Globally Optimal Consensus Set Maximization

Through Rotation Search

Jean-Charles Bazin
(1)
, Yongduek Seo
(2)
and Marc Pollefeys
(1)
(1): Institute for Visual Computing, ETHZ, Switzerland
(2): Department of Media Technology, Sogang University, South Korea
Abstract. A popular approach to detect outliers in a data set is to
nd the largest consensus set, that is to say maximizing the number
of inliers and estimating the underlying model. RANSAC is the most
widely used method for this aim but is non-deterministic and does not
guarantee to return the optimal solution. In this paper, we consider a
rotation model and we present a new approach that performs consensus
set maximization in a mathematically guaranteed globally optimal way.
We solve the problem by a branch-and-bound framework associated with
a rotation space search. Our mathematical formulation can be applied
for various computer vision tasks such as panoramic image stitching, 3D
registration with a rotating range sensor and line clustering and vanishing
point estimation. Experimental results with synthetic and real data sets
have successfully conrmed the validity of our approach.
1 Introduction
Outlier removal is a key issue in computer vision [1]. The main idea to detect the
outliers is to nd the largest consensus set, that is to say maximizing the number
of inliers and estimating the underlying model [14]. In this paper, we consider
rotation model, which is important in several computer vision applications such
as panoramic image stitching [5], line clustering and vanishing point estimation
[6] and 3D registration with a rotating range sensor. RANSAC [2] is the most
widely used and well-known method for the task of outlier removal. It starts by
randomly selecting a minimal number of data to hypothesize a model and then
counts the number of inliers verifying the hypothesized model up to a residual
threshold [1]. After a certain number of iterations, it returns the model maxi-
mizing the number of inliers and the inlier set. Several variants of RANSAC have
been proposed such as MLESAC, LO-RANSAC, preemptive-RANSAC, among
many others [710]. While great results can be obtained, RANSAC and its vari-
ants are not guaranteed to fulll the exact goal of consensus set maximization:
they might not return the optimal solution in terms of number of inliers. Fur-
thermore, their randomized nature make them somewhat unpredictable: dierent
runs might return dierent results.
Some works tried to overcome these issues by using algorithms based on
mathematical optimization. For example, Li [3] proposed a reformulation method
based on convex and concave envelops [11] with a branch-and-bound technique.
2 Jean-Charles Bazin, Yongduek Seo and Marc Pollefeys
However this technique can deal with only linear algebraic cost and linear con-
straints, i.e. it cannot handle rotation model and angular distance. Olsson et
al. proposed a post-validation step to check whether a candidate solution is the
optimal one [12]. However this approach does not give an upper bound on the
number of trials needed (or the number of candidate solutions to compute and
test) to reach the optimal solution.
The remaining of this paper is organized as follows. In Section 2, we introduce
our mathematical formulation of consensus set maximization with a rotation
model. In Section 3, we present our optimization procedure based on branch-
and-bound. Finally, experimental results are presented in Section 4.
2 Mathematical formulation
We note u
i
and v
i
the i
th
corresponding measurements (also called a match)
of two sets related by a rotation R. Let also note (u, v) the angle formed by
the vectors u and v, and in the range [0, ]. In noise-free and outlier-free cases,
Ru
i
= v
i
i and (Ru
i
, v
i
) = 0i. When noise and outliers exist, these relations
are not veried anymore. Similarly to the popular residual tolerance method
[2], we dene a match (u
i
, v
i
) an inlier if the distance is lower than a residual
tolerance , i.e. (Ru
i
, v
i
) . Otherwise the match is considered an outlier.
Let M represent the set of matches: M = {(u
i
, v
i
), i = 1 . . . N} where N is
the number of matches. The set M is partitioned into an inlier-set M
I
M
containing the inlier matches and an outlier-set M
O
S with M
O
= M
M
I
. The cardinality of M
I
corresponds to the number of inlier matches. The
consensus set maximization with a rotation model is formulated as:
max
M
I
,R
card(M
I
) (1a)
s.t. (Ru
i
, v
i
) , i M
I
M (1b)
R SO(3) (1c)
To solve this consensus set maximization problem, RANSAC is denitively the
most popular method, but as explained in Section 1, it is non-deterministic and
might not reach the optimal solution. In a numerical optimization point of view,
it is not simple to process the above formulation. Following the approach of Li
[3], let us introduce a binary variable y
i
for each match such that y
i
= 1 if the
i
th
match is an inlier and y
i
= 0 otherwise (i.e. outlier). We can now reformulate
system (1) in the following equivalent way:
max
y,R
N

