A Bayesian, Exemplar-Based Approach

to Hierarchical Shape Matching

Dariu M. Gavrila
AbstractThis paper presents a novel probabilistic approach to hierarchical, exemplar-based shape matching. No feature
correspondence is needed among exemplars, just a suitable pairwise similarity measure. The approach uses a template tree to
efficiently represent and match the variety of shape exemplars. The tree is generated offline by a bottom-up clustering approach using
stochastic optimization. Online matching involves a simultaneous coarse-to-fine approach over the template tree and over the
transformation parameters. The main contribution of this paper is a Bayesian model to estimate the a posteriori probability of the object
class, after a certain match at a node of the tree. This model takes into account object scale and saliency and allows for a principled
setting of the matching thresholds such that unpromising paths in the tree traversal process are eliminated early on. The proposed
approach was tested in a variety of application domains. Here, results are presented on one of the more challenging domains: real-time
pedestrian detection from a moving vehicle. A significant speed-up is obtained when comparing the proposed probabilistic matching
approach with a manually tuned nonprobabilistic variant, both utilizing the same template tree structure.
Index TermsHierarchical shape matching, chamfer distance, Bayesian models.

1 INTRODUCTION
O
BJECT detection is one of the central tasks in image
understanding. Among the various visual cues that can
be used to segment and compare objects, shape has the
advantage that it provides a powerful object discrimination
capability that is relatively stable to changes in lighting
conditions.
This paper presents a novel Bayesian approach for
hierarchical shape-based object representation and match-
ing. It integrates a number of desirable features: generality,
robustness, and efficiency. The generality refers to the
ability to deal with arbitrary shapes, whether parameterized
(e.g., polygons, ellipses) or not (e.g., outlines of pedes-
trians), whether involving closed contours or not. See Fig. 1.
Objects are described in terms of a set of training shapes or
exemplars, which cover the set of possible appearances due
to geometrical transformations (e.g., rotation, scale) and
intraclass variance (e.g., different pedestrians, different
poses). Nontraining object samples are covered by a defined
maximum allowable dissimilarity from a closest exemplar
in the training set. Thus, no feature correspondence is
required, only pairwise dissimilarities.
The proposed system is robust due to its use of template
matching. It copes relatively well with the effects of
suboptimal segmentation (e.g., edge gaps) or partial
occlusion by the use of correlation which integrates contribu-
tions at various image locations independently of each other.
Template-based systems are however notoriously com-
putationally intensive and, therefore, it is especially with
respect to efficiency that the proposed approach can make a
difference. It employs a combined coarse-to-fine approach
over a hierarchical shape representation and transformation
parameters, which results in significant speed-ups com-
pared to brute-force formulations; gains of several orders of
magnitude are typical. Central is its ability to employ
pruning techniques and to deal with object shape variations
by means of distance transforms.
The proposed object representation and matching ap-
proach contains the following components:
. a set of exemplars capturing object appearance,
. a pairwise similarity measure between exemplars,
. recursive clustering and prototype selection: offline
tree construction, and
. a (probabilistic) matching criterion: online tree
traversal.
This formulationis verygeneric andapplies toa large class
of object detection systems. In this paper, we consider a
particular instantiation based on shape-cues where the
pairwise similarity measure between shape exemplars is
the chamfer distance based on oriented edges [12], [23]. The
tree construction process consists of a recursive clustering
procedure where the objective function, the average in-
tracluster similarity, is optimized stochastically by simulated
annealing.
The main contribution of this paper is a Bayesian model
for estimating the a posteriori probability of the object class,
after a certain match at a node of the tree. This model takes
into account object scale and saliency, and allows for a
principled setting of the matching thresholds at the tree
nodes such that unpromising paths are pruned early on
during the tree traversal process. Fig. 2 illustrates the
overall approach.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 29, NO. 8, AUGUST 2007 1
. The author is with the Machine Perception Department of DaimlerChrysler
R&D, WilhelmRunge St. 11, 89081 Ulm, Germany and with the Intelligent
Systems Lab, Faculty of Science, University of Amsterdam, Kruislaan 403,
1098 SJ Amsterdam, The Netherlands.
E-mail: [email protected].
Manuscript received 24 Jan. 2006; revised 4 Aug. 2006; accepted 19 Sept.
2006; published online 18 Jan. 2007.
Recommended for acceptance by L. Van Gool.
For information on obtaining reprints of this article, please send e-mail to:
[email protected], and reference IEEECS Log Number TPAMI-0041-0106.
Digital Object Identifier no. 10.1109/TPAMI.2007.1062.
0162-8828/07/$25.00 2007 IEEE Published by the IEEE Computer Society
The outline of the paper is as follows: Section 2 discusses
previous work. Section 3 reviews the basic building blocks of
hierarchical shape matching: the use of distance transforms
for shape matching, the offline construction of the template
tree, and, finally, the online matching. Section 4 discusses
various quality measures of a shape-based exemplar repre-
sentation such as specificity, coverage, and compactness.
Furthermore, the effect of object scale is analyzed. This sets
the stage for the probabilistic hierarchical matching model
describedinSection5. The experiments are listedinSection6,
involving many thousands of images with groundtruth data.
Section 7 puts the proposed approach into context and
identifies areas of improvement. Finally, Section 8 lists the
conclusions.
2 PREVIOUS WORK
There is a large body of literature on shape representation
and matching, see, for example, a recent review by Zhang
and Lu [30]. One line of research has dealt with learning
shape models from a set of (closed-contour) training shapes.
Shape registration [13], [26] plays herein a central role. It
involves bringing the points across multiple shapes into
correspondence, factoring out variations due to geometrical
transformations between shapes (e.g., similarity) and
maintaining only those changes related to inherent shape
variation of the object class. The established point corre-
spondences allow embedding the training shapes into a
feature vector space, which in turn enables the computation
of various compact parametric shape representations based
on radial (mean-variance) [14] or modal (linear subspace,
PCA) [6] decompositions or combinations there of [15].
Automatic shape registration methods only stand a
reasonable chance of success if the respective shapes are
sufficiently similar. For example, known methods will fail to
correctly register a pedestrian shape viewed sideways with
the feet apart to one with the feet together. This has negative
implications in terms of the specificity of the derived shape
model, as physically implausible, interpolated shapes are
being represented. In order to cope with a larger set of shape
variations in the training set, Duta et al. [8] and Gavrila et al.
[10] combine shape registration and clustering and derive
from the training samples a representation in terms of
1 shape clusters, where only the (similar) shapes within a
cluster are embedded into the same vector space.
In this paper, we consider a shape representation and
matching approach that makes even weaker assumptions, in
the sense that it does not require closed contours and/or
2 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 29, NO. 8, AUGUST 2007
Fig. 1. Applications: detection of (a) traffic signs, (b) pedestrians,
(c) engine parts, (d) objects in range images, and (e) planes.
Fig. 2. Overview of the proposed Bayesian exemplar-based approach to hierarchical shape matching.
shape registration altogether. Instead, it relies only upon a
pairwise similarity measure between the shape exemplars.
Gavrila and Philomin [12] and Gdalyahu and Weinshall [13]
first explored the use of a hierarchical shape representation
built bottom-up from a set of shapes by dissimilarity-based
clustering. Gavrila and Philomin [12] introduced this
approach for the purpose of efficient exemplar-based object
detection. No particular constraints on the shape exemplars
were assumed. The tree was built by recursive partitional
clustering based on distance transforms (DTs). Simulated
annealing was used to obtain a good clustering solution at
each level of the tree. Gdalyahu and Weinshall [13]
considered closed contours and performed clustering based
