Diploma Thesis Toth

Universitt Karlsruhe (TH)
Fakultt fr Informatik
Interactive Systems Labs
Prof. Dr. A. Waibel
Diploma thesis
Combination of Classifier Cascades

and Training Sample Selection for
Robust Face Detection
Submitted by
Lorant Szasz-Toth
January 2009
Advisors
M.Sc. Hazm Kemal Ekenel, Universitt Karlsruhe (TH)
Dr. Jie Yang, School of Computer Science, Carnegie Mellon University
Dr. Wende Zhang, General Motors
Dr.-Ing. Rainer Stiefelhagen, Universitt Karlsruhe (TH)
Interactive Systems Labs
Institut fr Anthropomatik
Universitt Karlsruhe (TH)
Title: Combination of Classifier Cascades and Training Sample Selection for Robust Face
Detection
Author: Lorant Szasz-Toth
Lorant Szasz-Toth
Rudolfstr. 15
76131 Karlsruhe, Germany
email: LorantToth@gmail.com
Statement of authorship
I hereby declare that this thesis is my own original work which I created without illegit-
imate help by others, that I have not used any other sources or resources than the ones
indicated and that due acknowledgement is given where reference is made to the work of
others.
Karlsruhe, 30. January 2009

............................................
(Lorant Szasz-Toth)
Abstract
Face detection is one of the most fundamental tasks in human-computer-interaction,

surveillance, and, more recently, image retrieval. Determining the location and size of
faces in input images is a prerequisite for many other applications, including face recog-
nition. In recent years several breakthroughs have been made in this field. These days,
face detectors deliver high detection rates, low false alarm rates and run in real-time.
Despite the efforts and publicly available tools, training high-performance face detectors
from scratch remains a challenge. Mostly, because training time for a single cascade can be
in the order of days and various training parameters have to be chosen carefully. Usually,
training involves acquiring heuristics and a feeling for the intricacies of the training process
and the influence of training parameters. A substantial amount of time is spent training
classifiers iteratively and modifying parameters, while usually discarding intermediate
results.
The goal of this work is to overcome some of the problems of training cascade classifiers and
to promote the use of custom-trained classifiers. Specifically, two problems are addressed
in this work. First, an approach to combine several trained cascade classifiers into a single
cascade is presented and evaluated. Second, a technique to optimize the training set is
explored.
A major challenge during cascade training is the choice of training parameters. There
is no ideal way to choose these parameters and optimization is not feasible. Usually,
the process involves several attempts or guesses at the right parameters and, finally, the
best performing classifier is selected. Instead of discarding intermediate results, several
of these classifiers are combined into a single new classifier. Unlike previous work, the
base classifiers are not run in parallel but a fixed number of individual classifier stages are
optimized, selected and combined into a new classifier without added run-time overhead.
Experiments have shown the importance of a proper choice of training samples. Classifiers
trained with a reduced amount of well-chosen samples can outperform a classifier that was
trained on a far larger training set. The use of less training samples to achieve the same
performance decreases the required training time, especially with large training sets, where
results cannot be cached. Additionally, forcing the classifier to focus on difficult training
examples has shown to increase classification performance. Therefore, a method to select
an optimized set of training samples from a large set with the help of support vector
machines is explored.
The results of both presented approaches have been evaluated on the widely used, publicly
available CMU+MIT database. Both the SVM-based training sample selection and the
cascade combination approaches are shown to improve the performance over the base
classifiers. Cascade combination allows to generate a classifier within a single day that
performs nearly as well as a single, high-performance classifier trained in more than ten
days. Additionally, classifiers generated by cascade combination outperform the orignal
base cascades. Face detectors trained with SVM-based training set selection perform
better than equally trained base classifiers with a random choice of training samples. Both
presented approaches were able to produce cascade classifiers that clearly outperform the
publicly available OpenCV face detectors.
Zusammenfassung
Die Gesichtsdetektion ist eine der grundlegenden Aufgaben der Mensch-Maschine-Interaktion,

Videoberwachung und, neuerdings, der Verwaltung von digitalem Videomaterial. Das
Auffinden von Gesichtern ist die Grundvoraussetzung fr viele weitere Aufgaben, wie
zum Beispiel der Gesichtserkennung. Mehrere wissenschaftliche Durchbrche gelangen in
den letzten Jahren auf diesem Feld. Diese ermglichen heutigen Gesichtserkennern hohe
Detektionsraten in Echtzeit bei einer niedrigen Zahl falscher Detektionen.
Trotz einer Vielzahl wissenschaftlicher Verffentlichungen zu diesem Thema und der f-
fentlichen Verfgbarkeit von Toolkits zur Gesichtserkennung, bleibt das Training von
kaskaden-basierten Detektoren eine Herausforderung. Der Hauptgrund hierfr ist die hohe
Trainingszeit von meist mehreren Tagen, gepaart mit der Notwendigkeit mehrere Param-
eter manuell zu whlen. Normalerweise umfasst der Trainingsprozess daher das Erstellen
mehrer Detektoren, um ein Gefhl fr den Einfluss der Parameter und die Eigenheiten des
Trainingsprozesses zu erlangen. Viel Zeit vergeht daher mit dem wiederholten Training
dieser Klassifikatoren bis letztendlich der beste Parametersatz gewhlt wird.
Das Ziel dieser Arbeit ist es die genannten Probleme zu entschrfen. Folgende zwei An-
stze werden in dieser Arbeit im Besonderen behandelt. Erstens wird eine Technik zum
Zusammenfhren verschiedener, bereits trainierter Klassifikatoren in einen einzigen De-
tektor prsentiert. Der zweite Ansatz dient der Optimierung des Trainingsdatensatzes.
Da die Wahl optimaler Trainingsparameter schwierig und eine Optimierung nahezu un-
mgich ist, befasst sich der erste Ansatz mir der Kombination verschiedener, nicht optimal
trainierter Detektoren in einen einzigen finalen Klassifikator. Dieser kombinierte Detek-
tor verbessert die Leistungsfhigkeit der Ausgangsdetektoren und fhrt diese nicht einfach
parallel aus, sondern ist so schnell wie ein einziger Klassifikator.
Experimente haben die Bedeutung einer guten Wahl der Trainingsbeispiele gezeigt. Ein
kleiner, gut-gewhlter Trainingssatz kann einen Klassifikator hervorbringen, der die gleiche
Leistung erbringt wie ein mit grerem Datensatz trainierter Klassifikator. Dies widerum
reduziert den Trainingszeitbedarf. Auerdem hat sich gezeigt, dass es hilfreich ist das
Training des Detektors auf schwere Beispiele zu lenken. Darauf beruhend wird in dieser
Arbeit ein Ansatz zur optimierten Wahl der Trainingsbeispiele mittels SVMs vorgestellt.
Die Leistung beider vorgestellten Anstze wird auf dem vielfach genutzten, ffentlich ver-
fgbaren CMU+MIT Datensatz evaluiert. Sowohl die Optimierung des Trainingssatzes,
als auch die Kombination verschiedener Klassifikatoren, zeigen eine gesteigerte Leistung
im Vergleich zu herkmmlich trainierten Kaskaden. Die Kombination verschiedener Kaskaden
kann dazu genutzt werden innerhalb eines Tages einen Klassifikator zu generieren, der
mit einem ber den Zeitraum von zehn Tagen trainierten Detektor vergleichbar ist. Zum
Schluss dieser Arbeit folgt noch ein Vergleich der hier prsentierten Anstze mit ffentlich
verfgbaren Gesichtserkennern und State-of-the-Art Detektoren. Die resultierenden Klas-
sifikatoren beider vorgestellten Verfahren erreichen bessere Resultate als die ffentlich
verfgbaren OpenCV Kaskaden.
Acknowledgements
The completion of this thesis would not have been possible without the help and support
of several individuals. The author would like to thank his advisors Dr. Jie Yang, Carnegie
Mellon University, Dr. Wende Zhang, General Motors, M. Sc. Hazm K. Ekenel and Dr.-
Ing. Rainer Stiefelhagen from Universitt Karlsruhe (TH) for the valuable discussions,
suggestions and corrections.
The author would also like to thank everyone involved with the InterACT exchange pro-
gram for offering the opportunity to pursue the research at the Carnegie Mellon University
in Pittsburgh, PA.
This research was partially supported by General Motors through GM-CMU collaborative
Lab, NIH (under contract 1U01HL09173601) and by OSEO, French State agency for
innovation, as part of the Quaero Programme.
Contents
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Overview of Face Detection Approaches 5

2.1 Knowledge-based Face Detection . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Feature-based Face Detection . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Template-based Face Detection . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Appearance-based Face Detection . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 Cascades of Boosted Ensembles . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5.1 Alternative Boosting Algorithms . . . . . . . . . . . . . . . . . . . 13
2.5.2 Alternative Feature Sets . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.3 Weak Learners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5.4 Cascade Training Procedure . . . . . . . . . . . . . . . . . . . . . . 17
2.5.5 Cascade Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5.6 Training Time Improvement . . . . . . . . . . . . . . . . . . . . . . 19
2.5.7 Run-time Efficiency Improvement . . . . . . . . . . . . . . . . . . . 20
2.5.8 Classifier Combination . . . . . . . . . . . . . . . . . . . . . . . . . 21
3 Methodology 23
3.1 Face Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1.1 Haar Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1.2 Integral Image Representation . . . . . . . . . . . . . . . . . . . . . 24
3.1.3 Boosting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.4 Weak Learners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
I
3.1.5 Cascades of Boosted Ensembles . . . . . . . . . . . . . . . . . . . . 28
3.1.6 Optimal Thresholding for Cascaded Ensembles . . . . . . . . . . . . 29
3.1.7 Bootstrapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.8 Matrix-Structural Learning . . . . . . . . . . . . . . . . . . . . . . 30
3.2 Entropy and Mutual Information . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.1 Linear classification . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.2 Soft-margin linear classification . . . . . . . . . . . . . . . . . . . . 35
3.3.3 Non-linear classification . . . . . . . . . . . . . . . . . . . . . . . . 36
4 Cascade Combination and SVM-based Sample Selection 37

4.1 Cascade Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.2 Cascade Combination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2.1 Optimization Problem Formulation . . . . . . . . . . . . . . . . . . 39
4.2.2 Cascade Combination with Threshold Optimization . . . . . . . . . 41
4.2.3 Cascade Combination with Conditional Mutual Information Maxi-
mization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2.4 Cascade Combination with Mutual Information Maximization . . . 43
4.2.5 Cascade Combination with Correlation Maximization . . . . . . . . 44
4.2.6 Sample Extraction and Subsampling . . . . . . . . . . . . . . . . . 45
4.3 Training Set Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3.1 SVM-based Training Set Selection . . . . . . . . . . . . . . . . . . . 46
5 Experiments 49
5.1 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.1.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.1.2 Bounding Boxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.1.3 Face Matching Metric . . . . . . . . . . . . . . . . . . . . . . . . . 50
II
5.2 Receiver Operating Characteristic Curves . . . . . . . . . . . . . . . . . . . 51
5.3 Base Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.3.1 Training Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.3.2 Base Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.3.3 Training Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.3.4 Motivation for Training Set Selection . . . . . . . . . . . . . . . . . 58
5.3.5 Motivation for Cascade Combination . . . . . . . . . . . . . . . . . 60
5.3.6 Large Training Sets with Matrix-Structural Learning . . . . . . . . 61
5.4 Classifier Combination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.4.1 Threshold Optimization . . . . . . . . . . . . . . . . . . . . . . . . 62
5.4.2 Subsampling for Cascade Combination . . . . . . . . . . . . . . . . 64
5.4.3 Comparison of Cascade Combination Cost Functions . . . . . . . . 64
5.4.4 Combination Stage Count . . . . . . . . . . . . . . . . . . . . . . . 66
5.4.5 Comparison against Base Classifiers . . . . . . . . . . . . . . . . . . 67
5.5 SVM-based Training Set Selection . . . . . . . . . . . . . . . . . . . . . . . 69
5.5.1 Histogram Equalization vs. Raw Image Intensity Features . . . . . 69
5.5.2 SVM-based Training Set Resampling . . . . . . . . . . . . . . . . . 69
5.6 Combining SVM-based Resampling and Cascade Combination . . . . . . . 71
5.7 Comparison against other Cascade Classifiers . . . . . . . . . . . . . . . . 72
6 Conclusion 75
7 Future Work 77
III
List of Figures
1 Example Haar-like features . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2 Matrix-Structural Learning overview . . . . . . . . . . . . . . . . . . . . . 32
3 Maximum margin classification . . . . . . . . . . . . . . . . . . . . . . . . 34
4 Non-linear classification using kernel functions . . . . . . . . . . . . . . . . 36
5 ROC curve smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6 ROC curves 3000 samples . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
7 ROC curves 4500 samples . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
8 ROC curves 3000, 4500, 9000 samples . . . . . . . . . . . . . . . . . . . . . 55
9 ROC curve 20 vs. 25 stages . . . . . . . . . . . . . . . . . . . . . . . . . . 56
10 Training times during 25 stage training . . . . . . . . . . . . . . . . . . . . 58
11 Training times during 20 stage training . . . . . . . . . . . . . . . . . . . . 59
12 Comparison of identical cascade classifiers with training sets . . . . . . . . 59
13 Cascade combination of low-performance classifiers vs. high-performance
cascade . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
14 Cascade training with large training set via MSL . . . . . . . . . . . . . . 62
15 The influence of threshold optimization . . . . . . . . . . . . . . . . . . . . 63
16 Subsampling negative samples for cascade combination . . . . . . . . . . . 65
17 Comparison of cascade combination cost functions . . . . . . . . . . . . . . 66
18 Performance of increased final stage count through cascade combination . . 67
19 Performance of cascade combination from 20-stage base classifiers . . . . . 68
20 Performance of cascade combination from 25-stage base classifiers . . . . . 68
21 SVM bootstrap training . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
22 Performance of classifiers trained with resampled training set . . . . . . . . 71
23 Cascade combination and SVM-based training set selection vs 9000 sample
base cascade . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
24 Comparison against OpenCV default cascade . . . . . . . . . . . . . . . . . 73
25 Comparison against state-of-the-art . . . . . . . . . . . . . . . . . . . . . . 74
IV
List of Abbreviations
CART Classification and regression tree

ESS Enormous sample set
HCI Human-computer interaction
MSL Matrix-Structural Learning
ROC Receiver operating characteristics
RSAT Rotated summed area table
SAT Summed area table
SVM Support vector machine
ASM Active shape model
CT Modified census transform
FFS Forward feature selection
HMM Hidden Markov Model
ICA Independent component analysis
LDA Linear discriminant analysis
ML Machine learning
MLP Multi-layer perceptron
MVFD Multi-view face detection
PCA Principal component analysis
RBF Radial basis function
RNDA Recursive non-parametric discriminant analysis
V
VI
1
1 Introduction
Several reasons have lead to a tremendous amount of research in the field of face detection.
It is one of the classic problems in computer vision as it is the basis for many applications
such as human-computer-interaction, surveillance, identification and smart environments.
Despite the work that has been done in this field robust, real-world, real-time face detec-
tion remains an open research issue.
Face detection and recognition are integral parts of human-human interaction. Faces
convey a lot of information about the interaction partner, obviously identity but also
mood, age and gender. Humans have the innate ability to detect, recognize and interpret
faces.
The most important motivation to study face detection is, therefore, its integral role
in many human-computer interaction (HCI) tasks. HCI research tries to enhance the
usability of computer systems by providing other means of input and output than tradi-
tional keyboard and display-based systems. In order to ease the interaction with human
users, machines need to be able to perceive their environment in a similar way as their
users and supply more natural forms of interaction. Face detection is a key component of
many next-generation human-computer-interfaces. Current identification methods require
the user to identify themselves actively, by logging into a computer or by other means,
whereas during human-human-interaction the interaction partner is automatically rec-
ognized. Knowledge about the partner may ease interaction by allowing the computer
system to incorporate existing information to customize the interaction process. HCI ap-
plications have been continuously refined with the availability of more computing power
and cheap video recording devices for computers. Additionally, face detection is the first
processing step in solving many other HCI related problems like recognizing and identify-
ing people, gender or facial expressions. These tasks require a human face to be located
first. Once the face has been detected other classifiers can be applied to establish a
persons identity, emotional state and extract additional information with which smart
environments can enhance the users experience.
Face detection has also been in research focus due to its complex, non-trivial nature. It
attempts to find a non-rigid, transformable, 3-dimensional object in 2-dimensional images.
Faces exhibit great variance due to lighting, different camera parameters, in-plane head
and image rotation, facial expressions, occlusion by other objects or accessories like glasses
or scarfs and out-of-plane head rotation (pose). Albeit being more challenging than rigid
2 1 INTRODUCTION
object detection in controlled settings, faces exhibit limited amounts of deformation and
thus variance in appearance as opposed to, for example, hands. Therefore, face detection
is more challenging than other object detection tasks and remains a challenge, while
several promising advances have already been made and quick progress is possible.
Additionally, face detection poses a challenge to most machine learning algorithms due to
the high-dimensional raw data space of even small input images. The handling of large
training sets is necessary in order to cover most of the variance presented by faces as
discussed before. The use of more than 10,000 or even 100,000 training samples can be
mandatory in order to model all variations in the appearance of faces. These large data
sets and the complex nature of the classification task allow for testing and improving
general machine learning algorithms.
Early work on face detection was either based on the detection of facial components, skin-
color models or modeling facial appearance using the respective state-of-the-art machine
learning algorithms. These early approaches were either fast or achieved high detection
rates, but their computational requirements were immense. Consequently, the high com-
putational overhead limited their applications. While many machine learning algorithms
achieve good face/non-face classification results, they cannot meet real-time requirements
of real-world applications.
With the introduction of Haar feature-based cascade classifiers by Viola and Jones [64],
the first robust real-time applications became feasible and interest in the field of face
detection surged. Since then, many variations of and additions to their work have been
published. Viola and Jones contributed the combination of three important ideas. First,
the use of very basic Haar-like features whose computation is extremely efficient. Second,
the use of a cascade of classifiers for real-time detection and third, the use of AdaBoost
for the aggressive selection of discriminative features and ensemble classifier construction.
1.1 Motivation
Despite the recent work and breakthroughs in face detection systems, training high-
performance face detectors is non-trivial. A major challenge in training Haar feature-
based cascade classifiers is their high training time. In order to fully exploit the potential
of Haar feature-based cascade classifiers, several training parameters, i.e. the training set
and the number of stages, need to be hand-tuned. Implementations of Haar feature-based
1.2 Contribution 3
cascade classifiers are publicly available, e.g. the widely used OpenCV toolkit, along with
pretrained face detectors. Unfortunately, these face detectors are far behind the results
of state-of-the-art systems on publicly available databases, whereas most applications
would strongly benefit from retrained specific face detectors. The need to fine-tune many
parameters for optimal performance is a big challenge.
Since finding optimal parameters is not feasible computationally, the process usually in-
volves training several cascade classifiers with a varying number of final stages, different
training sets with various ratios of positive to negative training samples and several in-
dividual stage false acceptance and detection rate targets. Therefore, several cascade
classifiers with varying performance will be the result of determining optimal parameters,
but there are no guarantees that any of those classifiers will perform optimally.
Usually, a lot of time is thus spent on developing heuristics for cascade training. Several
cascade classifiers are generated in the process and, at some point, the cascade with the
best performance is selected. The other cascades are usually discarded. On the other
hand, various attempts were made to combine several trained classifiers into one superior
classifier. These methods usually run the base classifiers in parallel and then combine their
classification results. This combination can improve the performance, but the run-time
increases linearly with the number of combined classifiers.
This work explores two ways to overcome the difficulties of the training process and de-
livers improved cascade classifiers without increased run-time. The presented approaches
focus on optimizing the training set and combining multiple trained cascades to over-
come problems associated with the complexities of training. The approaches try to find a
smaller, optimized training set in order to reduce training time and try to combine several
trained cascade classifiers into a single classifier with improved classification performance.
The final combined classifier is a combination of individual classifiers stages, therefore
the base classifiers are not run in parallel.
1.2 Contribution
The main contribution of this work can be summed up as follows:
While classifier combination has been applied several times before, most approaches
run multiple classifiers in parallel and apply more or less complex arbitration and
4 1 INTRODUCTION
combination schemes. In the field of face detection, run-time efficiency is one of

