IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 33, NO.

7, JULY 2011 1415

Shape Analysis of Elastic Curves

in Euclidean Spaces
Anuj Srivastava, Eric Klassen, Shantanu H. Joshi, and Ian H. Jermyn

Abstract—This paper introduces a square-root velocity (SRV) representation for analyzing shapes of curves in euclidean spaces
under an elastic metric. In this SRV representation, the elastic metric simplifies to the IL2 metric, the reparameterization group acts by
isometries, and the space of unit length curves becomes the unit sphere. The shape space of closed curves is the quotient space of (a
submanifold of) the unit sphere, modulo rotation, and reparameterization groups, and we find geodesics in that space using a path
straightening approach. These geodesics and geodesic distances provide a framework for optimally matching, deforming, and
comparing shapes. These ideas are demonstrated using: 1) shape analysis of cylindrical helices for studying protein structure,
2) shape analysis of facial curves for recognizing faces, 3) a wrapped probability distribution for capturing shapes of planar closed
curves, and 4) parallel transport of deformations for predicting shapes from novel poses.

Index Terms—Elastic curves, Riemannian shape analysis, elastic metric, Fisher-Rao metric, square-root representations, path
straightening method, elastic geodesics, parallel transport, shape models.

1 INTRODUCTION
“closed” the curves along those geodesics to obtain
S HAPE is an important feature for characterizing objects in
several branches of science, including computer vision,
medical diagnostics, bioinformatics, and biometrics. The
deformations between closed curves. Klassen et al. [14]
restricted themselves to arc-length parameterized planar
variability exhibited by shapes within and across classes is curves and derived numerical algorithms for computing
often quite structured and there is a need to capture these geodesics between closed curves, the first ones to do so
variations statistically. One of the earliest works in directly on the space of closed curves and in a manner that
statistical analysis and modeling of shapes of objects came is invariant to reparameterization. Among other things,
from Kendall and colleagues [6], [12]. While this formula- they applied this framework to statistical modeling and
tion took major strides in shape analysis, its limitation was analysis using large databases of shapes [30]. Michor and
the use of landmarks in defining shapes. Since the choice of Mumford [18] and Mennucci [17], [32] have exhaustively
landmarks is often subjective and also because objects in studied several choices of Riemannian metrics on spaces of
images or in imaged scenes are more naturally viewed as planar curves for the purpose of comparing their shapes.
having continuous boundaries, there has been a recent Mio et al. [20] presented a family of elastic metrics that
focus on shape analysis of curves and surfaces, albeit in the quantified the relative amounts of bending and stretching
same spirit as Kendall’s formulation. Consequently, there needed to deform shapes into each other. Similarly, Shah
is now significant literature on shapes of continuous curves [27] derived geodesic equations for planar closed curves
as elements of infinite-dimensional Riemannian manifolds under different elastic metrics and different representations
called shape spaces. This highly focused area of research of curves. In all of these formulations, a shape space is
started with the efforts of Younes [33], who first defined typically constructed in two steps. First, a mathematical
shape spaces of planar curves and imposed Riemannian representation of curves with appropriate constraints leads
metrics on them. In particular, he computed geodesic paths to a preshape space. Then, one identifies elements of the
between curves under these metrics as open curves and preshape space that belong to the same orbits of shape-
preserving transformations (rotations, translations, and
scalings, as well as reparameterizations). The resulting
quotient space, i.e., the set of orbits under the respective
. A. Srivastava is with the Department of Statistics, Florida State
University, Tallahassee, FL 32306. E-mail: [email protected].
group actions, is the desired shape space. If a preshape
. E. Klassen is with the Department of Mathematics, Florida State space is a Riemannian (Hilbert) manifold, then the shape
University, Tallahassee, FL 32306. E-mail: [email protected]. space can inherit this Riemannian structure and become a
. S. Joshi is with the Laboratory of Neuro Imaging, Department of quotient manifold or an orbifold.
Neurology, UCLA School of Medicine, 635 Charles E. Young Drive The choices of shape representation and Riemannian
South, Suite 225, Los Angeles, CA 90095-7334.
E-mail: [email protected].
metric are critically important—for improved understand-
. I.H. Jermyn is with the Department of Mathematical Sciences, Durham ing, physical interpretations, and efficient computing. This
University, Science Laboratories, South Rd, Durham DH1 3LE, UK. paper introduces a particularly convenient representation
E-mail: [email protected]. that enables simple physical interpretations of the resulting
Manuscript received 4 Nov. 2009; revised 26 July 2010; accepted 10 Aug. deformations. This representation is motivated by the well-
2010; published online 30 Sept. 2010. known Fisher-Rao metric used previously for imposing a
Recommended for acceptance by N. Paragios. Riemannian structure on the space of probability densities.
For information on obtaining reprints of this article, please send e-mail to:
[email protected], and reference IEEECS Log Number Taking the positive square root of densities results in a
TPAMI-2009-11-0744. simple euclidean structure, where geodesics, distances, and
Digital Object Identifier no. 10.1109/TPAMI.2010.184. statistics are straightforward to compute [2], [28]. A similar
0162-8828/11/$26.00 ß 2011 IEEE Published by the IEEE Computer Society
Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
1416 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 33, NO. 7, JULY 2011

idea was introduced by Younes [33] and later used in that is natural, general, and efficient. We describe one such
Younes et al. [34] for studying shapes of planar curves under representation.
an elastic metric. The representation used in the current
paper is similar to these earlier ideas, but is sufficiently 2.1 SRV Representation and Preshape Space
different to be applicable to curves in arbitrary IRn . The Let  be a parameterized curve ( : D ! IRn ), where D is a
main contributions of this paper are as follows: certain domain for the parameterization. We are going to
restrict to those  that are absolutely continuous on D. In
1. Presentation of a square-root velocity (SRV) repre- general, D will be ½0; 1, but for closed curves, it will be more
sentation for studying shapes of elastic closed curves
natural to have D ¼ S1 . p We define
ffiffiffiffiffiffiffi
ffi a mapping: F : IRn ! IRn
in IRn , first introduced in the conference papers [8],
[9]. This has several advantages, as discussed later. according to F ðvÞ  v= kvk, if kvk 6¼ 0 and 0 otherwise.
2. The use of a numerical approach, termed path Here, k  k is the euclidean 2-norm in IRn ; note that F is a
straightening, for finding geodesics between shapes continuous map. For the purpose of studying the shape of
of closed elastic curves. It uses a gradient-based , we will represent it using the SRV function defined as
iteration to find a geodesic where, using the Palais q : D ! IRn , where
metric on the space of paths, the gradient is available qffiffiffiffiffiffiffiffiffiffiffiffiffi
in a convenient analytical form. _ _
qðtÞ  F ððtÞÞ ¼ ðtÞ= kðtÞk _ :
3. The use of a gradient-based solution for optimal
reparameterization of curves when finding geode- This representation includes those curves whose parame-
sics between their shapes. This paper compares the terization can become singular in the analysis. Also, for
strengths and weaknesses of this gradient solution every q 2 IL2 ðD; IRn Þ, there exists a curve  (unique up to a
versus the commonly used Dynamic Programming translation) such that the given q is the SRV function of that
(DP) algorithm. . In fact,
R t this curve can be obtained using the equation:
4. The application and demonstration of this frame- ðtÞ ¼ 0 qðsÞkqðsÞkds. The motivation for using this repre-
work to: sentation and comparisons with other such representations
is presented in the Section 3.1.
a. shape analysis of cylindrical helices in IR3 for
To remove the scaling variability, we rescale all curves to
use in studies of protein backbone structures,
be of unit length. This restriction to an orthogonal section of
b. shape analysis of 3D facial curves,
the full space of curves is identical to Kendall’s [12]
c. development of a wrapped normal distribution
to capture shapes in a shape class, and approach for removing the scale variability. The remaining
d. parallel transport of deformations from one transformations (rotation, translation, and reparameteriza-
shape to another. tion) will be dealt with differently. This is due to the
differences in the actions of scaling and other groups on the
The last item is motivated by the need to predict
representation space of curves, as described later. The
individual shapes or shape models for novel objects,
restriction
R that  is of
R unit length translates to the condition
or novel views of the objects, using past data. A
that D kqðtÞk2 dt ¼ D kkdt _ ¼ 1. Therefore, the SRV func-
similar approach has been applied to shape repre-
tions associated with these curves are elements of a unit
sentations using deformable templates [35] and for
hypersphere in the Hilbert manifold IL2 ðD; IRn Þ; we will use
studying shapes of 3D triangulated meshes [13].
Co to denote this hypersphere. According to Lang [15, p. 27],
The proposed representation spaces for curves are Co is a Hilbert submanifold in IL2 ðD; IRn Þ.
infinite-dimensional manifolds or, rather, their quotient
For studying shapes of closed curves, we impose an
spaces under the actions of infinite-dimensional groups.
additional condition that the curve starts and ends at the
The infinite dimensionality of such representations is an
important challenge. At a conceptual level, however, it may same point. In view of this condition, it is natural to have
help a reader to understand the proposed solutions on the domain D be the unit circle S1 for closed curves. For a
finite-dimensional manifolds at first and consider the issue certain placement of the origin on S1 , it can be identified
of infinite dimensionality later. Also, we clarify the use of with ½0; 1 using the function t 7! ðcosð2tÞ; sinð2tÞÞ. We
word geodesic in this paper. We refer to a path with a will use either one according to convenience. In terms of the
(covariantly) constant velocity (defined later in Section 4) as SRV
R function, this closure condition is given by
a geodesic and the shortest geodesic between any two points S1 qðtÞkqðtÞkdt ¼ 0. Thus, we have a space of unit length,
as a minimizing geodesic. closed curves represented by their SRV functions:
The paper is organized as follows: Section 2 introduces Z Z
the proposed elastic shape framework, while Section 3 c 2 1
C ¼ fq 2 IL ðS ; IR Þj n 2
kqðtÞk dt ¼ 1; qðtÞkqðtÞkdt ¼ 0g:
discusses its merits relative to existing literature. Section 4 S1 S1
describes a path straightening approach for finding
geodesics and a gradient-based approach for elastic curve The superscript c implies the closure condition. With the
registration. Section 5 presents four applications of this earlier identification of ½0; 1 with S1 , Cc  Co  IL2 ðD; IRn Þ.
framework. The paper ends with a short summary in What is the nature of the set Cc ? In the Appendix, we sketch
Section 6. a proof that Cc is a codimension-n submanifold of Co .
Now we have two submanifolds—Co and Cc —containing
all curves and only closed curves in IRn , respectively. They
2 SHAPE REPRESENTATION are called preshape spaces for their respective cases. We will
In order to develop a formal framework for analyzing call Co the preshape space of open curves just to emphasize
shapes of curves, one needs a mathematical representation that the closure constraint is not enforced here, even though

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
SRIVASTAVA ET AL.: SHAPE ANALYSIS OF ELASTIC CURVES IN EUCLIDEAN SPACES 1417

it does contain closed curves also, while Cc is purely the

preshape space of closed curves. To impose Riemannian
structures on these preshape spaces, we consider their
tangent spaces.