on the 1
2
norm of automatically registered points. Their
application context was fast retrieval of (already segmented)
shapes. Later, work by Srivastava et al. [26] considered
closed contours and shape retrieval as well, but involved
geodesics for establishing correspondence. This allowed the
computation of Karcher mean shapes. Olson and Huttenlo-
cher [23] and Amit et al. [1] also construct a hierarchical
representation; given their use of binary correlation in the
later matching stage, the shape prototype is selected so as to
capture overlapping pixels among the shape exemplars.
Olson and Huttenlocher [23] cluster by the chamfer distance,
whereas Amit et al. [1] use the Hamming distance. An
interesting alternative measure for shape similarity is the use
of shape contexts [3].
Given a hierarchical structure derived from a set of
2D shape exemplars by clustering, the next issue is how to
use it for matching. Srivastava et al. [26] suggest binary
hypothesis testing to distinguish between two probabilistic
shape models. Their idea is to start at the top, compare the
query with the shapes at each level, and proceed down
the branch following the best match. Assuming 1 possible
shapes at a particular level of the tree, this can be performed
by 1 1 binary tests. In experiments, this is simplified to
selecting the best matching shape. Amit et al. [1] introduce
successive approximations to likelihood tests arising from a
naive Bayesian statistical model for the edge maps extracted
from the original images.
Hierarchical approaches have also been used for speeding
up matching with a single shape exemplar. Borgefors [5] uses
multiple image resolutions for DT-based matching. Others
use a pruning[16], [24] or a coarse-to-fine approach[25] inthe
parameter space of relevant template transformations. The
latter approaches take advantage of the smooth similarity
measure associated with DT-based matching; one need not
match a template for each location, rotation, or other
transformation.
Exemplar-based shape representations have furthermore
been applied to tracking. Toyama and Blake [28] devise a
probabilistic framework, termed Metric Mixture, which
they use for tracking human bodies and mouths. Stenger et
al. [27] extend the hierarchical representation of [12] by
means of a Bayesian Filter.
Considering previous work, this paper is most related to
[27] and [1] in the sense that no constraints are imposed on
allowable shapes and in that a hierarchical exemplar
representation is combined with a probabilistic matching
model. However, the probabilistic model proposed here is
considerably different. Stenger et al. [27] consider a tracking
context, where the existence of the object class is implicitly
assumed. Their aimis howto efficiently compute the density
function over the state space and adapt this over time. In our
detection context, the existence of the object class needs to be
determined in the first place, thus, statistics regarding the
background also need to be considered. Furthermore, we do
not assume an underlying parameter space which we can
sample in coarse-to-fine fashion, and which defines a
hierarchical exemplar representation. Even when such
parameter space would be available, this approach is likely
to introduce redundancy in the representation, i.e., when
distinct parameter settings generate similar shape exemplars.
Here, the hierarchical structure is determined bottom-up, by
shape clustering directly based on pairwise dissimilarity.
Compared to the detection approach of [1], we do not
require 100 percent detection; instead, we define the
associated matching thresholds based on a Bayesian
a posteriori criterion. Hierarchy construction and matching
is furthermore differentiated by our use of distance trans-
forms as opposed to binary correlation with spread edges.
3 BASIC APPROACH
This section reviews the basic components of hierarchical
exemplar-based shape detection, as covered by our earlier
work [12]. It involves the definition of a pairwise similarity
measure betweenshape exemplars basedon DT(Section 3.1),
the offline construction of a hierarchical representation
(Section 3.2), and the online hierarchical traversal for
matching (Section 3.3).
3.1 Similarity Measure: Distance Transforms
Image matching with distance transforms (DTs) involves
two binary images, a segmented template T and a
segmented image 1, termed feature template and feature
image, respectively. The binary pixel values encode the
presence/absence of a feature (e.g., edges) at a particular
location. Matching T and 1 involves computing the DT of
the feature image 1. This transform converts a binary
feature image into a nonbinary image where each pixel
value denotes the distance to the nearest feature pixel. A
variety of DT algorithms exist, differing in their use of a
particular distance metric and the way distances are
computed [4]. Of particular interest is the class of sequential
DTs, or chamfer transforms, which approximate global
distances in the image by propagating local distances in
raster scan fashion. The chamfer-2-3 variant [2], which we
will use in the experiments, uses a 3 3 neighborhood with
values 2 and 3 to denote distances between horizontal/
vertical neighbors and diagonal neighbors, respectively.
After computing the DT, the template T is mapped onto
the DT image of 1 by a transformation G (e.g., translation,
rotation, scale); the matching measure 1
G
T. 1 is deter-
mined by the pixel values of the DT image which lie
under the feature pixels of the transformed template.
These pixel values form a distribution of distances of the
template features to the nearest features in the image. The
lower these distances are, the better the match between
image and template at this location. One possible matching
measure is the average directed chamfer distance [2]
1
c/oi)ci.G
T. 1
1
jTj
X
t2T
d
1
t. 1
where jTj denotes the number of features in T and d
1
t
denotes the chamfer distance between feature t in T and the
closest feature in 1. Other more robust and computationally
GAVRILA: A BAYESIAN, EXEMPLAR-BASED APPROACH TO HIERARCHICAL SHAPE MATCHING 3
intensive measures reduce the effect of missing features
(i.e., due to occlusion or segmentation errors) by using the
average truncated distance or the )th quantile value (the
Hausdorff distance) [16]. Further work weighs individual
pixel distance contributions based on probabilistic models
for pixel-adjacency [22].
Once DTs and match measure have been defined, the
task of a DT-based object detection system involves finding
a geometrical transformation G for which the distance
measure 1
G
T. 1 lies below a user-supplied dissimilarity
threshold 0
1
G
T. 1 < 0. 2
Fig. 3 illustrates the DT-based matching scheme for the
typical case of edge features. The advantage of matching a
template (Fig. 3b) with the DT image (Fig. 3d) rather than
with the edge image (Fig. 3c) is that the resulting similarity
measure will be smoother as a function of the template
transformation parameters. This enables the use of various
efficient search algorithms to lock onto the correct solution,
as will be discussed shortly. It also allows more variability
between a template and an object of interest in the image.
Matching with the unsegmented (gradient) image, on the
other hand, typically provides strong peak responses but
rapidly declining off-peak responses.
3.2 Construction of Template Tree: Recursive
Partitional Clustering
The basic idea for achieving an efficient shape representa-
tion is to group similar templates together and represent
them by two entities: a prototype template and a distance
parameter. The latter needs to capture the dissimilarity
between the prototype template and the templates it
represents. By matching the prototype with the images,
rather than the individual templates, a significant speed-up
can be achieved online. When applied recursively, this
grouping leads to a template tree, see Fig. 4.
The startingpoint is aset of templates (or shapeexemplars)
which cover inherent object shape variations and allowable
geometrical transformations other than translation (e.g.,
rotation, scale). The templates are assumed aligned with
respect to translation. A template tree is subsequently
constructed on top of these templates. The proposed
algorithm involves a bottom-up approach and a partitional
clustering step at each level of the tree. The input to the
algorithmis a set of templates t
1
. . . . . t
`
, a method to obtain a
prototype template p
/
from a subset of templates and the
desiredpartitionsize 1. The output is the 1-partition andthe
prototype templates p
1
. . . . . p
1
for each of the 1 groups
o
1
. . . . . o
1
. At the leaf level, the input is providedbythe shape
examples, whereas, at the nonleaf level, it is the prototypes
derived at the previous clustering step, one level lower.
1-way clustering is implemented as an iterative function
optimization process. Starting with an initial (random)
partition, templates are moved back and forth between
groups, whilethefollowingobjectivefunction1is minimized
1
X
1
/1
X
i
/
i1
1t
i
. p