the most important factors. Combination efforts negate the recent run-time per-
formance gains, while improving classification performance. The presented method
to cascade combination seeks to combine several stages of multiple pretrained cas-
cades into a single cascade. This method is, therefore, avoiding the need to run
several classifiers in parallel. Training of classifier cascades is not trivial. Especially
the optimization of the training parameters is time-consuming and requires a lot
of experience and possible hand-tuning of stage-thresholds. Additionally, negative
(and possibly positive) samples are bootstrapped, so the training sets of equally
trained cascades may vary. Therefore, several cascades trained with the same train-
ing parameters may exhibit different performances. So, careful selection of individ-
ual stages from different classifiers allows to overcome the deficiencies of individual
classifiers and create an improved classifier.
The other aspect of this work is training set resampling. While many aspects of the
design of cascaded classifiers have been examined, the choice of an ideal training
set has not received a lot of attention except for the work of Chen et al. [11]. We
examine a simple method to rank training samples and force a trained cascade to
concentrate on the classification boundary between face and non-face samples. This
approach is motivated by the already existing bootstrapping of the negative set for
each cascade stage, where the classifier is continually forced to concentrate on more
difficult examples. Also, this strategy has proved to be useful for classification in
support vector machines.
1.3 Outline
The rest of this work is organized as follows. Chapter 2 provides an overview of the
numerous face detection algorithms. Chapter 3 introduces the basic concepts used in the
presented methods, namely cascaded Haar feature-based face detectors, information the-
ory basics and support vector machines basics. Chapter 4 outlines the proposed methods
for cascade combination and training set resampling by means of support vector machines.
Chapter 5 presents several experiments that evaluate the performance of the presented
approaches and tries to present some general insights into training of cascaded classifiers.
Chapter 6 and 7 conclude by summarizing this work and presenting possible future work.
5
2 Overview of Face Detection Approaches
Face detection is defined as the task of determining the presence of faces in input images.
Given an image, face detection systems determine whether faces are present in the image
and, if present, supply the location and size of each face within the image. Overall at
least 150 approaches to face detection have been reported by 2002 [75] and many more
since then due to more recent breakthroughs.
Generally, face detection algorithms have been differentiated into the following four cat-
egories [75]:
1. Knowledge-based approaches try to apply human knowledge about the characteris-

tics of faces. These approaches mainly try to classify faces by the relation of facial
features such as eyes, nose and mouth.
2. Feature-based approaches aim to extract features invariant to changes in pose, view-

point and lighting.
3. Template-based approaches match one or more standard patterns of faces or facial

features and try to maximize the correspondence with a template.
4. Appearance-based methods classify parts of the image by applying models learned

from a set of training images that represent the face and non-face classes.
Due to the immense impact of boosted classifier cascade architectures on face detection
an additional subsection will detail the most recent developments in that area separately.
2.1 Knowledge-based Face Detection
Early attempts at face detection mainly consisted of knowledge-based methods. Re-

searchers tried to encode intuitive human knowledge into algorithms or rules. These rules
encode the features of faces and the relationships of the features. Care has to be taken
to not make these rules too strict or relaxed resulting in few correct detections or a high
number of false positives. Yang and Huang [72] developed a hierarchical knowledge-based
system. Their system consists of layers of rules, from coarse rules that describe the ap-
pearance of a face applied first, to more specific rules describing features once the input
image has been pre-scanned.
6 2 OVERVIEW OF FACE DETECTION APPROACHES
2.2 Feature-based Face Detection
Knowledge-based methods proved mostly only useful for face localization where only a
single face has to be located within an image. Therefore, research lead to approaches
that try to establish a faces location by finding features invariant to pose, viewpoint or
lighting changes. These methods are called feature-based approaches. They can further
be divided into methods that use facial features, texture, skin color or a combination of
the aforementioned to increase the robustness.
A number of facial feature based approaches have been suggested. Generally, detection
of local facial features is not very robust and leads to many false detections, while mul-
tiple passes over the image may be necessary to detect different features using different
classifiers.
Facial features: A method based on horizontal and vertical projections to locate bound-
aries of a face and the facial features has been used by Kotropoloulos and Pitas [31]. Large
changes in horizontal and vertical projections [29] are interpreted as the borders of faces.
Within the designated face region, local minima of those projections determine the loca-
tion of lips, eyes and nose. Another facial feature-based approach was devised by Leung
et al. [33]. They use local feature detectors to detect eyes, nostrils and the nose and lip
region. The face detection problem is expressed as a random graph matching problem
between a graph-model of a typical face that contains distributions of typical pair-wise
facial feature distances learnt from training samples and a graph constructed from face
candidates based on the local facial feature detections. Leung et al. [7] have refined their
approach to avoid problems associated with modeling mutual distances scale-, rotation-
and translation-invariantly by using shape statistics. Further feature-based methods, for
example, apply edge-maps and heuristics to locate faces [72]. Another method is based
on eye and eyebrow detection after morphological pre-processing and a geometrical face
model guided by eye detection [23]. Many further edge- or local feature-detector-based
approaches have been published.
Texture: The second category of feature-based approaches use the unique texture of a
face to distinguish it from other background objects. Using texture may be problematic
due to input images of different resolution and difficult lighting, complex backgrounds
may also pose a problem. Augusteijn et al. [2] developed a system for human face texture
2.2 Feature-based Face Detection 7
classification based on hair, skin and other features. They use second-order statistical
features (space gray-level dependence matrix) to model the textures and a cascade corre-
lation neural network for classification. Only texture classification results were published.
Dai and Nakano [15] use similar statistical features and embed color information into their
models to enhance skin-colored areas in images.
Skin color: Skin color has been successfully used to detect faces. Several studies have
been conducted and established the suitability of skin color for face detection, e.g. [73].
Different color spaces, including RGB, normalized RGB, HSV, YIQ and different models
for the skin-color distribution have been used. Unfortunately, skin color is sensitive to
changes in lighting and the resulting chrominance changes, so, it may have to be adapted.
On the other hand, the detection of skin-colored pixels and image patches is very ef-
ficient. The most basic approaches use thresholds to define a skin-color region within
the color space, e.g. Chai and Ngan [10] threshold Cr and Cb channels to classify skin-
color values, the thresholds are learnt from training data. Methods based on histograms
have also been used, e.g. histogram intersection in HSV space [49]. Other approaches
use unimodal Gaussian density functions [8], mixtures of Gaussians [74] with multimodal
Gaussian distributions to model the skin-color distribution. Additional work has been
done to overcome the lighting problem, i.e. that the color appearance varies highly due
to differences in lighting in different scenes. McKenna [40] et al. introduced an adap-
tive color model for tracking faces in complex environments, whereas Forsyth [18] derives
equations based on a physical model to address color constancy.
Combined approaches: Since most feature-based approaches are not robust by them-
selves, several systems have been proposed that combine several of the aforementioned
approaches. These systems tend to pre-process input images using large-scale features
such as skin color or shape and verify the resulting matches using local features such as
eyes, eyebrows, mouth or nostrils - Sobottka and Pitas present such a system [56]. HSV-
based skin-color segmentation is followed by a connected-component analysis. Then,
candidates that can be well represented by ellipses are selected and verified using facial
features as eyes and mouth that are assumed to be darker than the rest of the face.
2.3 Template-based Face Detection
Another rather intuitive way to detect faces is based on matching previously prepared pro-
totypical face templates against parts of the input image - as opposed to appearance-based
methods these templates are not learnt from training data, but specifically designed by
experts. Straight-forward template-based methods have difficulties dealing with changes
in pose, rotation and scale. Therefore, multi-resolution, multi-scale and deformable tem-
plates have been proposed to overcome these problems.
Static templates: Samal and Iyengar [47] presented a face detection systems using
silhouette templates. A set of face silhouettes, represented as arrays of bits, is generated
via principal component analysis (PCA) from selected training samples. The generalized
Hough-transform is then used in conjunction with the resulting principal silhouettes to
localize faces. Sinha [55] proposed the use of brightness differences between facial regions,
because, while the overall brightness may change, the relative brightness differences remain
similar. The brightness differences are further reduced to just the sign of the difference
to provide a robust feature. Then a template is generated by dividing a face into several
regions that roughly correspond to facial landmarks as eyes, nose, mouth, cheeks, eyebrows
and the mutual brightness differences between these regions are modeled using training
data. A face is detected if these brightness-difference constraints are met.
Deformable templates: To overcome some of the previously mentioned limitations,

deformable templates were introduced. They are parametrized templates which imple-
ment an elastic model that can be matched to different input face shapes. The most
prominent approaches were active shape models (ASMs). Cootes and Taylor [14] as well
as Kirby and Sirovich [30] use statistical models of shape and orientation of key features
to detect faces. Cootes and Taylor first select a set of key features and then create a
model of the statistics of the relative positions and orientations of these feature points to
create a deformable active shape model. Kwon and Vitoria Lobo [32] propose a method
using snakes. So-called snakelets are used to find edges, these are then used to match
ellipses to face candidate regions by checking which snakelets lie along the perimeter of
an ellipse via voting. These face candidates are verified with facial feature detectors using
a face template with deformable eye and face outline models.
2.4 Appearance-based Face Detection 9
2.4 Appearance-based Face Detection
As opposed to the previously presented template-based approaches, appearance-based

face detection systems try to learn the distinguishing qualities of faces and non-faces from
raw training examples without human assistance.
Machine learning algorithm-based approaches: Many appearance-based algorithms

use specific machine learning algorithms or statistical analysis tools to perform the pattern
recognition task of classifying face and non-face image patches. Most machine learning
(ML) algorithms have been successfully applied to face detection, usually in a supervised
learning context where all samples are labelled correctly before training. Simply applying
ML algorithms, though, is usually very time intensive, since in the appearance-based ap-
proach every possible sub-window, that is windows within the image of different location
and scale, has to be classified as face or non-face.
Distribution-based approaches: Sung and Poggio claim that appearance-based ap-

proaches have a few advantages over, for example, template-based, manually designed
systems [55, 61]. Appearance-based methods learn from large amounts of training data
and, therefore, reduce the dependence on domain-specific knowledge and reduce the risk
of mistakes due to incomplete or inaccurate knowledge. They also claim that, by adding
misclassified patterns to the training set, the number of falsely accepted non-faces and
the number of correctly detected faces can be arbitrarily tuned. The last claim is not
generally accepted as, although the quantity of training samples is important, quality of
training samples becomes more and more important as more samples are used to train a
system.
Sung and Poggio [61] presented a method to detect faces based on modeling the distri-
bution of face and non-face patterns. Their approach models the distribution of face and
non-face patterns as several multivariate Gaussian distributions using a cluster mean and
a co-variance matrix. These clusters are extracted via a modified k-means algorithm from
training data that is represented as vectors by concatenating image rows. Experiments
found 6 clusters for both face and non-face classes to perform best. Then, a multi-
layer perceptron (MLP) is used to perform the face/non-face classification using a special
distance metric. The distance consists of two components, the first is the normalized
Mahalanobis distance of the input pattern projected onto a subspace constructed from
the largest eigenvectors of the clusters. The second component is the Euclidean distance
between the test pattern and the reconstruction of its projection onto the subspace. An
important aspect of their work is the selection of training samples for the non-face class.
Since the non-face class contains substantially more possible patterns than the face class,
they select the negative samples in a bootstrap fashion. Starting with a small initial ran-
dom set of non-face samples and a set of face patterns an MLP face detector is trained,
then, misclassified non-face samples are added to the negative training set and the process
is repeated until a desired amount of samples has been collected. This method of selecting
negative training samples has since been used by other groups, e.g. Rowley et al. [45]
or, recently, boosted classifier cascades. While Sung and Poggio model the distribution of
global face patterns in a high-dimensional space, Schneiderman and Kanade [53] model
the joint probability of spatial relationships and local appearance of faces. They estimate
the posterior probability function of an input pattern being a face or non-face.
Linear transformation-based methods: Subspace methods project input patterns

into a lower-dimensional subspace in order to ease classification. Several projections are
known and have been used, principal component analysis (PCA), linear discriminant
analysis (LDA) and independent component analysis (ICA) are the most widely used
transformations. While PCA tries to retain as much variance of the input data as possible
and is, therefore, optimal for compression, LDA has been designed for classification by
producing a projection that best separates the different classes, i.e. it maximizes the ratio
of between-class scatter to within-class-scatter. Turk and Pentland presented a well-
known PCA-based face detection and recognition method [62]. They create a subspace
using PCA from face training samples and project face patterns into the subspace to
model their distribution. For classification, an input pattern is projected into that space
and the distance of the sample and the facespace is computed as the likeliness of the
pattern being a face.
Neural network face detectors: Neural networks have been widely used successfully
to perform many different pattern recognition tasks in the past and they have also been
applied to the face detection task. Neural networks are able to model highly non-linear
decision boundaries in the input data space. Therefore, neural networks allow the con-
struction of complex decision functions for face/non-face classification. On the other
hand, network type and network topology have to be carefully chosen in order to avoid
2.4 Appearance-based Face Detection 11
overfitting while retaining a network that is powerful enough to perform the classification
task. The optimization of these parameters is time-consuming and only aided by heuris-
tic guidelines. The most important neural network-based face detection approach has
been developed by Rowley et al. [45] - their approach is still cited in some more recent
publications as a baseline algorithm. They designed a multi-layer neural network face
detection system based on receptive fields that were chosen to allow the network to detect
individual or combined facial features. First, multiple neural networks are applied to the
input image to classify all possible sub-windows. Then, heuristics are applied to handle
overlapping detections of the same face at different scales and slightly different positions
to discard false positives that were found not to exhibit these multiple detections. Instead
of heuristics an arbitration between the detection results of the networks, like AND or OR
operations or an arbitration neural network, can be applied. Rowley et al. used a special,
faster pre-scanning network to identify likely face candidates quickly in order to speed up
the detection process. This first network was larger than the original networks and was
able to roughly locate face candidates quickly before the more extensive classifiers were
applied to the face candidates, this can be interpreted as an early cascaded architecture.
The presented approach cannot detect rotated faces, therefore Rowley et al. proposed an
extension of their method [46]. They used a prepended router network that was trained
to estimate the rotation of face image patches. These patches were then rotated to the
upright pose before they were passed to the original system. Other neural network based
approaches have been devised, Agui et al. [1] used two parallel networks fed with direct
intensity values and Sobel-filtered intensities, Soulie et al. [58] applied time-delay neural
networks on wavelet transformed input images for scale invariance.
Support vector machine-based approaches: Osuna et al. [43] were the first to
apply another successful machine learning algorithm to face detection, support vector
machines (SVMs). SVMs construct a linear hyperplane that separates two sets of input
data represented as image row concatenated vectors. As opposed to other linear classifiers,
SVMs aim to maximize the margin between the classes and the separating hyperplane
and thus minimize the expected generalization error instead of just the training error.
Additionally, the data points are projected into a high-dimensional space by means of
a kernel function before classification to allow for non-linear classification in the input
space. The system of Osuna et al. is trained in a similar manner as Sung and Poggio [61]
have done, by bootstrapping the negative non-face samples. They trained the classifier
on normalized, vectorized images (rows concatenated) 19x19 input images. The detection
is performed by applying the classifier to all possible sub-windows of the input image at
different scales. The SVM-based approach slightly outperforms the system of Sung and
Poggio. Another SVM-based system is presented by Heisele et al. [24], but their use of
overlapping appearance-based component classifiers, albeit not invariant to rotation and
scale, makes it a mixture of feature-based and appearance-based methods. They train 14
linear-SVM component classifiers that mostly cover the whole face. The component-based
layout is supposed to make the approach more robust to pose and illumination changes
than traditional holistic appearance-based approaches. Finally, an SVM is used to verify
that the configuration of the individual feature detectors constitutes a valid face.
Several other machine learning paradigms have been applied to face detection. For ex-
ample, Samaria [48] applied Hidden Markov Models (HMMs) for face detection by repre-
senting facial regions scanned line-by-line as states of the model, Yang et al. [76] applied
a sparse network of winnows as classifier.
Classifier cascades: More recently, a new face detection approach has been proposed
by Viola and Jones [64], referred to as cascades of boosted ensembles. As opposed to
previous works, their method was able to perform real-time face detection in a robust
manner achieving state-of-the-art performance on evaluation data sets. The success of
the approach is due to several contributions. In order to achieve high performance a so-
called cascaded classifier design was applied. The cascaded design is efficient and fast for
rare-event detection tasks, like face detection, by rejecting rather easy cases early with
only few feature computations. Additionally, conceptually simple and, by themselves, not
very powerful thresholded rectangle features were used. These are, on the other hand,
fast to evaluate using an integral image representation. Since a large quantity of these
features exists even within small scanned windows, a greedy approach based on AdaBoost
is used to select the most suitable features. A linear combination of these thresholded
features is used to build each stage of the classifier cascade. The combination coefficients
are chosen according to the AdaBoost algorithm. This seminal work has sparked high
interest in the research community and many alternative approaches to several aspects
of the work have been published. Fellow researchers have explored alternative features,
weak learners, methods to train the weak learners and boosting schemes.
2.5 Cascades of Boosted Ensembles 13
2.5 Cascades of Boosted Ensembles
The following Chapter will present some of the recent research in face detection associated
with cascades of boosted ensembles.
2.5.1 Alternative Boosting Algorithms
Most of the immediate work inspired by the publication of Viola and Jones [63] has been
the exploration of alternative boosting methods. Boosting algorithms combine several
simple learners into a single powerful classifier - see Chapter 3 for details. The boosting
algorithm has seen many improvements over the initially devised AdaBoost-based ap-
proach by Viola and Jones [63]. Several variants of boosting have been applied to object
detection tasks in conjunction with Haar-features and classifier cascades, e.g. Discrete
AdaBoost [19, 52] Real AdaBoost [26, 52], asymmetric AdaBoost [65], FloatBoost [79],
GentleBoost [20], Kullback-Leibler boosting[37] and WaldBoost [57]. Among the plethora
of boosting variants, Discrete AdaBoost and Real AdaBoost seem to be the most widely
used variants.
Initially, Viola and Jones [63] used discrete AdaBoost proposed by Freund and Schapire
[19, 52]. It was the first practical, polynomial-time boosting algorithm. AdaBoost works
by iteratively finding a simple weak-learner that best classifies the labelled, weighted
training data. Iteratively, the algorithm reweighs all samples, increasing weights for mis-
classified samples, and then finds a new best-performing classifier on the newly weighted
training set. At the end of each iteration the new classifier is added to the ensemble of
the classifiers with an appropriate weight.
Boosting algorithms seem prone to overfitting and learning noise, in order to overcome
that weakness Gentle AdaBoost has been proposed [19, 52]. Gentle AdaBoost tries to
reduce the susceptibility to noise by reducing the weights assigned to outliers. A study by
Lienhart et al. [35] features a comparison of Discrete AdaBoost, Real AdaBoost and Gen-
tleBoost and suggests that GentleBoost outperforms the other two boosting algorithms
for face detection.
Another variant of AdaBoost is Real AdaBoost [20], it was used by Lienhart et al. [35]
and a generalized version of Real AdaBoost for multi-class-multi-label classification was
used by Huang et al. [25, 26] in multi-view face detection. Real AdaBoost is an extension
of AdaBoost that uses real-valued weak learner outputs instead of simple binary class
labels in the basic two-class case. Real-valued outputs can improve the classification over
the coarse binary class label information.
Kullback-Leibler Boosting, or KLBoosting, has been proposed by Liu and Shum [37].
Feature-selection and boosting are based on Kullback-Leibler divergence, a measure of
distance between two probability distributions from information theory. The authors
seek to maximize the Kullback-Leibler divergence of histograms of face/non-face classes
that were generated by projecting the data to 1D-histograms of KL features that are
approximated with simple wavelets. Also, they use the Kullback-Leibler divergence to
derive the weight coefficients of weak classifiers when added to the ensemble.
AdaBoost turned out to be not ideal for applications in cascaded frameworks, since the
original algorithm optimizes the overall error, whereas a cascaded framework is mostly
concerned with the minimization of false negatives. The cascade architecture is strongly
dependent on high correct classification rates of individual stages, while modest false
acceptance rates per stage are tolerable. Viola and Jones [64] updated their face detec-
tion approach by introducing asymmetric AdaBoost that modifies the AdaBoost sample
reweighting scheme after each round of boosting to increase the weight of positive exam-
ples proportionally.
Li and Zhang proposed FloatBoost, an extended boosting strategy that includes back-
tracking after each iteration of boosting in order to remove ineffective or harmful weak
classifiers that increase the error rate [79].
For a near optimal trade-off between computational efficiency and minimal error rate,
Sochman and Matas [57] presented a new boosting algorithm based on Walds sequential
probability ratio test that optimizes the average decision time given a target error rate.
2.5.2 Alternative Feature Sets
The initial work by Viola and Jones [63, 64] used basic rectangle features, called Haar-
wavelets or Haar-filters. These are simple features consisting of a weighted sum of two
to four rectangular areas of summed image intensities - see Chapter 3 for details. This
large set of possible features forms an overcomplete set of bases that is able to represent
the original image intensities. Their advantage is their conceptual simplicity and the
ability to compute them very efficiently. Using an integral image representation only four
table lookups per rectangle are necessary. A disadvantage is their limited discriminational
ability for classification.
Lienhart et al. [35] introduced an extended set of features that include features rotated
by 45 that are equally efficient. The new features are a generalization of the concept
introduced by Viola and Jones [63] and some are inspired by biological features of human
vision. These features increase the set of available basis functions for classification and
allow for slightly better detection performance of about 10% decreased false alarm rate
at the same hit rate [35].
A simple modification for profile face detection has been proposed by Ishii et al. [27].
Instead of connected rectangle features, this approach allows the use of differences of two
disconnected rectangles to improve the classification of profile faces that are hard to model
with the traditional feature set due to the lack of many discriminative edges or lines as
in frontal faces.
Completely different feature sets were explored by Zhang et al. [78] and Wang and Ji
[67, 66]. Weak classifiers based on Haar-wavelets are fast to evaluate but may lack the
ability to discriminate well between face and non-face samples in later cascade training
stages, since the bootstrapping approach successively concentrates on hard samples. In
order to overcome this problem, different feature sets were suggested. On the other hand,
other features are usually computationally expensive, thus, Haar-wavelets are still used
in the initial stages. Zhang et al. [78] suggest the use of global PCA-based features at
later stages once the local Haar-features reach the discriminatory limit and error rates
approach 50%. After the face eigenvectors have been constructed via PCA from training
data, the boosting algorithm selects the new PCA-based global features based on their
discrimination ability using 1D feature value histograms, not their eigenvalue rank. Wang
and Ji [66] suggest the use of the recursive non-parametric discriminant (RNDA) analysis,
since PCA is designed for retaining the most variance in the data after projection and
is not optimal for discrimination. RNDA does not assume a Gaussian form of class
distributions and is thus a generalization of LDA and is also a global feature. Again,
feature histograms were used to model the distribution of features and, thereby, find the
most discriminative feature corresponding to eigenvectors of the RNDA process. Wang
and Ji applied their boosted RNDA framework to multi-view face detection [67].
Liu and Shum [37] used global projection features based on Kullback-Leibler-Features
(KL-features) that are linear projections that maximize the Kullback-Leibler divergence of
1D-histograms of the two classes. They approximated the optimal KL-features by linearly
combining a subset of Gabor-wavelet or Haar-wavelet features at different orientations,
scales and translations.
Another different feature type was used by Frba and Ernst [21]. They propose the
use of the modified census transform, a local transform that generates a string of bits
representing which neighborhood pixels have a lower intensity than the current location.
The modified census transform (CT) uses the mean neighborhood intensity of the current
locations as an anchor to be more expressive. The neighborhood size is theoretically
not limited, but sizes of 3 3 were used to capture local details and keep computational
overhead low. For every image pixel and its neighborhood a CT feature can be constructed
that describes the resemblance to a structural kernel. These features can be efficiently
evaluated by using the resulting bit-sequences as an index into a lookup table. A cascade
of boosted classifiers is then trained as in other approaches with a small number of image
locations selected as features, the last stage is trained to evaluate all locations. The
presented approach performs well on the CMU+MIT database and is very fast due to the
efficient and sparse calculation of CT features.
2.5.3 Weak Learners
Basic, thresholded Haar-features have been widely used as features for face detection.
Lienhart et al. [35] and Brubaker et al. [5] doubt that these simple, thresholded features
suffice for difficult classification tasks. Both used low height Classification and Regression
Trees (CARTs) [4] with Haar-wavelet-based nodes instead. Their results are somewhat
contradictory, Lienhart et al. [35] reported modest improvements whereas Brubaker et
al. [6] found significant improvements.
Huang et al.[25] introduced piece-wise functions as weak learners. Piece-wise functions

