Unsupervised Deep Learning for Structured Shape Matching

Jean-Michel Roufosse Abhishek Sharma Maks Ovsjanikov

LIX, École Polytechnique LIX, École Polytechnique LIX, École Polytechnique
jm.roufosse@gmail.com kein.iitian@gmail.com maks@lix.polytechnique.fr
arXiv:1812.03794v3 [cs.GR] 22 Aug 2019

Abstract

We present a novel method for computing correspon-

dences across 3D shapes using unsupervised learning. Our
method computes a non-linear transformation of given de-
scriptor functions, while optimizing for global structural
properties of the resulting maps, such as their bijectivity
or approximate isometry. To this end, we use the func-
tional maps framework, and build upon the recent FMNet
architecture for descriptor learning. Unlike that approach,
however, we show that learning can be done in a purely
unsupervised setting, without having access to any ground
truth correspondences. This results in a very general shape
matching method that we call SURFMNet for Spectral Un- Source descriptor before Target descriptor before
supervised FMNet, and which can be used to establish cor-
respondences within 3D shape collections without any prior
information. We demonstrate on a wide range of challeng-
ing benchmarks, that our approach leads to state-of-the-art
results compared to the existing unsupervised methods and
achieves results that are comparable even to the supervised
learning techniques. Moreover, our framework is an order
of magnitude faster, and does not rely on geodesic distance
computation or expensive post-processing.

1. Introduction
Shape matching is a fundamental problem in computer
vision and geometric data analysis, with applications in de-
formation transfer [42] and statistical shape modeling [6] Source descriptor after Target descriptor after
among other domains. During the past decades, a large
Figure 1: Given a pair of shapes with noisy descriptors
number of techniques have been proposed for both rigid
(top), our approach makes them more consistent (bottom)
and non-rigid shape matching [44]. The latter case is both
without the knowledge of the underlying map, and auto-
more general and more challenging since the shapes can
matically computes an accurate pointwise correspondence.
potentially undergo arbitrary deformations (See Figure 1),
which are not easy to characterize by purely axiomatic ap-
proaches. As a result, several recent learning-based tech- spondences between some shape pairs. In the simplest case,
niques have been proposed for addressing the shape cor- this can be formulated as a labeling problem, where differ-
respondence problem, e.g. [10, 25, 26, 51] among many ent points, e.g., in a template shape, correspond to labels
others. Most of these approaches are based on the idea to be predicted [51, 27]. More recently, several methods
that the underlying correspondence model can be learned have been proposed for structured map prediction, aiming
from data, typically given in the form of ground truth corre- to infer an entire map, rather than labeling each point in-

1
dependently [10, 23]. These techniques are based on learn- wise products of functions, which corresponds to functional
ing pointwise descriptors, but, crucially, impose a penalty maps arising from point-to-point correspondences [29, 28];
on the entire map, obtained using these descriptors, result- and slanted diagonal structure of functional map in the con-
ing in higher quality, globally consistent correspondences. text of partial shapes [36, 24] among others.
Nevertheless, while learning-based methods have achieved Similarly, several other regularizers have been proposed,
impressive performance, their utility is severely limited by including exploiting the relation between functional maps
requiring the presence of high-quality ground truth maps in different directions [14], the map adjoint [18], and pow-
between a sufficient number of training examples. This erful cycle-consistency constraints [17] in shape collec-
makes it difficult to apply such approaches to new shape tions to name a few. More recently constraints on func-
classes for which ground truth data is not available. tional maps have been introduced to promote map continu-
In our paper, we show that this limitation can be lifted ity [35, 34] and kernel-based techniques for extracting more
and propose a purely unsupervised strategy, which com- information from given descriptors [49] among others. All
bines the accuracy of learning-based methods with the gen- these methods, however, are based on combining first-order
erality of axiomatic techniques for shape correspondence. penalties that arise from enforcing descriptor preservation
The key to our approach is a bi-level optimization scheme, constraints with these additional desirable structural prop-
which optimizes for descriptors on the shapes, but imposes erties of functional maps. As a result, any artefact or in-
a penalty on the entire map, inferred from them. For this, we consistency in the pre-computed descriptors will inevitably
use the recently proposed FMNet architecture [23], which lead to severe map estimation errors. Several methods have
exploits the functional map representation [30]. However, been suggested to use robust norms [21, 20], which can help
rather than penalizing the deviation of the map from the reduce the influence of certain descriptors but still does not
ground truth, we enforce structural properties on the map, control the global map consistency properties.
such as its bijectivity or approximate isometry. This results Most recently, a powerful technique BCICP, for map op-
in a shape matching method that achieves state-of-the-art timization, was introduced in [35] that combines a large
accuracy among unsupervised methods and, perhaps sur- number of functional constraints with sophisticated post-
prisingly, achieves comparable performance even to super- processing, and careful descriptor selection. As we show
vised techniques. below our method is simpler, more efficient and achieves
superior accuracy even to this recent approach.
2. Related Work Learning-based Methods To overcome the inherent dif-
ficulty of axiomatic techniques, several methods have been
Computing correspondences between 3D shapes is a
introduced to learn the correct deformation model from data
very well-studied area of computer vision and computer
with learning-based methods. Some early approaches in
graphics. Below we only review the most closely related
this direction were used to learn either optimal parame-
methods and refer the interested readers to recent surveys
ters of spectral descriptors [25] or exploited random forests
including [46, 44, 5] for more in-depth discussions.
[38] or metric learning [11] for learning optimal constraints
Functional Maps Our method is built on the functional given some ground truth matches.
map representation, which was originally introduced in More recently, with the advent of deep learning methods,
[30] for solving non-rigid shape matching problems, and several approaches have been proposed to learn transforma-
then extended significantly in follow-up works, including tions in the context of non-rigid shape matching. Most of
[2, 21, 20, 9, 15, 36] among many others (see also [31] for the proposed methods either use Convolutional Neural Net-
a recent overview). works (CNNs) on depth maps, e.g. for dense human body
One of the key benefits of this framework is that it al- correspondence [51] or exploit extensions of CNNs directly
lows us to represent maps between shapes as small matri- to curved surfaces, either using the link between convolu-
ces, which encode relations between basis functions defined tion and multiplication in the spectral domain [7, 12], or di-
on the shapes. Moreover, as observed by several works in rectly defining local parametrizations, for example via the
this domain [30, 40, 21, 36, 9], many natural properties on exponential map, which allows convolution in the tangent
the underlying pointwise correspondences can be expressed plane of a point, e.g. [26, 8, 27, 33] among others.
as objectives on functional maps. This includes orthonor- These methods have been applied to non-rigid shape
mality of functional maps, which corresponds to the lo- matching, in most cases modeling it as a label prediction
cal area-preservation nature of pointwise correspondences problem, with points corresponding to different labels. Al-
[30, 21, 40]; commutativity with the Laplacian operators, though successful in the presence of sufficient training data,
which corresponds to intrinsic isometries [30], preservation such approaches typically do not impose global consistency,
of inner products of gradients of functions, which corre- and can lead to artefacts, such as outliers, requiring post-
sponds to conformal maps [40, 9, 50]; preservation of point- processing to achieve high-quality maps.
Learning for Structured Prediction Most closely related 3. Compute the optimal functional map C by solving the
to our approach are recent works that apply learning for following optimization problem:
structured map prediction [10, 23]. These methods learn a


