Object Detection by Contour Segment Networks

Vittorio Ferrari1 , Tinne Tuytelaars2 ⋆ , Luc Van Gool1,2

1
Computer Vision Group (BIWI), ETH Zuerich, Switzerland
2
ESAT-PSI, University of Leuven, Belgium
{ferrari,vangool}@vision.ee.ethz.ch, [email protected]

Abstract. We propose a method for object detection in cluttered real images,

given a single hand-drawn example as model. The image edges are partitioned
into contour segments and organized in an image representation which encodes
their interconnections: the Contour Segment Network. The object detection prob-
lem is formulated as finding paths through the network resembling the model
outlines, and a computationally efficient detection technique is presented. An ex-
tensive experimental evaluation on detecting five diverse object classes over hun-
dreds of images demonstrates that our method works in very cluttered images,
allows for scale changes and considerable intra-class shape variation, is robust to
interrupted contours, and is computationally efficient.

1 Introduction
We aim at detecting and localizing objects in real, cluttered images, given a single hand-
drawn example as model of their shape. This example depicts the contour outlines of
an instance of the object class to be detected (e.g. bottles, figure 1d; or mugs, composed
by two outlines as in figure 5a).
The task presents several challenges. The image edges are not reliably extracted
from complex images of natural scenes. The contour of the desired object is typically
fragmented over several pieces, and sometimes parts are missing. Moreover, locally,
edges lack specificity, and can be recognized only when put in the wider context of
the whole shape [2]. In addition, the object often appears in cluttered images. Clutter,
combined with the need for a ‘global view’ of the shape, is the principal source of
difficulty. Finally, the object shape in the test image can differ considerably from the
one of the example, because of variations among instances within an object class (class
variability).
In this paper, we present a new approach to shape matching which addresses all
these issues, and is especially suited to detect objects in substantially cluttered im-
ages. We start by linking the image edges at their discontinuities, and partitioning them
into roughly straight contour segments (section 3). These segments are then connected
along the edges and across their links, to form the image representation at the core of
our method: the Contour Segment Network (section 4). By recording the segment inter-
connections, the network captures the underlying image structure, and enables to cast
object detection as finding paths through the network resembling the model outlines.
We propose a computationally efficient matching algorithm for this purpose (section 5).
⋆
T. Tuytelaars acknowledges support by the Fund for Scientific Research Flanders (Belgium).
The resulting, possibly partial, paths are combined into final detection hypotheses by a
dedicated integration stage (section 6).
Operating on the Contour Segment Network brings two key advantages. First, even
when most of the image is covered by clutter segments, only a limited number is con-
nected to a path corresponding to a model outline. As we detail in section 5, this greatly
limits the choices the matcher has to make, thus allowing to correctly locate objects
even in heavily cluttered images. Besides, it also makes the computational complexity
linear in the number of test image segments, making our system particularly efficient.
Second, since the network connects segments also over edge discontinuities, the system
is robust to interruptions along the object contours, and to short missing parts.
Our method accommodates considerable class variability by a flexible measure of
the similarity between configurations of segments, which focuses on their overall spa-
tial arrangement. This measure first guides the matching process towards network paths
similar to the model outlines, and is then used to evaluate the quality of the produced
paths and to integrate them into final detections. As other important features, our ap-
proach can find multiple object instances in the same image, produces point correspon-
dences, and handles large scale changes.
In section 7 we report results on detecting five diverse object classes over hundreds
of test images. Many of them are severely cluttered, in that the object contours form a
small minority of all image edges, and they comprise only a fraction of the image. Our
results compare favorably against a baseline Chamfer Matcher.