divide the feature space into n equal bins and output a constant value for each bin that
describes the divergence of positive and negative samples feature values. The authors
claimed that piece-wise function learners allow the learning procedure to converge faster,
while allowing for more robust classification at later stages. It is not clear whether the
reported results are due to the new weak learner or the other novelty presented in the
paper, the Vector Boosting approach that will be explained later.
2.5.4 Cascade Training Procedure
The cascade training process as proposed by Viola and Jones [64] has a few draw-backs.
The most obvious problem is the ad-hoc choice of individual cascade training goals. The
authors chose the same hit rate and false alarm rate for all trained stages. These choices
may not be ideal. Luo [38] described a method to globally optimize the cascade architec-
ture post-training and set the final stage thresholds accordingly. Instead of simultaneously
optimizing stage classifiers and stage thresholds, the approach optimizes the thresholds
of a pre-trained cascade architecture. Two solutions were presented, a simple algorithm
assuming that stage classifiers can be optimized independently and a dependent greedy
search solution to find an optimal operating point by adjusting stage thresholds. Since
the approach does not address the adjustment of thresholds during the training phase
that influences bootstrapping, McCane et al. [39] suggested to model the cost of execu-
tion of individual stages. The cost function is two-dimensional and depends on both false
acceptance and detection rates and is therefore difficult to model. Hence, McCane et al.
[39] created a family of ROC curves based on an incrementally built monolithic classifier
that were used to estimate the parameters of the stage-execution-cost model. Then global
false acceptance and detection rates could be minimized using the cost function and op-
timal individual operating points for stages could be found. Brubaker et al. [5, 6, 59]
viewed the overall performance of the cascade as a random variable and trained individual
cascade stages with a minimal number of features so that the probability of meeting the
overall cascade goals was sufficiently high. Each stages performance iwas evaluated on an
independent evaluation set and the probability of meeting the overall cascade goals was
tested given the current results on the evaluation set. A cost function was defined based
on the probabilities of meeting the overall goals and minimized during training. Chen
and Yuille [13] proposed a method to learn a cascade that is as fast as possible given a
desired overall accuracy. They tried to minimize the average processing time, basically
trading off false acceptance rate against a stages execution time.
Dundar and Bi [16] proposed a different training architecture, called AND-OR learning.
Their method avoids the independent treatment of stages during the training process and
all stages were trained in a joint fashion. Instead of the usual greedy scheme that trains
classifiers sequentially, the approach optimizes all classifier stages in parallel by providing
mutual feedback between stages. Positive and negative sample are assigned different loss
functions as positive samples have to pass all stages (AND-operation) and negatives have
to be rejected by a single stage only (OR-operation). The approach iteratively optimizes

the overall cascade performance by adjusting a single stages parameters while fixing all
other stages parameters.
Chen et al. [11] presented an approach to optimize the training set for training a classifier
cascade. The choice of good training samples plays an important role in the quality
of a trained classifier. The presented approach covers generating a large training set
that should cover the whole face space and a resampling technique that selectes samples
as to densely and evenly cover the whole face space. An initial training set was first
expanded by using a genetic algorithm that crossed parts of faces (crossover operation)
and applied relighting (mutations) to generate new samples. After each round of the
genetic algorithms, the current generation of generated samples was checked against a
face classifier trained with the last-generation data. Samples with too large variations
from the last generation were discarded because of their low resemblance with faces.
The process was repeated several times. Once a sufficient amount of samples had been
generated, a manifold space was created by means of the Isomap algorithm to represent
the local distances of the face space in a lower-dimensional space. Then, sparse areas
of the face space were filled by interpolated samples. Finally, the set was resampled to
a sufficiently large set of samples that evenly covered the original face space without
dense or sparse areas. The final classifier presents one of the best reported results on
the CMU+MIT database. A one-class SVM was used as an additional last step to reject
further false positives, thereby lowering the false alarm rate compared to strict cascade
classifiers.
2.5.5 Cascade Architecture
Most of the complexity of current face detectors lies in their use of the cascaded classifier
architecture, several parameters like individual stage goals, number of stages, thresholds
and more have to be tuned to achieve state-of-the-art performance. Due to the training
time in the order of days, optimizing these parameters is difficult or impossible. Therefore,
several enhancements have been proposed to overcome the problem of threshold choices
by giving up a strict cascade architecture. Bourdev and Brandt [3] coined the term
soft cascade for an architecture that departs from strictly separated stages. Instead of
deciding whether to accept a sample after a sequence of features, i.e. after every stage, and
restarting the summation and thresholding, feature values are continuously thresholded
in a monolithic classifier so that a sample may be discarded after every feature. The
approach by Bourdev and Brandt [3] goes one step further than the Boosting Chain
presented by Xiao et al. [70]. The Boosting Chain avoids resetting the feature value
sums after thresholding at each stage. This is done in order to allow samples to pass
a stage they might have otherwise failed, because they had a large margin on previous
stages. Sochman and Matas [57] also implemented a similar monolithic design for making
decisions after every feature evaluation in their WaldBoost framework.
Another problem of the traditional cascade architecture design is the handling of multi-
view face detection (MVFD). MVFD has additional requirements in terms of run-time
performance and the classification task becomes more difficult. Huang et al. [25] propose
the use of width-first search trees in conjunction with a multi-class, multi-label version of
Real AdaBoost called Vector Boosting. The width-first search tree outperforms a simple
depth-first tree design because multiple paths may be taken simultaneously before making
a final decision. Trees in general allow for better multi-view classification, since the sample
sets at nodes become smaller and the discrimination from background samples is thus
easier due to less within-class variance. Other tree-based approaches are a pyramid-based
structure presented by Li et al. [34] and a tree-based approach by Viola and Jones [28]
with detectors trained for different views and rotations.
2.5.6 Training Time Improvement
Although cascades of boosted ensembles exhibit real-time run-time performance, train-

ing time ranges from days to weeks. Several approaches have been suggested to reduce
the required amount of time. The factors that affect the training time are the amount
of training samples and feature-set size that depends on the actual image dimensions of
training samples. The traditional training approach has a run-time of O (N T log (N ))
where N represents the number of samples and T is the number of features. The most
successful and straight-forward approach to reduce the training time is caching [68]. In
order to execute AdaBoost, the classification of each prospective classifier for each indi-
vidual sample has to be computed based on the current weight distribution. In the case
of Haar feature-based weak learners, the classification is a simple thresholded decision
on the scalar difference of rectangle areas. Therefore, for several iterations of AdaBoost,
only the weighting of samples but not the computed rectangle areas themselves change.
Precalculating and caching the differences of these rectangle sums for all samples is, thus,
the fastest way to train a face detector. Caching, on the other hand, is limited by the
amount of available memory, training with 40000 samples and a 20 20 window (~ 65000
basic features) requires several GB of RAM. Another method to reduce training time is
to avoid the constant reevaluation of feature values. A method that treats features as
high-dimensional random vectors and models the feature values with Gaussian distribu-
tions was presented by Pham et al. [44]. This approach avoids the recalculation of feature
values to determine optimal thresholds by using Gaussians to model the distribution of
feature values. Therefore the training time can be reduced to O(N d2 + T ), where d is the
amount of pixels within the probed feature region. Another possibility to reduce training
time is to select features directly to either decrease the false acceptance rate or to increase
the maximal hit rate directly using Forward Feature Selection as described in [69], but
the guarantees proved for boosting may not hold for FFS. Also, feature filtering has been
explored by Brubaker et al. [6], they explore methods to reduce the size of the feature
pool. They suggest the use of different filters. The simplest approach is a ranking filter
that eliminates features whose values have poor correlation with class labels. But that
may leave only redundant features whose combination is not useful. Therefore slower
filters based on pairwise mutual information of features have also been tried. They were
compared to a random filter selection of features. While ranking and mutual information
filters performed better than randomly selected features, there was a decrease in detection
rates of final classifiers with about factor four performance gains. The use of filtering to
eliminate Haar features did not meet the authors expectations.
Another approach to reduce training time is the Matrix-Structural Learning (MSL) ap-
proach by Yan et al. [71]. The approach extends bootstrapping to the positive set as well
as the negative set. Negative samples are bootstrapped at the end of each stages train-
ing, whereas positive samples are bootstrapped during the training of stages. Initially,
a cascade stage is trained with the current negative set and a random positive set, then
misclassified positive samples are added to the training set and the stage is retrained until
the detection rate on the overall global set is sufficient. Then, after bootstrapping the
negatives, the process is repeated for the next stage. See Chapter 3 for details.
2.5.7 Run-time Efficiency Improvement
Bourdev and Brandts [3] soft cascades allow to optimize both cascade performance and
run-time speed. Their approach produces monolithic classifiers that outperform compa-
rably fast cascaded classifiers. The trade-off between run-time efficiency and detection
performance can be optimized by analyzing receiver operating characteristics (ROC) sur-
faces instead of static ROC curves. Like the soft cascades, the work of Sochman and
Matas [57] and Xioa et al. [70] implicitly improves classification performance by allowing
for arbitrarily preempted points to stop execution of the classification.
McCane et al. [39] built an empirical cost model based on ROC families of a monolithic
classifier to model the trade-off between run-time and detection rate. They uses a cost
model that includes both false alarm and detection rates to optimize the cascade training
parameters.
Brubaker et al. [5] presented another method to optimize the run-time performance of a
cascade architecture by means of an abstract cascade execution model. This execution
cost model allows to decide the minimal amount of required weak hypothesis and a better
exploitation of the false alarm vs. correct detection rate trade-off by optimal choice of
stage thresholds. Additionally, existing stages can be split into smaller stages in order to
discard negative samples with the evaluation of less features. The work by McCane et al.
[39] and Brubaker et al. [5, 6] address the issue of choice of thresholds during training as
opposed to Luos [38] post-training threshold optimization strategy.
2.5.8 Classifier Combination
Classifier combination has been explored by Rowley et al. [45], Sung and Poggio [60] as
well as Viola and Jones [63]. Rowley et al. [45] were the first to build a face detection
system that incorporates a combination and arbitration of several classifiers. First, the
neural network-based face detector was run over the entire image and detections were
collected. Then, a heuristic was applied to these detections where detections were grouped
and single detections were removed, as face detections occur at slightly different positions
and scales whereas false positives usually do not. Finally, arbitration between multiple
networks was applied to further reduce the number of false alarms. The arbitration
schemes included logical AND, OR and voting operations as well as a separate network
trained to arbitrate detections of different classifiers.
Viola and Jones [63] first publication also contained a simple combination scheme to
improve classification performance. A simple majority voting was carried out between
three similarly trained cascades. They reported an increased hit rate and a reduction of
false alarms even though the cascade classifiers were similarly trained and thus redundant.
Grosvenor [22] presented a general object detection integration framework. The integra-
tion approach takes a set of cascades and statistical information of their interdependence
and creates a classification tree. The method is supposed to allow the automatic com-
bination of more specific detectors for problem subsets, e.g different poses in multi-view
face detection, into a general classifier that is more efficient than parallel evaluations of
the individual classifiers.
23
3 Methodology
This chapter presents the basic principles used within this work and explains their theo-
retical basis.
3.1 Face Detection
Face detection is the process of determining whether a face is present within an input
image and, if present, returning the location and size of the human faces. The following
sections explain the basic concepts that are used to perform face detection in this work.
One key criterion for face detection systems is real-time run-time performance. Viola and
Jones [63, 64] were first to present a state-of-the-art performance, real-time face detection
system. The concepts that allow for highly accurate object and face detection at high
speeds are explained in this Chapter.
Haar wavelets form the basic features for face detection, they are simple but fast to com-
pute. Face detectors are trained using AdaBoost, a powerful machine learning algorithm.
In order to further improve detector performance, a chain or so-called cascade of classifiers
is used to reduce the average execution time of classifiers.
3.1.1 Haar Features
Haar features are simple rectangular features reminiscent of Haar basis functions first
suggested for use in face detection by Papageogiou et al. in [41]. The over-complete
feature set consists of weighted combinations of two or more rectangles intensity sums.
Individual features can described as
X
f eatureI = i RecSum(ri ), (1)
iI={1,...,N }
with weights i , rectangles ri and number of rectangles N , the rectangle area intensity
sum function RecSum(r) will be defined later. The weights are restricted to opposite
signs in order to yield rectangle area differences, the number of rectangles is usually
restricted to a maximum of four, although more elaborate features are possible. The
features resemble Haar-basis functions, they are constructed to capture edges, lines and
24 3 METHODOLOGY
diagonals, sometimes center-surround features are added to the set. See Figure 1 for
examples and Lienhart et al. [35] for details.
3.1.2 Integral Image Representation
A contribution of Viola and Jones approach to object detection [65] was the use of integral
images to calculate rectangle sums necessary for the calculation of the Haar-like features,
thus allowing fast computation and evaluation. An integral image, also referred to as
summed area table (SAT), contains the sum of an input images pixel intensities within a
rectangle from the top left corner of the image to the current position (x, y). The integral
image value ii(x, y) at coordinates (x, y) is defined as
X
ii(x, y) = I(x0 , y 0 ), (2)
0x0 x,0y 0 y
with I(x, y) being the image pixel intensity at coordinate (x, y).
The integral image can be calculated with a single pass over the image as
ii(x, y) = ii(x, y 1) + ii(x 1, y) + I(x, y) ii(x 1, y 1), (3)
with ii(1, y) = ii(x, 1) = ii(1, 1) = 0.
Using the integral image representation rectangle area sums can be calculated using only
four lookups. The rectangle sum RecSum(r) with rectangle r = (x, y, w, h) at position
(x, y) with width w and height h, is defined as
RecSum(r) = ii(x 1, y 1) + ii(x + w 1, y + h 1)

ii(x 1, y + h 1) ii(x + w 1, y 1). (4)
Figure 1: Example Haar-like features consisting of two, three and four base rectangles.
3.1 Face Detection 25
Lienhart et al. [36] proposed an extension to the set of Haar-features to include rotated
features. The calculation overhead is only slightly increased as the rotated summed area
table can be calculated using in a single pass over the input image as well. The rotated
summed area table (RSAT) is defined in analogy to the summed area table above, but
represents the sum of pixel intensities of rectangles rotated by 45 whose left corner lies
at the image origin and the bottom corner at (x, y). The RSAT at point (x, y) is defined
as
X
RSAT (x, y) = I(x0 , y 0 ). (5)
x0 x,x0 x|yy 0 |
The calculation of RSAT elements can be performed using the following formula
RSAT (x, y) = RSAT (x 1, y 1) + RSAT (x + 1, y 1)

