Geometric Deep Learning

SIGGRAPH A SIA 2016 C OURSE N OTES

Organizers & Lecturers:

Jonathan Masci, Emanuele Rodolà, Davide Boscaini,

Michael M. Bronstein, Hao Li

Permission to make digital or hard copies of part or all of this work

for personal or classroom use is granted without fee provided that
copies are not made or distributed for profit or commercial
advantage and that copies bear this notice and the full citation on
the first page. Copyrights for third-party components of this work
must be honored. For all other uses, contact the Owner/Author.
Copyright is held by the owner/author(s).
SA '16 Courses, December 05-08, 2016, Macao
ACM 978-1-4503-4538-5/16/12.
http://dx.doi.org/10.1145/2988458.2988485
Contents

1 Introduction 1
1.1 Course goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 A brief history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 The challenges of geometric deep learning . . . . . . . . . . . . . . . . . . . 2

2 Basics of Deep Learning 4

2.1 Random forests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Neural networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Convolutional neural networks . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 Learning techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3 Extrinsic Deep Learning 11

3.1 Volumetric CNNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.1 3D Shape Representation . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.2 Network Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.3 Training and Data Collection . . . . . . . . . . . . . . . . . . . . . . 13
3.2 View-based CNNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.1 Classification Network . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 Human shape correspondence . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.1 Correspondence Computation . . . . . . . . . . . . . . . . . . . . . 17
3.3.2 Training Data Generation . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3.3 Network Design and Training . . . . . . . . . . . . . . . . . . . . . . 19
3.4 Performance Capture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4 Spectral Learning Methods 24

4.1 Spectral analysis on manifolds . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.2 Spectral descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.3 Optimal spectral descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.4 Deformable shape correspondence with random forests . . . . . . . . . . . 28
4.4.1 Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.4.2 Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.4.3 Shape matching via the label space . . . . . . . . . . . . . . . . . . . 29

5 Intrinsic Convolutional Neural Networks 31

5.1 Intrinsic CNNs in the spectral domain . . . . . . . . . . . . . . . . . . . . . 31
5.1.1 Spectral CNNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.1.2 Localized spectral CNNs . . . . . . . . . . . . . . . . . . . . . . . . 32
5.2 Intrinsic CNNs in the spatial domain . . . . . . . . . . . . . . . . . . . . . . 33

i
CONTENTS ii

5.2.1 Geodesic CNNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

5.2.2 Anisotropic diffusion CNNs . . . . . . . . . . . . . . . . . . . . . . . 35
5.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.3.1 Local descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.3.2 Correspondence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1

Introduction

1.1 Course goals

The goal of these course notes is to describe the main mathematical ideas behind geometric
deep learning and to provide implementation details for several applications in shape
analysis and synthesis, computer vision and computer graphics. The text in the course
materials is primarily based on previously published work including [71, 56, 86, 100, 10,
59, 97, 12, 11] among several others. With these notes we gather and provide a clear
picture of the key concepts and techniques that fall under the umbrella of geometric deep
learning, and illustrate the applications they enable. We also aim to provide practical
implementation details for the methods presented in these works, as well as suggest
further readings and extensions of these ideas.

1.2 Motivation
The past decade in computer vision research has witnessed the re-emergence of “deep
learning”, and in particular convolutional neural network (CNN) techniques, allowing to
learn powerful image feature representations from large collections of examples. CNNs
achieve a breakthrough in performance in a wide range of applications such as image
classification, segmentation, detection and annotation. Nevertheless, when attempting
to apply the CNN paradigm to 3D shapes (feature-based description, similarity, corre-
spondence, retrieval, etc.) one has to face fundamental differences between images and
geometric objects. Shape analysis and geometry processing pose new challenges that are
non-existent in image analysis, and deep learning methods have only recently started
penetrating into the 3D shape community. CNNs have been applied to 3D data in recent
works using standard (Euclidean) CNN architectures applied to volumetric or view-based
shape representations. Intrinsic versions of CNNs have also been proposed very recently
with the generalization of the CNN paradigm to non-Euclidean manifolds, allowing them
to deal with shape deformations. These “generalized” CNNs can be used to learn invariant
shape features and correspondence, allowing to achieve state-of-the-art performance in
several shape analysis tasks, while at the same time allowing for different shape rep-
resentations, e.g. meshes, point clouds, or graphs. The purpose of this short course is
to overview the foundations and the current state of the art on learning techniques for
3D shape analysis and geometry processing. Special focus will be put on deep learning
techniques (CNN) applied to Euclidean and non-Euclidean manifolds for tasks of shape
classification, retrieval, reconstruction and correspondence. The course will present in a
new light the problems of shape analysis, emphasizing the analogies and differences with

1
1. INTRODUCTION 2

the classical 2D setting, and showing how to adapt popular learning schemes in order to
deal with deformable shapes.
These course notes will assume no particular background, beyond some basic working
knowledge that is a common denominator for people in the field of computer graphics.
All the necessary notions and mathematical foundations will be described. The course is
targeted to graduate students, practitioners, and researchers interested in shape analysis,
synthesis, matching, retrieval, and big data.

1.3 A brief history

Deep learning methods have literally shaken many realms in the academia and industry
in the past few years. Technology giants like Apple, Google and Facebook have been
aggressively hunting for experts in the field and acquiring promising deep learning start-
up companies for unprecedented amounts, all of which are indicative of the enormous bets
the industry is currently placing on this technology. Nowadays, deep learning methods
are already widely used in commercial applications, including Siri speech recognition
in Apple iPhone, Google text translation, and Mobileye vision-based technology for
autonomously driving cars.
Though “deep learning” has become somewhat of a buzzword often taken out of
context and vaguely referring to artificial intelligence in general, the original term refers
to learning complicated concepts by a machine, by means of building them out of simpler
ones in a hierarchical or “multi-layer” manner. Artificial neural networks are a popular
realization of such deep multi-layer hierarchies inspired by the signal processing done
in the human brain. The first artificial neural network models are usually credited to
neuroscientists Hebb and Rosenblatt (late 1940s and 50s, respectively); deep architectures
and learning algorithms that resemble the modern ones can already be found in the
works of Ivahnenko and coauthors in the late 1960s, though the term “deep networks”
appeared much later [22, 3]. Research of artificial neural networks was thorny with cycles
of popularity and near oblivion. In the past few years, the growing computational power
of modern GPU-based computers, availability of large training datasets (“big data”),
and efficient stochastic optimization methods allowed creating and effectively training
complex network models with many layers and degrees of freedom. This allowed deep
neural networks achieve a qualitative breakthrough in performance on a wide variety of
tasks, from speech recognition and machine translation to image analysis and computer
vision, igniting the renaissance of the field.

1.4 The challenges of geometric deep learning

Dealing with signals such as speech, images, or video on 1D-, 2D- and 3D Euclidean
domains, respectively, has been the main focus of research in deep learning for the past
decades. However, in the recent years, more and more fields have to deal with data
residing on non-Euclidean geometric domains, which we call here “geometric data” for
brevity. For instance, in social networks, the characteristics of users can be modelled as
signals on the vertices of the social graph. Sensor networks are distributed interconnected
sensors, whose readings are modelled as time-dependent signals on graphs. In computer
graphics and vision, 3D shapes are modelled as Riemannian manifolds (surfaces) endowed
with properties such as color texture or motion field (e.g. dynamic meshes). Even more
complex examples include networks of operators, such as functional correspondences [66]
1. INTRODUCTION 3

Figure 1.1: Left: extrinsic methods such as volumetric CNNs [100] treat 3D geometric data
in its Euclidean representation. Such a representation is not invariant to deformations (e.g.,
in the shown example, the filter that responds to features on a straight cylinder would not
respond to a bent one). Right: in an intrinsic representation, the filter is applied on the
surface itself, thus being invariant to deformations.

or difference operators [74] in a collection of 3D shapes, or orientations of overlapping

cameras in multi-view vision (structure from motion) problems. Furthermore, modelling
high-dimensional data with graphs is an increasingly popular trend in general data science,
where graphs are used to describe the low-dimensional intrinsic structure of the data.
On the one hand, the complexity of geometric data and the availability of large datasets
(in the case of social networks, of the order of billions of examples) make it tempting
and very desirable to resort to machine learning techniques. On the other hand, the non-
Euclidean nature of such data implies that there are no such familiar properties as global
parametrization, common system of coordinates, vector space structure, or shift-invariance.
Consequently, basic operations such as linear combination or convolution that are taken
for granted in the Euclidean case, are even not well defined on non-Euclidean domains.
This happens to be a major obstacle that so far has precluded the use of successful deep
learning methods such as convolutional or recurrent neural networks on non-Euclidean
geometric data. As a result, the quantitative and qualitative breakthrough that deep
learning methods have brought into speech recognition, natural language processing, and
computer vision has not yet come to fields such as computer graphics or computational
sociology.
2

Basics of Deep Learning

In the recent years we have experienced a paradigm shift from axiomatic modeling to
data-driven modeling due to the unprecedented performance of Deep Neural Networks in
tasks such as image and speech recognition, machine translation, and image segmentation
and detection.
The main reasons to this success are to be found in: (1) the growing computational
power of modern GPU based computers, (2) larger annotated datasets, and (3) more
efficient stochastic optimization methods and architectural changes that allow to effectively
train much more complex networks with many layers and degrees of freedom.
These factors made the qualitative breakthrough possible, and the wide-spread interest
from various communities ranging from pure machine learning to system design rapidly
brought deep learning to be widely used in commercial applications, including Siri speech
recognition, Google text translation, and the Mobileye autonomous driving technology.
Key aspect of Deep Learning (DL) systems is that they are able to learn representations
directly from the raw input data without requiring any hand-crafted feature extraction
stage. They allow to perform representation learning at various levels of representation,
therefore in a hierarchical fashion, by composing simple non-linear building blocks usually
referred to as layers. This composition results in very powerful models which can leverage
the large quantities of annotated data available today.
This is the reason why we have experienced such an outstanding improvement in tasks
where all machine learning methods rely on a feature extraction stage, namely speech
signals, images, and everything related to perceptual data.
This is in sharp contrast with the now old paradigm of machine learning where
features had to be provided by the user. During the neural network “dark age”, when
methods such as SVM defined the gold standard for classification, very little effort was
devoted to the tweaking of their parameters and most of the work was instead devoted to
a good feature extraction strategy. As an example, the typical computer vision recognition
pipeline before DL consisted in computing SIFT, doing sparse vector quantization and
pooling to finally feed everything to an SVM classifier. Of course all of these stages are
completely arbitrary and sub-optimal to say the least. With DL one needs to have very
few assumptions on what is good for a given task and in most cases it is enough to have
the input image and its annotation; this leaves to the data-driven approach the task of
finding what is the right feature and the right feature aggregation strategy.
The ability to learn representations directly from the data with very little prior has
allowed many machine learning techniques to shine and to already surpass human
performance in some specific domains.
According to how many processing steps separate the input to the output of a given

4
2. BASICS OF DEEP LEARNING 5

model we obtain shallow and deep models. SVM are shallow models as they can be
mimicked by a single layer neural network (implementing the kernel function) whereas
models with more than a single layer start to be called deep. However, during the years
such term has had various meanings and currently it is used with architectures with at
least 10 or more layers.
Before introducing neural networks we first review the popular random forest paradigm.
Despite being composed of several decision steps, the random forest model is still consid-
ered to be shallow as the decisions at each level are taken on the original input space. In
deep networks each stage operates on the output of the previous one instead, therefore
performing feature learning.

2.1 Random forests

Random forests [13] are ensembles of decision trees that have become quite popular
in the computer vision and pattern recognition communities to solve classification and
regression problems. Applications of random forests range from object detection, tracking
and action recognition [29] to 3D pose estimation [90] to name just a few. The key idea
behind the random forest construction is that one can build a much stronger classifier
by combining several weak ones (the individual trees), which are trained independently
with a random sampling of the labelled training data [21]. This approach is particularly
appealing, since each tree can be trained efficiently, can be easily parallelized, and the
randomized feature selection allows to limit the correlation among trees and thus ensure
good generalization.
In these notes we will focus on the application of the random forest paradigm to solve
classification tasks, i.e., where the output of the forest is a discrete, categorical label – or in
its soft formulation, a probability distribution over some label space. In Section 4.4 we
will see how this model is useful to tackle shape matching problems.
As we mentioned, random forests are simple collections of decision trees, and the
output of a forest is an aggregation of the outputs of its constituent trees. Decision
trees are tree-structured classifiers that make a prediction by routing an input feature
sample along the interior nodes until it reaches a terminal node (or leaf), where the actual
classification takes place. Each interior node of the tree performs routing decisions based
on a (binary) test or “split” function ϕiΘi , parametrized by Θi . The test function is applied
to the incoming data, which is then sent to the left or right child depending on the
outcome of the test. In the inset figure, the path marked in bold illustrates the routing
of an input sample along a tree to reach the leaf `j ,
containing the final label distribution for that sample.
Designing a decision tree amounts to defining two
key ingredients: 1) the test functions ϕi associated to ϕi
each node, and 2) the label predictors associated to
each leaf (we refer the reader to Section 4.4 for exam-
ples). As is usually the case with machine learning
paradigms, we distinguish between a training and a
testing phase. `j
During training, a set of labelled training points is used to optimize the set of param-
eters {Θi } of the split functions. The training data that reaches a node is split into two
subsets (one per child node), such that the information gain incurred by the split is maxi-
mized (other energy functions are possible and depend on the task at hand). A key aspect
of the training process is the fact that randomness is injected into it: This is typically done
2. BASICS OF DEEP LEARNING 6

by using a random sample of all the training data, or by selecting the optimal parameters
at each node from a pool of randomly generated values. The training process is repeated
independently for each tree in the forest, leading to component trees that are different
from each other. This, in turn, leads to robustness to noisy data, de-correlation between
the individual tree predictions, and thus to improved generalization [21].
At testing time, a previously unseen data point is routed through each tree in the
forest, undergoing a number of predefined tests. When the data point reaches a leaf,
it is associated a class label (or a distribution over the space of labels). The final forest
prediction can be obtained, for example, by simply averaging the class posteriors of each
tree. Note that, differently from the training process, the testing phase is completely
deterministic once the trees are fixed.

