1602 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 28, NO.

10, OCTOBER 2006

Integral Invariants for Shape Matching

Siddharth Manay, Daniel Cremers, Member, IEEE, Byung-Woo Hong,
Anthony J. Yezzi Jr., Member, IEEE, and Stefano Soatto, Member, IEEE

Abstract—For shapes represented as closed planar contours, we introduce a class of functionals which are invariant with respect to
the Euclidean group and which are obtained by performing integral operations. While such integral invariants enjoy some of the
desirable properties of their differential counterparts, such as locality of computation (which allows matching under occlusions) and
uniqueness of representation (asymptotically), they do not exhibit the noise sensitivity associated with differential quantities and,
therefore, do not require presmoothing of the input shape. Our formulation allows the analysis of shapes at multiple scales. Based on
integral invariants, we define a notion of distance between shapes. The proposed distance measure can be computed efficiently and
allows warping the shape boundaries onto each other; its computation results in optimal point correspondence as an intermediate step.
Numerical results on shape matching demonstrate that this framework can match shapes despite the deformation of subparts, missing
parts and noise. As a quantitative analysis, we report matching scores for shape retrieval from a database.

Index Terms—Integral invariants, shape, shape matching, shape distance, shape retrieval.

1 INTRODUCTION

G EOMETRIC invariance is an important issue in computer

vision that has received considerable attention in the
past. The idea that one could compute functions of geometric
principled approach has so far been elusive. Nowadays, the
field is instead engaged in searching for invariant (or
insensitive) measures of photometric (rather than geometric)
primitives of the image that do not change under the various nuisances in the image formation process. Nevertheless, the
nuisances of image formation and viewing geometry was idea of computing functions that are invariant with respect to
appealing; it held potential for application to recognition, group transformations of the image domain remains im-
correspondence, 3D reconstruction, and visualization. The portant because it holds the promise to extract compact,
discovery that there exist no generic viewpoint invariants efficient representations for shape matching, indexing, and
was only a minor roadblock, as image deformations can be ultimately recognition.
approximated with homographies; hence, the study of
invariants to projective transformations and their subgroups 1.1 Why Shape Distances?
(affine, similarity, Euclidean) flourished. Toward the end of Our ultimate goal is to compare objects represented as
the last decade, the decrease in popularity of research on closed planar contours. This has obvious implications in
geometric invariance was sanctioned mostly by two factors: shape classification for object recognition, content-based
the progress on multiple view geometry (one way to achieve image retrieval, medical diagnosis, etc. At this level of
viewpoint invariance is to estimate the viewing geometry) generality, this is a monumental task that admits no simple
and noise. Ultimately, algorithms based on invariants did not meaningful solution [55]. Therefore, before we proceed any
meet expectations because most entailed computing various further, we need to specify what we mean by “objects,” how
derivatives of measured functions of the image (hence, the we describe their “shape” and concentrate our attention on
name “differential invariants”). As soon as noise was present particular ways in which they can “differ.” Within the scope
and affected the geometric primitives computed from the of this paper, by object we mean a closed planar contour1
images, the invariants were dominated by the small-scale with no self-intersections embedded in IR2 ; its shape is the
perturbations. Various palliative measures were taken, such equivalence class of objects obtained under the action of a
as the introduction of scale-space smoothing, but a more finite-dimensional group, such as the Euclidean, similarity,
affine, or projective group [42]. In other words, two objects
have the same shape if and only if one can be generated by
. S. Manay is with the Electronics Engineering Technologies Division, rotating, translating and scaling the other. However, in
Lawrence Livermore National Laboratory, PO Box 508, L-290 Livermore, comparing shapes, we want to be insensitive to certain
CA 94551-0508. E-mail: smanay.ece98@gtalumni.org. variations that can occur to an object; for instance, in Fig. 1,
. D. Cremers is with the Department of Computer Science, University of we want rotated, jagged, articulated, and occluded hands to
Bonn, Roemerstrasse 164, D-53117 Germany.
E-mail: dcremers@cs.uni-bonn.de. be judged as having shapes that are similar to that of the
. B.-W. Hong and S. Soatto are with the Computer Science Department, original hand. We prefer not to use the word “noise” when
University of California Los Angeles, 3811 Boelter Hall, Los Angeles, CA referring to these variations because, with the exception of
90095. E-mail: hong@cs.ucla.edu, soatto@ucla.edu. the jaggedness, they are not obtained with standard
. A.J. Yezzi Jr. is with the School of Electrical and Computer Engineering,
Georgia Institute of Technology, 777 Atlantic Drive NW, Atlanta, GA
additive, zero-mean, or small variance perturbations. For
30332-0250. E-mail: ayezzi@ece.gatech.edu. the case of the articulated and occluded objects, for instance,
Manuscript received 12 Apr. 2005; revised 16 Feb. 2006; accepted 21 Feb.
the perturbation can be quite significant in energy and
2006; published online 11 Aug. 2006. highly localized along the contour. Our goal is to define a
Recommended for acceptance by Y. Amit.
For information on obtaining reprints of this article, please send e-mail to: 1. Many of our considerations can be extended to compact surfaces
tpami@computer.org, and reference IEEECS Log Number TPAMI-0197-0405. embedded in IR3 .
0162-8828/06/$20.00 ß 2006 IEEE Published by the IEEE Computer Society
Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
MANAY ET AL.: INTEGRAL INVARIANTS FOR SHAPE MATCHING 1603

is defined as the integral (over the shape) of the difference

between the invariant values of corresponding points. By
minimizing an appropriate energy functional, we compute
the optimal correspondence, which is affected both by
differences in the local geometry of the two curves and by
Fig. 1. Variations of a sample shape. the amount of stretching or shrinking of the shapes’
parametrization required to map similar points to each other.
shape distance so that shapes that vary by Euclidean Given this dense correspondence, the concepts of shape
transformations have zero distance and shapes that vary by comparison, modeling, and interpolation can be naturally
scaling and moderate articulation and occlusion have small derived. We compute the optimal correspondence by casting
distance. In Fig. 1, all the hands should have low distance to the problem as one of identifying the shortest path in a graph
each other, but high distance to other classes of shapes. structure, the nodes of which label possible correspondences
The type of variations we want to be resistant to can be between the points of the two contours. Similar shortest-path
lumped in three categories: “small deformations” that result concepts were exploited in the context of shape matching and
in small set-symmetric differences between the interior of the warping in [65], [76], [30], [41], [84], [4], [85].
curves being compared, “high-frequency noise” that affects a In this paper, we briefly review the literature on shape
large portion of the contour and “localized changes” that analysis in this context (Section 2) before defining integral
significantly affect the total arclength of the contour but are invariants and giving a few examples (Section 3). We then
spatially localized, such as spikes or wedges. Many research- explore an optimization framework for computing shape
ers have addressed the comparison of objects under small distance and shape matching from invariants (Section 4), and
deformations in a way that is invariant with respect to various detail the implementation of this framework (Section 5). In
transformation groups (see Section 2); fewer have addressed Section 4.1, we discuss the extension of the proposed integral
the sensitivity to high-frequency noise and yet fewer have invariants to multiscale analysis. Finally, before concluding,
addressed localized changes [76], [65]. In this paper, we plan we demonstrate our method for computing correspondence
to develop a framework that will address all of these and shape distance on noisy shapes (Section 6).
variations simultaneously. To this end, we plan to employ a
representation of shape in terms of integral invariants, so that 2 PREVIOUS WORK AND OUR CONTRIBUTION
the distance between objects will by construction be invariant
with respect to the action of the chosen group; basing these Given the wealth of existing work on invariance, scale-space,
invariants on integral computations allows us to address and correspondence, our work naturally relates with a large
high-frequency noise in a principled way. Finally, establish- body of literature, as we describe in the next subsection. The
ing point correspondence (also called shape matching) among reader should notice that we consider each object as one entity
contours allows us to handle localized changes and small and perform no analysis or decomposition, so there is no
deformations. All these approaches are integrated into a notion of hierarchy or compositionality in our representation,
shape distance that is insensitive to all these nuisances. which is therefore intrinsically low-level.

1.2 Differential versus Integral Invariants 2.1 Shape and Shape Matching
Commonly, shape invariants are defined via differential In the literature, one finds various definitions of the term
operations. As a consequence, they are inherently sensitive to shape. Kendall, for example, defines shape as whatever
noise. As most practical applications of invariants require remains of point coordinates once you factor out a certain
some robustness to small perturbations of the shape, it is group transformation—for example, the similarity group
necessary to revert to smoothing and accept the unfortunate covering translation, rotation, and scaling. We refer to [26]
side effect that meaningful information will be lost as well. In for a short review of the history of shape research. In this
this work, we introduce invariants which are defined as work, we refer to shape as a closed planar contour modulo
integral functions of the shape. We restrict our analysis to certain group transformations. Moreover, we will denote by
Euclidean invariants, although extensions to the similarity shape matching the process of putting into correspondence
and affine groups fit within the framework we propose. These different parts of two given shapes. Applications of shape
integral invariants share the nice features of their differential matching in computer vision include the classification of
counterparts, being invariant to certain group transforma- objects and the retrieval of objects of the same class based
tions and being local descriptors, which make them well- on the similarity of the object boundary [41]. In medical
suited for matching under occlusions. Yet, in contrast to the imaging, a given anatomical structure may be modeled by a
differential invariants, the integral ones are inherently robust statistical shape representation [45], [21]. Statistical repre-
to noise and, therefore, do not require any preprocessing of sentations of shape may also be useful when modeling
the input data. In addition, they have the favorable feature complex shape deformations; for example, when observing
that varying the size of the integration kernel provides a the silhouette of a 3D object in various 2D views [18].
natural multiscale notion that, unlike differential scale Intermediate shapes between two objects can be generally
spaces, does not require destructive smoothing. obtained based on their correspondence [76].
There exists a vast literature on comparing shapes,
1.3 From Invariants to Shape Distances represented as a collection of points [3], [75], [87], [41], curves
Based on integral invariants, we define a shape distance [48], [91], [64], [5], [30], [92], [77], and continuous curves
between matching parts. Here, a meaningful shape matching, reduced to various types of graph representations [94], [78],
a dense correspondence mapping the parametrized domains [80], [63], [49], [38]; we represent curves as continuous objects
of one shape to another (and vice versa), is crucial, as distance living in infinite-dimensional spaces. (In Section 5, we sample