RSAT (x, y 2) + I(x, y) + I(x, y 1), (6)
with RSAT (x, y) = 0, x, y : x < 0, y < 0.
Rotated rectangles, rrot = (x, y, w, h), pixel intensity RecSum(rrot ) can thus be calculated
using
RecSum(rrot ) = RSAT (x h + w, y + w + h 1) + RSAT (x, y 1)

RSAT (x h, y + h 1) RSAT (x + w, y + w 1). (7)
3.1.3 Boosting
Boosting is a meta-algorithm used to improve the classification performance of a base

classifier, also called a weak classifier, by continuously combining weak classifiers trained
with different sample weights into a final, accurate classifier called a strong classifier. The
basic concept of boosting is the combination of simple base classifiers hi into an accurate
classifier f (x), where x denotes a sample:
26 3 METHODOLOGY
T
X
f (x) = t ht (x).
t=1
The t denote coefficients or weights of the individual weak classifiers in the final strong
classifier.
Boosting is based on the idea that many simple weak classifiers are easier to construct
than a single complex, strong classifier. Therefore, boosting algorithms combine several
weak classifiers into a strong classifier that is nonetheless able to perform the desired clas-
sification task well. Boosting algorithms iteratively run a base learning algorithm with a
set of training examples that is differently weighted each round to put emphasis on mis-
classified samples. The two main challenges of boosting are the choice of training sample
weights each round and the combination rule for building a strong classifier from weak
classifiers. Boosting algorithms put emphasis on training samples that are misclassified
by increasing those samples weights each round. The final classifier is a simple linear
combination of weak classifiers.
The boosting approach is independent of the base machine learning algorithm, also called
weak learner or weak classifier. Several algorithms have been used in conjunction with
boosting, for example decision trees have been found to work well for certain classification
tasks. Face detection applications use very primitive weak learners, simple thresholded
features. The threshold is set so that the feature values after projection separate the
classes as well as possible.
Several boosting algorithms have been devised, AdaBoost (Adaptive Boosting) by Schapire
and Freud [19, 51, 52] is the most popular boosting approach. Schapire and Freud were
the first to present a practical polynomial time algorithm that found wide acceptance. A
schematic overview of the boosting algorithm can be seen in Algorithm 1. AdaBoost is
given a training set (x1 , y1 ), . . . , (xn , yn ) as input, consisting of samples xi from the do-
main or instance space X and labels yi from the label set Y , usually {1, +1}. The initial
weight distribution is set so that all samples have equal weights 1/n. The parameter t
is chosen to reflect the importance of a weak learner ht , in the binary case usually

1 1 t
t = ln , (8)
2 t
where t is the error of the classifier ht , i.e. probability of ht misclassifying samples xi X

given the current weight distribution Dt . Each round, the distribution of weights Dt is
updated according to the rule in Algorithm 1.
By adding a weak classifier to the ensemble that minimizes the error given the current
weight distribution and weighting it according to Equation 8, AdaBoost decreases the
training error, i.e. the number of misclassified training samples, exponentially. It does
so by aggressively maximizing the so-called margin of misclassified training examples by
increasing their weight. The choice of Dt determines how fast the training error can be
reduced. A choice of Dt according to Algorithm 1 reduces the training error as fast as
possible. The margin of a training example xi is defined as
PT
yi t ht (x)
m(xi ) = Pt=1
T
. (9)
t=1 |t |
It lies between [1, +1] and correlates with the confidence in the classification of sample
xi . Therefore, it is positive if a sample xi is correctly classified. Large margins have been
proved to yield lower generalization errors on unseen testing data. AdaBoost has been
shown to maximize the margin aggressively, since it concentrates on those examples that
have the smallest margins. As can been be seen in Equations 9 and 8 t is chosen to
minimize the classification error of samples.
Algorithm 1 AdaBoost boosting algorithm[50]

Initialize D1 (i) = 1/n
For t = 1, . . . ,T:
Train base learner using distribution Dt .
Get base classifier ht : X R.
Choose t R,
Update:
Dt+1 (i) = Dt (i)exp(t yi ht (xi ))
Zt
, Zt is chosen so that Dt+1 will be a distribution
Final classifier:
P
T
H(x) = sign h
t=1 t t (x)
28 3 METHODOLOGY
3.1.4 Weak Learners
Boosting algorithms do not rely on a pre-determined base learning algorithm. Therefore,

the choice of a weak learner to be boosted remains. Popular choices are decision trees.
Viola and Jones [63] initially proposed the use of thresholded Haar-features as weak
learners. A Haar-feature is evaluated on the corresponding input image sample. If the
feature value exceeds a certain threshold, the output is 1, otherwise 0. The threshold is
chosen to satisfy the stages minimal detection rate on all training samples. Lienhart et
al. introduced Classification and Regression Trees (CART) as weak learners [35].
This work uses simple thresholded weak learners constructed from Haar-features. Using
CARTs can slightly improve performance, but studies do not agree [35, 39]. The perfor-
mance gain does increase the complexity of features and increases the difficulty of cascade
performance analysis. Therefore simple thresholded Haar-features were used.
3.1.5 Cascades of Boosted Ensembles
In order to decrease the average execution time of rare event detection systems, for ex-
ample face detection, the cascade architecture has been introduced. Rare event detection
tasks typically consist of a very imbalanced ratio of positive and negative samples during
testing, thus the majority of execution time is spent rejecting negative samples. Viola and
Jones [64] inspired the wide-spread use of cascaded detectors, especially in face detection.
The basic idea of cascaded detectors is to reduce the average execution time by applying
a chain of detectors that mimic a degenerate decision tree. For each presented testing
sample, several decisions are made whether the sample should be rejected as negative or
passed to the next element of the chain for further inspection. Only if the sample passes
all classifiers of the chain, it is accepted as a positive sample. The chain, or cascade,
architecture is organized so that the first classifiers mainly deal with simple samples and
let later stages handle more complicated decisions. Thus, early classifiers are kept simple
and the execution time low. Stages need to have very high detection rates, while fairly
high false acceptance rates around 50% are acceptable. That means that the first stage
is already able to reject 50% or more of all presented negative samples in a time-efficient
manner. There is no need to evaluate all stages as in a slow monolithic classifier. All
stages are only executed in the rare case of classifying a positive sample.
The global detection rate Dg and false acceptance rate Fg are the product of the individual,
chained stage detection and false acceptance rates di and fi of stage i. They are thus
defined as
YN
D= di , (10)
i=0
N
Y
F = fi (11)
i=0
in a cascade with N stages.
In face detection, cascades consist of classifiers trained via AdaBoost based on thresh-
olded Haar-feature weak learners. Each stage of the classifier is trained using a modified
AdaBoost approach. AdaBoost is used to select and combine the features into a strong
classifier. A weak classifier is constructed from a Haar-feature by thresholding the re-
sulting feature value h(x), so that the error on all training samples is minimized. Then,
weak classifiers are added to the stage ensemble until desired stage false acceptance and
detection rates are met. Equations 10 and 11 give rough ad-hoc guidelines for choosing
the detection rates. Viola and Jones [64] therefore chose all stage detection rates to be
set at di = D1/N and fi = F 1/N . Accordingly, each stage has to have a very high stage
detection rate, in excess of 99.9%, whereas false acceptance rates may be moderately high
around 50% for 20-stages architectures for example. These choices are not necessarily
optimal.
3.1.6 Optimal Thresholding for Cascaded Ensembles
A cascade architecture consists of a number of stages with individually defined operating

points. During training, several parameters have to be optimized - the stage classifiers,
i.e. the features, have to be constructed, optimal thresholds have to be determined for
each stage and the number of stages has to be decided. Viola and Jones [64] pointed
out that optimizing all these parameters simultaneously is not trivial. Therefore, ad-hoc
stage detection rate and false alarm rate targets, that is stage thresholds, were chosen that
may not be optimal globally. Also, the number of stages is fixed initially or the training
is aborted early if the goals cannot be met after a number of stages. Luo [38] devised
a high-level abstraction model for representing cascade architectures. In this model, a
cascade consists of n sequential stages, each stage consists of a thresholded node classifier
that projects input samples xi to a class label yi Y = {1, +1}. Each stage classifier
30 3 METHODOLOGY
is represented as a function C(t, xi ) Y , or simply C(t), whose decision boundary is

defined by a threshold t. The goal of the abstraction model is to devise a method to
optimize the stage thresholds ti for all stages i n, since the previous ad-hoc targets may
not be ideal. Reducing the threshold and false alarm rate, thus raising the detection rate,
of stage i and increasing the threshold of stage j, i 6= j, thus lowering its false alarm rate
may raise the overall detection rate if done correctly.
3.1.7 Bootstrapping
The term bootstrapping is derived from pulling oneself up by the strap of the boot
and describes a method to incrementally train a classifier from an exhaustively large
set of samples or, in general, a process that generates a complex system from a simple
initial system. It is an iterative method that enlarges an initial set of rules or samples
by generating new rules or acquiring new samples with help of the current state. In
classification, first, an initial population is randomly sampled from the large set of samples.
Then a classifier is trained with this initial set. In order to improve the performance, the
classifier is forced to focus on misclassified samples. Therefore, after each training step
the training set is augmented with a number of misclassified samples and the classifier is
retrained. This helps the classifier to correctly classify a larger set of samples and leads
to a steady improvement of classifier performance.
3.1.8 Matrix-Structural Learning
The main drawbacks of Haar feature-based face detection are the immense amount of
base features and the need for a large number of training samples. These necessities, in
conjunction with a boosting training scheme, result in long training times, in the order of
days. Several approaches have been presented to reduce training time, see Chapter 2.5.6.
Of these, the Matrix-Structural Learning [71] approach has the least limitations. Other
approaches discard possibly valuable information, require large amounts of memory, do
not evaluate all possible features or may break AdaBoosts proved generalization proper-
ties. The Matrix-Structural Learning approach extends the idea of bootstrapping to the
entire training set instead of just the negative samples as usually done. The approach is
illustrated in Figure 2.
Matrix-Structural Learning (MSL) works by alternating a positive and a negative boot-

3.2 Entropy and Mutual Information 31
strapping phase. In the original Viola and Jones [64] approach, the bootstrapping process
is limited to negative samples in between stages. This ensures that later stages have to
deal with a more challenging classification task, hence the specialization of later stages on
hard samples and their increased computational complexity. Usually, a training set con-
sists of a number of cropped face images and several large background images without any
faces. A very large amount of negative samples can be extracted from these background
images. Enormous positive training sets might make the Viola and Jones [64] training
approach unfeasible because of very long training times. Therefore, MSL introduces an
additional bootstrapping phase of positive samples.
Initially, from the original enormous sample set (ESS) with m positive samples a small
random positive set of n < m samples is drawn. C(i, j) will denote a trained classifier
for stage i that has been trained with a training set from the jth bootstrapping iteration.
This initial set P (1, 1) forms the positive training set for the first cascade stage. The
initialization of the negative set N1 is performed similarly to the original approach, by
random selection of samples from images known not to contain any face images. By means
of the initial training set consisting of P (1, 1) and N1 a first classifier is trained with a
target false alarm rate fmin and a target detection rate dmin . Iteratively, each newly
trained classifier C(i, j) is evaluated on the whole ESS. If the target detection rate dmin
is missed, a number of positive samples that have been misclassified is added to P (i, j)
to form the new positive training set P (i, j + 1) for the next training iteration. If the
detection rate target is met, the classifier C(i, j) is the final classifier for stage i, denoted
as C(i, Mi ). After a stage has been finalized, the negative set is bootstrapped by creating
a new negative training set Ni+1 that consists of samples gathered form misclassifications
of C(i, Mi ) on the background training images without faces. This alternating process
continues until the desired number of stages has been trained.
3.2 Entropy and Mutual Information
Information theory is concerned with the information content or randomness of signals,

i.e. the complexity of signals. A key concept of information theory is entropy, a measure
for the randomness or uncertainty of a signal. Instead of signals, more generally random
variables are quantified. So, given a set of n symbols or a random variable X with n out-
comes, i.e. X = {x1 , x2 , x3 , . . . , xn }, and corresponding probabilities for their occurrences
32 3 METHODOLOGY
Figure 2: Matrix-Structural Learning overview [71] - alternating positive and negative

sample bootstrapping phases.
{p1 , p2 , . . . , pn }, entropy is defined as the expected information content

n
X X
H(X) = pi log(pi ) = p(x) log (p(x)) .
i=0 xX
Conditional entropy measures the amount of uncertainty of a random variable X given

the value of another random variable Y. Therefore, conditional entropy is a measure for
the remaining entropy of X after the value of Y is known. Conditional entropy is defined
as
X X
H(X|Y ) = p(y) p(x|y) log (p(x|y))
yY xX

X p(x, y)
= p(x, y) log = H(X, Y ) H(X),
yY,xX
p(y)
where H(X, Y ) is the joint entropy of X and Y . Another important concept in information
3.3 Support Vector Machines 33
theory is mutual information that measures the information shared between two random
variables X and Y , i.e. the information of one variable that can be recovered by observing
the other. The mutual information I(X; Y ) is defined as

X p(x, y)
I(X; Y ) = p(x, y) log .
xX,yY
p(x) p(y)
Conditional mutual information represents the mutual information of two variables con-
ditioned on a third variable. It is defined as

X p(x, y|z)
I(X; Y |Z) = p(z) p(x, y|z) log .
zZ,yY,xX
p(x|z) p(y|z)
3.3 Support Vector Machines
Support vector machines (SVMs) are maximum margin binary classifiers that solve a
classification task using a linear separating hyperplane in a high-dimensional projection-
space. This hyperplane is chosen to maximize the distance between positive and negative
samples. Real-world problems seldom present linearly separable data, therefore a trans-
formation into a higher-dimensional space is applied before classification with hopes of
being able to linearly separate data in the new space. The advantage is that the hyper-
plane does not need to be projected down into the original space, instead the classification
takes place right in the high-dimensional space implicitly. SVMs are conceptually simple
yet powerful and the results are interpretable, all good reasons to employ SVMs. Further
introduction into SVMs can be found in [42, 54].
3.3.1 Linear classification
Let {(x1 , y1 ), (x2 , y2 ), . . . , (xm , ym )} again denote the training data, consisting of a training
vector xi and a corresponding classification class yi {1, 1}. Under the assumption that
the provided data can be separated linearly in an n-dimensional space, we can construct
an infinite amount of n 1 dimensional hyperplanes that correctly separate the training
data, because there are no restrictions on placement or orientation of the hyperplane as
long as the data is correctly classified. The idea of maximum margin classification is then
to choose the hyperplane with the maximum separating margin between the two classes
34 3 METHODOLOGY
Figure 3: Linear classifier and margins - the margin is proportional to the expected
generalization ability. Taken from [42].
because it can be expected that this maximum margin hyperplane is best at generalization,
i.e. the margin is proportional to generalization ability of the classifier.
A hyperplane can be described as
{xS : wx + b = 0, (w, b)S R} (12)
and the maximum margin separating hyperplane has to minimize the condition
mini=1...n |wxi + b| = 1 where xi are the training examples. Consequently, the distance
w(xi xj )
between two samples xi and xj relative to the hyperplane can be defined as kwk
.
Then the distance between the two classes is kwk 2
. So in order to classify training data
correctly the hyperplane can be found by maximizing kwk 2
or minimizing kwk2 under the
condition
yi (wxi + b) 1 for i = 1 . . . n, (13)
that guarantees that all samples are correctly classified. This minimization can be achieved
3.3 Support Vector Machines 35
using Lagrange multipliers once rewritten into

n
1 X
Lp = L(w, b, a) = kwk2 i (yi (wxi + b) 1) . (14)
2 i=1
with 1 , 2 , . . . , n being Lagrange multipliers. After solving the optimization prob-

lem most i are zero because their conditions are fulfilled. Those xi whose i > 0
are chosen as support vectors to represent the margins, they are the closest vectors
to the hyperplane. Then w can be computed as a linear combination of these i :
w = ni=1 i yi xi .
P
3.3.2 Soft-margin linear classification
In order to allow a certain number of misclassifications a soft-margin is introduced. The

optimization condition is changed to yi (wxi + b) 1 i for i = 1 . . . n, i 0, where i is
the sample xi s distance from the correct margin, it is sometimes also referred to as slack
term. Therefore if i > 1 the sample is misclassified, if 0 < i < 1 the sample is correctly
classified but within the margin (margin error) and if i = 0 the vector lies on the margin.
Consequently, the minimization term becomes
n
!
X
min kwk2 + C i (15)
w,b,i
i=1
where C is a weighting parameter that controls the rate of misclassifications. Small values
of C maximize the margin, large values of C yield few misclassifications.
So soft-margin classifiers allow a certain number of misclassifications and can therefore
better cope with data that is not exactly linearly separable.
36 3 METHODOLOGY
3.3.3 Non-linear classification
Given that most real-world data is not linearly separable, the data has to be altered in
some way in order for the previous linear maximum-margin classifiers to be useful. The
idea is to transform the data into a higher-dimensional space and scatter the data suitably
so that it can then be classified using linear separation.
The transformation is usually of the form : Rn Rm , m > n. But transformations and
computations in high-dimensional spaces are usually computationally expensive. There-
fore the so-called "kernel-trick" is employed, a kernel function is defined as the dot product
of the projection of two vectors - K(x, y) = (x) (y). As the previous equations ex-
clusively use dot products in the high-dimensional space there is no need to explicitly
transform the data or to transform the hyperplane. Instead, all equations can be evalu-
ated using kernel functions. Popular kernel functions are
Polynomial:
K(x, y) = (x y + c)d (16)
Radial basis functions:

kx yk2
K(x, y) = exp (17)
2 2
Sigmoid:
K(x, y) = tanh((x y) + ) (18)
c, d, , and are parameters and have to be chosen to optimize the classification.
Figure 4: Non-linear classification using kernel function. A kernel function is used to

transform data into a higher-dimensional space where the data is linearly separable.
37
4 Cascade Combination and SVM-based Training Sam-

ple Selection
Building real-time face detection systems is still an open research issue. The real-time
requirement of many applications requires fast detectors with high precision and low
false alarm rates. The work of Viola and Jones [63, 64] proposed a very promising and
now wide-spread approach. While many aspects of the approach have been in the focus of
research and several improvements were published, training state-of-the-art face detection
systems remains a challenge and achieving good results requires large amounts of time for
parameter tuning, training set creation and sample selection. Both the classifier training
parameters and the training set impact the final classifiers performance.
It is widely believed that training with a large training set improves classification results
due to the better coverage of variance in the data. But Chen et al. [12] have shown the
importance of selecting suitable training samples for face detection and that shrinking
the training set may actually be beneficial. By means of expanding and then down-
sampling positive training examples they have created high-performance face detectors.
Their algorithm utilizes a sophisticated method based on manifolds. Part of this work
attempts to achieve similar results by simpler means. The basic idea to sample selection is
not new. Especially in speech recognition and grammar parsing, sample selection played
an important role. Speech recognizers and grammar parsers require large quantities of
training examples, usually more than can be labeled. Therefore several sample selection
approaches were explored in order to pick the most promising training samples to be
labeled and used for the next training round. In face detection, Chen et al. have created
a large positive training set consisting of more than 100,000 samples overall. The problem
here though is not an insufficient amount of samples, but too many samples. As Chen et
al. have shown, using optimal training samples can improve the classifiers performance.
Another problem that this work tries to address is to overcome possibly non-optimal
selection of training samples by bootstrapping and non-optimal training parameters. Es-
pecially in asymmetric detection tasks as face detection, training heavily relies on boot-
strapping to select suitable negative examples from the millions of available samples.
Usually, according to the bootstrapping algorithm, those samples are added to the train-
ing set that are misclassified by the classifier of the previous training iteration. This
training scheme, along with the selection of positive training samples, introduces random-
38 4 CASCADE COMBINATION AND SVM-BASED SAMPLE SELECTION
ness into the training process. Two classifiers trained with identical parameters usually
exhibit different performance. Additionally, classifiers with different training parameters
and training sets may perform better in particular situations. In order to exploit these
findings, classifier combination has been explored by Rowley et al. [45]. They use simple
arbitration schemes of several neural networks to improve the overall classification perfor-
mance. Viola and Jones [63] also presented a system consisting of three parallel cascades
that were combined by voting, Sung [60] used neural fusion to combine classifiers. The
problem with these approaches of combining classifiers is their run-time performance im-
pact. They require to run several classifiers in parallel.
Training good classifiers with several stages is a time consuming task. Training time
increases tremendously as the amount of memory needed to cache feature values surpasses
the available memory. Then, many calculations have to be repeated for each feature
selection iteration. The cache size depends on the number of training samples and the
number of features, thus training several cascades in parallel on parts of the training
set would be a straight-forward way to minimize training time as that limit is reached.
Another possibility to reduce training time by combination stems from the fact that
stage training times increase with the number of samples and especially with the number
of stages. After a certain point, the negative bootstrapping process is the most time
consuming task.
4.1 Cascade Training
In order to evaluate the effect of large amounts of training data on classifier performance,
several classifiers had to be trained with many training samples. Training cascade classi-
fiers is tedious and time consuming, since a feeling for the influence of several parameters
has to be acquired first. The number of positive and negative samples, each stage target
hit rate, the ratio of positive and negative samples and the number of stages have to be
determined. In order to train classifiers, especially with a large amount of training samples
in excess of 100.000 samples, optimized training algorithms are necessary. Caching has
the largest impact on training time without having any side-effects on the final classifiers
performance, but it is limited by the amount of available memory, the number of sam-
ples and the number of features, i.e. feature set and window size. Once the memory for
caching runs out, each feature value for every sample has to be recomputed every feature
selection round. Thus, the training performance degrades sharply. In order to overcome
4.2 Cascade Combination 39
this limit, the Matrix-Structural Learning [71] approach was implemented.

