Workshop on 3D and 2D Face Analysis
and Recognition
Jan. 28, 2011
At Université de Lyon
Ecole Centrale de Lyon, Amphithéâtre 202
36 avenue Guy de Collongue, 69134 Ecully Cedex
Proposed by Pr. L.Chen (ECL Liris), Pr. Mohamed Daoudi (LIFL/Telecom Lille 1), Pr. J.L.Dugelay (Eurecom)
Access plan : www.ec-lyon.fr
Program
1)
9h15-9h30. Welcome
2)
9h30-10h05. “Facial Surfaces Analyzing by Using Riemannian
geometry”, by Prof. Mohamed Daoudi, LIFL, Telecom Lille 1
a.
In
this talk we explore the use of Riemannian geometry to analyze shapes of facial
surfaces. That is, we define a differentiable manifold, with a suitable
Riemannian metric, whose elements represent individual facial surfaces. In
particular, we obtain algorithms for computing geodesics, computing statistical
means, and stochastic clustering. We demonstrate these ideas in two application
contexts: biometric applications and facial expression recognition.
b.
Mohamed Daoudi is a Full Professor of Computer
Science in the Institut TELECOM ; TELECOM Lille 1, LIFL (UMR 8022). He received
the Ph.D. degree in Computer Engineering from the University of Lille 1 (USTL),
France, in 1993 and Habilitation à Diriger des Recherches from the
University of Littoral, France, in 2000. He was the founder of the MIIRE
research group of LIFL (UMR 8022). His research interests include pattern
recognition, image processing, three-dimensional analysis and retrieval and
more recently 3D face recognition. He has published more than 80 papers in
refereed journals and proceedings of international conferences. He is the
coauthor of the book "3D processing : Compression, Indexing and
Watermarking (Wiley, 2008)". He has served as a Program Commitee member
for the International Conference on Pattern Recognition (ICPR) in 2004 and the
International Conference on Multimedia and Expo (ICME) in 2004 and 2005. He was
a co-organizer and co-chair of ACM Workkshop on 3D Retrieval 2010 and
Eurographics 3D Object retrieval 2009. He has organized a special session on 3D
Face Analysis and Recognition at ICME 2008.He was an associate editor of the
Journal of Multimedia (2006-2009). He is a frequent reviewer for IEEE
Transactions on Pattern Analysis and Machine Intelligence and for IJCV, JMIV.
His research has been funded by ANR, RNRT and European Commission grants. He is
a Senior Member of IEEE.
3)
10h05-10h40. “3D face segmentation based
on high curvature edge detection for harmonic map alignment”,
by Przemyslaw Szeptycki, Liris, ECL
a.
In
this talk, we propose an automatic open mouth detection algorithm based on an analysis
of one of the principal curvatures. While facial surface deformation during
expression is assumed to be near-isometric, mouth opening significantly changes
the surface topology and introduces anisometry, thus making Harmonic Mapping inconsistent.
We show how removing the open mouth part and using an modified geodesic distance
for expression invariant face segmentation provides more reliable
data for further stages of the 3D Face processing pipeline, that require
consistently segmented faces, including the computation of Harmonic Maps. The
algorithm was evaluated on two 3D face datasets, thus testing performance both
on noisy models (FRGCv2) and for large facial expressions (Bosphorpus). In
order to evaluate the need for open mouth detection, the statistics of mouth
opening in terms of position and size were evaluated for a number of common
labeled facial actions.
b.
Przemyslaw
is a PhD student under the supervision of Dr.Mohsen Ardabilian and Prof.
L.Chen. His current interests include 3D face preprocessing, 3D face
landmarking and 3D face segmentation
4)
10h40-10h50.
Coffee break
5)
10h50-11h25. “Occlusion in 3D face recognition”,
by Dr Alessandro Colombo, University of Milano-Bicocca
a.
The
greater part of state of the art work in face recognition does not consider the
occlusion issue; only a few approaches which consider 2D images can be found in
the literature. In this seminar it will be presented an innovative three
dimensional face detection and face restoration strategy for the recognition of
three dimensional faces which may be partially occluded by unforeseen,
extraneous objects. No a-priori knowledge about the occluding objects is
required. These may be glasses, hats, scarves and the like, and differ greatly
in shape or size, introducing a high level of variability in appearance. The
restoration strategy is independent of the method used to detect occlusions and
can also be applied to restore faces in the presence of noise and missing
pixels due to acquisition inaccuracies. Results obtained on artificial and real
datasets will be presented.
b.
Alessandro
Colombo took his degree in Computer Science in 2004 and
his Ph.D. in Computer Science in 2008 at DISCo, Department of Information
Science, Systems Theory, and Communication at the University of Milano-Bicocca.
He is currently a post-doc researcher at the Imaging and Vision Laboratory (www.ivl.disco.unimib.it). His research interests cover the
processing, analysis and synthesis of 2D and 3D images. In particular, he
focused his research on 2D/3D detection and recognition of objects and faces.