transformation of given input descriptors, while optimizing Copt = arg min Edesc C12 + αEreg C12 , (1)
C12
for the deviation of the map computed from them using the
functional map framework, from ground truth correspon- where the first term aims at the descriptor preservation:

2
dences. By imposing a penalty on entire maps, and thus Edesc C12 = C12 A1 − A2 , whereas the second
evaluating the ultimate use of the descriptors, these methods term regularizes the map by promoting the correctness
have led to significant accuracy improvements in practice. of its overall structural properties. The simplest ap-
We note that concurrent to our work, Halimi et al. [16] also proach penalizes the failure of the unknown functional
proposed an unsupervised deep learning method that com- map to commute with the Laplace-Beltrami operators:
putes correspondences without using the ground truth. This 2
approach is similar to ours, but is based on computation of Ereg (C12 ) = C12 Λ1 − Λ2 C12 (2)
geodesic distances, while our method operates purely in the
where Λ1 and Λ2 are diagonal matrices of the Laplace-
spectral domain making it extremely efficient.
Beltrami eigenvalues on the two shapes.
Contribution Unlike these existing methods, we propose 4. Convert the functional map C to a point-to-point map,
an unsupervised learning-based approach that transforms for example using nearest neighbor search in the spec-
given input descriptors, while optimizing for structural map tral embedding, or using other more advanced tech-
properties, without any knowledge of the ground truth or niques [37, 15].
geodesic distances. Our method, which can be seen as a bi-
level optimization strategy, allows to explicitly control the One of the strengths of this pipeline is that typically
interaction between pointwise descriptors and global map Eq. (1) leads to a simple (e.g., least squares) problem with
consistency, computed via the functional map framework. k1 k2 unknowns, independent of the number of points on the
As a result, our technique is scalable with respect to shape shapes. This formulation has been extended using e.g. man-
complexity, leads to significant improvement compared to ifold optimization [22], descriptor preservation constraints
the standard unsupervised methods, and achieves compara- via commutativity [29] and, more recently, with kerneliza-
ble performance even to supervised approaches. tion [49] among many others (see also Chapter 3 in [31]).
3.2. Deep Functional Maps
3. Background & Motivation
Despite its simplicity and efficiency, the functional map
3.1. Shape Matching and Functional Maps estimation pipeline described above is fundamentally de-
pendent on the initial choice of descriptor functions. To al-
Our work is based on the functional map framework and
leviate this dependence, several approaches have been pro-
representation. For completeness, we briefly review the ba-
posed to learn the optimal descriptors from data [10, 23].
sic notions and pipeline for estimating functional maps, and
In our work, we build upon a recent deep learning-based
refer the interested reader to a recent course [31] for a more
framework, called FMNet, introduced by Litany et al. [23]
in-depth discussion.
that aims to transform a given set of descriptors so that the
Basic Pipeline Given a pair of shapes, S1 , S2 represented optimal map computed using them is as close as possible to
as triangle meshes, and containing, respectively, n1 and n2 some ground truth map given during training.
vertices, the basic pipeline for computing a map between Specifically, the approach proposed in [23] assumes, as
them using the functional map framework, consists of the input, a set of shape pairs for which ground truth point-wise
following main steps (see Chapter 2 in [31]) : maps are known, and aims to solve the following problem:
X
1. Compute a small set of k1 , k2 of basis functions on min lF (Sof t(Copt ), GT(S1 ,S2 ) ), where (3)
T
each shape, e.g. by taking the first few eigenfunctions (S1 ,S2 )∈Train
of the respective Laplace-Beltrami operators. Copt = arg min kCAT (D1 ) − AT (D2 ) k. (4)
C
2. Compute a set of descriptor functions on each shape
that are expected to be approximately preserved by Here T is a non-linear transformation, in the form of a neu-
the unknown map. For example, a descriptor function ral network, to be applied to some input descriptor functions
can correspond to a particular dimension (e.g. choice D, Train is the set of training pairs for which ground truth
of time parameter of the Heat Kernel Signature [43]) correspondence GT(S1 ,S2 ) is known, lF is the soft error
computed at every point. Store their coefficients in the loss, which penalizes the deviation of the computed func-
respective bases as columns of matrices A1 , A2 . tional map Copt , after converting it to a soft map Sof t(Copt )
 structural properties of the inferred functional maps, such
as their bijectivity or orthogonality. Importantly, we express
FM
Net all these desired properties, and thus the penalties during
optimization, purely in the spectral domain, which allows
us to avoid the conversion of functional maps to soft maps
during optimization as was done in [23]. Thus, in addition
to being purely unsupervised, our approach is also more ef-
ficient since it does not require pre-computation of geodesic
FM distance matrices or expensive manipulation of large soft
Net
map matrices during training.
To achieve these goals, we build on the FMNet model,
described in Eq. (3) and (4) in several ways: first, we pro-
Figure 2: Overview of our SURFMNet approach: given a pose to consider functional maps in both directions, i.e. by
pair of shapes and their descriptors D1 , D2 , we optimize for treating the two shapes as both source and target; second,
a non-linear transformation T using the FMNet architecture we remove the conversion from functional to soft maps;
so that the transformed descriptors lead to functional maps and, most importantly, third, we replace the soft map loss
that best satisfy the structural constraints. with respect to ground truth with a set of penalties on the
computed functional maps, which are described in detail
below. Our optimization problem can be written as:
from the ground truth correspondence, and AT (D1 ) de- X X
notes the transformed descriptors D1 written in the basis of min wi Ei (C12 , C21 ), where (5)
T
shape 1. In other words, the FMNet framework [23] aims (S1 ,S2 ) i∈penalties
to learn a transformation T of descriptors, so that the trans- C12 = arg min kCAT (D1 ) − AT (D2 ) k, (6)
formed descriptors T (D1 ), T (D2 ), when used within the C
functional map pipeline result in a soft map that is as close C21 = arg min kCAT (D2 ) − AT (D1 ) k. (7)
as possible to some known ground truth correspondence. C
Unlike methods based on formulating shape matching as Here, similarly to Eq. (3) above, T denotes a non-linear
a labeling problem this approach evaluates the quality of transformation in the form of a neural network, (S1 , S2 ) is
the entire map, obtained using the transformed descriptors, a set of pairs of shapes in a given collection, wi are scalar
which as shown in [23] leads to significant improvement weights, and Ei are the penalties, described below. Thus,
compared to several strong baselines. we aim to optimize for a non-linear transformation of in-
Motivation Similarly to other supervised learning methods, put descriptor functions, such that functional maps com-
although FMNet [23] can result in highly accurate corre- puted from transformed descriptors possess certain desir-
spondences, its applicability is limited to shape classes for able structural properties and are expressed via penalty min-
which high-quality ground truth maps are available. More- imization. Figure 2 illustrates our proposed method where
over, perhaps less crucially, the soft map loss in FMNet is we denote the total sum of all penalty terms in Eq. (5) as
based on the knowledge of geodesic distances between all Eglobal and back-propagation via grey dashed lines.
pairs of points, making it computationally expensive. Our When deriving the penalties used in our approach, we
goal, therefore, is to show that a similar approach can be exploit the links between properties of functional maps and
used more widely, without any training data, while working associated pointwise maps, that have been established in
purely in the spectral domain. several previous works [30, 40, 14, 29]. Unlike all these
methods, however, we decouple the descriptor preservation
4. SURFMNet constraints from structural map properties. This allows us
to optimize for descriptor functions, and thus, gain a very
4.1. Overview strong resilience in the presence of noisy or uninformative
descriptors, while still exploiting the compactness and effi-
In this paper, we introduce a novel approach, which we
ciency of the functional map representation.
call SURFMNet for Spectral Unsupervised FMNet. Our
method aims to optimize for non-linear transformations of 4.2. Deep Functional Map Regularization
descriptors, in order to obtain high-quality functional, and
In our work, we propose to use four regularization terms,
thus pointwise maps. For this, we follow the general strat-
by including them as a penalties in the objective function,
egy proposed in FMNet [23]. However, crucially, rather
all inspired by desirable map properties.
than penalizing the deviation of the computed map from
the known ground truth correspondence, we evaluate the Bijectivity Given a pair of shapes and the functional maps
in both directions, perhaps the simplest requirement is for and expressed in the full (hat basis), whereas Φ, Ψ are the
them to be inverses of each other, which can be enforced by fixed basis functions on the two shapes, and + denotes the
penalizing the difference between their composition and the Moore-Penrose pseudoinverse.
identity map. This penalty, used for functional map estima-
4.3. Optimization
tion in [14], can be written, simply as:
As mentioned in Section 4.1, we incorporate these four
E1 = kC12 C21 − Ik2 + kC21 C12 − Ik2 (8) penalties into the energy in Eq. (5). Importantly, the only
unknowns in this optimization are the parameters of the