See Chapter 3.1.8 for details of the Matrix-Structural Learning algorithm. The imple-
mentation was straight-forward. The training algorithm was extended with another loop
to bootstrap the positive sample set in between the negative bootstrap executions until
the target detection and false alarm rates are met on the extensive sample set.
4.2 Cascade Combination
Generally, cascade combination attempts to combine the stages of several pre-trained

cascades into a new stronger cascade. Cascade combination is supposed to bring the
benefits of combining several cascades in parallel, i.e. the benefits achieved by voting or
other arbitration techniques, without the need of running cascades in parallel. Also, when
combining cascades, several cascades have to be trained, but the amount of necessary
samples and stages is lower to achieve a similar performance. Therefore, the overall
training time can be less, depending on the extent of caching allowed by available memory,
feature set, and training set size.
4.2.1 Optimization Problem Formulation
The cascade combination task can be formulated as an optimization task. The goal of
cascade combination can be defined as selecting k optimal stages (1), . . . , (k) that max-
imize a cost function from a set of available stages taken from several cascades. Cascade
combination is quite similar to feature selection which has been studied widely in the
field of machine learning. Fleuret [17] has studied feature selection for face detection in
combination with several classifiers. Although cascade combination can be examined in
the same framework there are a few crucial differences. First, the choice of the optimal
feature or stage is slightly different.
1. In feature selection a good next feature has a large correlation with the class label
given all previous features classifications. But no additional information is gained
by adding another feature that highly correlates with the class labels and makes the
same errors as previously selected features, i.e. a feature that is redundant. Since
positive samples are only accepted by the cascade if they pass all stages, all stages
have to correctly classify positives. Thus, the decisions for positive samples need to
be coherent and redundant, as opposed to feature selection.
2. During feature selection, the final combination scheme of all selected features is
not usually easy to evaluate incrementally. The simple decision making of cascade
classifiers, on the other hand, is. Therefore, large lookup tables (caches) may be
computed that store each stages feature sum for every training sample. The clas-
sification vector can then be calculated by thresholding these stage feature sums
given stage thresholds. The combination of previous stages decisions is easy and
fast, the final decision is positive if all stages accept the sample, negative otherwise.
Due to these two differences, standard feature selection algorithms do not perform quite as
well in cascade stage combination. On the other hand, these properties may be exploited
to optimize the decisions.
Given a training data set consisting of m samples with label yi and training sample
xi , i.e. {(y1 , x1 ) , . . . , (ym , xm )}, n cascade stages Sn from different classifiers, we can
construct a class label vector Y and each stages classification vector sn . The class label
vector Y = (y1 , . . . , ym ) contains the binary class labels for all training samples and
each stages classification vector sn = (Hn (x1 ), . . . , Hn (xm )) denotes stage Sn s strong
classifiers output Hn (xi ) for sample xi . The class labels yi and the strong classifier
outputs Hn (xi ) are binary, i.e. yi , Hn (xi ) {0, 1}.
Therefore, cascade combination algorithms try to select k final stages S(1) , . . . , S(k) , each
iteration the next stage to select is determined by

(l) = argamxn cost(Y, sn , {s(1) , . . . , s(l1) }) .
The cost function takes the class labels Y, the target stage sn s classification vector and
all previously selected stages classification vectors as inputs. The following cost functions
were evaluated:
conditional mutual information maximization:
costCM IM (Y, sn , S) = minsk S I(Y ; sn |sk ),
mutual information maximization:

costM IM (Y, sn , S) = I(Y ; sn ),
correlation:
costcorrelation (Y, sn , S) = r(Y + , s+

n ).
sn is the classification vector of stage n only,
sn is the classification vector of stage Sn consisting of only negative samples that have
passed the selected n 1 stages and all positives, Y are the corresponding labels,
s+
n,Y
+
are the same as sn , Y but only positives that passed all previous stages are
included.
The corresponding cost functions will be explained in the following Chapters.
4.2.2 Cascade Combination with Threshold Optimization
Cascade stages are trained to be used in conjunction with certain other stages. Since each
stages negative training data is bootstrapped by using the previously trained stages, its
stage threshold has been chosen on training data that does not take previously discarded
samples into account. Additionally, several researchers have shown that the independent,
ad-hoc choice of stage thresholds as proposed by Viola and Jones [64] is not optimal.
Therefore, stage thresholds have to be re-adjusted when different cascades stages are
combined.
The optimization of stage thresholds is done using the same cost functions as outlined
in the previous section and described in more detail later. Instead of trying to find the
optimal next stage, here the threshold ti of the current stage Si is optimized. Let stn
denote the classification vector of stage Sn after its threshold has been modified to t. The
best threshold for the current stage maximizes the given cost function cost(), i.e.
ti = argamxt cost(Y, stn , {s(1) , . . . , s(l1) }) .

The range of possible values of t is restricted by the possible range of feature values
Hn (xi ) and extreme cases need not be evaluated. We performed a linear search to find
the optimal threshold for t [2.5, +2.5].
4.2.3 Cascade Combination with Conditional Mutual Information Maximiza-

tion
Conditional mutual information maximization (CMIM) has been used as a feature selec-
tion scheme in machine learning (ML) before. Fleuret [17] provides a good introduction.
In the CMIM scheme, in order to combine several stages of different cascades, we build a
new cascade from existing stages by iteratively choosing the next best stage to be removed
from the pool and added to the final cascade. We try to select the next stage sm which
shares the least information with all previously selected stages, but correlates well with
the true class labels Y , where m 1 stages have already been selected. As stated before,
a strong correlation with the class label is not useful if most stages correctly classify the
majority of samples, i.e. they are each individually good, but make the same, redundant
errors.
Ideally, we would maximize the term
(19)

max I Y |s(1) , s(2) , . . . , s(k)
that denotes the joint mutual information between the class label Y and all stages to
select. Optimizing that term, specifically estimating the required probabilities, is compu-
tationally infeasible. Maximizing the mutual information between each selected stages
classification vector and the class labels does not necessarily yield optimal results since
the chosen stages are examined completely independently, thus they may be redundant
and the final combination may not be ideal. A compromise is the use of CMIM.
Conditional mutual information is an estimate for the information that X1 and X2 share
when X3 is known - I(X1 , X2 |X3 ) = H(X1 |X3 ) H(X1 |X3 , X2 ) . So I is large if X3
carries information about X1 that was not already revealed by X2 , if X2 and X3 contain
the same information I is zero. In the cascade combination case, therefore, the optimal
combination of stages is the one that maximizes the conditional mutual information of all
selected stages, i.e. max(I(Y |Xs(1) , Xs(2) . . . . , Xs(n) )). Since that terms calculation is not
feasible, an approximation suggested by Fleuret [17] is to select each feature, or stage in
our case, so that

argmaxn minkn I(Y |sn , s(k) ) ,
i.e. select the stage that has the maximum conditional mutual information matched with
the worst case previously selected stage.
The first stage is simply selected by maximizing the mutual information between the class
labels and the stages classification, i.e.
argmaxn {I(Y |sn )} . (20)
While CMIM was shown to perform well for feature selection, the use of conditional
mutual information for cascade combination has limitations. The main problem is the
different decision process for positive samples when combining cascade stages. In feature
selection, different features ideally produce different classification results that are then
ultimately consistent with the class labels when combined. In cascade combination, the
non-redundancy of stages is only desirable for negative samples, whereas all classifiers
have to make the same decisions on positive samples in order to accept them.
4.2.4 Cascade Combination with Mutual Information Maximization
While CMIM has been shown to outperform mutual information maximization (MIM) for
feature selection tasks [17], the different prerequisites for cascade combination make it less
useful than MIM for cascade combination. Since the combination scheme of classifications
simple in cascade classifiers, the classification vector of all previously selected stages can
be obtained efficiently by caching stage classifications of samples. Therefore, instead
of optimizing the ideal global mutual information as in Equation 19, we can greedily
approximate its optimization by continually maximizing the mutual information between
the class label and the joint classification vector of all previously selected stages and the
current stage.
The optimization term then becomes
argmaxn {I(Y ; sn )} .
Y and sn require some more explanation. Since we cannot optimize Equation 19, an
approximation is necessary. The joint mutual information of k arbitrary stages that

need to be selected and the class label, as described in Equation 19, cannot be feasibly
optimized, since it requires the estimation of 2k+1 probabilities. Hence, the ideal choice of
the next stage is approximated by choosing the next stage, so that the classification result
of all previously selected stages and the candidate stage has maximum mutual information
with the class label. Samples that were rejected by previous stages should not have any
further influence on the current decision for the next candidate, since even a correct
classification would not make a difference in the overall classification. Positive samples on
the other hand are important and ideally stages should not discard any positive samples,
so we require the next stage to comply with all positive labels, not just the remaining.
Therefore, the terms Y and sn denote the class labels of all remaining samples that have
passed all previously selected stages and the corresponding classification vector of stage
Sn . The calculation of Y and sn is straight-forward, Y can be attained by discarding
labels for rejected samples and sn can be generated with the help of a list of remaining
samples to classify and evaluation of stage Sn on those samples to yield the classification
vector. Since these operations have to be repeated each iteration, caching the feature
values of all stages is very helpful. The first stage is selected according to Equation 20.
4.2.5 Cascade Combination with Correlation Maximization
So far the use of mutual information and conditional mutual information for selection
of stages in cascade combination has been explored. They have been used to measure a
correlation between stages classification vectors, i.e. classification results, and the true
class labels. Therefore, the formulation of the same problem with a simple correlation
metric is obvious. Instead of the mutual information between classification and class label,
the linear correlation is optimized, another widely used feature selection strategy. The
difference is that linear correlation tries to minimize the mean-square error of a line fit to
the data and class labels, thus considering the linear relationship between classification
vector and labels of individual samples. Linear correlation is expected to perform less
well but is useful for comparison. If it turns out to perform better than CMIM or MIM,
there is strong evidence for further possible optimizations.
The optimization term is
argmaxn {r(Y , sn )} .
Y and sn are defined as outlined in the previous subsection.

4.3 Training Set Selection 45
4.2.6 Sample Extraction and Subsampling
Since classifier combination is unlike the usual classifier training and more like an opti-
mization of pre-trained cascades, the samples used for the cascade combination algorithms
are different from those used for training. One reason to use a different sample set is to
avoid overfitting of the classifier. Thus, the training sets for training and combination
are completely independent. Because cascade classifiers are applied in a sliding window-
fashion, i.e. moving a window of different scales over the input image and testing for
positive detections, there are multiple acceptable windows for a single labeled face. Some
approaches, like Multiple-Instance-Pruning by Zhang and Viola [77], take advantage of
this. This work indirectly acknowledges this by extracting numerous samples from man-
ually labeled full input images, instead of using cropped faces, and maintaining multiple
positive samples for a single labeled face, i.e. slightly offset positions and slightly different
scales - see Chapter 5.1.3 for details on the face matching metric.
One problem that occurs is that the imbalance of positive and negative samples may
bias the cost function, e.g. the prior probability of class labels influences the entropy of
the different classes. In order to mitigate the imbalance, the negative samples have been
subsampled. Subsampling in this case means that only every k-th extracted negative
sample from the input image is added to the combination data set.
4.3 Training Set Selection
As outlined in the beginning of Chapter 4 resampling promises to improve classification

performance while reducing cascade training time as long as the sample selection itself does
not negate the gain. Chen et al. [11] have explored resampling. They have employed a
data-centric approach by modeling the distribution of samples using manifolds, which are
lower-dimensional spaces that preserve the local neighborhood of an initial space. Using
manifolds they resample the training set to have the same density at all locations in the
manifold, underpopulated regions are filled with interpolated samples. This approach can
be described as data-centric because its goal is a uniform coverage of the positive sample
space. The approach chosen by us is more classifier-centric. Support vector machines
have recently become one of the most-widely used classifiers. Their advantage lies in
their conceptional simplicity and solid theoretic foundation while achieving very good
classification rates. Another advantage is that their internal state, i.e. the choice of
the separating hyperplane, is interpretable as opposed to some other classifiers. Support

vector machines construct a hyperplane that separates positive and negative samples. A
kernel function allows to project samples into a higher-dimensional space so that non-
linear decision functions in the input space are possible. The projections of the closest
samples to the decision hyperplane are called support vectors. The separating hyperplane
is chosen so that all vectors of each class lie on either side of the hyperplane and the
support vectors have a distance of 1 from the hyperplane. This distance from hyperplane
is also called the margin. Therefore, samples further away from the decision hyperplane
(higher margin) can be thought of as being easier to classify than those closer to it. The
hyperplane margin can then be used as an indicator for the difficulty or importance of
samples for the classification.
Unfortunately, support vector machines do not perform as well as classifiers of boosted
ensembles for face recognition. Their main drawback is the evaluation speed. Thus the
overall approach is to train an SVM-based face detector, use the distance-from-hyperplane
of samples to quantify each samples difficulty and then train a cascaded classifier based
on the resulting ranking of samples.
4.3.1 SVM-based Training Set Selection
Each stage of a cascade classifier architecture has the purpose of rejecting a certain subset
of samples as negatives. The bootstrapping process used to gather negative training sam-
ples forces later stages continually focus more and more on the decision surface between
positive and negative samples. SVM-based training set selection attempts to construct
a positive set so that the classifier is focusing on the decision hyperplane on the positive
side as well [43].
SVM-based Face Detector. In order to resample the training set with the help of sup-
port vector machines, a suitable face detector has to be built first. There have been several
approaches for face detection with SVMs. Heisele et al. [24] built a component-based face
detection system, where local feature-detectors based on linear SVMs are trained with
artificially generated samples and the components relative positions are verified with
another SVM. Earlier approaches used SVMs on raw input data [43].
The SVM-based face detector used for resampling is based on raw image intensity inputs,
no feature extraction is applied. The training set consists of a number of positive and
4.3 Training Set Selection 47
negative training samples whose raw intensity values are vectorized and used for training.
The positive samples are manually extracted and cropped face images, the same as used
for cascade classifier training. Since an almost arbitrarily large amount of negative samples
can be extracted from the training data, a bootstrapping approach such as proposed by
Sung and Poggio [61] and used by Rowley et al. [45] was also used here. First, an initial
set is randomly selected from large background images that are verified not to contain any
faces. Then, an SVM is trained with the given positive and negative training samples.
Next, misclassified negative samples are added to the training set. Once an upper bound
on sample set size is reached, bootstrapped samples are randomly replaced, in order to
avoid too strong imbalances.
Support vector machines have several parameters - the kernel projection and the corre-
sponding kernel parameters. In order to find the best parameters, a simple grid search was
used without bootstrapping the negative samples. A grid search simply tries all pairwise
combination of parameters, in this case RBF ((u, v) = exp (ku vk2 )) or polyno-
d
mial kernel ((u, v) = uT v ), and the kernel parameters for RBF, the exponent
d {2, 3} and for polynomial kernels and the misclassification cost parameter C that
allows for a number of misclassifications in order to avoid overfitting. A polynomial kernel
with d = 2, C = 10, = 1 worked best according to the grid-search. All components of
the input feature vectors were mapped from image grey-value intensities [0, 255] to [0, 1].
Lighting normalization has been widely employed in face detection and face recognition.
Therefore, normalization approaches were explored, namely histogram equalization. His-
togram equalization tries to increase the global contrast of images by spreading the image
intensities in the histogram. It is computationally inexpensive.
Support vector machines trained with histogram equalization performed better than those
trained on raw input data, the detection rate was higher for a given false alarm rate and,
after the same number of iterations, the SVM trained on normalized data achieved a
higher detection rate.
SVM-based Training Set Selection. Once the support vector machine is boot-
strapped and trained sufficiently, samples have to be selected for the following cascade
classifier training. The presented approach tries to focus the attention of classifier train-
ing on the decision boundary between positive and negative samples. Boosted classifiers,
among others, tend to overfit training data when not carefully chosen due to their property
of concentrating on misclassified data. Using only samples along the decision boundary,
a single negative outlier within a positive region would most likely split that region and
enforce a negative region within the region of positives. Therefore, while concentrating
on the decision boundary, sampling the whole population is helpful to reduce the effect
of outliers.
Two strategies for selecting samples were examined. Both try to force the trained clas-
sifier to focus on the decision boundary. The first strategy simply takes the l hardest
samples, where is l < n and n is the number of overall samples. The second strategy tries
to focus on hard examples without solely selecting those samples. When examining the
resulting SVM-based ranking of training samples, about 0.1% were misclassified as nega-
tive (distance-from-hyperplane dSV M < 1), 23.5% were not clearly negative or positive
(dSV M [1, 1]) and 76.4% were correctly classified as faces, dSV M was in the range of
[3.2, 5.2], dSV M < 0 for 0.6% and dSV M 0 for 99.4%. Therefore, the second strategy
for selecting l < n samples simply divides the range of distances into l segments and
selects a sample closest to the segment boundary for each segment. This approach covers
the whole range of distances, not just the hard samples. As most samples have positive
distances and only a limited amount of segments will cover positive dSV M , relatively more
samples with dSV M < 0 will be selected compared to the original distribution of samples,
thus focusing on hard training samples. In other words, more segments fall onto the fewer
negatively misclassified samples, therefore increasing their probability of being included
in the final set.
Cascade Retraining. Once the training set was resampled, the cascade was trained
in a straight-forward manner using Matrix-Structural Learning to speed up the learning
process. The resulting cascaded classifier has the same properties as any other trained
classifier, since the training process remains exactly the same. The only difference is the
choice of training samples that has had more attention.
49
5 Experiments
This chapter describes the experiments that were conducted in order to evaluate the
effectiveness of the proposed algorithms.
5.1 Experimental setup
This section describes the basic setup of the experiments, i.e. the data set, how face
bounding boxes for matching faces are created and how a detection is determined to be
correct.
5.1.1 Dataset
The data set used to evaluate the classifiers that were trained with the proposed ap-
proaches is the CMU and MIT database (CMU+MIT). The dataset consists of images
collected at CMU, Pittsburgh, PA, and MIT, Boston, MA. These images were initially
used to evaluate the neural network-based face detector developed by Rowley et al. [45].
Since then, the dataset has become the most widely used benchmark in face detection.
Most groups report results on the CMU+MIT dataset, thus the same set was used to al-
low for comparison with related work. Usually, results on 507 labeled faces are reported,
while the ground-truth labels obtained from the CMU neural network-face detection page
contain 512 faces. Potential extreme side-views or non-human faces may have been re-
moved in other work but specific information could not be found. The dataset consists
of 117 images of medium to high resolution, containing one to several faces. The images
are constituted of mixed subjects, from soccer teams, to theater pamphlets, to photos of
graduation classes. The provided CMU+MIT ground-truth labels mark positions of the
eyes, nose, as well as left and right mouth corners and mouth center.
5.1.2 Bounding Boxes
Our face classifiers return face bounding boxes, (x, y, w, h) consisting of the location (x, y)
and width and height (w, h) of the face detection. Since the ground-truth labels do not
explicitly provide face bounding boxes, these have to be generated from the data labels.
Face labels are provided as (x, y) locations of eyes, nose and mouth left and right corner
50 5 EXPERIMENTS
and the mouth center. These locations are denoted as leye, reye, nose, lmouth, rmouth
and cmouth with added subscripts x,y for the coordinate. A face bounding box is obtained
from the labels by taking
eyedistx = reyex leyex ,

