preliminary simulations, we chose to run our ac-
tual acquisition experiments with a measurement cost of 1/4 and the CS-Both
optimization functional.
We have implemented the temporal coding of compressive structured light for

recovering inhomogeneous participating media. The spatial coding is not im-
plemented currently due to its extensive calibration requirement, as discussed
in §8. As shown in Fig. 1c, our system consists of a 1024 × 768 DLP projec-
tor and a 640×480 Dragonfly Express 8-bit camera, positioned at right angles,
both viewing the inhomogeneous participating medium (milk drops in water).
The projector and the camera are synchronized and both operate at 360fps.6
Using 24 coded light patterns, we are able to recover a 1283 volume at 15fps.
These light patterns consist of 128 vertical stripes. Each stripe is assigned 0 or
1 randomly with the probability of 0.5. In this way, about half amount of the
light is turned on for each measurement. We also tried alternative light pat-
terns such as Hadamard codes, and found the random binary codes have better
performance.
6
The camera’s resolution is set to 320×140 in order to achieve 360fps.
Without Attenuation Correction Without Attenuation Correction
(a) Photograph (c) View 1 View 2 View 3 (a) Photograph (c) View 1 View 2 View 3
With Attenuation Correction With Attenuation Correction
(b) Coded Image (d) View 1 View 2 View 3 (b) Coded Image (d) View 1 View 2 View 3
Fig. 6. Reconstruction results of LEFT: an object consisting of two glass slabs with
powder where the letters “EC” are on the back slab and “CV” on the front slab, and
RIGHT: point cloud of a face etched in a glass cube. Both examples show: (a) a
photograph of the objects, (b) one of the 24 images captured by the camera, and re-
constructed volumes at different views with (c) and without (d) attenuation correction.
We used this system to recover several types of inhomogeneous participating

media, including, multiple translucent layers, a 3D point cloud of a face etched
in a glass cube, and the dynamic process of milk mixing with water. The recon-
structed volumes are visualized with the ray casting algorithm [23] in which the
opacity function is set to the volume density.
We first perform reconstruction on static volumes. Fig. 6(left) shows the re-
sults of an object consisting of two glass slabs with powder on both. The letters
“EC” are drawn manually on the back plane and “CV” on the front plane by
removing the powder. Thus we create a volume in which only two planes have
non-zero density. A photograph of the object is shown in Fig. 6a. We then re-
construct the volume using the proposed method. Fig. 6 shows one of the 24
captured images as well as the reconstructed volume at different views with and
without attenuation correction. It shows that attenuation correction improves
the results by increasing the density on the back plane.
Similarly, Fig. 6(right) show the reconstruction for a 3D point cloud of a
face etched in a glass cube. As shown, our method also achieved good recon-
struction of the volume. In this example, multiple scattering and attenuation
within the point cloud are much stronger than the previous example. Thus in
the reconstructed volume, the half of the face not directly visible to the camera
has a lower estimated density (e.g., the relative darker area of the right eye in
Fig. 6).
Finally, we use our system to reconstruct time-varying volumes. We take the
dynamic process of milk drops dissolving in water as an example. We use a
syringe to drip milk drops into a water tank as shown in the adjacent figure.
With the proposed method, we are able to reconstruct time-varying volumes with
high spatial resolution (128 × 128 × 250) at 15fps, which recovers the interesting
patterns of the dynamic process (see Fig. 7).
856 J. Gu et al.
0.0 Photograph View 1 View 2 View 3
y
y
y
z x
z z
x x
1.0
Time (sec.)
2.0
3.0
4.0
Fig. 7. Reconstruction results of milk drops dissolving in water. 24 images are used
to reconstruct the volume at 128 × 128 × 250 at 15fps. The reconstructed volumes are
shown in three different views. Each row corresponds to one instance in time. The
leftmost column shows the corresponding photograph (i.e., all projector pixels emit
white) of the dynamic process.
8 Limitations
Multiple Scattering. Although utilizing more light elements increases the effi-
ciency of the acquisition, it will increase multiple scattering as well, which will
cause biased reconstruction, as the artifacts shown in Fig. 6. One potential way
to alleviate this problem is to separate multiple/single scattering by using more
complex light codes in a similar way to Nayar et al. [24].
Calibration for the Spatial Coding Method. The spatial coding seems more desir-
able than the temporal coding due to its high temporal resolution (i.e., volume
reconstruction from one single image) and the easy access of high spatial resolu-
tion devices. However, it requires highly accurate calibration both geometrically
and radiometrically. The defocus of both the projector and the camera needs to
be considered as well. In contrast, the temporal coding method is more robust
to noise and defocus and easy to calibrate.
9 Conclusions
We proposed compressive structured light for recovering the volume densities of
inhomogeneous participating media. Unlike conventional structured light range
finding methods where coded light patterns are used to establish correspondence
for triangulation, compressive structured light uses coded light as a way to gen-
erate measurements which are line-integrals of volume density. By exploiting the
sparsity of the volume density, the volume can be accurately reconstructed from
a few measurements. This makes the acquisition highly efficient both in acquisi-
tion time and illumination power, and thus enables the recovery of time-varying
volumetric phenomena.
We view compressive structured light as a general framework for coding the
3D light function L(x, y, t) for reconstruction of signals from line-integral mea-
surements. In this light, existing methods such as laser sheet scanning and laser
line interpolation, as well as the spatial coding and temporal coding discussed in
this paper, can be considered as special cases. One interesting future direction is
to design more complex coding strategies to improve the performance or apply
the method to new problems.
Acknowledgments. We would like to thank Tim Hawkins for providing their

smoke data and anonymous reviewers for their valuable comments. This work
was supported in part by the NSF (ITR-03-25867, CCF-05-41259, IIS-04-12759,
IIS-05-28402, CNS-06-14770, and CCF-06-43268), a Sloan Research Fellowship
BR-4485, and an ONR Young Investigator award N00014-07-1-0900.
References
1. Salvi, J., Pages, J., Batlle, J.: Pattern codification strategies in structured light
systems. Pattern Recognition 37, 827–849 (2004)
2. Narasimhan, S., Nayar, S., Sun, B., Koppal, S.: Structured light in scattering media.
In: ICCV 2005, pp. 420–427 (2005)
858 J. Gu et al.
3. Hawkins, T., Einarsson, P., Debevec, P.: Acquisition of time-varying participating

media. In: SIGGRAPH 2005, pp. 812–815 (2005)
4. Deusch, S., Dracos, T.: Time resolved 3D passive scalar concentration-field imag-
ing by induced fluorescence (LIF) in moving liquids. Measurement Science and
Technology 12(2), 188–200 (2001)
5. Fuchs, C., Chen, T., Goesele, M., Theisel, H., Seidel, H.: Density estimation for
dynamic volumes. Computers and Graphics 31(2), 205–211 (2007)
6. Candes, E.J., Romberg, J.: Sparsity and incoherence in compressive sampling. In-
verse Problems 23(3), 969–985 (2007)
7. Donoho, D.: Compressed sensing. IEEE Trans. on Information Theory 52(4), 1289–
1306 (2006)
8. Schechner, Y.Y., Nayar, S.K., Belhumeur, P.N.: A theory of multiplexed illumina-
tion. In: ICCV 2003, pp. 808–815 (2003)
9. Willett, R., Gehm, M., Brady, D.: Multiscale reconstruction for computational
spectral imaging. In: Computational Imaging V at SPIE Electronic Imaging (2007)
10. Lustig, M., Donoho, D., Pauly, J.M.: Sparse MRI: The application of compressed
sensing for rapid MRI imaging. Magnetic Resonance in Medicine 58(6), 1182–1195
(2007)
11. Wright, J., Yang, A., Ganesh, A., Sastry, S., Ma, Y.: Robust face recognition via
sparse representation (2008)
12. Hasinoff, S., Kutulakos, K.: Photo-consistent reconstruction of semi-transparent
scenes by density sheet decomposition. IEEE Trans. on PAMI 29(5), 870–885 (2007)
13. Ihrke, I., Magnor, M.: Image-based tomographic reconstruction of flames. In: SCA,
pp. 361–375 (2004)
14. Ihrke, I., Magnor, M.: Adaptive grid optical tomography. Graphical Models 68(5),
484–495 (2006)
15. Trifonov, B., Bradley, D., Heidrich, W.: Tomographic reconstruction of transparent
objects. In: EGSR, pp. 51–60 (2006)
16. Ihrke, I., Kutulakos, K.N., Lensch, H.P.A., Magnor, M., Heidrich, W.: State of the
art in transparent and specular object reconstruction. In: STAR Proc. of Euro-
graphics (2008)
17. Candes, E.J., Romberg, J., Tao, T.: Stable signal recovery from incomplete and in-
accurate measurements. Communications on Pure and Applied Mathematics 59(8),
1207–1223 (2006)
18. Ishimaru, A.: Wave Propagation and Scattering in Random Media. IEEE Press,
New York (1978)
19. GLPK: The Gnu Linear Programming Kit, http://www.gnu.org/software/glpk
20. Olshausen, B.A., Field, D.J.: Emergence of simple-cell receptive field properties by
learing a sparse code for natural images. Nature 381, 607–608 (1996)
21. Simoncelli, E.P.: Statistical models for images: compression restoration and syn-
thesis. In: Proc. Asilomar Conference on Signals, Systems and Computers, pp.
673–678 (1997)
22. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation noise removal algo-
rithm. Physica D 60, 259–268 (1992)
23. Schroeder, W., Martin, K., Lorensen, B.: The Visualization Toolkit: An Object-
Oriented Approach to 3D Graphics, 4th edn. Pearson Eduction, Inc., London
(2006)
24. Nayar, S.K., Krishnan, G., Grossberg, M.D., Raskar, R.: Fast separation of direct
and global components of a scene using high frequency illumination. SIGGRAPH
2006, 935–944 (2006)
Passive Reflectometry
Fabiano Romeiro, Yuriy Vasilyev, and Todd Zickler
School of Engineering and Applied Sciences,

Harvard University, Cambridge MA 02138, USA
romeiro@fas.harvard.edu
Abstract. Different materials reflect light in different ways, so re-

flectance is a useful surface descriptor. Existing systems for measuring
reflectance are cumbersome, however, and although the process can be
streamlined using cameras, projectors and clever catadioptrics, it gener-
ally requires complex infrastructure. In this paper we propose a simpler
method for inferring reflectance from images, one that eliminates the
need for active lighting and exploits natural illumination instead. The
method’s distinguishing property is its ability to handle a broad class of
isotropic reflectance functions, including those that are neither radially-
symmetric nor well-represented by low-parameter reflectance models.
The key to the approach is a bi-variate representation of isotropic re-
flectance that enables a tractable inference algorithm while maintaining
generality. The resulting method requires only a camera, a light probe,
and as little as one HDR image of a known, curved, homogeneous surface.
1 Introduction
Different surfaces modulate light in different ways, and this leads to distinc-
tive lightness, gloss, sheen, haze and so on. Thus, like shape and color, surface
reflectance can play a significant role in characterizing objects.
Computationally, surface reflectance is represented by the bi-directional re-
flectance distribution function, or BRDF; and the task of inferring the reflectance
of a surface is formulated as that of inferring a BRDF from radiometric mea-
surements. According to conventional methods, measuring surface reflectance
requires the use of controlled, active lighting to sample the double-hemisphere of
input and output directions that constitute the BRDF domain. These approaches
demand complex infrastructure, including mechanical rotation and translation
stages, digital cameras and projectors, and custom catadioptrics.
Perceptual studies suggest that humans can also infer reflectance information
from image data, but that they do so in a very different manner. While the
vast majority of machine measurement systems rely on illumination by a single
moving point source, humans rely on images captured under complex, natural
lighting [1]. The human approach has clear practical advantages: it is a passive
technique that eliminates the need for controlled lighting, and it substantially
reduces the measurement burden.
In this paper we present a passive system for inferring bi-directional surface
reflectance that also exploits natural lighting. The approach is general in that,

860 F. Romeiro, Y. Vasilyev, and T. Zickler
90 90
60 60
30 30
0 0
Fig. 1. Reflectometry using only a camera and a light probe (bottom left). Using a
bivariate representation of reflectance, the constraints induced by a single HDR image
(top left) of a known shape are sufficient to recover a non-parametric BRDF (mid-
dle). The recovered BRDF summarizes the object’s reflectance properties and is an
important material descriptor. Here, its accuracy is demonstrated through its use in
rendering a synthetic image of a novel shape (right).
besides assuming isotropy, it can handle a rather unrestricted class of materials.