1. Open Curves: Since Co is a sphere in IL2 ð½0; 1; IRn Þ, its

tangent space at a point q is given by Tq ðCo Þ ¼ fv 2
IL2 ð½0; 1; IRn Þjhv; qi ¼ 0g. Here, hv; qi,R denotes the
1
inner product in IL2 ð½0; 1; IRn Þ: hv; qi ¼ 0 hvðtÞ; qðtÞidt.
2. Closed Curves: The tangent space to Cc at a point q Fig. 1. Reparameterizations of open and closed curves using orienta-
tion-preserving diffeomorphisms.
is, of course, a subset of IL2 ðS1 ; IRn Þ. Since Cc is a
qffiffiffiffiffiffiffiffi
submanifold, this subset is often defined using the _ _ ;
R F ðððtÞÞ _
ðtÞÞ ¼ qððtÞÞ ðtÞ
differential of the map q 7! GðqÞ ¼ S1 qðtÞkqðtÞkdt. In
fact, the tangent space Tq ðCc Þ at a point q 2 Cc is where q is the SRV of .pffiffiffiThis gives us the right action
given by the kernel of the differential of G at that C   ! C, ðq; Þ ¼ ðq  Þ _ . In order for our shape
point [19]. Therefore, it is often easier to specify the comparison to be invariant to these transformations, it is
normal space, i.e., the space of functions in important for these groups to act by isometries. We note the
IL2 ðS1 ; IRn Þ that are perpendicular to Tq ðCc Þ. This following properties of these actions:
normal space is found using the directional deriva- Lemma 1. The actions of SOðnÞ and  on C commute.
tives of G, and is given by Proof. It follows from the definition. u
t
  
c qi ðtÞ
Nq ðC Þ ¼ span qðtÞ; qðtÞ þ kqðtÞkei ; Therefore, we can form a joint action of the product
kqðtÞk pffiffiffi
 ð1Þ group SOðnÞ   on C according to ððO; Þ; qÞ ¼ Oðq  Þ _ .
i ¼ 1; . . . ; n : Lemma 2. The action of the product group   SOðnÞ on C is by
isometries with respect to the chosen metric.
H e n c e , Tq ðCc Þ ¼ fv 2 IL2 ðS1 ; IRn Þjhv; wi ¼ 0; 8w 2 Proof. For a q 2 C, let u; v; 2 Tq ðCÞ. Since hOuðtÞ; OvðtÞi ¼
Nq ðCc Þg. huðtÞ; vðtÞi for all O 2 SOðnÞ and t 2 D, the proof for
The standard metric on IL2 ðD; IRn Þ restricts to the two SOðnÞ follows. For the  part, fix an arbitrary element
manifolds Co and Cc to form Riemannian structures on  2 , and define a map  : C ! C by ðqÞ ¼ ðq; Þ. A
them. These structures can then be used to determine glance at the formula for ðq; Þ confirms that  is a linear
geodesics and geodesic lengths between elements of these transformation. Hence, its derivative d has the same
spaces. Let C be a Riemannian manifold denoting either Co formula as . In other words, the mapping pffiffiffi d : Tq ðCÞ !
or Cc , and let  : ½0; 1 ! C be a parameterized path such that Tðq;Þ ðCÞ is given by u 7! u~  ðu  Þ _ . The Riemannian
ð0Þ ¼ q0 and metric after the transformation is h~ u; v~i ¼
R 1 ð1Þ ¼ q1 . 1=2 Then, the length of  is defined to
Z Z
be L½ ¼ 0 hðÞ; _ _
ðÞi d, and  is a said to be a pffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffi
minimizing geodesic if L½ achieves the infimum over all h~
uðtÞ; v~ðtÞidt ¼ huððtÞÞ ðtÞ
_ ; vððtÞÞ ðtÞ _ idt
D
such paths. The length of this geodesic becomes a distance ZD
dc ðq0 ; q1 Þ ¼ inf f:½0;1!Cjð0Þ¼q0 ;ð1Þ¼q1 g L½. The computation of ¼ huðÞ; vðÞid; with  ¼ ðtÞ:
D
geodesics in Co is straightforward since it is a sphere, but the
case of Cc is more complicated and requires numerical Putting these two results together, the joint action of  
methods described in Section 4. SOðnÞ on C is by isometries with respect to the chosen
metric. u
t
2.2 Shape Space as Quotient Space
Since the action of the product group is by isometries, we
By representing a parameterized curve R  by its SRV can form a quotient space of C modulo   SOðnÞ and try to
function q, and imposing the constraint D hqðtÞ; qðtÞidt ¼ 1, inherit the Riemannian metric from C to that quotient space.
we have taken care of the translation and the scaling The orbit of a function q 2 C is given by
variability, but the rotation and the reparameterization pffiffiffi
variability still remain. A rotation is an element of SOðnÞ, ½q ¼ fOðq  Þ _ Þjð; OÞ 2   SOðnÞg:
the special orthogonal group of n  n matrices, and a
An orbit is associated with a shape uniquely and comparisons
reparameterization is an element of , the set of all between shapes are performed by comparing the orbits of the
orientation-perserving diffeomorphisms of D. In the follow- corresponding curves, and thus the need for a metric on the
ing discussion, C stands for either Co or Cc . set of orbits. We would like to use the basic fact that if a
The rotation and reparameterization of a curve  are compact Lie group H acts freely on a Riemannian manifold M
denoted by the actions of SOðnÞ and  on its SRV. While the (i.e., no elements of M are fixed by h 2 H unless h is the
action of SOðnÞ is the usual SOðnÞ  C ! C, ðO; qðtÞÞ ¼ identity) by isometries, and if the orbits are closed, then the
OqðtÞ, the action of  is derived as follows: For a  2 , the quotient M=H is a manifold and inherits a Riemannian metric
composition    denotes its reparameterization (as shown from M. The trouble is that while the group   SOðnÞ acts by
in Fig. 1); the SRV of the reparameterized curve is isometries, the orbits are not closed. The reason for this is that

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
1418 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 33, NO. 7, JULY 2011

the space of diffeomorphisms is not closed with respect to Definition 1 (Elastic Metric). Let a and b be positive real
either the IL2 or the Palais metric since a sequence of numbers. For ðu1 ; v1 Þ; ðu2 ; v2 Þ 2 Tð;Þ ð  Þ, define an
diffeomorphisms might approach a map which is not a inner product
diffeomorphism under either of these two metrics. To resolve Z
this theoretical difficulty, we propose that instead of modd- hðu1 ; v1 Þ; ðu2 ; v2 Þið;Þ ¼ a2 u1 ðtÞu2 ðtÞeðtÞ dt
D
ing out by the orbits, we mod out by the closures of these Z ð3Þ
orbits. Thus, if there is a sequence qi in the orbit ½q, and this þ b2 hv1 ðtÞ; v2 ðtÞieðtÞ dt:
sequence converges to a function q~ in Co (with respect to the D