Orthogonality As observed in several works [30, 40] a neural network applied to the descriptor functions. The
point-to-point map is locally area preserving if and only functional maps C12 and C21 are fully determined by the
if the corresponding functional map is orthonormal. Thus, optimized descriptors via the solution of the optimization
for shape pairs, approximately satisfying this assumption, a problems in Eq. (6) and Eq. (7). Note that although stated
natural penalty in our unsupervised pipeline is: as optimization problems, both Eq. (6) and Eq. (7) reduce
to solving a linear system of equations. This is easily differ-
E2 = kC> 2 >
12 C12 − Ik + kC21 C21 − Ik
2
(9)
entiable using the well-known closed-form expression for
derivatives of matrix inverses [32]. Moreover, the function-
Laplacian commutativity Similarly, it is well-known that
ality of differentiating a linear system of equations is im-
a pointwise map is an intrinsic isometry if and only if
plemented in TensorFlow [1] and we use it directly, in the
the associated functional map commutes with the Laplace-
same way as it was used in the original FMNet work. Fi-
Beltrami operator [39, 30]. This has motivated using the
nally, all of the penalties E1 , E2 , E3 , E4 are differentiable
lack of commutativity as a regularizer for functional map
with respect to the functional maps C12 , C21 . This means
computations, as mentioned in Eq. (2). In our work, we use
that the gradient of the total energy can be back-propagated
it to introduce the following penalty:
to the neural network T in Eq. (5), allowing us to optimize
E3 = C12 Λ1 − Λ2 C12
2
+ C21 Λ2 − Λ1 C21
2
(10) for the descriptors while penalizing the structural properties
of the functional maps.
where Λ1 and Λ2 are diagonal matrices of the Laplace-
Beltrami eigenvalues on the two shapes. 5. Implementation & Parameters
Descriptor preservation via commutativity The previ-
Implementation details We implemented 1 our method in
ous three penalties capture desirable properties of point-
TensorFlow [1] by adapting the open-source implementa-
wise correspondences when expressed as functional maps.
tion of FMNet [23]. Thus, the neural network T used
Our last penalty promotes functional maps that arise from
for transforming descriptors in our approach, in Eq. (5) is
point-to-point maps, rather than more general soft corre-
exactly identical to that used in FMNet, as mentioned in
spondences. To achieve this, we follow the approach pro-
Eq. (3). Namely, this network is based on a residual archi-
posed in [29] based on preservation of pointwise products
tecture, consisting of 7 fully connected residual layers with
of functions. Namely, it is known that a non-trivial linear
exponential linear units, without dimensionality reduction.
transformation T across function spaces corresponds to a
Please see Section 5 in [23] for more details.
point-to-point map if and only if T (f h) = T (f ) T (h)
for any pair of functions f, h. Here denotes the point- Following the approach of FMNet [23], we also sub-sample
wise product between functions [41], i.e. (f h)(x) = a random set of 1500 points at each training step, for ef-
f (x)h(x). When f is a descriptor function on the source ficiency. However, unlike their method, sub-sampling is
and g is the corresponding descriptor on the target, the au- done independently on each shape, without enforcing con-
thors of [29] demonstrate that this condition can be rewrit- sistency. Remark that our network is fully connected on
ten in the reduced basis as follows: CMf = Mg C, where the dimensions of the descriptors, not across vertices them-
Mf = Φ+ Diag(f )Φ, and Mg = Ψ+ Diag(g)Ψ. This leads selves. For example, the first layer has 352 × 352 weights
to the following penalty, in our setting: (not 1500×352 weights) where 352 and 1500 are the dimen-
X sions of the SHOT descriptors, and no. of sampled vertices
E4 = ||C12 Mfi − Mgi C12 ||2 respectively. Indeed, in exactly the same way as in FM-
(fi ,gi )∈Descriptors Net, our network is applied on the descriptors of each point
(11) independently, using the same (learned) weights, and differ-
+||C21 Mgi − Mfi C21 ||2 ,
ent points on the shape only communicate through the func-
Mfi = Φ+ Diag(fi )Φ, Mgi = Ψ+ Diag(gi )Ψ.
tional map estimation layer, and not in the MLP layers. This
In this expression, fi and gi are the optimized descriptors 1 Code available at https://github.com/
on source and target shape, obtained by the neural network, LIX-shape-analysis/SURFMNet.
 Methods E1+E2+E3+E4 E3 E1 E2 E4 We stress that although we also evaluated on the origi-
Geodesic Error 0.020 0.073 0.083 0.152 0.252 nal FAUST dataset, we view the remeshed datasets as more
realistic, providing a more faithful representation of the ac-
Table 1: Ablation study of penalty terms in our method on curacy and generalization power of different techniques.
the FAUST benchmark.
Ablation study We first evaluated the relative importance
ensures invariance to permutation of shape vertices. We also of the different penalties in our method on the FAUST
randomly sub-sample 20% of the optimized descriptors for shape dataset [6]. We evaluated the average correspondence
our penalty E4 at each training step to avoid manipulating a geodesic error with respect to the ground truth maps.
large set of operators. We observed that this sub-sampling Table 1 summarizes the quality of the computed corre-
not only helps to gain speed but also robustness during opti- spondences between shapes in the test set, using different
mization. Importantly, we do not form large diagonal matri- combination of penalties. We observe that the combination
ces explicitly, but rather define the multiplicative operators of all four penalties significantly out-performs any other
M in objective E4 directly via pointwise products and sum- subsets. Besides, among individual penalties used indepen-
mation using contraction between tensors. dently, the Laplacian commutativity gives the best result.
Finally, we convert functional maps to pointwise ones For more combinations of penalty terms, we refer to a more
with nearest neighbor search in the spectral domain, fol- detailed ablation study in the supplementary material.
lowing the original approach [30].
Baselines We compared our method to several techniques,
Parameters Our method takes two types of inputs: the in- both supervised and fully automatic. For conciseness, we
put descriptors, and the scalar weights wi in Eq. (5). In refer to SURFMNet as Ours in the following text. For a
all experiments below, we used the same SHOT [45] de- fair comparison with FMNet, we evaluate our method in
scriptors as in FMNet [23] with the same parameters, which two settings: Ours-sub and Ours-all. For Ours-sub, we
leads to a 352-dimensional vector per point, or equivalently, split each dataset into training and test sets containing 80
352 descriptor functions on each shape. For the scalar and 20 shapes respectively, as done in [23]. For Ours-all,
weights, wi , we used the same four fixed values for all ex- we optimize over all the dataset and apply the optimized
periments below (namely, w1 = 103 , w2 = 103 , w3 = 1 network on the same test set as before. We stress that unlike
and w4 = 105 ), which were obtained by examining the rel- FMNet, our method does not use any ground truth in either
ative penalty values obtained throughout the optimization setting. We use the notation Ours-sub only to emphasize the
on a small set of shapes, and setting the weights inversely split of dataset into train and test since the “training set” was
proportionally to those values. We train our network with a only used for descriptor optimization with the functional
batch size of 10 for 10 000 iterations using a learning rate map penalties introduced above without any ground truth.
of 0.001 and ADAM optimizer [13]. Since the original FMNet work [23] already showed very
strong improvement compared to existing supervised learn-
6. Results ing methods we primarily compare to this approach. For
reference, we also compare to the Geodesic Convolutional
Datasets We evaluate our method on the following datasets: Neural Networks (GCNN) method of [26] on the remeshed
the original FAUST dataset [6] containing 100 human datasets, which were not considered in [23]. GCNN is a rep-
shapes in 1-1 correspondence and the remeshed versions of resentative supervised method based on local shape param-
SCAPE [3] and FAUST [6] datasets, made publicly avail- eterization, and as FMNet assumes, as input, ground truth
able recently by Ren et al. [35]. These datasets were ob- maps between a subset of the training shapes. For super-
tained by independently re-meshing each shape to approx- vised methods, we always split the datasets into 80 (resp.
imately 5000 vertices using the LRVD re-meshing method 60) shapes for training and 20 (resp. 10) for testing in the
[52], while keeping track of the ground truth maps within FAUST and SCAPE datasets respectively.
each collection. This results in meshes that are no longer Among fully automatic methods, we use the Product
in 1-1 correspondence, and indeed can have different num- Manifold Filter method with the Gaussian kernel [48] (PMF
ber of vertices. The re-meshed datasets therefore offer sig- Gauss) and its variant with the Heat kernel [47] (PMF Heat).
nificantly more variability in terms of shape structures, in- We also compare to the recently proposed BCICP [35],
cluding e.g. point sampling density, making them more which achieved state-of-the-art results among axiomatic
challenging for existing algorithms. Let us note also that methods. With a slight abuse of notation, we denote these
the SCAPE dataset is slightly more challenging since the non-learning methods as Unsupervised in Figure 4 since
shapes are less regular (e.g., there are often reconstruction none of these methods use ground truth. Finally, we also
artefacts on hands and feet) and have fewer features than evaluated the basic functional map approach, based on di-
those in FAUST. rectly optimizing the functional maps as outlined in Sec-
 Figure 3: Quantitative evaluation of pointwise correspondences comparing our method with Supervised Methods.