It eliminates the need for active illumination and requires only a camera, a light
probe, and as little as one HDR image of a known curved shape (Fig. 1).
The success of the approach hinges on its representation of surface reflectance.
We avoid the restrictions of low-parameter BRDF models (Lambertian, Lafor-
tune, Ward, Cook-Torrance, etc.) by using general bivariate functions. This ap-
proach is motivated by recent empirical studies [2,3], and our evaluations show
that when used appropriately, this new representation provides a fine balance
between tractability and generality. It enables reflectance to be reliably inferred
from as little as one image, and at the same time, it is able to capture impor-
tant phenomena such as specular and off-specular reflection, retro-reflection, and
Fresnel effects. In particular, it is not constrained by any low-parameter analytic
BRDF model; and unlike approaches based on de-convolution, it is not limited
to radially-symmetric (i.e., one-dimensional) BRDFs.
1.1 Background and Related Work

The BRDF describes the manner in which incident radiant flux is modulated by a
uniform surface patch. It is a positive function of four angular dimensions and can
be written f (u, v), where u and v are unit vectors on the hemisphere centered
about the patch normal. These are the directions of incident and reflected flux,
respectively, and they are often expressed in spherical coordinates: (θu , φu ) and
(θv , φv ).
One can measure the BRDF of a planar material by sampling the double
hemisphere of input and output directions with a gonioreflectometer. Since this
is extremely slow, and since a slight loss of accuracy is often acceptable, a number
of camera-based alternatives have been proposed. When a camera is used with
a curved mirror [4] or a curved material sample [5], one image provides a dense
sampling of a 2D slice of the BRDF. To recover the complete BRDF domain,
these can be combined with a moving light source (e.g., [5]) or a projector [6].
Passive Reflectometry 861
These camera-based systems significantly reduce measurement time, but they

also require special-purpose hardware and precise lighting control.
Passive methods for reflectometry that require only natural lighting provide
an attractive alternative. In the computer graphics community, the inference of
reflectance from natural images has been studied under the banner of ‘inverse
rendering’. Ramamoorthi et al. [7] derive an elegant framework for inverse ren-
dering by interpreting the rendering equation as a convolution. This yields an
important theoretical tool that, among other things, enables the recovery of re-
flectance through de-convolution. Unfortunately, this approach can only yield
general isotropic BRDFs when the full 4D output light field is observed. More
typically, one has access to a small number of images; and when this is the case,
de-convolution can only yield radially-symmetric BRDFs1 , which are incapable
of representing off-specular peaks and important grazing-angle effects [2,3].
Inverse rendering can also be formulated, as it is here, directly in the angular
domain. Many approaches exist, and almost all of them rely on low-parameter
BRDF models (Phong, Cook-Torrance, etc.) to make the problem tractable.
Low-parameter BRDF models impose strong constraints on reflectance, and as
a result, one can exploit them to recover more than just reflectance information
from a set of input images. For example, there are methods for handling global
illumination effects and anisotropic reflectance [8,9], spatial reflectance variation,
and the simultaneous recovery of illumination and/or shape (e.g., [10,11]). (Pa-
tow et al. [12] provide a review.) Every parametric approach suffers from limited
accuracy, however, because the expressiveness of existing low-parameter BRDF
models is quite restricted [2,3]. This situation is unlikely to improve in the short
term. Given the diversity of the world’s materials, designing ‘general purpose’
low-parameter models that are simultaneously accurate, flexible and amenable
to tractable analysis has proven to be a very difficult problem.
Unlike these existing approaches, our goal is to recover general reflectance
information without the restrictions of radial symmetry or low-parameter mod-
els. By avoiding these restrictions, we can handle a broader class of materials.
To maintain this generality, we find it necessary to assume isotropic reflectance,
ignore global illumination effects, and require that shape and illumination be
known a priori. While the tools we develop can likely be applied to other inverse
rendering problems (see discussion in Sect. 5), we leave this for future work.
2 A Bivariate BRDF for Reflectometry
Passive reflectometry is not well-posed without some constraints on the BRDF.

Indeed, a BRDF is a function of four (angular) dimensions, while an input image
is a function of two. What we require is a way to constrain the BRDF without
surrendering our ability to represent important phenomena. Here, we present an
approach based on a bivariate representation for isotropic surface reflectance.
1
A radially-symmetric BRDF is one that, like the Phong model, is radially symmetric
about the reflection vector. It’s angular domain has dimension one.
For many materials, the dimension of the BRDF domain can be reduced with-
out incurring a significant loss of detail. The domain can be folded in half, for
example, because reciprocity ensures that BRDFs are symmetric about the di-
rections of incidence and reflection: f (u, v) = f (v, u). In many cases, the domain
(θu , φu , θv , φv ) can be further ‘projected’ onto the 3D domain (θu , θv , φu − φv )
and then folded onto (θu , θv , |φu − φv |). The projection is acceptable whenever
a BRDF exhibits little change for rotations of the input and output directions
(as a fixed pair) about the surface normal; and additional folding is acceptable
whenever there is little change when reflecting the output direction about the
incident plane. Materials that satisfy these two criteria—for some definition of
‘little change’—are said to satisfy isotropy and bilateral symmetry, respectively.
(It is also common to use the term isotropy to mean both.)
It is convenient to parameterize the BRDF domain in terms of halfway and
difference angles [13]. Accordingly, the complete 4D domain is written in terms
of the spherical coordinates of the halfway vector h = (u + v)/||u − v|| and
those of the input direction with respect to the halfway vector: (θh , φh , θd , φd ).
See Fig. 2. In this parameterization, the folding due to reciprocity corresponds
to φd → φd + π, and the projection due to isotropy (without bilateral symmetry)
is one onto (θh , θd , φd ) [13]. While undocumented in the literature, it is straight-
forward to show that bilateral symmetry enables the additional folding φd →
φd + π/2 which gives the 3D domain (θh , θd , φd ) ⊂ [0, π/2]3 .
Here, we consider an additional projection of the BRDF domain, one that
reduces it from three dimensions down to two. In particular, we project
(θh , θd , φd ) ⊂ [0, π/2]3 to (θh , θd ) ∈ [0, π/2]2 . A physical interpretation is de-
picted in Fig. 2, from which it is clear that the projection is acceptable whenever
a BRDF exhibits little change for rotations of the input and output directions
(as a fixed pair) about the halfway vector. This is a direct generalization of
isotropy, bilateral symmetry and reciprocity, which already restrict the BRDF
to be π2 -periodic for the same rotations. We refer to materials that satisfy this
requirement (again, for some definition of ‘little change’) as being bivariate.
The accuracy of bivariate representations of the materials in the MERL BRDF
Fig. 2. Domain reduction for reciprocal, isotropic, bilaterally-symmetric, and bivari-

ate BRDFs. Isotropic BRDFs are unchanged by rotations about the surface normal
(i.e., changes in φh ), while reciprocity and bilateral symmetry impose periodicity for
rotations about the halfway vector (i.e., changes in φd ). Here we consider bivariate
BRDFs, which are constant functions of φd .
0.9
0.8
ORIGINAL
0.7
Relative RMS BRDF Error
0.6
BIVARIATE
0.5
0.4
0.3
0.2
0.1
0
black−oxidized−steel
two−layer−silver
two−layer−gold
grease−covered−steel
black−fabric
polyethylene
green−latex
black−obsidian
nickel
black−phenolic
specular−yellow−phenolic
specular−black−phenolic
yellow−matte−plastic
ipswich−pine−221
polyurethane−foam
alumina−oxide
silver−metallic−paint
pink−felt
blue−fabric
blue−metallic−paint
special−walnut−224
blue−metallic−paint2
pink−fabric2
pickled−oak−260
green−fabric
white−fabric2
nylon
red−fabric
alum−bronze
green−metallic−paint
beige−fabric
specular−blue−phenolic
orange−paint
aventurnine
white−marble
brass
silicon−nitrade
green−metallic−paint2
dark−blue−paint
blue−rubber
white−fabric
red−fabric2
red−specular−plastic
specular−green−phenolic
color−changing−paint3
natural−209
white−diffuse−bball
specular−orange−phenolic
pink−jasper
gold−paint
colonial−maple−223
specular−red−phenolic
green−plastic
chrome−steel
tungsten−carbide
aluminium
gold−metallic−paint
light−red−paint
steel
specular−white−phenolic
teflon
specular−violet−phenolic
dark−specular−fabric
ss440
red−phenolic
white−paint
red−metallic−paint
red−plastic
blue−acrylic
fruitwood−241
specular−maroon−phenolic
chrome
pure−rubber
hematite
pink−plastic
neoprene−rubber
delrin
green−acrylic
dark−red−paint
pearl−paint
violet−acrylic
purple−paint
cherry−235
Fig. 3. Accuracy of bivariate representations of materials in the MERL BRDF

database. Materials are in order of increasing accuracy, and representative renderings
are shown for comparison. Most materials in the database are well-represented by a
bivariate function. (Images embedded at high resolution; please zoom in.)
database [14] are shown in Fig. 3, where they are sorted by relative RMS BRDF
error:
⎛ ⎞ 12
¯
(f (θh , θd , φd ) − f (θh , θd )) ⎠
2
Erms = ⎝ , (1)
(f (θh , θd , φd ))2
θh ,θd ,φd
with
1
f¯(θh , θd ) = f (θh , θd , φd ).
|Φ(θh , θd )|
Φ(θh ,θd )
Here, Φ(θh , θd ) is the set of valid φd values given fixed values of θh and θd .
The figure also shows synthetic images of materials that are more and less well-
represented by a bivariate BRDF. Overall, our tests suggest that the overwhelm-
ing majority of the materials in the database are reasonably well-represented by
bivariate functions. We even find that the bivariate reduction has positive effects
in some cases. For example, the original green-acrylic BRDF has lens flare ar-
tifacts embedded in its measurements2 , and these are removed by the bivariate
reduction (see Fig. 3).
Motivation for a bivariate representation is provided by the work of Stark et
al. [2] who show empirically that a carefully-selected 2D domain is often sufficient
2
W. Matusik, personal communication.
for capturing (off-)specular reflections, retro-reflections, and important Fresnel

effects. The 2D domain (θh , θd ) that is introduced above is homeomorphic to
that of Stark et al., which is why it posesses these same properties. Stark et
al. propose the ‘ασ-parameterization’ for two-dimensional BRDFs, and this is
related to (θh , θd ) by
1
α = sin2 θd , σ = (1 + cos 2θd ) sin2 θh .
2
For this reason, Figs. 2 and 3 can be seen as providing a new interpretation
and validation for their model. (The original paper examined Cornell BRDF
data [15], which is arguably more accurate but also quite sparse.)
One important advantage of our (θh , θd ) parameterization is that it provides
an intuitive means for controlling how the 2D domain is sampled. This is explored
next, where we use it for reflectometry.
3 Passive Reflectometry
We assume that we are given one or more images of a known curved surface,
and that these images are acquired under known distant lighting, such as that
measured by an illumination probe. In this case, each pixel in the images pro-
vides a linear constraint on the BRDF, and our goal is to infer the reflectance
function from these constraints. While the constraints from a single image are
not sufficient to recover a general 3D isotropic BRDF [7], we show that they
often are sufficient to recover plausible bivariate reflectance.
To efficiently represent specular highlights, retro-reflections and Fresnel ef-
fects, we can benefit from a non-uniform sampling of the 2D domain. While
‘good’ sampling patterns can be learned from training data [16], this approach
may limit our ability to generalize to new materials. Instead, we choose to man-
ually design a sampling scheme that is informed by common observations of
reflectance phenomena. This is implemented by defining continuous functions
s(θh , θ.
d ) and t(θh , θd ) and sampling uniformly in (s, t). Here we use s = 2θd /π,
t = 2θh /π which increases the sampling density near specular reflections
(θh ≈ 0). With this in mind, we write the rendering equation as