2 Previous work

The construction of our Contour Segment Network (sections 3 - 4) is rooted in earlier

perceptual organization works [14, 12]. However, unlike these, we do not seek to single
out salient edge groups. Instead, we connect all subsequent segments in a single, global
network which comprises all possible contour paths. This enables our main contribu-
tion: to perform object class detection as path search on the network.
Much previous work on shape matching has focused on class variability. Several
measures of shape similarity have been proposed [2, 1]. They can distinguish objects
of different classes, while allowing for variations and deformations within a class. How-
ever, these works assume the object to be in a clean image, thereby avoiding the problem
of localization, and the difficulties of contour detection. Hence, the rest of this review
focuses on methods handling clutter.
Our algorithm of section 5 is related to “local search” [4] and “interpretation
trees” [11], as it iteratively matches model features to test image features. However,
at each iteration it meets an approximately constant, low number of matching candi-
dates (only those connected to the latest matched segment, section 5). Interpretation
Trees / Local Search approaches instead, need consider a large number of test features
(often all of them [4]). As a consequence, our method is far less likely to be confused
by clutter, and has lower computational complexity (linear in the number of test seg-
ments), thus it can afford processing heavily cluttered images (with typically about 300
clutter segments, compared to only 30 in [4]). Besides, both [4, 11] expect the model to
transform rigidly to the test image, while our method allows for shape variations.
 c2 c1
c2

c2
c1
c1
a b c d e

Fig. 1. (a-c) Example links between edgel-chains. (a) Endpoint-to-endpoint link. (b) Tangent-
continuous T-junction link. (c) Tangent-discontinuous link. (d) 8 segments on a bottle-shaped
edgel-chain. (e) A segment (marked with an arc) bridging over link b).

Deformable template matching techniques deform a template shape so as to min-

imize some energy function, e.g. diffusion-snakes [7], elastic matching [5], and ac-
tive shape models [6]. These approaches require rough initialization near the object to
be found. Additionally, several such methods need multiple examples with registered
landmark points [6], and/or do not support scale changes [7]. Chamfer matching meth-
ods [10] can detect shapes in cluttered images, but, as pointed out by [17, 13], they
need a large number of templates to handle shape variations (a thousand in [10]), and
are prone to produce rather high false-positive rates (1-2 per image in [10]). Recently
Berg et al. [3] proposed a powerful point-matching method based on Integer Quadratic
Programming. However, the nature and computational complexity of the optimization
problem require to explicitly set rather low limits on the maximal portion of clutter
points, and on the total number of points considered from the test image (via a sam-
pling scheme). This is not suitable when the objects’ edge points are only a fraction of
the total in the image. Besides, [3] uses real images as models, so it is unclear how it
would perform when given simpler, less informative hand-drawings. The same holds
for [16], whose approach based on edge patches seems unsuited in our setting. Felzen-
szwalb [8] applies Dynamic Programming to find the optimal locations of the vertices
of a polygonal model on a regular image grid. Since the computational complexity is
quadratic in the number of grid points, it is intractable to have a high resolution grid,
which is necessary when the object covers a small portion of the image (while [8] has a
60 × 60 grid, taking 5 minutes, using a 180 × 180 grid would be 81 times slower).
In contrast to previous contributions, our method combines the attractive proper-
ties of dealing with highly cluttered images, allowing for shape variations and large
scale changes, working from a single example, being robust to broken edges, and being
computationally efficient.

3 Early processing

Detecting and linking edgel-chains. Edgels are detected by the excellent Berkeley
natural boundary detector [15], which was recently successfully applied to object recog-
nition [3]. Next, edgels are chained and a smoothing spline curve is fit to each edgel-
chain, providing estimates of the edgels’ tangent orientations.
 c2 c2
c2
c3
c2

s s

c1

c1
c1
c1
rule 1 rule 2 rule 3 rule 4 rule 5 rule 6

Fig. 2. The six rules to build the Contour Segment Network. They connect (arrows) regular seg-
ments and bridging segments (marked with an arc). Rules 2-6 connect segments over different
edgel-chains ci .