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
1604 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 28, NO. 10, OCTOBER 2006

the curve for implementation purposes.) Within this choice, To overcome the limitations of differential invariants,
many have addressed matching curves under various types there have been attempts to derive invariants based on
of motions [3], [75], [87] and deformations [48], [91], [64], [5], integral computations. A statistical approach to describe
[30], [22], [92], [15], [82], [77], some involving a mapping from invariants was introduced using moments in [35]. Moment
one curve to another that has some optimality property [6], invariants under affine transformations were derived from
[30], [48], [64], [91], [15], [82], [5], [92], [85], [77]. the classical moment invariants in [27]. They have a limitation
The role of invariants in computer vision has been in that high-order moments are sensitive to noise which
advocated for various applications ranging from shape results in high variances. The error analysis and analytic
representation [57], [7] to shape matching [6], [46], quality characterization of moment descriptors were studied in [47].
control [88], [14], and general object recognition [66], [1]. The Fourier transform was also applied to obtain integral
Consequently, a number of features that are invariant under invariants [93], [52], [2]. A closed curve was represented by a
specific transformations have been investigated [24], [39], set of Fourier coefficients and normalized Fourier descriptors
[25], [33], [56], [81], [74]. were used to compute affine invariants. In this method, high-
In particular, one can construct primitive invariants of order Fourier coefficients are involved and they are not stable
algebraic entities such as lines, conics, and polynomial with respect to noise. Several techniques have been devel-
curves, based on a global descriptor of shape [59], [28]. oped to restrict the computation to local neighborhoods. The
In addition to invariants to transformation groups, wavelet transform was used for affine invariants using the
considerable attention has been devoted to invariants with dyadic wavelet in [83] and potentials were also proposed to
respect to the geometric relationship between 3D objects and preserve locality [34]. Alternatively, semilocal integral in-
their 2D views. Invariant features can be computed from a variants are presented by integrating object curves with
collection of coplanar points or lines [67], [68], [32], [9], [29], respect to arc length [73]. More recently, attempts to develop
[95], [1], [79], [40]. invariants with the locality properties, but without the
An invariant descriptor of a collection of points that relates sensitivity, of differential invariants have resulted in func-
to our approach is the shape context introduced by Belongie tions of curves that are based not on differential operators, but
et al. [6], which consists of a radial histogram of the relative on integral operators applied to the contour or the character-
coordinates of the rest of the shape at each point. istic function of its interior [50], [65].
Differential invariants to actions of various Lie groups In this paper, we introduce a framework for integral
have been addressed thoroughly [44], [36], [17], [58], [76], invariants along with two general classes of integral
[30], [48], [64], [91], [85], [37]. An invariant is defined by an invariants. We use the resulting invariant descriptions to
unchanged subset of the manifold which the group define a notion of distance between shapes and we illustrate
transformation is acting on. In particular, an invariant the potential of our representation on several experiments
signature which pairs curvature and its first derivative with simulated and real images.
avoids parameterization in terms of arc length [13], [60].
Calabi and coworkers suggested numerical expressions for 2.2 Implicit versus Explicit Contour
curvature and first derivative of curvature in terms of joint Representations
invariants. However, it is shown that the expression for the In the context of image segmentation, the implicit repre-
first derivative of curvature is not convergent and modified sentation of closed contours as the zero-crossing of
formulas are presented in [8]. corresponding embedding functions has become increas-
In order to reduce noise-induced fluctuations of the ingly popular. The level set method [23], [61] provides a
signature, semidifferential invariant methods are intro- framework to elegantly propagate boundaries in a way
duced by using first derivatives and one reference point which allows for topological changes of the embedded
instead of curvature, thus avoiding the computation of high- contour and does not require reparameterization. Recently,
order derivatives [62], [31], [43]. Another semi-invariant is shape dissimilarity measures and statistical shape models
given by transforming the given coordinate system to a have been formulated on the basis of the level set
canonical one [89]. representation [45], [86], [69], [20], [19]. Yet, such implicit
A useful property of differential and (some) semi-differ- representations do not provide inherent support for
ential invariants is that they can be applied to match shapes pointwise correspondences. In order to model the notion
despite occlusions, due to the locality of the signature [11], of corresponding features and parts and, therefore, take
[10]. However, the fundamental problem of differential these correspondences into account in a model of shape
invariants is that high-order derivatives have to be computed, similarity (quantified by shape distance), we revert to
amplifying the effect of noise. There have been several explicit parameterizations of closed contours.
approaches to decrease sensitivity to noise by employing
scale-space via linear filtering [90]. The combination of
invariant theory with geometric multiscale analysis is 3 INTEGRAL INVARIANTS
investigated by applying an invariant diffusion equation for In this section, we focus on the definition and examples of
curve evolution [71], [72], [16]. A scale parameter is another integral invariants.
way to build a scale-space which is determined by the size of Throughout this section, we indicate with C : SS1 ! IR2 a
the differencing interval used to approximate derivatives closed planar contour with infinitesimal arclength ds and G
using finite differences [12]. In [54], a curvature scale-space a group acting on IR2 , with dx the area form on IR2 . We also
was developed for a shape matching problem. A set of use the formal notation C to indicate the interior of the
Gaussian kernels was applied to build a scale-space of region bounded by C.  is either the curve C itself (a one-
curvature whose extrema were observed across scales. dimensional object) or C (a two-dimensional object) and

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
MANAY ET AL.: INTEGRAL INVARIANTS FOR SHAPE MATCHING 1605

Fig. 2. Demonstration of the effect of noise on different invariants. (a) A rectangular shape. (b) A rectangular shape with noise. (c) Curvature of (a).
(d) Curvature of (b). (e) Distance Integral Invariant of (a). (f) Distance Integral Invariant of (b). (g) Local Distance Integral Invariant of (a). (h) Local
Distance Integral Invariant of (b). (i) Local Area Integral Invariant of (a). (j) Local Area Integral Invariant of (b).

dðxÞ the corresponding measure, i.e., the area form dx or transformation of SEð2Þ, which will be considered as a
the infinitesimal arclength ds, respectively. transformation group G for the following invariants. How-
Definition 1. Let G be a transformation group acting on IR2 . A ever, as shown in Fig. 2d, it is sensitive to noise because the
function I : IR2 ! IR is a G-invariant if it satisfies calculation of the curvature is dependent on second-order
derivatives. Thus, we introduce an invariant that is robust to
IðCÞ ¼ Iðg  CÞ; 8g 2 G: ð1Þ noise by employing integral operations for its calculation. We
begin with a general notion of integral invariant.
The function IðÞ associates to each point on the contour a
real number. In particular, if the point p 2 C is parameterized Definition 2. A function IC ðpÞ : IR2 ! IR is an integral
by arclength, the invariant can be interpreted as a function G-invariant if there exists a kernel k : IR2  IR2 ! IR such that
from ½0; L, where L is the length of the curve, to the reals: Z
IC ðpÞ ¼ kðp; xÞdðxÞ; ð4Þ
fC : SS1 ! IR2 g7!fIC ðpðsÞÞ : ½0; L ! IRg: ð2Þ 

Similarly, if p 2 C is parameterized from ½0; 1, the invariant where kð; Þ satisfies
can be interpreted as a function from ½0; 1 to the reals: Z Z
kðp; xÞdðxÞ ¼ kðgp; xÞdðxÞ 8 g 2 G; ð5Þ
fC : SS1 ! IR2 g7!fIC ðpðsÞÞ : ½0; 1 ! IRg: ð3Þ  g
:
We abuse this generalized notation in our discussions. g¼fgx j g 2 G; x 2 g.
This formal definition of an invariant includes some very
familiar examples, such as curvature. The definition can be extended to vector invariants2 or to
Example 1 (Curvature). For G ¼ SEð2Þ, the curvature  of C multiple integrals. Note that the point p does not necessarily
is G-invariant. lie on the contour C, as long as there is an unequivocal way
The profiles of the curvature for the rectangular shape in of associating p 2 IR2 to C (e.g., the centroid of the curve).
Fig. 2a and its noisy version in Fig. 2b are shown in Figs. 2c and Note that a regularized version of curvature, or, in general,
2d, respectively. The curvature is called differential invariant a curvature scale-space, can be interpreted as an integral
since its calculation is based on differential operations. The
2. While not discussed in this paper, vector-valued invariants could be
curvature is a useful feature for describing shapes at composed of, e.g., integral invariants at multiple scales, invariants of
matching due to its invariant property under a group increasing derivative or integral order, or even unrelated invariants.

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
1606 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 28, NO. 10, OCTOBER 2006

Fig. 4. (a) A rectangular shape with two mark points  and 4. (b) Local
Distance Integral Invariant of (a) and corresponding mark points, which
have the same invariant value even though they have very different
shapes (i.e., a corner and a straight line). (c) Local Area Integral
Invariant of (a) and corresponding mark points.