I(v, n) = L(Rn−1 u)f (s(u, Rn v), t(u, Rn v)) cos θu du, (2)
Ω
where v is the view direction, n is the surface normal corresponding to a given

pixel, and Rn is the rotation that sends the surface normal to the z-axis and the
view direction to the xz-plane. We use overloaded notation for s and t, which
depend on the incident and reflected directions indirectly through (θh , θd ).
At each pixel, this integral is computed over the visible hemisphere of light
directions Ω. Our use of a bivariate BRDF induces a ‘folding’ of this hemisphere
because light directions u and u that are symmetric about the view/normal
plane correspond to the same point in our 2D BRDF domain. When the lighting
and surface shape are known, we obtain a constraint from each pixel, and each
Fig. 4. Constraints on bivariate reflectance from natural lighting. Each pixel of an input
image (middle) captured under distant illumination (left) gives a linear constraint that
can be interpreted as an inner product of the 2D BRDF (right, first argument) and
a visible hemisphere of lighting that is weighted, warped and folded across the local
view/normal plane (right, second argument).
constraint can be interpreted as an inner product between the unknown BRDF

and a hemisphere of illumination that is weighted by cos θu , folded across the
local view/normal plane, and warped onto the st-plane. See Fig. 4.
To infer the BRDF from these constraints, we create a uniform grid S =
{(si , ti )} in the BRDF domain and approximate the rendering equation by a
sum over a discrete set Ωd of lighting directions on the hemisphere:
⎛ ⎞
2π ⎝
I(v, n) ≈ αi,j L(Rn uk )f (si , tj )⎠ cos θuk ,
k −1
(3)
|Ωd |
uk ∈Ωd si ,tj ∈Nk
where Nk is the set of the four BRDF grid points that are closest to s(uk , Rn v),
t(uk , Rn v), and αki,j is the coefficient of the bilinear interpolation associated
with these coordinates and si , tj . (We find a piecewise linear approximation of
the BRDF to be adequate.) This equation can be rewritten as
2π
I(v, n) ≈ f (si , tj ) αki,j L(Rn−1 uk ) cos θuk , (4)
|Ωd |
(si ,tj )∈S uk ∈binij
to emphasize its interpretation as an inner product.

Observations of distinct normals n1 . . . nN obtained from one or more images
provide constraints that are combined into a system of equations
I = Lf (5)
where I = [I(v, n1 ), . . . , I(v, nN )] and L is a lighting matrix whose rows are

given by the non-BRDF terms in Eq. 4. The goal is then to find f such that
these constraints are satisfied. While this may work well in the noiseless case,
in practice we require regularization to handle noise caused by the sensor, the
bivariate approximation, the discretization of the rendering equation, and errors
in the assumed surface shape.
As with general 4D BRDFs, bivariate BRDFs vary slowly over much of their
domain. Regularization can therefore be implemented in the form of a smooth-
ness constraint in the st-plane. There are many choices here, and we have found
spatially-varying Tikhonov-like regularization to be especially effective. Accord-
ing to this design choice, the optimization becomes

D D2 D −1 D2
argmin I − Lf 2 + α DΛ−1 D D D
2
s Ds f 2 + Λ t Dt f 2 (6)
f
subject to f ≥ 0,
where Ds and Dt are |S| × |S| derivative matrices, and α is a tunable scalar
regularization parameter. The matrices Λs and Λt are diagonal |S|× |S| matrices
that affect non-uniform regularization in the bivariate BRDF domain. Their
diagonal entries are learned from the MERL database by setting each to the
variance of the partial derivative at the corresponding st domain point, where
the variance is computed across all materials in the database. Probabilistically,
this approach can be interpreted as seeking the MAP estimate with independent,
zero-mean Gaussian priors on the bivariate BRDF’s partial derivatives.
There are many possible alternatives for regularization. For example, one
could learn a joint distribution over the entire bivariate domain, perhaps by
characterizing this distribution in terms of a small number of modes of variation.
However, we have found that the simple approach in Eq. 6 provides reasonable
results, does not severely ‘over-fit’ the MERL database, and is computationally
quite efficient (it is a constrained linear least squares problem).
3.1 Adequate Illumination

There is a question of when an environment is adequate for reflectometry to
be well-posed and well-conditioned. An algebraic condition is readily available;
we simply require the rank of the illumination matrix L to be sufficiently large
(i.e., to approach |S|). More intuitively, we require sufficient observations of all
portions of the BRDF domain, with regions corresponding to specular reflec-
tions (θh ≈ 0), retro-reflections (θd ≈ 0), and grazing angles (θd ≈ π/2) being
particularly important. In particular, we do not expect good results from simple
environments composed of a small number of isolated point sources. This is in
agreement with perceptual studies showing that humans are also unable to infer
reflectance under such simple and ‘unrealistic’ conditions [1].
It is interesting to compare our approach to the convolution framework of
Ramamoorthi et al. [7]. That approach enables a frequency domain analysis and
provides very clear conditions for adequacy. For radially-symmetric BRDFs, for
example, we know that an environment is adequate only if its band-limit exceeds
that of the BRDF [7]. A frequency domain analysis is difficult to apply in the
present case, however, because Eq. 5 does not represent a convolution. While an
analysis of the conditions for adequate illumination in the bivariate case may
be for worthwhile direction of future work, we focus instead on an empirical
investigation here. We show that while the quality of the result depends on the
environment, accurate reflectometry is achievable in many cases.
4 Evaluation and Results
We begin with an evaluation that uses images synthesized with tabulated BRDF
data from the MERL database [14], measured illumination3 , and a physically
based renderer4 . Using these tools, we can render images for input to our al-
gorithm as well as images with the recovered BRDFs for direct comparison to
ground truth. In all cases, we use complete 3D isotropic BRDF data to create
the images for input and ground-truth comparison, since this is closest to a real-
world setting. Also, we focus our attention on the minimal case of a single input
image; with additional images, the performance can only improve. It is worth
emphasizing that this data is not free of noise. Sources of error include the fact
that the input image is rendered with a 3D BRDF as opposed to a bivariate one,
that normals are computed from a mesh and are stored at single precision, and
that a discrete approximation to the rendering equation is used.
Given a rendered input image of a defined shape (we use a sphere for sim-
plicity), we harvest observations from 8,000 normals uniformly sampled on the
visible hemisphere to create an observation vector I of length 8,000. We discard
normals that are at an angle of more than 80◦ from the viewing direction, since
the signal to noise ratio is very low at these points. The bivariate BRDF domain
is represented using a regular 32 × 32 grid on the st-plane, and our observation
matrix L is therefore M × 1024, where M is the number of useable normals. The
entries in L are computed using Eq. 4 with 32,000 points uniformly distributed
on the illumination hemisphere. With I and L determined, we can solve for the
unknown BRDF as described in the previous sections.
We find it beneficial to use a small variant of the optimization in Eq. 6: we
solve the problem twice using two separate pairs of diagonal weight matrices
(Λs , Λt ). One pair gives preference to diffuse reflectance, while the other gives
preference to gloss. This provides two solutions, and we choose the one with low-
est residual. Using this procedure, we were able to use the same weight matrices
and regularization parameter (α) for all results in this paper. In every case, the
optimizations were initialized with a Lambertian BRDF.
Results are shown in Fig. 5. The two left columns show results using a single
input image synthesized with the Grace Cathedral environment. The recovered
bivariate BRDFs are compared to the (3D) ground truth by synthesizing images
in another setting (St. Peter’s Basilica). Close inspection reveals very little no-
ticeable difference between the two images, and the recovered BRDF is visually
quite accurate. There are numerical differences, however, and these have been
scaled by 100 for visualization. Note that some of this error is simply due to the
bivariate approximation (see Fig. 6). The next two columns similarly show the
recovery of the yellow-matte-plastic and green-acrylic materials, this time using
the Cafe environment and the St. Peter’s Basilica environment (alternately) for
input and comparison to ground truth.
3
Light probe image gallery: http://www.debevec.org/Probes/
4
PBRT: http://www.pbrt.org/
WHITE ACRYLIC NICKEL YELLOW MATTE PLASTIC GREEN ACRYLIC ENVIRONMENTS

INPUT
RECOVERED
GROUND TRUTH
100 x DIFFERENCE
Fig. 5. Visual evaluation with MERL BRDF data. A bivariate BRDF is estimated
from a single input image (top), and this estimate is used to render a new image under
novel lighting (second row ). Ground truth images for the novel environments are shown
for comparison, along with difference images scaled by 100. Few noticeable differences
exist. Far right: Environment maps used in the paper, top to bottom: St. Peter’s
Basilica, Grace Cathedral, Uffizi Gallery, Cafe and Corner Office.
In addition to these visual comparisons, we can also evaluate the recovered

BRDFs quantitatively using scatter plots and RMS errors. The top of Fig. 6
shows incident-plane scatter plots for the red channels of three recovered BRDFs
from Fig. 5, as well as the recovered colonial-maple BRDF from Fig. 1. While
the scatter plots reveal clear deviations from ground truth, they suggest that the
approach provides reasonable approximations for a variety of materials. This is
true even though just a single image is used as input—many fewer than the 300
images that were used to collect the original data [14].
The bottom of the figure displays relative RMS errors for these four recov-
ered BRDFs, along with corresponding results for all materials in the BRDF
database. Shown is the accuracy (Eq. 1) of the bivariate BRDF for each ma-
terial as estimated from one input image. This is done twice—once each using
the Grace Cathedral and St. Peter’s environments—and the curves are superim-
posed on the graph from Fig. 3, which shows the accuracy of the ‘ground truth’
bivariate reduction. (Note that the materials have been re-sorted for display
Fig. 6. Quantitative evaluation with MERL BRDF data. Top: Incident plane scatter-
plots for the four materials in Fig. 5, each showing: original 3D BRDF (blue); ‘ground
truth’ bivariate BRDF (green); and BRDF recovered from one input image (red ). Bot-
tom: Relative RMS BRDF errors for all materials in the MERL database when each is
recovered using a single image under the Grace Cathedral or St. Peter’s environments.
Vertical red lines match the scatterplots above.
purposes). The discrepancy between the results for the two different environ-
ments is expected in light of the discussion from Sect. 3.1. To further emphasize
this environment-dependence, Fig. 7 compares estimates of yellow-matte-plastic
using two different input images. The Uffizi Gallery environment (top left) does
not provide strong observations of grazing angle effects, so this portion of the
BRDF is not accurately estimated. This leads to noticeable artifacts near grazing
angles when the recovered BRDF is used for rendering, and it is clearly visible
in a scatter plot. When the Cafe environment is used as input, however, more
accurate behavior near grazing angles is obtained.
4.1 Captured Data