Due to the well-known brittleness of edge detection, a contour is often broken into
several edgel-chains. Besides, the ideal contour might have branchings, which are not
captured by simple edgel-chaining. We counter these issues by linking edgel-chains: an
edgel-chain c1 is linked to an edgel-chain c2 if any edgel of c2 lies within a search area
near an endpoint of c1 (figure 1). The search area is an isosceles trapezium. The minor
base rests on the endpoint of c1 , and is perpendicular to the curve’s tangent orientation,
while the height points away from c1 1 . This criterion links c1 to edgel-chains lying
in front of one of its endpoints, thereby indicating that it could continue over c2 . The
trapezium shape expresses that the uncertainty about the continuation of c1 ’s location
grows with the distance from the breakpoint . Note how c1 can link either to an endpoint
of c2 , or to an interior edgel. The latter allows to properly deal with T-junctions, as it
records that the curve could continue in two directions (figure 1b). Besides, we point
out that it is not necessary for the end of c1 to be oriented like the bit of c2 it links to (as
in figure 1b). Tangent-discontinuous links are also possible (figure 1c).
The edgel-chain links are the backbone structure on which the Contour Segment
Network will be built (section 4).
Contour segments. The elements composing the network are contour segments. These
are obtained by partitioning each edgel-chain into roughly straight segments. Figure 1d
shows the segmentation for a bottle-shaped edgel-chain. In addition to these regular
segments, we also construct segments bridging over tangent-continuous links between
edgel-chains. The idea is to bridge the breaks in the edges, thus recovering useful seg-
ments missed due to the breaks.

4 Building the Contour Segment Network

Equipped with edgel-chain links and contour segments, we are ready to build the im-
age representation which lies at the heart of this paper: the Contour Segment Network
(or just network, for short). To this end, we connect segments along edgel-chains, and
across links between edgel-chains. Thanks to the explicit modeling of the edgel-chains’
interconnections, the network supports robust matching of shapes in cluttered images.
1
The dimensions of the trapezium are fixed, and the same in all experiments.
Definitions. Before explaining how to build the network, we give a few definitions.
First, every segment is directed, in that it has a back and a front. This only serves to
differentiate the two endpoints, they have no semantic difference. As a convention, the
front of a segment is followed by the back of the next segment on the edgel-chain.
Second, every edgel-chain link is directed as well: the edgel-chain c1 , on which the
trapezium search-area rests, is at the back, while the other edgel-chain c2 is at the
front. This also defines the front and back endpoints of a segment bridging between
two edgel-chains. For clarity, we use the word links between edgel-chains, and connec-
tions between segments.
Rules. The network is built by applying the following rules, illustrated in figure 2.
These connect the front of each segment to a set of segments, and its back to another
set of segments. Thus the network structure is unconstrained and its complexity adapts
to the image content.
1. The front of a segment is connected to the back of the next segment on the same
edgel-chain.
2. When two edgel-chains c1 , c2 are linked at endpoints, the segment of c1 before the
link is connected to the segment of c2 after the link.
3. Consider a T-junction link (i.e. from an endpoint of c1 to the interior of c2 ). The
segment of c1 before the link is connected to the two segments of c2 with the closest
endpoints. As can be seen in figure 2.3, this records that the contour continues in
both directions.
4. Let s be a segment bridging over a link from c1 to c2 . s is connected to the segment
of c2 coming after its front endpoint, and to the segment of c1 coming before its
back endpoint.
5. Two bridging segments which have consecutive endpoints on the same edgel-chain
are connected. Here ‘consecutive’ means that no other segment lies inbetween.
6. Consider a bridging segment s without front connection, because it covers the front
edgel-chain c2 until its end. If c2 is linked to another edgel-chain c3 , then we con-
nect s to the segment of c3 coming after its front endpoint. An analogue rule applies
if s lacks the back connection.
Although they might seem complex at first sight, the above rules are pretty natural. They
connect two segments if the edges provide evidence that they could be connected on an
ideal edge-map, where all edges would be detected and perfectly chained. Notice how
the last three rules, dedicated to bridging segments, create connections analog to those
made by the first three rules for regular segments. Therefore, both types are treated
consistently.
Since each edgel-chain is typically linked to several others, the rules generate a
complex branching structure, a network of connected segments. The systematic con-
nections across different edgel-chains, together with the proper integration of bridging
segments, make the network robust to incomplete or broken edgel-chains, which are
inevitable in real images. Figure 3 shows a segment on a bottle outline, along with all
connected segments up to depth 8 (those reachable following up to 8 connections). Al-
though there is no single edgel-chain going all around the bottle, there is a path doing
so, by spanning several edgel-chains. It is the task of the forthcoming matching stage to
discover such desired paths.
 s