i
y
i
(2a)
s.t. y
i
(Ru
i
, v
i
) y
i
, i = 1 . . . N (2b)
y
i
{0, 1}, i = 1 . . . N (2c)
R SO(3) (2d)
When y
i
= 1, the inlier/outlier constraint at Eq (2b) corresponds to Eq (1b).
When y
i
= 0, the constraint is simply eliminated (i.e. always true since > 0).
Globally Optimal Consensus Set Maximization Through Rotation Search 3
System (2) still constitutes a challenging optimization problem. First, it con-
tains two families of unknowns: binary variables (y) and continuous geometric
model parameters (R). Second, it is non-linear because of the distance and the
unknowns (y and R) are multiplied to each other at Eq (2b). Third, it is non-
convex. Finally, fullling the rotation constraint at Eq (2d) requires specic care.
3 Rotation Search Algorithm
In this section, we explain our algorithm to solve system (2), and thus system (1),
in a globally optimal way. It is based on branch-and-bound framework and a ro-
tation space search. We start by a brief introduction to branch-and-bound, then
explain how to compute the bounds and nally present our search procedure.
3.1 Branch-and-bound
Branch-and-bound (noted B&B) is a general framework for global optimization
[13]. The main idea is the following: the denition space of the model to be
estimated is partitioned into smaller sub-spaces that are iteratively discarded or
rened. When it is certain, via a feasibility test, that a sub-space does not contain
the optimal solution, then this sub-space is discarded. On the contrary, if it might
contain the solution, then it is partitioned into smaller sub-spaces. Iteratively,
the size of the sub-spaces decreases and the estimated solution converges to the
optimal solution. The process stops when the desired accuracy level is reached.
B&B has gained popularity quite recently in computer vision [1420].
3.2 Bounds
The feasibility test is generally conducted by the computation of bounds, that is
the best (and optionally the worst) solution that can be obtained in a sub-space.
In the following, we explain how to compute the bounds for our application.
Computation By adopting the angle-axis representation, any rotations can be
represented by a point in a ball B

of radius . For example, a point in this ball

represents the rotation whose axis is / and angle [20]. Let decompose
B

into a set of cubes. The question becomes: how to compute the bounds of
the inlier cardinality that can be obtained for any rotations in a cube D. The
answer is: the lower and upper bounds are the solutions of the systems (3) and
(4) respectively, where

R is the (known) rotation corresponding to the center of
the cube D and is the half length side of D (cf proof below).
max
y
N

i
y
i
(3a)
s.t. y
i
(

Ru
i
, v
i
) y
i
(3b)
y
i
{0, 1}, i = 1 . . . N (3c)
max
y
N

i
y
i
(4a)
s.t. y
i
(

Ru
i
, v
i
) y
i
(
+

3
)
(4b)
y
i
{0, 1}, i = 1 . . . N (4c)
4 Jean-Charles Bazin, Yongduek Seo and Marc Pollefeys
The unknowns of the systems (3) and (4) are the binary y. It is interesting to
note that an Integer Programming solver, which is computationally expensive,
does not have to be run. Indeed, we can simply nd wether y
i
is 1 or 0 by
checking whether a matching pair veries the inlier distance constraint or not:
(

Ru
i
, v
i
) for the lower bound at system (3)
(

Ru
i
, v
i
) +

3 for the upper bound at system (4)

Proof Let us now prove that systems (3) and (4) provide valid lower and upper
bounds in the current cube D, that is to say that the optimal inlier cardinality
c that can be obtained in D is between these lower l and upper u bounds, i.e.
l c u. For writing simplication, let note R
opt
the rotation of D providing c.
Proving that the lower bound is valid is trivial: since c is the optimal cardi-
nality with respect to a residual threshold among all R D and since