A version of the invariant IC that preserves locality can

be obtained by weighting the integral in (4) with a kernel
: R
qðp; xÞ, so that IC ðpÞ¼ C kðp; xÞdsðxÞ where
:
kðp; xÞ¼qðp; xÞdðp; xÞ: ð7Þ
The kernel qð; Þ is free for the designer to choose depending
Fig. 3. (a) Distance integral invariant defined in (6), made local by means of
on the final goal. This local integral invariant can be thought
a kernel as described in (7). (b) Integral local area invariant defined by (9).
of as a continuous generalization of “shape context,” which
invariant since regularized curvature is an algebraic function was designed for discretized shapes represented as points [6].
of the first and second-regularized derivatives [54]. There- The difference is that the shape context signature is a local
fore, integral invariants are quite general and contain radial histogram of neighboring points, whereas in our case,
regularized differential invariants as a subset. However, the we only store the mean of their distance. This allows
spirit of this manuscript is to avoid the computation of extension to continuous representations of shapes and
derivatives of contour data, so we will not explore this subset. obviates the need to choose histogram granularity, etc.
The local distance integral invariant is a local descriptor
Example 2 (Distance integral invariant). Consider G ¼
SEð2Þ and the following function, computed at every provided by the integral kernel restricted on a circular
point p 2 C: neighborhood. It is also robust to noise as shown in Figs. 2g
Z and 2h. Thus, it may be effective for both noise and occlusion.
: However, this invariant is not discriminative in that it can
IC ðpÞ¼ jp  xjdsðxÞ; ð6Þ
C have the same value for different geometric features. This
where jy  xj is the Euclidean distance in IR2 . This is drawback is demonstrated in Fig. 4. The two points marked
illustrated in Fig. 3a. by  and 4 on different geometric features of the shape in
It is immediate to show that this is an integral Fig. 4a have the same local distance integral invariant as
Euclidean invariant, since Euclidean transformations shown in Fig. 4b. This is a motivation to introduce the
preserve distance. We note that, unlike curvature, the following invariant.
range of values for the distance invariant is IRþ (since the
Euclidean distance is always nonnegative). Example 3 (Area integral invariant). Define a ball Br ðpÞ as a
The profiles of the distance integral invariant for the function Br : IR2  IR2 7!f0; 1g to be an indicator function
shapes in Figs. 2a and 2b are shown in Figs. 2e and 2f, on the interior of a circle with radius r centered at p,
respectively. The distance integral invariant is robust to noise, 
1 jp  xj  r
the effect of which is reduced, as shown in Fig. 2f. However, it Br ðp; xÞ ¼ ð8Þ
0 otherwise:
is a global descriptor in that a local change of a shape affects
the values of the distance integral invariant for the entire For any given radius r, the corresponding integral
shape. invariant,

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
MANAY ET AL.: INTEGRAL INVARIANTS FOR SHAPE MATCHING 1607

Fig. 6. Construction of approximation of the area integral invariant.

Fig. 5. Bottom row: Multiscale local area integral invariant for the shapes
(a) The computation of . (b) the approximation of ICr ðpÞ.
(top row). (a) A rectangular shape. (b) A rectangular shape with noise.
Z 3.1 Relation of Local Area Integral Invariant to
:
ICr ðpÞ¼ Br ðp; xÞdx; ð9Þ Curvature
C
In this section, we note a formal connection between
can be thought of as a function from the interval ½0; L to traditional differential invariants and the Local Area
IRþ (since area is always nonnegative), bounded above Integral Invariant in a limiting case. Curvature provides a
by the area of the region bounded by the curve C. This is useful descriptor for shape matching due to its invariance
illustrated in Fig. 3b and examples are shown in Fig. 2. and locality. It is considered a complete invariant in the sense
that it allows the recovery of the original curve up to the
As shown in Figs. 2i and 2j, the local area integral
action of the symmetry group. Furthermore, all differential
invariant is robust to noise and has a locality property
invariants of any order on the plane are functions of
similar to the local distance integral invariant. In addition, it curvature [89] and, therefore, linking our integral invariant
has a strong descriptive power with respect to the marked to curvature would allow us to tap into the rich body of
points on the shape shown in Fig. 4c. Thus, the local area results on differential invariants without suffering from the
integral invariant is an effective descriptor for shape shortcomings of high-order derivatives at its computation.
matching and we rely on a variant of this integral invariant We first assume that the curve C is smooth,3 so that a
throughout this work. notion of curvature is well-defined and the curvature can be
We note again that this integral invariant is a case of approximated locally by the osculating circle shown in Fig. 6.
“shape context,” this time with one histogram bin. In The invariant Ir ðpÞ denotes the area of the intersection of a
addition to allowing extension to continuous representa- circle with radius r with the interior of C, and it can be
tions of shape, local area invariants are explicity rotationally approximated to first-order by the area of the shaded sector in
invariant, while shape context achieves rotational invar- Fig. 6, i.e., Ir ðpÞ ’ 2r2 . Now, the angle  can be computed as a
iance by computing circular shifts of the radial descriptor. function of r and R using the cosine law: cosðÞ ¼ r=2R. Since
Naturally, if we plot the value of ICr ðpðsÞÞ for all values of s curvature  is the inverse of R we have
and r ranging from zero to a maximum radius so that the local  
1
kernel encloses the entire curve Br ðpÞ  C 8p 2 C (at which Ir ðpÞ ’ 2r2 cos1 rðpÞ : ð11Þ
2
point the invariant would be a constant), we can generate a
graph of a function that can be interpreted as a multiscale Now, since cos1 ðxÞ is an invertible function, to the extent in
integral invariant, as shown in Fig. 5. We will return to this which the approximation above is valid (which depends on r),
idea in Section 4.1. Furthermore, Br ðp; xÞ can be substituted we can recover curvature from the integral invariant. The
approximation above is valid in the limit when r ! 0.
by a more general kernel, for instance, a Gaussian centered at
p with  ¼ r.
Note also that the integral invariant can also be defined 4 SHAPE MATCHING AND DISTANCE
as normalized by the area of Br ðpÞ for convenience. Given two shapes represented by curves C1 ; C2 , we want to
Example 4 (Normalized area integral invariant). compute their shape distance, a scalar that quantifies the
R similarity of the two contours. Basing this computation on a
r : R C Br ðp; xÞdx group invariant will ensure that the shape distance is not
IC ðpÞ¼ : ð10Þ
IR2 Br ðp; xÞdx
affected by group actions on the shape; further basing it on an
integral invariant will make the distance computation robust
The corresponding integral invariant is then bounded to noise and local deformations of the contour. Naively, we
between 0 and 1. Because it mimics the qualities of the
3. Notice that our invariant does not require that the shape be smooth,
Area Integral Invariant discussed above, this is the invariant and that this assumption is made only to relate our results to the literature
we favor in the remainder of this work. on differential invariants.

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
1608 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 28, NO. 10, OCTOBER 2006

could define the shape distance to be the difference between

the invariant functions, but upon further reflection we see
that this distance is meaningful only if the computation
somehow compares similar parts of the two shapes. If we
compare one (for example) rabbit’s ears to another’s leg, we
will decide (incorrectly) that the two shapes are very
different. Yet this will be the effect of computing a shape
distance without first establishing a dense correspondence
between the points of the contours. Computing the difference
of invariant values between corresponding points implies
comparing one rabbit’s ears to the second rabbit’s ears, a
much more meaningful metric.
Thus, we wish to find an optimal correspondence between
the contours and concurrently measure the shape distance Fig. 7. Correspondences between two invariant signals I1 ðsÞ (top) and
based on the correspondence. We will express the correspon- I2 ðsÞ (bottom) with different values for the control parameter  in the
dence with a continuous disparity function dðsÞ : SS 7!IR that energy functional. Smaller values of the parameter  in (15) will facilitate
reparameterizes two curves C1 and C2 (and their integral contour shrinking and stretching in the matching process. (a)  ¼ 100,
invariants I1 and I2 ). Given two reparameterizations h1 ; h2 : (b)  ¼ 30, and (c)  ¼ 1.
SS 7!SS where C1 ðh1 ðsÞÞ C2 ðh2 ðsÞÞ ( denotes correspon-
dence), the disparity function is The second term E2 of the energy functional is the elastic
energy of the disparity function dðsÞ that penalizes stretching
h1 ðsÞ  h2 ðsÞ
dðsÞ ¼ : ð12Þ or shrinking of the mapped curve length. When dðsÞ ¼ 0, the
2
parameterizations of the two contours instruct the matching
We require that h1 ; h2 are nondecreasing functions of s, directly (i.e., points on the contour with the same parameter
from which we derive 1  d0 ðsÞ  1. Below, we define an
value correspond). dðsÞ such that d0 ðsÞ ¼ 0 indicates circular
energy functional Eð. . . ; dÞ; the disparity function d
that
minimizes an energy functional “shifts” of the correspondence. Other values of dðsÞ “stretch”
or “shrink” the length of segments of one contour onto the
d
ðsÞ ¼ arg min Eð. . . ; dÞ ð13Þ other; it is this action of dðsÞ that the E2 energy term penalizes.
dðsÞ
To demonstrate the effect of the control parameter  in the
describes the optimal point correspondence between the energy functional, one example of the optimal correspon-
two curves,
dence between two integral invariants with various values of
C1 ðs  d
ðsÞÞ C2 ðs þ d
ðsÞÞ; 8s 2 SS: ð14Þ the control parameter is shown in Fig. 7. One integral
Intuitively, two corresponding points on two contours invariant is represented by a straight line shown on the
should have similar invariant values, which leads us to bottom and the other integral invariant is represented by a
define the energy functional EðI1 ; I2 ; dÞ for the discrepancy line with a spike shown on the top in each figure. The larger
between two integral invariants I1 ; I2 , in terms of the the control parameter , the more correspondence is
disparity function dðsÞ, as follows: regularized, as shown in Fig. 7a. Fig. 7c shows that a feature
EðI1 ; I2 ; dÞ ¼ E1 ðI1 ; I2 ; dÞ þ E2 ðd0 Þ characterized by the spiculation in one integral invariant on
Z 1 the top is mapped to an infinitesimal portion in the other
¼ kI1 ðs  dðsÞÞ  I2 ðs þ dðsÞÞk2 ð15Þ integral invariant on the bottom. The difference of geome-
0
trical features is emphasized more with a small .
þ kd0 ðsÞk2 ds; Ultimately, a notion of shape distance should be
where  > 0 is a constant. The first term E1 of the energy symmetric. It is generally undesirable to privilege one
functional measures the similarity of two curves by integrat- shape rather than the other when matching two shapes. The
ing the local difference of the integral invariant at correspond- energy functional defined in (15) is designed to satisfy a
ing points. A cost functional based on a local comparisons symmetry property that gives
minimizes the impact of articulations and local changes of a
contour because the difference in invariants is proportionally d
ðsÞ ¼ arg min EðI1 ; I2 ; dÞ () d
ðsÞ ¼ arg min EðI2 ; I1 ; dÞ
dðsÞ dðsÞ
localized in the domain of the integral; contrast this with a