/
. 3
Here, 1t
i
. p

/
denotes the distance measure between the
ith element of group / and the prototype p

/
for that group at
the current iteration. i
/
denotes the current size of group /.
Given the availability of shape correspondence, p

/
involves
the analytically computed mean shape [6], [26]. In the more
general case of no shape correspondence, p

/
can be taken as
the template with the smallest mean dissimilarity to the other
templates in a group. This allows the offline computation of
1 to be stored as a dissimilarity matrix.
Alow1-value is desirable since it implies a tight grouping;
this lowers the distance thresholdthat will berequiredduring
matching(see(6)), which, inturn, likelydecreases thenumber
of locations which one needs to consider during matching.
Simulated Annealing (SA) [18] is used to perform the
minimization of 1. SA is a well-known stochastic optimiza-
tion technique where, during the initial stages of the search
procedure, moves can be accepted which increase the
objective function. The aim is to do enough exploration of
the searchspace, before resortingtogreedymoves, inorder to
avoidlocal minima. Candidate moves are acceptedaccording
to probability j:
4 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 29, NO. 8, AUGUST 2007
Fig. 3. (a) Original image. (b) Template. (c) Edge image. (d) DT image.
Fig. 4. A hierarchical structure for pedestrian shapes (partial view).
j
1
1 c
1
T
. 4
where T is the temperature parameter which is adjusted
according to a certain cooling schedule (we use an
exponential schedule [18]).
Other algorithms could have been used for partitional
clustering based on similarity values, e.g., [9], [29]. SA has
some appealing theoretical properties, such as convergence
to the global minimum in the limiting case (e.g., sufficient
high initial temperature, infinitesimal small temperature
decrements). Although the underlying optimality conditions
cannot be met in practice, we selected SA because it
nevertheless tends to outperform deterministic approaches
at still manageable large iteration counts. The drawback of its
computational cost is not a major issue, considering the
template tree is constructed offline. It is worth allocating
substantial resources to devise an efficient representation
offline (in the sense of minimizing 1) because this translates
in online computational gains. See Fig. 4 for a typical partial
view. Observe howthe shape similarity increases toward the
leaf level.
3.3 Hierarchical Matching
Online matchingcanbe seenas traversingthe tree structure of
templates. Processing a node involves matching the corre-
sponding (prototype) template p with the image at some
interest locations. For the locations where the distance
measure between template and image is below a user-
supplied threshold 0
j
, the child nodes are added to the list
of nodes to be processed. For locations where the distance
measure is above-threshold, search does not propagate to the
subtree; it is this pruning capability that brings large
efficiency gains.
The above coarse-to-fine approach is combined with a
coarse-to-fine approach over the transformation parameters
(i.e., translation). Image locations where matching is success-
ful foraparticular nonleaf nodegiverisetoanewset of interest
locations for thechildnodes onafiner gridinthevicinityof the
original locations. See Figs. 5 and 6. At the root, the interest
locations lie on a uniformgrid over the image. By following a
pathinthe tree towardthe leaf node, bothtemplate suitability
and template localization increase. Final detections are the
successful matches at the leaf level of the tree.
Let p be the template corresponding to the node currently
processed during traversal at level | and let C ft
1
. . . . . t
c
g
be the set of templates corresponding to its childnodes. Let c
j
be the maximum distance between p and the elements of C.
c
j
max
t
i
2C
1p. t
i
. 5
Let o
|
be the size of the underlying uniform grid at level | in
grid units and let j denote the distance along the diagonal
of a single unit grid element. Furthermore, let t
to|
denote the
allowed shape dissimilarity value between template and
image at a correct location. Then, by having
0
j
t
to|
c
j