2.2 Neural networks

Neural networks are learning methods able to extract hierarchical representations from
the data via several processing stages, defining a mapping from input to output which
exhibits multiple levels of abstraction.
Each stage is referred to as a layer, or basic building block, computing a very simple
function such as linear combinations and point-wise non-linearities. The concatenation of
such blocks into a DAG (directed acyclic graph) constitutes a feed-forward neural network,
whereas in the case of a DCG (directed cyclic graph) this results in recurrent models, of
which LSTM [41] is the most prominent example.
Each layer can be thought of as a function fi , where index i identifies the layer in the
hierarchy, parametrized by the (possibly empty) set Θi . Deep feed-forward networks are
composed of fully connected layers of the form

fc(x) = σ(Wx + b), (2.1)

where σ is a point-wise non-linearity such as ReLU, σ(x) = max(0, x). Other common
choices are s-shaped functions such as hyperbolic tangent and logistic. The learnable
parameters are the projection matrix W and the bias vector b, which for simplicity are
concatenated into the set Θ = {W, b}.
A deep feed-forward network is therefore a complex non-linear function resulting by
the composition of several such fully connected layers:

FΘ (x) = fn (. . . f1 (x, θ1 ), θn ), (2.2)

where Θ = {θi |i = 1..n} is the set of all parameters in each of the utilized layers.
In this document we will focus only on supervised feed-forward learning methods,
the most used approach in computer vision tasks, where the desired targets are known
in advance. In the cases where the input-to-output mapping incurs in a long list of
intermediate transformations, the models are said to be deep, otherwise they are referred to
as shallow. It is important to note that the meaning of the word deep has greatly changed
over the past few years given the recent advances in the field.

Unsupervised learning From a machine learning standpoint a very interesting and rela-
tively unexplored area is the one of unsupervised learning. In this setting the model knows
only the input signal and aims at modeling the generating input distribution that better
explains the data: the unknown function y is generating the input xi , and the function itself
is the target of the learning process. Unsupervised models usually aim at reconstructing
2. BASICS OF DEEP LEARNING 7

the input while limiting the representational power of the model via sparsity penalties for
example. Models in this family are called autoencoders. Another very popular paradigm
for unsupervised learning which does not rely on the ill-posed reconstruction cost is the
one of Predictability Minimization (PM) and Generative Adversarial Networks (GAN)
which are able to sample very realistic natural images and whose representations are often
competitive with fully supervised ones. A GAN model is composed by two networks,
the generator and the discriminator. The generator is trained to produce plausible input
samples from some input distribution, for example Gaussian noise, so that it makes the
discriminator fail. The discriminator is trained to classify samples from the training set
from samples generated by the generator. The result is a system able to learn features with
no supervision which compete with the ones obtained with fully supervised techniques
and quite impressive generative capabilities, in particular for scene and face generation
when the generator is a deep convolutional network [31, 69].

Supervised learning In a generic supervised learning setting we are given a training

dataset composed as follows
D = {(xi , ti )}i∈I , (2.3)
where xi ∈ Rn and the ground truth ti is generated by an unknown function y : Rn → Y;
ti = y(xi ). I is an index set of samples in the dataset. In a supervised classification task
the signal ti can be thought of as a discrete quantity with values in Y, and with cardinality
equal to the number of possible classes whereas for regression Y = Rm .
More formally, given a model and a training dataset, one needs to define a way of
assessing how good a solution is (by solution we mean a given Θ). This is done via the
definition of a loss function L, and the fitting of the parameters of the model is done
through minimization of such function:

Θ∗ = arg min LΘ (D) . (2.4)

Θ

For classification tasks, where each sample can be assigned to one of K mutually
exclusive classes, the common choice is the multinomial logistic loss
K
X
LΘ (xi , ti ) = − ti log(FΘ (xi )) (2.5)
k=1

where ti is one-hot encoded (vector containing a 1 at the position of the ground truth label
and 0 otherwise) and the output of the network F (xi ) is K-dimensional with sof tmax
activation; this allows to interpret the output as posterior probability of the class given the
input p(F (xi ) = ti |xy , Θ).
The minimization of the loss function is performed through some form of gradient
descent and the gradient of the parameters of the network is computed with the backprop-
agation algorithm [72, 99].

2.3 Convolutional neural networks

Convolutional Networks are a particular class of neural network models whose main
building blocks are convolution (C) and pooling (P).
What is nowadays known as convolutional network dates back to the Neocognitron
of Fukushima [28] who first proposed to perform local operations with weight sharing.
The author did not consider backpropagation, but a form of Hebbian learning instead.
2. BASICS OF DEEP LEARNING 8

LeCun and colleagues [50] later introduced gradient based learning and advanced the
state-of-the-art for this class of models since then. We can say that what is now known as
convolutional network is the latter variant of LeCun et al.
The convolutional layer C acts on a p-dimensional input f (x) = (f1 (x), . . . , fp (x)) and
applies a filter-bank W = (wl,l0 ∈ Rrc ), where r and c indicate the number of rows and
columns in the convolutional kernel; l and l0 indicate the number of input and output
maps:
p
X
gl (x) = (fl0 ? wl,l0 )(x). (2.6)
l0 =1

After a convolutional layer, as with fully-connected layers, a point-wise non-linear

activation function is applied; a common choice is to use ReLU (rectified linear units) of
the form σ(x) = max(0, x).
Convolutional filters have a local support, usually of very few pixels, however recently
some authors [102, 107] have revised the classical dilated convolution which operates at
increasing spatial resolutions and that can be efficiently learned using the well known
algorithme à trous [58].
In order to introduce some invariance to small translations a pooling layer P can follow
the convolution. It performs the following operation

gl (x) = G(fl (x0 ) : x0 ∈ Neigh(x)), (2.7)

where Neigh is a neighborhood around the point x and G is a permutation invariant

aggregation function, usually Lp -norm with p = ∞ being the most popular and effective
choice resulting in max-pooling. While pooling has been introduced to subsample the
signal over non-overlapping neighborhoods – tiling the input image into regions of size
2x2 or 3x3 for example – it has become popular also to do overlapping pooling which
results in a morphological dilation operation.
While initially popularized for pure classification tasks CNN thanks to the capability
to reduce dimensionality through pooling operators they have immediately found appli-
cation on tasks where the dimensionality does not need to be reduced, such as semantic
segmentation where they excel by far the non-DL based approaches. In order to avoid
losing spatial resolution deconvolution operations have been introduced, among which
the simplest being the up-sampling with linear interpolation. Thanks to this many authors
have started to tackle tasks where a prediction for every pixel in the input image needs
to be taken, this comprehends optical flow estimation, semantic segmentation, image
generation, volumetric based shape synthesis [24, 57, 105, 25].
Spatial Transformer Networks are a different class of models which aim to learn
a perspective transformation of the input image, more generally a distortion field in
the input space, so that the classification and detection tasks can be carried out more
easily because of the suppression of unwanted signals such as the background and to the
eventual rescaling and centering of the object of interest. They have been shown to deliver
excellent performance in digit classification without data augmentation [44].
Attention based models instead [85, 62, 32, 101] aim at inferring an attention mask
over the input or the hidden states of a network to emphasize only certain aspect of it.
These models have been successfully used for tasks such as image captioning, generation,
and speech recognition.
2. BASICS OF DEEP LEARNING 9

2.4 Learning techniques

Obtaining the parameters of deep learning models by minimizing a task-specific cost
seems straightforward as these models are differentiable and therefore any gradient based
technique could be used. This learning process is non-convex and it has always been linked
with bad local minima; several techniques have been investigated to indeed mitigate such
eventuality. A very interesting approach is the one called flat minima search [40] which
aims at finding regions in the weight spaces where a small perturbation does not change
the loss; in a bad local minima a tiny change of one parameter may in fact lead to huge
differences in the loss. However it turns out that simpler stochastic optimization methods,
in conjunction with much larger and overparameterized models, offer excellent empirical
performance. Understanding the structure of the optimization problems, considering the
growing interest in these methods, has of course let to a very active area of research [17, 45].
Even with the latest advances deep models are still much harder to train than shallow
ones but attain much better performance when properly trained. This has opened up
many interesting research directions to allow gradients to flow more easily as in Highway
and Residual Networks [83, 34], better behaved activation functions as the Exponential
Linear Unit [18], and understanding the training process itself from a “model of models”
standpoint [84].
Having high capacity models allows to escape better local minima, and allows more
complex mappings which could better capture the true important factors in the input data.
However this makes possibly successful models prone to severe overfitting, meaning that
the discrepancy between training performance and test performance is high and that they
do not generalize well to unseen data.
Many techniques that helped breaking many of the records in pattern recognition aim
indeed at reducing overfitting and are implemented as an additional weighted term to the
loss function.
Here we briefly list them and refer the reader to the respective publications for details.

Pre-training Several works have shown, in the early days of deep learning, that initial-
izing the weights by the ones obtained via unsupervised learning is beneficial and can
lead improved and more robust results [27, 60, 51, 38, 93]. This process involves though a
first training phase which can take lot of time and that has a diminishing return when the
amount of labeled samples increases and has become less popular in practical applications.
What is common in computer vision though is to take a model trained to classify objects in
ImageNet [23] such as the VGG [81] or Inception [89] and to use its features or predictions
for other higher level tasks.

Initialization is a critical point for deep networks and finding good initializations was
one of the good arguments of pre-training. While initializing the weights from a normal
distribution with zero-mean and small standard deviation is often sufficient to achieve
good convergence properties, several authors have shown that ad-hoc initializations are
often more powerful and can lead to better solutions [30, 35].

Weight decay is perhaps the most popular regularization technique for deep learning.
It simply penalizes weights with large magnitude thus favoring smooth solutions and is
defined by the regularizer kΘk22
2. BASICS OF DEEP LEARNING 10

Dropout is a very successful technique to prevent co-adaptation of features [39]; it is

among the simplest techniques to implement (a layer with no parameters) and because
of this it is widely used. At training time each dimension at the input is dropped with a
given probability p thus enforcing features to code for different things and to not rely on
the presence of the other features. It can be seen as well as a very effective way of doing
model averaging over an exponential number of networks, each one identified by the
zero-one mask produced by the sampling. At test time, instead of doing an expensive
averaging over an exponential number of possible masks, the input is rescaled by p; this
cheap trick works reasonably well in feed forward networks and it’s the classical way of
applying dropout at test time. Given its popularity several follow up works have shown
either improvements in terms of speed [96], applicability to recurrent networks [67], and
theory [7, 47].

Batch Normalization [43] is a regularization that is implemented in form of a layer

in deep networks and it is capable of reducing training times of very large models con-
siderably, while also improving generalization. It normalizes each mini-batch during
stochastic optimization to have zero mean and unit variance (estimates are build with
moving averages during training), and then performs a parametric affine transformation
to produce the output. After training, one can re-estimate the statistics on the test set or
simply keep the training set estimates.
3

Extrinsic Deep Learning

3.1 Volumetric CNNs

Convolutional Neural Networks have demonstrated to outperform hand-crafted features
and domain specific techniques for various core tasks in computer vision. These applica-
tions include classification [48], segmentation [57, 64, 75], regression [65], and synthesis
tasks [25]. A wide range of applications have been recently introduced that process 3D
geometric shapes, such as highly effective 3D object detection for RGB-D data [82], object
classifcation of point clouds [61], 3D local feature matching [106], and 3D deformation
flows [103].
A natural extention to classic CNNs that process 2D images is to process 3D data using
a volumetric represencation and 3D convolutions. Wu and coworkers [100] presented 3D
ShapeNets, a 3D deep learning framework for modeling shapes using a voxel discretization
of 3D shapes, and demonstrated how spatial features can be learned directly in 3D. The
approach consists of representing a geometric 3D shape as a probabilistic distribution of
binary variables on a voxel grid. A convolutional deep belief network is used to learn
the joint distribution of all 3D voxels based on massive amounts of 3D training data
such as the CAD model database, ModelNet, and real-world scans from the NYU depth
dataset [80]. In this work, they have demonstrated highly accurate object recognition from
input depth maps and the ability to synthesize the missing parts of single-view input data.