global descriptor where local changes influence the descriptor EðI1 ; I2 ; d Þ ¼ EðI2 ; I1 ; d Þ:
everywhere. Representing the correspondence with dðsÞ ð16Þ
ensures that the E1 term is both symmetric (we make this
precise below) and independent of the arclength of dðsÞ.4 The shape distance DðC1 ; C2 Þ between two curves C1 ; C2 is
measured via the optimal correspondence d
ðsÞ in the
4. Correspondence representations such as the “warping function”
introduced in Section 5 implicitly couple the E1 term to the corresponden-
energy functional E between their integral invariants I1 ; I2
ce’s arclength and, thus, require normalization terms. We favor separating as defined by
the energy functional into an arclength-free penalty computed on
corresponding points in the E1 term, and an explicit warping penalty in DðC1 ; C2 Þ ¼ EðI1 ; I2 ; d
Þ: ð17Þ
the E2 term.

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
MANAY ET AL.: INTEGRAL INVARIANTS FOR SHAPE MATCHING 1609

Fig. 8. Scalogram of the shape in Fig. 2a and trace of local extrema

across scales. (a) Scalogram of the curvature scale-space. Note the
dislocation of the extrema. (b) Scalogram of the integral invariant
scale-space.

Fig. 10. Demonstration of correspondences between two rectangular

shapes with spikes at different scales. The figures show the optimal point
correspondence determined by our algorithm for increasing size of the
kernel width r in (10). The two spikes are identified as “corresponding” on
a fine scale only. (a) Fine scale, (b) intermediate scale, and (c) coarse
scale.

versions of the shape. This obviates the need to compute

Fig. 9. Integral Invariant for a noisy shape computed at different scales. correspondences between scales.
R ¼ 120 for this shape. A multiscale integral invariant, demonstrated in Fig. 9,
could also be used in a hierarchical description of features.
Since the energy functional is symmetric (up to the sign of The matching using the integral invariants at a fine scale
d
), the shape distance is as well. In Section 5, we outline the provides a correspondence taking into account detail features
computation of d
and DðC1 ; C2 Þ. on the shapes while matching at a coarse scale considers large
features on the shapes. This is demonstrated in Fig. 10 on the
4.1 Shape Matching with Multiscale Integral fine scale peak, which in accounted for in the fine-scale
Invariants matching but ignored in the coarser scale matchings.
The integral invariant intrinsically introduces the notion of With any method that allows a choice of scale, the question
scale, varying the size of the kernel naturally forms a of which scale is optimal for shape analysis arises. While the
multiscale invariant. While we do not exploit the multiscale choice of an optimal scale is a subject of continuing
properties of these invariants or multiresolution methods in investigation, we offer the following comment. In Section 3,
this work, this section highlights some of the features and we noted that the kernel radius of the Local Area Integral
potential of the integral invariant framework. Invariant is bounded above by a value R, for which the kernel
In [53], a curvature scale-space is used in a coarse-to-fine encompasses the entire shape. At this scale, the value of the
procedure to establish a sparse matching between inflection invariant is constant. So, rather than express scale in absolute
points of two shapes. Matches at a given scale are computed terms parameterized by the kernel radius r, scale can also be
from matches at coarser scales. However, mismatching in the parameterized in relative terms measured by the normalized
first stage causes fatal errors. Further, since curvature scale- kernel radius r=R, as in Fig. 9. This provides a common context
space is derived from Gaussian smoothing, the inflection to parameterize scale even among shapes of different sizes.
points move with increased blurring, and reparameteriza-
tion is required to find correspondence between these 5 IMPLEMENTATION
inflection points across scale. In Fig. 8, the curvature scale- In Section 4, we presented a distance between invariants (and
space and a multiscale integral invariant (more specifically, therefore shape) that depends on the choice of a disparity
the local area invariant with varying kernel radius) for the function dðsÞ. To complete the calculation of distance and to
shape in Fig. 2a are compared. The dislocation of the extrema establish a local correspondence between the curves, we must
points occurs across scales in the curvature scale-space, as optimize distance with respect to dðsÞ. This section briefly
shown in Fig. 8a. In contrast, the location of the extrema outlines the implementation of the computation of the local
points stays the same across scales in the multiscale integral area integral invariant and a well-known approach to
invariant, as in Fig. 8b since features at various scales are globally optimize the correspondence for a discrete repre-
observed based on the original shape rather than on blurred sentation of the curves as ordered sets of points.

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
1610 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 28, NO. 10, OCTOBER 2006

Fig. 12. Schematic view of a graph used to compute the correspondence

for two curves with M ¼ N ¼ 5. The dots are the nodes, and the arrows
are the directed edges. The dashed arrows are the “wrap around”
edges. A typical portion of this graph is shown enlarged in Fig. 13.
Fig. 11. Optimal path through the graph. The path warps the
parameterization of the hand (on the bottom of the graph) and the
parameterization of the noisy occluded (four-fingered) hand (on the left Differentiating (18) with respect to s yields
side of the graph). Both shapes are shown in Fig. 1. The gray levels
indicate the dissimilarity between points; lighter shade indicates higher d
dissimilarity. See text for more details. 1  d0 ðsÞ ¼ h01 ðÞ
ds ð20Þ
0 0 d
To efficiently compute the local area integral invariant, 1 þ d ðsÞ ¼ h2 ðÞ :
ds
consider the binary image ðCÞ  (an indicator function defined
Then, the original energy functional in (15) becomes
as 1 on the interior of the contour and 0 elsewhere) and
:
convolve it with the kernel kðp; xÞ¼Br ðp  xÞ, where p 2 IR2 , e 1 ; I2 ; h1 ; h2 Þ ¼
EðI
not just the curve C. Evaluating the result of this convolution Z Lh  h0 ðÞ þ h0 ðÞ
2 1 2
on p 2 C yields ICr , without the need to parameterize the kI1 ðh1 ðÞÞ  I2 ðh2 ðÞÞk d
0 2 ð21Þ
curve. However, we retain a parametrized representation of Z Lh   !  0 
h02 ðÞ  h01 ðÞ2 h1 ðÞ þ h02 ðÞ
the curve for the computation of the correspondence. þ   d:
h0 ðÞ þ h0 ðÞ 2
Our implementation is based on dynamic programming 0 1 2
approaches similar to those employed by many in the In this way, the warping function hðÞ derives a formula for
shape, stereo, and registration (for medical imaging) the energy functional in terms of ðh1 ðÞ; h2 ðÞÞ,
communities [70], [65], [76], [30], [41], [84], [4], [85].
In order to adopt a graph search framework, the e 1 ; I2 ; h1 ; h2 Þ:
h : EðI1 ; I2 ; dÞ ! EðI ð22Þ
representation of the correspondence needs to be recast as Then, finding an optimal disparity function d ðsÞ in the

a parameterized path. Thus, we exchange the disparity energy functional E becomes equivalent to finding an
function dðsÞ for a warping function,5 hðÞ ¼ ðh1 ðÞ; h2 ðÞÞ for optimal warping function h
ðÞ ¼ ðh
1 ðÞ; h
2 ðÞÞ in the
curves ðC1 ðsÞ; C2 ðsÞÞ in the energy functional defined in (15) e as follows:
energy functional E
by setting
d
ðsÞ ¼ arg min EðI1 ; I2 ; dÞ
dðsÞ
s  dðsÞ ¼ h1 ðÞ
ð18Þ m ð23Þ
s þ dðsÞ ¼ h2 ðÞ:
ðh
1 ðÞ; h
2 ðÞÞ e 1 ; I2 ; h1 ; h2 Þ:
¼ arg min EðI
Since hi : ½0; Lh 7!SS1 for i 2 f1; 2g (with hi ð0Þ ¼ hi ðLh Þ), the h1 ðÞ;h2 ðÞ

warping function h : ½0; Lh 7!SS1  SS1 is a path in SS1  SS1 , To exploit the dynamic programming framework, we
parameterized by its length Lh , that represents the pointwise must discretize the curve by sampling it at uniform
correspondence between the curves C1 ðsÞ; C2 ðsÞ: intervals. The result is an ordered set of points.
In the discrete case, an intuitive algorithm to compute
C1 ðh1 ðÞÞ C2 ðh2 ðÞÞ; 8 2 ½0; Lh   IR: ð19Þ shape matching would be as follows: We first find an initial
correspondence between a point on each curve, say C1 ½i
5. The representation of the reparameterization of two curves to achieve C2 ½j (more on this below). The “next” correspondence should
pointwise correspondence has been approached in several different ways. be the choice of action that minimizes the energy ((15)); the
For example, in [76], symmetric and asymmetric “alignment functions” are possible actions are 1) locally contracting the first curve onto
used. All of these methods are somewhat similar, with one main distinction:
symmetric representations result in symmetric energy and distance the second, setting C1 ½i þ 1 C2 ½j, 2) locally contracting the
functions, while asymmetric representations do not. Due to our focus on second curve onto the first, setting C1 ½i C2 ½j þ 1, or
shape distance, we restrict ourselves to symmetric reparameterizations: the 3) locally mapping the discrete points as one-to-one, setting
disparity function because it results in an energy functional that does not
depend implicitly on pathlength, and the warping function for its simplicity C1 ½i þ 1 C2 ½j þ 1 (Fig. 13). By induction we can now
to implement in a discrete dynamic programming framework. compute correspondences for every point on the curves. This