1
2
jo
|
. 6
one has the desirable property that, using untruncated
distance measures such as the chamfer distance, one can
guarantee that the above coarse-to-fine approach using the
template tree will not miss a solution.
Comparing the above hierarchical matching approach
with an equivalent brute-force method, one observes that,
given image width \, image height H, and 1 templates, the
brute-force version would require \ H 1 correlations.
Inthe presentedhierarchical version, bothfactors \ H and
1are pruned(by a coarse-to-fine approachintransformation
space and in template space). It is not possible to provide an
analytical expressionfor thespeed-up, since it depends onthe
actual image data and template distribution. We measured
gains of several orders of magnitude in the applications we
considered.
4 SHAPE EXEMPLAR REPRESENTATION
The hierarchical structure discussed previously is motivated
by efficiency considerations. The best achievable detection
performance, i.e., correct versus false detections, is however
cappedbythe matchingresults obtainedat the leaf level. This,
in turn, depends on how well the shape exemplars and
associateddissimilaritythresholds represent object variation.
Here, we consider some qualitative aspects of a non-
hierarchical, flat exemplar representation. We refer to the
coverage and specificity of a particular exemplar-based
representation as the degree that possible object and
nonobject instantiations lie within a given dissimilarity
interval from a nearby shape exemplar, respectively. We
refer to compactness as to the degree that possible object
instantiations lie within the dissimilarity interval from
multiple shape exemplars. See Fig. 7 for a visualization,
simplified in the sense that, in reality, the exemplars are not
necessarily embedded in a common feature vector space.
Increasing the number of exemplars is generally favorable
GAVRILA: A BAYESIAN, EXEMPLAR-BASED APPROACH TO HIERARCHICAL SHAPE MATCHING 5
Fig. 5. Intermediate matching results for a three-level template tree:
Templates matched successfully at levels 1, 2, 3 (leaf) are shown in white,
gray, and black, respectively.
Fig. 6. Illustration of expanded interest locations on a coarse-to-fine grid
as search goes from top (large light gray dots) to intermediate (medium,
dark gray dots) and leaf level (small black dots) in a three level template
tree.
from the detection performance point of view. An enlarged
(representative) exemplar set allows decreasing the dissim-
ilarity thresholds, increasing specificity without decreasing
coverage. These improvements in detection performance
need, however, to be balanced with increased memory and
computational cost, especially when the compactness of a
representation degrades.
An important issue is how to match at multiple object
scales, given that the exemplar-based representation is not
scale-invariant. One possibility is to maintain the shape
exemplars at a single scale and resize the image accordingly.
This approach avoids the memory cost of storing exemplars
at multiple scales. However, this comes at the expense of
possible lower matching performance, when, due to lower
image resolution, segmentation is degraded (e.g., edge
segmentation in Section 3.1).
Inthe remainder of this paper, we consider the case where,
due to efficiency reasons, shape exemplars are pregenerated
at multiple scales. Matching such multiscale representation
could simply involve scaling-up a dissimilarity threshold
accordingly. However, when scaling up, increasing the
distance thresholds in many cases results in a degradation
of specificity of the representation. This is because of the
presence of spurious (edge) features in the background,
whose density is independent of the scale of the object, and
which are increasingly mismatched. In order to maintain a
particular detection performance, it will be necessary to
counteract this effect by increasing the number of exemplars
in the training set.
We experimentally determine how detection perfor-
mance is influenced by the number of shape exemplars
and how this depends on increasing object scale. This allows
an appropriate choice for the shape exemplars at the leaf
level, i.e., on which the tree is built, see Section 6.1.
5 PROBABILISTIC MATCHING
5.1 Probabilistic Model
One important questionis howtoset the matchingthresholds
associated with each node in the template tree in a principled
manner. Manual parameter setting is not practical for trees
that can contain hundreds if not thousands of nodes. One
possibility is to use (6), but the resulting thresholds are, in
practice, very conservative. In many applications, one can
lower the thresholds to speed up matching at the cost of
possibly missing a solution. In this section, we are interested
to derive ana posteriori probabilitycriteriononwhichto base
our decision rule (thresholds).
First, we introduce some notation. Binary state random
variable A 2 fO. `g denotes the presence of an object O or
background ` at a particular node and image location. At
the leaf level of the tree, the object class O occurs for the best
matching template at the best location. Furthermore, the
object class O
|
occurs at level | for the (optimal) path from
the root to the best matching leaf level node, together with
the associated locations on the coarse-to-fine image grid,
e.g., Fig. 6 (for notational simplicity, we do not include in
the remainder the subscripts regarding image location). The
dissimilarity measurement obtained at the |th level of the
tree, associated with random variable 1
|
, is denoted by
d
|
2 <. Define d
1:|
fd
i
g
|
i1
to be the measurements from
the top level up to level |, along a particular path in the tree.
Desired is a Bayesian framework for modeling the
a posteriori probability of the object class at a particular node
of the tree, given (dissimilarity) measurements along the
path to that node. Given the Bayes rule
jO
|
jd
1:|

jO
|
jd
1:|
jO
|

jd
1:|

7
and
jd
1:|
jO
|
jd
1:|
jO
|
j`
|
jd
1:|
j`
|
. 8
one obtains
jO
|
jd
1:|

jO
|
jd
1:|
jO
|

jO
|
jd
1:|
jO
|
j`
|
jd
1:|
j`
|

1
1
j`
|

jO
|

jd
1:|
j`
|

jd
1:|
jO
|

.
9
Assuming the following Markov property along the path
from the root to the current node,
jd
|
jd
1:|1
A
|
jd
|
jd
|1
A
|
. 10
and considering three possible transitions from a parent
node at level | 1 to a current node at level |,
1. O
|1
O
|
: both parent and current node lie on the
optimal path,
2. O
|1
`
|
: parent lies on the optimal path but current
node does not, and
3. `
|1
`
|
: parent does not lie on the optimal path (and,
consequently, neither does current node).
We arrive to the following recursive form of the posterior
(see the Appendix for derivation):
6 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 29, NO. 8, AUGUST 2007
Fig. 7. Representation of object manifold by exemplars. (a) Good coverage, bad specificity, bad compactness. (b) Bad coverage, good specificity,
good compactness. (c) Reasonable trade-off between coverage, specificity, and compactness.
jO
|
jd
1:|

1
1 c
|
11
with for | 1
Let )
AY
d
|
jd
|1
and )
A
d
1
denote the conditional
probability functions associated with jd
|
jd
|1
AY and
jd
1
jA, respectively. Approximations for the various
) values are derived from histogramming dissimilarity
measurements at the nodes of the template tree. For
example, )
OO
d
|
jd
|1
is derived by collecting dissimilarity
measurements in training images at nonleaf nodes along the
path from the top to the best matching template at the leaf
level. )
O`
d
|
jd
|1
is derived by collecting the dissimilarity
measurements at the nodes and locations which, at the
current level, deviate from this optimal path. )
``
d
|
jd
|1
is
derived by collecting dissimilarity measurements not on the
optimal path at the current or previous level. See Fig. 8.
5.2 Model Instantiation
In practice, it is possible to collect sufficient data for a good
approximation of )
AY
at the higher levels of the tree, where
the nodes are frequently accessed. When examples are scarce
(e.g., typically pertaining to the object class), the aggregation
of dissimilarity measurements at various nodes of the tree
and/or the use of parametric models becomes necessary.
Denote a particular node by its level | in the tree and by
shape t and scale : of underlying template. We model
)
|.t.:
O
d
|
)
|.:
O
d
|
. | 1
)
|.t.:
OO
d
|
jd
|1
)
|.:
OO
d
|
jd
|1
. | 1.
(
12
Thus, given the presence of the object class, dissimilarity
measurements observed at a node are assumed dependent of
the level (accounting for the varying search grid size and
number of prototypes at a level) andobject scale (as discussed
in Section 4); they are assumed to be independent of the
particular template shape. Similarly, we model
)
|.t.:
O`
d
|
jd
|1
)
|.:
O`
d
|
jd
|1
. | 1.
n
13
For a transition within the nonobject class, we make no such
assumptions and maintain
)
|.t.:
`
d
|
| 1
)
|.t.:
``
d
|
jd
|1
| 1.
(
14
Thus, in addition to level and object scale, dissimilarities
are also assumed dependent on template shape (introdu-
cing the aspect of template saliency).
The chi-square and exponential distribution were used
earlier to model )
O
d in the nonhierarchical context [28].
Our experiments indicated an appreciable imprecision in
modeling the tail of various distributions. We therefore
chose to incorporate additional degrees of freedom by
means of the gamma distribution. The gamma probability
density function, parameterized by o and /, is given by
y )djo. /
1
/
o
o
d
o1
c

d
/
o 0. / 0. 15
where d 2 0. 1. The gamma function is defined by the
integral
o
Z
1
0
r
o1
c
r
dr. o 0. 16
The chi-square and exponential distribution are special
cases of the gamma distribution, namely, for / 2
and o / 1, respectively. The experiments show that
distributions )
|.:
O
d
|
. | ! 1 are very well fitted by the
gamma distribution, see Section 4. The same applies for
)
|.t.:
AA
d
|
jd
|1
. | 1, given a discretization of d
|1
.
The sole distribution not fitted well by the gamma
distribution (or by other well-known parametric distribu-
tions) in our preliminary experiments was )
1.t.:
`
d
1
. We
chose to fit a nonparametric model using normal kernel
smoothing [20].
Finally, given that a parent node at level | 1 has
C children and each candidate location is expanded into
1 new locations (on a finer grid, see Fig. 6), we model:
jO
|
jO
|1