Figure 4: Quantitative evaluation of pointwise correspondences comparing our method with Unsupervised Methods.

tion 3.1, but using all four of our energies for regulariza- Remark that the remeshed datasets are significantly
tion. This method, which we call “Fmap Basic” can be harder for both supervised and unsupervised methods, since
viewed as a combination of the approaches of [14] and [28], the shapes are no longer identically meshed and in 1-1 cor-
as it incorporates functional map coupling (via energy E1 ) respondence. We have observed this difficulty also while
and descriptor commutativity (via E4 ). Unlike our tech- training supervised FMNet and GCNN techniques with
nique, however, it operates on fixed descriptor functions, very slow convergence during training. On both of these
and uses descriptor preservation constraints with the origi- datasets, our approach achieves the lowest average error, re-
nal and noisy descriptors. ported in Figure 3 and 4. Note that on the remeshed FAUST
For fairness of comparison, we used SHOT descriptors dataset, as shown in Figure 3, only GCNN [7] produces a
[45] as input to all methods, except BCICP [35], which uses similarly large fraction of correspondences with a small er-
carefully curated WKS [4] descriptors. Furthermore, we ror. However, this method is supervised. On the remeshed
consider the results of FMNet [23] before and after applying SCAPE dataset, our method leads to the best results across
the PMF-based post-processing as suggested in the original all measures, despite being purely unsupervised.
article. We also report results with ICP post-processing in-
troduced in [30]. Besides the accuracy plots shown in Fig- Postprocessing Results As shown in Figures 3 and 4 our
ures 3 and 4, we also include statistics such as maximum method can often obtain high quality results even without
and 95th percentile in supplementary material. any post-processing. Nevertheless, in the challenging cases
such as the SCAPE remeshed dataset, when trained on a
6.1. Evaluation and Results subset of shapes, it can also benefit from an efficient ICP-
based refinement. This refinement, does not require com-
Figure 3 summarizes the accuracy obtained by super- puting geodesic distances and does not require the shapes
vised methods on the three datasets whereas Figure 4 com- to have the same number of points, thus maintaining the
pares with unsupervised methods, using the evaluation pro- flexibility and efficiency of our pipeline.
tocol introduced in [19]. Note that in all cases, our network
SURFMNet, (Ours-all), when optimized on all shapes Correlation with actual Geodesic loss We further investi-
achieves the best results even compared to the recent state- gated if there is a correlation between the value of our loss
of-the-art method in [35]. Furthermore, our method is com- and the quality of correspondence. Specifically, whether
parable even to supervised learning techniques, GCNN [7] minimizing our loss function, mainly consisting of regular-
and FMNet [23] despite being purely unsupervised. ization terms on estimated functional maps, corresponds to
 Source Ground-Truth SURFMNet BCICP PMF (heat) PMF (gauss)

Figure 5: Comparison of our method with Unsupervised methods for texture transfer on the SCAPE remeshed dataset. Note
that BCICP is roughly 7 times slower than our method and its shortcomings are marked with red circles.

Source Ground-Truth SURFMNet+ICP SURFMNet FMNet FMNet + PMF GCNN

Figure 6: Comparison of our method with Supervised method for texture transfer on the SCAPE remeshed dataset.