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
MANAY ET AL.: INTEGRAL INVARIANTS FOR SHAPE MATCHING 1611

Fig. 13. (a) Portions of two curves with possible pointwise correspondences shown (and labeled with graph notation, e.g., vði; jÞ , C1 ½i C2 ½j).
(b) Typical portion of graph in Fig. 12 showing “current” correspondence node vði; jÞ and the three possible “next” nodes.

sketch of the algorithm lends itself to a graph formulation, based on a weighting function wðvt ; vtþ1 Þ adapted from (15),
where each node of a directed graph is a correspondence defined by
between a point on each curve, and each edge represents one
ðh0 þ h02 Þ
of the possible actions, linking the current node to the possible wðvði; jÞ; vðk; lÞÞ ¼ kI1 ðks1 Þ  I2 ðls2 Þk2 1
“next” nodes. Assuming C1 and C2 are sampled with N and 2
 0  ð27Þ
M points, respectively, the graph is a regular grid with h2  h01 2 ðh01 þ h02 Þ
þ h0 þ h0 
 :
NM nodes and 3NM edges. Because the curves are defined 1 2 2
on a periodic domain, the graph has edges that “wrap
The regular nature of the graph allows us to simplify the
around” from the top of the grid to the bottom, and from the
computations of h01;2 :
right side of the grid to the left, as shown in Fig. 12.6 The edges
are weighted by the distance between the invariants s
h01 ¼s1 ¼1; h02 ¼s0 ¼0; if k¼mod ði;MÞþ1; l¼j;
1 2
associated with the “next” node, cf., (15). We formalize and h01 ¼s0 ¼0;
s
h02 ¼s2 ¼1; if k¼i; l¼modðj;NÞþ1;
1 2
extend this concept in the remainder of this section. h01 ¼pffiffiffiffiffiffiffiffiffiffiffi
1 s
; h02 ¼pffiffiffiffiffiffiffiffiffiffiffi
2s
; if k¼mod ði;MÞþ1; l¼mod ðj;NÞþ1:
s2 þs2 s2 þs2
We sample the curves C1 ; C2 on the discrete domains 1 2 1 2

1 ; 2 with equal spacing as follows: ð28Þ

1 ¼½0; s1 ; 2s1 ; . . . ; Ms1 ¼ 1; The direction of edges in the graph is constrained so that
the warping function hðÞ is monotonic. The monotonicity of
1 ð24Þ
s1 ¼ ; M 2 INþ ; the warping function prevents cross correspondence that
M causes a topological change in matching. (This is equivalent to

2 ¼½0; s2 ; 2s2 ; . . . ; Ns2 ¼ 1;

1 ð25Þ
s2 ¼ ; N 2 INþ :
N
The weighted, directed graph G ¼ ðV ; EÞ is formed based on
a grid structure of the discrete domain  ¼ 1  2 , as shown
in Fig. 12. Each node vði; jÞ 2 V in the graph represents a
pointwise correspondence C1 ðis1 Þ C2 ðjs2 Þ, where i 2
½0; M  INþ and j 2 ½0; N  INþ . The adjacency relation of
nodes is defined by an edge eðvði; jÞ; vðk; lÞÞ that represents a
directed relation vði; jÞ ! vðk; lÞ indicating the following
correspondence vðk; lÞ given the current correspondence
vði; jÞ.
The minimization of the energy functional E e is equiva-
lent to finding a shortest path p ¼< v0 ; v1 ; v2 ; . . . ; vL > that Fig. 14. Shape correspondence with increasing noise perturbation.
gives a minimum weight from v0 ¼ vð0; 0Þ to vL ¼ vðM; NÞ, (a) Correspondence of Shape 24 to noisy instances of Shape 20
computed, (b) via integral invariants, and (c) via differential invariants.
X
L1 Some salient points on the contour are labeled on Shape 24, and the
wðpÞ ¼ wðvt ; vtþ1 Þ ð26Þ corresponding points are labeled on the remaining contours. Since the
t¼0 integral invariant is more robust to noise than the differential one, it
consistently identifies the corresponding parts, even for contours which
are strongly perturbed by noise. For the differential invariant, on the
6. The graph domain can also be considered to be embedded in a two- other hand, the algorithm fails to capture the correct correspondence
dimensional torus. when   2:5.

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
1612 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 28, NO. 10, OCTOBER 2006

Fig. 15. Shape distance as a function of noise for the shapes in Fig. 14. Fig. 17. Histograms of shape distance between Shape 24 and
While the shape distance measure based on differential invariants 1,000 perturbations of Shape 20 with noise at variance  ¼ 2:5. The
strongly varies with noise, the distance based on integral invariants is dark bars on the left represent shape distances computed with integral
much more insensitive to noise. (See text for details.) invariants; the lighter bars on the right represent shape distances
computed via differential invariants. At a fixed noise variance, the
computed shape distance based on integral invariants remains
essentially constant over all trials, in contrast to the distance based on
differential invariants. (To scale the histograms for presentation, two
experiments were removed from the differential invariant trials, with
shape distances 2.8 and 9.7.)

Fig. 16. Shape correspondence for several perturbations (with noise of

scale  ¼ 2:5). (a) Correspondence of Shape 24 to noisy instances of
Shape 20 computed, (b) via integral invariants and (c) via differential
invariants. Some salient points on the contour are labeled on Shape 24,
and the corresponding points are labeled on the remaining contours. In
contrast to the distance based on differential invariants, the integral
invariant distance consistently provides the correct correspondence.

enforcing the constraint in Section 4 that 1  dðsÞ  1.)

Further, each point in each curve must have at least one
corresponding point in the other curve. These constraints are
implied by the existence of an edge between nodes vði; jÞÞ and
vðk; lÞ only if i 2 fmodðk; MÞ þ 1g and j 2 fl; modðl; NÞ þ 1g.
Dijkstra’s algorithm is used for finding a shortest path
from a single source node to a single destination node in a
graph. Let p be a sequence for the shortest path in the graph
G ¼ ðV ; EÞ such as

p ¼< v0 ; v1 ; v2 ; . . . ; vP > Fig. 18. Shapes used in matching experiments.

ð29Þ
¼< vði0 ; j0 Þ; vði1 ; j1 Þ; vði2 ; j2 Þ; . . . ; vðiP ; jP Þ >;
No fast algorithm exists to determine the best choice of the
where P is the number of nodes in the path p, and v0 ¼ vP . initial correspondence. Previous implementations (cited
Then, the optimal warping function hðÞ ¼ ðh1 ðÞ; h2 ðÞÞ is above) choose a fixed a point on the first curve and pair it
given by with all possible choices of points on the second curve,
calculating the path for each pair to determine the shortest.
h1 ðÞ ¼ ði0 s1 ; i1 s1 ; i2 s1 ; . . . ; iP 1 s1 Þ
ð30Þ (This process can be thought of as exhaustively searching
h2 ðÞ ¼ ðj0 s2 ; j1 s2 ; j2 s2 ; . . . ; jP 1 s2 Þ: among all possible values of dð0Þ or h1 ð0Þ.) Searching near an
An example of the result of this algorithm is shown in Fig. 11. initial point estimated as in [51] will obviate the need for an
Finally, the shape distance is computed via (26). exhaustive search. Alternately, the exhaustive search can be

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
MANAY ET AL.: INTEGRAL INVARIANTS FOR SHAPE MATCHING 1613

Fig. 19. Noisy shape recognition from a database of 24 shapes. The upper number in each cell is the distance computed via the local-area integral
invariant; the lower number is the distance computed via curvature invariant. The bold, italic number in each row represents the best match for the
noisy shape at the left of that row; the four remaining italic numbers represent the next four best matches. See text for more details.

avoided by observing that strong features, such as corners or correspondence and compute the warping function h and its
convex/concave points, provide a heuristic way to propose associated distance for two curves with 100 points each in less
point correspondences. These points are easily classified in than 1 second using MATLAB on a computer with an Intel
the invariant space. For instance, for the local area invariant, 800 MHz processor.
points with little or no curvature are in a ball around I ¼ :5. We
find a subset of points outside this ball, fsj jI1 ðsÞ  :5j > T g,
where T is some threshold (typically T ¼ :1). These points,
6 EXPERIMENTAL RESULTS
with their nearest neighbors on I2 , form a set of likely initial This section presents experiments that show that the locality
correspondences. In this way, we can find an initial and noise robustness properties of the integral invariant

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
1614 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 28, NO. 10, OCTOBER 2006

Fig. 20. Shape distance between noisy shapes (across top) and original shapes (along left side) via integral invariant. Lighter shade indicates higher
distance. See text for more details.