1
C 1
. 17
Trivially, j`
|
jO
|1
1 jO
|
jO
|1
.
We are now in the position to derive node-specific
dissimilarity thresholds based on three different criteria.
The first two are based on dissimilarity values directly, the
third on a probability criterion. As a first option, one can
specify a desired object throughput rate 0
|
at each level of the
tree. The associated dissimilarity thresholds c are selected
such that
0
|
1
|.:
O
c
|.:
. 18
where 1
|.:
O
is the cumulative distribution function associated
with )
|.:
O
. Similarly, when specifying a nonobject throughput
rate 0
|
for a certain tree level, one obtains
0
|
1
|.t.:
`
c
|.t.:
. 19
Alternatively, one can specify a threshold 0
|
on the
minimum a posteriori probability by (11). Obviously, the
three criteria cannot be set independently. The last criterion
GAVRILA: A BAYESIAN, EXEMPLAR-BASED APPROACH TO HIERARCHICAL SHAPE MATCHING 7
Fig. 8. Collecting distance measurements during training for the purpose
of estimating )
O
d
1
and )
OO
d
|
jd
|1
(solid black), )
O`
d
|
(solid gray)
and, )
`
d
1
and )
``
d
|
jd
|1
(dotted gray). The best matching solution
at the leaf level is marked by a rectangle. Figure does not capture
multiple image locations.
has the advantage that it allows direct control of the
efficiency of the hierarchical matching process, avoiding the
exploration of unpromising paths in the tree. It is the
criterion we will use at the experiments in next section.
6 EXPERIMENTS
We tested the basic version of the hierarchical shape
detector (Section 3) in a wide range of applications, from
the detection of traffic signs and pedestrians from a moving
vehicle, plane detection in aerial images, engine detection
for visual inspection, to 3D object localization in depth
images for robot vision. See Fig. 1.
Given the large shape variation, the lack of an explicit
model andthe difficult segmentationproblem, the pedestrian
applicationis certainlythemost challengingamongthese. For
example, we had to use more than 100 times as many shape
exemplars for the pedestrian application as for the traffic sign
applicationinorder toobtainadecent performance; thetraffic
sign application involves rigid objects of few standardized
dimensions and shapes. Furthermore, segmenting the edges
of pedestrians is more challenging than those of traffic signs
because of less pronounced contrast. Considering finally the
relevance of pedestrian detection for a number of important
application settings (e.g., driver assistance systems, visual
surveillance), we selected this task to illustrate the concepts
discussed in Sections 4 and 5.
The data sets used in this section involved a wide variety
of pedestrian appearances with different poses (standing
versus running), clothes, and ages of pedestrians and
various (day time) lighting conditions. The pedestrians
were not significantly occluded. Ground truth data was
obtained by manually labeling the pedestrian contours. The
training and test sets were separated.
In all experiments, we used the average directed chamfer-
2-3 distance (1) as the dissimilarity measure because of
efficiency considerations. To alleviate the effects of missing
data, distance contributions of individual pixels were
truncated before being averaged. Chamfer images and
distances were separately computed for eight different edge
orientation-intervals following [12], [23]; matching results
for the individual edge orientation intervals were summed
to an overall match measure.
6.1 Nonhierarchical Exemplar Pedestrian
Representation
In order to obtain an indication of the appropriate number of
exemplars needed at the leaf level of the tree, we first
conducted tests on a nonhierarchical representation. ROC
detection performance was related to the number of
exemplars and object scale (see Section 4). The training set
consisted of a set of 5,749 binary images representing
manually labeled pedestrian shapes. Partitional clustering
was applied on this data set for various values of 1, i.e.,
1 50, 150, 500, 1,500, and 5,749, see Section 3.2. The
resulting 1 shape prototypes were subsequently selected as
the exemplars of a nonhierarchical, flat pedestrian repre-
sentation. The test set consistedof 4,070 and9,770 rectangular
image regions containing object and nonobject class, respec-
tively. See Fig. 9. Both training and test sets were scaled such
that the patterns were of the same size; this was done
separately for sizes of 20, 50, 80, and 140 pixels.
To obtain detection performance at a particular object
scale, we iterated over the samples in the test set (object and
nonobject class separately) and histogrammed the mini-
mum distance to the elements of the training set. From the
two cumulative histograms, we derived the corresponding
ROC curve by considering a particular distance threshold
(x-axis of histogram) and identifying the fraction of object
and nonobject samples which have lower distance values.
The resulting ROC curves are shown in Fig. 10.
Several observations can be made from Fig. 10. At small
object sizes (Fig. 10a), performance is comparatively low;
the exemplar-based shape representation is not able to
capture sufficient object detail to allow an effective
discrimination between object and background. As the
object size increases, performance increases (Figs. 10b and
10c) up to a point, after which, performance decreases again
(Fig. 10d). Here, the number of exemplars is no longer
sufficient to cover the possible shape variation.
A second observation is that, at large false alarm rates,
there is little difference in the performance obtained with
the various values of 1. Evidently, the coverage and
sensitivity obtained with few exemplars is similar to that
reached by using many exemplars, given one uses a large
(relaxed) distance threshold. This is the case of Fig. 7a.
Given that one uses smaller distance thresholds, gaps in
coverage of the pedestrian manifold start appearing, and
8 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 29, NO. 8, AUGUST 2007
Fig. 9. Examples of the object and nonobject class in the test set, shown in the top and bottom row, respectively.
representations with larger numbers of exemplars start to
outperform the ones with fewer exemplars. This is the case
of Fig. 7b. Furthermore, the divergence in performance at
lower false alarms increases for larger object scale (i.e.,
compare Figs. 10a and 10d).
6.2 Hierarchical Pedestrian Detection
The detection experiments involved a training set of
2,666 pedestrian instances and a test set of 2,254 pedestrian
instances (1,306 images). The number of the templates in the
original training set was doubled by mirroring the template
shapes across the y-axis. On the resulting set, a four-level
pedestrian tree was built, following Section 3.2.
The tree construction process was performed separately
for eachof the nine template scales (height range 36-84 pixels,
increments of sixpixels) that were used. At the leaf level of the
scale-specific trees, all available shape exemplars were used
from the training set, appropriately scaled. At a nonleaf
level |, we select the number of template nodes to increase
quadratically with template height / as
`
|./

/
/
iii

2
`
|./
iii
. 20
where `
|./
iii
is the number of templates at level | for the
smallest height /
iii
(36 pixels). We set
`
3./
iii
100. `
2./
iii