R is a
particular element of D, then it is obvious that the cardinality l obtained by

R
with the same residual threshold is such that l c.
For the proof of the upper bound, it has been shown by Hartley and Kahl [20]
that max
R
((

Ru, Ru))

3 where R is any rotation in the cube D of half

length side and u any vectors. Let (u
i
, v
i
) be an inlier match by R
opt
, i.e.
(R
opt
u
i
, v
i
) . Therefore we can write:
(

Ru
i
, v
i
) (R
opt
u
i
, v
i
) +(

Ru
i
, R
opt
u
i
) (5)
+

3 (6)
It means that if a match is an inlier by R
opt
with respect to the residual threshold
, then it is also an inlier by

R with respect to the residual threshold +

3.
Therefore system (4) provides at least c inliers, i.e. u c.
Finally, B&B is guaranteed to converge to the globally optimal solution when
the bounds are valid [13]. Since we proved the bounds are valid, it proves, in
turn, that our proposed B&B-based algorithm is globally optimal.
3.3 Search procedure
Our B&B procedure is formalized in Alg1 using a breadth-rst search. The cube
list L is simply initialized with the rotation ball B

. At each B&B iteration,

each cube of L is subdivided into 8 sub-cubes of half size and the associated
bounds are computed. Then we remove from the list all the cubes whose upper
bound is lower than the maximum lower bound l

obtained so far because it

means that, even in the best case, it is sure that they cannot obtain an inlier
cardinality higher than l

. The procedure stops when at least one cube whose

lower bound equals the maximum upper bound is obtained. Finally the rotation
corresponding to the center of that cube is returned, i.e. this is a rotation that
globally maximizes the number of inliers.
4 Experiments
In Section 3.2, we proved that our algorithm is mathematically guaranteed to
obtain the globally optimal solution. In addition, we present, in the following,
Globally Optimal Consensus Set Maximization Through Rotation Search 5
Algorithm 1 B&B for inlier maximization through rotation search
Initialize the cube list L with the rotation ball B

repeat
Subdivision ( /2) of each cube D
i
of L
for each cube D
i
of L do
Get the rotation R
D
i
of the cube center
Compute the lower l
i
and upper u
i
bounds (cf Section 3.2)
end for
l

= max
i
l
i
, i

= arg max
i
u
i
, u

= u
i
, R

= R
D
i

Remove all the cubes from L such that u

i
< l

until i such that l

i
= u

(i.e. at least one cube whose lower-bound is u

)
Return: R

(i.e. the rotation maximizing the nb of inliers)

experimental results for both synthesized and real data to illustrate the feasibil-
ity and convergence of our approach. The algorithm has been implemented in
Matlab and run on a laptop equipped with an Intel i7-2820QM 2.3GHz CPU (a
single core was used) and 8GB RAM. We applied the algorithm on the whole
rotation space B