3.1.1 3D Shape Representation

To use a volumetric convolutional neural network, a 3D shape in the form of a mesh can
be represented as a probability distribution of binary values on a 3D grid. We define the
inside of a mesh surface using a voxel value being 1 and the outside (or empty space)
as 0. A relatively small grid resolution of 30 × 30 × 30 is used in the experiments of
Wu et al. [100]. A deep belief network (DBN) as introduced by Hinton et al. [37], which
describes the joint probabilistic distribution over all pixels and labels in a 2D image, can
be directly extended to the 3D domain by replacing pixels with voxels. Since the memory
requirement increases drastically with the resolution of the 3D voxel volume, a fully
connected convolutional DBN would require a massive amount of parameters to train.
Wu et al. [100] therefore proposed to use a weight sharing approach to reduce the number
of model parameters.
The convolutional deep learning model does not use pooling as it could yield greater

11
3. EXTRINSIC DEEP LEARNING 12

Figure 3.1: Left: network architecture of 3D ShapeNets for high-dimensional feature

learning of 3D voxel input. Right: for every neuron, the 100 training examples with the
highest response are averaged and their volume cropped inside the receptive field (gray).
An implicit surface can be extracted via zero-crossing (red). The network captures the
structures of low-level surface features (L1) to object parts (L2,L3) and entire objects (L4).
Figure reproduced from [100].

uncertainty and the energy E for each convolutional layer is computed as:
XX f 
hj (W f ? v)j + cf hfj −
X
E(v, h) = − bl vl (3.1)
f j l

hfj
where vl is a visibility unit, a hidden unit in the feature channel f , W f a convolutional
filter, and ? a convolution operation. To help with the reconstruction, each visibility
unit vl is associated with a unique bias term bl and all hidden units hfj in the same
convolution channel share the same bias term cf . The extracted features can then be used
for classification, recognition, and reconstruction tasks and similar to classic 2D CNNs,
they significantly outperform existing methods as shown in [100].

3.1.2 Network Architecture

The 3D shapes in [100] have relatively low resolution and are represented by 24 × 24 × 24
voxels and 3 additional cells for padding in order to reduce convolution artifacts. The
network architecture of 3D ShapeNets is depicted in Figure 3.1, where only one filter is
visualized for each convolutional layer for simplicity. Layer 1 consists of 48 filters of size
6 with stride 2, layer 2 has 160 filters of size 5 with stride 2, layer 3 has 512 filters of size
4, and each convolution filter is connected to all feature channels in the previous layer.
Layer 4 is a fully connected Restricted Boltzmann Machine (RBM) which consists of 1200
hidden units. Layer 5 consists of 4000 hidden units which takes as input a combination of
multinomial label variables and Bernouilli feature variables.
3. EXTRINSIC DEEP LEARNING 13

network training (offline) correspondence computation (online)

SCAPE MIT Yobi3D input output

data set data set data set depth map correspondences

labels depth map labels depth map labels depth map

feature per pixel closest
extraction descriptor feature
feature feature feature shared
extraction extraction extraction network
parameters
per vertex
descriptor descriptor descriptor input descriptors
full model

classification classification classification classification classification loss

... ... function feature
prediction 1 prediction N prediction 1 prediction N prediction depth depth extraction
generator maps and
averaging

Figure 3.2: A neural network can be trained to extract a feature descriptor and predict
the corresponding segmentation label on the human body surface for each point in the
input depth maps. Per-vertex descriptors are generated for 3D models by averaging the
feature descriptors in their rendered depth maps. The extracted features can then be used
for example to compute dense correspondences.

3.1.3 Training and Data Collection

The model is first pre-trained layer-by-layer using a contrastive divergence approach [36],
then fine-tuned using a wake sleep approach as detailed in [37]. In the pre-training phase
of the first layer, learning signals are only collected in non-empty receptive fields. In
this way we can minimize learning distraction in largely unoccupied cells for the RBM.
The learned filters are shown in Figure 3.1. As with any deep neural network, a large
collection of training data is required to effectively handle intra-class variations. Wu and
colleagues [100] have introduced the database ModelNet which combines objects (vehicles,
furniture, items, etc.) from 3D Warehouse, Yobi3D, Princeton Shape Benchmark, and the
SUN database. Amazon Mechanical Turk is used to remove mis-labeled, unrealistic,
duplicate, and irrelevant objects. The final database contains 151K 3D CAD models from
660 categories. During training, 40 common object categories are selected from ModelNet,
each containing 100 unique CAD models. The data is augmented using a combination
of rotation and translations resulting in 12 poses per object. Pre-training takes roughly 2
days on an Intel XEON E5-2690 CPU with an NVIDIA K40c GPU.

3.2 View-based CNNs

This section introduces a deep learning framework to compute high dimensional feature
descriptors for classification tasks. We will show later that such classification network can
be used for dense correspondence computation across full or partial human shapes. Such
system is trained with depth maps of humans in arbitrary pose and with varying clothing.
Categories of data other than clothed humans are also possible.
Traditional classification neural networks tend to separate the embedding of surface
points lying in different but nearby classes. Thus, using such learned feature descriptors
for correspondence matching between deformed surfaces often results in significant
outliers at the segmentation boundaries. In order to alleviate this problem, the idea is
to leverage repeated mesh segmentations to produce smoother embeddings into feature
space. This technique maps shape points that are geodesically close on the surface of their
corresponding 3D model to nearby points in the feature space. As a result, not only are
outliers considerably reduced during deformable shape matching, but it can be also shown
that the amount of training data can be drastically reduced compared to conventional
3. EXTRINSIC DEEP LEARNING 14

learning methods.
Given depth maps of two humans I1 , I2 , one important application is to determine
which two regions Ri ⊂ Ii of the depth maps come from corresponding parts of the body,
and to find the correspondence map φ : R1 → R2 between them. The strategy for doing so
is to formulate the correspondence problem first as a classification problem: first, a feature
descriptor f : I → Rd , which maps each pixel in a single depth image to a feature vector,
is learned. These feature descriptors are then used to establish correspondences across
depth maps (see Figure 3.2).

3.2.1 Classification Network

The feature vector output of a CNN for the classification task of 3D objects should satisfy
two properties:

1. f depends only on the pixel location on the human body, so that if two pixels are
sampled from the same anatomical location on depth scans of two different humans,
their feature vector should be nearly identical, irrespective of pose, clothing, body
shape, and angle from which the depth image was captured;

2. kf (p) − f (q)k is small when p and q represent nearby points on the human body, and
large for distant points.

The literature takes two different approaches to enforcing these properties when learning
descriptors using convolutional neural networks. Direct methods include in their loss
functions terms penalizing failure of these properties (by using e.g. Siamese or triplet-loss
energies). However, it is not trivial how to sample a dense set of training pairs or triplets
that can all contribute to training [76]. Indirect methods instead optimize the network
architecture to perform classification. The network consists of a descriptor extraction tower
and a classification layer, and peeling off the classification layer after training leaves the
learned descriptor network (for example, many applications use descriptors extracted from
the second-to-last layer of the AlexNet). This approach works since classification networks
tend to assign similar (dissimilar) descriptors to the input points belonging to the same
(different) class, and thus satisfy the above properties implicitly. An indirect approach
is taken, as the experiments suggest that an indirect method that uses an ensemble of
classification tasks has better performance and computational efficiency.

Descriptor learning as ensemble classification. There are two challenges to learning a

feature descriptor for depth images of human models using this indirect approach. First,
the training data is heterogenous: between different human models, it is only possible
to obtain a sparse set of key point correspondences, while for different poses of the
same person, we may have dense pixel-wise correspondences (e.g., SCAPE [5]). Second,
smoothness of descriptors learned through classification is not explicitly enforced. Even
though some classes tend to be closer to each other than the others in reality, the network
treats all classes equally.
To address both challenges, per-pixel descriptors can be learned for depth images
by first training a network to solve a group of classification problems, using a single
feature extraction tower shared by the different classification tasks. This strategy allows
to combine different types of training data as well as designing classification tasks for
various objectives. Formally, suppose there are M classification problems Ci , 1 ≤ i ≤ M .
Denote the parameters to be learned in classification problem Ci as (wi , w), where wi and
3. EXTRINSIC DEEP LEARNING 15

training mesh segmentation 1 segmentation 2 segmentation 3

Figure 3.3: To ensure smooth descriptors, [97] defines a classification problem for multiple
segmentations of the human body. Nearby points on the body are likely to be assigned
the same label in at least one segmentation.

w are the parameters corresponding to the classification layer and descriptor extraction
tower, respectively. The descriptor learning is defined as minimizing a combination of
loss functions of all classification problems:
M
X
{wi? }, w? = arg min L(wi , w). (3.2)
{wi },w i=1

After training, the optimized descriptor extraction tower becomes the output. It is easy to
see that when wi , w are given by convolutional neural networks, Eq. (3.2) can be effectively
optimized using stochastic gradient descent through back-propagation.
To address the challenge of heterogenous training sets, two types of classification tasks
are included in this ensemble: one for classifying key points, used for iter-subject training
where only sparse ground-truth correspondences are available, and one for classifying
dense pixel-wise labels, e.g., by segmenting models into patches (See Figure 3.3), used for
intra-subject training. Both contribute to the learning of the descriptor extraction tower.
To ensure descriptor smoothness, instead of introducing additional terms in the loss
function, a simple yet effective strategy is proposed that randomizes the dense-label
generation procedure. Specifically, as shown in Figure 3.3, multiple segmentations of the
same person are considered, and a classification problem for each is introduced. Clearly,
identical points will always be associated with the same label and far-apart points will be
associated with different labels. Yet for other points, the number of times that they are
associated with the same label is related to the distance between them. Consequently, the
similarity of the feature descriptors are correlated to the distance between them on the
human body resulting in a smooth embedding satisfying the desired properties discussed
in the beginning of the section.

3.3 Human shape correspondence

The computation of correspondences between geometric shapes is a fundamental building
block for many important tasks in 3D computer vision, such as reconstruction, tracking,
analysis, and recognition. Temporally-coherent sequences of partial scans of an object can
be aligned by first finding corresponding points in overlapping regions, then recovering
the motion by tracking surface points through a sequence of 3D data; semantics can be
3. EXTRINSIC DEEP LEARNING 16

0 1 2 3 4 5 6 7 8 9 10
layer image conv max conv max 2×conv conv max 2×conv int conv
filter-stride - 11-4 3-2 5-1 3-2 3-1 3-1 3-2 1-1 - 3-1
channel 1 96 96 256 256 384 256 256 4096 4096 16
activation - relu lrn relu lrn relu relu idn relu idn relu
size 512 128 64 64 32 32 32 16 16 128 512
num 1 1 4 4 16 16 16 64 64 1 1

Table 3.1: The end-to-end network architecture generates a per-pixel feature descriptor and
a classification label for all pixels in a depth map simultaneously. From top to bottom
in column: The filter size and the stride, the number of filters, the type of the activation
function, the size of the image after filtering and the number of copies reserved for
up-sampling.

extracted by fitting a 3D template model to an unstructured input scan. With the popu-
larization of commodity 3D scanners and recent advances in correspondence algorithms
for deformable shapes, human bodies can now be easily digitized [53, 63, 26] and their
performances captured using a single RGB-D sensor [52, 91].
Most techniques are based on robust non-rigid surface registration methods that can
handle complex skin and cloth deformations, as well as large regions of missing data due
to occlusions. Because geometric features can be ambiguous and difficult to identify and
match, the success of these techniques generally relies on the deformation between source
and target shapes being reasonably small, with sufficient overlap. While local shape
descriptors [73] can be used to determine correspondences between surfaces that are far
apart, they are typically sparse and prone to false matches, which require manual clean-
up. Dense correspondences between shapes with larger deformations can be obtained
reliably using statistical models of human shapes [5, 9], but the subject has to be naked [8].
For clothed bodies, the automatic computation of dense mappings [46, 54, 71, 16] has
been demonstrated on full surfaces with significant shape variations, but are limited to
compatible or zero-genus surface topologies.
The problem of estimating accurate dense correspondence between deformable partial
shapes has been investigated recently in the works of Rodolà and colleagues [70, 20, 55],
and has been the target of two recent SHREC’16 Correspondence benchmarks [19, 49].
Despite the recent interest, however, this problem has remained relatively unexplored in
the literature due to the several challenges it entails.
Wei and colleagues [97] recently introduced a deep neural network structure for com-
puting dense correspondences between shapes of clothed subjects in arbitrary complex
poses. The input surfaces can be a full model, a partial scan, or a depth map, maximiz-
ing the range of possible applications. Their system is trained with a large dataset of
depth maps generated from the human bodies of the SCAPE database [5], as well as
from clothed subjects of the Yobi3D [2] and MIT [94] dataset. While all meshes in the
SCAPE database are in full correspondence, they manually labeled the clothed 3D body
models. They combine both training datasets and learned a global feature descriptor using
a network structure that is well-suited for the unified treatment of different training data
(bodies, clothed subjects). Similar to the unified embedding approach of FaceNet [76], the
AlexNet [48] classification network can be used to learn distinctive feature vectors for dif-
ferent subregions of the human body. While the performance of this dense correspondence
computation is comparable to state of the art techniques between two full models, they
also demonstrate that learning shape priors of clothed subjects can yield highly accurate
matches between partial-to-full and partial-to-partial shapes. Their examples include fully
3. EXTRINSIC DEEP LEARNING 17

clothed individuals in a variety of complex poses and the effectiveness of this approach
has been demonstrated on a template based performance capture application that uses a
single RGB-D camera as input.

