IEEE TRANS.

PAMI 1

Shape recognition with spectral distances

Michael M. Bronstein Member, IEEE, Alexander M. Bronstein Member, IEEE,

Abstract—Recent works have shown the use of diffusion geometry for various pattern recognition applications, including non-rigid
shape analysis. In this paper, we introduce spectral shape distance as a general framework for distribution-based shape similarity and
show that two recent methods for shape similarity due to Rustamov and Mahmoudi & Sapiro are particular cases thereof.

Index Terms—diffusion distance, commute time, spectral distance, eigenmap, Laplace-Beltrami operator, heat kernel, distribution,
global point signature, non-rigid shapes, similarity.

1 I NTRODUCTION In Section 4, we sketch standard approaches to the

Recent works [1], [9], [23], [15], [6], [16] have shown the discretization of diffusion geometry, and in Section 5
use of diffusion geometry for various pattern recognition show experimental results. Finally, Section 6 concludes
applications, including analysis of non-rigid shape and the paper.
surfaces. In particular, diffusion geometry arising from
heat propagation on a surface [24], [3] allows defining 2 BACKGROUND
shape similarity criteria which are intrinsic, i.e., invari- Let X be a shape, modeled as a compact Riemannian
ant to inelastic deformations of the surface. The latter manifold. In the following, we denote by µ the stan-
property is especially important in the comparison and dard area measure on X. The L2 norm of a function
retrieval of non-rigid shapes possessing a high degree f on X withZ respect to the measure µ is defined by
of flexibility [4], [22], [11]. Also, diffusion geometry kf kL2 (X) = f (x)dµ(x). Equipped with a metric d :
appears to be robust to topological noise [6], unlike other X
X×X → R+ , the pair (X, d) forms a metric space and
geometries, e.g., geodesic [5].
the triplet (X, d, µ) a metric measure space.
In this paper, we propose the spectral shape distance,
based on distributions of generic diffusion distances,
2.1 Diffusion kernels
and show that this framework generalizes previously
proposed approaches. In particular, our formulation al- A function k : X × X → R is called a diffusion kernel if it
lows to compare the method of Rustamov [23] (modeling satisfies the following properties:
shapes as distribution of commute time distances) and (K1) Non-negativity: k(x, x) ≥ 0.
of Mahmoudi and Sapiro [15] (modeling shapes as dis- (K2) Symmetry: k(x, y) = k(y, x).
tributions of diffusion distances) and show the relation (K3) Positive-semidefiniteness: for every bounded f ,
between them. Z Z
k(x, y)f (x)f (y)dµ(x)dµ(y) ≥ 0.
The rest of the paper is organized as follows. In Sec-
tion 2, we introduce the mathematical background (for a Z Z
comprehensive overview of the notions in diffusion ge- (K4) Square integrability: k 2 (x, y)dµ(x)dµ(y) < ∞.
Z
ometry in the continuous and discrete case, we refer the (K5) Conservation: k(x, y)dµ(y) = 1.
reader to [9], [12] and [20], respectively). In Section 3, we
The value of k(x, y) can be interpreted as a transition
present the construction of the spectral shape distances
probability from x to y by one step of a random walk
and discuss their invariance properties. A discussion is
on X.
dedicated to the comparison of the methods in [23], [15],
Diffusion kernel defines a linear operator
which are regarded as particular cases of our framework. Z
Kf = k(x, y)f (y)dµ(y), (1)
• The authors are with the Department of Computer Science, Technion –
Israel Institute of Technology, Haifa, 32000, Israel. which is known to be self-adjoint. Because of (K4), K
Email: bronstein@ieee.org, alexbron@ieee.org.
has a finite Hilbert norm and therefore is compact.
IEEE TRANS. PAMI 2