1
f acecenterx = reyex + eyedistx ,
2
moutheyey = max(reyey, leyey ) cmouthy ,
x = f acecenterx eyedistx ,
w = 2 eyedistx ,
1
y = max(reyey , leyey ) moutheyey ,
2
h = 2 moutheyey .
5.1.3 Face Matching Metric
The definition of a correct or false detection is not trivial and most papers seldom report
exact face matching metrics. While detection rate by itself is not conclusive, since effi-
ciency and speed are also important factors, most authors solely report detection rates on
the CMU+MIT dataset. The detection rate d is the fraction of positive samples that were
correctly classified as face samples and the false acceptance rate f is the number of nega-
tive samples that were also classified as faces. The absolute number of false acceptances
is denoted as F . Given n test samples, npos positive and nneg negative of which mpos were
correctly detected and mneg were falsely accepted, detection and false acceptance rates are
defined as
mpos
d= ,
npos
mneg
f= .
nneg
Once face bounding boxes from face labels have been constructed, the problem of defining
a correct detection still remains. According to Lienhart et al. [35] a face is considered as
5.2 Receiver Operating Characteristic Curves 51
correctly detected, if
the distance of centers of bounding boxes of a detection and a labeled face is less
than 30% of the actual face width and
the size difference between the two bounding boxes, in our case both width and
height, is less than 50% of the actual face width.
The same detection metric was used in this work.
5.2 Receiver Operating Characteristic Curves
The receiver operating characteristics curve (ROC curve) is defined as a plot of detection
rate d vs. false alarm rate f . It provides a way of comparing binary classifiers, as well
as setting desired operating points where, for example, only a certain maximum false
acceptance rate is acceptable.
Usually, ROC curves are generated by modifying the classification threshold of the binary
classifier. Since cascades of boosted ensembles consist of several stages with an own
threshold each, this simple scheme cannot be applied straight-forward.
Since the work of Rowley et al. [45] heuristic post-processing steps have been widely
used. Neighboring detections are grouped and a face detection is only reported if at least
n neighbors are in a group. This method provides an effective way to reduce f with a
slight impact on d. Another way to influence the detection rate d indirectly is modifying
the scale s at which the sliding window is scaled up, allowing for a finer-scaled search
of faces. Additionally, the number of stages used by the classifier can be reduced and
thus d and f rise. Varying thresholds is more complicated, since m stages interact and
each stage i < m has an individual threshold ti . In order to explore all possibilities,
an exponentially growing number of thresholds would have to be evaluated. Therefore,
we vary the stage thresholds ti individually in k steps, within a range [ti t, ti + t].
In our experiments, t = 1.0, k = 4, n {0, . . . , 8}, s {1.01, 1.1} and the number of
stages was not reduced. We chose small scales s since larger scales barely contribute
to the interesting region of the ROC curve with values of F < 50. Since the cascade
architecture does not allow to vary the thresholds in a straight-forward way, several of the
aforementioned approaches have to be combined to estimate as many operating points as
52 5 EXPERIMENTS
possible. Overall, 96 parameters are evaluated per stage per image, resulting in several
thousand operating points. Nonetheless, many of these operating points lie at high false
alarm rates. Therefore, the following plots may not be perfectly convex because the
combination of the aforementioned parameters could not capture an optimal operating
point at a certain false alarm rate.
In order to smooth the plots, the following procedure was applied and is demonstrated in
Figure 5. Each point of the ROC curve (xi , yi ) has to have a y value that is greater than
the value of a point on the line connecting the previous point on the curve with any later
point, i.e. it has to satisfy
(yj yi1 )
yi (xi xi1 ) , j > i, i > 1.
(xj xi1 )
Thus, the smoothing operation removes data points whose value is below the connecting
line of the previous point with any following point. The reasoning is that these points can
be removed because the ROC generation algorithm only produced sub-optimal results.
A better data point should be producible with an optimal combination of the above
parameters and thresholds. The generation of these optimal points is impractical due to
the immense computational requirements necessary to find these points.
0.82
0.8
0.78
0.76
0.74
Hit Rate
0.72
0.7
0.68
0.66
0.64 smooth
unsmooth
0.62
0 10 20 30 40 50
False Acceptances
Figure 5: ROC curve smoothing. The originally generated ROC curve is compared against
the smoothed curve. Data points with smaller values than a line connecting any previous
and following points are removed.
5.3 Base Classifiers 53
5.3 Base Classifiers
5.3.1 Training Set
Several base classifiers have been trained to be used for classifier combination and as a
baseline to compare against for SVM-based resampling method. The global data set used
for training of these classifiers, unless otherwise noted, consists of 9000 positive face images
and 339 large background images. Negative samples are sampled via bootstrapping from
the background images that do not contain faces. Our training set is a subset of the
100,000 sample positive set created by Chen et al. [12].
The 9000 positive samples have been randomly partitioned into subsets consisting of 4500
(two sets) and 3000 samples (three sets), all subsets are disjunctive. These sets were used
to train the following base classifiers.
5.3.2 Base Classifiers
The following base classifiers were trained from the aforementioned training sets by using
Matrix-Structural Learning. The number of negative samples is equal to the number of
positive samples. Therefore, one set with 9000 negative and positive samples, two sets
with 4500 samples each and three sets with 3000 used samples each were available. The
classifiers were trained to produce 20 stages, width and height of the resulting classifier
were 20 20, a basic feature set with 41910 features overall was used and the minimal hit
rate per stage was dmin = 0.999 . No tree-based weak classifiers were used, only simple
tree stump classifiers which correspond to thresholded single features.
Figure 6, 7 and 8 show the ROC curves of classification results on the CMU+MIT dataset
of cascade classifiers trained on three different training set sizes. These classifiers were
trained with 3000, 4500 and 9000 positive and negative samples. If multiple sets of
training data were available, multiple cascades were trained. All resulting classifiers for
all available training data sets are shown below. Even though the training parameters
are the same, the cascades exhibit different performances. This is most likely due to the
different positive training sets and the bootstrapping procedure used for training. Since
the bootstrapping process starts with an initial random set of negative samples, these
classifiers were also trained with different negative sets. All training sets were of the same
size. These results indicate that a careful choice of the training has an influence on the
performance of the classifier.
54 5 EXPERIMENTS
0.8
0.7
0.6
0.5
Hit Rate
0.4
0.3
0.2 3000 pos+neg base cascade 1 - 20 stg

3000 pos+neg base cascade 2 - 20 stg
0.1
0 10 20 30 40 50
False Acceptances
Figure 6: ROC curve for cascades trained on 3000 positive and negative samples.
0.8
0.75
0.7
0.65
Hit Rate
0.6
0.55
0.5
0.45

0.35
0 10 20 30 40 50
False Acceptances
Figure 7: ROC curve for cascades trained on 4500 positive and negative samples.
Figure 8 compares all cascades that were obtained by training on the original training
set and all derived subsets. The cascades were trained with the same parameters as
described above. It is interesting to note that while training with 4500 samples yields an
improvement over 3000 samples, training with full 9000 samples results in no performance
improvement. Additionally, 3000 samples base cascade 2 seems to perform almost as well
0.8
0.7
0.6
0.5
Hit Rate
0.4

0.1
0 10 20 30 40 50
False Acceptances
Figure 8: ROC curve for cascades trained on 3000, 4500 and 9000 positive and negative
samples.
as the 4500 sample cascades. There is also a substantial difference in cascade performances
trained with the same parameters and amount of samples. Out of the cascade classifiers
trained with 3000 samples, the detection rate d of cascade 2 is 5% higher than the rivalling
cascades at higher false alarm rates, 4500 cascade 2 performs about 3% better than cascade
1. These considerations were, along with the work of Chen et al. [12], the motivations to
attempt to resample the positive training set.
The improvement of classification results with additional stages has been shown in Figure
9. Two of the three 25-stage cascades show an improvement over the 20-stage cascades. As
the figure illustrates, adding more cascade stages raises the ROC curve as is to be expected
since adding stages mostly means rejecting further negative samples while retaining a large
majority of positives.
56 5 EXPERIMENTS
0.9
0.8
0.7
0.6
Hit Rate
0.5
0.4
0.1
0 10 20 30 40 50
False Acceptances
Figure 9: Comparison of ROC curves of cascades with 20 and 25 stages.

5.3.3 Training Time
The need to train several cascades led to an interesting observation. Figure 10 shows that
after training about 20 stages, the bootstrapping time rises almost exponentially, whereas
training time itself (feature selection and threshold estimation) rises nearly linearly. The
values fluctuate considerably in Figure 10, in order to clarify the effect, the fluctuation
was reduced in Figure 11 and the linear increase is more visible. As can be seen from
the figures, overall training time is mostly dominated by the bootstrapping process and
not the stage training itself after about 20 stages. Two possible bottlenecks for the
negative bootstrapping have been examined, network-stored image files with slow access
and possibly inefficient selection of search windows within background images by frequent
changes between several background images. Slow file access via a networked file system
has been avoided by moving the files to a local hard disk volume. The bootstrapping
time only slightly decreased, loading the files is therefore not a critical factor. Then,
the code that samples negative examples from image files was checked not to perform
too many unnecessary disk accesses. A new file is loaded after about 20,000 patches
have been samples which should not warrant a high overhead. Hence, the reason for
the high bootstrapping time is the need to sample many random image patches before
a suitable negative sample can be added to the bootstrapped training set. Many image
patches have to be extracted and checked against the currently trained cascade stages,
only misclassified samples are added. One explanation for the disproportional rise would
be the size of the negative background images set. These 339 full-size images with diverse,
cluttered background may not contain enough suitable samples. So finding misclassified
patches becomes very time-consuming with random sampling. Therefore, a larger set may
solve the issue.
Figure 10 shows that the training time does not increase linearly but fluctuates. Therefore,
the number of selected features was also plotted to verify that the classifier is constantly
growing in complexity. The fluctuations in training time are due to the MSL-style learning,
where positive samples are bootstrapped. Feature reuse was not implemented, therefore
after every bootstrapping run all features have to be reselected. In some cases, a single
positive bootstrap run was enough to fulfill the stage goal on the whole set whereas the
next stage required, e.g., 4 bootstrap runs, thus taking about four times as long. Figure
11 shows training time, bootstrapping time and number of features as well for a cascade
trained with 1500 (instead of 500 before) of 3000 positive samples already in the initial
58 5 EXPERIMENTS
120
100 3000 pos+neg, 25 stg
# Features
80 3000 pos+neg 2, 25 stg

60
40
20
0
0 5 10 15 20 25
Stage
7000
6000 3000 pos+neg, 25 stg
Training time
5000 3000 pos+neg 2, 25 stg

4000
3000
2000
1000
0
0 5 10 15 20 25
Stage
Bootstrapping time
140000
120000 3000 pos+neg, 25 stg
100000 3000 pos+neg 2, 25 stg
80000
60000
40000
20000
0
0 5 10 15 20 25
Stage
Figure 10: Training time, bootstrapping time and number of features for 25 stage cascade
training. The bootstrapping time increases exponentially after 20 stages were trained.
MSL-style learning causes the fluctuations in training time.
bootstrap set, i.e. reduced MSL bootstrapping. The training time curve is smoother and
the exponential increase is still visible towards the end, but it is more pronounced in
Figure 10.
5.3.4 Motivation for Training Set Selection
An optimal training set can have a great influence on the classifiers performance. In order
to examine the effect of different training sets, the 9000 sample base set was randomly
divided into three 3000 sample sets as described in Section 5.3.1. Figure 12 demonstrates
the resulting classifiers performances trained with these different sets. The classifiers were
all trained with the same training parameters. Figure 12 clearly shows that while cascade
1 and 2 roughly have the same performance, cascade 3 underperforms in comparison.
Training set selection is supposed to make sure that cascades trained with an optimized
set will perform better.
70
60 3000 pos+neg 1, 20 stg
# Features
50 3000 pos+neg 2, 20 stg
40
30
20
10
0
0 2 4 6 8 10 12 14 16 18 20
Stage
6000
5000 3000 pos+neg 1, 20 stg
Training time
4000 3000 pos+neg 2, 20 stg

3000
2000
1000
0
0 2 4 6 8 10 12 14 16 18 20
Stage
Bootstrapping time
2500
2000 3000 pos+neg 1, 20 stg
3000 pos+neg 2, 20 stg
1500
1000
500
0
0 2 4 6 8 10 12 14 16 18 20
Stage
Figure 11: Training time, bootstrapping time and number of features for a 20 stage
cascade with minimal MSL bootstrapping. The exponential increase in bootstrapping
time begins around stage 18. The variations in training time are caused by the MSL
bootstrapping of positive samples.
0.85
0.8
0.75
0.7
Hit Rate
0.65
0.6
0.55
0.5
0.4
0 10 20 30 40 50
False Acceptances
Figure 12: Comparison of three cascades trained with identical parameters but different
training sets of the same size. The performance of cascade 3 is far below that of cascade
1 and 2 and must be due to the choice of training set since all other parameters were the
same.
60 5 EXPERIMENTS
5.3.5 Motivation for Cascade Combination
Figure 13 demonstrates another advantage of cascade combination. Cascade combina-

tion can be used to improve the results of low-performance base cascades and therefore
generate cascade classifiers that are comparable to classifiers that were trained more ex-
tensively. The extensive classifiers were trained with more final stages and higher stage
target detection rate dmin = 0.9999, as opposed to 20 stages and dmin = 0.999 for the
base cascades. Figure 13 compares the performance of a 20-stage combined cascade clas-
sifier constructed from three 20-stage base cascades with an extensively trained 25-stage
cascade classifier. As was shown in Section 5.3.3, training time increases sharply after
twenty stages. Section 5.3.2, on the other hand, shows that substantial performance gains
are possibly with extended training.
Figure 13 shows the resulting classifiers performances. The plot displays the best 20-stage
base cascade, the best extensively trained 25-stage cascade and the resulting combined
cascade classifier. In order to avoid clutter, only the best classifiers for 20 and 25 stages
are shown, although three of each were trained - see Section 5.3.2 for details. The results
indicate that cascade combination was able to produce a cascade classifier that substan-
tially outperforms the base classifiers, as can be seen in Section 5.4.5. More importantly,
the combined cascades can almost match the performance of the 25-stage base cascade
that took considerably longer to train.
Overall, the training time for the depicted extensively trained, high-performance cascade
was 267 hours. The training time for the depicted low-performance base cascade was only
6 hours and 58 minutes. This striking difference in training times is due to an effect shown
in Section 5.3.3, the exponential increase in bootstrapping time after 20 stages. Overall,
the three 20-stage base cascades were trained in about 22 hours, the combination requires
another 30 minutes to an hour, depending on the amount of extracted samples. Therefore,
the total training time for the combined cascades was about one day as opposed to 11
days for the extensively trained cascade. That is a substantial improvement and should
outweigh the fact that the 25-stage cascade performs slightly better, two to five percent
increased detection rate for a given false alarm rate.
0.85
0.8
0.75
0.7
Hit Rate
0.65
0.6
0.55
Fused cascade CMIM - 20 stg base to 20 stg
Fused cascade MIM - 20 stg base to 20 stg
0.45
0 10 20 30 40 50
False Acceptances
Figure 13: Cascade combination of three 20-stage, low-performance cascades trained with
3000 positive and negative samples vs. 25-stage, high-performance cascades trained on
the same training sets. The resulting combined classifers performance approaches the
performance of the high-performance cascades. Training time for all three 20-stages base
cascades and the combination process is less than a day, whereas the 25-stage cascades
were trained for more than 11 days each. Cascade combination improves the base cascade
classifiers significantly while keeping training time low. Only the best base and 25-stage
cascade classifiers are shown for comparison.
5.3.6 Large Training Sets with Matrix-Structural Learning
Figure 14 compares three cascades with different training set sizes trained to twenty stages.
Three cascades were trained with the same training parameters but different training set
sizes. Two 20-stage cascades were trained with small training sets of 1,000 and 2,000
samples whereas a third 20-stage cascade was trained with a large 40,000 sample set. The
40,000 sample set classifier clearly outperforms the other classifiers as was to be expected.
The increase in performance is at the cost of training time. Training with 40,000 samples
requires a cache size of about 8GB of RAM and is therefore not always practical. If the
amount of available memory is insufficient to cache feature values for all samples, the
training time will increase strongly. Almost similar performance can be achieved with a
25-stage classifier trained with only 3,000 training samples.
62 5 EXPERIMENTS
0.9
0.8
0.7
0.6
Hit Rate
0.5
0.4
0.3
0.2
MSL 1000 samples
0.1 MSL 2000 samples
MSL 40000 samples
0
0 10 20 30 40 50
False Acceptances
Figure 14: Comparison of cascades trained with different amount of samples. Increasing
the training set size improves the performance at the cost of training time and memory
consumption.
5.4 Classifier Combination
The combination of cascade classifiers is analyzed in the following subsections. First,