The procedure outlined above was applied without change to captured data.
Figure 8 shows the results for a number of materials. As before, each BRDF is
recovered from a single input image (left), and the recovered BRDFs are used to
render synthetic images of the same object from a novel viewpoint. The synthetic
images are directly compared to real images captured in the same novel positions.
Captured data contains at least three significant sources of noise in addition to
what exists in the rendered data above: 1) errors in the assumed surface geometry;
Fig. 7. Dependence on environment used for capture. An input image under the Uffizi
Gallery environment (top left) does not contain strong observations of grazing angle
effects, and as a result, the recovered BRDF is inaccurate. This is visible in a scatter
plot (bottom right, black curves) and causes noticeable artifacts when used to render in
a novel setting. If a different environment is used as input (bottom left) these artifacts
are largely avoided.
Fig. 8. Results using captured data. A BRDF is estimated from a single input image
(top) under a known environment. This recovered BRDF is used to render a synthetic
image for novel view within the same environment (middle). An actual image for the
same novel position is shown for comparison (bottom). Despite the existence of non-
idealities such as surface mesostructure and spatial inhomogeneity, plausible BRDFs
are recovered.
2) surface mesostructure (e.g., the green sphere); and 3) spatial reflectance vari-
ations (e.g., the grey sphere). Presently, surface shape is computed by assuming
the camera to be orthographic and estimating the center and radius of the sphere
in the camera’s coordinate system. Errors in this process, coupled with errors in
the alignment with the illumination probe, lead to structured measurement noise.
Despite this, our results suggest that plausible BRDFs can be recovered for a di-
versity of materials.
5 Discussion
This paper presents a technique for ‘light-weight’ reflectometry that eliminates

the need for active illumination and requires minimal infrastructure for ac-
quisition. This is enabled by reducing the domain of isotropic bi-directional
reflectance functions from three dimensions to two. We provide an empirical
evaluation of this reduced representation that compliments recent work [2].
The proposed approach has clear advantages over existing inverse rendering
techniques that recover reflectance from 2D images using de-convolution or low-
parameter BRDF models. These existing methods recover reflectance functions
that are one-dimensional (radially-symmetric) or zero-dimensional (parametric),
respectively. In contrast, the method presented here recovers a two-dimensional
reflectance function, and thereby matches the dimension of the output with that
of the input. For this reason, it can be applied to a much broader class of surfaces.
One of the important things we give up in exchange for generality is the intu-
ition provided by the convolution framework. It becomes difficult to characterize
the necessary conditions for adequate illumination, and this suggests a direction
for future work. In particular, it may be possible to clarify the role that ‘envi-
ronment foldings’ (Fig. 4) play in reducing redundancy in L and ‘enhancing the
adequacy’ of an environment.
There are a number of additional directions for future work. We presented one
of many possible regularization schemes, and it is possible that others are more
suitable. In exploring this possibility, one must be wary of ‘overfitting’ existing
BRDF databases, since these may provide descriptions of only a fraction of the
world’s interesting materials. We have largely avoided this in our approach, but
even so, we expect our method to be less successful for highly retro-reflective
surfaces, which are not well represented in the MERL database.
Our focus in this work is the recovery of general reflectance functions, mean-
ing those that are not necessarily well-represented by low-parameter models and
those that are not radially-symmetric. For this reason, we considered the case
in which the surface is homogeneous, its shape is known, and the illumination
environment is also known. Relaxing these conditions is perhaps the most inter-
esting direction for future work, and it is quite likely that the tools presented
here will prove useful elsewhere (see [17] for a reconstruction application).
In this vein, the proposed framework provides an opportunity to explore the
joint recovery of reflectance and illumination (f and L in Eq. 6), or at least
the recovery of reflectance when lighting is unknown. Using our framework, this
essentially becomes a blind de-convolution problem. It is possible that this line of
research may eventually yield computational systems that can match the human
ability to infer reflectance in uncontrolled conditions [1].
Acknowledgements
We thank Wojciech Matusik for helpful discussions regarding the MERL

database. Support comes from an NSF CAREER award and a Sloan Foundation
fellowship.
References
1. Fleming, R., Dror, R.O., Adelson, E.H.: Real-world illumination and the perception
of surface reflectance properties. Journal of Vision 3 (2003)
2. Stark, M., Arvo, J., Smits, B.: Barycentric parameterizations for isotropic BRDFs.
IEEE Transactions on Visualization and Computer Graphics 11, 126–138 (2005)
3. Ngan, A., Durand, F., Matusik, W.: Experimental analysis of brdf models. In:
Eurographics Symposium on Rendering, pp. 117–126 (2005)
4. Ward, G.: Measuring and modeling anisotropic reflection. Computer Graphics
(Proc. ACM SIGGRAPH) (1992)
5. Marschner, S., Westin, S., Lafortune, E., Torrance, K., Greenberg, D.: Image-based
BRDF measurement including human skin. In: Proc. Eurographics Symposium on
Rendering, pp. 139–152 (1999)
6. Ghosh, A., Achutha, S., Heidrich, W., O’Toole, M.: BRDF acquisition with basis
illumination. In: Proc. IEEE Int. Conf. Computer Vision (2007)
7. Ramamoorthi, R., Hanrahan, P.: A signal-processing framework for inverse render-
ing. In: Proceedings of ACM SIGGRAPH, pp. 117–128 (2001)
8. Boivin, S., Gagalowicz, A.: Image-based rendering of diffuse, specular and glossy
surfaces from a single image. In: Proceedings of ACM SIGGRAPH (2001)
9. Yu, Y., Debevec, P., Malik, J., Hawkins, T.: Inverse global illumination: recover-
ing reflectance models of real scenes from photographs. In: Proceedings of ACM
SIGGRAPH (1999)
10. Georghiades, A.: Incorporating the Torrance and Sparrow model of reflectance in
uncalibrated photometric stereo. In: Proc. IEEE Int. Conf. Computer Vision, pp.
816–823 (2003)
11. Hara, K., Nishino, K., Ikeuchi, K.: Mixture of spherical distributions for single-
view relighting. IEEE Trans. Pattern Analysis and Machine Intelligence 30, 25–35
(2008)
12. Patow, G., Pueyo, X.: A Survey of Inverse Rendering Problems. Computer Graph-
ics Forum 22, 663–687 (2003)
13. Rusinkiewicz, S.: A new change of variables for efficient BRDF representation. In:
Eurographics Rendering Workshop, vol. 98, pp. 11–22 (1998)
14. Matusik, W., Pfister, H., Brand, M., McMillan, L.: A data-driven reflectance model.
ACM Transactions on Graphics (Proc. ACM SIGGRAPH) (2003)
15. Westin, S.: Measurement data, Cornell University Program of Computer Graphics
(2003), http://www.graphics.cornell.edu/online/measurements/
16. Matusik, W., Pfister, H., Brand, M., McMillan, L.: Efficient isotropic BRDF mea-
surement. In: Proc. Eurographics Workshop on Rendering, pp. 241–247 (2003)
17. Alldrin, N., Zickler, T., Kriegman, D.: Photometric stereo with non-parametric and
spatially-varying reflectance. In: Proc. CVPR (2008)
Fusion of Feature- and Area-Based Information for
Urban Buildings Modeling from Aerial Imagery
Lukas Zebedin1 , Joachim Bauer1 , Konrad Karner1 , and Horst Bischof2

1
Microsoft Photogrammetry
irstname.lastname@microsoft.com
2
Graz University of Technology
bischof@icg.tugraz.at
Abstract. Accurate and realistic building models of urban environments are in-
creasingly important for applications, like virtual tourism or city planning. Initia-
tives like Virtual Earth or Google Earth are aiming at offering virtual models of
all major cities world wide. The prohibitively high costs of manual generation of
such models explain the need for an automatic workflow.
This paper proposes an algorithm for fully automatic building reconstruction
from aerial images. Sparse line features delineating height discontinuities and
dense depth data providing the roof surface are combined in an innovative man-
ner with a global optimization algorithm based on Graph Cuts. The fusion pro-
cess exploits the advantages of both information sources and thus yields superior
reconstruction results compared to the indiviual sources. The nature of the al-
gorithm also allows to elegantly generate image driven levels of detail of the
geometry.
The algorithm is applied to a number of real world data sets encompassing
thousands of buildings. The results are analyzed in detail and extensively evalu-
ated using ground truth data.
1 Introduction
Algorithms for the semi- or fully automatic generation of realistic 3D models of urban
environments from aerial images are subject of research for many years. Such models
were needed for urban planning purposes or for virtual tourist guides. Since the advent
of web-based interactive applications like Virtual Earth and Google Earth and with the
adoption of 3D content for mashups the demand for realistic models has significantly
increased. The goal is to obtain realistic and detailed 3D models for entire cities.
This poses several requirements for the algorithm: First, it should not require any
manual interaction because this would induce high costs. This restriction also dissuades
the use of cadastral maps as they vary in accuracy, are not readily available everywhere
and require careful registration towards the aerial data. Additionally such a dependency
increases the cost at large scale deployment. Second, the algorithm should be flexible
enough to generate accurate models for common urban roof structures without limiting
itself to one specific type, like gabled roofs or rectangular outlines for example. This

This work has been supported by the FFG project APAFA (813397) under the FIT-IT program.

874 L. Zebedin et al.
also includes the requirement to be able to deal with complex compositions of roof
shapes if those happen to be adjacent. Third, the algorithm should have a certain degree
of efficiency as it is targeted at thousands of cities with millions of buildings in total.
Last, the algorithm should be robust: the visual appearance should degrade gracefully
under the presence of noise or bad input data quality.
In the following a survey and assessment of existing algorithms is given, which fail
to meet one or more of the above mentioned requirements.
Among the early approaches are feature based modelling methods ([1,2,3,4,5])
which show very good results for suburban areas. The drawback of those methods is
their reliance on sparse line features to describe the complete geometry of the build-
ing. The fusion of those sparse features is very fragile as there is no way to obtain the
globally most consistent model.
The possibility of using additional data (cadastral maps and other GIS data in most
cases) to help in the reconstruction task is apparent and already addressed in many
publications ([6,7,8]). Such external data, however, is considered manual intervention
in our work and thus not used.
A different group of algorithms concentrates on the analysis of dense altimetry data
obtained from laser scans or dense stereo matching ([9,10]). Such segmentation ap-
proaches based solely on height information, however, are prone to failure if buildings
are surrounded by trees and require a constrained model to overcome the smoothness
of the data at height discontinuities. Guhno and Downman ([11]) combined the eleva-
tion data from a LIDAR scan with satellite imagery using rectilinear line cues. Their
approach was, however, limited to determining the outline of a building. In our work
we develop this approach further and embed it into a framework which overcomes the
problems described above.
In [12] we have proposed a workflow to automatically derive the input data used
in this paper. The typical aerial images used in the workflow have 80% along-strip
overlap and 60% across-strip overlap. This highly redundant data is utilized in this
paper. Similar approaches have been proposed by others ([13,14]), which demonstrate
that it is possible to automatically derive a digital terrain model, digital elevation model,
land use classification and orthographic image from aerial images. Figure 1 illustrates
(a) (b) (c) (d)

Fig. 1. These figures depict the data which is used for the reconstruction process: (a) height
field, (b) building mask and (c) 3D line segments. Image (d) shows the obtained model by our
algorithm.
Fusion of Feature- and Area-Based Information 875
the available data which is used for the reconstruction algorithm and also shows the
result of the proposed algorithm.
Our proposed method does not need any manual intervention and uses only data
derived from the original aerial imagery. It combines dense height data together with
feature matching to overcome the problem of precise localization of height discontinu-
ities. The nature of this fusion process separates discovery of geometric primitives from
the generation of the building model in the spirit of the recover-and-select paradigm
([15]), thus lending robustness to the method as the global optimal configuration is cho-
sen. The integration of the theory of instantaneous kinematics ([16]) allows to elegantly
detect and estimate surfaces of revolution which describe a much broader family of roof
shapes. A major feature of the proposed method is the possibility to generate various
levels of geometric detail.
The rest of the paper is structured as follows: Chapter 2 gives a general overview of
the method. In Chapter 3 we will describe the discovery of geometric primitives which
are used to approximate the roof shape, whereas Chapter 4 discusses the building seg-
mentation. Chapter 5 gives details about the fusion process which combines line fea-
tures and dense image data. Results and experiments are outlined in Chapter 6. Finally,
conclusions and further work are described in Chapter 7.
2 Overview of the Method