As the result, it admits a discrete eigendecomposition frequency, K(λ) can be thought of as the transfer function
Kψi = αi ψi with eigenfunctions {ψi }∞i=0 and eigenvalues of a low-pass filter.1 Using this signal processing analogy,
∞
{αi }i=0 . αi ≥ 0 by virtue of property (K3), and αi ≤ 1, the kernel k(x, y) can be interpreted as the point spread
by virtue of (K5) and consequence of the Perron-Frobenis function at a point y, and the action of the diffusion
theorem. operator Kf on a function f on X can be thought
By the spectral theorem, the diffusion kernel admits of as the application of the point spread function by
the following spectral decomposition property means of a non shift-invariant version of convolution.
∞
X The transfer function of the diffusion operator Kt is
k(x, y) = αi ψi (x)ψi (y). (2) K t (λ), which can be interpreted as multiple applications
i=0
of the filter K(λ). Such multiple applications decrease
Since ψi form an orthonormal basis of L2 (X), the effective bandwidth of the filter and, consequently,
Z Z X∞ increase its effective support in space. Because of this
k 2 (x, y)dµ(x)dµ(y) = αi2 , (3) duality, we will freely interchange between k(x, y) and
i=0
K(λ) referring to both as kernels.
a fact sometimes referred to as Parseval’s theorem. Using
An important particular case of is the heat oper-
these results, properties (K3–5) can be rewritten in the
∞
X ator Ht defined by the family of transfer functions
spectral form as 0 ≤ αi ≤ 1 and αi2 < ∞. H t (λ) = e−tλ and the associated heat kernels ht (x, y) =
i=0
P∞ −tλi
An important property of diffusion operators is the i=0 e φi (x)φi (y). The heat kernel ht (x, y) is the so-
fact that for every t ≥ 0, the operator Kt is also a diffu- lution of the heat equation with point heat source at x
sion operator with the eigenbasis of K and correspond- at time t = 0, i.e., the heat value at point y after time t.
t
ing eigenvalues {αit }∞
i=0 . The kernel of K expresses the
transition probability by random walk of t steps. This 2.3 Diffusion distances
allows to define a scale space of kernels, {kt (x, y)}t∈T ,
Since a diffusion kernel k(x, y) measures the degree of
with the scale parameter t.
proximity between x and y, it can be used to define a
metric
2.2 Heat diffusion
d2 (x, y) = kk(x, ·) − k(y, ·)k2L2 (X) , (6)
There exists a large variety of possibilities to define a
diffusion kernel and the related diffusion operator. Here, on X, dubbed the diffusion distance by Coifman and La-
we restrict our attention to operators describing heat fon [9], [12]. Another way to interpret the latter distance
diffusion. Heat diffusion on surfaces is governed by the is by considering the embedding Ψ : x 7→ L2 (X) by
heat equation, µ ¶ which each point x on X is mapped to the function
∂
∆X + u = 0, (4) Ψ(x) = k(x, ·). The embedding Ψ is an isometry between
∂t
X equipped with diffusion distance and L2 (X) equipped
where u is the distribution of heat on the surface and
with the standard L2 metric, since
∆X is the positive-semidefinite Laplace-Beltrami opera-
tor, a generalization of the Laplacian to non-Euclidean d(x, y) = kΨ(x) − Ψ(y)kL2 (X) . (7)
domains. For compact surfaces, the Laplace-Beltrami
Because of spectral duality, the diffusion distance can
operator has discrete eigendecomposition of the form
also be written as
∆X φi = λi φi .
∞
X
Diffusion operators associated with heat propagation
d2 (x, y) = K 2 (λi )(φi (x) − φi (y))2 . (8)
processes are diagonalized by the eigenbasis of the i=0
Laplace-Beltrami operator, namely K’s having ψi = φi .
Here as well we can define an isometric embedding
The corresponding diffusion kernels can be expressed as
Φ : x 7→ `2 with Φ(x) = {K(λi )φi (x)}∞ i=0 , termed as
∞
X the diffusion map by Lafon. The diffusion distance can be
k(x, y) = K(λi )φi (x)φi (y), (5)
i=0 cast as d(x, y) = kΦ(x) − Φ(y)k`2 .
where K(λ) is some function such that αi = K(λi ). Since 1. For an insightful discussion of the physical interpretation of the
the Laplace-Beltrami eigenvalues can be interpreted as Laplace-Beltrami operator, the reader is referred to [13].
IEEE TRANS. PAMI 3

The same way a diffusion operator Kt defines a scale and for p = ∞,

space, a family of diffusion metrics can be defined for
DL∞ (F, G) = sup |f − g|. (12)
t ≥ 0 as δ∈[0,∞)

d2t (x, y) = kΦt (x) − Φt (y)k2`2 (9) Kullback-Leibler dissimilarity:

X∞ Z ∞µ ¶
f g
= K 2t (λi )(φi (x) − φi (y))2 , DKL (F, g) = f log + g log dδ. (13)
i=0 0 g f

where Φt (x) = {K t (λi )φi (x)}∞ Bhattcharyya dissimilarity:

i=0 . Interpreting diffusion
Z ∞ p
processes as random walks, dt can be related to the
DBhatt (F, G) = − log f g dδ. (14)
“connectivity” of points x and y by walks of length t 0
(the more such walks exist, the smaller is the distance). χ2 dissimilarity:
Z ∞
(f − g)2
2.4 Distance distributions Dχ2 (F, G) = 2 dδ. (15)
0 f +g
Though diffusion metrics contain significant amount of Earth mover’s distance (EMD), which for normalized
information about the geometry of the underlying shape, univariate distributions is known to be equivalent to
direct comparison of metrics is problematic since it re- Z ∞
quires computation of correspondence between shapes. DEMD (F̂ , Ĝ) = |F̂ − Ĝ|dδ. (16)
0
A common way to circumvent the need of correspon-
dence is by representing a metric by its distribution, and 3 S PECTRAL SHAPE DISTANCES
measuring the similarity of two shapes by comparing the
In this paper, we propose the spectral distance distri-
distributions of the respective metrics.
butions framework for shape similarity, which will be
A metric d on X naturally pushes forward the product
shown to generalize several state-of-the-art approaches.
measure µ × µ on X × X (i.e., the measure defined by
Figure 1 schematically describes the proposed data
d(µ×µ)(x, y) = dµ(x)dµ(y)) to the measure F = d∗ (µ×µ)
flow. Given two shapes, the eigendecomposition of their
on [0, ∞) defined as F (I) = (µ × µ)({(x, y) : d(x, y) ∈ I})
Laplace-Beltrami operators is performed. The eigenval-
for every measurable set I ⊂ [0, ∞). F can be fully de-
ues and eigenfunctions are used to define a diffusion
scribed by means of a cumulative distribution function,
kernel k, which in turn, is used to define a family of
denoted by
Z δ Z scale-dependent diffusion distances between points on
F (δ) = dP = χd(x,y)≤δ dµ(x)dµ(y) (10) the shapes. Next, the distributions of pair-wise distances
0 are computed. The dissimilarity of two shapes is thus
with some abuse of notation (here χ is the indicator converted into dissimilarity of distributions. Multiple
function). F (δ) defined this way is the measure of pairs scales are merged into a single spectral shape distance
of points the distance between which in no larger than by means of an aggregation operation, which can be
δ; F (∞) = µ2 (X) is the squared area of the surface X. performed after either of the above stages (possible
The density function (histogram) can be defined as the locations of the integration operator are denoted in red
d
derivative f (δ) = dδ F (δ). Sometimes, it is convenient in the figure). We individuate four main components in
to work with normalized distributions, F̂ = F/F (∞) this framework:
and the corresponding density functions, fˆ, which can Kernel: According to (5), the diffusion kernel k(x, y) is
be interpreted as probabilities. described by the transfer function K(λ). The dependence
Using this idea, comparison of two metric measure on the scale parameter is introduced by taking the
spaces reduces to the comparison of measures on [0, ∞), powers K t . The kernel can be selected to satisfy certain
or equivalently, comparison of un-normalized or nor- invariance properties, as detailed in the next section.
malized distributions, which is carried out using one of Norm: A generalized diffusion distance dt (x, y) is
the standard distribution dissimilarity criteria used in defined as a cross-talk between the diffusion kernels,
statistics: dt (x, y) = kkt (x, ·) − kt (x, ·)k, using some norm. For the
Lp dissimilarity: for 1 ≤ p < ∞, choice of L2 (X), the diffusion distance has an explicit
µZ ∞ ¶1/p
p
expression according to (9). It is remarkable that some
DLp (F, G) = |f − g| dδ ; (11)
0
selections of transfer functions may result in kernels
IEEE TRANS. PAMI 4

Kernel Distance Distribution

{λi } {kt } {dt } {ft }
K(λ) A(kt ) k·k A(dt ) hist A(ft ) Dissimilarity
{Dt }
D A(Dt )
{gt }
K(λ) A(kt ) k·k A(dt ) hist A(gt )

Fig. 1. Data flow in the proposed approach: for each of the compared shapes, a transfer function K(λ) and the norm
k · k define a family of pair-wise distances. Shape dissimilarity is cast as distribution dissimilarity using the criterion D.
Aggregation of different scales can be performed at any of these stages as denoted in dashed red.