3.3.1 Correspondence Computation

This trained classification network can be used to extract per-pixel feature descriptors for
depth maps. For full or partial 3D scans, the authors first render depth maps from multi-
ple viewpoints and compute a per-vertex feature descriptor by averaging the per-pixel
descriptors of the depth maps. These descriptors are used to establish correspondences
simply by a nearest neighbor search in the feature space (see Figure 3.2).
For applications that require deforming one surface to align with the other, one can
fit the correspondences described in this chapter into any existing deformation method
to generate the alignment. One possibility is to use the efficient as-rigid-as possible
deformation model described in [52].
This section demonstrates the performance of deep learning-based correspondence
computation evaluated on various real and synthetic datasets, naked and clothed subjects,
as well as full and partial matching for challenging examples as illustrated in Figure 3.7.
The real capture data examples (last column) are obtained using a Kinect One (v2) RGB-D
sensor and demonstrate the effectiveness for real life scenarios. Each partial data is a
single depth map frame with 512 × 424 pixels and the full template model is obtained
using the non-rigid 3D reconstruction algorithm of [53]. All examples include complex
poses (side views and bended postures), challenging garment (dresses and vests), and
props (backpacks and hats).
Four different synthetic datasets are used to provide a quantitative evaluation. The
3D models from both SCAPE and MIT databases are part of the training data of the
deep neural network described in the previous Sections, while the FAUST and Mixamo
models [1] are not used for training. The SCAPE and FAUST data sets are exclusively
naked human body models, while the MIT and Mixamo models are clothed subjects.
For all synthetic examples, the partial scans are generated by rendering depth maps
from a single camera viewpoint. The Adobe Fuse and Mixamo softwares [1] were used
to procedurally model realistic characters and generate complex animation sequences
through a motion library provided by the software.
The correspondence colorizations validate the accuracy, smoothness, and consistency
of the dense matchings in extreme situations, including topological variations between
source and target. Notice how the correspondences between front and back views are
being correctly identified in the real capture 1 example for the full-to-partial matchings.
Popular skeleton extraction methods from single-view 3D captures such as [78, 98, 92]
often have difficulties resolving this ambiguity.

Comparisons. Surface matching techniques which are not restricted to naked human
body shapes are currently the most suitable solutions for handling subjects with clothing.
Though robust to partial input scans such as single-view RGB-D data, cutting edge non-
rigid registration techniques [42, 52] often fail to converge for large scale deformations
without additional manual guidance as shown in Figure 3.4. When both source and
target shapes are full models, an automatic mapping between shapes with considerable
deformations becomes possible as shown in [46, 54, 71, 16]. This method is compared with
the recent work of Chen et al. [16] and computes correspondences between pairs of scans
sampled from the same (intra-subject) and different (inter-subject) subjects. Chen et al.
evaluate a rich set of methods on randomly sampled pairs from the FAUST database [9]
3. EXTRINSIC DEEP LEARNING 18

intra AE intra WE inter AE inter WE

Chen et al. (unsupervised) 4.49 10.96 5.95 14.18
Wei et al. (learning-based) 2.00 9.98 2.35 10.12

Table 3.2: Comparison between the method described in this Section and the recent work
of Chen et al. [16] by computing correspondences for intra- and inter-subject pairs from
the FAUST data set. We show the average error on all pairs (AE, in centimeters) and
the average error on the worst pair for each technique (WE, in centimeters). While the
learning-based technique may introduce worse WE, overall accuracies are improved in
both cases.

source / target [Wei et al. 16] [Li et al. 09] [Huang et al. 08]

Figure 3.4: Compared to other non-rigid registration algorithms, larger deformations

between a full template and a partial scan can be handled.

and report the state of the art results for their method. For a fair comparison, this method is
also evaluated on the same set of pairs. As shown in Table 3.2, the learning-based method
described here improves the average accuracy for both the intra- and the inter-subject
pairs. Note that by using simple AlexNet structure, an average accuracy of 10 cm can be
achieved. However, if multiple segmentations are not adapted to enforce smoothness, the
worst average error can be up to 30 cm.

Limitations. Like any supervised learning approach, this framework cannot handle
arbitrary shapes as the prior is entirely based on the class of training data. Despite
superior performance compared to the state of the art, the current implementation is
far from perfect. For poses and clothings that are significantly different than those from
the training data set, this method still produces wrong correspondences. However, the
outliers are often grouped together due to the enforced smoothness of the embedding,
which could be advantageous for outlier detection. Due to the limited memory capacity
of existing GPUs, this approach requires downsizing of the training input, and hence the
correspondence resolutions are limited to 512 × 512 depth map pixels.

Performance. The shown experiments are performed on a 6-core Intel Core i7-5930K
Processor with 3.9 GHz and 16GB RAM. Both offline training and online correspondence
computation run on an NVIDIA GeForce TITAN X (12GB GDDR5) GPU. While the com-
plete training of the neural network takes about 250 hours of computation, the extraction
of all the feature descriptors never exceeds 1 ms for each depth map. The subsequent
correspondence computation with these feature descriptors varies between 0.5 and 1 s,
depending on the resolution of the input data.
3. EXTRINSIC DEEP LEARNING 19

SCAPE MIT Yobi3D Yobi3D Yobi3D

Figure 3.5: Sparse key point annotations of 33 landmarks across clothed human models of
different datasets.

3.3.2 Training Data Generation

Collecting 3D Shapes. To generate the training data for the network, 3D models are
collected from three major resources: the SCAPE [5], the MIT [94], and the Yobi3D [2] data
sets. The SCAPE database provides 71 registered meshes of one person in different poses.
The MIT dataset contains the animation sequences of three different characters. Similar to
SCAPE, the models of the same person have dense ground truth correspondences. All
animation sequences are used except for the samba and swing ones, which can be reserved
for evaluation. Yobi3D is an online repository that contains a diverse set of 2000 digital
characters with varying clothing. Note that the Yobi3D dataset covers the shape variability
in local geometry, while the SCAPE and the MIT datasets cover the variability in pose.
Simulated Scans. Each model is rendered from 144 different viewpoints to generate
training depth images. A depth image resolution of 512 × 512 pixels is used, where the
rendered human character covers roughly half of the height of the depth image. This
setup is comparable to those captured from commercial depth cameras; for instance, the
Kinect One (v2) camera provides a depth map with resolution 512 × 424, where a human
of height 1.7 meters standing 2.5 meters away from the camera has a height of around 288
pixels in the depth image.
Key-point annotations. Human experts are used to annotate 33 key points across the
input models as shown in Figure 3.5. These key points cover a rich set of salient points
that are shared by different human models (e.g. left shoulder, right shoulder, left hip,
right hip etc.). Note that for shapes in the SCAPE and MIT datasets, only one rest-shape
is annotated and the ground-truth correspondences are used to propagate annotations.
The annotated key points are then propagated to simulated scans, providing 33 classes for
training.
500-patch segmentation generation. Each distinctive model in the model collection is di-
vided into multiple 500-patch segmentations. Each segmentation is generated by randomly
picking 10 points on each model, and then adding the remaining points via furthest point-
sampling. In total, 100 pre-computed segmentations are used. Each such segmentation
provides 500 classes for depth scans of the same person (with different poses).

3.3.3 Network Design and Training

The neural network structure for training consists of a descriptor extraction tower and a
classification module.
3. EXTRINSIC DEEP LEARNING 20

Extraction tower. The descriptor extraction tower takes a depth image as input and
extracts for each pixel a dimension d (d = 16 in this paper) descriptor vector. A popular
choice is to let the network extract each pixel descriptor using a neighboring patch (c.f.[33,
104]). However, such a strategy is too expensive in this setting as this has to be computed
for dozens of thousands of patches per scan.
The strategy is to design a network that takes the entire depth image as input and
simultaneously outputs a descriptor for each pixel. Compared with the patch-based
strategy, the computation of patch descriptors are largely shared among adjacent patches,
making descriptor computation fairly efficient in testing time.
Table 3.1 describes the proposed network architecture. The first 7 layers are adapted
from the AlexNet architecture. Specifically, the first layer downsamples the input image
by a factor of 4. This downsampling not only makes the computations faster and more
memory efficient, but also removes salt-and-pepper noise which is typical in the output
from depth cameras. Moreover, a similar strategy described in [77] is adapted to modify
the pooling and inner product layers so that the original image resolution can be recov-
ered through upsampling. The final layer performs upsampling by using neighborhood
information in a 3-by-3 window. This upsampling implicitly performs linear smoothing
between the descriptors of neighboring pixels. It is possible to further smooth the descrip-
tors of neighboring pixels in a post-processing step, but as shown in the results, this is not
necessary since the network is capable of extracting smooth and reliable descriptors.

Classification module. The classification module receives the per-pixel descriptors and
predicts a class for each annotated pixel (i.e., either key points in the 33-class case or all
pixels in the 500-class case). Note that one layer for each segmentation of each person is
introduced in the SCAPE and the MIT datasets and one shared layer for all the key points.
Similar to AlexNet, softmax is used when defining the loss function.

Training. The network is trained using a variant of stochastic gradient descent. Specif-
ically, a task (i.e., key points or dense labels) is randomly picked for a random partial
scan and fed into the network for training. If the task is dense labels, a segmentation is
randomly selected among all possible segmentations. The network can be tuned with
200,000 iterations using a batch size of 128 key points or dense labels which may come
from multiple datasets.

3.4 Performance Capture

The presented corrrespondence computation can be used for template based performance
capture applications using a depth map sequence captured from a single RGB-D sensor.
The complete geometry and motion is reconstructed in every sequence by deforming a
given template model to match the partial scans at each incoming frame of the perfor-
mance. Unlike existing methods [88, 52, 95, 91] which track a template using the previous
frame, here the template model is always deformed from its canonical rest pose using
the computed full-to-partial correspondences in order to avoid potential drifts. Defor-
mation is achieved using the robust non-rigid registration algorithm presented in Li et
al. [52], where the closest point correspondences are replaced with the ones obtained from
the presented method. Even though the correspondences are computed independently
in every frame, we observe a temporally consistent matching during smooth motions
without enforcing temporal coherence as with existing performance capture techniques
as shown in Figure 3.6. Since the deep learning framework does not require source and
3. EXTRINSIC DEEP LEARNING 21

target shapes to be close, it can effectively handle large and instantaneous motions. For
the real capture data, we visualize the reconstructed template model at every frame and
for the synthetic model we show the error to the ground truth.
3. EXTRINSIC DEEP LEARNING 22

real capture

input depth map

dense correspondence

dynamic geometry reconstruction

Mixamo (synthetic)

input depth map

dense correspondence

error (in cm): 0 20

Figure 3.6: Geometry and motion reconstruction is performed by deforming a template

model to captured data at each frame using the correspondences computed by the deep
learning-based method. Even though temporal coherence is not enforced explicitly, we
observe faithful and smooth reconstructions. Shown here are examples with both real and
synthetic data.
3. EXTRINSIC DEEP LEARNING 23

error (in cm): 0 20

full-to-full

real capture 1
SCAPE

FAUST
source target error source target error source target

real capture 2
Mixamo
MIT

source target error source target error source target

full-to-partial

real capture 1
SCAPE

FAUST

source target error source target error source target

real capture 2
Mixamo
MIT

source target error source target error source target

partial-to-partial
real capture 1
SCAPE

FAUST

source target error source target error source target

real capture 2
Mixamo
MIT

source target error source target error source target

Figure 3.7: Deep learning-based correspondence computation can handle full-to-full,

partial-to-full, and partial-to-partial matchings between full 3D models and partial scans
generated from a single depth map. This method is evaluated on various real and synthetic
datasets.
4

Spectral Learning Methods

4.1 Spectral analysis on manifolds

We model a 3D shape as a connected smooth compact two-dimensional manifold (sur-
face) X, possibly with a boundary ∂X. Locally around each point x the manifold is
homeomorphic to a two-dimensional Euclidean space referred to as the tangent plane and
denoted by Tx X. The collection (disjoint union) of tangent spaces at all points is referred
to as the tangent bundle and denoted by T X. A Riemannian metric is an inner product
h·, ·iTx X : Tx X × Tx X → R on the tangent space depending smoothly on x.
It is important to note that the definition of a Riemannian manifold is completely
abstract and does not require a geometric realization in any space. However, a Riemannian
manifold can be realized as a subset of a Euclidean space (in which case it is said to
be embedded in that space) by using the structure of the Euclidean space to induce a
Riemannian metric. The Nash Embedding Theorem guarantees that any sufficiently smooth
Riemannian manifold can be realized in a Euclidean space of sufficiently high dimension.
An embedding is not necessarily unique; two different realizations of a Riemannian metric
are called isometries.