IL2 -metric), then we identify q with q~ in this quotient Note that h; i in the second integral on the right denotes
construction. As evidence that this idea has merit, one can the standard dot product in IRn . This elastic metric,
prove that in this situation, if we let  and ~ be the curves introduced in [20], has the interpretation that the first
corresponding to q and q~, both  and ~ contain exactly the integral measures the amount of “stretching” since u1 and
~
same points. (This is assuming that we set ð0Þ ¼ ð0Þ.) With u2 are variations of the log speed  of the curve, while the
a slight abuse of notation, we will use ½q to denote the closure second integral measures the amount of “bending” since v1
of the orbit of q. Define the quotient space S as the set of all and v2 are variations of the direction  of the curve. The
such closed orbits associated with the elements of C, i.e., constants a2 and b2 are weights that we choose depending
S ¼ f½qjq 2 Cg. on how much we want to penalize these two types of
Since we have a quotient map from C to S, its differential deformations.
induces a linear isomorphism between T½q ðSÞ and the Perhaps the most important property of this Riemannian
normal space to ½q at any point q~ 2 ½q. The Riemannian metric is that the groups SOðnÞ and  both act by isometries.
metric on C (i.e., the IL2 inner product) restricts to an inner To elaborate on this, recall that O 2 SOðnÞ acts on a curve  by
product on the normal space, which in turn induces an ðO; ÞðtÞ ¼ OðtÞ, and  2  acts on  by ð; ÞðtÞ ¼ ððtÞÞ.
inner product on T½q ðSÞ. The fact that   SOðnÞ acts by Using our identification of the set of curves with the space
isometries implies that the resulting inner product on T½q ðSÞ    results in the following actions of these groups. O 2
is independent of the choice of q~ 2 ½q. In this manner, S SOðnÞ acts on ð; Þ by ðO; ð; ÞÞ ¼ ð; OÞ and  2  acts on
inherits a Riemannian structure from C. Consequently, the ð; Þ by ð; ð; ÞÞ ¼ ð   þ ln ;
_   Þ.
geodesics in S correspond to those geodesics in C that We now need to understand the differentials of these
are perpendicular to all of the orbits they meet in C and the group actions on the tangent spaces of   . SOðnÞ is easy;
geodesic distance between any two points in S is given by since each O 2 SOðnÞ acts by the restriction of a linear
pffiffiffi transformation on   L2 ðD; IRn Þ, it acts in exactly the same
ds ð½q0 ; ½q1 Þ ¼ inf dc ðq0 ; Oðq1  Þ _ Þ: ð2Þ way on the tangent spaces ðO; ðu; vÞÞ ¼ ðu; OvÞ, where ðu; vÞ 2
ð;OÞ2SOðnÞ
Tð;Þ ð  Þ and ðu; OvÞ 2 Tð;OÞ ð  Þ. The action of  2 
We state without proof that if q0 and q1 lie in two different given in the above formula is not linear, but affine linear,
orbits which are not in each other’s closure, then this because of the additive term ln . _ Hence, its action on the
distance is strictly positive. tangent space is the same, but without this additive term
ð; ðu; vÞÞ ¼ ðu  ;   Þ, where ðu; vÞ 2 Tð;Þ ð  Þ, and
ðu  ;   Þ 2 Tð;ð;ÞÞ ð  Þ. Combining these actions of
3 MOTIVATION AND COMPARISONS SOðnÞ and  with the above inner product on   , it is an
We first motivate the choice of SRV and the elastic metric easy verification that these actions are by isometries, i.e.,
for shape analysis and then compare our choice with
previous ideas. hðO; ðu1 ; v1 ÞÞ; ðO; ðu2 ; v2 ÞÞiðO;ð;ÞÞ ¼ hðu1 ; v1 Þ; ðu2 ; v2 Þið;Þ ;
hð; ðu1 ; v1 ÞÞ; ð; ðu2 ; v2 ÞÞið;ð;ÞÞ ¼ hðu1 ; v1 Þ; ðu2 ; v2 Þið;Þ :
3.1 Motivation for the SRV Representation
Let  : D ! IRn be a curve in IRn . Assume that for all t 2 D, Since we have identified the space of curves with   , we
_ 6¼ 0 (this is only for comparing with past works, and our
ðtÞ may identify the space of shapes with the quotient space
method does not require it). We then define  : D ! IR by ð  Þ=ðSOðnÞ  Þ. Furthermore, since these group actions
_
ðtÞ ¼ lnðkðtÞkÞ, and  : D ! Sn1 by ðtÞ ¼ ðtÞ=k _ _
ðtÞk. are by isometries with respect to all of the metrics we
Clearly,  and  completely specify _ since, for all t, introduced above, no matter what values we assign to a and b,
_ ¼ eðtÞ ðtÞ. Thus, we have defined a map from the space
ðtÞ we get a corresponding two-parameter family of metrics on
of open curves in IRn to   , where  and  are sets of the quotient space ð  Þ=ðSOðnÞ  Þ. Note that in
smooth maps. This map is surjective; it is not injective, but distinguishing between the structures (for example, geode-
sics) associated to these metrics, only the ratio of a to b is
two curves are mapped to the same pair ð; Þ if and only if
important since if we multiply both by the same real number,
they are translates of each other, i.e., if they differ by an
we just rescale the metric, which results in the same geodesics.
additive constant. In physical terms,  is the (log of the) speed This is not the only consideration, however. The issue of
of traversal of the curve, while  is the direction of the curve at computing geodesics between curves for different choices
each t. of c ¼ b=2a remains, especially once we restrict attention to
The tangent space of    at any point ð; Þ is given by the space of unit length curves. One can ask: Is there some
Tð;Þ ð  Þ ¼   fv 2 IL2 ðD; IRn ÞjvðtÞ ? ðtÞ; 8t 2 Dg. We particular choice of weights which will be especially natural
now define a Riemannian metric on   . and which will result in the geodesics becoming easier to

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
SRIVASTAVA ET AL.: SHAPE ANALYSIS OF ELASTIC CURVES IN EUCLIDEAN SPACES 1419

compute? We now show that the SRV representation . When c 6¼ 1: For n ¼ 2, the representation space of
provides an answer to this question. curves is flat except at qc ¼ 0, where it is singular; for
1
In terms of ð; Þ, SRV is given by qðtÞ ¼ e2ðtÞ ðtÞ. A n > 2, the curvature is again singular at qc ¼ 0;
simple derivation shows that if ðu; vÞ 2 Tð;Þ ð  Þ, then otherwise, it is nonflat (the curvature is not zero).
the corresponding tangent vector to IL2 ðD; IRn Þ at q is given . When c ¼ 1: The curvature is identically zero for all n;
1 1
by f ¼ 12 e2 u þ e2 v. Now let ðu1 ; v1 Þ and ðu2 ; v2 Þ denote two the space of curves is flat.
elements of Tð;Þ ð  Þ, and let f1 and f2 denote the The euclidean coordinates thus exist for all n only when
c ¼ 1: These coordinates are the SRV representation. We
corresponding tangent vectors to IL2 ðD; IRn Þ at q. Comput-
conjecture that this situation continues to hold in the infinite-
ing the IL2 inner product of f1 and f2 yields dimensional case. This would mean that the SRV represen-
Z   tation occupies a unique position among curve representa-
1 1 1 1 1 1
hf1 ; f2 i ¼ e2 u1  þ e2 v1 ; e2 u2  þ e2 v2 dt tions. In addition to providing a more stable representation
D 2 2
for the n ¼ 2 case, when compared to Younes et al. [34], it
Z   ð4Þ
1  also covers the case n > 2 that has not been studied before.
¼ e u1 u2 þ e hv1 ; v2 i dt:
D 4

In this computation, we have used the fact that hðtÞ; ðtÞi ¼ 1 4 COMPUTATION OF GEODESICS
since ðtÞ is an element of the unit sphere, and that In this section, we focus on the task of computing geodesics
hðtÞ; vi ðtÞi ¼ 0 since each vi ðtÞ is a tangent vector to the unit between any given pair of shapes in a shape space. This task
sphere at ðtÞ. This expression, when compared with (3), is accomplished in two steps. First, we develop tools for
shows that the IL2 metric on the space of SRV representations computing geodesics in the preshape spaces, Co or Cc , and
corresponds precisely to the elastic metric on   , with then we remove the remaining shape-preserving transfor-
a ¼ 1=2 and b ¼ 1. However, expressed in terms of the SRV mations to obtain geodesics in the shape spaces. In the case
functions, the IL2 -metric is the “same” R at every point of of Co , the underlying space is a sphere and the task of
IL2 ðD; IRn Þ (it is simply hf1 ; f2 i ¼ D hf1 ðtÞ; f2 ðtÞi dt, which computing geodesic paths there is straightforward. For any
does not depend on the point at which these tangent vectors two points q0 and q1 in Co , a geodesic connecting them is
are defined), and we will thus have access to more efficient given by  : ½0; 1 ! Co ,
ways of computing geodesics in our preshape and shape
spaces using the SRV formulation. We emphasize again that 1
ðÞ ¼ ðsinðð1  ÞÞq0 þ sinðÞq1 Þ; ð6Þ
this is true for curves in arbitrary dimension. sinðÞ

3.2 Comparison with Prior Work where  ¼ cos1 ðhq0 ; q1 iÞ is the length of the geodesic.
The previous section showed that the SRV representation However, we will use a path straightening approach to
provides euclidean coordinates for the space of parameter- compute geodesics in Cc .
ized curves in IRn equipped with the elastic metric. In this Notationally, we are using  to parameterize paths on
section, we compare the SRV representation to previous spaces of curves and t to parameterize individual curves.
work, and provide evidence that this is the only case for 4.1 Path Straightening Method: Theory
which euclidean coordinates can be found.
For any two closed curves, denoted by q0 and q1 in Cc , we
When n ¼ 1, there is no  component R and the elastic are interested in finding a geodesic path between them in
metric in (3) takes the form hu1 ; u2 i ¼ D u1 ðtÞu2 ðtÞeðtÞ dt.
Cc . We start with an arbitrary path ðÞ connecting q0 and
This is called the Fisher-Rao metric and has been used for
q1 , i.e.,  : ½0; 1 7! Cc such that ð0Þ ¼ q0 and ð1Þ ¼ q1 .
imposing a Riemannian structure on the space of prob-
Then, we iteratively “straighten”  until it achieves a local
ability density functions on D [1], [2], [4]. Note that for a
minimum of the energy:
curve of unit length, eðtÞ can be interpreted as a probability
density function. It is well known, at least since 1943 [2], Z  
1 1 d d
that under the square-root representation, i.e., for EðÞ  ðÞ; ðÞ d ð7Þ
1
2 0 d d
qðtÞ ¼ e2ðtÞ , this metric reduces to the IL2 metric, given by
(4) with n ¼ 1. over all paths from q0 to q1 . It can be shown that a critical
To discuss n > 1, it is useful to use a slightly different point of E is a geodesic on Cc . However, it is possible that
_ _ 12c1 . For vc ; wc there are multiple geodesics between a given pair q0 and q1 ,
representation. Let us define qc ¼ ðtÞ=k ðtÞk
and a local minimum of E may not correspond to a
in the tangent space at qc , the elastic metric becomes minimizing geodesic. Therefore, this approach has the
Z limitation that it finds a geodesic between a given pair
hvc ; wc iqc ¼ b2 kqc ðtÞkð2c2Þ hvc ðtÞ; wc ðtÞi dt: ð5Þ but may not reach the minimizing geodesic, if it exists. A
D
cartoon illustration of this method for a unit two-sphere is
Notice that when c ¼ 1, the integrand is the euclidean shown in Fig. 2.
metric on IRn ; otherwise, it is not. If we use a discrete Let H be the set of all paths in Cc and H0 be the subset of
representation of curves, say using N points sampled on H of paths that start at q0 and end at q1 . The tangent space of
each curve, one can calculate the curvature of the resulting H i s T ðHÞ ¼ fwj8 2 ½0; 1; wðÞ 2 TðÞ ðCc Þg, w h e r e
finite-dimensional representation space (details are TðÞ ðCc Þ is specified as a set orthogonal to NðÞ ðCc Þ (defined
omitted). This calculation shows that: in (1)). A tangent w is actually a tangent vector field along 