Fig. 3. Network connectedness. All black segments are connected to S, up to depth 8. They include
a path around the bottle (thick).

5 Basic matching
By processing the test image as described before, we obtain its Contour Segment Net-
work. We also segment the contour chains of the model, giving a set of contour segment
chains along the outlines of the object.
The detection problem can now be formulated as finding paths through the net-
work which resemble the model chains. Let’s first consider a subproblem, termed basic
matching: find the path most resembling a model chain, starting from a basis match
between a model segment and a test image segment. However we do not know a priori
where to start from, as the test image is usually covered by a large majority of clutter
segments. Therefore, we apply the basic matching algorithm described in this section,
starting from all pairs of model and test segment with roughly similar orientations. The
resulting paths are then inspected and integrated into full detection hypotheses in the
next section.
We consider the object transformation from the model to the test image to be com-
posed of a global pose change, plus shape variations due to class variability. The pose
change is modeled by a translation t and a scale change σ, while class variability is
accommodated by a flexible measure of the similarity between configurations of seg-
ments.
The basic matching algorithm. The algorithm starts with a basis match between a
model segment bm and a test segment bt , and then iteratively matches the other model
segments, thereby tracing out a path in the network. The matched path P initially only
contains {bm , bt }.
1. Compute the scale change σ of the basis match.
2. Move to the next model segment m. Points 3-6 will match it to a test segment.
3. Define a set C of candidate test segments. These are all successors 2 of the current
test segment in the network, and their successors (figure 4a). Including successors
at depth 2 brings robustness against spurious test segments which might lie along
the desired path.
2
All segments connected at its free endpoint, i.e. opposite the one connecting to P.
 4. Evaluate the candidates. Each candidate is evaluated according to its orientation
similarity to m, how well it fits in the path P constructed so far, and how strong its
edgels are (more details below).
5. Extend the path. The best candidate cbest is matched to m and {m, cbest } is added
to P.
6. Update σ. Re-estimate the scale change over P (more details below).
7. Iterate. The algorithm iterates to point 2, until the end of the model segment chain,
or until the path comes to a dead end (C = ∅). At this point, the algorithm restarts
from the basis match, proceeding in the backward direction, so as to match the
model segments lying before the basis one.

For simplicity, the algorithm is presented above as greedy. In our actual implementa-
tion, we retain the best two candidates, and then evaluate their possible successors. The
candidate with the best sum of its own score and the score of the best successor wins.
As the algorithm looks one step ahead before making a choice, it can find better paths.
Evaluate the candidates. Each candidate test segment c ∈ C is evaluated by the
following cost function 3

qc = q(m, c, P) = wla Dla (m, c, P) + wld Dld (m, c, P) + wθ Dθ (m, c) (1)

The last term Dθ (m, c) ∈ [0, 1] measures the difference in orientation between m and
c, normalized by π.
The other terms consider the location of c in the context of test segments matched
so far, and compare it to the location of m within the matched model segments. The
first such spatial relation is
1 X
−→i , −
→
Dla (m, c, P) = Dθ (−
mm cti )
|P|
{mi ,ti }∈P

the average difference in direction between vectors −−→i going from m’s center to the
mm
−→
centers of matched model segments mi , and corresponding vectors cti going from c
to the matched test segments ti (see figure 4d). The second relation is analogous, but
focuses on the distances between segments
1 X
σk−