Calculus on manifolds A scalar field is a smooth real function f : X → R on the manifold.

A tangent vector field F : X → T X is a mapping attaching a tangent vector F (x) ∈ Tx X
to each point x. We define the Hilbert spaces of scalar and vector fields on manifolds,
denoted by L2 (X) and L2 (T X), respectively, with the following inner products:
Z
hf, giL2 (X) = f (x)g(x)dx; (4.1)
ZX
hF, GiL2 (T X) = hF (x), G(x)iTx X dx; (4.2)
X

dx denotes here a volume element induced by the Riemannian metric.

In calculus, the notion of derivative describes how the value of a function changes with
an infinitesimal change of its argument. One of the big differences distinguishing classical
calculus from differential geometry is a lack of vector space structure on the manifold,
prohibiting us from naïvely using expressions like f (x + dx). The conceptual leap that is
required to generalize such notions to manifolds is the need to work locally in the tangent
space.
The differential of f as an operator df : T X → R acting on tangent vector fields. At each
point x, the differential can be identified with a linear form df (x) = h∇X f (x), · iTx X acting
on tangent vectors F (x) ∈ Tx X, which model a small displacement around x. The change

24
4. SPECTRAL LEARNING METHODS 25

of the function value as the result of this displacement is given by applying the form to the
tangent vector, df (x)F (x) = h∇X f (x), F (x)iTx X , and can be thought of as an extension of
the notion of the classical directional derivative. The operator ∇X f : L2 (X) → L2 (T X) is
called the intrinsic gradient, and is similar to the classical notion of the gradient defining the
direction of the steepest change of the function at a point. Similarly, the intrinsic divergence
is an operator divX F : L2 (T X) → L2 (X) acting on tangent vector fields and (formal)
adjoint to the gradient operator

hF, ∇X f iL2 (T X) = h−divX F, f iL2 (X) . (4.3)

Finally, the Laplace-Beltrami operator (LBO) ∆X : L2 (X) → L2 (X) is an operator

∆X f = −divX (∇X f ) (4.4)

acting on scalar fields. From (4.3) it follows that the LBO is self-adjoint,

h∇X f, ∇X f iL2 (T X) = h∆X f, f iL2 (X) = hf, ∆X f iL2 (X) . (4.5)

The LBO is intrinsic, i.e., expressible entirely in terms of the Riemannian metric. As a
result, it is invariant to isometric (metric-preserving) deformations of the manifold.

Spectral analysis on manifolds The LBO of a compact manifold admits an eigendecom-

position ∆X φk = λk φk with a countable set of real eigenvalues 0 = λ1 ≤ λ2 ≤ . . . and
the corresponding eigenfunctions φ1 , φ2 , . . . form an orthonormal basis on L2 (X). This
basis is a generalization of the Fourier basis to non-Euclidean domains:1 given a function
f ∈ L2 (X), it can be represented as the Fourier series
X
f (x) = hf, φk iL2 (X) φk (x), (4.6)
k≥1

where the analysis fˆk = hf, φk iL2 (X) can be regarded as the forward Fourier transform
and the synthesis k≥1 fˆk φk (x) is the inverse one; the eigenvalues {λk }k≥1 play the role
P

of frequencies.
Note that in the Euclidean case, the eigenfunctions of the 1D Laplacian takes the form
d2 iωx
− dx2 e = ω 2 eiωx , and the classical Fourier transform as the inner product between the
signal f (x) and the Laplacian eigenfunctions e−iωx , i.e.
Z ∞
fˆ(ω) = hf (x), e−iωx iL2 (R) = f (x)e−iωx dx. (4.7)
−∞

Heat diffusion on manifolds is governed by the diffusion equation,

∂
 
∆X + f (x, t) = 0; (4.8)
∂t
f (x, 0) = f0 (x), (4.9)

where f (x, t) denotes the amount of heat at point x at time t, f0 (x) is the initial heat
distribution; if the manifold has a boundary, appropriate boundary conditions must be
1
It is easy to verify that the classical Fourier basis functions eiωx are eigenfunctions of the Euclidean
d2 iωx
Laplacian operator − dx 2e = ω 2 eiωx .
4. SPECTRAL LEARNING METHODS 26

added. The solution of (4.8) is obtained by applying the heat operator H t = e−t∆X to the
initial condition,
Z
t
f (x, t) = H f0 (x) = f0 (x0 )ht (x, x0 )dx0 , (4.10)
X

Since H t has the same eigenfunctions as ∆X with the eigenvalues {e−tλk }k≥1 , we can
express the solution of (4.8) in the Fourier domain as
Z
f0 (x0 ) e−tλk φk (x)φk (x0 ) dx0 ,
X
f (x, t) = (4.11)
X k≥1
| {z }
ht (x,x0 )

where ht (x, x0 ) is the heat kernel. Interpreting the LBO eigenvalues as ‘frequencies’, the
coefficients e−tλ play the role of a transfer function corresponding to a low-pass filter
sampled at {λk }k≥1 .

Discretization In the discrete setting, the surface X is sampled at n points x1 , . . . , xn .

On these points, we construct a triangular mesh (V, E, F ) with vertices V = {1, . . . , n},
in which each interior edge ij ∈ E is shared by exactly two triangular faces ikj and
jhi ∈ F , and boundary edges belong to exactly one triangular face. The set of vertices
{j ∈ V : ij ∈ E} directly connected to i is called the 1-ring of i. A real-valued function
f : X → R on the surface is sampled on the vertices of the mesh and can be identified with
an n-dimensional vector f = (f (x1 ), . . . , f (xn ))> . The discrete version of the LBO is given
as an n × n matrix L = A−1 W, where

(cot αij + cot βij )/2

 ij ∈ E;
wij = − k6=i wik (4.12)
P
i = j;


0 else;

αij , βij denote the angles ∠ikj, ∠jhi of the triangles sharing the edge ij, and A =
diag(a1 , . . . , an ) with ai = 31 jk:ijk∈F Aijk being the local area element at vertex i and
P

Aijk denoting the area of triangle ijk [68].

The first k ≤ n eigenfunctions and eigenvalues of the LBO are computed by performing
the generalized eigendecomposition WΦ = AΦΛ, where Φ = (φ1 , . . . , φk ) is an n × k
matrix containing as columns the discretized eigenfunctions and Λ = diag(λ1 , . . . , λk ) is
the diagonal matrix of the corresponding eigenvalues.

4.2 Spectral descriptors

Heat Kernel Signature (HKS) Sun et al. [87] proposed a construction of intrinsic dense
descriptors by considering the diagonal of the heat kernel,

e−tλk φ2k (x),

X
ht (x, x) = (4.13)
k≥0

also known as the autodiffusivity function. The physical interpretation of autodiffusivity

is the amount of heat remaining at point x after time t. Geometrically, autodiffusivity
is related to the Gaussian curvature K(x) by virtue of the Taylor expansion ht (x, x) =
4. SPECTRAL LEARNING METHODS 27

1
4πt+ K(x)
12π + O(t). Sun et al. [87] defined the heat kernel signature (HKS) of dimension Q at
point x by sampling the autodiffusivity function at some fixed times t1 , . . . , tQ ,

f (x) = (ht1 (x, x), . . . , htQ (x, x))> . (4.14)

The HKS has become a very popular approach in numerous applications due to several
appealing properties. First, it is intrinsic and hence invariant to isometric deformations of
the manifold by construction. Second, it is dense. Third, the spectral expression (4.13) of
the heat kernel allows efficient computation of the HKS by using the first few eigenvectors
and eigenvalues of the Laplace-Beltrami operator.
At the same time, a notable drawback of HKS stemming from the use of low-pass
filters is poor spatial localization (by the uncertainty principle, good localization in the
Fourier domain results in a bad localization in the spatial domain).

Wave Kernel Signature (WKS) Aubry et al. [6] considered a different physical model
of a quantum particle on the manifold, whose behavior is governed by the Schrödinger
equation,
∂
 
i∆X + ψ(x, t) = 0, (4.15)
∂t
where ψ(x, t) is the complex wave function capturing the particle behavior. Assuming
that the particle oscillates at frequency λ drawn from a probability distribution π(λ), the
solution of (4.15) can be expressed in the Fourier domain as
X
ψ(x, t) = eiλk t π(λk )φk (x). (4.16)
k≥1

The probability of finding the particle at point x is given by

Z T X
p(x) = lim |ψ(x, t)|2 dt = π 2 (λk )φ2k (x), (4.17)
T →∞ 0
k≥1

and depends on the initial frequency distribution

 π(λ). Aubry et al. [6] considered a
log ν−log λ
log-normal frequency distribution πν (λ) = exp 2σ 2
with mean frequency ν and
standard deviation σ. They defined the Q-dimensional wave kernel signature (WKS)

f (x) = (pν1 (x), . . . , pνQ (x))> , (4.18)

where pν (x) is the probability (4.17) corresponding to the initial log-normal frequency
distribution with mean frequency ν, and ν1 , . . . , νQ are some logarithmically-sampled
frequencies.
While resembling the HKS in its construction and computation, WKS is based on
log-normal transfer functions that act as band-pass filters and thus exhibits better spatial
localization.

4.3 Optimal spectral descriptors

Litman and Bronstein [56] considered generic descriptors of the form

X K
X
f (x) = τ (λk )φ2k (x) ≈ τ (λk )φ2k (x) (4.19)
k≥1 k=1
4. SPECTRAL LEARNING METHODS 28

where τ (λ) = (τ1 (λ), . . . , τQ (λ))> is a bank of transfer functions acting on LBO eigenval-
ues, and used parametric transfer functions
M
X
τq (λ) = aqm βm (λ) (4.20)
m=1

in the B-spline basis β1 (λ), . . . , βM (λ), where aqm (q = 1, . . . , Q, m = 1, . . . , M ) are the

parametrization coefficients. Plugging (4.20) into (4.19) one can express the qth component
of the spectral descriptor as
X M
X X
fq (x) = τq (λk )φ2k (x) = aqm βm (λk )φ2k (x), (4.21)
k≥1 m=1 k≥1
| {z }
gm (x)

where g(x) = (g1 (x), . . . , gM (x))> is a vector-valued function referred to as geometry vector,
dependent only on the intrinsic geometry of the shape. Thus, (4.19) is parametrized by the
Q × M matrix A = (alm ) and can be written in matrix form as f (x) = Ag(x). The main
idea of [56] is to learn the optimal parameters A by minimizing a task-specific loss which
reduces to a Mahalanobis-type metric learning.

4.4 Deformable shape correspondence with random forests

In this Section we see how the classification forest paradigm (Section 2.1) can be employed
to predict a dense correspondence between two given shapes M and N . The key idea, as
introduced in [71], is to learn from examples a canonical transformation, i.e., a transformation
from the points of an input shape (say M ) to a canonical label set L. We will first show
how to predict such a transformation from each individual shape M , N to the label set. A
correspondence between M and N will be then obtained by a careful composition of the
two mappings.