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
1420 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 33, NO. 7, JULY 2011

2.2
the backward parallel translation of a tangent vector
2.1
w1 2 Tð1Þ ðCc Þ, along , when, for ðÞ
~  ð1  Þ, u is the
2

forward parallel translation of w1 along .~ It must be noted

Path Length
1.9

1.8

1.7
that parallel translations, forward or backward, lead to
1.6
vector fields that are covariantly constant.
1.5
1 2 3 4
iterations
5 6 7

According to Lemma 3, to project the gradient u into

Fig. 2. An example of path straightening method for computing T ðH0 Þ, we simply need to subtract off a covariantly linear
geodesics between two points on S2 . The right panel shows the vector field which agrees with u at  ¼ 0 and  ¼ 1 (recall
decrease in the path length.
that uð0Þ ¼ 0). Clearly, the correct covariantly linear field is
simply  u~ðÞ, where u~ðÞ is the covariantly constant field
such that wðÞ is tangent to Cc at ðÞ. Similarly,
obtained by parallel translating uð1Þ backward along .
T ðH0 Þ ¼ fw 2 T ðHÞjwð0Þ ¼ wð1Þ ¼ 0g. To ensure that 
Hence, we have proven the following theorem:
stays at the desired end points, the allowed vector field on 
has to be zero at the ends. Theorem 2. Let  : ½0; 1 7! Cc be a path,  2 H0 . Then, for u as
Our study of paths on H requires the use of covariant defined in Theorem 1, the gradient of the energy function E
derivatives and integrals of vector fields along these paths. restricted to H0 is wðÞ ¼ uðÞ   u~ðÞ, where u~ is the vector
For a given path  2 H and a vector field w 2 T ðHÞ, the field obtained by parallel translating uð1Þ backward along .
covariant derivative of w along  is the vector field
obtained by projecting dw d ðÞ onto the tangent space To finish this discussion, we show that the critical points
TðÞ ðCc Þ, for all , and is denoted by Dw d ðÞ. Similarly, a of E are geodesics.
vector field u 2 T ðHÞ is called a covariant integral of w Lemma 4. For a given pair q0 ; q1 2 Cc , a critical point of E on H0
along  if the covariant derivative of u is w, i.e., Du d ¼ w. is a geodesic on Cc connecting q0 and q1 .
To make H a RiemannianR1 manifold, an obvious metric
would be hw1 ; w2 i ¼ 0 hw1 ðÞ; w2 ðÞid, for w1 ; w2 2 T ðHÞ. Proof. Let  be a critical point of E in H0 . That is, the
Instead, we use the Palais metric [22], which is gradient of E is zero at . Since the gradient vector field
Z 1  is given by uðÞ   u~ðÞ, we have that uðÞ ¼  u~ðÞ for all
Dw1 Dw2 Dð u~Þ
. Therefore, d Du
d ¼ d ¼ d ¼ u ~. Since u~ is a parallel
hhw1 ; w2 ii ¼ hw1 ð0Þ; w2 ð0Þi þ ðÞ; ðÞ d;
0 d d translation of uð1Þ, it is covariantly constant, and there-
fore, the velocity field d d is covariantly constant. By
where h; i is the chosen metric on Cc . The reason for using
the Palais metric is that with respect to this metric, T ðH0 Þ is definition, this implies that  is a geodesic. u
t
a closed linear subspace of T ðHÞ, and H0 is a closed subset 4.2 Path Straightening Method: Implementation
of H. Therefore, any vector w 2 T ðHÞ can be uniquely
We present some numerical procedures for computing
projected into T ðH0 Þ. This enables us to derive the gradient
geodesic paths between curves represented by q0 and q1 in
of E as a vector field on .
Cc . There are two basic items that are used repeatedly in
Our goal is to find the minimizer of E in H0 , and we will
use a gradient flow to do that. Therefore, we wish to find these procedures: 1) For projecting arbitrary points in
the gradient of E in T ðH0 Þ. To do this, we first find the IL2 ðS1 ; IRn Þ into Cc , and 2) For projecting arbitrary points
gradient of E in T ðHÞ and then project it into T ðH0 Þ. in IL2 ðS1 ; IRn Þ into Tq ðCc Þ for some q 2 Cc .
Item 1: The projection from IL2 ðD; IRn Þ to Co is simple
Theorem 1. The gradient vector of E in T ðHÞ is given by the
q 7! q=kqk. The further projection from Co to Cc is realized as
unique vector field u such that Du=d ¼ d=d and uð0Þ ¼ 0.
follows: Recall the mapping G : Co ! IRn given by GðqÞ ¼
In other words, u is the covariant integral of d=d with zero R 2 n
initial value at  ¼ 0. 0 qðtÞkqðtÞkdt 2 IR . Our idea is to iteratively update q in
such a way that GðqÞ becomes ð0; . . . ; 0Þ. The update is
Proof. Please refer to the Appendix. u
t
performed in the normal space Nq ðCc Þ since changing q
We will introduce some additional properties of vector along the tangent space Tq ðCc Þ does not change its G value.
fields along  that are useful in our construction. A vector The question is—which particular normal vector should be
field w is called covariantly constant if Dw=d is zero at all used in this update?
points along . Similarly, a path  is called a geodesic if its
1. CRa l c u l a t e t h e J a c o b i a n m a t r i x , Ji;j ¼ ij þ
velocity vector field is covariantly constant. That is,  is a
D d 3 S1 qi ðsÞqj ðsÞds, i; j ¼ 1; 2; . . . ; n. Here, ij ¼ 1 if
geodesic if d ð d Þ ¼ 0 for all . Also, a vector field w along
i ¼ j, else it is zero.
the path  is called covariantly linear if Dw=d is a
2. Compute the residual r ¼ gðqÞ and solve the equa-
covariantly constant vector field.
tion J ¼ r for P2 IRn .
Lemma 3. The orthogonal complement of T ðH0 Þ in T ðHÞ is the 3. Update q ¼ q þ ni¼1 i bi , > 0, where fbi ji ¼
space of all covariantly linear vector fields w along . 1; . . . ; ng form an orthonormal basis of the normal
Proof. Please refer to the Appendix. u
t space Nq ðCc Þ given in (1). Rescale using q 7! q=kqk.
4. If krðqÞk <
, stop. Else, go to Step 1.
A vector field u is called the forward parallel translation Item 2: For the second item, take the orthonormal basis
of a tangent vector w0 2 Tð0Þ ðCc Þ, along , if and only if fbi g of the normal space Nq ðCc Þ and project the given vector
Pnþ1
uð0Þ ¼ w0 and DuðÞ
d ¼ 0 for all  2 ½0; 1. Similarly, u is called w using w 7! w  i¼1 hbi ; wibi .

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
SRIVASTAVA ET AL.: SHAPE ANALYSIS OF ELASTIC CURVES IN EUCLIDEAN SPACES 1421

With these two items, we can address the task of

straightening paths into geodesics. Let fð=kÞ :  ¼
0; 1; 2; . . . ; kg be a given path between q0 and q1 in Cc . First,
we need to compute the velocity vector d d at discrete points
along .
d
Algorithm 1. [Compute d along ]
For all  ¼ 0; 1; . . . ; k,

1. Compute: cð=kÞ ¼ kðð=kÞ  ðð  1Þ=kÞÞ. This

difference is computed in IL2 ðS1 ; IRn Þ.
2. Project cð=kÞ into Tð=kÞ ðCc Þ using Item 2 to get an
approximation for d
d ð=kÞ.
Next, we want to approximate the covariant integral of
d
dt along , using partial sums, i.e., we want to add the Fig. 3. (a) Computing geodesics in the quotient space C=ð  SOðnÞÞ.
current sum, say uðð  1Þ=kÞ, to the velocity d d ð=kÞ. (b) The mapping from u 2 T1 ðÞ to the tangent vector in Tq~1 ð½q1 Þ in two
However, these two quantities are elements of two different steps.
tangent spaces and cannot be added directly. Therefore, we
project uðð  1Þ=kÞ into the tangent space at the point 2. Initialize a path  between q0 and q1 in Co using (6) and
ð=kÞ first and then add it to d
dt ð=kÞ to estimate uð=kÞ. project each point along it in Cc using Item 1.
Algorithm 2. [Compute covariant integral of d 3. Compute the velocity vector field d=d along the path 
d along ]
Set uð0Þ ¼ 0 2 Tð0Þ ðCc Þ. For all  ¼ 1; 2; . . . ; k, using Algorithm 1.
4. Compute the covariant integral of d=d, denoted by u,
1. Project uðð  1Þ=kÞ into the tangent space Tð=kÞ ðS c Þ using Algorithm 2.
(Item 2) and rescale to the original length to result in 5. Compute the backward parallel transport of the vector
uk ðð  1Þ=kÞ. uð1Þ along  using Algorithm 3; denote it by u~.
k
2. Set uð=kÞ ¼ k1 d
d ð=kÞ þ u ðð  1Þ=kÞ. 6. Compute the full gradient vector field of the energy E
Next, we compute an estimate for the backward parallel along the path , denoted by w, using wðÞ ¼
transport of uð1Þ: uðÞ   u~ðÞ (Algorithm 4).
7. Update
Pk  along the vector field w using Algorithm 5. If
Algorithm 3. [Backward parallel transport of uð1Þ]
hwðÞ; wðÞi is small, then stop. Else, return to
Set u~ð1Þ ¼ uð1Þ and l ¼ kuð1Þk. For all  ¼ k  1; ¼1
Step 3.
k  2; . . . ; 0,
In these implementations, each curve is represented by
1. Project u~ðð þ 1Þ=kÞ into Tð=kÞ ðCc Þ using Item 2 to its coordinates at some sampled points and the algorithm
obtain cð=kÞ. smoothly interpolates between them when needed. The
2. Set u~ð=kÞ ¼ lcð=kÞ=kcð=kÞk. derivatives are approximated using symmetric finite differ-
Now, we can compute the desired gradient: ences and integrals are approximated using summations.
Algorithm 4. [Gradient vector field of E in H0 ] 4.4 Removing Shape-Preserving Transformations
For all  ¼ 1; 2; . . . ; k, compute wð=kÞ ¼ uð=kÞ 
Now that we have procedures for constructing geodesics
ð=kÞ~
uð=kÞ.
between points in a preshape space C (Co or Cc ), we focus on
By construction, this vector field, w, is zero at  ¼ 0 and the same task for shape spaces. Toward this goal, we need
 ¼ k. As a final step, we need to update the path  in a to solve the joint minimization problem on ð; OÞ stated in
direction opposite to the gradient of E. (2), with the cost function being H :   SOðnÞ ! IR,
pffiffiffi
Algorithm 5. [Path update] Hð; OÞ ¼ dc ðq0 ; Oðq1  Þ _ Þ. This optimization problem is
Select a small
> 0 as the update step size. For all depicted using a cartoon diagram in Fig. 3a. Our strategy is
 ¼ 0; 1; . . . ; k, perform to fix one variable and iteratively optimize over the other. In
case of Co , this procedure is simple since the solutions to
1. Compute the gradient update 0 ð=kÞ ¼ ð=kÞ 