result in a shape description that is less sensitive to occlusions Note that the distance computed via the integral invariant
or localized deformations, when compared to similarly increases as  increases, but, in general, added noise affects
implemented differential invariant methods. We compute the shape distance only slightly. Contrast this with the
the differential invariant using the method outlined in [13], distance computed via curvature, which increases drastically
[8]. We will begin with experiments demonstrating the as a function of  until the curves are so noisy that a
computation of shape distance and correspondences be- meaningful correspondence cannot consistently be com-
tween two shapes before demonstrating the retrieval of puted using differential invariants (e.g., the  ¼ 2:5 column
matches for noisy shapes from a database [76]. of Fig. 14). The drastic decrease in shape distance for values of 
In this section, we use a subset of shapes from the Kimia beyond this “breakdown value” further demonstrates the
silhouette database.7 The contours are extracted from the dangers of relying on differential invariants; even though the
silhouettes using the pixel coordinates as a reference frame. distance value indicates that this correspondence is optimal,
The result is that the x and y-values of the contours are in the correspondence is subjectively incorrect.
½0; 99. Empirically, we sample the contours at 100 evenly Fig. 16 again shows the noise robustness of the integral
spaced points, and set r ¼ 15 and  ¼ :01. invariant, compared to differential invariants, when used for
Fig. 14 shows the shape matching induced via the local- shape matching. The variance of the noise is held constant at
area integral invariant and via curvature between two  ¼ 2:5, however, the experiment is repeated 1,000 times.
different bunnies8 despite increasing noise. Noise is added Shape matching via the integral invariant provides a
by perturbing all points on the contour in the normal direction consistent correspondence (as shown with the labeled
by a distance drawn from a zero-mean Gaussian random features in the second row) and a consistent shape distance
variable with specified . We indicate the correspondence by shown in the histogram in Fig. 17. Computation of shape
matching and shape distance via curvature results in a
showing the mapping of the numbered landmarks onto the
correspondence that varies with the noise, as shown in the
noisy shape, although we emphasize that invariant values
third row, and a more erratic shape distance, shown in Fig. 17.
from everywhere on the curve, and not just at feature points,
In Fig. 19, the results of matching and retrieving noisy
are used to compute shape matching and distance. Fig. 15 is a
shapes (shown on the left side) from a database (shown across
plot of the shape distance for the matching, shown in Fig. 14.
the top) are shown. (Large displays of the original and noisy
7. Available at http://www.lems.brown.edu/vision/software/index. shapes are shown in Fig. 18.) We especially highlight several
html. pairs where representation by differential invariants leads to
8. Although the two bunnies look similar, closer examination shows that
the noisy bunny has a thicker body and a longer snout, in addition to mismatches, such as the third, fourth, and fifth fish (in the
differences in position. first group). Due to the differential invariant’s sensitivity to

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
MANAY ET AL.: INTEGRAL INVARIANTS FOR SHAPE MATCHING 1615

Fig. 21. Shape distance between noisy shapes (across top) and original shapes (along left side) via differential invariant. Lighter shade indicates
higher distance.

noise, these fish have a smaller shape distance to the rabbits, occlusions and configuration changes. We embed both of
where the shape distance based on integral invariants orders these concepts in a formulation of a shape distance, and
the shapes correctly (fish are closer to themselves than to outline how distance and optimal correspondence are jointly
rabbits). Examination of the data shows several such cases computed via efficient dynamic programming algorithms.
where integral invariants are more robust than curvature on On an experimental level, we demonstrate robustness of
noisy shapes. the integral invariant distances for shape matching and
Fig. 20 and Fig. 21 show this same data in a more aggregate identifying corresponding shape parts under perturbation in
way, using shades of gray to indicate distance. The distance increasing amounts of noise.
matrix computed using integral invariants (Fig. 20) has low
distances on the diagonal and the block-diagonal structure, as
expected in a database with grouped classes. Contrast this
ACKNOWLEDGMENTS
with the curvature-based distance matrix (Fig. 21) which This work was performed under the auspices of the US
lacks clearly lower distance on the diagonal and has a Department of Energy by the University of California
vertically banded structure opposed to the desired block- Lawrence Livermore National Laboratory under contract
diagonal structure, indicating that the added noise influ- No. W-7405-Eng-48. This research was supported by grants
enced the shape distance more than the shape itself. NIH U54 RR021813, AFOSR F49620-03-1-0095/E-16-V91-G2,
and ONR N00014-03-1-0850/N00014-02-1-0720.

7 DISCUSSION AND CONCLUSIONS

In this paper, we address one of the key disadvantages of
REFERENCES
[1] R. Alferez and Y.F. Wang, “Geometric and Illumination Invariants
differential invariants for shape matching—namely, their
for Object Recognition,” IEEE Trans. Pattern Analysis and Machine
inherent sensitivity to noise. We introduce a new distance for Intelligence, vol. 21, no. 6, pp. 505-536, June 1999.
2D shapes which is based on the notion of integral group- [2] K. Arbter, W.E. Snyder, H. Burkhardt, and G. Hirzinger,
invariant descriptions of shape. Both theoretically and “Applications of Affine-Invariant Fourier Descriptors to Recogni-
tion of 3-D Objects,” IEEE Trans. Pattern Analysis and Machine
experimentally we relate these integral invariant shape Intelligence, vol. 12, no. 7, pp. 640-646, July 1990.
distances to previously proposed shape distances which are [3] N. Ayache and O. Faugeras, “HYPER: A New Approach for the
based on differential invariants. Recognition and Positioning of Two-Dimensional Objects,” IEEE
While integral invariants are employed for robustness to Trans. Pattern Analysis and Machine Intelligence, vol. 8, no. 1, pp. 44-
54, 1986.
high-frequency noise and small deformations, shape warp- [4] M. Bakircioglu, U. Grenander, N. Khaneja, and M.I. Miller, “Curve
ing by the computation of an optimal reparameterization Matching on Brain Surfaces Using Frenet Distances,” Human Brain
allows us to account for large localized changes such as Mapping, vol. 6, pp. 329-333, 1998.

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
1616 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 28, NO. 10, OCTOBER 2006