4.4.1 Inference
We start by describing the inference (or testing)
step. In the context of shape matching, a decision
tree routes a point x ∈ M along the tree and
to a leaf node, where a probability distribution
defined on a discrete label set L is assigned to
the point. Each label ` ∈ L identifies a tuple
of corresponding points from the collection of
training shapes; in the inset figure, points having
the same color are associated to the same label
in L. This way, the collection of probability distributions over L, which are predicted
for each point x ∈ M , can be interpreted as defining a soft map from shape M to some
reference shape from the training set. Note that such a reference is not a specific shape
from the collection, but rather an abstraction that allows us to think of the label space as a
physical object. We will discuss this aspect with more detail in Section 4.4.3.
According to this inference procedure, each tree t ∈ F of a forest F provides a
posterior probability P (`|x, t) of label ` ∈ L, given a point x ∈ M . The prediction of
the whole forest F can be obtained by averaging the predictions of the single trees as
P (`|x, F) = |F1 | t∈F P (`|x, t).
P
4. SPECTRAL LEARNING METHODS 29

4.4.2 Learning
During the learning phase, the structure of the trees, the split functions and the leaf
posteriors are determined from a training set. The latter consists of a collection of shapes
with point-wise ground truth matches among them; for the sake of simplicity, we assume
that all shapes have the same number of points, such that for each shape Ri in the training
set we have a bijective mapping (or canonical transformation) Ti : Ri → L. The training
set of labelled data is then given by {(x, Ti (x)) | x ∈ Ri }i . It remains to define the label
predictions associated to each leaf, and the test functions associated to the interior nodes
of the forest.
A straightforward way to assign a label distribution to each leaf node is to measure,
for each label ` ∈ L, the proportion of training samples (x, `) among all training samples S
that have reached the leaf:
|{(x, `) ∈ S}|
P (`|S) = . (4.22)
|S|
The probability distribution P (·|S) will thus become the posterior probability during
inference for every shape point reaching the leaf.
As splitting functions for the interior nodes, Rodolà et al. [71] proposed to consider
binary tests of the form fΘ (x) > τ , where function fΘ (x) is a local shape descriptor
computed at x ∈ M and parametrized by Θ, and τ is a randomly chosen threshold.
Since the parameters Θ are optimized during training, the idea is to consider an existing
descriptor but let the forest automatically determine its discriminative features based on
the training examples. The baseline descriptor can be chosen depending on the matching
problem at hand. The WKS was considered in [71] for classical shape matching, while the
HKS was used in [19, 49] due to its better resilience to missing shape parts.

4.4.3 Shape matching via the label space

Once the random forest is fixed, the inference step for a given test shape M provides us
with a mapping T : M → P (L). However, as briefly mentioned in Section 4.4.1, such a
mapping carries no geometric meaning by itself. The transformation can be endowed
with a meaningful interpretation if we consider some reference shape Ri from the training
set, for which we know the bijection Ti : Ri → L, and instead consider the soft map
(Ti−1 ◦ T ) : M → P (R)i . We can now interpret the forest prediction as a probabilistic
correspondence between the input test shape and a reference shape from the training set.
This construction is useful, for example, to visualize the quality of the labeling and was
utilized in [71] to devise a regularization step for the forest prediction.
In a typical matching problem, we are given two TM
shapes M and N among which we seek to estab- L TN
lish a dense correspondence. One is then interested
in P (y|x), the probability that a point x ∈ M cor-
responds to a point y ∈ N . Using the law of total
−1
probability, we can write TN ◦ TM

X
P (y|x) = P (`|x) · P (y|`) . (4.23)
`∈L
M N
Note that this calculation is very simple to carry
out in practice. Let XM , XN denote two matrices
containing the label predictions for the two shapes, i.e., for each point xi ∈ M and each
4. SPECTRAL LEARNING METHODS 30

label ` ∈ L the probability P (`|xi ) is given by (XM )`i and similarly for N . Since XM and
XN are left-stochastic matrices, (4.23) can be written as

(XN )j` (XM )`i = X̃>

X
P (y|x) = N XM , (4.24)
`∈L

where X̃>N denotes column normalization after transpose. In other words, it stores the
probabilities of a point yj ∈ N being the pre-image of a label ` ∈ L:

P (`|y)
(X̃>
N )j` = P (y|`) = P .
y∈N P (`|y)

The resulting matrix X̃>

N XM can now be interpreted as a soft map between the two input
shapes, and it can be processed in different ways depending on the specific application.
5

Intrinsic Convolutional Neural

Networks

We have seen in Chapter 3 that deep learning methods, in particular, convolutional neural
architectures, could be applied to geometric data that is treated as a Euclidean 3D object
(volume or range image). An alternative way of treating geometric data intrinsically as
manifolds, was discussed in Chapter 4. The main goal of this chapter is to generalize
convolutional neural networks to this latter setting. In particular, we will focus on the
intrinsic definition of the convolution operation.

5.1 Intrinsic CNNs in the spectral domain

A fundamental result of classical Euclidean signal processing, states that the Fourier
transform diagonalizes the convolution operator: the convolution f ? g of two functions
in the spectral domain can be expressed as the element-wise product of their Fourier
transforms (4.7),
\
(f ? g)(ω) = fˆ(ω) · ĝ(ω). (5.1)
Property (5.1) can be used to define convolution on non-Euclidean domains in the spectral
domain as X
(f ? g)(x) = hf, φk iL2 (X) hg, φk iL2 (X) φk (x). (5.2)
k≥0

Such an operation can be interpreted as a non-linear filtering of the signal f . The key
difference from the classical convolution is the lack of shift-invariance, which makes the
filter kernel to change depending on its position.

5.1.1 Spectral CNNs

In [15], the spectral definition of convolution (5.2) was used to generalized CNNs to
graphs, by representing filters in the spectral domain. In our context, the fundamental
drawbacks of this formulation is its limitation to a single given domain, since the spectral
representation of the filters is basis dependent. It implies that if we learn a filter w.r.t. a basis
on one domain, and then try to apply it on another domain with another basis, the result
could be very different (see Figure 5.1).

31
5. INTRINSIC CONVOLUTIONAL NEURAL NETWORKS 32

Figure 5.1: An illustration of the poor generalization of spectral filtering across non-
Euclidean domains. Left: a function defined on a manifold; middle: result of the applica-
tion of a filter in the frequency domain on the same manifold; right: the same filter applied
on the same function but on a different (nearly-isometric) domain produces a completely
different result.

5.1.2 Localized spectral CNNs

One of the key drawbacks of the Fourier transformation is the fact that it is global (in
other words, the basis functions have a global support). A localized or windowed Fourier
transform (WFT) is used in signal processing for local space-frequency analysis of signal.
The main idea of the WFT is to analyze the frequency content of a signal that is localized
by means of multiplication by a window. Given a function f ∈ L2 (R) and some ‘mother
window’ g localized at zero, the classical WFT is defined as
Z ∞
0
(Sf )(x, ω) = f (x0 )g(x − x0 )e−ix ω dx0 .
−∞

Alternatively, it can be defined as an inner product with a translated and modulated

window,
(Sf )(x, ω) = hf, Mω Tx giL2 (R) ,
where Tx and Mω denotes the translation and modulation operator respectively. The
translated and modulated window Mω Tx g is referred to as the WFT atom. In the Euclidean
setting the translation operator is defined simply as (Tx0 f )(x) = f (x − x0 ), while the mod-
ulation operator is a multiplication by a Laplacian eigenfunction (Mω f )(x) = eiωx f (x),
which amounts to a translation in the frequency domain (M \ ˆ 0
ω 0 f ) = f (ω − ω ).
The notion of translation to a point x0 , x − x0 , is not well defined in the non-Euclidean
setting. In [79, 10] these operations were defined in the frequency domain. Translation to
x0 is replaced by convolution (5.2) with a delta-function centered at x0 , yielding
X
(Tx0 f )(x) = (f ? δx0 )(x) = hf, φk iL2 (X) hδx0 , φk iL2 (X) φk (x)
k≥1

hf, φk iL2 (X) φk (x0 )φk (x).

X
=
k≥1

Note that such a translation is not shift-invariant in general, i.e., the window would change
when moved around the manifold (see Figure 5.2). The modulation operator is defined as
5. INTRINSIC CONVOLUTIONAL NEURAL NETWORKS 33

ĝ1
0.20
0.15
0.10
0.05
0
0 0.01 0.02 0.03
eigenvalues

ĝ2
0.20
0.15
0.10
0.05
0
0 0.01 0.02 0.03
eigenvalues

Figure 5.2: Examples of different WFT atoms gx,k using different windows (top and bottom
rows; window Fourier coefficients are shown on the left), shown in different localizations
(second and third columns) and modulations (fourth and fifth columns).

(Mk f )(x) = φk (x)f (x), where φk is the kth eigenfunction of the Laplace-Beltrami operator.
Combining the two operators together, the WFT atom (see examples in Figure 5.2) becomes

ĝi φi (x)φi (x0 ).

X
gx0 ,k (x) = (Mk Tx0 g)(x) = φk (x)
i≥1

Note that the ‘mother window’ is defined here in the frequency domain by the coefficients
ĝi . Finally, the WFT of a signal f ∈ L2 (X) can be defined as

(Sf )(x0 , k) = hf, gx0 ,k iL2 (X) = gˆi φi (x0 )hf, φi φk iL2 (X) .
X
(5.3)
i≥1

The WFT (Sf )(x, k) performs a filtering of the signal f at the point x at the frequency
k. By collecting its behavior over different frequencies, the content of the signal f in a local
support around x is extracted, reproducing in this way the window extraction on images.
The localized spectral convolution layer can thus be defined as
p X
X K
gl (x) = wl,k,l0 |(Sfl0 )(x, k)|,
l0 =1 k=1

where fl0 , l0 = 1, . . . , p is the input signal, W = (wl,k,l0 ) is a p × K × q tensor representing

the learnable weights, and gl , l = 1, . . . , q is the output signal. An additional degree of
freedom is the possibility to learn the window itself as well.
A drawback of this approach is its memory and computation requirements, as each
window ĝi in Equation (5.3) needs to be explicitly produced.

5.2 Intrinsic CNNs in the spatial domain

An alternative definition of an intrinsic equivalent of convolution is in the spatial domain.
A classical convolution can be thought of as a template matching with filter, operating
as a sliding window: e.g. in an image, one extracts a patch of pixels, correlates it with a
5. INTRINSIC CONVOLUTIONAL NEURAL NETWORKS 34

Figure 5.3: Construction of local geodesic polar coordinates on a manifold. Left: exam-
ples of local geodesic patches, center and right: example of angular and radial weights,
respectively (red denotes larger weights).

template, and moves the window to the next position. In the non-Euclidean setting, the
lack of shift-invariance makes the patch extraction operation position-dependent. The
patch operator Dj (x) acting on the point x ∈ X can be defined as a re-weighting of the
input signal f by means of some weighting kernels {wi (x, ·)}i=1,...,J spatially localized
around x, i.e. Z
Dj (x)f = f (x0 )wj (x, x0 )dx0 , j = 1, . . . , J. (5.4)
X
The intrinsic convolution can be defined as
X
(f ? g)(x) = gj Dj (x)f, (5.5)
j

where gj denotes the filter coefficients applied on the patch extracted at each point.
Different spatial-domain intrinsic convolutional layers amounts for a different definition
of the patch operator D. In the following we will see two examples.

5.2.1 Geodesic CNNs

In [59], the authors propose to define the patch operator as a combination of gaussian
weights defined on a local intrinsic polar system of coordinates. Given a point x on the
shape X, the local polar system of coordinates specifies the coordinates of the surrounding
points in terms of radial and angular components (ρ(x), θ(x)). The radial coordinate is
constructed as ρ-level sets {x0 : dX (x, x0 ) = ρ} of the geodesic distance function dX for
ρ ∈ [0, ρ0 ], where ρ0 is the radius of the geodesic disc. The angular coordinate θ(x) is
constructed as a set of equispaced geodesics Γ(x, θ) emanating from x in direction θ in a
way that they are perpendicular to the geodesic distance level sets.
Once the local geodesic system of coordinates is extracted, the geodesic patch operator
is defined as
Z
(D(x)f )(θ, ρ) = f (x0 )wθ (x, x0 )wρ (x, x0 )dx0 ; (5.6)
X
−d2X (Γ(x,θ),x0 )/2σθ2
wθ (x, x0 ) = e ; (5.7)
−(dX (x,x0 )−ρ)2 /2σρ2
wρ (x, x0 ) = e , (5.8)
5. INTRINSIC CONVOLUTIONAL NEURAL NETWORKS 35

Figure 5.4: Visualization of different heat kernels (red represent high values). Leftmost:
example of an isotropic heat kernel. Remaining: examples of anisotropic heat kernels for
different rotation angles θ and anisotropy coefficient α.

where wθ , wρ are the angular and radial weights, respectively (see Figure 5.3 center and
right). Note that the choice of the origin of the angular coordinate is arbitrary, and therefore
it can vary from point to point. To overcome this problem, an angular max pooling was
used in [59], leading to the following definition of the geodesic convolution
Z
(f ? w)(x) = max w(θ + ∆θ, ρ)(D(x)f )(θ, ρ) dθdρ, (5.9)
∆θ∈[0,2π)

5.2.2 Anisotropic diffusion CNNs

In section 4.1, we saw how the heat propagation on a shape X is governed by the heat
diffusion equation (4.8). In particular, given as initial heat distribution a delta function
centered on x, the heat distribution on X after some time t is represented by the heat
kernel ht (x, ·). Such a heat kernel is isotropic, i.e., diffuses heat equally in all directions
(Figure 5.4, left). The diffusion strength is controlled by the diffusion time t.
A more general anisotropic heat diffusion is described by the anisotropic diffusion
equation,
ft (x, t) = −div(A(x)∇f (x, t)), (5.10)
where the thermal conductivity matrix A(x) specifies the heat conductivity properties at
each point on the shape X. In particular, A(x) contains both positional and directional
information. This more general diffusion model was considered in [4] for shape analysis
tasks. In [11], the authors defined the thermal conductivity matrix as
!
α
Aαθ (x) = Rθ (x) Rθ (x)> ,
1

where the matrix Rθ (x) performs rotation of θ w.r.t. to some reference (e.g. the maximum
curvature) direction and α > 0 is a parameter controlling the degree of anisotropy (α = 1
corresponds to the classical isotropic case).
5. INTRINSIC CONVOLUTIONAL NEURAL NETWORKS 36

filter bank 1
p filters

Nθ rotations
... Σ

max
Σ ξ

...