wð=kÞ. This update is performed in the ambient individual optimizations are well known. For a fixed , the
space IL2 ðS1 ; IRn Þ. optimization of H ¼ Hð; Þ over SOðnÞ is obtained using
2. Project 0 ð=kÞ to Cc using Item 1 to obtain the the SVD while, for a fixed O, the optimization of HO ¼
updated ð=kÞ. Hð; OÞ over  is performed using the DP algorithm.
In case of Cc , these direct solutions do not apply and we
4.3 Path Straightening Algorithm
resort to a gradient-based approach. Let  ðmÞ ¼ 1  2     
Now, we describe an algorithm for computing geodesics in
Cc using path straightening. The subalgorithms referred to m and OðmÞ ¼ O1  O2      Om be the cumulative group
here are listed in the previous section. elements, and at the mth iteration, we seek the increments
Path Straightening Algorithm. To find a geodesic ðmþ1 ; Omþ1 Þ that minimize Hð ðmþ1Þ ; Oðmþ1Þ Þ. Let q~1 denote
between two curves 0 and 1 in Cc . ðmÞ
qffiffiffiffiffiffiffiffiffi element of the orbit ½q1 , i.e., q~1 ¼ O ðq1 
the current
1. Compute their representations q0 and q1 in Cc . _ and let  : ½0; 1 ! C be a geodesic from q to q~ .
 ðmÞ Þ  ðmÞ 0 1

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
1422 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 33, NO. 7, JULY 2011

6 1.4

So, _ 1 is the velocity vector at q~1 and define v  ð1Þ=

_ 5
1.2

1
4
0.8

kð1Þk.
_ This v is precisely the gradient of dc ðq0 ; q~1 Þ with 3

2
0.6

0.4

respect to q~1 .
0.2

0
0 2 4 6 0 50 100 150 200

6 2.5

o o
1. Rotations: In the case of C , since C is a sphere, the 5

4
2

1.5

geodesic length is given by an arc-length, and 3

2
1

minimizing arc length is the same as minimizing 1

0 2 4 6
0.5

0
0 50 100 150 200

the corresponding chord length. Therefore, the

optimal rotation is directly written as
^mþ1 ¼ argmin kq0  O~
O q1 k ¼ UV T ; ð8Þ
Fig. 4. (a) The original shape represented by q0 . p ffiffiffi An arbitrary  2 .
(b)
(c) The second shape formed using q0 ¼ ðq1  Þ _ . (d) Evolution of H
O2SOðnÞ in matching q~1 with q0 . (e) Final curve represented by q~1 .
R
T
where UV ¼ svdðBÞ and B ¼ D q0 ðtÞ~ q1 ðtÞT dt. If Z   
q ðtÞ
d~
the detðBÞ < 0, then the last column of V T changes r HO ðuÞ ¼ vðtÞ; 2 1 uðtÞ þ q~1 ðtÞuðtÞ dt:
D dt
sign before multiplication.
In the case of Cc , the update uses the gradient of Since T1 ðÞ is an infinite-dimensional space, we
H . The tangent space to the rotation orbit at identity can approximate the gradient of HO with respect to
the -component by considering a finite-dimen-
i s fA~ q1 jA 2 IRnn ; A þ AT ¼ 0g. L e t E1 ; E2 ; . . . ;
sional subspace of T1 ðÞ, as follows: Form a
Enðn1Þ=2 be an orthonormal basis for the space of subspace of T1 ðÞ ¼ ff : D ! IRjhf; 1i ¼ 0g using
n  n skew-symmetric matrices. The gradient up- fðp1ffiffi sinð2ntÞ; p1ffiffi cosð2ntÞÞjn ¼ 1; 2; . . . ; m=2g. Then,
dates for rotation are performed by projecting v in approximate the partial
P P derivative of H with
this space to obtain A ¼ i hEi q~1 ; viEi and updating respect to using c ¼ m i¼1 r HO ðci Þci , where the
using Omþ1 ¼ e o A q~1 for a step size o > 0. ci s are the basis elements of that subspace. Then,
2. Reparameterizations: In case of Co , the optimization update the component according to 1 7! kþ1 
c
over HO can be performed using the DP algorithm cosð g kckÞ1 þ sinðð g kckÞ kck for a step size g > 0.
but for Cc , we use the following gradient iteration. Since  is a hypersphere, this update is simply the
exponential map on that sphere, at the point 1, and
We seek the incremental mþ1 that minimizes HO .
applied to the tangent vector R t c. This mþ1 in turn
There are two possibilities: One is to take the
gives mþ1 ðtÞ ¼ mþ1 ð0Þ þ 0 mþ1 ðsÞ2 ds, and thus
gradient of HO ððmþ1Þ Þ directly with respect to mþ1  ðmþ1Þ .
and use it to update  ðmþ1Þ . The other possibility, the
We can now state the algorithm for computing geodesics on
one that we have used in this paper, is to use a shape spaces.
square-root representation of _ that often simplifies
pffiffiffiffiffiffiffiffiffiffi Shape Geodesic Algorithm. Find a geodesic between
its analysis. Define mþ1 ¼ _ mþ1 and reexpress
shapes of two parameterized curves 0 and 1 in S (S o or
mþ1 as the pair ðmþ1 ð0Þ; mþ1 Þ. With a slight abuse
S c ). Compute the representations of each curve in C; denote
of notation, let HO be a function of ðmþ1 ð0Þ; mþ1 Þ. them by q0 and q1 , respectively. Set q~1 ¼ q1 .
Note that the space  of all -functions is the unit
hypersphere in IL2 ðD; IRÞ. We initialize with 1. Compute the geodesic  between q0 and q~1 in the preshape
0 ðtÞ ¼ t, with the corresponding representation space. For Co , use the analytical expression, while for Cc ,
being ð0; 1Þ and 1 being the constant function with use the path straightening algorithm given in the previous
value one. At the iteration m, we take the gradients section.
2. Removal of nuisance variables:
of HO , with respect to mþ1 ð0Þ and mþ1 , and update
these individually. The derivative with respect to
R a. Rotation: For Co , use the SVD-based solution ((8)).
mþ1 ð0Þ, evaluated at ð0; 1Þ, is @@H O
mþ1 ð0Þ
¼ D hvðtÞ; For Cc , compute A, the derivative of H with respect
q1 ðtÞ
d~
dt idt. To obtain the derivative with respect to
to SOðnÞ, and form the rotation update Omþ1 .
mþ1 , consider the sequence of maps b. Reparameterization: For Co , one can use the DP
Rt 2 algorithm. More generally, compute the derivatives of
0
ðsÞ ds  HO with respect to mþ1 and mþ1 ð0Þ, and for the
7!
 7! r; mþ1 .
reparameterization, updatepffiffiffiffiffiffiffiffiffiffi
pffiffiffi 3. Update q~1 7! Omþ1 ð~ q1  mþ1 Þ _ mþ1 .
where r  ðÞ ¼ ð~ q1  Þ _ , as shown in Fig. 3b.
For the constant function 1 2  and a tangent 4. If the norms of the increments are small, then stop. Else,
return to step 1.
u 2 T1 ðÞ, the differential of the first mapping at 1
Rt The two rows in Fig. 4 shows two examples of optimiza-
is uðtÞ 7! 2 uðtÞ ¼ 2 0 uðsÞds and, for a tangent
tion over . In each case, we start with a parameterized curve,
w 2 Tid ðÞ, the differential of the second mapping
shown in Fig. 4a and represented by q1 , generate a random
at id is w 7!  ðwÞ  d~ q1 1
dt w þ 2 q _ Concatenating
~1 w.  2  (shown in Fig. pffiffiffi4b), and form a reparameterized curve
these two linear maps, we obtain the directional using q0 ¼ ðq1  Þ _ (shown in Fig. 4c). Then, we use the
partial derivative of HO in a direction u 2 T1 ðÞ as gradient approach described above to find an optimal

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
SRIVASTAVA ET AL.: SHAPE ANALYSIS OF ELASTIC CURVES IN EUCLIDEAN SPACES 1423