−→i k − k−
→ 
Dld (m, c, P) = mm cti k
σdm |P|
{mi ,ti }∈P

where dm is the diagonal of the model’s bounding-box, and hence σdm is a normal-
ization factor adapted to the current scale change estimate σ. Thus, all three terms of
function (1) are scale invariant.
The proposed cost function grows smoothly as the model transformation departs
from a pure pose change. In particular the Dla term captures the structure of the spatial
arrangements, while still allowing for considerable shape variation. Function (1) is low
when c is located and oriented in a similar way as m, in the context of the rest of the
shape matched so far. Hence, it guides the algorithm towards a path of test segments
with an overall shape similar to the model.
3
In all experiments, the weights are wla = 0.7, wld = 0.15, wθ = 1 − wla − wld = 0.15.
 2 1

cbest 3 bm
2 1 4

bt bt
c
3
m m1

4
m2

a b d

Fig. 4. Basic matching. (a) Iteration 1: basis segment bt , candidates C with qc ≤ 0.3 (black
thin), and best candidate cbest (thick). (b) Matched path P after iteration 4. (c) Model, with basis
segment bm and segments matched at iteration 1-4 labeled. (d) Example vectors used in Dla , Dld .

Analyzing the values of qc over many test cases reveals that for most correct can-
didates qc < 0.15. In order to prevent the algorithm from deviating over a grossly
incorrect path when no plausible candidate is available, we discard all candidates with
qc above the loose threshold qth = 0.3. Hence: C ← {c|qc ≤ qth }.
In addition to the geometric quality qc of a retained candidate c, we also consider
its relevance, in terms of the average strength of its edgels ▽c ∈ [0, 1]. Hence, we set
the overall cost of c to qc · (1 − ▽c ). Experiments show a marked improvement over
treating edgels as binary features, when consistently exploiting edge strength here and
in the path evaluation score (next section).
Update σ. After extending P the scale change σ is re-estimated as follows. Let δm
be the average distance between pairs of edgels along the model segments, and δt be
the corresponding distance for the test segments. Then, set σ = δδmt . This estimation
considers the relative locations of the segments, together with their individual transfor-
mations, and is robust to mismatched segments within a correct path (unlike simpler
measures such as deriving σ from the bounding-box areas). Thanks to this step, σ is
continuously adapted to the growing path of segments, which is useful for computing
Dld when matching segments distant from the basis match. Due to shape variability and
detection inaccuracies, the scale change induced by a single segment holds only locally.
Properties. The basic matching algorithm has several attractive properties, due to op-
erating on the Contour Segment Network. First and foremost, at every iteration it must
chose among only a few candidates (about 4 on average), because only segments con-
nected to the previous one are considered. Since it meets only few distractors, it is likely
to make the right choices and thus find the object even in substantially cluttered images.
The systematic exploitation of connectedness is the key driving force of our system. It
keeps the average number of candidates D low, and independent of the total number of
test segments T . As another consequence, the computational complexity for processing
all basis matches is O(T M D log2 (M )), with M the number of model segments. In
contrast to “local search” [4] and “interpretation trees” [11], this is linear in T , mak-
ing it possible to process images with a very large number of clutter segments (even
thousands). Second, the spatial relations used in Dla , Dld can easily be pre-computed
for all possible segment pairs. During basic matching, evaluating a candidate takes but
a few operations, making the whole algorithm computationally efficient. In our Matlab
implementation, it takes only 10 seconds on average to process the approximately 1000
basis matches occurring when matching a model to a typical test image. Third, thanks
to the careful construction of the network, there is no need for the object contour to be
fully or cleanly detected. Instead, it can be interrupted at several points, short parts can
be missing, and it can be intertwined with clutter contours.