...
Σ
...
f1 f1 g1

...

...

...

...
filter bank q
... Σ

max
Σ ξ

...

...
Σ
...
fm fp gq
Input LIN ReLU GC AMP Output
m-dim q-dim

Figure 5.5: A simple example of an intrinsic convolutional neural network architecture.

The network takes as input an off-the-shelf m-dimensional hand-crafted local shape
descriptor for each vertex, then a linear dimensionality reduction layer is applied to
reduce the input dimension to p < m, followed by a ReLU non-linearity. Finally, a
geodesic convolution layer (2.6) with q banks of p filters is followed by an angular max
pooling layer to remove the ambiguity of the choice of angular coordinate origin, which
produces a q-dimensional output feature for each vertex.

Analogously to Equation (4.11), the anisotropic heat kernel is given by

hαθt (x, x0 ) = e−tλαθ,k φαθ,k (x)φαθ,k (x0 ),

X

k≥0

where φαθ,k (x), λαθ,k are the eigenfunctions and eigenvalues of the anisotropic Laplacian
∆αθ = −div(Aαθ (x)∇). The anisotropic heat kernel hαθt depends on two additional
parameters, the coefficient α and the rotation angle θ. Figure 5.4 shows some examples of
anisotropic heat kernels computed at different rotations θ and anisotopies α. In [11], such
kernels were used as the weighting functions for the construction of patch operator (5.4),
Z
(D(x)f )(θ, t) = hαθt (x, x0 )f (x0 )dx0 ,
X

mapping the values of f around point x to a local polar-like system of coordinates (θ, t).

5.3 Applications
Similarly to the Euclidean CNNs, an intrinsic convolutional neural network consists of
several layers that are applied subsequently, i.e. the output of the previous layer is used
as the input into the subsequent one. The convolutional layer (2.6) is used with the only
difference that the convolution operation is replaced by an intrinsic analogy. Figure 5.5
shows a toy example of an intrinsic CNN architecture with one intrinsic convolutional
layer.
Intrinsic CNN is a non-linear hierarchical parametric map of the form

ΨΘ = ψθnn ◦ · · · ◦ ψθ22 ◦ ψθ11 ,

where ψi , i = 1, . . . , n, represents the ith layer with parameters θi . The parameters of the
model Θ = {θi : i = 1, . . . , n} are the set of all the parameters of each layer. The function
5. INTRINSIC CONVOLUTIONAL NEURAL NETWORKS 37

is applied to point-wise input data (e.g. some simple geometric descriptors) and produces
some point-wise output. In the following, we will see how intrinsic CNNs can be applied
to two basic problems in computer graphics: the computation of local descriptors and
correspondences.

5.3.1 Local descriptors

The result of applying an intrinsic CNN in a point-wise manner on some input feature
vector f (x) is a feature map g(x) that can be regarded to as a dense local descriptor at
point x. Ideally, a local descriptor should be as similar as possible at corresponding points
(positives) across a collection of shapes, and as dissimilar as possible at non-corresponding
points (negatives). Learning optimal descriptors is possible using the siamese architecture
[14], composed of two identical copies of the same intrinsic CNN model sharing the same
parameterization and fed by pairs of knowingly similar or dissimilar samples, where the
training procedure tries to minimize the siamese loss

|T+ | |T− |
(µ − kΨΘ (fi ) − ΨΘ (fi− )k)2+ .
X X
L(Θ) = (1 − γ) kΨΘ (fi ) − ΨΘ (fi+ )k2 + γ
i=1 i=1

Here, λ ∈ [0, 1] is a parameter trading off between the positive and negative losses, µ is
a margin, (·)+ = max{0, ·} and T± = {(fi , fi± )} denotes the sets of positive and negative
pairs, respectively.
In [59], the authors used the geodesic CNN architecture shown in Figure 5.5 to produce
dense intrinsic pose- and subject-invariant descriptors on human shapes. A qualitative
evaluation of the goodness of the learned descriptors is reported in Figure 5.6, where
the Euclidean distance in the descriptor space between the descriptor at a selected point
and the rest of the points on the same shape as well as its transformations is depicted.
The descriptors produced by the geodesic CNN (GCNN) manifest both good localization
(better than HKS) and are more discriminative (less spurious minima than WKS and OSD),
as well as robustness to different kinds of noise, including isometric and non-isometric
deformations, geometric and topological noise, different sampling, and missing parts.

5.3.2 Correspondence
As we discussed in Section 4.4.3 of these notes, finding the correspondence in a collection
of shapes can be posed as a labelling problem, where one tries to label each vertex of a
given query shape X with the index of a corresponding point on some common reference
shape Y [71]. Let n and m denote the number of vertices in X and Y , respectively. For
a point x on a query shape, the output of an intrinsic CNN ΨΘ (x) is m-dimensional and
is interpreted as a probability distribution (‘soft correspondence’) on Y . The output of
the network at all the points of the query shape can be arranged as an n × m matrix with
elements of the form ψΘ (x, y), representing the probability of x mapped to y.
Let us denote by y ∗ (x) the ground-truth correspondence of x on the reference shape.
We assume to be provided with examples of points from shapes across the collection
and their ground-truth correspondence, T = {(x, y ∗ (x))}. The optimal parameters of the
network are found by minimizing the multinomial regression loss

log ψΘ (x, y ∗ (x)),

X
Lreg (Θ) = −
(x,y ∗ (x)∈T )
5. INTRINSIC CONVOLUTIONAL NEURAL NETWORKS 38

Heat kernel signature (HKS)

Wave kernel signature (WKS)

Optimal spectral descriptor (OSD)

GCNN
Figure 5.6: Normalized Euclidean distance between the descriptor at a reference point
on the shoulder (white sphere) and the descriptors computed at the rest of the points for
different transformations (shown left-to-right: near isometric deformations, non-isometric
deformations, topological noise, geometric noise, uniform/non-uniform subsampling,
missing parts). Cold and hot colors represent small and large distances, respectively;
distances are saturated at the median value. Ideal descriptors would produce a distance
map with a sharp minimum at the corresponding point and no spurious local minima at
other locations.

Figure 5.7: Examples of correspondence on the FAUST humans dataset obtained by the
anisotropic diffusion CNN. Shown is the texture transferred from the leftmost reference
shape to different subjects in different poses by means of our correspondence. The
correspondence is nearly perfect (only very few minor artifacts are noticeable).

which represents the Kullback-Leibler divergence between the probability distribution

produced by the network and the ground-truth distribution δy∗ (x). Figures 5.7, 5.8 and 5.9
show a qualitative evaluation of the quality of the correspondences learned by anisotropic
diffusion CNN.
5. INTRINSIC CONVOLUTIONAL NEURAL NETWORKS 39

0.1

0
Blended Intrinsic Map

geodesic CNN

anisotropic diffusion CNN

Figure 5.8: Pointwise geodesic error (in % of geodesic diameter) of different correspon-
dence methods (top to bottom: Blended intrinsic maps, geodesic CNN, anisotropic diffu-
sion CNN) on the FAUST dataset. For visualization clarity, the error values are saturated
at 10% of the geodesic diameter. Hot colors correspond to large errors. Note the different
behavior of different approaches: BIM produces large distortions with very few accurate
matches; geodesic CNN produces many near-perfect matches but also many matches
with large distortion; anisotropic diffusion CNN produces very few matches with large
distortion and many near-perfect matches.
5. INTRINSIC CONVOLUTIONAL NEURAL NETWORKS 40

0.1

anisotropic diffusion CNN

Random Forest
Figure 5.9: Examples of partial correspondence on the dog shape from the SHREC’16
Partial (holes) dataset. First row: correspondence produced by anisotropic diffusion
CNN. Corresponding points are shown in similar color. Reference shape is shown on the
left. Second and third rows: pointwise geodesic error (in % of geodesic diameter) of the
anisotropic diffusion CNN and RF correspondence, respectively. For visualization clarity,
the error values are saturated at 10% of the geodesic diameter. Hot colors correspond to
large errors.
Bibliography

[1] 3D Animation Online Services, 3D Characters, and Character Rigging - Mixamo.

https://www.mixamo.com/. Accessed: 2015-10-03.

[2] Yobi3D - free 3D model search engine. https://www.yobi3d.com. Accessed:

2015-11-03.

[3] I. Aizenberg, N. Aizenberg, and J. Vandewalle. Multi-valued and universal binary

neurons. Springer, 2000.

[4] M. Andreux, E. Rodolà, M. Aubry, and D. Cremers. Anisotropic Laplace-Beltrami

operators for shape analysis. In Proc. NORDIA, 2014.

[5] D. Anguelov, P. Srinivasan, D. Koller, S. Thrun, J. Rodgers, and J. Davis. SCAPE:

shape completion and animation of people. Trans. Graphics, pages 408–416, 2005.

[6] M. Aubry, U. Schlickewei, and D. Cremers. The wave kernel signature: A quantum
mechanical approach to shape analysis. In Proc. 4DMOD, 2011.

[7] P. Baldi and P. Sadowski. The dropout learning algorithm. Artificial Intelligence,
210:78–122, 2014.

[8] F. Bogo, M. J. Black, M. Loper, and J. Romero. Detailed full-body reconstructions of

moving people from monocular RGB-D sequences. In Proc. ICCV, 2015.

[9] F. Bogo, J. Romero, M. Loper, and M. J. Black. FAUST: Dataset and evaluation for
3D mesh registration. In Proc. CVPR, 2014.

[10] D. Boscaini, J. Masci, S. Melzi, M. M. Bronstein, U. Castellani, and P. Vandergheynst.

Learning class-specific descriptors for deformable shapes using localized spectral
convolutional networks. Computer Graphics Forum, 34(5):13–23, 2015.

[11] D. Boscaini, J. Masci, E. Rodolà, and M. Bronstein. Learning shape correspondence

with anisotropic convolutional neural networks. In Proc. NIPS, 2016.

[12] D. Boscaini, J. Masci, E. Rodolà, M. Bronstein, and D. Cremers. Anisotropic diffusion

descriptors. Computer Graphics Forum, 35(2):431–441, 2016.

[13] L. Breiman. Random forests. In Machine Learning, volume 45, pages 5–32, 2001.

[14] J. Bromley et al. Signature verification using a “Siamese” time delay neural network.
In Proc. NIPS. 1994.

[15] J. Bruna, W. Zaremba, A. Szlam, and Y. LeCun. Spectral networks and locally
connected networks on graphs. In Proc. ICLR, 2014.

41
BIBLIOGRAPHY 42

[16] Q. Chen and V. Koltun. Robust nonrigid registration by convex optimization. In

IEEE ICCV, 2015.

[17] A. Choromanska, M. Henaff, M. Mathieu, G. B. Arous, and Y. LeCun. The loss

surfaces of multilayer networks. In Proc. AISTATS, 2015.

[18] D. Clevert, T. Unterthiner, and S. Hochreiter. Fast and accurate deep network
learning by exponential linear units (elus). In Proc. ICLR, 2015.

[19] L. Cosmo, E. Rodolà, M. M. Bronstein, et al. SHREC’16: Partial matching of de-

formable shapes. In Proc. 3DOR, 2016.

[20] L. Cosmo, E. Rodolà, J. Masci, A. Torsello, and M. Bronstein. Matching deformable

objects in clutter. In Proc. 3DV, 2016.

[21] A. Criminisi, J. Shotton, and E. Konukoglu. Decision forests: A unified frame-

work for classification, regression, density estimation, manifold learning and semi-
supervised learning. Foundations and Trends in Computer Graphics and Vision, 7(2-
3):81–227, 2012.

[22] R. Dechter. Learning while searching in constraint-satisfaction-problems. In Proc.

AAAI, 1986.

[23] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A large-scale
hierarchical image database. In Proc. CVPR, 2009.

[24] A. Dosovitskiy, P. Fischer, E. Ilg, P. Hausser, C. Hazirbas, V. Golkov, P. van der

Smagt, D. Cremers, and T. Brox. FlowNet: Learning optical flow with convolutional
networks. In Proc. ICCV, 2015.

[25] A. Dosovitskiy, J. T. Springenberg, and T. Brox. Learning to generate chairs with

convolutional neural networks. In Proc. CVPR, 2015.

[26] M. Dou, J. Taylor, H. Fuchs, A. Fitzgibbon, and S. Izadi. 3d scanning deformable

objects with a single rgbd sensor. In IEEE CVPR, 2015.