[5] R. Basri, L. Costa, D. Geiger, and D. Jacobs, “Determining the [31] L. Van Gool, T. Moons, E. Pauwels, and A. Oosterlinck, “Semi-
Similarity of Deformable Shapes,” Vision Research, vol. 38, Differential Invariants,” Geometric Invariance in Computer Vision,
pp. 2365-2385, 1998. pp. 193-214, 1992.
[6] S. Belongie, J. Malik, and J. Puzicha, “Shape Matching and Object [32] L. Van Gool, T. Moons, and D. Ungureanu, “Affine/Photometric
Recognition Using Shape Contexts,” IEEE Trans. Pattern Analysis Invariants for Planar Intensity Patterns,” Proc. European Conf.
and Machine Intelligence, vol. 24, no. 4, pp. 509-522, Apr. 2002. Computer Vision, vol. 1, pp. 642-651, 1996.
[7] A. Bengtsson and J.-O. Eklundh, “Shape Representation by [33] J.H. Grace and A. Young, The Algebra of Invariants. Cambridge
Multiscale Contour Approximation,” IEEE Trans. Pattern Analysis Univ. Press, 1903.
and Machine Intelligence, vol. 13, no. 1, pp. 85-93, Jan. 1991. [34] C.E. Hann and M.S. Hickman, “Projective Curvature and Integral
[8] M. Boutin, “Numerically Invariant Signature Curves,” Int’l J. Invariants,” Int’l J. Computer Vision, vol. 40, no. 3, pp. 235-248, 2000.
Computer Vision, vol. 40, no. 3, pp. 235-248, 2000. [35] M.K. Hu, “Visual Pattern Recognition by Moment Invariants,” IRE
[9] R.D. Brandt and F. Lin, “Representations that Uniquely Char- Trans. IT, vol. 8, pp. 179-187, 1961.
acterize Images Modulo Translation, Rotation and Scaling,” [36] K. Kanatani, Group Theoretical Methods in Image Understanding.
Pattern Recognition Letters, vol. 17, pp. 1001-1015, 1996. Springer, 1990.
[10] A. Bruckstein, N. Katzir, M. Lindenbaum, and M. Porat, [37] E. Klassen, A. Srivastava, M. Mio, and S.H. Joshi, “Analysis of
“Similarity Invariant Signatures for Partially Occluded Planar Planar Shapes Using Geodesic Paths on Shape Spaces,” IEEE Trans.
Shapes,” Int’l J. Computer Vision, vol. 7, no. 3, pp. 271-285, 1992. Pattern Analysis and Machine Intelligence, vol. 26, no. 3, pp. 372-383,
[11] A.M. Bruckstein, R.J. Holt, A.N. Netravali, and T.J. Richardson, Mar. 2004.
“Invariant Signatures for Planar Shape Recognition under Partial [38] P.N. Klein, S. Tirthapura, D. Sharvit, and B. Kimia, “A Tree-Edit-
Occlusion,” J. Computer Vision, Graphics, and Image Processing, Distance Algorithm for Comparing Simple, Closed Shapes,” Proc.
vol. 58, no. 1, pp. 49-65, 1993. Symp. Discrete Algorithms, pp. 696-704, 2000.
[12] A.M. Bruckstein, E. Rivlin, and I. Weiss, “Scale-Space Semi-Local [39] E.P. Lane, Projective Differential Geometry of Curves and Surfaces.
Invariants,” Image and Vision Computing, vol. 15, no. 5, pp. 335-344, Univ. of Chicago Press, 1932.
1997. [40] J. Lasenby, E. Bayro-Corrochano, A.N. Lasenby, and G. Sommer,
[13] E. Calabi, P. Olver, C. Shakiban, A. Tannenbaum, and S. Haker, “A New Framework for the Formation of Invariants and Multiple-
“Differential and Numerically Invariant Signature Curves Ap- View Constraints in Computer Vision,” Proc. Int’l Conf. Image
plied to Object Recognition,” Int’l J. Computer Vision, vol. 26, Processing, pp. 313-316, 1996.
pp. 107-135, 1998. [41] L.J. Latecki and R. Lakämper, “Shape Similarity Measure Based on
[14] D. Chetverikov and Y. Khenokh, “Matching for Shape Defect Correspondence of Visual Parts,” IEEE Trans. Pattern Analysis and
Detection,” Lecture Notes in Computer Science, vol. 1689, no. 2, Machine Intelligence, vol. 22, no. 10, pp. 1185-1190, Oct. 2000.
pp. 367-374, 1999.
[42] H. Le and D.G. Kendall, “The Riemannian Structure of Euclidean
[15] I. Cohen, N. Ayache, and P. Sulger, “Tracking Points on Shape Spaces: A Novel Environment for Statistics,” Annals of
Deformable Objects Using Curvature Information,” Proc. European Statistics, vol. 21, no. 3, pp. 1225-1271, 1993.
Conf. Computer Vision, pp. 458-466, 1992.
[43] G. Lei, “Recognition of Planar Objects in 3-D Space from Single
[16] T. Cohignac, C. Lopez, and J.M. Morel, “Integral and Local Affine Perspective Views Using Cross Ratio,” Robotics and Automation,
Invariant Parameter and Application to Shape Recognition,” Proc.
vol. 6, no. 4, pp. 432-437, 1990.
Int’l Conf. Pattern Recognition, vol. 1, pp. 164-168, 1994.
[44] R. Lenz, “Group Theoretical Methods in Image Processing,”
[17] J.B. Cole, H. Murase, and S. Naito, “A Lie Group Theoretical
Lecture Notes in Computer Science, vol. 413, 1990.
Approach to the Invariance Problem in Feature Extraction and
Object Recognition,” Pattern Recognition Letters, vol. 12, pp. 519- [45] M.E. Leventon, W.E.L. Grimson, and O. Faugeras, “Statistical
523, 1991. Shape Influence in Geodesic Active Contours,” Proc. Conf. Computer
Vision and Pattern Recognition, vol. 1, pp. 316-323, June 2000.
[18] D. Cremers, T. Kohlberger, and C. Schnörr, “Shape Statistics in
Kernel Space for Variational Image Segmentation,” Pattern [46] S.Z. Li, “Shape Matching Based on Invariants,” Progress in Neural
Recognition, vol. 36, no. 9, pp. 1929-1943, 2003. Networks: Shape Recognition, vol. 6, pp. 203-228, 1999.
[19] D. Cremers, S.J. Osher, and S. Soatto, “Kernel Density Estimation [47] S. Liao and M. Pawlak, “On Image Analysis by Moments,” IEEE
and Intrinsic Alignment for Knowledge-Driven Segmentation: Trans. Pattern Analysis and Machine Intelligence, vol. 18, no. 3,
Teaching Level Sets to Walk,” Pattern Recognition, Sept. 2004. pp. 254-266, Mar. 1996.
[20] D. Cremers and S. Soatto, “A Pseudo-Distance for Shape Priors in [48] H. Liu and M. Srinath, “Partial Shape Classification Using Contour
Level Set Segmentation,” Proc. IEEE Second Int’l Workshop Matching in Distance Transforms,” IEEE Trans. Pattern Analysis and
Variational, Geometric, and Level Set Methods, pp. 169-176, 2003. Machine Intelligence, vol. 12, no. 2, pp. 1072-1079, Feb. 1990.
[21] R. Davies, C. Twining, T. Cootes, J Waterton, and C Taylor, “A [49] T. Liu and D. Geiger, “Approximate Tree Matching and Shape
Minimum Description Length Approach to Statistical Shape Similarity,” Proc. Int’l Conf. Computer Vision, pp. 456-462, 1999.
Modeling,” IEEE Trans. Medical Imaging, vol. 21, no. 5, pp. 525- [50] S. Manay, B. Hong, A. Yezzi, and S. Soatto, “Integral Invariant
537, 2002. Signatures,” Proc. European Conf. Computer Vision, May 2004.
[22] A. DelBimbo and P. Pala, “Visual Image Retrieval by Elastic [51] J.S. Marques and A.J. Abrantes, “Shape Alignment—Optimal
Matching of User Sketches,” IEEE Trans. Pattern Analysis and Initial Point and Pose Estimation,” Pattern Recognition Letters,
Machine Intelligence, vol. 19, no. 2, pp. 121-132, Feb. 1997. vol. 18, no. 1, pp. 49-53, 1997.
[23] A. Dervieux and F. Thomasset, “A Finite Element Method for the [52] T. Miyatake, T. Matsuyama, and M. Nagao, “Affine Transform
Simulation of Raleigh-Taylor Instability,” Springer Lecture Notes in Invariant Curve Recognition Using Fourier Descriptors,” Informa-
Math., vol. 771, pp. 145-158, 1979. tion Processing Soc. Japan, vol. 24, no. 1, pp. 64-71, 1983.
[24] L.E. Dickson, Algebraic Invariants. John Wiley and Sons, 1914. [53] F. Mokhtarian and A.K. Mackworth, “Scale-Based Description and
[25] J. Dieudonne and J. Carrell, Invariant Theory: Old and New. Recognition of Planar Curves and Two-Dimensional Shapes,”
Academic Press, 1970. IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 8, no. 1,
[26] I.L. Dryden and K.V. Mardia, Statistical Shape Analysis. Wiley, 1998. pp. 34-43, Jan. 1986.
[27] J. Flusser and T. Suk, “Pattern Recognition by Affine Moment [54] F. Mokhtarian and A.K. Mackworth, “A Theory of MultiScale,
Invariants,” Pattern Recognition, vol. 26, no. 1, pp. 167-174, 1993. Curvature-Based Shape Representation for Planar Curves,” IEEE
[28] D.A. Forsyth, J.L. Mundy, A. Zisserman, and C.M. Brown, Trans. Pattern Analysis and Machine Intelligence, vol. 14, no. 8,
“Projectively Invariant Representations Using Implicit Algebraic pp. 789-805, Aug. 1992.
Curves,” Image and Vision Computing, vol. 9, no. 2, pp. 130-136, 1991. [55] D. Mumford, “Mathematical Theories of Shape: Do They Model
[29] D.A. Forsyth, J.L. Mundy, A.P. Zisserman, C. Coelho, A. Heller, Perception,” Geometric Methods in Computer Vision, vol. 1570, pp. 2-
and C.A. Othwell, “Invariant Descriptors for 3-D Object Recogni- 10, 1991.
tion and Pose,” IEEE Trans. Pattern Analysis and Machine [56] D. Mumford, J. Fogarty, and F.C. Kirwan, Geometric Invariant
Intelligence, vol. 13, no. 10, pp. 971-991, Oct. 1991. Theory, third ed. Springer-Verlag, 1994.
[30] Y. Gdalyahu and D. Weinshall, “Flexible Syntactic Matching of [57] D. Mumford, A. Latto, and J. Shah, “The Representation of
Curves and Its Application to Automatic Hierarchical Classication Shape,” Proc. IEEE Workshop Computer Vision, pp. 183-191, 1984.
of Silhouettes,” IEEE Trans. Pattern Analysis and Machine Intelli- [58] Geometric Invariance in Computer Vision. J.L. Mundy and A. Zisser-
gence, vol. 21, no. 12, pp. 1312-1328, Dec. 1999. man, eds. MIT Press, 1992.