6)
11h25-12h00. “3D face analysis”, by Prof.Wang Yunhong, Beihang University, China
a.
As one of the most important organ of human beings,
face conveys a significant amount of information, including the identity,
emotion, gender, age, ethnicity etc., which plays an important role in
face-to-face communications. Human can acquire this information easily, but
it’s difficult for computers to possess the same ability. Enabling computers to
understand various information that human face conveys is the target always
pursued by the computer vision and pattern recognition communities. Due to its
non-intrusive and user friendly advantages, automatic interpretation of face
information has promising applications in public security, access control,
human-computer interaction, image retrieval, computer animation, etc. Although
studies have shown that facial attributes such as the aging, gender and
ethnicity are revealed not only by the 2D textures but also by the 3D
morphology of human faces, most of the proposed works on face analysis are 2D
image based, which are sensitive to illumination and pose variations. In order
to deal with this problem, and with the rapid development of 3D capturing
equipments, 3D face analysis is attracting more and more attentions. This talk
will present our ongoing research work on 3D facial feature point localization
and registration, feature extraction and matching, multimodal gender and
ethnicity classification, and face aging.
b.
Yunhong Wang is Professor of Computer Science at
Beihang University. She received her PhD in 1998 and worked in the National Lab
of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences
from1998 to 2004. She transferred to Beihang University at 2004. She has been
working on pattern recognition and image processing. She served as program
committee members of many important academic conferences such as International
Conference on Pattern Recognition (ICPR), International Conference on
Biometrics (ICB) etc. Her current research interests include digital
watermarking, pattern recognition and image processing.
7)
12h00-13h30.
Buffet
8)
13h30-14h05. “Asymmetric
3D/2D Face recognition”, By Huang Di, Liris, ECL
a.
3D Face recognition has been considered as a major
solution to deal with these unsolved issues for reliable 2D face recognition in
the recent years, i.e. illumination and pose variations. However, 3D based
technique is currently limited by its high registration and computation cost.
In this talk, asymmetric 3D-2D face recognition is presented, enrolling people
in textured 3D environment but performing identification in 2D automatically.
The goal is to limit the use of 3D data to where it really helps to improve
recognition accuracy. The proposed method consists of two separate matching
processes: Sparse Representation Classifier (SRC) is applied in 2D-2D matching,
while Canonical Correlation Analysis (CCA) is utilized to learn the mapping
between range LBP face (3D) and texture LBP face (2D). Both matching scores are
further combined for final decision. Moreover, new designed preprocessing step
enhances its robustness to illumination and pose effects. The proposed approach
achieves much better experimental results in the FRGC v2.0 database than 2D
algorithms do, while avoiding the high cost of data acquisition and computation
in 3D approaches.
b.
Huang Di is a PhD student at Liris ECL under the
supervision of Dr.Mohsen Ardabilian, Prof. Yunhong Wang and Prof. Liming Chen.
His current research interests include 2D/3D facial representations, 2D/3D
asymmetric face recognition, lighting normalization
9)
14h05-14h40. “Biometric Application for 3D face model”, by
Dr Stéphane Gentric, Morpho, Safran Group, France
a.
The limitation of considering a face as a 2D object
appears in most of biometric applications. This talk will present how 3D face
morphable models are used in two different products, what are the problems that
are currently solved and what are the remaining issues. The first application
is one of the tools proposed for face images enhancement. It allows an operator
to simply drive the fitting of 3D face model on multiple bad quality images, in
order to build a frontal synthetic face image, by merging textures. We will
show how much it improves the biometric performances. The second application is
Face-On-The-Fly, an automatic gate for border crossing system. From four video
streams, it builds a frontal face of a subject, during the crossing the gate,
without any stop and any cooperation. Fitting of a 3D face model, image
synthesis, feature extraction, matching against passport image and door opening
decision are done in real time. For most of the face recognition issues (pose,
expression and illumination), the use of 3D models open new perspectives and
more complex challenges.
b.
Stéphane Gentric is Research Unit Manager at Morpho (www.morpho.com). He receives his PhD in 1999, on Pattern Recognition
at UPMC. From 1999 to 2002, he worked mainly on fingerprint algorithms. From
2002, he focused on Face Recognition, then Iris Recognition. He is now, team
leader for both biometries, driving all algorithmic aspects, from Acquisition
Device to Large Scale Matching System. He was involved in most of Morpho’s
projects in biometrics of the past 10 years, such as Smartgate Australian
border crossing System as well as NIST benchmarks, or the UIDAI project. His
current research interests stay pattern recognition for improvement of
biometric systems.
10)
14h40-15h15 .“A unified Neural Scheme for
facial Image Understanding”, by Prof. Christophe Garcia, Liris, INSA-Lyon
a.