Runtime 7. Conclusion & Future Work

Methods Pre-processing Training Testing Post-processing Total
FMNet 60s 1500s 0.3s N/As 1650s
FMNet + PMF 60s 1500s 0.3s 30s 1680s We presented an unsupervised method for computing
Fmap Basic 10s N/A 60s N/A 120s
BCICP N/A N/A 60s 180s 240s
correspondences between shapes. Key to our approach is
SURFMNet 10s 25s 0.3s N/A 35s a bi-level optimization formulation, aimed to optimize de-
SURFMNet + ICP 10s 25s 0.3s 10s 45s scriptor functions, while promoting the structural properties
of the entire map, obtained from them via the functional
Table 2: Runtime of different methods averaged over 190 maps framework. Remarkably, our approach achieves sim-
shape pairs. ilar, and in some cases superior performance even to super-
vised correspondence techniques.
minimizing the geodesic loss with respect to the unknown
ground truth map. We found strong correlation between the In the future, we plan to incorporate other penalties on
two and share a plot in the supplementary material. functional maps, e.g., those arising from recently-proposed
kernalization approaches [49], or for promoting orientation
Qualitative and Runtime Comparison Figures 5 and 6 preserving maps[35] and also incorporate cycle consistency
show examples shape pairs and maps obtained between constraints [17]. Finally, it would be interesting to extend
them using different methods, visualized via texture trans- our method to partial and non-isometric shapes and match-
fer. Note the continuity and quality of the maps obtained ing other modalities, such as images or point clouds, since it
using our method, compared to other techniques (more re- opens the door to linking the properties of local descriptors
sults in supplementary material). One further advantage of to global map consistency.
our method is its efficiency, since we do not rely on the
computation of geodesic matrices and operate entirely in Acknowledgements Parts of this work were supported by
the spectral domain. Table 2 compares the run-time of the the ERC Starting Grant StG-2017-758800 (EXPROTEA),
best performing methods on an Intel Xeon 2.10GHz ma- KAUST OSR Award No. CRG-2017-3426, and a gift from
chine with an NVIDIA Titan X GPU. Note that our method Nvidia. We are grateful to Jing Ren, Or Litany, Emanuele
is over an order of magnitude faster than FMNet and signif- Rodolà and Adrien Poulenard for their help in performing
icantly faster than the currently best unsupervised BCICP. quantitative comparisons and producing qualitative results.
References [12] Michaël Defferrard, Xavier Bresson, and Pierre Van-
dergheynst. Convolutional neural networks on graphs
[1] Martı́n Abadi, Ashish Agarwal, and Paul Barham et with fast localized spectral filtering. In Advances in
al. TensorFlow: Large-scale machine learning on het- Neural Information Processing Systems, pages 3844–
erogeneous systems, 2015. Software available from 3852, 2016. 2
tensorflow.org. 5
[13] J. Ba D.P. Kingma. Adam: A method for stochastic
[2] Yonathan Aflalo, Anastasia Dubrovina, and Ron Kim- optimization. In ICLR, 2015. 6
mel. Spectral generalized multi-dimensional scal-
[14] Davide Eynard, Emanuele Rodola, Klaus Glashoff,
ing. International Journal of Computer Vision,
and Michael M Bronstein. Coupled functional maps.
118(3):380–392, 2016. 2
In 3D Vision (3DV), pages 399–407. IEEE, 2016. 2, 4,
[3] Dragomir Anguelov, Praveen Srinivasan, Daphne 5, 7
Koller, Sebastian Thrun, Jim Rodgers, and James
[15] Danielle Ezuz and Mirela Ben-Chen. Deblurring and
Davis. SCAPE: Shape Completion and Animation of
denoising of maps between shapes. In Computer
People. In ACM Transactions on Graphics (TOG),
Graphics Forum, volume 36, pages 165–174. Wiley
volume 24, pages 408–416. ACM, 2005. 6
Online Library, 2017. 2, 3
[4] Mathieu Aubry, Ulrich Schlickewei, and Daniel Cre- [16] Oshri Halimi, Or Litany, Emanuele Rodol‘a, Alex
mers. The wave kernel signature: A quantum mechan- Bronstein, and Ron Kimmel. Unsupervised learning
ical approach to shape analysis. 31(4), Nov. 2011. 7 of dense shape correspondence. In CVPR, 2019. 3
[5] Silvia Biasotti, Andrea Cerri, A Bronstein, and M [17] Qixing Huang, Fan Wang, and Leonidas Guibas.
Bronstein. Recent trends, applications, and perspec- Functional map networks for analyzing and exploring
tives in 3d shape similarity assessment. In Computer large shape collections. ACM Transactions on Graph-
Graphics Forum, volume 35, pages 87–119, 2016. 2 ics (TOG), 33(4):36, 2014. 2, 8
[6] Federica Bogo, Javier Romero, Matthew Loper, and [18] Ruqi Huang and Maks Ovsjanikov. Adjoint map rep-
Michael J. Black. FAUST: Dataset and evaluation resentation for shape analysis and matching. In Com-
for 3D mesh registration. In Proceedings IEEE Conf. puter Graphics Forum, volume 36, pages 151–163.
on Computer Vision and Pattern Recognition (CVPR), Wiley Online Library, 2017. 2
Piscataway, NJ, USA, June 2014. IEEE. 1, 6 [19] Vladimir G Kim, Yaron Lipman, and Thomas
[7] Davide Boscaini, Jonathan Masci, Simone Melzi, Funkhouser. Blended intrinsic maps. In ACM Trans-
Michael M Bronstein, Umberto Castellani, and Pierre actions on Graphics (TOG), volume 30, page 79.
Vandergheynst. Learning class-specific descriptors for ACM, 2011. 7
deformable shapes using localized spectral convolu- [20] Artiom Kovnatsky, Michael M Bronstein, Xavier
tional networks. In Computer Graphics Forum, vol- Bresson, and Pierre Vandergheynst. Functional cor-
ume 34, pages 13–23. Wiley Online Library, 2015. 2, respondence by matrix completion. In Proceedings of
7 the IEEE conference on computer vision and pattern
[8] Davide Boscaini, Jonathan Masci, Emanuele Rodola, recognition, pages 905–914, 2015. 2
and Michael M. Bronstein. Learning shape correspon- [21] Artiom Kovnatsky, Michael M Bronstein, Alexan-
dence with anisotropic convolutional neural networks. der M Bronstein, Klaus Glashoff, and Ron Kimmel.
In Proc. NIPS, pages 3189–3197, 2016. 2 Coupled quasi-harmonic bases. In Computer Graph-