&
1
10
`
3./
iii

. `
1./
iii

&
1
10
`
2./
iii

.
21
Fig. 11 illustrates the Simulated Annealing optimization
approach corresponding to shape clustering at the leaf levels
of these nine trees. The figure plots the objective function as a
function of the iteration count. All plots show the same
typical behavior: With an increasing number of iterations,
both short-term variance and mean of the objective function
decrease as the temperature parameter tends toward zero
following the exponential annealing schedule.
Inorder to improve the compactness of the representation,
the leaf level of the original tree was discarded, resulting in a
three-level tree used for matching. Following the above
choices for `
|./
, the new leaf levels of the scale-specific trees
GAVRILA: A BAYESIAN, EXEMPLAR-BASED APPROACH TO HIERARCHICAL SHAPE MATCHING 9
Fig. 10. ROC performance of nonhierarchical exemplar representation as a function of number of exemplars for different object sizes. (a) Size 20.
(b) Size 50. (c) Size 80. (d) Size 140.
contain between100 and544 exemplars, for object scales 36 to
84 pixels, respectively. This corresponds approximately to
the grayROCs (50-150 shapes) inFig. 10 a andFig. 10bandthe
dotted black ROC (500 shapes) in Fig. 10c. It thus indeed
represents a reasonable trade-off between ROC performance
and computational/memory cost, considering the experi-
ments from last section.
The resulting scale-specific trees were subsequently
merged into one overall template tree, which contained 27,
267, and 2,666 templates at the first, second, and leaf level,
respectively. An increase in computational efficiency was
obtained by subsampling the template points, based on the
level of the corresponding node in the tree. We used a point
sampling rate of 6, 3, 1 for the three levels fromtop to bottom,
respectively. The spatial grid sizes on which templates were
matched with the image were o 9. 3. 1 pixels, respectively
(see Fig. 6).
Independently of the particular DT-based dissimilarity
measure used, we found that having essentially only one
edge segmentation threshold was not always appropriate. A
restrictive value would result in sufficient edges to guide the
searchat the coarser level of the tree, but matching at the finer
level would suffer. Setting the edge threshold to include all
edges needed for a fine-level match would be computation-
ally intensive and degrade the underlying coarse-to-fine
concept. In the experiments, we set multiple edge thresholds
and compute the associated distance images based on the
level of the tree where matching was conducted.
With the representational structure and matching logic
in place, we now turn our attention toward selecting the
appropriate dissimilarity thresholds, following the prob-
abilistic approach described in Section 5. The nine different
template scales were aggregated to four scale intervals 36-
48, 54-60, 66-72, 78-84 (index : 1. . . . . 4) for the purpose of
computing the various distribution functions.
Fig. 12a shows the cumulative distribution function of the
distance values at the top level of the tree for the pedestrian
and nonpedestrian class. Recall that, for the object class,
distance distributions were aggregated by object scale (12),
whereas, for the nonobject class, separate distributions were
maintained for each node (14). The four curves associated
with1
1.:
O
d
1
(: 1. . . . . 4) are those inFig. 12a whichhave the
strongest slope upwards. The other 27 curves represent
1
1.t.:
`
d
1
for the nodes at the toplevel. The curves inFig. 12 are
furthermoregray-coded; those correspondingtosmallest and
largest object scale (: 1 and : 4) are shown in light and
dark gray, respectively; the intermediate object scales (: 2
and : 3) are plotted in black.
Fig. 12b shows the computed posterior jOjd
1
at the top
level. We would like to visually verify that the proposed
probabilistic model indeed captures a measure of object
saliency. We indicated in the figure two objects at the same
object scale, which correspond to the maximum and
minimum a posteriori probability for a given distance
value. One observes that the most salient object is one
which involves a pedestrian with feet apart, whereas the
least salient is one which has the feet closed. Indeed, with
10 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 29, NO. 8, AUGUST 2007
Fig. 11. Shape clustering by simulated annealing, objective function
(average intracluster distance) as a function of iteration. Curves
correspond to clustering shape exemplars of various scales.
Fig. 12. At the top level of the template tree. (a) Cumulative distributions 1
1.:
O
d
1
and 1
1.t.:
`
d
1
. (b) Posterior p(Oj d
1
) (scaling factor p(O) set to 1), for
scales : 1. . . . . 4 and nodes i 1. . . . . 27. Plots corresponding to : 1 and : 4 are shown in light and dark gray, respectively.
quite a few diagonally oriented edges, the former pattern
arises less likely by accident in the data set depicting urban
traffic scenes than the pattern which essentially consists of
two major vertical lines. The latter is more likely to match
upon man-made structures in the image.
Fig. 12b furthermore nicely illustrates the need for a
nontrivial adjustment of the distance thresholds withincreas-
ing object scale (see previous discussion in Section 4).
Consider the light gray plots at a posterior value of 0.6, for
example. The associated distance threshold is about 5. A
linear scalingof thedistancethresholdtoobtainanobject scale
comparativelytothedarkgrayplotswouldresult inadistance
thresholdof roughly5 2 (factor 2 because the darkgrayplot
stands for an average template height of 81 pixels while the
light gray plot stands for an average template height of
42 pixels). But, as canbe seenfromFig. 12b, this setting would
result in an a posterior value below0.1; thus, it is significantly
lower than the 0.6 value obtained at the smaller object scale.
Fig. 12b also illustrates what may appear to be a counter-
intuitive result, namely, that for decreasing distances starting
from d
|
1, the posterior decreases back to zero. This might
be considered an aberration given the scarcity of data in that
range, leadingtoamanual specificationof the posterior as 1in
that range. However, a plausible explanation for this result is
that, if a template matches very well (d
|
< 1), this is much
more likely to be the effect of strong edge clutter in the
background than of a very good matching template. In the
experiments, we choose the latter interpretation, not revert-
ing to some ad hoc logic.
Fig. 13 shows the cumulative distributions for the object
class at various scales and levels of the tree. It captures the
decrease of the distance values along the path from the top