TABLE 1
Timing Analysis of Gradient-Based Reparameterization
and Comparison with DP Algorithm

reparameterization of q1 that best matches this q0 by

minimizing the cost function HO . The evolution of the cost
function HO is shown in Fig. 4d, and the final reparameter-
ized curve q~1 is shown in Fig. 4e. In these examples, since q0 is
simply a reparameterization of q1 , the minimum value of HO
should be zero. Note that in the top row, where the original 
is closer to the identity, the cost function goes to zero, but in Fig. 5. Examples of planar elastic geodesics.
the bottom case, where  is rather drastic, the algorithm
converges to a final value of H that is not close to zero. We and 3D shapes [24], shape classification of point clouds [29],
conjecture that this can be mitigated by an improved and joint gait-cadence analysis for human identification in
numerical implementation of the basic procedure. videos [11].
To illustrate the strengths and limitations of a gradient-
based approach with respect to a common DP algorithm [7],
5.1 Shapes Analysis of 3D Helices
[26], we present a comparison of computational costs (using As the first example, we will study shapes of helical
Matlab on a 2.4 GHz Intel processor) and performance in curves in IR3 by matching and deforming one into
Table 1. In this experiment, we consider the shape space S o another. One motivation for studying shapes of cylindrical
since DP is not applicable for optimization in the case of helices comes from protein structure analysis. A primary
closed curves. The computational complexity of the gradient structure in a protein is a linked chain of carbon, nitrogen,
approach is OðT mkÞ, where T is the number of samples on the and oxygen atoms known as the backbone, and the
curve, k is the number of basis functions, and m is the number geometry of the backbone is often a starting point in the
of iterations, while that of DP algorithm is OðT 2 Þ. The table is structural analysis of proteins. These backbones contain
generated for T ¼ 100 and m ¼ 200. As a measure of certain distinct geometrical pieces; one prominent type is
matching performance, we also present the relative final cost the so-called -helix. In analyzing shapes of backbones, it
as a percentage (ðHO ðfinalÞ=HO ðinitialÞÞ  100). This table seems important to match not only their global geometries
shows that while the DP algorithm is very accurate in but also the local features (such as -helices) that appear
estimating the unknown , its computational cost is along these curves. We suggest the use of elastic shape
relatively high. One gets to solutions, albeit approximate, analysis of curves as a framework for studying shapes of
much faster when using the gradient method. An important protein backbones and present some results involving
limitation of the gradient method is that its solution is both synthetic and real data.
always local. Shown in Fig. 7 are two examples of geodesics between
Fig. 5 shows some elastic geodesics between several pairs some cylindrical helices. In each case, Fig. 7a and Fig. 7b show
of shapes. We have drawn ticks on these curves to show the two helices, and Fig. 7c is the optimal matching between them
optimal reparametrizations. The spacings between the ticks obtained using the estimated  function shown in Fig. 7d. The
are uniform in the leftmost shapes (q0 ) but have been resulting geodesic paths in S o between these curves are
adjusted for the other shapes during the minimization of H. shown in the bottom row. It is easy to see the combination of
The reader can see that the combinations of bending and bending and stretching/compression that goes into
stretching used in these deformations are successful in the
deforming one shape into another. In the left example,
sense that geometrical features are well-preserved.
where the turns are quite similar and the curves differ
Fig. 6 compares the elastic geodesics in S c with the
nonelastic method of Klassen et al. [14], where the only in the placements of these turns along the curve, a
representation is restricted to arc-length parameterizations.
The resulting deformation is purely bending and no
stretching is allowed. We observe that the elastic shape
analysis results in a better matching of features across shapes
and a more natural deformation along the geodesic path.

5 APPLICATIONS
In this section, we illustrate the proposed elastic shape
analysis using some applications. Some additional applica- Fig. 6. In each case, the top row shows a nonelastic geodesic ([14]),
tions have been presented elsewhere: symmetry analysis of 2 while the bottom rows the elastic geodesic between the same shapes.

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
1424 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 33, NO. 7, JULY 2011

Fig. 7. (a), (b) Original curves, (c) optimal registration between them,
and (d) optimal  . Bottom: corresponding geodesic paths.
Fig. 10. (a) Two facial surfaces represented by indexed collections of
facial curves. (b) Geodesics between shapes of corresponding curves.

Fig. 8. A set of helices with different numbers and placements of spirals

and their clustering using the elastic distance function. Fig. 11. Elastic geodesics between facial profiles and the corresponding
matching function in the far right.

5.2 3D Face Recognition

Human face recognition is a problem of great interest in
homeland security, client access systems, and several other
areas. Since recognition performance using 2D images has
been limited, there has been a push toward using shapes of
facial surfaces, obtained using weak laser scanners, to
recognize people. The challenge is to develop methods
and metrics that succeed in classifying people despite
changes in shapes due to facial expressions and measure-
Fig. 9. Two proteins: (a) 1CTF and (b) 2JVD and the elastic geodesic ment errors. Samir et al. [23], [31] have proposed an
between their shapes. approach that: 1) computes a function on a facial surface
as the shortest path distance from the tip of the nose (similar
simple stretching/compression is sufficient to deform one to [3], [21]), 2) defines facial curves to be the level curves of
into another. However, in the right example, where the that function, and 3) represents the shapes of facial surfaces
number of turns is different, the algorithm requires both using indexed collections of their facial curves. Fig. 10a
bending and stretching. shows two facial surfaces overlaid with facial curves. These
Fig. 8 shows an example of using the elastic distances facial curves are closed curves in IR3 and their shapes are
invariant to rigid motions of the original surface. We
between curves for clustering and classification. In this
compare shapes of facial surfaces by comparing shapes of
example, we study 12 cylindrical helices that contain
the corresponding facial curves, using geodesics between
different numbers, radii, and placements of turns. The first
them in S c . As an example, Fig. 10b shows geodesics in S c
three helices have only one turn, the next three have two
between the two sets of facial curves. For display, these
turns, and so on. Using the elastic geodesic distances intermediate curves have been rescaled and translated to the
between them in S o and the dendrogram clustering original values and, through reconstruction, they result in a
program in Matlab, we obtain the clustering shown in the geodesic path such that points along that path approximate
right panel. This clustering demonstrates the success of the full facial surfaces. These geodesic paths can be used to
proposed elastic metric in that helices with similar numbers compute average faces or facial parts, or to define metrics for
of turns are clustered together. human recognition [5].
Finally, in Fig. 9, we present an example of comparing Another example of elastic shape analysis of faces, this
real protein backbones. In this experiment, we use two time using facial profiles, is shown in Fig. 11.
simple proteins—1CTF and 2JVD—that contain three and
two -helices, respectively. The top row of this figure shows 5.3 Elastic Models for Planar Shapes
depictions of the two backbones, while the bottom row An important application of this elastic shape framework is in
shows the geodesic path between them in S o . These results developing probability models for capturing the variability
suggest a role for elastic shape analysis in protein structure present in the observed shapes. For example, the left panel
analysis. Additional details and experiments are presented of Fig. 12 shows examples of 20 observed 2D shapes of a
in [16]. “runner” taken from the Kimia database. Our goal is derive

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
SRIVASTAVA ET AL.: SHAPE ANALYSIS OF ELASTIC CURVES IN EUCLIDEAN SPACES 1425

P
form Pa random vector i zi Ui and then the random shape
exp ð i zi Ui Þ. The right panel of Fig. 12 shows 20 such
random shapes. It is easy to see the success of this wrapped
model in capturing the shape variability exhibited in the
original 20 shapes.

5.4 Transportation of Shape Deformations

One difficulty in using images for recognizing 3D objects is
that their 2D appearance changes with viewing angle. Since a
large majority of imaging technology is oriented toward 2D
images, there is a striking focus on planar shapes, their
analysis, and modeling, despite the viewing variability.
Within this focus area, there is an interesting problem of
predicting shapes of 3D objects from novel viewing angles.
(The problem of predicting full appearances, using pixels, has
Fig. 12. The left panel shows a set of 20 observed shapes of a “runner”
been studied by Savarese and Li [25] and others.) Our solution
from the Kimia data set. The middle panel shows their Karcher mean, to the problem of shape prediction is the following: If we
and the right panel shows a random sample of 20 shapes from the know how a known object deforms under a viewpoint
learned wrapped nonparametric model on S c . The bottom three rows change, perhaps we can apply the “same” deformation to a
show eigenvariations of shapes in three dominant directions around the similar (yet novel) object and predict its deformation under
mean, drawn from negative to positive direction and scaled by the
corresponding eigenvalues. the same viewpoint change. The basic technical issue is to be
able to transport the required deformation from the first
object to the second object before applying that deformation.
a probability model on the shape space S c so that we can
Since shape spaces are nonlinear manifolds, the deformations
use this model in future inferences. Using ideas presented
of one shape cannot simply be applied to another.
in earlier papers [6], [30], we demonstrate a simple model
The mathematical statement of this problem is as
where we: 1) first compute the sample Karcher mean [10] of
follows: Let ½q1a  and ½q1b  be the shapes of an object O1 when
the given shapes, 2) learn a probability model on the
viewed from two viewing angles a and b , respectively.
tangent space (at the mean) by mapping the observations to
The deformation in contours in going from ½q1a  to ½q1b 
that tangent space, and 3) wrap the probability model back
depends on some physical factors, the geometry of O1 and
to S c using the exponential map. In this paper, we the viewing angles involved. Consider another object O2
demonstrate the model using random sampling: Random which is similar but not identical to O1 in geometry. Given
samples are generated in the tangent space and mapped its shape ½q2a  from the viewing angle a , our goal is to
back to S c . P predict its shape ½q2b  from the viewing angle b . Our solution
Let  ¼ argmin½q2Sc ni¼1 ds ð½q; ½qi Þ2 be the Karcher mean is based on taking the deformation that deforms ½q1a  to ½q1b 
of the given shapes q1 ; q2 ; . . . ; qn , where ds is the geodesic and applying it to ½q2a  after some adjustments.
distance on S c . The Karcher mean of the 20 observed shapes
is shown in the middle panel of Fig. 12. Once we have  we 1. Let 1 ðÞ be a geodesic between ½q1a  and ½q1b  in S c
can map ½qi  into T ðS c Þ using the inverse exponential map and v1  _1 ð0Þ 2 T½q1a  ðS c Þ be its initial velocity.
½qi  7! vi  exp1
 ð½qi Þ. Since the tangent space is a vector 2. We need to transport v1 to ½q2a ; this is done using
space, we can perform more standard statistical analysis. forward parallel translation. Let 12 ðÞ be a geodesic
The infinite dimensionality of T ðS c Þ is not a problem since from ½q1a  to ½q2a  in S c . Construct a vector field wðtÞ such
we usually have only a finite number of observations. For that wð0Þ ¼ v1 and Dw d ¼ 0 for all points along 12 . This
instance, one can perform PCA on the set fvi g to find is accomplished in practice using Algorithm 2 in
dominant directions and associated observed variances. Section 4.2. Then, v2  wð1Þ 2 T½q2a  ðS c Þ is a parallel
One can study these dominant directions of variability as translation of v1 .
shapes by projecting vectors along these directions to the 3. Construct a geodesic starting from ½q2a  with v2 as the
shape space. Let ði ; Ui Þ be the singular values and singular initial velocity.
directions in the tangent space. Then, the mapping Fig. 13 shows two examples of this idea. In the top case, a
i Ui 7! exp ði Ui Þ helps visualize these principal modes hexagon (½q1a ) is deformed into a square (½q1b ) using an
as shapes. The three principal components of the 20 given elastic geodesic; this deformation is then transported to a
shapes are given in the lower three rows of Fig. 12, each row circle (½q2a ) and applied to it to result in the prediction ½q2b . A
displaying some shapes from  ¼ 1 to  ¼ 1. similar transport is carried out in the bottom example.
In terms of probability models, there are many choices Next, we consider an experiment involving the M60 tank
available. For the coefficients fzi g defined with respect to the as O1 and the T72 as O2 . Given shapes for different
basis fUi g, one can use any appropriate model from multi- azimuthal pose (fixed elevation) of M60 and one azimuth
variate statistics. In this experiment, we try a nonparametric for the T72, we would like to predict shapes for the T72
approach where a kernel density estimator, with a Gaussian from the other azimuthal angles. Since both objects are
kernel, is used for each coefficient zi independently. One of tanks, they have similar but not identical geometries. For
the ways to evaluate this model is to generate random instance, both have mounted guns but the T72 has a longer
samples from it. Using the inverse transform method to gun than the M60. In this experiment, we select a ¼ 0 and
sample zi s from their estimated kernel densities, we can predict the shape of the T72 for several b . The results are