6 Hypothesis integration

Basic matching produces a large set H = {Pi } of matched paths Pi , termed hypothe-
ses. Since there are several correct basis matches to start from along the object contour,
there are typically several correct hypotheses on an object instance (figure 5b+c+d). In
this section we group hypotheses likely to belong to the same object instance, and fuse
them in a single integrated hypothesis. This brings two important advantages. First,
hypotheses matching different parts of the same model contour chain, are combined
into a single, more complete contour. The same holds for hypotheses covering different
model chains, which would otherwise remain disjoint (figure 5d). Second, the presence
of (partially) repeated hypotheses is a valuable indication of their correctness (i.e. that
they cover an object instance and not clutter). Since the basic matcher prefers the cor-
rect path over others, it produces similar hypotheses when starting from different points
along a correct path (figure 5b+c). Clutter paths instead, grow much more randomly.
Hence, hypothesis integration can accumulate the evidence brought by overlapping hy-
potheses, thereby separating them better from clutter.
Before proceeding with the hypothesis integration stage, we evaluate the quality of
each hypothesis P ∈ H. Each segment match {m, t} ∈ P is evaluated with respect to
the others using function (1): q (m, t, P\{m, t}). Whereas during basic matching only
segments matched before were available as reference, here we evaluate {m, t} in the
context of the entire path. The score of {m, t} is now naturally defined by setting the
maximum value qth of q as roof: qth − q (m, t, P\{m, t}). Finally, the total score of
P is the sum over the component matches’ scores, weighed by their relevance (edgel
strength ▽)
1 X
φ(P) = ▽t · (qth − q (m, t, P\{m, t}))
qth
{m,t}∈P

the normalization by q1th makes φ range in [0, |P|]. In order to reduce noise and speedup
further processing, we discard obvious garbage hypotheses, scoring below a low thresh-
old φth = 1.5: H ← {P|φ(P) ≥ φth }.
Hypothesis integration consists of the following two phases:
Grouping phase.
1. Let A be a graph with nodes the hypotheses H, and arcs (Pi , Pj ) weighed by the
(in-)compatibility csim between the pose transformations of Pi , Pj : csim (Pi , Pj ) =
1
2 (c(Pi , Pj ) + c(Pj , Pi )) , with
 
|ti − tj | σi σj
c(Pi , Pj ) = · max ,
d m σi σj σi
 a b c d e

Fig. 5. Hypothesis integration. a) mug model, composed of an outer and an inner chain (hole). b-
d) 3 out of 14 hypotheses in a group. b) and c) are very similar, and arise from two different basis
matches along the outer model chain. Instead, d) covers the mug’s hole. e) All 14 hypothesis are
fused into a complete integrated hypothesis. Thanks to evidence accumulation, its score (28.6) is
much higher than that of individual hypotheses (b scores 2.8). Note the important variations of
the mug’s shape w.r.t the model.

The first factor measures the translation mismatch, normalized by the scale change
σ, while the second factor accounts for the scale mismatch.
2. Partition A using the Clique Partitioning algorithm proposed by [9]. Each resulting
group contains hypotheses with similar pose transformations. The crux is that a
group contains either hypotheses likely to belong to the same object instance, or
some clutter hypotheses. Mixed groups are rare.
Integration phase. We now combine the hypotheses within each group G ⊂ A into a
single integrated hypothesis.
1. Let the central hypothesis Pc of G be the one maximizing
 
X  \ 
φ(Pi ) ·  Pi Pj  · φ(Pj )
Pj ∈{G\Pi }
T
where |Pi Pj | is the number of segment matches present in both Pi and Pj .
The central hypothesis best combines the features of having a good score and being
similar to the others. Hence, it is the best representative of the group. Note how
the selection of Pc is stable w.r.t. fluctuations of the scores, and robust to clutter
hypotheses which occasionally slip into a correct group.
2. Initialize the integrated hypothesis as Gint = Pc , and add the hypothesis B resulting
in the highest combined score φ(Gint ). This means adding the parts of B that match
model segments unexplained by Gint (figure 5d, with initial Gint in 5b). Iteratively
add hypotheses until φ(Gint ) increases no further.
3. Score the integrated hypothesis by taking into account repetitions within the group,
so as to accumulate the evidence for its correctness. φ(Gint ) is updated by multi-
plying the component matches’ scores by the number of times they are repeated.
Evidence accumulation raises the scores of correct integrated hypotheses, thus im-
proving their separation from false-positives.
In addition to assembling partial hypotheses into complete contours and accumulating
evidence, the hypothesis integration stage also enables the detection of multiple object
instances in the same test image (delivered as separate integrated hypotheses). More-
over, the computational cost is low (1-2 seconds on average).
The integrated hypotheses Gint are the final output of the system (called detections).
In case of multiple detections on the same image location, we keep only the one with
the highest score.