of the tree toward a correct solution on the leaf level.
Fig. 14 illustrates the application of parametric models
using the gamma distribution, as discussed in Section 5.2.
Fig. 14a involves various approximations of 1
1.:
O
d
1
at
the top level. Fig. 14b shows a typical result for
)
3.t.:
``
d
3
jd
2
at a leaf level node.
Fig. 15 illustrates some final detection results. Consider-
ing the difficulty of the problem at hand, performance is
quite favorable, with correct detections in a wide range of
scenes. The system is far from flawless, however, with its
main shortcomings being the production of false positives
in heavy textured image regions (e.g., see fourth row, first
and second columns) and nondetections in image areas of
low contrast and occlusion (e.g., see fourth row, third and
fourth columns). The last row shows detection results for a
single image sequence.
We compared the performance of probabilistic hierarch-
ical shape detection, where per-level thresholds involved
the a posteriori criterion, to an earlier, nonprobabilistic
version [12] where the per-level thresholds involved
distance values, properly tuned. Detections were consid-
ered correct if the four corners of the bounding boxes
associated with the found shape template were all within
20 pixels of the manually labeled location. The outcome of
this comparison is summarized in Table 1. As can be seen,
at approximately equal detection and false positive rates,
GAVRILA: A BAYESIAN, EXEMPLAR-BASED APPROACH TO HIERARCHICAL SHAPE MATCHING 11
Fig. 13. Distributions 1
|.:
O
d
|
for level | 1 (light gray), | 2 (dark gray),
and | 3 (black), each plotted for object scale : 1. . . . . 4.
Fig. 14. Distributions and parametric fits. (a) 1
1.:
O
d
1
for object scale : 1. . . . . 4 shown in black, gamma fit in gray. (b) )
3.t.:
``
d
3
jd
2
for particular t and
: and three values of d
2
(vertical lines) shown in black, gamma fit in gray.
the proposed approach manages to reduce computational
cost (determined by the number of pixel correlations) by a
significant factor. The hierarchical shape detector runs
image at 7-15 Hz on a 2.4 GHz Pentium IV processor.
7 DISCUSSION
The previous section has, among other things, shown that a
surprising large variation in object shape can be captured
by a discrete set of shape exemplars when represented in a
hierarchical fashion. This beneficial effect has its limits,
undoubtedly. As was seen in Section 6.1, a strongly
increasing number of exemplars are needed to maintain a
certain ROC performance as the pedestrian comes closer, up
to the point where the approach is not practical anymore
due to storage and processing requirements.
One possible solution is to utilize a hybrid discrete-
continuous shape representation. This could involve match-
ing first with the discrete hierarchical exemplar representa-
tion and, at the leaf level, switching to a more compact
continuous representation, such as the linear subspace shape
model (PDM) employed by Cootes et al. [6]. The premise of
obtaining a sound PDMmodel, namely, very similar training
12 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 29, NO. 8, AUGUST 2007
Fig. 15. Hierarchical probabilistic shape-based pedestrian detection.
TABLE 1
Detection Performance and Computational Cost of State-of-the-Art
Hierarchical Shape Detector versus Proposed Probabilistic Extension
shapes to allow successful automatic shape registration,
would be met at the leaf node of the tree.
Another solution for counteracting the unfavorable
complexity of exemplar-based approaches is the use of
component-based approaches (e.g., [19], [21]). In our case,
separate hierarchical representations could be built for
object parts, and the detection results be merged, taking into
account spatial relationships.
So far, we considered the hierarchical shape detector in
isolation. In a typical application, the shape detector is
combined with other modules for additional robustness and
efficiency. A particular worthwhile combination is the use of
the shape-based detector and a texture-based pattern
classifier for object recognition. Pattern classifiers [17] that
work on pixel values (or derived filter coefficients) tend to be
sensitive to spatial misalignment of a ROI. Applying them
exhaustively over the image, is on the other hand, typically
not an option due to large computational cost. The idea is to
use a shape detector to efficiently localize candidate object
instances, which are subsequently verified with a more
powerful pattern classifier, based on richer texture cues. We
in fact employ this combined approach for the applications
depicted in Fig. 1, e.g., see Fig. 16. In the pedestrian
application, the use of the proposed shape detector further-
more has the advantage that it can index onto a set of
specialized (body pose-specific) texture classifiers. The
resulting mixture-of-experts classifier scheme manages to
reduce the false positives by an order of magnitude, without
appreciably reducing the correct detection rate [11].
Another worthwhile possibility is to precede the shape
detector with an additional attention focusing mechanism.
For example, in the pedestrian system by Gavrila and
Munder [11], stereo vision is usedto quickly identify obstacle
regions in front of a vehicle before initiating shape-based
pedestrian detection. The use of the additional depth cue
furthermore manages to reduce the number of false detec-
tions by a further order of magnitude. The performance
shown in Table 1 is thus in practice significantly enhanced by
the use of preceeding/following modules based on comple-
mentary visual cues. A comparison of state-of-the-art pedes-
trian systems [7], [11], [19], [21], using the same data set and
performance metrics, is worthwhile for future work.
8 CONCLUSIONS
This paper presented a novel probabilistic hierarchical
approachfor shape-basedobject detection. ABayesianmodel
was developed to estimate the a posteriori probability of the
object class at the various node of a tree structure, built
automatically from examples. The model took into account
several object characteristics such as scale and saliency.
In the context of pedestrian detection, this paper
provided an experimental answer to the question of how
many pedestrian exemplars one needs to obtain a certain
detection performance and how this depends on object
scale. The paper furthermore demonstrated the appeal of
utilizing the a posteriori probability criterion at each tree
node in order to directly control the efficiency of
hierarchical shape matching. It showed a significant
speed-up versus a nonprobabilistic matching variant,
where dissimilarity thresholds were manually tuned, one
per tree level.
APPENDIX
For the object class,
jd
1:|
jO
|
jd
1:|
jO
|
O
|1
jO
|1
jO
|