to show the worst case scenario (i.e. no a priori or initial

solution). We compared our approach to RANSAC and the quality of the results
is measured by the number of detected inliers. As RANSAC is not determinis-
tic, we repeated the experiments 1000 times with the same input to obtain the
inlier cardinality distribution and where the number of RANSAC iterations is
automatically computed with the true outlier ratio (if not available, e.g. for real
data, set to 60%), a guaranteed accuracy of 99% and the minimal sampling (2
points for rotation) [1]. Our algorithm runs in a fully automatic manner and
converges in a couple of seconds.
4.1 Synthesized data
We randomly generated two sets of N 3D points related by any random 3D
rotation in the whole SO(3). We applied a Gaussian noise to the coordinates of
the points and corrupted a certain percentage p of the points to create outliers.
Representative generated data is shown in Fig 1.
Figure 2 compares the number of inliers obtained by the proposed approach
and RANSAC. Our method obtains 30 inliers, which corresponds to the number
of synthesized inliers. On the contrary, RANSAC spans between 2 and 30 inliers,
and manages to detect the 30 inliers in only 1% of the experiments. This limited
performance of RANSAC might be surprising, especially given the high value of
the guaranteed accuracy. This can be explained by two reasons. First, the success
of RANSAC is related to the data noise: we performed experiments without any
noise and in this case, the RANSAC success rate corresponded to the guaranteed
accuracy. Second, RANSAC can hypothesize only models directly supported by
the randomly selected minimal points. Therefore if the optimal rotation cannot
be hypothesized by the minimal points then RANSAC cannot return the opti-
mal number of inliers. We performed extensive experiments with various data
amounts (N = 20 500 3D points) and proportion of outliers (p = 0% 80%).
6 Jean-Charles Bazin, Yongduek Seo and Marc Pollefeys
0
1
2
3
4
5
6
0.5
0
0.5
1
1
0.5
0
0.5
1
1.5
X
Z
Y
X
Y
Z
Fig. 1. Representative result of registration of two sets of 3D points related by a ro-
tation and containing outliers. The green (reciprocally red) lines correspond to the
inlier (reciprocally outlier) matches automatically found by our approach. For a better
visualization, only 30 matches are displayed and a set of points is translated.
In all the experiments, our algorithm always converged to (at least
1
) the number
of synthesized inliers and our estimated number of inliers was always higher or
equal than the one obtained by RANSAC.
In Figure 2-a, one might also note that the lower and upper bounds con-
verge which illustrates the convergence of the proposed algorithm and proves
the bounds are valid. For completeness, Figure 2-(c,d) illustrate the evolution of
the volume and the number of cubes. It shows the volume drops continuously
and the number of cubes remains limited. Note that this number of cubes does
not have to be strictly decreasing but must remain relatively low for memory
eciency, which is our case.
4.2 Real data: registration
We tested our algorithm on two parts of the Bunny 3D model from the Stan-
ford 3D Scanning Repository
2
in the context of rotational registration. These
two parts have a limited overlap and are related by a rotation. We applied the
method of Darom and Keller [21] on the two 3D sets to extract 3D feature points
and their associated descriptors, which provides candidate matches between the
two 3D sets. Figure 3-a represents the inlier and outlier matches automatically
detected by the proposed algorithm with an accuracy of 0.5

. Figure 3-(b,c)
illustrate the registration of the two sets by the estimated rotation (without any
additional renement, e.g. [22], in order to show the quality of our B&B solu-
tion). Figure 4 compares the number of inliers obtained by the proposed method
1
in rare cases, it might happen that the optimal number of inliers (returned by our
algorithm) is higher than the number of synthesized inliers because some outliers
might create a new dominant model, especially for high outlier ratio.
2
http://graphics.stanford.edu/data/3Dscanrep
Globally Optimal Consensus Set Maximization Through Rotation Search 7
1 2 3 4 5 6
0
10
20
30
Convergence of bounds
Iterations
U
p
p
e
r

a
n
d

l
o
w
e
r

b
o
u
n
d
s


upper
lower
(a)
5 10 15 20 25 30
0
5
10
15
20
Nb of detected inliers
p
e
r
c
e
n
t
a
g
e

o
f

e
x
p
e
r
i
m
e
n
t
s
distribution of the nb of inliers


(b)
1 2 3 4 5 6
10
4
10
2
10
0
10
2
Convergence of volume
Iterations
r
e
m
a
i
n
i
n
g

v
o
l
u
m
e

(
i
n
%
)
(c)
1 2 3 4 5 6
10
1
10
2
10
3
evolution of the nb of cubes
Iterations
N
u
m
b
e
r

o
f

c
u
b
e
s
(d)
Fig. 2. Experimental results on synthesized data. (a): convergence of the bounds of
the proposed B&B approach. (b): distribution of the number of inliers detected by
RANSAC on the same problem instance. (c): convergence of the volume of the search
space. (d): evolution of the number of cubes.
and RANSAC. Our method converges to 8 inliers, whereas RANSAC spans be-
tween 2 and 8 inliers. RANSAC performs poorly on this dataset because of the
large proportion of outliers (8 inliers out of 1191 points) and the data noise. In
contrast our approach can handle this challenging case without any diculties.
4.3 Real data: panorama stitching
We now present results in the context of panoramic image stitching [5]. When
the scene is far away, two images are related by the innite homography H