The workflow of the proposed method is outlined in Figure 2. Three types of informa-
tion are necessary as input for the algorithm: Dense height data is generated by a dense
image matching algorithm ([17]) (Figure 1a, represented as a height field) and gives a
good estimate of the elevation, but suffers from oversmoothing at height discontinuities
([18]). Additionally a rough segmentation of the building is required (Figure1b) which
could be directly deduced from the height data for example. The third component are
sparse 3D line segments (Figure1c) which are obtained from line matching over multi-
ple views ([1]).
Fig. 2. Illustration of the single steps of the proposed method: height data and building mask
are used to obtain a set of geometric primitives; In parallel the 3D lines are used to generate a
segmentation of the building. Finally, a labeled segmentation is produced.
The building mask is combined with the dense height data, thus filtering out all 3D
points which do not belong to the building. Afterwards the remaining points are grouped
into geometric primitives. The geometric primitives are the basic building blocks for
assembling the roof shape.
The 3D line segments are projected into the height field and used to obtain a line-
based segmentation of the building. The 2D lines of the segmentation form polygons
which are then assigned to one of the geometric primitives. Therefore, it is important
that the 3D lines capture the location of the height discontinuities as each polygon is
treated as one consistent entity which can be described by one geometric primitive. By
extruding each of the 2D polygons to the assigned geometric primitive a 3D model of
the building is generated.
Note that the algorithm presented in this paper makes no assumptions about the roof
shape. Façades are modeled as vertical planes, because the oblique angle of the aerial
images does not allow a precise reconstruction of any details.
3 Geometric Primitives
Geometric primitives form the basic building blocks which are used to describe the roof
shape of a building. Currently two types of primitives, namely planes and surfaces of
revolution, are used, but the method can be trivially extended to support other primitives.
It is important to note, that the detection of geometric primitives is independent from
the composition of the model. This means that an arbitrary amount of hypotheses can
be collected and fed into later stages of the algorithm. As the order of discovery of the
primitives is not important, weak and improbable hypotheses are also collected as they
will be rejected later in the fusion step. If a primitive is missed, the algorithm selects
another detected primitive instead which minimizes the incurred reconstruction error.
3.1 Planes
Efficiently detecting planes in point clouds for urban reconstruction is well studied and
robust algorithms are readily available ([9]). Thanks to the independence of hypothesis
discovery and model selection, a region growing process is sufficient in our workflow
for the discovery of planes. Depending on the size of the building a number of ran-
dom seed points are selected, for which the normal vector is estimated from the local
neighbourhood. Starting from the seed points, neighbours are added which fit the ini-
tial plane estimate. This plane is regularly refined from the selected neighbours. Small
regions are rejected to improve the efficiency of the optimization phase. Due to their
frequency, close to horizontal planes are modified to make them exactly horizontal, the
other oblique ones are left unchanged.
3.2 Surfaces of Revolution
Planar approximations of certain roof shapes (domes and spires for example) obtained
from plane fitting algorithms, however, are not robust, visually displeasing and do not
take the redundancy provided by the symmetrical shape into account. Therefore it is
necessary to be able to deal with other shapes as well and combine them seamlessly to
obtain a realistic model of the building.
Surfaces of revolution are a natural description of domes and spires and can be ro-
bustly detected. Mathematically such surfaces can be described by a 3D curve which
moves in space according to an Euclidean motion. Instantaneus kinematics gives a re-
lationship ([19]) between that Euclidean motion parameters and the corresponding ve-
locity vector field. Using that connection it is possible to estimate the parameters of
the Euclidean motion in a least squares sense given the normal vectors of the resulting
surface.
The equation
v(x) = c̄ + c × x (1)
describes a velocity vector field with a constant rotation and constant translation defined
by the two vectors c, c̄ ∈ R3 . If a curve sweeps along that vector field, the normal
vectors of all points on the resulting surface have to be perpendicular to the velocity
vector at the associated point. Thus
n(x)v(x) = 0 (2)
n(x) (c̄ + c × x) = 0
holds, where n(x) gives the normal vector at point x. With equation (2) it is possi-
ble to estimate the motion parameters given at least six point and normal vector pairs
(x, n(x)) lying on the same surface generated by such a sweeping curve. In the case
of point clouds describing an urban scene the parameter can be constrained by requir-
ing the rotation axis to be vertical. This already reduces the degrees of freedom to two
(assuming that z is vertical) and makes the problem easily solvable:
c̄ = (0, x, y)T c = (0, 0, 1)T
(a) (b) (c)
Fig. 3. Illustrations how starting with the dense height data the 3D curve is derived which gen-
erates the dome if it rotates around a vertical axis. (a) Raw height field with the detected axis,
(b) all inliers are projected into the halfplane formed by axis and a radial vector, (c) the moving
average algorithm produces a smooth curve.
where c̄ gives the position of the axis and c denotes the vertical rotation axis. The re-
maining two unknown parameters are estimated by transforming each 3D point with
the estimated normal vector (x, n(x)) into a Hough space ([20]). Local maxima in the
accumulation space indicate axes for surfaces of revolution. For each axis all inliers
are computed and projected into the halfplane spanned by the rotation axis and an ar-
bitrary additional radial vector. The redundancy of the symmetrical configuration can
be exploited by a moving average algorithm in order to estimate a smooth curve which
generates the surface containing the inliers. Figure 3 illustrates those steps with a point
cloud describing the shape of a spire.
4 Segmentation
The goal of the segmentation is to represent the general building structure - not only a
rectangular shape - as a set of 2D polygons.
The approach of Schmid and Zisserman ([21]) is used for the generation of the 3D
line set that is then used for the segmentation of the building into 2D polygons. A 3D
line segment must have observations in at least four images in order to be a valid hypoth-
esis. This strategy ensures that the reliability and geometric accuracy of the reported 3D
line segments is sufficiently high. The presence of outliers is tolerable since the purpose
of the 3D lines is to provide a possible segmentation of the building. Any 3D line that
does not describe a depth discontinuity can be considered as an unwanted outlier which
will contribute to the segmentation, but will be eliminated in the fusion stage.
The matched 3D line segments are used to obtain a 2D segmentation of the building
into polygons by appying an orthographic projection. The 2D lines cannot be used di-
rectly to segment the building, however, as the matching algorithm often yields many
short line segments describing the same height discontinuity. A grouping mechanism
merges those lines to obtain longer and more robust lines. A weighted orientation
(a) (b) (c)
Fig. 4. Segmentation into polygons: (a) The matched 3D lines are projected into the 2 12 D height
field, (b) outliers are eliminated by a weighted orientation histogram which helps to detect princi-
pal directions of the building. (c) Along those directions lines are grouped, merged and extended
to span the whole building.
histogram - the weights correspond to the length of each line - is created. The prin-
cipal orientations are detected by finding local maxima in the histogram. Along those
directions quasi parallel lines are grouped and merged thus refining their position.
Each grouped line is extended to span the whole building in order to simplify the
segmentation process. The lines are splitting the area into a number of polygons. Each
polygon is considered to be one consistent entity where the 3D points can be approxi-
mated by one geometric primitive.
Figure 4 illustrates this concept. The advantage of this approach is that no assumption
or constraint of the shape, angles and connectivity of the building is necessary.
5 Information Fusion
Each polygon resulting from the segmentation is assigned to one geometric primitive
(plane or surface of revolution, see Chapter 3). This labeling allows to create a piece-
wise planar reconstruction of the building - surfaces of rotation are approximated by
a rotating polyline and therefore also yield piecewise planar surfaces in the polyhedral
model.
The goal of the fusion step is to approximate the roof shape by the geometric prim-
itives in order to fullfill an optimization criterion. In this paper we use the Graph Cuts
algorithm with alpha-expansion moves ([22,23]), but other techniques like belief propa-
gation are suited as well. The goal of this optimization is to select a geometric primitive
for each polygon of the segmentation and to find an optimal trade-off between data
fidelity and smoothness.
5.1 Graph Cuts Optimization

The Graph Cuts algorithm finds a very good approximation of the globally optimal so-
lution for a broad range of tasks which can be stated as an energy minimization problem
of the following form:

E(f ) = Dp (fp ) + λ · Vp,q (fp , fq ) (3)
p∈P {p,q}∈N
where Vp,q (fp , fq ) is called the smoothness term for the connected nodes p and q which
are labeled fp and fq and Dp (fp ) is called the data term which measures a data fidelity
obtained by assigning the label fp to node p.
In our approach the segmentation induces a set P of polygons, where each polygon
represent a node of the graph. The neighboorhood relationship is reflected by the set
N , which contains pairs of adjacent polygons, ie. polygons sharing an edge. The set of
labels used in the optimization process represent the geometric primitives (planes and
surfaces of revolution):
L = {plane1 , plane2 , ..., surface-of-revolution1 , surface-of-revolution2 , ...} (4)
Thus fp ∈ L reflects the label (current geometric primitve) assigned to node (polygon)
p ∈ P.
The optimization using polygons is much faster than optimizing for each individual
pixel because there are much fewer polygons than pixels. On the other hand it also
exploits the redundancy of the height data because it is assumed that all pixels in one
polygon belong to the same geometric primitive.
In our context the smoothness term measures the length of the border between two
polygons and the data term measures the deviation between the observed surface (ob-
tained from the dense image matching algorithm) and the fitted primitive. The following
formulae are used to calculate those two terms:

Dp (fp ) = heightobs (x) − heightfp (x) (5)
x∈p
5
length(border(p, q)) iffp =
fq
Vp,q (fp , fq ) = (6)
0 iffp = fq
where p and q denote two polygons and fp is the current label of polygon p. The preset
constant λ can be used to weight the two terms in the energy functional. The data term
Dp calculates an approximation of the volume between the point cloud (heightobs(x))
and primitive fp (heightfp (x)) by sampling points x which lie within the polygon p.
This sampling strategy allows to treat all geometric primitives similarly. because they
are reduced to the incurred difference in volume and induced border to other polygons
assigned to another geometric primitive. The smoothness term Vp,q penalizes neigh-
bouring polygons with different labels depending on their common border, thus favour-
ing homogeneous regions.
The alpha-expansion move is used in order to efficiently optimize the labeling of
all polygons with respect to all discovered primitives. The initial labeling can either be
random or a labeling which minimizes only the data term for each individual polygon.
After a few iterations (usually less than 5), the optimization converges and all 2D poly-
gons can be extruded to the respective height of the assigned primitive to generate a
polyhedral model of the building.
5.2 Levels of Detail

The second term in Equation (3) regularizes the problem and favors smooth solutions.
Depending on the actual value of λ in Equation (3) different results are obtained. Higher
values result in fewer and shorter borders at the cost of larger volumetric differences
Table 1. The impact of the smoothness parameter λ on the reconstructed model. The number of
unique labels used after the Graph Cuts optimization iterations decreases as well as the number of
triangles in the polygonal model. Δ Volume denotes the estimated difference in volume between
the surface obtained by dense image matching and the reconstruced model (data term). The last
column refers to the accumulated length of all borders in the final labeling (smoothness term).
λ #Labels #Triangles Δ Volume [m3 ] Border Length [m]