violating properties of (K1–5) and lead to distances that becomes

∞
are not metrics. Comparison of distributions of such 1 X
k 0 (x, y) = K(λ/α2 )φi (x)φi (y). (17)
distances may still be very useful. α2 i=0
Dissimilarity D(F, G) is used to compare the distance
When, for example, the Lp norm is used, the correspond-
distributions, as detailed in Section 2.4.
ing diffusion distance is given by
Aggregation: the specific way to aggregate the scales Z
and its position in the pipeline, as detailed in Section 3.2. d0 (x, y) = |k 0 (x, z) − k 0 (y, z)|p dµ0 (z) (18)
¯
Z ¯X ¯
∞ ¯p
¯ ¯
3.1 Invariance = ¯ α2/p−2 K(λ/α2 )φi (z)(φi (x) − φi (y))¯ dµ(z).
¯ ¯
i=0
The described framework is very generic and different In order to obtain a scale-invariant distance, the transfer
choices of the transfer function K and the scale integra- function K(λ) must satisfy K(λ/α2 ) = α2−2/p K(λ).
tion method lead to different shape similarity criteria. Functions satisfying this property are K(λ) = λ−1/p .
In practical applications, specific choices are driven by In the particular case p = 2, the kernel K(λ) = λ−1/2
desired invariance considerations: it is possible to design defines a distance
K in such a way that it is invariant to certain classes of Z X ∞
1 2
shape transformation. d2CT (x, y) = φi (z)(φi (x) − φi (y))2 dµ(z)
λ
i=1 i
Deformation invariance: Since the Laplace-Beltrami ∞
X 1
operator is intrinsic (i.e., fully expressible in terms of = (φi (x) − φi (y))2 (19)
the Riemannian metric of the surface), its eigenfunctions λ
i=1 i

and eigenvalues are invariant to inelastic bendings. As a (note that the singularity at λ0 = 0 can be removed
result, diffusion kernels based on φi , λi are deformation- starting the summation from i = 1, since φ0 = const).
invariant, and so are distributions of the related dis- Such a distance is known in spectral graph theory as
tances. the commute time distance, and can be thought of as the
Topological invariance: In many practical applica- average time of a walk starting at x to go through y
tions, shapes suffer from “topological noise”, resulting and return back to x (called the commute time, hence
R∞
in different local connectivity. This is a typical situation the name). Invoking the relation 0 e−λt dt = λ1 , the
when shapes are scanned by a 3D range camera. The commute time distance can interpreted as the integral
change in connectivity alters the diffusion distances by of the squared heat diffusion distance d2t over the entire
introducing “shortcuts”; however, these artifacts have time scale,
different effect at different scale. For large t, the influence Z ∞ X∞ Z ∞
2
of the shortcuts is smaller. dt (x, y) = e−2λi t dt(φi (x) − φi (y))2 (20)
0 i=0 0
Scale invariance: When a shape X 0 is obtained by X∞
1 1
uniformly scaling X by a factor α > 0, the eigenvalues = (φi (x) − φi (y))2 = d2CT (x, y).
2λ 2
of the Laplace-Beltrami operator are scaled according i=0

to λ0i = α−2 λi , and the corresponding L2 -normalized Equivalently, the commute time kernel KCT (λ) = λ−1/2 can
eigenfunctions become φ0i (x) = α−1 φi (x). A diffusion 2
be though of as the integral of the heat kernel, KCT (λ) =
R∞
kernel defined by a transfer function K(λ) therefore 0
H t (λ)dt.
IEEE TRANS. PAMI 5

3.2 Scale aggregation denotes some set of time scale. Such a distance has the
The derivation of the commute time kernel and the advantage of comparing shape properties at multiple
commute time distance from the corresponding heat scales, considering each scale separately.
kernel and diffusion distance can be interpreted as the dDSS is also spectral shape distance with the heat
application of a scale aggregation functional on {Ht (λ)}t>0 kernel K t (λ) = e−λt , diffusion distance defined with
or {dt (x, y)}t>0 . Examples of such aggregation function- the L2 (X) norm, and L2 distribution dissimilarity. The
als are aggregation is performed on distribution dissimilarities
P
Sum or, more generally, weighted sum A(ft ) = using sum t∈T over a finite set of scales T .
R
w(t)ft dt where w is some scale-dependent weighting Global point signature: In [23], Rustamov pro-
T
function and T is some range of scales. posed the global point signatures (GPS) representation for
Sum of squares or, more generally, sum of non- shapes based on the eigenmap of the form ΨX (x) =
R −1/2
linearities A(ft ) = T ξ(ft ) dt, where ξ is a scalar non- {λi φi (x)}∞i=1 . Given two shapes X, Y , Rustamov