7 Results and conclusions

We present results on detecting five diverse object classes (bottles, swans, mugs, gi-
raffes, apple logos) over 255 test images 4 covering several kinds of scenes. In total, the
objects appear 289 times, as some images contain multiple instances. As all images are
collected from Google Images and Flickr, they are taken under varying, uncontrolled
conditions. While most are photographs, some paintings, drawings, and computer ren-
derings are included as well. The target objects appear over a wide range of scales.
Between the smallest and the largest detected swan there is a scale factor of 4, while for
the apple logos class, there is a factor of 6. The system is given only a single hand-drawn
example of each class (figure 7, i2-j3), and its parameters are always kept fixed.
Figures 6 and 7 show example detections. In many test cases the object is success-
fully and accurately localized in spite of extensive clutter, and even when it comprises
only a small portion of the image (e.g. b1, b3, e1, h2). The dominant presence of clutter
edges is illustrated in a2, b2, c2, with the edge-maps for cases a1, b3, c3. The object
contours form only a small minority of all image edges (about 1/30). The capacity of
handling large scale changes is demonstrated in d1 and e1, where the mug sizes dif-
fer by a scale factor of 3. Moreover, the individual shapes of the detected objects vary
considerably, and differ from the models, hence showing the system’s tolerance to class
variations. Compare d3 and e2 to the bottle model, or the variations among different
mugs. In d1 we overlay the model after applying the best possible translation and scale.
Five of the six mugs imaged in figure c1 are found by the system, proving its ability
to detect multiple object instances in the same image. As examples of the accuracy of
our method, figures d2 and g2 display the image contours matched to the object for
cases d1, d3, and g1 (the other cases are reported as the bounding-boxes of the matched
contours).
We quantitatively assess performance as the number of correct detections (bounding-
box on an instance of the target object class) and false positives (other detections). All
five models have been matched to all 255 test images. The thick curves on plots i2-j3
depict the percentage of correct detections (detection-rate) versus the incidence of false-
positives (number of false-positives per image FPPI). The system performs well on all
five classes, and achieves a remarkable 82% average detection rate at the moderate rate
of 0.4 FPPI. For a baseline comparison, we processed the dataset also with a simple
Chamfer Matching algorithm 5 . The model is shifted over the image at several scales,
and the local maxima of the Chamfer distance give detection hypotheses. In case of
multiple overlapping hypotheses, only the strongest one is retained. As the plots show
4
The dataset is available on our website: www.vision.ee.ethz.ch/∼ferrari
5
While this does not include multiple orientation planes, we believe they would improve per-
formance only moderately [13].
(thin curves) the Chamfer Matcher performs markedly worse than our approach, and
reaches an average detection-rate of only 39% at 0.4 FPPI. As also pointed out by [13],
the reason is that the Chamfer distance is about as low on clutter edgels areas as it is on
the target object, resulting in many false-positives hardly distinguishable from correct
detections. The problem is particularly outspoken in our setting, where only a single
template shape is given [17]. Our approach instead, is much more distinctive and thus
brings satisfactory performance even in these highly cluttered images.
In conclusion, the experiments confirm the power of the presented approach in deal-
ing with extensive clutter, large scale changes, and intra-class shape variability, while
taking only a single hand-drawn example as input. Moreover, it is robust to discontin-
uous edges, and is computationally efficient (the complexity is linear in the number
of image segments). As one limitation, models cannot self-cross or branch, therefore
excluding some objects (e.g. chairs, text). Nevertheless, every object with a distinctive
silhouette can be modeled by a set of disjoint outlines, even if a detailed drawing would
feature crossing/branchings (e.g. most animals, tools, and logos). Future work aims at
addressing this issue, as well as learning the class variability from a few examples, to
apply it for constraining the matching.