5 7 79 1210.79 710.4
10 6 69 1677.19 349.4
20 4 42 1699.31 337.0
100 3 33 2293.36 290.4
between observed height values and reconstructed models. This feature can be used to
generate different models with varying smoothness, trading data fidelity for geometric
simplificationa as smaller details of the building are omitted. An example of such a
simplification is shown in Figure 6. The relevant numbers for that building are given in
Table 1.
6 Experiments
The first illustrative experiment was conducted on a test data set of a Graz. The ground
sampling distance of the aerial imagery is 8cm. The examined building features four
small cupolas at the corners. Additionally one façade is partially occluded by trees.
Figure 5 shows the results of the reconstruction process. The texture of the façades is
well aligned, implying that their orientation was accurately estimated by the 3D line
matching. The domes are smoothly integrated into the otherwise planar reconstruction.
Even the portion occluded by the tree has been straightened by the extension of the
matched 3D lines.
The next example is taken from a data set of Manhattan, New York. This building
shows that the reconstruction algorithm is not limited to façades perpendicular or par-
allel to each other. Figure 6 illustrates the effect of the smoothness term in the global
optimization energy function. Various runs with different values for λ yield a reduced
triangle count as the geometry is progressively simplified. Table 1 gives details about
the solution for different values of λ. The Graph Cuts algorithm allows to find a glob-
ally optimal tradeoff between data fidelity and generalization. Those properties are ex-
pressed by the decreased length of borders and number of labels (which translate in
general to fewer triangles) at the cost of an increase of the average difference between
reconstructed and observed surface.
(a) (b) (c)
Fig. 5. The stages of the reconstruction are illustrated by means of the building of the Graz Uni-
versity of Technology: (a) Segmented height field, (b) labeled polygons after the Graph Cuts
optimization, (c) screenshot of the reconstructed model (λ = 5)
(a) 7 Primitives, λ = 5 (b) 4 Primitives, λ = 20 (c) 3 Primitives, λ = 100
Fig. 6. Levels of Detail: The same building was reconstructed with different values for λ. The
number of geometric primitives used to approximate the shape of the roof is decreasing with
higher values for λ. In the upper row a screenshot of the reconstruction is depicted, below are
illustrations of the matching labeling obtained by the Graph Cuts optimization.
Apart from judging the visual appearance of the resulting models, we assess the
quality of the reconstructed models by comparing them to a ground truth which was
obtained manually from the same imagery. For this purpose we use a stereoscopic de-
vice to trace the roof lines in 3D. Those roof lines are connected to form polygons and
then extruded to the ground level. Those manually reconstructed models are considered
ground truth data in this paper. Using this procedure the whole data set from Manhattan
(consisting of 1419 aerial images at 15cm ground sampling distance) was processed
yielding 1973 buildings.
A comparison of manual and automatic reconstruction for one building is illustrated
in Figure 7. Both building models are converted into a height field with a ground sam-
pling distance of 15cm. This makes it easy to determine and illustrate their differences.
Figure 8 gives a break down of the height differences as a cummulative probabilty
distribution. Those graphs give the percentage of pixels where the height difference be-
tween manual and automatic reconstruction is lower than a certain threshold. Analysis
of this chart shows that for the whole data set of Manhattan (1973 buildings) 67.51%
of the pixels have a height difference smaller than 0.5m, 72.85% differ by less than
1m and 86.91% are within 2m. There are two main reasons for discrepancies of height
values: On the one hand there are displacement errors of roof edges which lead to large
height differences, depending on the height of the adjacent roof. On the other hand the
human operator is able to recognize small superstructurial details on the roofs like el-
evator shafts and air conditioning units which cause height differences usually below
2m. Those small features are sometimes missed by the automatic reconstruction.
(a) (b) (c)
Fig. 7. Quality assessment with a manually generated ground truth: In (a) and (b) the height fields
for the manually and automatically reconstructed building are shown, in (c) the height differences
are shown. The largest difference in the placement of edges is about two pixels, which is about
30cm.
Fig. 8. The cummulative probabilty distribution of the height difference for manual and automatic
reconstruction. The graph shows the error distribution for 1973 buildings from a data set of Man-
hattan, New York. The left image shows the graphs for height differences up to 100 meters; the
right graph zooms on differences up to five meters.
Detailed views of typical results from the Manhattan data set are shown in Figure 9.
The reconstruction of rectangular buildings is very successful, even though huge por-
tions of their façades are occluded by trees. The integration of surfaces of revolution
realistically models domes and spires (see 9b and 9d). It is important to note that for the
purpose of visualization the surfaces of revolution are converted to triangle meshes by
sampling them regularly (2m radially with 45 degrees of angular separation).
7 Conclusions and Future Work

In this paper we proposed a novell approach to reconstruct building models from aerial
images by combining 3D line segments and dense image matching algorithms with a
global optimization technique. The framework is able to use arbitrary basic geomet-
ric building blocks to describe the roof shape. The proposed surfaces of revolution
elegantly describe domes and spires which are difficult to recover with an approach
based on planes only. The combination of line based features and dense image matching
(a) (b)
(c) (d)
Fig. 9. Four detailed views of typical results for different types of buildings from the Manhat-
tan data set: (a) rectangular buildings, (b) rectangular building with nicely integrated dome, (c)
skyscrapers in downtown and (d) skyscraper with a spire
algorithms using a global optimization technique is very promising and is not restricted
to the reconstruction of urban scenes from aerial imagery. Additionally it allows for the
generation of different globally optimal levels of detail.
Future work will involve the investigation of other geometric primitives and methods
to exploit symmetries encountered in common roof shapes like gabled roofs. Further re-
search will be needed to evaluate the possibilities of this approach in other applications
like streetside imagery.
References
1. Baillard, C., Zisserman, A.: Automatic Line Matching And 3D Reconstruction Of Buildings
From Multiple Views. In: ISPRS Conference on Automatic Extraction of GIS Objects from
Digital Imagery, vol. 32, pp. 69–80 (1999)
2. Bignone, F., Henricsson, O., Fua, P., Stricker, M.A.: Automatic Extraction of Generic House
Roofs from High Resolution Aerial Imagery. In: European Conference on Computer Vision,
Berlin, Germany, pp. 85–96 (1996)
3. Fischer, A., Kolbe, T., Lang, F.: Integration of 2D and 3D Reasoning for Building Recon-
struction using a Generic Hierarchical Model. In: Workshop on Semantic Modeling for the
Acquisition of Topographic Information, Munich, Germany, pp. 101–119 (1999)
4. Taillandier, F., Deriche, R.: Automatic Buildings Reconstruction from Aerial Images: a
Generic Bayesian Framework. In: Proceedings of the XXth ISPRS Congress, Istanbul,
Turkey (2004)
5. Vosselman, G.: Building Reconstruction Using Planar Faces in Very High Density Height
Data. In: ISPRS Conference on Automatic Extraction of GIS Objects from Digital Imagery,
Munich, vol. 32, pp. 87–92 (1999)
6. Baillard, C.: Production of DSM/DTM in Urban Areas: Role and Influence of 3D Vectors.
In: ISPRS Congress, Instanbul, Turkey, vol. 35, p. 112 (2004)
7. Haala, N., Anders, K.H.: Fusion of 2D-GIS and Image Data for 3D Building Reconstruction.
In: International Archives of Photogrammetry and Remote Sensing, vol. 31, pp. 289–290
(1996)
8. Suveg, I., Vosselman, G.: Reconstruction of 3D Building Models from Aerial Images and
Maps. ISPRS Journal of Photogrammetry and Remote Sensing 58(3-4), 202–224 (2004)
9. Haala, N., Brenner, C.: Generation of 3D City Models from Airborne Laser Scanning Data.
In: 3rd EARSEL Workshop on Lidar Remote Sensing on Land and Sea, Tallinn, Estonia, pp.
105–112 (1997)
10. Maas, H.G., Vosselman, G.: Two Algorithms for Extracting Building Models from Raw
Laser Altimetry Data. In: ISPRS Journal of Photogrammetry and Remote Sensing, vol. 54,
pp. 153–163 (1999)
11. Sohn, G., Dowman, I.: Data Fusion of High-Resolution Satellite Imagery and LIDAR Data
for Automatic Building Extraction. ISPRS Journal of Photogrammetry and Remote Sens-
ing 62(1), 43–63 (2007)
12. Zebedin, L., Klaus, A., Gruber-Geymayer, B., Karner, K.: Towards 3D Map Generation from
Digital Aerial Images. ISPRS Journal of Photogrammetry and Remote Sensing 60(6), 413–
427 (2006)
13. Chen, L.C., Teo, T.A., Shaoa, Y.C., Lai, Y.C., Rau, J.Y.: Fusion of Lidar Data and Optical
Imagery for Building Modeling. In: International Archives of Photogrammetry and Remote
Sensing, vol. 35(B4), pp. 732–737 (2004)
14. Hui, L.Y., Trinder, J., Kubik, K.: Automatic Building Extraction for 3D Terrain Reconstruc-
tion using Interpretation Techniques. In: ISPRS Workshop on High Resolution Mapping from
Space, Hannover, Germany, p. 9 (2003)
15. Leonardis, A., Gupta, A., Bajcsy, R.: Segmentation of Range Images as the Search for Geo-
metric Parametric Models. International Journal of Computer Vision 14(3), 253–277 (1995)
16. Pottmann, H., Leopoldseder, S., Hofer, M.: Registration without ICP, vol. 95, pp. 54–71
(2004)
17. Klaus, A., Sormann, M., Karner, K.: Segment-Based Stereo Matching Using Belief Propa-
gation and a Self-Adapting Dissimilarity Measure. In: Proceedings of the 18th International
Conference on Pattern Recognition, vol. 3, pp. 15–18. IEEE Computer Society Press, Wash-
ington (2006)
18. Scharstein, D., Szeliski, R.: A Taxonomy and Evaluation of Dense Two-Frame Stereo Cor-
respondence Algorithms. International Journal of Computer Vision 47, 7–42 (2002)
19. Pottmann, H., Leopoldseder, S., Hofer, M.: Simultaneous Registration of Multiple Views of
a 3D Object. In: Archives of the Photogrammetry, Remote Sensing and Spatial Information
Sciences, vol. 34, Part 3A (2002)
20. Illingworth, J., Kittler, J.: A Survey of the Hough Transform. Computer Vision, Graphics and
Image Processing 44(1) (1988)
21. Schmid, C., Zisserman, A.: Automatic Line Matching Across Views. In: IEEE Conference
on Computer Vision and Pattern Recognition, pp. 666–671 (1997)
22. Boykov, Y., Veksler, O., Zabih, R.: Fast Approximate Energy Minimization Via Graph Cuts.
In: International Conference on Computer Vision, Kerkyra, Corfu, vol. 1, pp. 377–384 (1999)
23. Kolmogorov, V., Zabih, R.: What Energy Functions Can Be Minimized Via Graph Cuts? In:
European Conference on Computer Vision, Copenhagen, Denmark, vol. 3, pp. 65–81 (2002)
Author Index
Aach, Til I-509 Boiman, Oren IV-30

Agarwala, Aseem IV-74 Boné, Romuald II-392
Agrawal, Motilal IV-102 Bougleux, Sébastien II-129, III-57
Ahmed, Amr III-69 Bouthemy, Patrick I-113
Ai, Haizhou I-697 Bowden, Richard I-222
Ali, Asem M. III-98 Boyer, Edmond II-30
Ali, Saad II-1 Bronstein, Alexander M. II-143
Alvino, Christopher I-248 Bronstein, Michael M. II-143
Åström, Kalle IV-130 Brostow, Gabriel J. I-44
Athitsos, Vassilis I-643 Brox, Thomas I-739
Authesserre, Jean-Baptiste III-400 Bujnak, Martin III-302
Burgeth, Bernhard III-521
Babakan, Sevkit III-224 Burkhardt, Hans II-239
Babenko, Boris I-193, II-211 Byröd, Martin IV-130
Bach, Francis III-43
Bagon, Shai IV-30 Calonder, Michael I-58
Bai, Xiang IV-788 Campbell, Neill D.F. I-766
Bălan, Alexandru O. II-15 Cernuschi-Frı́as, Bruno I-113
Baldrich, Ramon IV-1 Cevher, Volkan II-155
Baraniuk, Richard G. II-155 Chai, Jinxiang I-657
Barbu, Adrian IV-465 Chan, Syin IV-817
Barinova, Olga II-100 Chang, Shih-Fu IV-270
Barreto, João P. IV-609 Charpiat, Guillaume III-126
Bartoli, Adrien III-196 Chellappa, Rama II-155
Basu, Anup II-554 Chen, Daozheng IV-116
Bauer, Joachim IV-873 Chen, Jianing I-671
Belhumeur, Peter N. IV-116, Chen, Jingni III-15, III-725
IV-340, IV-845 Chen, Tsuhan I-441, II-446
Belongie, Serge I-193, II-211 Chen, Yuanhao II-759
Berclaz, Jérôme III-112 Cheng, Irene II-554
Berroir, Jean-Paul IV-665 Cheong, Loong-Fah III-330
Berthoumieu, Yannick III-400 Chi, Yu-Tseh IV-256
Betke, Margrit I-643 Chia, Liang-Tien IV-817
Beveridge, J. Ross II-44 Chli, Margarita I-72
Bhat, Pravin II-114 Cho, Minsu IV-144
Bhusnurmath, Arvind IV-638 Chung, Albert C.S. IV-368
Bibby, Charles II-831 Chung, Ronald II-733
Bischof, Horst I-234, III-588, III-792, Cipolla, Roberto I-44, I-290, I-766
IV-677, IV-873 Cohen, Laurent D. II-129, II-392, III-57,
Black, Michael J. II-15, III-83 III-628
Blake, Andrew I-99, IV-15 Cohen, Michael II-114
Blas, Morten Rufus IV-102 Collins, Brendan I-86
Blaschko, Matthew B. I-2 Collins, Robert T. II-474, III-140
Bogoni, Luca IV-465 Comaniciu, Dorin I-711, IV-465
888 Author Index
Cooper, David B. IV-172 Fauqueur, Julien I-44