linearity. looked at the corresponding ΨX and ΨY as Euclidean

Maximum: A(ft ) = supt∈T ft . objects with distances defined by dX (x, y) = kΨX (x) −
Here, ft denotes a family of scale-dependent objects. De- ΨX (y)kL2 (X) (similarly for dY ), and computed the dis-
pending on the stage at which aggregation is performed, tance between the shapes X, Y as the L2 dissimilarity of
aggregation functionals can be classified as the distributions of the distances dX , dY , dGPS (X, Y ) =
Kernel aggregation: A acting on {K t (λ)} or {kt (x, y)}. kfX − fY k2 . This method can be considered as an ap-
Distance aggregation: A acting on {dt (x, y)}. plication of the shape distributions approach [17] to the
Distribution aggregation: A acting on {Ft } or {ft }. embeddings ΨX , ΨY .
Dissimilarity aggregation: A acting on {D(Ft , Gt )}. The GPS distance can be cast as a particular setting
This resembles the spirit of the aggregation of multi-scale of the spectral shape distance with the commute time
dissimilarities of histograms-based descriptors proposed kernel K(λ) = λ−1/2 , L2 (X) norm, L2 distribution
in [14]. dissimilarity, and no scale aggregation, or, alternatively,
as a the spectral distance with the heat kernel, L2 (X)
3.3 Relation to other methods norm, and aggregation of diffusion distances using the
sum of squares on t ∈ [0, ∞).
Diffusion distance distributions: In [15], Mahmoudi
Considering all the three aforementioned methods
and Sapiro proposed applying the shape distribution
from the standpoint of our framework, we can clearly
approach to heat diffusion distances. Given two shapes
see their equivalence up to position of the scale inte-
X and Y with distribution densities fX and fY of the
gration operation. In dMS,t , the heat diffusion distance is
corresponding diffusion distances dX,t , dY,t , Mahmoudi
computed at a single scale t, then distribution distance is
and Sapiro defined a similarity criterion as dMS,t (X, Y ) =
computed, without doing any scale integration. In dGPS ,
kfX −fY k2 . This criterion can be regarded as a particular
integration is performed before computing the distribu-
case of our spectral shape distance, with the heat kernel
tion. In dDSS , integration is performed after computing
K t (λ) = e−λt , L2 (X) norm, and the L2 distribution
the distribution distances.
dissimilarity. The diffusion distances are taken at some
Scale invariant heat kernel signature: In [8], it was
single fixed scale t; no scale aggregation is performed.
observed that using the logarithmic scale τ = log t, the
Choosing different t allows describing small- and large-
heat kernel on X and its counterpart on X 0 uniformly
scale properties of the shape (for example, two shapes
scaled by a factor α > 0 are related by log kτ0 = log kτ +β +
can be similar at small scale and dissimilar on large
β, where β = −2 log α. Taking the derivative of log kτ
scales or vice versa). At the same time, the scale depends
with respect to τ undoes the additive constant β, while
on the shape size and the choice of t can often be
the shift in τ can be undone in the Fourier domain by
problematic [20].
canceling the phase. The resulting family of kernels,
Diffusion scale-space distance: In [7], we proposed µ¯ µ ¶¯¶
a way to address the problem of scale selection by ¯ ∂ ¯
k̂τ (x, y) = F −1 ¯¯F log kτ (x, y) ¯¯ , (21)
integrating the Mahmoudi-Sapiro distance on a set of ∂τ
scales, resulting in what was termed the diffusion scale- is scale invariant. From the point of view of the proposed
P
space distance, dDSS (X, Y ) = t∈T dMS,t (X, Y ), where T framework, the above transformation can be thought of
IEEE TRANS. PAMI 6