we analyze whether potential performance gains are due to combination or solely due to
the associated threshold optimization. Then, the necessity and effects of subsampling
the amount of negative samples for cascade combination are examined. Finally, differ-
ent cascade combination cost functions are evaluated and additional experiments include
determining the effect of combining stages from more than two cascades.
5.4.1 Threshold Optimization
Threshold optimization of pre-trained cascade classifiers has been studied before, i.e. by
Luo [38]. The published works have shown substantial improvements over the initial base
cascade by adjustinging stage thresholds. The algorithms proposed in this work do not
specifically aim at improving the stage threshold, but in order to successfully recombine
stages, the thresholds need to be adjusted as well. Since the stages were trained to be
used at a certain position within a cascade, the threshold determined during training may
not be ideal anymore. Therefore, each stage is threshold ti is optimized based on the
5.4 Classifier Combination 63
same cost function that is used to combine the classifiers. Thresholds are adjusted using a
coarse-to-fine search within a certain range from the original value. The search interval is
[ti 0.5, ti +0.5], which has been found empirically. This subsection analyzes the influence
of the threshold optimization on the overall performance to ensure that the gains are not
solely due to choosing better thresholds.
Figure 15 shows the initial base cascade, the cascade after threshold optimization and the
CMIM combination of three base cascades stages with the same amount of final stages.
As can be seen from the plot, threshold optimization by itself provides an increase in
the detection rate for a given false alarm rate. Using a combination of stages of different
base classifiers can further improve results. Therefore, the performance gains of cascade
combination are not solely due to the integrated threshold optimization.
0.8
0.75
0.7
0.65
0.6
Hit Rate
0.55
0.5
0.45
0.4
Base cascade 1 - 20 stg
0.35 Fused classifier - 20 stg
Thresh optimization - 20 stg
0.3
0 10 20 30 40 50
False Acceptances
Figure 15: Cascade combination vs. threshold optimization of a single cascade. Threshold
optimization improves the classification results. Additional performance gains can be
achieved with cascade combination.
64 5 EXPERIMENTS
5.4.2 Subsampling for Cascade Combination
As noted before, due to the strong imbalance of negative and positive samples extracted
for the optimal cascade combination and the bias of cost functions regarding class sample
sizes, there is a need to reduce the imbalance for the combination to perform best. If the
optimization set contains too many negative samples, the cost function may prefer stages
that correctly reject more negative samples at the expense of also rejecting more positives
and degrading the overall performance. On the other hand, forcing the combination pro-
cess to adopt a slight bias may be helpful. Once the optimal parameter for subsampling
has been determined, that parameter does not need to be changed much for different cas-
cades trained with different parameters or count of stages. Figure 16 shows the influence
of subsampling on each of the combination cost functions. As it was assumed, correlation
is not highly influenced by subsampling because the class priors play no role. CMIM and
MIM are more dependent on subsampling, especially MIM. Figure 16 shows the varying
classification performance for different subsampling parameters. The MIM cost function
does not handle imbalanced data sets well, whereas the other two methods seem to be
less influenced. The necessity to determine optimal subsampling parameters complicates
the combination process and is therefore not very desirable, this is a major draw-back for
the MIM cost-function. For the other cost functions, subsampling turns out to be a vast
speed improvement since, after subsampling, much fewer samples have to be analyzed
when selecting each stage.
5.4.3 Comparison of Cascade Combination Cost Functions
Figure 17 shows a comparison of the three different cost functions. The MIM cost function
outperforms both correlation and CMIM cost functions. While the CMIM cost function
performs better than the correlation cost function. The graph was generated with optimal
subsampling parameters for each cost function, all other parameters were the same.
0.8 0.65
0.7 0.6
0.6 0.55
0.5 0.5
Hit Rate
Hit Rate
0.4 0.45
0.3 Fused MIM 10 subsamp, 800 samples, 25 stg 0.4 Fused Correlation 10 subsamp, 800 samples, 25 stg
Fused MIM 200 subsamp, 800 samples, 25 stg Fused Correlation 20 subsamp, 800 samples, 25 stg
0.2 Fused MIM 320 subsamp, 800 samples, 25 stg 0.35 Fused Correlation 200 subsamp, 800 samples, 25 stg
0.1 0.3
0 0.25
0 10 20 30 40 50 0 10 20 30 40 50
False Acceptances False Acceptances
0.8
0.7
0.6
0.5
Hit Rate
0.4
0.3 Fused CCMIM 10 subsamp, 800 samples, 25 stg

Fused CMIM 20 subsamp, 800 samples, 25 stg
0.2 Fused CMIM 200 subsamp, 800 samples, 25 stg
0.1
0
0 10 20 30 40 50
False Acceptances
Figure 16: Cascade combination, the effect of subsampling for MIM, CMIM and correla-
tion cost functions. While CMIM and correlation cost functions are not very sensitive to
imbalances of positive and negative sets, the MIM cost function is.
66 5 EXPERIMENTS
0.85
0.8
0.75
Hit Rate
0.7
0.65
0.6 Fusion CMIM, 3x3000 samp

Fusion Correlation 3x3000 samp
Fusion MIM, 3x3000 samp
0.55
0 10 20 30 40 50
False Acceptances
Figure 17: Comparison of the MIM, CMIM and correlation cost functions for combined
3000 sample trained cascades. The MIM cost function outperforms the other cost func-
tions.
5.4.4 Combination Stage Count
The use of cascade combination would allow to construct a final cascade with more stages
than the base cascades were made up of. Theoretically, adding more stages to the cascade
should further reduce the false alarm rate - at the cost of slightly decreasing the correct
detection rate. Figure 18 demonstrates the result of the associated experiment. Reducing
the final stage count roughly keeps the detection rate at higher false alarm rates but
reduces the detection rate at low false alarm rates. Adding more stages seems to yield a
mild increase of the detection rate at low false alarm rates, whereas the detection rate at
higher false alarm rates suffers.
The false alarm rate of a cascade classifier is determined by its stages. It is possible to
reduce the false alarm rate by tightening the thresholds so that fewer negative samples are
accepted as positive. This also decreases the correct detection rate. Another way to reduce
the false alarm rate is to add additional stages. In ordinary cascades, later stages are
trained with more difficult training examples, therefore they tend to be more complex. In
cascade combination, none of these complex stages trained with more challenging training
examples are available. Thus, when adding new stages, the only possibility to decrease
false alarm rates is to utilize stricter thresholds for later stages. Figure 18 shows that
adding more stages is not beneficial in this case - the same effect could be achieved with
stricter thresholds on earlier stages.
0.9
0.8
0.7
0.6
Hit Rate
0.5
0.4
0.3
0.2 Fused MIM 3000 smp, 25 stg to 20 stg

Fused MIM 3000 smp, 25 stg to 25 stg
Fused MIM 3000 smp, 25 stg to 30 stg
0.1
0 10 20 30 40 50
False Acceptances
Figure 18: Combining 25 stage base cascades into 20, 25 and 30 stage final cascades.
Increasing the number of final stages increases the hit rate at low false alarm rates but
also decreases the overall hit rate slightly.
5.4.5 Comparison against Base Classifiers
Figures 19 and 20 show a comparison of combined classifiers versus the base classifiers
from which the combination was constructed.
Figure 19 shows a clear improvement of both combined cascades over the initial base
cascades. The detection rate, especially at low false alarm rates, is improved for a given
false alarm rate. The CMIM and MIM cost function perform almost similarly. The base
cascades were 20-stage classifiers as described in Section 5.3.2.
Figure 20 shows the same comparison for more extensively trained cascades. The base
cascades whose results are shown in Figure19 were trained in about seven to eight hours
each, the 25-stage cascades in Figure 20 were trained for about ten days each. The
extensive training seems to have improved the cascades significantly. Compared to Figure
19, the detection rate at 50 false acceptances rises slightly. But especially in lower false
alarm regions, the detection rate has been improved substantially. The curve is near the
maximum detection rate even for lower false alarm rates and drops slower than before.
68 5 EXPERIMENTS
Therefore, the combination approach does not seem able to produce any substantial gains
over the 25-stage base cascades anymore. Although there is a slight improvement, the
performance gains are much lower than those seen in Figure 19.
0.8
0.7
0.6
0.5
Hit Rate
0.4
0.3
Base cascade 1 - 3000 samples, 20 stg
0.2 Base cascade 3 - 3000 samples, 20 stg
Fused cascade CMIM
Fused cascade MIM
0.1
0 10 20 30 40 50
False Acceptances
Figure 19: Comparison of 20-stage base cascades and combined cascades with 20 stages.
A substantial improvement is possible here with both MIM and CMIM cost functions.
0.9
0.8
0.7
0.6
Hit Rate
0.5
0.4
0.3 3000 smp base cascade 1

3000 smp base cascade 2
0.2 3000 smp base cascade 3
Fused CMIM 3000 - 25 stg
Fused MIM 3000 - 25 stg
0.1
0 10 20 30 40 50
False Acceptances
Figure 20: Comparing cascade combination against well-trained 25-stage base classifiers.
Only marginal improvements can be achieved over well-trained base cascades.
5.5 SVM-based Training Set Selection 69
5.5 SVM-based Training Set Selection
The following experiments examine the effect of resampling the positive training set with
use of support vector machines. First, we compare the training of an SVM-based face
detector trained using histogram equalization-based features and raw image intensity-
based features. The next experiment performs an analysis of the possible performance
gains by SVM-based sample set selection.
5.5.1 Histogram Equalization vs. Raw Image Intensity Features
In order to resample the positive training set with the help of support vector machines,
an SVM has to be trained first. The SVM is trained by bootstrapping the negative
training samples. In face detection, the SVM has to distinguish face and non-face patterns.
Since face detection is a complex task with a lot of variance, partially due to different
lighting situations, normalization techniques have been shown to improve classification
when applied before-hand. Histogram equalization worked well and Figure 21 shows that
classifiers trained with histogram equalized data performed better. False alarm rates were
lower and detection rates higher when compared to raw intensity features after the same
number of bootstrapping iterations.
5.5.2 SVM-based Training Set Resampling
Figure 22 illustrates the performance of classifiers trained with SVM-based training set
resampling compared to the best similar base cascade. The figure compares the base cas-
cades trained with 3000 positive and negative samples against two cascades with samples
selected based on SVM distance to hyperplane. An SVM-based face detector was used to
create a ranking of 9,000 initial training samples. These 9,000 samples are the same that
were partitioned and used to train the base classifiers. The SVM-based feature ranking
was then used to select 3000 positive samples with which cascade classifiers were trained.
The label hardest describes the sample set comprised of 3000 samples closest to the
SVM hyperplane, equidist describes samples selected from the whole range of samples.
Both resampling techniques achieve similar performance. Covering the whole range of
samples seems slightly advantageous, as was to be expected, because the whole face space
is covered. But the difference is minimal and might as well be due to the bootstrapping
of negative samples. In experiments, two iterations of bootstrapping were found to be
70 5 EXPERIMENTS
20000
18000
16000
14000
False Alarms
12000
10000
8000
6000
4000 SVM bootstrapping with hist norm
2000 SVM bootstrapping raw image intensities
0
1 1.5 2 2.5 3 3.5 4 4.5 5
bootstrapping run
0.6
0.55
0.5
0.45
Hit rate
0.4
0.35
0.3
SVM bootstrapping with hist norm
0.25
SVM bootstrapping raw image intensities
0.2
1 1.5 2 2.5 3 3.5 4 4.5 5
bootstrapping run
Figure 21: SVM bootstrap training - 3000 positive, 5000 negative samples, with and
without histogram normalization of input images.
5.6 Combining SVM-based Resampling and Cascade Combination 71
enough to produce a useful ranking. More bootstrapping iterations had little to no impact
on the classifiers performance.
0.8
0.7
0.6
0.5
Hit Rate
0.4
0.3
0.2 Base cascade 3 - 3000 samples, 20 stg
Cascade SVM sample selection equal dist - 2 bs runs 20 stg
Cascade SVM sample selection hardest samples - 2 bs runs 20 stg
0.1
0 10 20 30 40 50
False Acceptances
Figure 22: Cascades trained with SVM-based bootstrapping, 3000 training samples and
20 stages vs. 20-stage base cascades. Cascades trained with SVM-based training sample
selection outperform the base cascades.
5.6 Combining SVM-based Resampling and Cascade Combina-

tion
In this subsection, the application of both training set optimization and cascade combi-
nation has been studied. Figure 23 compares the base cascade trained with 9000 training
samples with a combination of two cascades whose 3000 training samples were acquired
with both different SVM-based resampling methods. Before applying the cascade combi-
nation algorithm with the MIM cost function, two cascade classifiers were trained with
3000 training samples each. These training samples were selected using the SVM-based
training sample selection as outlined in Section 4.3.1. The figure shows that it is possible
to achieve almost similar and in part better performance with combination of two care-
fully trained 3000 sample cascades as with a single 9000 sample cascade. The combined
72 5 EXPERIMENTS
cascade resulted in 20 stages, the two base cascades were trained to 20 stages and the
9000 sample cascade also contained 20 stages. The 3000 training samples for the cascades
were selected from the 9000 sample set used to train the larger cascade.
0.8
0.7
0.6
0.5
Hit Rate
0.4
0.3
0.2
9000 smp, 20 stg base cascade
Fused MIM, SVM bs 1+2 3000 smp, 500 subsamp
0.1
0 10 20 30 40 50
False Acceptances
Figure 23: Comparison of a base cascade trained with 9000 positive and negative samples
with a combined cascade. The combined cascades input classifiers were trained with
SVM-based sample selection, 3000 samples for training were selected from 9000 initial
samples. The two input cascades samples were selected with the help of an SVM and two
different sample selection strategies. The MIM cost function was used. Both the 9000
sample base cascade and the combined cascade have 20 stages.
5.7 Comparison against other Cascade Classifiers
This section compares the presented methods to other publicly available classifiers. Figure
24 shows a comparison of several classifiers that were created for this work against publicly
available OpenCV cascades. The figure demonstrates that our cascades clearly outperform
the OpenCV face detectors. Some classifiers are not visible in the graph, because, in order
to preserve details in that most important part of the ROC, the window was focused on
false alarm rates in [0, 50]. The invisible classifiers were unable to produce any hits at false
alarm rates below 50. The comparison was done under the same premises, all cascades
were evaluated with the ROC generation approach outlined in Section 5.2 and the same
parameters.
5.7 Comparison against other Cascade Classifiers 73
0.9
0.8
0.7
0.6
Hit Rate
0.5
0.4
Base Cascade 2 - 3000 samples, 20 stg
Base Cascade 2 - 3000 samples, 25 stg
0.3
Fused Cascade - 3000 samples, 25 stg, MIM
Open CV - haarcascadefrontalfacealt
0.2
Open CV - haarcascadefrontalfacealt2
Open CV - haarcascadefrontalfacealttree-log-res.txt
0.1 Open CV - haarcascadefrontalfacedefault
SVM resampled cascade - 300 samples, 20 stg, equal dist
0
0 10 20 30 40 50
False Acceptances
Figure 24: Comparison against OpenCV default cascade
Figure 25 compares our cascades against a state-of-the-art classifier by Luo [38]. Luos
classifiers [38] are constructed by automatic threshold optimization and outperform the
presented classifiers. It should be noted that the base classifier used for the threshold
optimization already performed much better than our base classifiers, even our final clas-
sifiers. Luos base cascade [38] before the threshold optimization, which is not shown in
the figure, has a detection rate that is about 10% higher on average compared to our
base classifiers. Our results indicate that cascade combination can improve classification
results but they cannot achieve state-of-the-art performance. For better comparison, base
cascades would have to be trained that match the performance of the initial cascades in
other publications.
74 5 EXPERIMENTS
0.9
0.8
0.7
Hit Rate
0.6
0.5
0.4
0.3
Fused cascade CMIM
Fused cascade MIM
0.2 Luo dep thresholding
Luo indep thresholding
0.1
0 10 20 30 40 50
False Acceptances
Figure 25: Comparison against state-of-the-art classifiers [38].

75
6 Conclusion
This work presents two methods to improve the training of face detection cascades. The
first method, cascade combination, attempts to combine the advantages of classifier com-
bination without the additional overhead of running several classifiers in parallel. Second,
SVM-based training set selection seeks to optimize the training set in a conceptually
simple fashion in order to force the final classifier to focus on the decision boundary.
Both methods have been shown to improve the classification performance over base clas-
sifiers. All experiments were conducted on the widely used, public CMU+MIT benchmark
database in order to allow comparison with other work.
In this study, several base classifiers were trained so as to create a baseline for evaluating
the performance of the proposed cascade optimization techniques. An initial training
set was divided into several smaller sets for various experiments. Several classifiers were
trained based on these sub-sets and used as base classifiers for cascade combination and
as a comparison baseline for the training set resampling.
A framework to represent cascade combination as an optimization and three methods

for cascade combination have been presented. The first method is based on conditional
mutual information as a cost function and is akin to a successful feature selection approach
that has been studied. Another method is based on simple mutual information between
each stages classification combined with previously selected stage classifications and the
true sample class labels. These methods seek to directly optimize the correlation between
the set of selected stages plus a prospective next stage and the class labels. Another
method based on linear correlation has also been explored for comparison.
The cascade combination approach was shown to improve the performance of the base
classifiers by utilizing and combining several stages of different cascade classifiers. It was
also shown that cascade combination can be used to reduce training times by combining
several stages of quickly trained classifiers into a combined classifier with superior perfor-
mance that can achieve comparable performance as an extensively trained classifier. The
required overall time could be reduced from more than ten days to a single day, with only
a slight performance drop. Cascade combination is especially useful for quickly trained
cascades where large performance benefits can be achieved. Extensively trained cascades
on the other hand do not have as much potential for improvement.
The second method explored in this work is SVM-based training set selection. By resam-
76 6 CONCLUSION
pling the training set, emphasis is put on samples near the decision hyperplane, where
most of the errors occur. An SVM is trained and used to generate a ranking of samples
by difficulty, by using their distance from the hyperplane of the SVM. Two resampling
techniques have been studied, one selects samples closest to the decision boundary, the
other selects samples from the whole set. Both resampling techniques perform almost
equally well with a slight advantage for the set sampled from the whole initial training
set. Cascade classifiers trained with a resampled training set outperformed comparable
base cascades trained with random training sets.
Both the SVM-based training set resampling and cascade combination methods proved
useful for increasing the detection rate of cascade classifiers. Additionally, a combination
of both methods can be used to create a cascade classifier that is trained with 3000 samples
that can achieve the performance of a classifier that was trained with 9000 samples. This
property proves to be useful for cascade classifiers trained with large training sets and
many stages as the bootstrapping time increases exponentially and large amounts of
memory would be necessary in order to cache feature value.
77
7 Future Work
Several improvements to the proposed methods are possible and could be addressed in
future work.
Cost functions: Other cost functions may have a positive impact on the performance
of combined classifiers and may be less prone to the imbalances of the validation set used
for combination. A cost function that can handle unbalanced sets eliminates another
training parameter that has to be optimized in a time-consuming fashion.
Improved base cascades: In order to improve the overall performance of the combined
cascades better base cascade classifiers are necessary. Although a substantial amount of
time was spent on training and evaluating base cascades, the presented cascades are unable
to match the base cascades used in state-of-the-art methods, like those in [38]. Generally,
cascade training is not trivial and requires a lot of know-how and engineering effort.
Heuristics and a feeling for the importance of the many parameters must be acquired
which is a lengthy process. A survey of publicly available cascades by Castrilln-Santana
et al. [9] suggests that publicly available Haar cascade classifiers based on OpenCV
tend to be inferior to state-of-the-art implementations. The biggest problem so far is
the exponential increase in bootstrapping time after about 20 stages. Most publications
report cascades trained with 30-40 stages, but in our experiments, even after 10 days, only
28-stage cascade classifiers had been trained. More stages should decrease the false alarm
rate. But the decrease in false alarms may come at a reduced detection rate. Therefore,
it has to be seen whether the benefit of adding stages is not negated by the decrease of
the detection rate. Another possibility to increase the performance of base cascades is the
use of a larger training set, although care has to be taken to have sufficient free memory
for caching. Otherwise training time will increase dramatically.
Threshold optimization: Besides improved subsampling, an improved threshold opti-

mization algorithm could also help improve the detection rates. As demonstrated before,
part of the improvement is due to better choices of thresholds than in the original cascades.
Several threshold optimization algorithms have been proposed in other publications. An
approach based on the work by Luo [38] would be a prime candidate for implementation
in future work. Instead of using an iterative algorithm as suggested by Luo [38] that
78 7 FUTURE WORK
shifts ROC mass from weaker stages to more powerful stages, thresholds could be directly
chosen to be optimal using cached feature values and classification tables.
Bootstrapping of cascade combination stage negative sets: Since in cascade

combination, as during cascade training, stage thresholds have to be adapted for a specific
point in a cascade, bootstrapping may be helpful to adjust the stages better for each
selection round. Early stages are mostly meant to reject trivial negatives; later stages
should reject more difficult examples. It may be helpful if the validation set is chosen
accordingly.
Reuse of stage feature value sums: The idea of a boosting chain or soft cascade
[3, 70] and the implementation of multiple instance pruning [77] should help to increase
the classification performance. The experiments and the cascade combination framework
were initially designed to be used in conjunction with some form of continuous feature
sums that are not reset after stage boundaries as introduced by Bourdev and Brant [3] or
Xiao et al. [70]. Several publications that use continuous feature sums show substantial
improvements in classification performance.
Reducing training time: As stated in Section 5.3.3, in the training of late stages
the training time itself for feature selection and combination of weak classifiers becomes
negligible compared to the bootstrapping time. Presumably, the high bootstrapping times
could be due to the fact that the background images do not contain enough difficult image
patches. Although the background images seem to contain a large variety of motives with
clutter, adding more background images or more complex images may be necessary.
REFERENCES 79
References
[1] T. Agui, Y. Kokubo, H. Nagashashi, and T. Nagao. Extraction of face recognition
from monochromatic photographs using neural networks. Proc. Second Intl Conf.
Automation, Robotics, and Computer Vision, 18:8185, 1992.
[2] M.F. Augusteijn and T.L. Skujca. Identification of human faces through texture-
based feature recognition and neural network technology. Proc. IEEE Conf. Neural
Networks, pages 392398, 1993.
[3] L. Bourdev and J. Brandt. Robust object detection via soft cascade. Proc. IEEE
Conf. Computer Vision and Pattern Recognition, 2:236243, 2005.
[4] L. Brieman, J. Friedman, R. Olshen, and C. Stone. Classification and Regression