Cour, Timothee IV-158 Fehr, Janis II-239
Cremers, Daniel I-332, I-739, I-752, Fei-Fei, Li I-86, III-602, IV-527
III-792, IV-677 Feiner, Steven IV-116
Criminisi, Antonio I-99 Ferencz, Andras IV-527
Crivelli, Tomás I-113 Figl, Michael IV-492
Cui, Jinshi III-642 Fleischmann, Oliver II-638
Curless, Brian II-114 Fleuret, François III-112, IV-214
Foroosh, Hassan I-318
Dambreville, Samuel II-169 Fossati, Andrea IV-200
Damen, Dima III-154 Fradet, Matthieu III-210
Daniilidis, Kostas IV-553 Frahm, Jan-Michael I-427, II-500
Darzi, Ara IV-492 Franke, Uwe I-739
Davis, Larry S. I-16, II-610, IV-423 Freeman, William T. III-28, IV-88
Davison, Andrew J. I-72 Fritz, Mario II-527
de Campos, Cassio P. III-168 Fua, Pascal I-58, II-405, III-112, IV-200,
Delmas, Patrice II-350 IV-214, IV-567, IV-581
Deng, Jia I-86 Fulkerson, Brian I-179
Denis, Patrick II-197 Fundana, Ketut III-251
Dexter, Emilie II-293 Fusiello, Andrea I-537
Didas, Stephan III-521
Dinerstein, Michael II-321
Dinh, Thang Ba II-678 Galleguillos, Carolina I-193
Doermann, David II-745, III-752 Gammeter, Stephan II-816
Dollár, Piotr II-211 Gao, Jizhou II-624
Donner, Yoni IV-748 Garbe, Christoph III-290
Doretto, Gianfranco IV-691 Gaspar, José António IV-228
Douze, Matthijs I-304 Ge, Weina III-140
Drummond, Tom III-372 Georgiev, Todor III-224
Du, Wei II-225 Geusebroek, Jan-Mark III-696
Duarte, Marco F. II-155 Gevers, Theo I-208
Durand, Frédo IV-88 Gijsenij, Arjan I-208
Gilbert, Andrew I-222
Ecker, Ady I-127 Gimel’farb, Georgy L. II-350, III-98
Eden, Ibrahim IV-172 Gleicher, Michael IV-437
Edwards, Philip IV-492 Goh, Alvina III-238
Efros, Alexei A. IV-354 Goldman, Dan B IV-74
Elder, James H. II-197 Gong, Shaogang III-574, IV-383
Elmoataz, Abderrahim III-668 Gong, Yihong II-419, III-69
Enqvist, Olof I-141 González, Germán IV-214
Escobar, Maria-Jose IV-186 Gosch, Christian III-251
Ess, Andreas II-816 Graber, Gottfried III-792
Estrada, Francisco J. II-197 Grabner, Helmut I-234, III-588
Estrin, Deborah III-276 Grady, Leo I-248, II-252
Gray, Douglas I-262
Fan, Lixin III-182 Grinspun, Eitan IV-845
Farag, Aly A. III-98 Grossmann, Etienne IV-228
Farenzena, Michela III-196 Gu, Jinwei IV-845
Farhadi, Ali I-154, IV-451 Gu, Leon I-413
Farzinfar, Mahshid I-167 Gu, Xianfeng III-1
Author Index 889
Gupta, Abhinav I-16 Jiang, Xiaoyue IV-284

Gupta, Raj II-265 Jin, Hailin I-576
Jordan, Chris IV-158
Haines, Tom S.F. III-780 Josephson, Klas IV-130
Han, Bohyung IV-527 Junejo, Imran N. I-318, II-293
Han, Junwei IV-242 Jung, Ho Yub II-307, IV-298
Hartley, Richard I-276
Hasinoff, Samuel W. IV-45 Kahl, Fredrik I-141
Hebert, Martial III-43, III-481 Kanade, Takeo I-413
Hébert, Patrick I-454 Karlinsky, Leonid II-321
Heitz, Geremy I-30 Karner, Konrad IV-873
Herlin, Isabelle IV-665 Kidode, Masatsugu III-681
Hernández, Carlos I-290, I-766 Kim, Tae Hoon III-264
Heyden, Anders III-251 Kjellström, Hedvig II-336
Ho, Jeffrey IV-256 Klein, Georg II-802
Hofmann, Matthias III-126 Klodt, Maria I-332
Hogg, David III-154 Ko, Teresa III-276
Hoi, Steven C.H. III-358, III-766 Kobayashi, Takumi I-346
Hoiem, Derek II-582 Koch, Reinhard IV-312
Horaud, Radu II-30 Koenderink, Jan J. I-1
Hu, Weiming IV-396 Kohli, Pushmeet II-582
Hu, Yiqun IV-817 Koike, Hideki III-656
Hua, Gang I-441 Kolev, Kalin I-332, I-752
Huang, Chang II-788 Koller, Daphne I-30
Huang, Haoda II-759 Kolmogorov, Vladimir II-596
Huang, Jianguo IV-284 Komodakis, Nikos III-806
Huang, Kaiqi III-738 Kondermann, Claudia III-290
Huang, Qingming IV-541 Kong, Yuk On IV-284
Huang, Thomas II-419 Konolige, Kurt IV-102
Huang, Xinyu II-624 Konushin, Anton II-100
Huttenlocher, Daniel P. II-379, III-344 Konushin, Vadim II-100
Koppal, Sanjeev J. IV-830
Ikeuchi, Katsushi IV-623 Korah, Thommen I-359
Illingworth, John I-222 Kornprobst, Pierre IV-186
Intwala, Chintan III-224 Köser, Kevin IV-312
Irani, Michal IV-30 Kragić, Danica II-336
Isambert, Till IV-665 Krahnstoever, Nils IV-691
Krajsek, Kai IV-326
Jacobs, David W. IV-116 Kress, W. John IV-116
Jäggli, Tobias II-816 Krueger, Matthias II-350
Jain, Arpit I-483 Kukelova, Zuzana III-302
Jebara, Tony IV-270 Kumar, Neeraj II-364, IV-340
Jegou, Herve I-304 Kumar, Sanjiv III-316
Jepson, Allan D. I-127 Kuthirummal, Sujit IV-60, IV-74
Jermyn, Ian H. III-509 Kutulakos, Kiriakos N. I-127, IV-45
Ji, Qiang II-706, III-168 Kwon, Dongjin I-373
Jia, Jiaya I-671, IV-775 Kwon, Junseok I-387
Jiang, Hao II-278
Jiang, Shuqiang IV-541 Lai, Shang-Hong I-589, III-468
Jiang, Wei IV-270 Lalonde, Jean-François IV-354
890 Author Index
Lampert, Christoph H. I-2 Loxam, James III-372

Langer, Michael S. I-401 Lu, Le IV-465
Lao, Shihong I-697 Lucassen, Marcel P. I-208
Laptev, Ivan II-293 Lui, Yui Man II-44
Latecki, Longin Jan IV-788 Lumsdaine, Andrew III-224
Law, Max W.K. IV-368 Luo, Yiwen III-386
Lazebnik, Svetlana I-427 Lyu, Michael R. III-766
Lee, Hyunjung I-780
Lee, KeeChang II-100 Mairal, Julien III-43
Lee, Kyong Joon I-373 Makadia, Ameesh III-316
Lee, Kyoung Mu I-387, II-307, III-264, Makram-Ebeid, Sherif III-628
IV-144, IV-298 Mandal, Mrinal II-554
Lee, Sang Uk I-373, II-307, III-264, Marszalek, Marcin IV-479
IV-298 Martin, David R. II-278
Lee, Sang Wook I-780 Martı́nez, David II-336
Leibe, Bastian II-816 Matsushita, Yasuyuki II-692, III-656,
Leistner, Christian I-234 IV-623
Lempitsky, Victor IV-15 McKenna, Stephen J. IV-242
Leordeanu, Marius III-43 McMillan, Leonard I-711
Lepetit, Vincent I-58, II-405, IV-581 Medioni, Gérard II-678
Levi, Dan II-321 Mégret, Rémi III-400
Levin, Anat IV-88 Mei, Lin IV-492
Lewis, J.P. III-83 Mensink, Thomas II-86
Lézoray, Olivier III-668 Menzel, Marion I. IV-326
Li, Jian IV-383 Mester, Rudolf III-290
Li, Kai I-86 Metaxas, Dimitris I-685
Li, Shimiao III-330 Mezouar, Youcef III-196
Li, Shuda I-631 Migita, Tsuyoshi III-412
Li, Xi IV-396 Milborrow, Stephen IV-504
Li, Xiaowei I-427 Mille, Julien II-392
Li, Yi II-745 Miltsakaki, Eleni IV-158
Li, Yuan IV-409 Mittal, Anurag I-483, II-265
Li, Yunpeng II-379, III-344 Mordohai, Philippos IV-553
Liang, Jianming IV-465 Moreels, Pierre III-426
Liang, Lin II-72 Moreno-Noguer, Francesc II-405,
Liang, Wei II-664 IV-581
Lim, Hwasup II-100 Mori, Greg III-710
Lin, Chenxi II-759 Mory, Benoit III-628
Lin, Zhe IV-423 Murray, David II-802
Ling, Haibin IV-116
Liu, Ce III-28 Nagahara, Hajime IV-60
Liu, David I-441 Namboodiri, Anoop III-616
Liu, Feng IV-437 Narasimhan, Srinivasa G. IV-354,
Liu, Jianzhuang I-603, III-358 IV-830
Liu, Qingshan I-685 Nayar, Shree K. II-364, IV-60, IV-74,
Liu, Wei III-358 IV-340, IV-845
Liu, Yanxi II-474 Nevatia, Ramakant II-788, IV-409
Loeff, Nicolas IV-451 Nickel, Kai IV-514
Lopez, Ida IV-116 Nicolls, Fred IV-504
Loui, Alexander C. IV-270 Niebles, Juan Carlos IV-527
Author Index 891
Ning, Huazhong II-419 Rasmussen, Christopher I-359