as a scale aggregation functional acting on the kernel. the Arnoldi algorithm used in MATLAB function eigs.
It can be further followed by selection of a single value This should be taken into consideration when choosing
of τ or aggregation of several scales in the previously the transfer function K(λ): ideally, it should decay fast
discussed way. Though the kernel k̂τ is not a valid enough to allow taking a small number of eigenvalues,
diffusion kernel (e.g., it can assume negative values), thus reducing the computation complexity. Specifically
the resulting “distance” is still informative. We intend in the particular cases discussed in Section 3.3, numerical
to investigate this interesting relation in future studies. approximation of the commute time distance (19) is less
favorable compared to the heat diffusion distance,since
the decay of the terms e−λt is faster than λ−1/2 for typical
4 N UMERICAL IMPLEMENTATION
choice of t.
For numerical computation, the shape X is represented
by a finite number of samples X̂ = {x1 , ..., xN } ⊆ X. The
5 R ESULTS
discrete approximation of the Laplace-Beltrami operator
has the following generic form We used a subset of 486 shapes from the Shape-
Google database [18], containing different shapes with
1 X
(∆X̂ u)i = wij (ui − uj ), (22) simulated transformations. The transformations include
ai j
null (no transformation), isometry (near-isometric bend-
where u = (u1 , ..., uN ) is a scalar function defined on ing), topology (connectivity change obtained by welding
X̂, wij are weights, and ai are normalization coeffi- some of the shape vertices), isometry+topology (com-
cients. In matrix notation, Equation (22) can be written bination of the previous two), triangulation (different
as ∆³X̂ u = A´−1 Lf , where A = diag(ai ) and L = meshing of the same shape), and partiality (missing
P
diag l6=i wil − (wij ) are sparse matrices. The eigen- information, obtained by making holes and cutting parts
values and eigenfunctions of the discretized Laplace- off the shape). Multiple instances of each transformation
Beltrami operator are computed by solving the gener- are present for each shape class.We performed a leave-
alized eigendecomposition problem AΦ = ΛLΦ, where Λ one-out retrieval experiment, in which each class of
is the (k + 1) × (k + 1) diagonal matrix of the smallest transformations was queried against a database con-
eigenvalues λ0 , ..., λk , and Φ is an N × (k + 1) matrix of taining the remaining transformations. Matches were
corresponding eigenfunctions [13]. consider correct between different transformations of
Discrete Laplace-Beltrami operator. Different dis- the same shape. The average precision (area under the
cretizations of the Laplace-Beltrami operator lead to precision-recall curve) was measured for each query
different choice of A and W [21]. For triangular meshes, shape, and the mean average precision (mAP) was used as
a popular choice is the cotangent weight scheme [19], in a single number quantifying the retrieval performance
which wij = cot αij + cot βij (αij and βij are the two for each class of transformations. As an additional qual-
angles opposite to the edge between vertices i and j ity measure, for each class of transformations the receiver
in the two triangles sharing the edge) for j in the 1- operating characteristic (ROC) curve was computed, rep-
ring neighborhood of vertex i and zero otherwise, and resenting a tradeoff between the percentage of similar
ai is one third of the area of the triangles sharing the shapes correctly identified as similar (true positives rate
vertex xi . This discretization preserves many important or TPR) and the percentage of dissimilar shapes wrong-
properties of the continuous Laplace-Beltrami operator, fully identified as similar (false positive rate or FPR). For
such as positive semi-definiteness, symmetry, and lo- additional details, the reader is referred to [18].
cality, and in addition it is numerically consistent. For Experiments. We performed several experiments to
shapes represented as point clouds, the Laplace-Beltrami reveal the influence of the choice of the different pro-
operator can be approximated using [2]. cessing stages in Figure 1. While these experiments
Approximation quality and complexity. In practice, do not attempt to entirely cover the entire space of
numerical inaccuracy due to the discretization of the possible choices, we believe they provide a sufficiently
Laplace-Beltrami operator imposes a limit on the number interesting insight about the importance of each ingredi-
k of eigenvalues and eigenvectors that can be accu- ent. We denote each choice as Kernel-Norm-Dissimilarity-
rately computed. Typical complexity of computing k first Aggregation, according to Section 3. As Kernel, we used
eigenvectors of a sparse symmetric matrix is O(N k) with K(λ) = e−λ (heat kernel, abbreviated as H), λ−1/2
IEEE TRANS. PAMI 7

TABLE 1
mAP (%) of Heat-L2 -Dissimilarity -Single for different distribution dissimilarities.

Transform. L1 L2 nL1 nL2 K-L Bhatt χ2 EMD

Null 96.19 95.59 96.19 95.82 93.96 94.81 95.10 93.83
Isometry 94.62 93.27 94.62 94.07 91.59 92.71 93.42 90.56
Iso+Topo 89.43 87.32 89.43 87.56 85.73 86.89 87.31 85.13
Partiality 90.49 90.92 90.48 91.38 84.74 88.62 90.06 78.78
Topology 83.40 83.82 83.39 83.73 65.80 68.57 70.07 69.31
Triang. 92.54 91.56 92.55 91.81 90.38 91.04 91.23 84.68
Noise 89.14 89.26 89.16 89.34 85.03 86.67 87.45 86.06
All 90.93 90.22 90.94 90.56 86.12 87.86 88.62 84.23