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
MANAY ET AL.: INTEGRAL INVARIANTS FOR SHAPE MATCHING 1617

[59] L. Nielsen and G. Saprr, “Projective Area-Invariants as an [86] A. Tsai, A. Yezzi, W. Wells, C. Tempany, D. Tucker, A. Fan, E.
Extension of the Cross-Ratio,” J. Computer Vision, Graphics, and Grimson, and A. Willsky, “Model-Based Curve Evolution Tech-
Image Processing, vol. 54, no. 1, pp. 145-159, 1991. nique for Image Segmentation,” Proc. Conf. Computer Vision and
[60] P.J. Olver, Equivalence, Invariants and Symmetry. Cambridge, 1995. Pattern Recognition, pp. 463-468, 2001.
[61] S.J. Osher and J.A. Sethian, “Fronts Propagation with Curvature [87] S. Umeyama, “Parameterized Point Pattern Matching and Its
Dependent Speed: Algorithms Based on Hamilton-Jacobi For- Application to Recognition of Object Families,” IEEE Trans. Pattern
mulations,” J. Computer Physics, vol. 79, pp. 12-49, 1988. Analysis and Machine Intelligence, vol. 15, no. 2, pp. 136-144, Feb. 1993.
[62] T. Pajdla and L. Van Gool, “Matching of 3-D Curves Using Semi- [88] J. Verestoy and D. Chetverikov, “Shape Detection in Ferrite
Differential Invariants,” Proc. Int’l Conf. Computer Vision, pp. 390- Cores,” Machine Graphics and Vision, vol. 6, no. 2, pp. 225-236, 1997.
395, 1995. [89] I. Weiss, “Noise Resistant Invariants of Curves,” IEEE Trans.
[63] M. Pelillo, K. Siddiqi, and S.W. Zucker, “Matching Hierarchical Pattern Analysis and Machine Intelligence, vol. 15, no. 9, pp. 943-948,
Structures Using Association Graphs,” IEEE Trans. Pattern Sept. 1993.
Analysis and Machine Intelligence, vol. 21, no. 11, pp. 1105-1120, [90] A.P. Witkin, “Scale-Space Filtering,” Proc. Int’l Joint Conf. Artificial
Nov. 1999. Intelligence, pp. 1019-1021, 1983.
[64] A. Pikaz and I. Dinstein, “Matching of Partially Occluded Planar [91] H. Wolfson, “On Curve Matching,” IEEE Trans. Pattern Analysis
Curves,” Pattern Recognition, vol. 28, no. 2, pp. 199-209, 1995. and Machine Intelligence, vol. 12, no. 5, pp. 483-489, May 1990.
[65] A. Pitiot, H. Delingette, A. Toga, and P. Thompson, “Learning [92] L. Younes, “Optimal Matching between Shapes via Elastic
Object Correspondences with the Observed Transport Shape Deformations,” Image and Vision Computing, vol. 17, pp. 381-389,
Measure,” Proc. Conf. Information Processing in Medical Imaging, 2003. 1999.
[66] T.H. Reiss, “Recognizing Planar Objects Using Invariant Image [93] C.T. Zahn and R.Z. Roskies, “Fourier Descriptors for Plane Closed
Features,” Lecture Notes in Computer Science, vol. 676, 1993. Curves,” IEEE Trans. Computers, vol. 21, pp. 269-281, 1972.
[67] C. Rothwell, A. Zisserman, D. Forsyth, and J. Mundy, “Canonical [94] S. Zhu and A. Yuille, “Forms: A Flexible Object Recognition and
Frames for Planar Object Recognition,” Proc. European Conf. Modeling System,” Int’l J. Computer Vision, vol. 20, no. 3, pp. 187-
Computer Vision, pp. 757-772, 1992. 212, 1996.
[68] C. Rothwell, A. Zisserman, D. Forsyth, and J. Mundy, “Planar [95] A. Zisserman, D.A. Forsyth, J.L. Mundy, C.A. Rothwell, and J.S.
Object Recognition Using Projective Shape Representation,” Int’l J. Liu, “3D Object Recognition Using Invariance,” Artificial Intelli-
Computer Vision, vol. 16, pp. 57-99, 1995. gence, vol. 78, pp. 239-288, 1995.
[69] M. Rousson and N. Paragios, “Shape Priors for Level Set
Representations,” Proc. European Conf. Computer Vision, May 2002. Siddharth Manay received the PhD degree in
[70] H. Sakoe and S. Chiba, “Dynamic Programming Algorithm 2003 from the School of Electrical Engineering,
Optimization for Spoken Word Recognition,” IEEE Trans. Acous- Georgia Institute of Technology, Atlanta. In
tics, Speech, and Signal Processing, vol. 26, no. 1, pp. 43-49, 1978. 2003, he spent a year as a postdoctoral fellow
[71] G. Sapiro and A. Tannenbaum, “Affine Invariant Scale Space,” with the Vision Lab at the Computer Science
Int’l J. Computer Vision, vol. 11, no. 1, pp. 25-44, 1993. Department, University of California at Los
[72] G. Sapiro and A. Tannenbaum, “Area and Length Preserving Angeles. He is currently with Lawrence Liver-
Geometric Invariant Scale-Spaces,” IEEE Trans. Pattern Analysis more National Laboratory in Livermore, Califor-
and Machine Intelligence, vol. 17, no. 1, pp. 67-72, Jan. 1995. nia. His research interests are in the field image
[73] J. Sato and R. Cipolla, “Affine Integral Invariants for Extracting processing and computer vision, with a particu-
Symmetry Axes,” Image and Vision Computing, vol. 15, no. 8, lar focus on models and invariants of shape.
pp. 627-635, 1997.
[74] H. Schulz-Mirbach, “Invariant Features for Gray Scale Images,” Daniel Cremers received the BS degree in
Proc. 17th Ann. Symp. German Assoc. for Pattern Recognition, pp. 1- mathematics (1994) and physics (1994), and the
14, 1995. MS degree (Diplom) in theoretical physics
[75] J. Schwartz and M. Sharir, “Identification of Partially Obscured (1997) from the University of Heidelberg. In
Objects in Two and Three Dimensions by Matching Noisy 2002, he obtained the PhD degree in computer
Characteristic Curves,” Int’l J. Robotic Research, vol. 6, no. 2, science from the University of Mannheim,
pp. 29-44, 1987. Germany. Subsequently, he spent two years
[76] T. Sebastian, P. Klein, and B. Kimia, “On Aligning Curves,” IEEE as a postdoctoral researcher at the University of
Trans. Pattern Analysis and Machine Intelligence, vol. 25, no. 1, California at Los Angeles and one year as a
pp. 116-125. Jan. 2003. permanent researcher at Siemens Corporate
[77] E. Sharon and D. Mumford, “2D-Shape Analysis Using Conformal Research in Princeton, New Jersey. Since October 2005, he has
Mapping,” Proc. IEEE Conf. Computer Vision and Pattern Recogni- headed the Research Group for Computer Vision and Pattern
tion, June 2004. Recognition at the University of Bonn, Germany. His research is
focused on statistical and variational methods for computer vision. He
[78] D. Sharvit, J. Chan, H. Tek, and B. Kimia, “Symmetry-Based
received several awards, among them the Best Paper of the Year 2003
Indexing of Image Databases,” Proc. IEEE Workshop Content-Based
by the Pattern Recognition Society, the Olympus Award 2004, and the
Access of Image and Video Libraries, pp. 56-62, 1998.
UCLA Chancellor’s Award for Postdoctoral Research 2005. He is a
[79] A. Shashua and N. Navab, “Relative Affine Structure: Canonical
member of the IEEE.
Model for 3D From 2D Geometry and Applications,” IEEE Trans.
Pattern Analysis and Machine Intelligence, vol. 18, no. 9, pp. 873-883,
Sept. 1996. Byung-Woo Hong received the BSc degree in
[80] K. Siddiqi, A. Shokoufandeh, S.J. Dickinson, and S.W. Zucker, mathematics from Chung-Ang University, Korea
“Shock Graphs and Shape Matching,” Proc. Int’l Conf. Computer in 1995, the MSc degree in computer science
Vision, pp. 222-229, 1998. from the Weizmann Institute of Science, Israel in
2001, and the DPhil degree in information
[81] C.E. Springer, Geometry and Analysis of Projective Spaces. San
engineering from the University of Oxford,
Francisco: Freeman, 1964.
United Kingdom in 2005. He joined the computer
[82] H. Tagare, D. O’Shea, and A. Rangarajan, “A Geometric
vision lab at the University of California at Los
Correspondence for Shape-Based NonRigid Correspondence,”
Angeles as a postdoctoral researcher in 2005.
Proc. Int’l Conf. Computer Vision, pp. 434-439, 1995.
His research interests include shape analysis,
[83] Q.M. Tieng and W.W. Boles, “Recognition of 2D Object Contours
registration, segmentation, and medical imaging.
Using the Wavelet Transform Zero-Crossing Representation,”
IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 19, no. 8,
pp. 910-916, Aug. 1997.
[84] C. Tomasi and R. Manduchi, “Stereo without Search,” Proc.
European Conf. Computer Vision, pp. 452-465, 1996.
[85] A. Trouve and L. Younes, “Diffeomorphic Matching Problems in
One Dimension: Designing and Minimizing Matching Functions,”
Proc. European Conf. Computer Vision, pp. 573-587, 2000.

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.
1618 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 28, NO. 10, OCTOBER 2006

Anthony J. Yezzi Jr. received the PhD degree Stefano Soatto received the PhD degree in
in 1997 through the Department of Electrical control and dynamical systems from the Califor-
Engineering at the University of Minnesota. After nia Institute of Technology in 1996. He currently
completing a postdoctoral research position in is a professor of computer science at the
the Laboratory for Information and Decision University of California at Los Angeles (UCLA).
Systems (LIDS) at the Massachusetts Institute He joined UCLA after being an assistant and
of Technology, he joined the faculty of the then associate professor of electrical and bio-
School of Electrical and Computer Engineering medical engineering at Washington University,
at Georgia Institute of Technology in 1999 where and a research associate in applied sciences at
he currently holds the position of associate Harvard University. He was also Ricercatore in
professor. Dr. Yezzi has also consulted for a number of medical the Department of Mathematics and Computer Science at the University
imaging companies including GE, Picker, and VTI, and has been an of Udine, Italy. He received the DIng degree (highest honors) from the
IEEE member since 1999. His research lies primarily within the fields of University of Padova, Italy in 1992. His general research interests are in
image processing and computer vision. He has worked on a variety of computer vision and nonlinear estimation and control theory. In
problems including image denoising, edge-detection, segmentation and particular, he is interested in ways for computers to use sensory
grouping, shape analysis, multiframe stereo reconstruction, tracking, information (e.g., vision, sound, touch) to interact with humans and the
and registration. Some central themes of his research include curve and environment. Dr. Soatto is the recipient of the David Marr Prize for work
surface evolution theory, differential geometry, and partial differential on Euclidean reconstruction and reprojection up to subgroups. He also
equations. received the Siemens Prize with the Outstanding Paper Award from the
IEEE Computer Society for work on optimal structure from motion. He
received the US National Science Foundation Career Award and the
Okawa Foundation Grant. He is an associate editor of the IEEE
Transactions on Pattern Analysis and Machine Intelligence (TPAMI),
and a member of the editorial board of the International Journal of
Computer Vision (IJCV) and of the journal Foundations and Trends in
Computer Graphics and Vision. He is a member of the IEEE.

. For more information on this or any other computing topic,

please visit our Digital Library at www.computer.org/publications/dlib.

Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on October 6, 2008 at 18:57 from IEEE Xplore. Restrictions apply.