Nishino, Ko III-440 Ravichandran, Avinash II-514
Nistér, David II-183 Ravishankar, Saiprasad I-483
Novatnack, John III-440 Reddy, Dikpal II-155
Reid, Ian II-831
Ogino, Shinsuke III-412 Reisert, Marco II-239
Okada, Ryuzo II-434 Ren, Xiaofeng III-533
Oliensis, John I-562 Ribnick, Evan III-546
Orabona, Francesco IV-228 Rittscher, Jens IV-691
Otsu, Nobuyuki I-346 Robert, Philippe III-210
Ouellet, Jean-Nicolas I-454 Romeiro, Fabiano IV-859
Romero, Javier II-336
Pajdla, Tomas III-302 Ross, David A. III-560
Pal, Christopher J. I-617 Roth, Stefan III-83
Pan, Gang I-603 Rother, Carsten II-596, IV-15
Pan, Wei-Hau III-468 Rousseau, François I-497
Pang, Junbiao IV-541 Rueckert, Daniel IV-492
Pantofaru, Caroline III-481 Russell, David III-574
Papadopoulo, Théodore II-486
Papanikolopoulos, Nikolaos III-546 Saffari, Amir III-588
Paragios, Nikos III-806 Salganicoff, Marcos IV-465
Parikh, Devi II-446 Salzmann, Mathieu IV-581
Paris, Sylvain II-460 Samaras, Dimitris III-1
Park, Minwoo II-474 Sandhu, Romeil II-169
Patterson, Alexander IV IV-553 Sankaranarayanan, Aswin II-155
Sato, Yoichi III-656
Pavlovic, Vladimir III-316
Pele, Ofir III-495 Savarese, Silvio III-602
Scharr, Hanno I-509, IV-326
Peng, Ting III-509
Schiele, Bernt II-527, IV-733
Pérez, Patrick II-293, III-210
Schikora, Marek I-332
Perona, Pietro I-523, II-211, III-426
Schindler, Konrad II-816
Peyré, Gabriel II-129, III-57
Schmid, Cordelia I-304, III-481, IV-479
Piater, Justus II-225
Schnieders, Dirk I-631
Pilet, Julien IV-567
Schnitzspan, Paul II-527
Piovano, Jérome II-486
Schnörr, Christoph III-251
Piriou, Gwenaelle I-113
Schoenemann, Thomas I-332, III-792
Pizarro, Luis III-521
Schölkopf, Bernhard III-126
Pock, Thomas III-792, IV-677
Schuchert, Tobias I-509
Pollefeys, Marc II-500
Sclaroff, Stan I-643
Ponce, Jean III-43
Sebastian, Thomas IV-691
Prinet, Véronique III-509
Seitz, Steven M. II-541
Pylvänäinen, Timo III-182
Seo, Yongduek I-780
Shah, Mubarak II-1
Quan, Long III-15, III-725 Shahed, S.M. Nejhum IV-256
Shakunaga, Takeshi III-412
Rabe, Clemens I-739 Sharma, Avinash III-616
Rabinovich, Andrew I-193 Sharp, Toby I-99, IV-595
Raguram, Rahul II-500 Shen, Chunhua IV-719
Ramamoorthi, Ravi IV-116, IV-845 Sheorey, Sameer IV-116
Ranganathan, Ananth I-468 Shi, Jianbo II-774, IV-760
892 Author Index
Shin, Young Min IV-144 Ukita, Norimichi III-681

Shotton, Jamie I-44 Ullman, Shimon II-321
Simon, Ian II-541
Singh, Meghna II-554 van de Weijer, Joost IV-1
Sivic, Josef III-28 van Gemert, Jan C. III-696
Smeulders, Arnold W.M. III-696 Van Gool, Luc II-650, II-816
Soatto, Stefano I-179, II-434, III-276, Varanasi, Kiran II-30
IV-705 Vasilyev, Yuriy IV-859
Sommer, Gerald II-638 Vaudrey, Tobi I-739
Somphone, Oudom III-628 Vazquez, Eduard IV-1
Song, Xuan III-642 Vedaldi, Andrea I-179, IV-705
Sorokin, Alexander I-548 Veenman, Cor J. III-696
Spain, Merrielle I-523 Veksler, Olga III-454
Stewénius, Henrik II-183 Veltkamp, Remco C. IV-652
Stiefelhagen, Rainer IV-514 Verbeek, Jakob II-86
Strecha, Christoph IV-567 Vidal, René I-276, II-514, III-238
Sturm, Peter IV-609 Vogiatzis, George I-290, I-766
Sugano, Yusuke III-656
Sun, Deqing III-83
Wang, Fei II-568
Sun, Jian II-72, IV-802
Wang, Hongzhi I-562
Sun, Yi II-58
Wang, Jingbin I-643
Syeda-Mahmood, Tanveer II-568
Wang, Lei IV-719
Szummer, Martin II-582
Wang, Liang I-576
Ta, Vinh-Thong III-668 Wang, Liming II-774
Tabrizi, Mostafa Kamali I-154 Wang, Qiang II-720
Takamatsu, Jun IV-623 Wang, Ruixuan IV-242
Tan, Tieniu III-738 Wang, Shu-Fan I-589
Tang, Xiaoou I-603, II-720, III-386, Wang, Xianwang II-624
IV-802 Wang, Yang III-1, III-710
Tannenbaum, Allen II-169 Wang, Yueming I-603
Tao, Dacheng I-725 Wedel, Andreas I-739
Tao, Hai I-262 Wei, Shou-Der III-468
Tarlow, Daniel III-560 Wei, Xiaolin K. I-657
Taskar, Ben IV-158 Weickert, Joachim III-521
Taylor, Camillo Jose IV-638 Weinman, Jerod J. I-617
Teoh, Eam Khwang I-167 Wen, Fang II-72
ter Haar, Frank B. IV-652 Werman, Michael III-495
Toldo, Roberto I-537 White, Sean IV-116
Tong, Yan II-706, III-168 Wietzke, Lennart II-638
Torralba, Antonio III-28 Willems, Geert II-650
Torresani, Lorenzo II-596 Wilson, Richard C. III-780
Tran, Du I-548 Wojek, Christian IV-733
Tran, Lam I-617 Wolf, Lior IV-748
Tran, Son D. II-610 Wolf, Matthias IV-465
Trobin, Werner IV-677 Wong, Kwan-Yee K. I-631
Tsuji, Ryosuke III-681 Wu, Bo II-788
Tu, Peter IV-691 Wu, Changchang I-427
Tu, Zhuowen II-211, IV-788 Wu, Yang II-774, IV-760
Tuytelaars, Tinne II-650 Wu, Zheng I-643
Author Index 893
Xiang, Tao IV-383 Zemel, Richard S. III-560

Xiao, Jianxiong III-15, III-725 Zeng, Wei III-1
Xiao, Rong I-603, II-72 Zeng, Yun III-1
Xing, Eric III-69 Zerubia, Josiane III-509
Xu, Li I-671, IV-775 Zha, Hongbin III-642
Xu, Wei II-419, III-69 Zhang, Jingdan I-711
Xu, Zenglin III-766 Zhang, Lei II-706
Xue, Zhong I-167 Zhang, Li II-364
Zhang, Ling IV-116
Yakubenko, Anton II-100 Zhang, Shuwu II-664
Yamazaki, Shuntaro IV-830 Zhang, Tianhao I-725
Yang, Jie I-725
Zhang, Wei II-720
Yang, Ming-Hsuan I-468, IV-256
Zhang, Weiwei IV-802
Yang, Peng I-685
Zhang, Xiaoqin IV-396
Yang, Ruigang I-576, II-624
Zhang, Yanning IV-284
Yang, Wuyi II-664
Zhang, Zhang III-738
Yang, Xingwei IV-788
Zhang, Zhongfei IV-396
Yao, Bangpeng I-697
Zhang, Ziming IV-817
Yao, Jian-feng I-113
Zhao, Huijing III-642
Yeung, Dit-Yan III-15, III-725
Zhao, Ming II-733
Yezzi, Anthony II-169
Zhao, Rongchun IV-284
Yin, Lijun II-58
Zheng, Nanning IV-760
Yin, Xiaotian III-1
Yu, Kai III-69 Zheng, Yefeng III-752
Yu, Qian II-678 Zhou, Changyin IV-60
Yu, Ting IV-691 Zhou, Luping IV-719
Yu, Xiaodong II-745 Zhou, Shaohua Kevin I-711
Yuen, Jenny II-692, III-28 Zhu, Guangyu II-745, III-752
Yuille, Alan II-759 Zhu, Jianke III-766
Yun, Il Dong I-373 Zhu, Long (Leo) II-759
Zhu, Qihui II-774, IV-760
Zach, Christopher I-427 Zickler, Todd IV-859
Zaharescu, Andrei II-30 Zitnick, C. Lawrence II-114, II-446
Zebedin, Lukas IV-873 Zwanger, Michael IV-326

Computer Vision - ECCV 2008

Uploaded by

Copyright:

Available Formats

Computer Vision - ECCV 2008

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Computer Vision - ECCV 2008

Uploaded by

Copyright:

Available Formats

Lecture Notes in Computer Science 5305

Commenced Publication in 1973

10th European Conference on Computer Vision

Library of Congress Control Number: 2008936989

CR Subject Classification (1998): I.4, I.2.10, I.5.4, I.5, I.7.5

LNCS Sublibrary: SL 6 – Image Processing, Computer Vision, Pattern Recognition,

Welcome to the 2008 European Conference on Computer Vision. These proceed-

August 2008 Andrew Zisserman

Program Chair Support

Richard Szeliski Microsoft Research, USA

James Coughlan Christopher Geyer Esther Koller-Meier

Ram Nevatia Radim Šára John Tsotsos

Mark Drew Mike Langer Michael Ross

Image Segmentation by Branch-and-Mincut . . . . . . . . . . . . . . . . . . . . . . . . . 15

What Is a Good Image Segment? A Uniﬁed Approach to Segment

Flexible Depth of Field Photography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

Understanding Camera Trade-Oﬀs through a Bayesian Analysis of

Searching the World’s Herbaria: A System for Visual Identiﬁcation of

A Column-Pivoting Based Strategy for Monomial Ordering in

Riemannian Anisotropic Diﬀusion for Tensor Valued Images . . . . . . . . . . . 326

FaceTracer: A Search Engine for Large Collections of Images with

What Does the Sky Tell Us about the Camera? . . . . . . . . . . . . . . . . . . . . . . 354

Three Dimensional Curvilinear Structure Detection Using Optimally

Scene Segmentation for Behaviour Correlation . . . . . . . . . . . . . . . . . . . . . . . 383

Robust Visual Tracking Based on an Eﬀective Appearance Model . . . . . . 396

Key Object Driven Multi-category Object Recognition, Localization

A Pose-Invariant Descriptor for Human Detection and Segmentation . . . 423

Texture-Consistent Shadow Removal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437

Scene Discovery by Matrix Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451

Simultaneous Detection and Registration for Ileo-Cecal Valve Detection

Constructing Category Hierarchies for Visual Recognition . . . . . . . . . . . . . 479

Locating Facial Features with an Extended Active Shape Model . . . . . . . 504

Dynamic Integration of Generalized Cues for Person Tracking . . . . . . . . . 514

Extracting Moving People from Internet Videos . . . . . . . . . . . . . . . . . . . . . . 527

Multiple Instance Boost Using Graph Embedding Based Decision Stump

Object Detection from Large-Scale 3D Datasets Using Bottom-Up and

Making Background Subtraction Robust to Sudden Illumination

Closed-Form Solution to Non-rigid 3D Surface Registration . . . . . . . . . . . . 581

Implementing Decision Trees and Forests on a GPU . . . . . . . . . . . . . . . . . . 595

General Imaging Geometry for Central Catadioptric Cameras . . . . . . . . . . 609

Estimating Radiometric Response Functions from Image Noise

Solving Image Registration Problems Using Interior Point Methods . . . . . 638

3D Face Model Fitting for Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 652

A Multi-scale Vector Spline Method for Estimating the Fluids Motion

Continuous Energy Minimization Via Repeated Binary Fusion . . . . . . . . . 677

Uniﬁed Crowd Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691

Quick Shift and Kernel Methods for Mode Seeking . . . . . . . . . . . . . . . . . . . 705

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 887

Eduard Vazquez, Joost van de Weijer, and Ramon Baldrich

Computer Vision Center \ Dpt. Ciencies de la Computacio. Ediﬁci O Universitat

Abstract. The segmentation method proposed in this paper is based

Image segmentation is a computer vision process consisting in the partition of an

perform the segmentation. Finally, hybrid techniques combine methods of the

f (x) = mb (x) cb + mi (x) ci (1)

in which f = {R, G, B}, cb is the body reﬂectance, ci the surface reﬂectance,

To illustrate the diﬃculty of extracting the distributions of DCs consider

In literature several methods have explicitly used the dichromatic model to

2 A Ridge Based Distribution Analysis Method (RAD)

2.1 First Step: Ridge Extraction

Multilocal Creaseness: Finding Candidates and Enhancing Connectiv-

LM P (Ω(x, y)) = {(x, y)|(∇Ω(x, y) = 0), λ1 < 0, λ2 < 0} (5)

T RP (f (x, y)) = {(x, y)|∇Ω(x, y) = 0, λ1 < 0, ∇Ω(x, y) · ω1 = 0,

SP (f (x, y)) = {(x, y)|∇Ω(x, y) = 0, λ1 · λ2 < 0} (7)

Shai Bagon, Oren Boiman, and Michal Irani