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
1426 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 33, NO. 7, JULY 2011

TABLE 2
Classification Rate with (Bold Fonts) and without (Normal Fonts)
Use of Predicted Shapes for the T72

Fig. 13. In each case, a geodesic from the template shape (hexagon) to then perform prediction for a novel view using joint
the training shape (a) and deformation of the test shape (circle) with the
information from these views.
transported deformation (b).

shown in Fig. 14. The first and the third rows show the 6 SUMMARY
shapes for ½q1a  and ½q2a , respectively, the shapes for the M60 We have presented a new representation of curves that
and the T72 looking from head on. The second row shows facilitates an efficient elastic analysis of their shapes and is
½q1b  for different b given in the last column, while the fourth applicable to IRn for all n. Its most important advantage is
row shows the predicted shapes for the T72 from those b . that the elastic metric reduces to a simple IL2 metric.
How can we evaluate the quality of these predictions? Geodesics between shapes of closed curves are computed
We perform a simply binary classification with and without using a path straightening approach. This framework is
the predicted shapes and compare results. Here is the illustrated using several applications: shape analysis of
experimental setup. We have 62 and 59 total azimuthal helical curves in IR3 with applications in protein backbone
views of the M60 and the T72, respectively. Of these, we structure analysis, shapes of 3D facial curves with applica-
randomly select 31 views of M60 and one view of the T72 as tions in biometrics, wrapped probability models for
the training data; the remaining 31 (58) views of the M60 capturing shape variability, and parallel transport of
(the T72) are used for testing. The classification results, deformation models to predict shapes of 3D objects from
using the nearest neighbor classifier and the elastic distance novel viewpoints.
ds ((2)), are shown in the table below. While the classifica-
tion for the M60 is perfect, as expected, the classification for APPENDIX
the T72 is 46.55 percent. (Actually, this number is somewhat
higher than expected—we would expect a smaller perfor- Proof that Cc Is a Submanifold of Co . This proof is based on
mance with only one training shape.) Now, we generate [15, pp.R 25-27]. Let G : Co ! IRn be a map defined as
additional 31 shapes for the T72 using the prediction GðqÞ ¼ S1 qðtÞkqðtÞkdt. First, we need to check that its
method described earlier. Using the 31 training shapes of differential, dGq : Tq ðCo Þ ! IRn , is surjective at every
n
the M60, we generate 31 corresponding shapes of the T72 q 2 G1 ð0Þ;
R 0 2 IR is the origin. For the ith component
using parallel transport. The a used here was 90 . The Gi ðqÞ ¼ S1 qi ðtÞkqðtÞkdt, i ¼ 1; 2; . . . ; n, its directional
classification result after including the 31 predicted shapes derivative in a direction w 2 IL2 ðS1 ; IRn Þ is given by
is found to be 60.34 percent, a 15 percent increase in the Z
performance when using shape predictions. We performed qi ðtÞ
dGi ðwÞ ¼ hwðtÞ; qðtÞ þ kqðtÞkei idt;
the same experiment for another azimuth, a ¼ 0 , and the S1 kqðtÞk
results are listed under experiment 2 in Table 2. In this case, where ei is a unit vector in IRn along the ith coordinate
we improve the classification performance from 6.8 to
axis. To show that G is surjective, we need to show the
17.2 percent, an increase of almost 11 percent, using the qi ðtÞ
functions fkqðtÞk qðtÞ þ kqðtÞkei ; i ¼ 1; 2; . . . ; ng are linearly
predicted shapes of the T72. While this experiment was
performed with only one training shape, one can repeat this independent. Suppose not. This implies that there exists
idea using multiple given shapes for the novel object and a constant vector b ¼ ðb1 ; b2 ; . . . ; bn Þ such that, for all t,
P qi ðtÞ
i bi ðkqðtÞk qðtÞ þ kqðtÞkei Þ ¼ 0. This in turn implies that
for all t, qðtÞ is in the same direction as a constant vector
Pn
i¼1 bi ei . This proves that for any q function that does
not lie in a single 1D subspace, the mapping G is
surjective. So, the space Cc is a manifold except at those
points. These exceptional functions correspond to curves
that lie entirely in a straight line in IRn . This collection of
curves is a “very small” subset of Co , and we conclude
that G is a submersion at the remaining points of G1 ð0Þ.
Therefore, using [15], Cc is a codimension-n submanifold
Fig. 14. Shape predictions for novel pose. In each column, the first two of Co , for all points except those in this measure zero
are given shapes of the M60 from a ¼ 0 and b . The deformation subset. We will ignore this subset since there is
between these two is used to deform the T72 shape in the third row and
obtain a predicted shape in the fourth row. The accompanying pictures essentially a zero probability of encountering it in real
show the true shapes of the T72 at those views. problems. We conclude that Cc , with the earlier proviso,

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
SRIVASTAVA ET AL.: SHAPE ANALYSIS OF ELASTIC CURVES IN EUCLIDEAN SPACES 1427