.
When the intrinsic calibration matrix K is known [1], H

reduces to a rotation
that we aim to estimate. To obtain correspondences between the two images, we
extracted and matched SIFT features [23] and applied K
1
to compute the asso-
ciated ray matches. Figure 5-a illustrates the inlier and outlier matches found by
the proposed approach. Once the inliers are found, any reprojection error mini-
mization method can be applied [5, 22]. The nal stitching result is shown in Fig-
ure 5-b. Figure 5-(c,d) compare the number of inliers obtained by the proposed
8 Jean-Charles Bazin, Yongduek Seo and Marc Pollefeys
(a)
(b) (c)
Fig. 3. Registration result between two 3D point sets of the Bunny model. (a): inlier
(green lines) and outlier (red lines) matches detected by the proposed approach. (b,c):
3D views after 3D registration of the two 3D point sets by the estimated rotation,
without any additional renement.
2 4 6 8
0
20
40
60
80
100
120
Convergence of bounds
Iterations
U
p
p
e
r

a
n
d

l
o
w
e
r

b
o
u
n
d
s


upper
lower
1 2 3 4 5 6 7 8
0
20
40
60
80
Nb of detected inliers
p
e
r
c
e
n
t
a
g
e

o
f

e
x
p
e
r
i
m
e
n
t
s
distribution of the nb of inliers


Fig. 4. Experimental results for the Bunny model. (a): convergence of the bounds of
the proposed B&B approach. (b): distribution of the number of inliers by RANSAC.
approach and RANSAC. Our approach nds 654 inliers while RANSAC spans
between 651 and 654 inliers and manages to detect 654 inliers in less than 1%
of the RANSAC runs.
Globally Optimal Consensus Set Maximization Through Rotation Search 9
(a)
(b)
2 4 6 8
0
200
400
600
800
Convergence of bounds
Iterations
U
p
p
e
r

a
n
d

l
o
w
e
r

b
o
u
n
d
s


upper
lower
(c)
651 652 653 654
0
20
40
60
80
100
Nb of detected inliers
p
e
r
c
e
n
t
a
g
e

o
f

e
x
p
e
r
i
m
e
n
t
s
distribution of the nb of inliers


(d)
Fig. 5. Result of panoramic image stitching. (a): inlier (green lines) and outlier (red
lines) matches detected by our approach between the two input images. (b): stitching
of the two images to build a panoramic view. (c): convergence of the bounds of our
B&B approach. (d): distribution of the number of inliers detected by RANSAC.
10 Jean-Charles Bazin, Yongduek Seo and Marc Pollefeys
4.4 Real data: line clustering
Given a set of input lines, line clustering aims to nd which line belongs to which
vanishing point (VP) and estimate the VPs [24]. Very recently, Bazin et al. [25]
proposed a method to maximize the number of clustered lines in this context. It
provides interesting results but is based on interval analysis [26], which provides
loose bounds and, similarly to the big-M method [11], is subject to numerical
inaccuracies. The formulation of system (2) can be modied so that our frame-
work can be applied for line clustering. In the following, we consider intrinsically
calibrated images and orthogonal vanishing points (Manhattan world [6, 27]).
Mathematical formulation: let n
i
be the normal vector of the i
th
input
line with i = 1 . . . N and v
j
be the j
th
VP with j = 1 . . . M where 1 M 3. A
line-VP pair composed of (n
i
, v
j
) is considered an inlier if

(n
i
, v
j
)

2

.
We say that n
i
is matched (clustered) to v
j
and we note (n
i
, v
j
) M
I
. The
consensus set maximization can be written:
max
M
I
,{v}
card(M
I
) (7a)
s.t.

(n
i
, v
j
)

2

, (i, j) M
I
M (7b)
and v
j
= 1, j and v
T
j
v
k
= 0, j, k with j = k (7c)
As observed by Bazin et al. [6], this system can be simplied by representing
the set of M orthogonal VPs by a single rotation R: v
j
= Re
j
where e
j
represents
the j
th
axis of the canonical basis, i.e. e
1
= (1, 0, 0), e
2
= (0, 1, 0) and e
3
=
(0, 0, 1). Additionally, we introduce the clustering variables y
i,j
set to 1 if the
i
th
line is clustered to the j
th
VP and to 0 otherwise. This permits to dene the
following system:
max
y,R
N

i
M

j
y
i,j
(8a)
s.t. y
i,j

(n
i
, Re
j
)