[9] Oliver Burghard, Alexander Dieckmann, and Rein- ics Forum, volume 32, pages 439–448, 2013. 2
hard Klein. Embedding shapes with Green’s functions [22] Artiom Kovnatsky, Klaus Glashoff, and Michael M
for global shape matching. Computers & Graphics, Bronstein. MADMM: a generic algorithm for non-
68:1–10, 2017. 2 smooth optimization on manifolds. In Proc. ECCV,
[10] Etienne Corman, Maks Ovsjanikov, and Antonin pages 680–696. Springer, 2016. 3
Chambolle. Supervised descriptor learning for non- [23] Or Litany, Tal Remez, Emanuele Rodolà, Alexan-
rigid shape matching. In Proc. ECCV Workshops der M. Bronstein, and Michael M. Bronstein. Deep
(NORDIA), 2014. 1, 2, 3 functional maps: Structured prediction for dense
[11] Luca Cosmo, Emanuele Rodola, Jonathan Masci, An- shape correspondence. 2017 IEEE International Con-
drea Torsello, and Michael M Bronstein. Matching ference on Computer Vision (ICCV), pages 5660–
deformable objects in clutter. In 3D Vision (3DV), 5668, 2017. 2, 3, 4, 5, 6, 7
2016 Fourth International Conference on, pages 1–10. [24] Or Litany, Emanuele Rodolà, Alex M Bronstein, and
IEEE, 2016. 2 Michael M Bronstein. Fully spectral partial shape
 matching. In Computer Graphics Forum, volume 36, correspondences via functional maps. ACM Transac-
pages 247–258. Wiley Online Library, 2017. 2 tions on Graphics (TOG), 37(6), 2018. 2, 6, 7, 8
[25] Roee Litman and Alexander M Bronstein. Learning [36] Emanuele Rodolà, Luca Cosmo, Michael M Bron-
spectral descriptors for deformable shape correspon- stein, Andrea Torsello, and Daniel Cremers. Partial
dence. IEEE transactions on pattern analysis and ma- functional correspondence. In Computer Graphics Fo-
chine intelligence, 36(1):171–180, 2014. 1, 2 rum, volume 36, pages 222–236. Wiley Online Li-
brary, 2017. 2
[26] Jonathan Masci, Davide Boscaini, Michael Bronstein,
and Pierre Vandergheynst. Geodesic convolutional [37] Emanuele Rodolà, M Moeller, and Daniel Cremers.
neural networks on riemannian manifolds. In Proceed- Point-wise map recovery and refinement from func-
ings of the IEEE international conference on computer tional correspondence. In Proc. Vision, Modeling and
vision workshops, pages 37–45, 2015. 1, 2, 6 Visualization (VMV), 2015. 3
[27] Federico Monti, Davide Boscaini, Jonathan Masci, [38] Emanuele Rodolà, Samuel Rota Bulo, Thomas Wind-
Emanuele Rodolà, Jan Svoboda, and Michael M. heuser, Matthias Vestner, and Daniel Cremers. Dense
Bronstein. Geometric deep learning on graphs and non-rigid shape correspondence using random forests.
manifolds using mixture model cnns. In CVPR, pages In Proceedings of the IEEE Conference on Computer
5425–5434. IEEE Computer Society, 2017. 1, 2 Vision and Pattern Recognition, pages 4177–4184,
2014. 2
[28] Dorian Nogneng, Simone Melzi, Emanuele Rodolà,
[39] Steven Rosenberg. The Laplacian on a Riemannian
Umberto Castellani, M Bronstein, and Maks Ovs-
manifold: an introduction to analysis on manifolds,
janikov. Improved functional mappings via prod-
volume 31. Cambridge University Press, 1997. 5
uct preservation. In Computer Graphics Forum, vol-
ume 37, pages 179–190. Wiley Online Library, 2018. [40] Raif Rustamov, Maks Ovsjanikov, Omri Azencot,
2, 7 Mirela Ben-Chen, Frederic Chazal, and Leonid
Guibas. Map-based exploration of intrinsic shape
[29] Dorian Nogneng and Maks Ovsjanikov. Informative
differences and variability. ACM Trans. Graphics,
descriptor preservation via commutativity for shape
32(4):72:1–72:12, July 2013. 2, 4, 5
matching. Computer Graphics Forum, 36(2):259–
267, 2017. 2, 3, 4, 5 [41] R K Singh and J S Manhas. Composition Operators
on Function Spaces. ELSEVIER, 1993. 5
[30] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon,
[42] Robert W Sumner and Jovan Popović. Deformation
Adrian Butscher, and Leonidas Guibas. Functional
transfer for triangle meshes. In ACM Transactions on
Maps: A Flexible Representation of Maps Between
Graphics (TOG), volume 23, pages 399–405. ACM,
Shapes. ACM Transactions on Graphics (TOG),
2004. 1
31(4):30, 2012. 2, 4, 5, 6, 7
[43] Jian Sun, Maks Ovsjanikov, and Leonidas Guibas. A
[31] Maks Ovsjanikov, Etienne Corman, Michael Bron-
Concise and Provably Informative Multi-Scale Signa-
stein, Emanuele Rodolà, Mirela Ben-Chen, Leonidas
ture Based on Heat Diffusion. In Computer graphics
Guibas, Frederic Chazal, and Alex Bronstein. Com-
forum, volume 28, pages 1383–1392, 2009. 3
puting and processing correspondences with func-
tional maps. In ACM SIGGRAPH 2017 Courses, SIG- [44] Gary KL Tam, Zhi-Quan Cheng, Yu-Kun Lai, Frank C
GRAPH ’17, pages 5:1–5:62, 2017. 2, 3 Langbein, Yonghuai Liu, David Marshall, Ralph R
Martin, Xian-Fang Sun, and Paul L Rosin. Registra-
[32] K. B. Petersen and M. S. Pedersen. The matrix cook- tion of 3d point clouds and meshes: a survey from
book, 2012. 5 rigid to nonrigid. IEEE transactions on visualization
[33] Adrien Poulenard and Maks Ovsjanikov. Multi- and computer graphics, 19(7):1199–1217, 2013. 1, 2
directional geodesic neural networks via equivariant [45] Federico Tombari, Samuele Salti, and Luigi Di Ste-
convolution. arXiv preprint arXiv:1810.02303, 2018. fano. Unique signatures of histograms for local sur-
2 face description. In International Conference on Com-
[34] Adrien Poulenard, Primoz Skraba, and Maks Ovs- puter Vision (ICCV), pages 356–369, 2010. 6, 7
janikov. Topological function optimization for contin- [46] Oliver Van Kaick, Hao Zhang, Ghassan Hamarneh,
uous shape matching. In Computer Graphics Forum, and Daniel Cohen-Or. A survey on shape correspon-
volume 37, pages 13–25. Wiley Online Library, 2018. dence. In Computer Graphics Forum, volume 30,
2 pages 1681–1707, 2011. 2
[35] Jing Ren, Adrien Poulenard, Peter Wonka, and Maks [47] Matthias Vestner, Zorah Lähner, Amit Boyarski, Or
Ovsjanikov. Continuous and orientation-preserving Litany, Ron Slossberg, Tal Remez, Emanuele Rodola,
 Alex Bronstein, Michael Bronstein, Ron Kimmel, and Unsupervised Loss vs Average Geodesic Error 10 4
2.2