References
1. R. Basri, L. Costa, D. Geiger, D. Jacobs, Determining the Similarity of Deformable Shapes,
Vision Research, 1998.
2. S. Belongie, J. Malik, J. Puzicha, Shape Matching and Object Recognition Using Shape
Contexts, PAMI, 24:4, 2002.
3. A. Berg, T. Berg and J. Malik, Shape Matching and Object Recognition using Low Distortion
Correspondence, CVPR, 2005.
4. J. R. Beveridge and E. M. Riseman, How Easy is Matching 2D Line Models Using Local
Search?, PAMI, 19:6, 1997
5. A. Del Bimbo, P. Pala, Visual Image Retrieval by Elastic Matching of User Sketches, PAMI,
19:2, 1997.
6. T. Cootes, C. J. Taylor, D. H. Cooper, and J. Graham, Active Shape Models - Their Training
and Application, CVIU, 61:1, 1995.
7. D. Cremers, C. Schnorr, and J. Weickert, Diffusion-Snakes: Combining Statistical Shape
Knowledge and Image Information in a Variational Framework, Workshop on Variational
and Levelset Methods, 2001.
8. P. F. Felzenszwalb, Representation and Detection of Deformable Shapes, CVPR’03
9. V. Ferrari, T. Tuytelaars, and L. Van Gool, Real-time Affine Region Tracking and Coplanar
Grouping, CVPR, 2001.
10. D. Gavrila, V. Philomin, Real-time Object Detection for Smart Vehicles, ICCV’99
11. W. Grimson, T. Lozano-Perez, Localizing Overlapping Parts by Searching the Interpretation
Tree, PAMI, 9:4, 1987.
12. D. Jacobs, Robust and Efficient Detection of Convex Groups, PAMI, 18:1, 1996.
13. B. Leibe, B. Schiele, Pedestrian detection in crowded scenes, CVPR, 2005.
14. D. Lowe, T. Binford, Perceptual organization as a basis for visual recognition, AAAI, 1983
15. D. Martin, C. Fowlkes and J. Malik, Learning to detect natural image boundaries using local
brightness, color, and texture cues, PAMI, 26(5):530-549, 2004.
16. A. Selinger, R. Nelson, A Cubist approach to Object Recognition, ICCV, 1998.
17. A. Thayananthan, B. Stenger, P. Torr, R. Cipolla, Shape Context and Chamfer Matching in
Cluttered Scenes, CVPR, 2003.
 1 2 3

Fig. 6. Results (first page). See text for discussion.

 1 2 3

1 1

0.9 0.9

0.8 0.8

0.7 0.7
Detection rate

Detection rate

0.6 0.6

i 0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.25 0.5 0.75 1 1.25 1.5 0 0.25 0.5 0.75 1 1.25 1.5
False−positives per image False−positives per image
1 1 1

0.9 0.9 0.9

0.8 0.8 0.8

0.7 0.7 0.7

Detection rate

Detection rate

Detection rate

0.6 0.6 0.6

j 0.5 0.5 0.5

0.4 0.4 0.4

0.3 0.3 0.3

0.2 0.2 0.2

0.1 0.1 0.1

0 0 0
0 0.25 0.5 0.75 1 1.25 1.5 0 0.25 0.5 0.75 1 1.25 1.5 0 0.25 0.5 0.75 1 1.25 1.5
False−positives per image False−positives per image False−positives per image

Fig. 7. Results (second page). See text for discussion.