jd
1:|
jO
|
`
|1
j`
|1
jO
|

jd
1:|
jO
|
O
|1
jd
1:|1
jO
|1
jd
|
jd
|1
O
|
O
|1

jO
|1
jd
1:|1

jd
1:|1

jO
|1

jd
|
jd
|1
O
|
O
|1

jO
|1
jd
1:|1

jd
1:|1

jO
|

jd
|
jd
|1
O
|
O
|1
jO
|
jO
|1

22
given jO
|1
jO
|
1, j`
|1
jO
|
0, and
jO
|
,jO
|1
jO
|
jO
|1
.
For the nonobject class,
jd
1:|
j`
|
jd
1:|
j`
|
O
|1
jO
|1
j`
|

jd
1:|
j`
|
`
|1
j`
|1
j`
|

jd
1:|1
jO
|1
jd
|
jd
|1
`
|
O
|1
jO
|1
j`
|

jd
1:|1
j`
|1
jd
|
jd
|1
`
|
`
|1
j`
|1
j`
|

jO
|1
jd
1:|1

jd
1:|1

jO
|1

jd
|
jd
|1
`
|
O
|1

j`
|
jO
|1

jO
|1

j`
|

j`
|1
jd
1:|1

jd
1:|1

j`
|1

jd
|
jd
|1
`
|
`
|1
j`
|
j`
|1

j`
|1

j`
|

jO
|1
jd
1:|1

jd
1:|1

j`
|

jd
|
jd
|1
`
|
O
|1
j`
|
jO
|1

j`
|1
jd
1:|1

jd
1:|1

j`
|

jd
|
jd
|1
`
|
`
|1

23
given j`
|
j`
|1
1.
Substituting (22) and (23) in (9), we obtain the recursive
form of the Bayes rule of (11).
ACKNOWLEDGMENTS
The author wouldlike tothankM. Hofmannfor his assistance
at the experiments of Section 6.1. He also appreciates the
many interesting discussions with S. Munder.
GAVRILA: A BAYESIAN, EXEMPLAR-BASED APPROACH TO HIERARCHICAL SHAPE MATCHING 13
Fig. 16. Pedestrian classificationdetections shown in white, solutions
classified as pedestrians marked by STOP sign.
REFERENCES
[1] Y. Amit, D. Geman, and X. Fan, A Coarse-to-Fine Strategy for
Multiclass Shape Detection, IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 26, no. 12, pp. 1606-1621, Dec. 2004.
[2] H. Barrow et al., Parametric Correspondence and Chamfer
Matching: Two New Techniques for Image Matching, Proc. Intl
Joint Conf. Artificial Intelligence, pp. 659-663, 1977.
[3] S. Belongie, J. Malik, and J. Puzicha, Shape Matching and Object
Recognition Using Shape Contexts, IEEE Trans. Pattern Analysis
and Machine Intelligence, vol. 24, no. 4, pp. 509-522, May 2002.
[4] G. Borgefors, Distance Transformations in Digital Images,
J. Computer Graphics, Vision, Image Processing, vol. 34, no. 3,
pp. 344-371, June 1986.
[5] G. Borgefors, Hierarchical Chamfer Matching: A Parametric
Edge Matching Algorithm, IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 10, no. 6, pp. 849-865, Nov. 1988.
[6] T. Cootes, C. Taylor, D. Cooper, and J. Graham, Active Shape
ModelsTheir Training and Applications, Computer Vision and
Image Understanding, vol. 61, no. 1, pp. 38-59, 1995.
[7] N. Dalal and B. Triggs, Histograms of Oriented Gradients for
Human Detection, Proc. Conf. Computer Vision and Pattern
Recognition, pp. pp. 886-893, 2005.
[8] N. Duta, A.K. Jain, and M.-P. Dubuisson-Jolly, Automatic
Construction of 2D Shape Models, IEEE Trans. Pattern Analysis
and Machine Intelligence, vol. 23, no. 5, pp. 433-446, May 2001.
[9] C. Fowlkes, S. Belongie, F. Chung, and J. Malik, Spectral Grouping
Using the Nystrom Method, IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 26, no. 2, pp. 214-225, Feb. 2004.
[10] D.M. Gavrila, J. Giebel, and H. Neumann, Learning Shape
Models from Examples, Proc. German Assoc. Pattern Recognition
Conf., pp. 369-376, 2001.
[11] D.M. Gavrila and S. Munder, Multi-Cue Pedestrian Detection
and Tracking from a Moving Vehicle, Intl J. Computer Vision,
vol. 73, no. 1, pp.41-59, June 2007.
[12] D.M. Gavrila and V. Philomin, Real-Time Object Detection
for Smart Vehicles, Proc. Intl Conf. Computer Vision, pp. 87-
93, 1999.
[13] Y. Gdalyahu and D. Weinshall, Flexible Syntatic Matching of
Curves and Its Application to Automatic Hierarchical Classifica-
tion of Silhouettes, IEEE Trans. Pattern Analysis and Machine
Intelligence, vol. 21, no. 12, pp. 1312-1328, Dec. 1999.
[14] C. Goodall, Procrustes Methods in the Statistical Analysis of
Shape, J. Royal Statistical Soc. B, vol. 53, no. 2, pp. 285-339, 1991.
[15] T. Heap and D. Hogg, Improving the Specificity in PDMs Using a
Hierarchical Approach, Proc. British Machine Vision Conf., 1997.
[16] D. Huttenlocher, G. Klanderman, and W.J. Rucklidge, Compar-
ing Images Using the Hausdorff Distance, IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 15, no. 9, pp. 850-863, Sept.
1993.
[17] A. Jain, R. Duin, and J. Mao, Statistical Pattern Recognition: A
Review, IEEE Trans. Pattern Analysis and Machine Intelligence,
vol. 22, no. 1, pp. 4-37, Jan. 2000.
[18] S. Kirkpatrick Jr., C.D. Gelatt, and M.P. Vecchi, Optimization by
Simulated Annealing, Science, vol. 220, pp. 671-680, 1993.
[19] B. Leibe, E. Seemann, and B. Schiele, Pedestrian Detection in
Crowded Scenes, Proc. Conf. Computer Vision and Pattern
Recognition, pp. 878-885, 2005.
[20] MathWorks Matlab, Function ksdensity, 2005.
[21] A. Mohan, C. Papageorgiou, andT. Poggio, Example-BasedObject
Detection in Images by Components, IEEE Trans. Pattern Analysis
and Machine Intelligence, vol. 23, no. 4, pp. 349-361, Apr. 2001.
[22] C.F. Olson, A Probabilistic Formulation for Hausdorff Match-
ing, Proc. Conf. Computer Vision and Pattern Recognition, 1998.
[23] C.F. Olson and D.P. Huttenlocher, Automatic Target Recognition
by Matching Oriented Edge Pixels, IEEE Trans. Image Processing,
vol. 6, no. 1, pp. 103-113, Jan. 1997.
[24] D.W. Paglieroni, G.E. Ford, and E.M. Tsujimoto, The Position-
Orientation Masking Approach to Parametric Search for Template
Matching, IEEE Trans. Pattern Analysis and Machine Intelligence,
vol. 16, no. 7, pp. 740-747, July 1994.
[25] W. Rucklidge, Locating Objects Using the Hausdorff Distance,
Proc. Intl Conf. Computer Vision, pp. 457-464, 1995.
[26] A. Srivastava, S.H. Joshi, W. Mio, and X. Liu, Statistical Shape
Analysis: Clustering, Learning, and Testing, IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 27, no. 4, pp. 590-602, Apr.
2005.
[27] B. Stenger, A. Thayananthan, P. Torr, and R. Cipolla, Model-
Based Hand Tracking Using a Hierarchical Bayesian Filter, IEEE
Trans. Pattern Analysis and Machine Intelligence, vol. 28, no. 9,
pp. pp.1372-1385, Sept. 2006.
[28] K. Toyama and A. Blake, Probabilistic Tracking with Exemplars in
a Metric Space, Intl J. Computer Vision, vol. 48, no. 1, pp. 9-19, 2002.
[29] M. Yang and K. Wu, A Similarity-Based Robust Clustering
Method, IEEE Trans. Pattern Analysis and Machine Intelligence,
vol. 26, no. 4, pp. 434-448, Apr. 2004.
[30] D. Zhang and G. Lu, Review of Shape Representation and
Description Techniques, Pattern Recognition, vol. 37, pp. 1-19, 2004.
Dariu M. Gavrila received the MSc degree in
computer science from the Free University in
Amsterdamin 1990. He received the PhDdegree
in computer science from the University of
Maryland at College Park in 1996. He was a
visiting researcher at the MIT Media Laboratory
in 1996. Since 1997, he has been a senior
research scientist at DaimlerChrysler Research
in Ulm, Germany. In 2003, he was named
professor in the Faculty of Science at the
University of Amsterdam, chairing the area of Intelligent Perception
Systems (part time). Over the last decade, Professor Gavrila has
specialized in visual systems for detecting human presence and
recognizing activity, with application to intelligent vehicles and surveil-
lance. He has published more than 20 papers in this area in leading vision
conferences and journals. His personal Web site is www.gavrila.net.
> For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.
14 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 29, NO. 8, AUGUST 2007