2

y
i,j
, i = 1 . . . N, j = 1 . . . M (8b)
y
i,j
{0, 1}, i = 1 . . . N, j = 1 . . . M (8c)
0
M

j
y
i,j
1, i = 1 . . . N (8d)
R SO(3) (8e)
The (optional) constraint at Eq (8d) enforces that a line can belong to at most
one VP. To solve system (8), we employ a strategy similar to Section 3.2 where
the rotation cube center is used. Concretely, the variables y
i,j
are set to 0 or 1
by simply testing the constraints:

(n
i
,

Re
j
)

2

for the lower bound

(n
i
,

Re
j
)

2

3 for the upper bound

The (optional) at-most-one constraint at Eq (8d) can be enforced afterwards
by simply setting only one y
i,j
to 1 given i, for example for the nearest VP:
Globally Optimal Consensus Set Maximization Through Rotation Search 11
Fig. 6. Representative line clustering results obtained by our algorithm on the York
urban database [28] for indoor and outdoor images, with 2 and 3 VPs. Top: input line
segments. Bottom: output line clustering obtained by the proposed algorithm.
Fig. 7. Distribution of the execution time of our algorithm for line clustering on all
the images of the York urban database [28].
j

= arg min
j

(n
i
, Re
j
)

2

. Finally we apply the B&B procedure presented

in Section 3.3. It returns the rotation maximizing the number of clustered lines,
the associated VPs, the clustering information (i.e. which line belongs to which
VP) and the outlier set (i.e. the lines belonging to none of the M VPs).
Results: we tested our algorithm on the York urban database [28] which
is composed of 102 images acquired in indoor and outdoor man-made environ-
ments. Each image contains the camera intrinsic calibration parameters, a set of
line segments manually extracted and the ground truth line clustering. For ro-
bustness, we removed the segments shorter than 30 pixels. Our algorithm takes
12 Jean-Charles Bazin, Yongduek Seo and Marc Pollefeys
as input the set of lines and returns the line clustering. We applied our algorithm
to each image and compared our results to the ground truth. This comparison
showed that, for the entire database, each line clustered by our algorithm corre-
sponds to its ground truth clustering, which successfully validates our approach.
Some clustering results are shown in Fig 6. The distribution of the execution
time is illustrated in Fig 7: it shows our algorithm runs in just a few seconds.
5 Conclusion
This paper was dedicated to the problem of inlier/outlier detection under a
rotation model. RANSAC is a popular approach to solve this task but is non-
deterministic (dierent runs might lead to dierent results) and might not return
the optimal solution. In contrast, we proposed a method that is mathematically
guaranteed to obtain the largest consensus set, i.e. the highest number of inliers,
in a globally optimal manner, which constitutes our main contribution. Extensive
experiments on synthesized data and challenging real images have successfully
demonstrated the validity of the proposed method.
Acknowledgments
This research, which has been partially carried out at BeingThere Centre, is
supported by the Singapore National Research Foundation under its Interna-
tional Research Centre @ Singapore Funding Initiative and administered by the
IDM Programme Oce. It has also been funded in part by the ECs Seventh
Framework Programme (FP7/2007-2013) / ERC grant #210806 4D Video and
the Sogang University Research Grant of 2012 (201210029.1).
References
1. Hartley, R.I., Zisserman, A.: Multiple View Geometry in Computer Vision. Second
edn. Cambridge University Press (2004)
2. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model
tting with applications to image analysis and automated cartography. In: Com-
munications of the ACM. (1981)
3. Li, H.: Consensus set maximization with guaranteed global optimality for robust
geometry estimation. In: Proc. International Conference on Computer Vision.
(2009) 10741080
4. Agarwal, S., Snavely, N., Simon, I., Seitz, S.M., Szeliski, R.: Building Rome in a