[27] D. Erhan, Y. Bengio, A. Courville, P.-A. Manzagol, P. Vincent, and S. Bengio. Why
does unsupervised pre-training help deep learning? JMLR, 11, 2010.

[28] K. Fukushima. Neocognitron: A self-organizing neural network for a mechanism of

pattern recognition unaffected by shift in position. Biological Cybernetics, 36(4):193–
202, 1980.

[29] J. Gall, A. Yao, N. Razavi, L. Van Gool, and V. Lempitsky. Hough forests for object
detection, tracking, and action recognition. Trans. PAMI, 33(11):2188–2202, 2011.

[30] X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward
neural networks. In Proc. Conf. Artificial Intelligence and Statistics, 2010.

[31] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair,

A. Courville, and Y. Bengio. Generative adversarial nets. In Proc. NIPS. 2014.

[32] K. Gregor, I. Danihelka, A. Graves, D. J. Rezende, and D. Wierstra. DRAW: A

recurrent neural network for image generation. In Proc. ICML, 2015.
BIBLIOGRAPHY 43

[33] X. Han, T. Leung, Y. Jia, R. Sukthankar, and A. C. Berg. MatchNet: Unifying feature
and metric learning for patch-based matching. In Proc. CVPR, 2015.
[34] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition.
arXiv:1512.03385, 2015.
[35] K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing
human-level performance on ImageNet classification. In Proc. ICCV, 2015.
[36] G. E. Hinton. Training products of experts by minimizing contrastive divergence.
Neural Computation, 14(8):1771–1800, 2002.
[37] G. E. Hinton, S. Osindero, and Y.-W. Teh. A fast learning algorithm for deep belief
nets. Neural Computation, 18(7):1527–1554, July 2006.
[38] G. E. Hinton and R. R. Salakhutdinov. Reducing the dimensionality of data with
neural networks. Science, 313(5786):504–507, 2006.
[39] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov.
Improving neural networks by preventing co-adaptation of feature detectors.
arXiv:1207.0580, 2012.
[40] S. Hochreiter and J. Schmidhuber. Simplifying neural nets by discovering flat
minima. In Proc. NIPS, 1995.
[41] S. Hochreiter and J. Schmidhuber. Long Short-Term Memory. Neural Computation,
9(8):1735–1780, 1997.
[42] Q. Huang, B. Adams, M. Wicke, and L. J. Guibas. Non-rigid registration under
isometric deformations. In Proc. SGP, 2008.
[43] S. I. and C. S. Batch normalization: Accelerating deep network training by reducing
internal covariate shift. arXiv:1502.03167, 2015.
[44] M. Jaderberg, K. Simonyan, A. Zisserman, and K. Kavukcuoglu. Spatial transformer
networks. In Proc. NIPS, 2015.
[45] K. Kawaguchi. Deep learning without poor local minima. In Proc. NIPS, 2016.
[46] V. G. Kim, Y. Lipman, and T. Funkhouser. Blended Intrinsic Maps. Trans. Graphics,
30(4), 2011.
[47] D. P. Kingma, T. Salimans, and M. Welling. Variational dropout and the local
reparameterization trick. In Proc. NIPS. 2015.
[48] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep
convolutional neural networks. In Proc. NIPS. 2012.
[49] Z. Lähner, E. Rodolà, M. M. Bronstein, et al. SHREC’16: Matching of deformable
shapes with topological noise. In Proc. 3DOR, 2016.
[50] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to
document recognition. Proc. IEEE, 86(11):2278–2324, 1998.
[51] H. Lee, R. Grosse, R. Ranganath, and A. Y. Ng. Convolutional deep belief networks
for scalable unsupervised learning of hierarchical representations. In Proc. ICML,
2009.
BIBLIOGRAPHY 44

[52] H. Li, B. Adams, L. J. Guibas, and M. Pauly. Robust single-view geometry and
motion reconstruction. Trans. Graphics, 28(5), 2009.

[53] H. Li, E. Vouga, A. Gudym, L. Luo, J. T. Barron, and G. Gusev. 3D self-portraits.

Trans. Graphics, 32(6), 2013.

[54] Y. Lipman and T. Funkhouser. Möbius voting for surface correspondence. Trans.
Graphics, 28(3), 2009.

[55] O. Litany, E. Rodolà, A. M. Bronstein, M. M. Bronstein, and D. Cremers. Non-rigid

puzzles. Computer Graphics Forum, 35(5), 2016.

[56] R. Litman and A. M. Bronstein. Learning spectral descriptors for deformable shape
correspondence. Trans. PAMI, 36(1):170–180, 2014.

[57] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic
segmentation. Proc. CVPR, 2015.

[58] S. Mallat. A Wavelet Tour of Signal Processing. Academic Press, 2008.

[59] J. Masci, D. Boscaini, M. M. Bronstein, and P. Vandergheynst. Geodesic convolutional

neural networks on riemannian manifolds. In Proc. 3dRR, 2015.

[60] J. Masci, U. Meier, D. Cireşan, and J. Schmidhuber. Stacked convolutional auto-

encoders for hierarchical feature extraction. In Proc. ICANN, 2011.

[61] D. Maturana and S. Scherer. VoxNet: A 3D convolutional neural network for

real-time object recognition. In Conf. Intelligent Robots and Systems, 2015.

[62] V. Mnih, N. Heess, A. Graves, and K. Kavukcuoglu. Recurrent models of visual

attention. In Proc. NIPS. 2014.

[63] R. A. Newcombe, D. Fox, and S. M. Seitz. DynamicFusion: Reconstruction and

tracking of non-rigid scenes in real-time. In Proc. CVPR, 2015.

[64] H. Noh, S. Hong, and B. Han. Learning deconvolution network for semantic
segmentation. arXiv:1505.04366, 2015.

[65] K. Olszewski, J. J. Lim, S. Saito, and H. Li. High-fidelity facial and speech animation
for VR HMDs. Trans. Graphics, 35(6), 2016.

[66] M. Ovsjanikov, M. Ben-Chen, J. Solomon, A. Butscher, and L. J. Guibas. Functional

maps: a flexible representation of maps between shapes. Trans. Graphics, 31(4), 2012.

[67] V. Pham, C. Kermorvant, and J. Louradour. Dropout improves recurrent neural

networks for handwriting recognition. arXiv:1312.4569, 2013.

[68] U. Pinkall and K. Polthier. Computing discrete minimal surfaces and their conju-
gates. Experimental Mathematics, 2(1):15–36, 1993.

[69] A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with

deep convolutional generative adversarial networks. In Proc. ICLR. 2016.

[70] E. Rodolà, L. Cosmo, M. M. Bronstein, A. Torsello, and D. Cremers. Partial functional

correspondence. Computer Graphics Forum, 2016.
BIBLIOGRAPHY 45

[71] E. Rodolà, S. Rota Bulò, T. Windheuser, M. Vestner, and D. Cremers. Dense non-rigid
shape correspondence using random forests. In Proc. CVPR, 2014.

[72] D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Learning representations by

back-propagating errors. Nature, 323(6088):533–536, 1986.

[73] S. Rusinkiewicz, B. Brown, and M. Kazhdan. 3D scan matching and registration. In

Proc. ICCV Courses, 2005.

[74] R. M. Rustamov, M. Ovsjanikov, O. Azencot, M. Ben-Chen, F. Chazal, and L. J.

Guibas. Map-based exploration of intrinsic shape differences and variability. Trans.
Graphics, 32(4):72, 2013.

[75] S. Saito, T. Li, and H. Li. Real-time facial segmentation and performance capture
from RGB input. In Proc. ECCV, 2016.

[76] F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face
recognition and clustering. arXiv:1503.03832, 2015.

[77] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun. OverFeat:

Integrated recognition, localization and detection using convolutional networks.
arXiv:1312.6229, 2013.

[78] J. Shotton, R. Girshick, A. Fitzgibbon, T. Sharp, M. Cook, M. Finocchio, R. Moore,

P. Kohli, A. Criminisi, A. Kipman, and A. Blake. Efficient human pose estimation
from single depth images. Trans. PAMI, 35(12):2821–2840, 2012.

[79] D. I. Shuman, B. Ricaud, and P. Vandergheynst. Vertex-frequency analysis on graphs.

arXiv:1307.5708, 2013.

[80] N. Silberman, D. Hoiem, P. Kohli, and R. Fergus. Indoor segmentation and support
inference from RGBD images. In Proc. ECCV, 2012.

[81] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale
image recognition. arXiv:1409.1556, 2014.

[82] S. Song and J. Xiao. Deep sliding shapes for amodal 3D object detection in RGB-D
images. arXiv:1511.02300, 2015.

[83] R. K. Srivastava, K. Greff, and J. Schmidhuber. Training very deep networks. In

Proc. NIPS, 2015.

[84] R. K. Srivastava, J. Masci, F. J. Gomez, and J. Schmidhuber. Understanding locally

competitive networks. In Proc. ICLR, 2015.

[85] M. Stollenga, J. Masci, F. Gomez, and J. Schmidhuber. Deep networks with internal
selective attention through feedback connections. In Proc. NIPS, 2014.

[86] H. Su, S. Maji, E. Kalogerakis, and E. G. Learned-Miller. Multi-view convolutional

neural networks for 3D shape recognition. In Proc. ICCV, 2015.

[87] J. Sun, M. Ovsjanikov, and L. J. Guibas. A concise and provably informative multi-
scale signature based on heat diffusion. In Proc. SGP, 2009.

[88] J. Süssmuth, M. Winter, and G. Greiner. Reconstructing animated meshes from

time-varying point clouds. Computer Graphics Forum, 27(5):1469–1476, 2008.
BIBLIOGRAPHY 46

[89] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke,

and A. Rabinovich. Going deeper with convolutions. Technical report, 2014.

[90] J. Taylor, J. Shotton, T. Sharp, and A. Fitzgibbon. The vitruvian manifold: Inferring
dense correspondences for one-shot human pose estimation. In Proc. CVPR, 2012.

[91] A. Tevs, A. Berner, M. Wand, I. Ihrke, M. Bokeloh, J. Kerber, and H.-P. Seidel. Ani-
mation cartography – intrinsic reconstruction of shape and motion. Trans. Graphics,
31(2):12:1–12:15, 2012.

[92] J. Tompson, M. Stein, Y. LeCun, and K. Perlin. Real-time continuous pose recovery
of human hands using convolutional networks. Trans. Graphics, 33(5), 2014.

[93] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.-A. Manzagol. Stacked denoising
autoencoders: Learning useful representations in a deep network with a local
denoising criterion. JMLR, 11:3371–3408, 2010.

[94] D. Vlasic, I. Baran, W. Matusik, and J. Popović. Articulated mesh animation from
multi-view silhouettes. Trans. Graphics, 27(3), 2008.

[95] M. Wand, B. Adams, M. Ovsjanikov, A. Berner, M. Bokeloh, P. Jenke, L. Guibas, H.-P.

Seidel, and A. Schilling. Efficient reconstruction of nonrigid shape and motion from
real-time 3D scanner data. Trans. Graphics, 28(2), 2009.

[96] S. Wang and C. Manning. Fast dropout training. In Proc. ICML, 2013.

[97] L. Wei, Q. Huang, D. Ceylan, E. Vouga, and H. Li. Dense human body correspon-
dences using convolutional networks. In Proc. CVPR, 2016.

[98] X. Wei, P. Zhang, and J. Chai. Accurate realtime full-body motion capture using a
single depth camera. Trans. Graphics, 31(6), 2012.

[99] P. J. Werbos. Beyond Regression: New Tools for Prediction and Analysis in the Behavioral
Sciences. PhD thesis, Harvard University, 1974.

[100] Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao. 3D Shapenets: A
deep representation for volumetric shapes. In Proc. CVPR, 2015.

[101] K. Xu, J. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhudinov, R. Zemel, and Y. Bengio.
Show, attend and tell: Neural image caption generation with visual attention. In
Proc. ICML, 2015.

[102] F. Yu and V. Koltun. Multi-scale context aggregation by dilated convolutions. In

Proc. ICLR. 2016.

[103] M. E. Yumer and N. J. Mitra. Learning semantic deformation flows with 3d convo-
lutional networks. In Proc. ECCV, 2016.

[104] S. Zagoruyko and N. Komodakis. Learning to compare image patches via convolu-
tional neural networks. In Proc. CVPR, 2015.

[105] S. Zagoruyko, A. Lerer, T.-Y. Lin, P. O. Pinheiro, S. Gross, S. Chintala, and P. Dollár.
A multipath network for object detection. In Proc. BMVC, 2016.

[106] A. Zeng, S. Song, M. Nießner, M. Fisher, and J. Xiao. 3DMatch: Learning the
matching of local 3D geometry in range scans. arXiv:1603.08182, 2016.
BIBLIOGRAPHY 47

[107] S. Zhou, J. Wu, Y. Wu, and X. Zhou. Exploiting local structures with the kronecker
layer in convolutional networks. arXiv:1512.09194, 2015.