(commute time kernel, abbreviated as CT), and the Step TABLE 2

mAP (%) of Kernel-Normal-nL1 for different spectral distances.
function. As Norm, we used L1 , L2 , and L∞ . The Kernel-
Norm pair fully defines the spectral distance used in
Transform. H-L2 H-L1 H-L∞ CT-L2 Step-L2 Geo
the distribution computation. For reference, we also Null 96.19 95.41 94.61 94.41 96.35 94.85
show retrieval using distributions of geodesic distances Isometry 94.62 92.84 90.31 91.89 95.18 90.98
Iso+Topo 89.43 91.48 87.17 83.69 93.41 84.88
(denoted as Geo) as proposed in [10]. As Dissimilarity,
Partiality 90.48 82.27 82.85 78.05 87.68 84.29
we used the L1 , L2 , normalized L1 (nL1 ), normalized Topology 83.39 98.29 85.73 86.65 98.92 96.24
L2 (nL2 ), KL, Bhattcharyya, χ2 , and EMD, as detailed in Triang. 92.55 91.59 91.57 89.78 92.93 90.14
Noise 89.16 84.68 81.70 92.21 82.53 85.33
Section 2.4. For scale Aggregation, we tested a few simple
All 90.94 89.87 87.17 87.29 91.65 88.32
choices applied at the last stage aggregating distribution Scale 30.55 15.58 28.28 100.00 17.27 31.07
dissimilarities at different scales, using Single scale selec-
tion, Sum, Sum2 , and Max, as detailed in Section 3.2. TABLE 3
In all experiments, distributions were computed in mAP (%) of H-L2 -nL1 -Aggregation, for different aggregation
the interval from 0 to the 95-percentile of the distances functionals.
on the entire corpus of shapes. The Parzen window
Transform. t = 1024 t = 4096 Sum Sum2 Max
technique was used to produce the histograms with the
Null 96.19 93.07 96.42 96.54 96.51
1
window bandwidth set to 256 of the interval length. The Isometry 94.62 89.06 93.90 94.06 94.24
histograms were sampled on 256 evenly spaced bins. Iso+Topo 89.43 82.42 89.03 89.20 91.75
Partiality 90.48 80.22 88.75 89.06 90.37
Dissimilarity choice. In this experiment, we evalu- Topology 83.39 65.49 83.37 87.03 98.23
ated the choice of different dissimilarity functions used Triang. 92.55 84.43 92.70 92.83 92.91
Noise 89.16 86.48 92.31 92.43 92.45
to compare between distributions. The Heat-L2 spectral
All 90.94 83.37 90.90 91.37 93.05
distance was fixed with a single time scale t = 1024,
evaluating the retrieval performance of different choices
of the form Heat-L2 -Dissimilarity-Single. Table 1 summa- heat kernel and L1 , L2 and L∞ norms, the commute
rizes the mean average precision of each dissimilarity time and step kernels with the L2 norm. For compari-
function in different transformation classes. While there son, the performance of Geo-nL1 -Single is given. Since
is no absolute winner in all classes, the normalized L1 geodesic distances are known to be sensitive to large-
dissimilarity performs the best in four out of seven scale changes in topology, it is usually helpful to re-
classes, close to the best in other three, and wins in strict the distances within a certain radius. The reported
overall performance. This behavior is observed consis- results were obtained with d ≤ 100 which maximized
tently for other choices of spectral distances. Based on performance. Table 2 summarizes the measured mAP
this observation, we use the normalized L1 dissimilarity broken down by transformation class. Step kernel with
in all further experiments. the L2 norm performs the best in five out of seven
Kernel-Norm choice. In this experiment, we fixed the classes and achieves the highest overall performance.
distribution dissimilarity to nL1 and evaluated the per- The heat kernel with the L∞ norm achieves the low-
formance of different spectral distances. We evaluated est overall performance followed by the commute time
the performance of Kernel-Normal-nL1 -Single with the kernel and geodesic distances. As another reference,
IEEE TRANS. PAMI 8