Daniel Cremers. Efficient deformable shape corre- 0.11

Average Geodesic Distance Error
Unsupervised Loss
spondence via kernel matching. In Proc. 3DV, 2017. 0.1 2

6 0.09

Average Geodesic Distance Error

0.08 1.8
[48] M. Vestner, R. Litman, E. Rodolà, A. Bronstein, and

Unsupervised Loss
0.07

D. Cremers. Product manifold filter: Non-rigid shape 0.06 1.6

correspondence via kernel density estimation in the 0.05

product space. In Proc. CVPR, pages 6681–6690, 0.04 1.4

2017. 6 0.03

0.02 1.2

[49] Larry Wang, Anne Gehre, Michael M Bronstein, and 0.01

Justin Solomon. Kernel functional maps. In Computer 1

0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Graphics Forum, volume 37, pages 27–36. Wiley On- Number of Mini-Batch Iterations

line Library, 2018. 2, 3, 8

Figure 7: Correlation with average geodesic loss computed
[50] Yuexuan Wang, B Liu, K Zhou, and Yu Tong. Vec-
from ground truth correspondences.
tor field map representation for near conformal sur-
face correspondence. In Computer Graphics Forum,
volume 37, pages 72–83. Wiley Online Library, 2018.
2 C. Sensitivity to number of basis functions
[51] Lingyu Wei, Qixing Huang, Duygu Ceylan, Etienne Figure 8 shows the sensitivity of our network SURFM-
Vouga, and Hao Li. Dense human body correspon- Net on the SCAPE remeshed dataset as the number of eigen
dences using convolutional networks. In Proceedings functions are varied from 20 to 150. We train the network
of the IEEE Conference on Computer Vision and Pat- each time with 10000 mini batch steps. As evident, we
tern Recognition, pages 1544–1553, 2016. 1, 2 obtain best result using 120. However, when trained on
[52] Dong-Ming Yan, Guanbo Bao, Xiaopeng Zhang, and an individual dataset and tested on a different one, we see
Peter Wonka. Low-resolution remeshing using the lo- over-fitting when using a large eigen-basis. We attribute
calized restricted voronoi diagram. IEEE transactions this phenomenon to the initialization of our descriptors with
on visualization and computer graphics, 20(10):1418– SHOT which is a very local descriptor and is not robust to
1427, 2014. 6 very strong mesh variability. However, over-fitting is mini-
mal when we train together on a relatively larger subset of
SCAPE and FAUST and test on a different subset of shapes
8. Supplement from both datasets, with smaller eigen basis.
A. Correlation with actual geodesic loss
To support the claim made in the subsection ’Evaluation 0.15
SCAPE remeshed
and Results’, we include a plot here to visualize the cor-
relation between our loss and the actual geodesic loss. As
evident in Figure 7, there is a strong correlation between our
Average Geodesic Distance Error

0.1
loss value and the quality of correspondence as measured by
average geodesic error.