is a submanifold of the Hilbert space Co , and thus, [3] A.M. Bronstein, M.M. Bronstein, and R. Kimmel, “Three-Dimen-
sional Face Recognition,” Int’l J. Computer Vision, vol. 64, no. 1,
IL2 ðS1 ; IRn Þ. u
t pp. 5-30, 2005.
[4] 
N.N. Cencov, Statistical Decision Rules and Optimal Inferences. Am.
Proof of Theorem 1. Define a variation of  to be a smooth Math. Soc., 1982.
function, hð; sÞ with h : ½0; 1  ð
;
Þ ! H such that [5] H. Drira, B. Ben Amor, A. Srivastava, and M. Daoudi, “A
Riemannian Analysis of 3d Nose Shapes for Partial Human
hð; 0Þ ¼ ðÞ for all  2 ½0; 1. The variational vector field Biometrics,” Proc. IEEE Int’l Conf. Computer Vision, 2009.
corresponding to h is given by vðÞ ¼ hs ð; 0Þ, where s [6] I.L. Dryden and K.V. Mardia, Statistical Shape Analysis. John Wiley
denotes the second argument in h. Thinking of h as a & Sons, 1998.
[7] M. Frenkel and R. Basri, “Curve Matching Using Fast Marching
path of curves in H, indexed by s, we define EðsÞ as the Method,” Proc. Fourth Int’l Workshop Energy Minimization Methods
energy of the curve obtained by restricting h to in Computer Vision and Pattern Recognition, pp. 35-51, 2003.
R1 [8] S.H. Joshi, E. Klassen, A. Srivastava, and I. Jermyn, “Removing
½0; 1  fsg. That is, EðsÞ ¼ 12 0 hh ð; sÞ; h ð; sÞid. We Shape-Preserving Transformations in Square-Root Elastic (SRE)
now compute Framework for Shape Analysis of Curves,” Proc. Sixth Int’l Conf.
Energy Minimization Methods in Computer Vision and Pattern
Z 1  Recognition, pp. 387-398, 2007.
_ Dh
Eð0Þ ¼ ð; 0Þ; h ð; 0Þ d [9] S.H. Joshi, E. Klassen, A. Srivastava, and I.H. Jermyn, “A Novel
0 ds Representation for Riemannian Analysis of Elastic Curves,” Proc.
Z 1  Z 1  IEEE Conf. Computer Vision and Pattern Recognition, pp. 1-7, 2007.
Dhs Dv d
¼ ð; 0Þ; h ð; 0Þ d ¼ ðÞ; ðÞ d [10] H. Karcher, “Riemannian Center of Mass and Mollifier Smooth-
0 d 0 d d ing,” Comm. Pure and Applied Math., vol. 30, no. 5, pp. 509-541,
1977.
since h ð; 0Þ is simply d d ðÞ. Now, the gradient of E [11] D. Kaziska and A. Srivastava, “Joint Gait-Cadence Analysis for
_
should be a vector field u along  such that Eð0Þ ¼ hhv; uii. Human Identification Using an Elastic Shape Framework,” Comm.
R 1 Dv Du Statistics—Theory and Methods, vol. 39, no. 10, pp. 1817-1831, 2010.
_
That is, Eð0Þ ¼ hvð0Þ; uð0Þi þ 0 h d ; d id. From this ex- [12] D.G. Kendall, “Shape Manifolds, Procrustean Metrics and Com-
pression, it is clear that u must satisfy the initial condition plex Projective Spaces,” Bull. of the London Math. Soc., vol. 16, no. 2,
pp. 81-121, 1984.
uð0Þ ¼ 0 and the ordinary (covariant) differential equation [13] M. Kilian, N.J. Mitra, and H. Pottmann, “Geometric Modeling in
Du d
d ¼ d . u
t Shape Space,” Proc. ACM SIGGRAPH, 2007.
[14] E. Klassen, A. Srivastava, W. Mio, and S.H. Joshi, “Analysis of
Proof of Lemma 3. Suppose v 2 T ðH0 Þ (i.e., vð0Þ ¼ Planar Shapes Using Geodesic Paths on Shape Spaces,” IEEE
vð1Þ ¼ 0), and w 2 T ðHÞ is covariantly linear. Then, Trans. Pattern Analysis and Machine Intelligence, vol. 26, no. 3,
using (covariant) integration by parts pp. 372-383, Feb. 2004.
[15] S. Lang, Fundamentals of Differential Geometry. Springer, 1999.
Z 1  [16] W. Liu, A. Srivastava, and J. Zhang, “Protein Structure Alignment
DvðÞ DwðÞ Using Elastic Shape Analysis,” Proc. ACM Conf. Bioinformatics and
hhv; wii ¼ ; d
0 d d Computational Biology, Aug. 2010.
  Z 1   [17] A.C.G. Mennuci, Metrics of Curves in Shape Optimization and
DwðÞ 1 D DwðÞ Analysis. 2009.
¼ v;  vðÞ; d ¼ 0:
d 0 0 d d [18] P.W. Michor and D. Mumford, “Riemannian Geometries on
Spaces of Plane Curves,” J. European Math. Soc., vol. 8, pp. 1-48,
Hence, T ðH0 Þ is orthogonal to the space of covariantly 2006.
[19] J.W. Milnor, Topology from the Differentiable Viewpoint. Princeton
linear vector fields along  in T ðHÞ. This proves that the Univ. Press, 1997.
space of covariantly linear vector fields is contained in [20] W. Mio, A. Srivastava, and S.H. Joshi, “On Shape of Plane Elastic
Curves,” Int’l J. Computer Vision, vol. 73, no. 3, pp. 307-324, 2007.
the orthogonal complement of T ðH0 Þ. To prove that [21] I. Mpiperis, S. Malassiotis, and M.G. Strintzis, “3-D Face
these two spaces are equal, observe first that, given any Recognition with the Geodesic Polar Representation,” IEEE Trans.
choice of tangent vectors at ð0Þ and ð1Þ, there is a Information Forensics and Security, vol. 2, no. 3, pp. 537-547, Sept.
2007.
unique covariantly linear vector field interpolating them. [22] R.S. Palais, “Morse Theory on Hilbert Manifolds,” Topology, vol. 2,
It follows that every vector field along  can be uniquely pp. 299-349, 1963.
[23] C. Samir, A. Srivastava, and M. Daoudi, “Three-Dimensional Face
expressed as the sum of a covariantly linear vector field Recognition Using Shapes of Facial Curves,” IEEE Trans. Pattern
and a vector field in T ðH0 Þ. The lemma follows. u
t Analysis and Machine Intelligence, vol. 28, no. 11, pp. 1858-1863,
Nov. 2006.
[24] C. Samir, A. Srivastava, M. Daoudi, and S. Kurtek, “On Analyzing
Symmetry of Objects Using Elastic Deformations,” Proc. Int’l Conf.
Computer Vision Theory and Applications , Feb. 2009.
ACKNOWLEDGMENTS [25] S. Savarese and F.-F. Li, “View Synthesis for Recognizing Unseen
This work was partially supported by the US Air Force Poses of Object Classes,” Proc. 10th European Conf. Computer Vision,
2008.
Office of Scientific Research FA9550-06-1-0324, the US [26] T.B. Sebastian, P.N. Klein, and B.B. Kimia, “On Aligning Curves,”
Office of Naval Research N00014-09-1-0664, and the US IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 25, no. 1,
National Science Foundation (NSF) DMS-0915003, and by pp. 116-125, Jan. 2003.
the INRIA/FSU Associated Team “SHAPES.” [27] J. Shah, “An H2 Type Riemannian Metric on the Space of Planar
Curves,” Proc. Workshop Math. Foundations of Computational
Anatomy, 2006.
[28] A. Srivastava, I. Jermyn, and S.H. Joshi, “Riemannian Analysis of
REFERENCES Probability Density Functions with Applications in Vision,” Proc.
[1] S. Amari, Differential Geometric Methods in Statistics, vol. 28. IEEE Conf. Computer Vision and Pattern Recognition, pp. 1-8, June
Springer, 1985. 2007.
[2] A. Bhattacharya, “On a Measure of Divergence between Two
Statistical Populations Defined by Their Probability Distribu-
tions,” Bull. of Calcutta Math. Soc., vol. 35, pp. 99-109, 1943.

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.
1428 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 33, NO. 7, JULY 2011

[29] A. Srivastava and I.H. Jermyn, “Looking for Shapes in Two- Shantanu H. Joshi received the BE degree in
Dimensional, Cluttered Point Clouds,” IEEE Trans. Pattern electronics and telecommunication in 1998 from
Analysis and Machine Intelligence, vol. 31, no. 9, pp. 1616-1629, the University of Pune in India. He also received
Sept. 2009. the MS and PhD degrees in electrical engineer-
[30] A. Srivastava, S.H. Joshi, W. Mio, and X. Liu, “Statistical Shape ing from Florida State University. He is currently
Anlaysis: Clustering, Learning and Testing,” IEEE Trans. Pattern a postdoctoral research fellow in the Laboratory
Analysis and Machine Intelligence, vol. 27, no. 4, pp. 590-602, Apr. of Neuro Imaging, Department of Neurology,
2005. University of California, Los Angeles.
[31] A. Srivastava, C. Samir, S.H. Joshi, and M. Daoudi, “Elastic Shape
Models for Face Analysis Using Curvilinear Coordinates,” J. Math.
Imaging and Vision, vol. 33, no. 2, pp. 253-265, Feb., 2009.
[32] G. Sundaramoorthi, A.C.G. Mennucci, S. Soatto, and A. Yezzi, “A
New Geometric Metric in the Space of Curves, and Applications
to Tracking Deforming Objects by Prediction and Filtering,” 2010. Ian H. Jermyn received the BA Honors degree
[33] L. Younes, “Computable Elastic Distance between Shapes,” SIAM (First Class) in physics from Oxford University in
J. Applied Math., vol. 58, no. 2, pp. 565-586, 1998. 1986, and the PhD degree in theoretical physics
[34] L. Younes, P.W. Michor, J. Shah, D. Mumford, and R. Lincei, “A from the University of Manchester, United King-
Metric on Shape Space with Explicit Geodesics,” Matematica e dom, in 1991. After working for a total of three
Applicazioni, vol. 19, no. 1, pp. 25-57, 2008. years at the International Centre for Theoretical
[35] L. Younes, A. Qiu, R.L. Winslow, and M.I. Miller, “Transport of Physics in Trieste, Italy, he began study for a
Relational Structures in Groups of Diffeomorphisms,” J. Math. PhD degree in computer vision in the Computer
Imaging and Vision, vol. 32, no. 1, pp. 41-56, 2008. Science Department of the Courant Institute of
Mathematical Sciences at New York University,
Anuj Srivastava received the MS and PhD receiving the PhD degree in July 2000. He joined the Ariana Research
degrees in electrical engineering from Washing- Group at INRIA Sophia Antipolis as a postdoctoral researcher in August
ton University in St. Louis in 1993 and 1996, 2000. From September 2001 to August 2010, he was a senior research
respectively. After spending the year 1996-1997 scientist in the Ariana Group. Since September 2010, he has been a
at Brown University as a visiting researcher, he reader in statistics in the Department of Mathematical Sciences at the
joined Florida State University (FSU) as an University of Durham in the United Kingdom. His main research interests
assistant professor in 1997. He is a professor include the statistical modeling of shape and texture, and information
of statistics at FSU in Tallahassee. He has geometry as applied to inference.
received the Developing Scholar and the Grad-
uate Faculty Mentor Awards at FSU. His
research is focused on pattern-theoretic approaches to problems in . For more information on this or any other computing topic,
image analysis, computer vision, and signal processing. In particular, he please visit our Digital Library at www.computer.org/publications/dlib.
has developed computational tools for performing statistical inferences
on certain nonlinear manifolds and has published more than 130 journal
and conference articles in these areas.

Eric Klassen received the PhD degree from

Cornell University in 1987 in the field of low-
dimensional topology. After postdoctoral posi-
tions at Caltech and the University of California,
San Diego, he joined the Department of Mathe-
matics at Florida State University in 1991 as a
professor. He has worked in topology, geometry,
gauge theory, and Riemann surfaces, as well as
on applications to computer vision and pattern
recognition.

Authorized licensed use limited to: Universite de Geneve. Downloaded on March 30,2023 at 11:17:28 UTC from IEEE Xplore. Restrictions apply.