Trees. Wadsworth, 1984.
[5] S.C. Brubaker, M.D. Mullin, and J.M. Rehg. Towards optimal training of cascaded
detectors. European Conference on Computer Vision, 1:325337, 2006.
[6] S.C. Brubaker, J. Wu, J. Sun, M.D. Mullin, , and J.M. Rehg. On the design of
cascades of boosted ensembles for face detection. Intl Journal of Computer Vision,
(77):6586, 2008.
[7] M. C. Burl, T. K. Leung, and P. Perona. Face localization via shape statistics. Proc.
First Intl Workshop Automatic Face and Gesture Recognition, pages 154159, 1995.
[8] J. Cai, A. Goshtasby, and C. Yu. Detecting human faces in color images. Proc. 1998
Intl Workshop Multi-Media Database Management Systems, pages 124131, 1998.
[9] M. Castrilln-Santana, O. Dniz-Surez, L. Antn-Canals, and J. Lorenzo-Navarro.

Face and facial feature detection evaluation performance evaluation of public domain
haar detectors for face and facial feature detection. VISAPP, pages 167172, 2008.
[10] D. Chai and K.N. Ngan. Locating facial region of a head-and-shoulders color image.
Proc. Third Intl Conf. of Automatic Face and Gesture Recognition, pages 124129,
1998.
[11] J. Chen, X. Chen, J. Yang, S. Shan, R. Wang, and W. Gao. Optimization of a

training set for more robust face detection. Pattern Recognition Journal.
80 REFERENCES
[12] J. Chen, R. Wang, S. Yan, S. Shan, X. Chen, and W. Gao. Enhancing human
face detection by resampling examples through manifolds. IEEE Transactions on
Systems, Man, and Cybernetics, 6:10171028, 2007.
[13] X. Chen and A.L. Yuille. A time-efficient cascade for real-time object detection:
With applications for the visually impaired. IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, pages 2026, 2005.
[14] T.F. Cootes and C.J. Taylor. Locating faces using statistical feature detectors. Proc.
Second Intl Conf. Automatic Face and Gesture Recognition, pages 204209, 1996.
[15] Y. Dai and Y. Nakano. Face-texture model based on sgld and its application in face
detection in a color scene. Pattern Recognition, 29(6):238242, 1996.
[16] M. Dundar and M.J. Bi. Joint optimization of cascaded classifiers for computer aided
detection. IEEE Computer Society Conference on Computer Vision and Pattern
Recognition, pages 18, 2007.
[17] F. Fleuret. Fast binary feature selection with conditional mutual information. The
Journal of Machine Learning Research, 5:15311555, 2004.
[18] D. Forsyth. A novel approach to color constancy. Intl Journal of Computer Vision,
5(1):536, 1990.
[19] Y. Freund and R. E. Schapire. Experiments with a new boosting algorithm. In

International Conference on Machine Learning, pages 148156, 1996.
[20] J. Friedman, T. Hastie, , and R. Tibshirani. Additive logistic regression: a statistical

view of boosting. Annals of Statistics, 28:337407, 2000.
[21] B. Frba and A. Ernst. Face detection with the modified census transform. IEEE
International Conference on Automatic Face and Gesture Recognition, pages 9196,
2004.
[22] D. Grosvenor. Integrating object detectors. Technical Report HPL-2007-66, Media

Technologies Laboratory HP Laboratories Bristol, 2007.
[23] C.-C. Han, H.-Y.M. Liao, K.-C. Yu, and L.-H. Chen. Face face detection via
morphology-based pre-processing. Proc. Ninth Intl Conf. Image Analysis and Pro-
cessing, pages 469476, 1998.
REFERENCES 81
[24] B. Heisele, T. Serre, M. Pontil, and T. Poggio. Component-based face detection.

Proc. IEEE Conf. Computer Vision and Pattern Recognition, 1:pp. 657, 2001.
[25] C. Huang, H. Ai, Y. Li, and Sh. Lao. Vector boosting for rotation invariant multi-
view face detection. In Proc. IEEE International Conference on Computer Vision,
pages 446453, 2005.
[26] C. Huang, H. Z. Ai, Y. Li, and S. H. Lao. High-performance rotation invariant

multiview face detection. IEEE Trans. Pattern Anal. Mach. Intell., 29(4):671686,
2007.
[27] Y. Ishii, K. Imagawa, E. Fukumiya, K. Iwasa, and Y. Ogura. Profile face detection
using block difference feature for automatic image annotation. Proc. IEEE Interna-
tional Conference on Consumer Electronics, pages 337 338, 2006.
[28] M. Jones and P. Viola. Fast multi-view face detection. MERL TR2003 96, July 2003.
[29] T. Kanade. Picture Processing by Computer Complex and Recognition of Human

Faces. PhD thesis, Kyoto University, 1973.
[30] M. Kirby and L. Sirovich. Application of the karhunen-loeve procedure for the char-
acterization of human faces. IEEE Trans. Pattern Anal. Mach. Intell., 12(1):103108,
1990.
[31] C. Kotropoulos and I. Pitas. Rule-based face detection in frontal views. Proc. of
IEEE Int. Conf. on Acoustics, Speech and Signal Processing, 4:25372540, 1997.
[32] Y. H. Kwon and N. da Vitoria Lobo. Face detection using templates. Proc. Intl
Conf. Pattern Recognition, pages 764767, 1994.
[33] T.K. Leung, M.C. Burl, and P. Perona. Finding faces in cluttered scenes using
random labeled graph matching. Proc. Fifth IEEE Intl Conf. Computer Vision,
pages 637644, 1995.
[34] S. Z. Li, L. Zhu, and Z. Q. Zhang. Statistical learning of multi-view face detection.
Proc. European Conference on Computer Vision, pages 6781, 2002.
[35] R. Lienhart, A. Kuranov, and V. Pisarevsky. Empirical analysis of detection cas-

cades of boosted classifiers for rapid object detection. In DAGM03, pages 297304,
Madgeburg, Germany, 2003.
82 REFERENCES
[36] R. Lienhart and J. Maydt. An extended set of haar-like features for rapid object
detection. Proc. IEEE International Conference on Image Processing, 1:900903,
2002.
[37] C. Liu and H.Y. Shum. Kullback-leibler boosting. Proc. IEEE Conf. Computer
Vision and Pattern Recognition, (587-594), 2003.
[38] H. Luo. Optimization design of cascaded classifiers. Proc. IEEE Conf. Computer
Vision and Pattern Recognition, pages 480485, 2005.
[39] B. McCane, K. Novins, and M. Albert. Optimizing cascade classifiers. Unpublished.
[40] S. McKenna, Y. Raja, and S. Gong. Tracking colour objects using adaptive mixture
models. Pattern Recognition, 31(12):18831892, 1998.
[41] A. Mohan, C. Papageorgiou, and T. Poggio. Example-based object detection in

images by components. IEEE Trans. Pattern Anal. Mach. Intell., 23(4):349361,
2001.
[42] K.R. Muller, S. Mika, G. Ratsch, K. Tsuda, and B. Schoelkopf. An introduction to

kernel-based learning algorithms. IEEE Trans. on Neural Networks, 12(2):181201,
2001.
[43] E. Osuna, R. Freund, and F. Girosi. Training support vector machines: An applica-
tion to face detection. Proc. IEEE Conf. Computer Vision and Pattern Recognition,
pages 130136, 1997.
[44] M.-T. Pham and T.-J. Cham. Online learning asymmetric boosted classifiers for ob-
ject detection. IEEE Computer Society Conference on Computer Vision and Pattern
Recognition (CVPR07), pages 18, 2007.
[45] H. Rowley, S. Baluja, and T. Kanade. Neural network-based face detection. Proc.
IEEE Conf. Computer Vision and Pattern Recognition, pages 203208, 1996.
[46] H. A. Rowley, S. Baluja, and T. Kanade. Rotation invariant neural network-based

face detection. Proc. IEEE Conf. Computer Vision and Pattern Recognition, pages
3844, 1998.
[47] A. Samal and P.A. Iyengar. Human face detection using silhouettes. Intl J. Pattern
Recognition and Artificial Intelligence, 9(6):845867, 1995.
REFERENCES 83
[48] F.S. Samaria. Face Recognition Using Hidden Markov Models. PhD thesis, Univ. of
Cambridge, 1994.
[49] D. Saxe and R. Foulds. Toward robust skin identification in video images. Proc.
Second Intl Conf. Automatic Face and Gesture Recognition, pages 379384, 1996.
[50] R. Schapire. The boosting approach to machine learning: An overview. MSRI

Workshop on Nonlinear Estimation and Classification, 2002.
[51] R. E. Schapire, Y. Freund, P. Barlett, and W. S. Lee. Boosting the margin: A new
explanation for the effectiveness of voting methods. In International Conference on
Machine Learning, pages 322330, 1997.
[52] R.E. Schapire and Y.Singer. Improved boosting algorithms using confidence-rated
predictions. Machine Learning, 37:297336, 1999.
[53] H. Schneiderman and T. Kanade. Probabilistic modeling of local appearance and

spatial relationships for object recognition. Proc. IEEE Conf. Computer Vision and
Pattern Recognition, pages 4551, 1998.
[54] B. Schoelkopf. Support vector learning. DAGM99, 1999.
[55] P. Sinha. Object recognition via image invariants: A case study. Investigative Oph-
thalmology and Visual Science, 35(4):17351740, 1995.
[56] K. Sobottka and I. Pitas. Face localization and feature extraction based on shape and
color information. Proc. IEEE Intl Conf. Image Processing, pages 483486, 1996.
[57] J. Sochman and J. Matas. Waldboost learning for time constrained sequential
detection. Proc. IEEE Conf. Computer Vision and Pattern Recognition, pages 150
156, 2005.
[58] F. Soulie, E. Viennet, and B. Lamy. Multi-modular neural network architectures:

Pattern recognition applications in ptical character recognition and human face recog-
nition. Intl J. Pattern Recognition and Artificial Intelligence, 7(4):721755, 1993.
[59] J. Sun, J. M. Rehg, and A. Bobick. Automatic cascade training with perturbation
bias. IEEE Computer Society Conference on Computer Vision and Pattern Recogni-
tion, 2:276283, 2004.
84 REFERENCES
[60] K.-K. Sung. Learning and Example Selection for Object and Pattern Detection. PhD
thesis, EECS, MIT, February 1996.
[61] K.-K. Sung and T. Poggio. Example-based learning for view-based human face de-
tection. IEEE Trans. Pattern Anal. Mach. Intell., 20(1):3951, 1998.
[62] M. Turk and A. Pentland. Eigenfaces for recognition. Journal of Cognitive Neuro-
science, 3(1):7186, 1991.
[63] P. Viola and M. Jones. Rapid object detection using a boosted cascade of simple
features. Proc. IEEE Conf. Computer Vision and Pattern Recognition, pages 511
518, 2001.
[64] P. Viola and M. Jones. Fast and robust classification using asymmetric adaboost and
a detector cascade. Advances in Neural Information Processing Systems, 14:1311
1318, 2002.
[65] P. Viola and M. Jones. Fast and robust classification using asymmetric adaboost and
a detector cascade. Advances in Neural Information Processing Systems, 14:1311
1318, 2002.
[66] P. Wang and Q. Ji. Learn discriminant features for multi-view face and eye detection.
In Proc. IEEE International Conference on Computer Vision, pages 373379, 2005.
[67] P. Wang and Q. Ji. Multi-view face and eye detection using discriminant features.
Computer Vision and Image Understanding, pages 99111, 2007.
[68] J. Wu, S.C. Brubaker, M.D. Mullin, and J.M. Rehg. Fast asymmetric learning for
cascade face detection. In IEEE Transactions on Pattern Analysis and Machine
Intelligence, volume 30, pages 369382, 2007.
[69] J. Wu, J. Rehg, and M. Mullin. Learning a rare event detection cascade by direct
feature selection. Advances in Neural Information Processing Systems, 16:15231530,
2003.
[70] R. Xiao, L. Zhu, and H.-J. Zhang. Boosting chain learning for object detection. Proc.
Fifth IEEE Intl Conf. Computer Vision, 1:709715, 2003.
REFERENCES 85
[71] S. Yan, S. Shan, X. Chen, W. Gao, and J. Chen. Matrix-structural learning (msl) of
cascaded classifier from enormous training set. Proc. IEEE Conf. Computer Vision
and Pattern Recognition, pages 17, 2007.
[72] G. Yang and T.S. Huang. Human face detection in complex background. Pattern
Recognition, 27(1):5363, 1994.
[73] J. Yang and A. Waibel. A real-time face tracker. Proc. Third Workshop Applications
of Computer Vision, pages 142147, 1996.
[74] M.-H. Yang and N. Ahuja. Gaussian mixture model for human skin color and its
application in image and video databases. Proc. ACM Human Factors in Computing
Systems Conf. (CHI 98), pages 142147, 1998.
[75] M.-H. Yang, David J. Kriegman, and Narendra Ahuja. Detecting faces in images: A
survey. IEEE Trans. Pattern Anal. Mach. Intell., 24(1):3558, January 2002.
[76] M.-H. Yang, D. Roth, and N. Ahuja. A snow-based face detector. Advances in Neural
Information Processing Systems, pages 855861, 2000.
[77] C. Zhang and P. Viola. Multiple-instance pruning for learning efficient cascade de-
tectors. Advances in Neural Information Processing Systems, 2007.
[78] D. Zhang, S.Z. Li, and D. Gatica-Perez. Real-time face detection using boosting in
hierarchical feature spaces. Proc. Intl Conf. Pattern Recognition, 2:411414, 2004.
[79] Z. Q. Zhang, M. Li, S. Z. Li, and H.J. Zhang. Multi-view face detection with float-
boost. Proc. Sixth IEEE Workshop on Applications of Computer Vision, page pp.
184, 2002.

Diploma Thesis Toth

Uploaded by

Copyright:

Available Formats

Diploma Thesis Toth

Uploaded by

Document Information

Original Description:

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

Diploma Thesis Toth

Uploaded by

Copyright:

Available Formats

Universitt Karlsruhe (TH)

Combination of Classifier Cascades

Karlsruhe, 30. January 2009

Face detection is one of the most fundamental tasks in human-computer-interaction,

Die Gesichtsdetektion ist eine der grundlegenden Aufgaben der Mensch-Maschine-Interaktion,

2 Overview of Face Detection Approaches 5

4 Cascade Combination and SVM-based Sample Selection 37

CART Classification and regression tree

The main contribution of this work can be summed up as follows:

combination schemes. In the field of face detection, run-time efficiency is one of

2 Overview of Face Detection Approaches

1. Knowledge-based approaches try to apply human knowledge about the characteris-

2. Feature-based approaches aim to extract features invariant to changes in pose, view-

3. Template-based approaches match one or more standard patterns of faces or facial

4. Appearance-based methods classify parts of the image by applying models learned

2.1 Knowledge-based Face Detection

Early attempts at face detection mainly consisted of knowledge-based methods. Re-

2.2 Feature-based Face Detection

2.3 Template-based Face Detection

Deformable templates: To overcome some of the previously mentioned limitations,

2.4 Appearance-based Face Detection

As opposed to the previously presented template-based approaches, appearance-based

Machine learning algorithm-based approaches: Many appearance-based algorithms

Distribution-based approaches: Sung and Poggio claim that appearance-based ap-

Linear transformation-based methods: Subspace methods project input patterns

2.5 Cascades of Boosted Ensembles

2.5.1 Alternative Boosting Algorithms

2.5.2 Alternative Feature Sets

scales and translations.

2.5.3 Weak Learners

Huang et al.[25] introduced piece-wise functions as weak learners. Piece-wise functions

2.5.4 Cascade Training Procedure

to be rejected by a single stage only (OR-operation). The approach iteratively optimizes

2.5.5 Cascade Architecture

2.5.6 Training Time Improvement

Although cascades of boosted ensembles exhibit real-time run-time performance, train-

2.5.7 Run-time Efficiency Improvement

2.5.8 Classifier Combination

3.1 Face Detection

3.1.1 Haar Features

Individual features can described as

3.1.2 Integral Image Representation

ii(x, y) = ii(x, y 1) + ii(x 1, y) + I(x, y) ii(x 1, y 1), (3)

with ii(1, y) = ii(x, 1) = ii(1, 1) = 0.

RecSum(r) = ii(x 1, y 1) + ii(x + w 1, y + h 1)

RSAT (x, y) = RSAT (x 1, y 1) + RSAT (x + 1, y 1)

with RSAT (x, y) = 0, x, y : x < 0, y < 0.

RecSum(rrot ) = RSAT (x h + w, y + w + h 1) + RSAT (x, y 1)

Boosting is a meta-algorithm used to improve the classification performance of a base

where t is the error of the classifier ht , i.e. probability of ht misclassifying samples xi X

Algorithm 1 AdaBoost boosting algorithm[50]

Train base learner using distribution Dt .

Get base classifier ht : X R.

3.1.4 Weak Learners

Boosting algorithms do not rely on a pre-determined base learning algorithm. Therefore,

3.1.5 Cascades of Boosted Ensembles

in a cascade with N stages.

3.1.6 Optimal Thresholding for Cascaded Ensembles

A cascade architecture consists of a number of stages with individually defined operating

where t is the error of the classifier ht , i.e. probability of ht misclassifying samples xi X

{xS : wx + b = 0, (w, b)S R} (12)