B. Detailed Tabular Quantitative Comparison

0.05

Besides the average geodesic error reported for quanti-

tative comparison in Figures 3 and 4, we provide detailed
statistics in Table 3. Note that Table 3 also includes ’Fmap
Ours Opt’ which is equivalent to “Fmap Basic” but uses 0
20 40 60 80 100 120 140 160
Number of Eigenvectors
the learned descriptors instead of original ones. Its compet-
itive performance across all datasets proves quantitatively Figure 8: Accuracy of our method on the SCAPE remeshed
the utility of learning descriptors. Figures 13 and 14 illus- dataset as the number of eigenfunctions is varied from 20 to
trate this further. For completeness, in Table 4, we also pro- 150.
vide a detailed ablation study with different combinations
of penalties.
 (Results are ×10−3 ) FAUST 7k FAUST 5k SCAPE 5k
Supervised Methods Mean 95th Percentile Maximum Mean 95th Percentile Maximum Mean 95th Percentile Maximum
FMNet 25.01 63.11 1207.8 112.8 451.8 1280.6 172.6 543.8 1399.6
SURFMNet Subset 19.83 52.11 1204.0 92.09 493.6 1279.4 60.32 329.8 1068.7
FMNet + PMF 2.98 14.10 1222.7 83.61 395.7 1576.4 63.00 159.8 1561.5
SURFMNet-sub + PMF 5.33 22.90 1302.4 74.80 408.5 1619.3 51.03 111.5 1555.6
FMNet + ICP 11.16 27.91 1206.8 47.53 237.3 1348.6 81.76 341.4 1226.5
SURFMNet-sub + ICP 11.79 35.76 1088.4 30.47 95.64 1277.3 23.00 54.76 73.18
GCNN 50.49 206.3 1578.2 71.85 374.2 1523.7
Unsupervised Methods
BCICP 15.46 53.27 572.4 31.08 64.51 1149.9 22.28 50.60 107.5
PMF (Gaussian Kernel) 29.42 83.80 1168.1 75.13 236.9 1632.7 54.68 156.9 465.1
PMF (Heat Kernel) 17.26 25.06 1168.1 31.08 64.51 1150.0 47.23 133.4 802.1
Fmap Basic 457.56 1171.4 1568.4 366.2 1159.0 1549.1 383.0 1043.7 1280.3
Fmap Ours Opt 9.75 30.02 420.2 20.19 53.24 1169.5 13.98 31.16 86.45
SURFMNet-all 7.89 26.01 572.4 18.56 50.25 1156.3 17.50 42.50 228.8

Table 3: Quantitative comparison on all three benchmark datasets for shape correspondence problem.

Methods E1+E2+E3+E4 E3 E1+E2+E3 E1+E3+E4 E1 E2+E3+E4 E1+E2+E4 E2 E4 FMNet Ours-Sub Ours-all

Mean Geodesic Error 0.044 0.073 0.081 0.077 0.111 0.079 0.126 0.135 0.330 0.025 0.020 0.008

Table 4: Ablation study of penalty terms in our method and comparison with the supervised FMNet on the FAUST benchmark.

Source Ground-Truth SUFMNet + ICP SUFMNet FMNet FMNet + ICP GCNN

Figure 9: Comparison of our method with Supervised methods for texture transfer on the FAUST remeshed dataset.

D. More Qualitative Comparison

In Figures 9 and 12 , we provide more qualitative com-
parisons of SURFMNet on the FAUST remeshed datasets
whereas Figures 10 and 11 provide a comparison on the
SCAPE remeshed dataset. In all cases, our method pro-
duces the highest quality maps.
 Source Ground-Truth SURFMNet-all BCICP PMF (heat) PMF (gauss)
Figure 10: Comparison of our method with Unsupervised methods for texture transfer on the SCAPE remeshed dataset.

Source Ground-Truth SURFMNet + ICP SURFMNet FMNet FMNet + ICP GCNN

Figure 11: Comparison of our method with Supervised methods for texture transfer on the SCAPE remeshed dataset.

Source Ground-Truth SURFMNet BCICP PMF (heat) PMF (gauss)

Figure 12: Comparison of our method with Unsupervised methods for texture transfer on the FAUST remeshed dataset. Note
that BCICP is roughly 7 times slower when compared to our method. We highlight the shortcomings of BCICP matching
with red circles.
 Supervised Methods on FAUST Original Supervised Methods on FAUST Remesh Supervised Methods on SCAPE Remesh
1 1 1

0.9 0.9 0.9

0.8 0.8 0.8

Fraction of correspondences

Fraction of correspondences

Fraction of correspondences
0.7 0.7 0.7

0.6 0.6 0.6

0.5 0.5 0.5

0.4 0.4
FMNet : 0.1128 0.4
FMNet : 0.1726
FMNet : 0.0250 FMNet + PMF : 0.0836 FMNet + PMF : 0.0630
FMNet + PMF : 0.0030 FMNet + ICP : 0.0475 FMNet + ICP : 0.0818
0.3 0.3 0.3
FMNet + ICP: 0.0112 Ours-sub : 0.0921 Ours-sub : 0.0603
Ours-sub : 0.0198 Ours-sub + PMF : 0.0748 Ours-sub + PMF : 0.0510
0.2 Ours-sub + PMF : 0.0053 0.2 Ours-sub + ICP : 0.0305 0.2 Ours-sub + ICP : 0.0230
Ours-sub + ICP : 0.0118 GCNN : 0.0505 GCNN : 0.0718
0.1 Fmap Basic : 0.1453 0.1 Fmap Basic : 0.2199 0.1 Fmap Basic : 0.2291
Fmap Ours Opt: 0.0181 Fmap Ours Opt: 0.0610 Fmap Ours Opt: 0.0433
0 0 0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Geodesic error Geodesic error Geodesic error

Figure 13: Quantitative evaluation of pointwise correspondences comparing our method with Supervised Methods.

Unsupervised Methods on FAUST Original Unsupervised Methods on FAUST Remesh Unsupervised Methods on SCAPE Remesh
1 1 1

0.9 0.9 0.9

Fraction of correspondences

Fraction of correspondences

0.8
Fraction of correspondences

0.8 0.8

0.7 0.7 0.7

0.6 0.6 0.6

0.5 0.5 0.5

0.4 0.4 0.4

PMF Gauss : 0.0294
PMF Heat : 0.0173 PMF Gauss : 0.0389 PMF Gauss : 0.0547
0.3 0.3 PMF Heat : 0.0381 0.3 PMF Heat : 0.0472
BCICP : 0.0155
BCICP : 0.0501 BCICP : 0.0223
0.2 Ours-all : 0.0079 0.2 0.2
Ours-all: 0.0185 Ours-all : 0.0175
Fmap Basic : 0.4576 Fmap Basic : 0.3830
0.1 Fmap Basic : 0.3662 0.1
Fmap Ours Opt : 0.0098 0.1
Fmap Ours Opt : 0.02019 Fmap Ours Opt : 0.0140
0 0 0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Geodesic error Geodesic error Geodesic error

Figure 14: Quantitative evaluation of pointwise correspondences comparing our method with Unsupervised Methods.