day. Proc. International Conference on Computer Vision (2009) 7279
5. Brown, M., Lowe, D.: Automatic panoramic image stitching using invariant fea-
tures. International Journal of Computer Vision 74 (2007) 5973
6. Bazin, J.C., Demonceaux, C., Vasseur, P., Kweon, I.: Rotation estimation and
vanishing point extraction by omnidirectional vision in urban environment. The
International Journal of Robotics Research 31 (2012) 6381
7. Torr, P., Zisserman, A.: MLESAC: A new robust estimator with application to
estimating image geometry. Computer Vision and Image Understanding (2000)
8. Nister, D.: Preemptive RANSAC for live structure and motion estimation. In:
Proc. International Conference on Computer Vision. Volume 1. (2003) 199
Globally Optimal Consensus Set Maximization Through Rotation Search 13
9. Raguram, R., Frahm, J.M., Pollefeys, M.: A comparative analysis of RANSAC
techniques leading to adaptive real-time random sample consensus. In: Proc. Eu-
ropean Conference on Computer Vision. (2008) 500513
10. Chum, O., Matas, J.: Optimal randomized RANSAC. IEEE Transactions on
Pattern Analysis and Machine Intelligence 30 (2008) 14721482
11. McCormick, G.: Computability of global solutions to factorable nonconvex pro-
grams: part I - convex underestimating problems. Mathematical Programming 10
(1976) 147175
12. Olsson, C., Enqvist, O., Kahl, F.: A polynomial-time bound for matching and
registration with outliers. Proc. IEEE Conference on Computer Vision and Pattern
Recognition (2008) 18
13. Horst, R., Tuy, H.: Global optimization: deterministic approaches. Springer Verlag
(2006)
14. Amberg, B., Vetter, T.: Optimal landmark detection using shape models and
branch and bound. In: Proc. IEEE Conference on Computer Vision and Pattern
Recognition. (2011) 455462
15. Sun, M., Telaprolu, M., Lee, H., Savarese, S.: An ecient branch-and-bound algo-
rithm for optimal human pose estimation. In: Proc. IEEE Conference on Computer
Vision and Pattern Recognition. (2012)
16. Lehmann, A., Leibe, B., Gool, L.V.: Fast PRISM: Branch and bound Hough
transform for object class detection. International Journal of Computer Vision 94
(2011)
17. Lampert, C.H., Blaschko, M.B., Hofmann, T.: Ecient subwindow search: A
branch and bound framework for object localization. IEEE Transactions on Pat-
tern Analysis and Machine Intelligence 31 (2009) 21292142
18. Thakoor, N., Gao, J., Devarajan, V.: Multibody structure-and-motion segmenta-
tion by branch-and-bound model selection. IEEE Transactions on Image Process-
ing 19 (2010)
19. Yu, C., Seo, Y., Lee, S.W.: Global optimization for estimating a BRDF with mul-
tiple specular lobes. In: Proc. IEEE Conference on Computer Vision and Pattern
Recognition. (2010) 18
20. Hartley, R., Kahl, F.: Global optimization through rotation space search. Inter-
national Journal of Computer Vision 82 (2009) 6479
21. Darom, T., Keller, Y.: Scale invariant features for 3D mesh models. IEEE Trans-
actions on Image Processing 21 (2012) 27582769
22. Horn, B.K.P.: Closed-form solution of absolute orientation using unit quaternions.
Journal of the Optical Society of America. A 4 (1987) 629642
23. Lowe, D.: Distinctive image features from scale-invariant keypoints. International
Journal of Computer Vision 20 (2003) 91110
24. Forsyth, D., Ponce, J.: Computer vision: a modern approach. (2002)
25. Bazin, J.C., Seo, Y., Demonceaux, C., Vasseur, P., Ikeuchi, K., Kweon, I., Pollefeys,
M.: Globally optimal line clustering and vanishing point estimation in Manhattan
world. Proc. IEEE Conference on Computer Vision and Pattern Recognition (2012)
26. Moor, R.: Interval Analysis. (1966)
27. Coughlan, J., Yuille, A.: The Manhattan world assumption: Regularities in scene
statistics which enable bayesian inference. Conference on Neural Information Pro-
cessing Systems (2000)
28. Denis, P., Elder, J.H., Estrada, F.J.: Ecient edge-based methods for estimating
manhattan frames in urban imagery. In: Proc. European Conference on Computer
Vision. (2008) 197210