we show the mAP of the compared distances under [8] M. M. Bronstein and I. Kokkinos, Scale-invariant heat kernel signa-
global scale transformations ranging from 125% to 200%. tures for non-rigid shape recognition, Proc. CVPR, 2010.
[9] R. R. Coifman and S. Lafon, Diffusion maps, Applied and Compu-
Poor performance is obtained by all methods with the tational Harmonic Analysis 21 (2006), 5–30.
exception of commute time kernel that scores 100% mAP. [10] A. B. Hamza and H. Krim, Geodesic matching of triangulated
This gives an experimental evidence to scale invariance surfaces, IEEE Trans. Imag. Proc. 15 (2006), no. 8, 2249.
[11] A. Ion, N.M. Artner, G. Peyré, W.G. Kropatsch, and L.D. Cohen,
of commute time distances and lack of scale invariance Matching 2D and 3D Articulated Shapes using Eccentricity, (2009).
of geodesic and diffusion distances. [12] S. Lafon, Diffusion Maps and Geometric Harmonics, Ph.D. disserta-
Aggregation choice. In this experiment, we evaluated tion, Yale University (2004).
[13] B. Lévy, Laplace-Beltrami eigenfunctions towards an algorithm that
the importance of aggregation of distribution dissimilar- “understands” geometry, Int’l Conf. Shape Modeling and Applica-
ities at different scales. The nL1 dissimilarity of the Heat- tions, 2006.
L2 spectral distance distribution was computed at six [14] H. Ling and K. Okada, Diffusion distance for histogram comparison,
Proc. CVPR, 2006.
time scales t ∈ {1024, 1351, 1782, 2352, 3104, 4096}. The [15] M. Mahmoudi and G. Sapiro, Three-dimensional point cloud recog-
performance of Heat-L2 -nL1 -Aggregation was evaluated nition via distributions of geometric distances, Graphical Models 71
for two single scales t = 1024 and 4096 as well as for (2009), no. 1, 22–31.
[16] F. Mémoli, Spectral Gromov-Wasserstein distances for shape matching,
the sum, sum of squares, and maximum aggregation Proc. NORDIA, 2009.
functions. Table 3 summarizes the results. We conclude [17] R. Osada, T. Funkhouser, B. Chazelle, and D. Dobkin, Shape
that using the maximum function achieves the best distributions, ACM Transactions on Graphics (TOG) 21 (2002),
no. 4, 807–832.
overall mAP. An even bigger gain is seen from the ROC [18] M. Ovsjanikov, A.M. Bronstein, M.M. Bronstein, and L.J. Guibas,
curves in Figure 2, where performance is evaluated in Shape Google: a computer vision approach to invariant shape retrieval,
terms of false positive and negative rates. Proc. NORDIA, 2009.
[19] U. Pinkall and K. Polthier, Computing discrete minimal surfaces and
their conjugates, Experimental mathematics 2 (1993), no. 1, 15–36.
6 C ONCLUSION [20] H. Qiu and ER Hancock, Clustering and embedding using commute
times, Trans. PAMI 29 (2007), no. 11, 1873–1890.
We presented a generic framework for computing the [21] M. Reuter, S. Biasotti, D. Giorgi, G. Patanè, and M. Spagnuolo,
similarity of non-rigid shapes based on the distributions Discrete Laplace–Beltrami operators for shape analysis and segmenta-
tion, Computers & Graphics 33 (2009), no. 3, 381–390.
of diffusion distances. Our framework allows building in
[22] M.R. Ruggeri and D. Saupe, Isometry-invariant matching of point set
explicit invariance properties, and also generalizes pre- surfaces, Proc. 3DOR, 2008.
vious approached due to Rustamov [23], Mahmoudi and [23] R. M. Rustamov, Laplace-Beltrami eigenfunctions for deformation
invariant shape representation, Proc. SGP, 2007, pp. 225–233.
Sapiro [15], and [7]; the presented formulation brings in-
[24] J. Sun, M. Ovsjanikov, and L. Guibas, A concise and provably
sights on the relations and equivalence of these methods. informative multi-scale signature based on heat diffusion, Proc. SGP,
An important question left undiscussed in this paper is 2009.

the construction of an optimal task-specific kernel K(λ).

We plan to address this problem in follow-up works.

R EFERENCES
[1] M. Belkin and P. Niyogi, Laplacian eigenmaps for dimensionality
reduction and data representation, Neural Computation 13 (2003),
1373–1396.
[2] M. Belkin, J. Sun, and Y. Wang, Discrete laplace operator on meshed
surfaces, Proc. Symp. Computational Geometry, 2009, pp. 278–287.
[3] A. Bronstein, M. Bronstein, M. Ovsjanikov, and L. Guibas, Shape
google: a computer vision approach to invariant shape retrieval, Proc.
NORDIA, 2009.
[4] A. M. Bronstein, M. M. Bronstein, and R. Kimmel, Numerical
geometry of nonrigid shapes, Springer, 2008.
[5] , Topology-invariant similarity of nonrigid shapes, Int’l J. Com-
puter Vision (IJCV) (2008), to appear.
[6] A. M. Bronstein, M. M. Bronstein, M. Mahmoudi, R. Kimmel, and
G. Sapiro, A Gromov-Hausdorff framework with diffusion geometry for
topologically-robust non-rigid shape matching, IJCV (2010).
[7] M. M. Bronstein and A. M. Bronstein, On a relation between
shape recognition algorithms based on distributions of distances, CIS
Technical Report 2009-14, Technion, 2009.
IEEE TRANS. PAMI 9

null isometry isometry+topology partiality topology triangulation noise all

0.9

0.8
TPR

0.7

0.6

0.5
(a) (b) (c)
0.4

0.9

0.8
TPR

0.7

0.6

0.5
(d) (e) (f )
0.4
−3 −2 −1 0 −3 −2 −1 0 −3 −2 −1 0
10 10 10 10 10 10 10 10 10 10 10 10
FPR FPR FPR

Fig. 2. ROC curves of (a) H-L2 -nL1 -Single, t = 1024; (b) H-L2 -nL1 -Single, t = 4096; (c) H-L2 -nL1 -Max,
t ∈ [1024, 4096]; (d) Geo-nL1 -Single; (e) CT-L2 -nL1 -Single; and (F) Step-L2 -nL1 -Single, t = 1024.