Over the last decade, facial image processing has
become a very active research field due to the large number of possible
applications, such as model-based video coding, image retrieval, surveillance
and biometrics, visual speech understanding, virtual characters for e-learning,
entertainment and intelligent human-computer interaction. With the introduction
of new powerful machine learning techniques, statistical classification methods
and complex deformable models, recent progress has been made on face detection
and tracking, person identification, facial expression and emotion recognition,
gender classification, face coding and virtual face synthesis. However, much
progress is still to be made to provide more robust systems, in order to cope
with the variability of facial image appearances caused by lighting conditions,
poses and expressions, image noise and partial occlusions, in an unconstrained,
real-world context. Among these machine learning approaches,
Convolutional Neural Networks (ConvNets) are powerful models that tightly
couple local feature detection, global model construction and classification in
a single architecture where all parameters are conjointly learnt. They
alleviate the limitations of the traditional hand-designed feature extraction
and selection steps, by automatically learning optimal filters and classifiers
that are very robust to noise. In this presentation, we will show that ConvNets
are very effective for facial image processing by presenting different
architectures and learning schemes, designed for face detection, facial feature
detection, face alignment, gender classification and face recognition.
b.
Christophe received his PhD degree in computer vision
from the University of Lyon I, France, in 1994 and his "Habilitation
à Diriger des Recherches" (HDR) from Insa-Lyon / University of Lyon
I, in 2009. From 1992 to 1997, he has been involved in various computer vision
and robotics research projects at the IBM Vision Automation Group, France, the
Computer Vision Center of the Autonomous University of Barcelona, and the
German National Research Center (now Fraunhofer Institute). From 1997 to 2000,
he has been a researcher at the Foundation for Research and Technology Hellas
(FORTH), Greece, where he was involved in several advanced EU projects in the
field of video and image analysis. From 2000 to 2002, he was a visiting
Professor at the Computer Science Department of the University of Crete,
Greece, where he was teaching Artificial Neural Networks and Pattern
Recognition. In 2003, he spent 10 months at IRISA-INRIA, Rennes, France,
working in the field of automatic video structuring and indexing. From
2004 to 2010, Christophe has been working in France Telecom R&D / Orange
Labs, as a Fellow Expert Researcher in Pattern Recognition, Neural Networks and
Image Indexing and manager of the Multimedia Content Analysis Technologies
group. Since November 2010, He is a Full Professor at INSA-LYON, working
in the “Laboratoire d'InfoRmatique en Image et Systèmes d'information”
(LIRIS). His current technical and research activities are in the areas of
neural networks, pattern recognition, image and video indexing, and computer
vision. He holds 22 industrial patents and has published more than 110 articles
in international conferences and journals. He is currently associate editor of
the Int. Journal of Visual Communication and Image Representation (Elsevier),
Image and Video Processing (Hindawi), Pattern Analysis and Application
(Springer-Verlag) and Pattern Recognition (Elsevier).
11)
15h15-15h30. Coffee break
12)
15h30-16h05. “Face Recognition with Patterns of
Oriented Edge Magnitudes”, by Son VU & Prof. Alice Caplier
a.
This talk addresses the question of
computationally inexpensive yet discriminative and robust feature sets for
real-world face recognition. The proposed descriptor named Patterns of Oriented
Edge Magnitudes (POEM) has desirable properties: POEM (1) is an oriented, spatial
multi-resolution descriptor capturing rich information about the original
image; (2) is a multi-scale self-similarity based structure that results in
robustness to exterior variations; and (3) is of low complexity and is
therefore practical for real-time applications. Briefly speaking, for every
pixel, the POEM feature is built by applying a self-similarity based structure
on oriented magnitudes, calculated by accumulating a local histogram of
gradient orientations over all pixels of image cells, centered on the
considered pixel. The robustness and discriminative power of the POEM
descriptor is evaluated for face recognition on both constrained (FERET) and
unconstrained (LFW) datasets. Experimental results show that our algorithm
achieves better performance than the state-of-the-art representations. More
impressively, the computational cost of extracting the POEM descriptor is so
low that it runs around 20 times faster than just the first step of the methods
based upon Gabor filters. Moreover, its data storage requirements are 13 and 27
times smaller than those of the LGBP (Local Gabor Binary Patterns) and HGPP (Histogram
of Gabor Phase Patterns) descriptors respectively
b.
Son VU is a PhD student under the
supervision of Prof. Alice Caplier
13)
16h05-16h40. “Face
recognition in the wild: verification and caption-based recognition”, by Dr
Jakob Verbeek, INRIA Rhône-Alpes
a.
In this talk I will present our recent work on face verification and
recognition in uncontrolled settings. We work with images taken from Yahoo News
and the associated captions, from which we automatically extract faces and
names respectively. The data was manually annotated, yielding over 30.000 faces
of over 5.000 individuals. We compute high dimensional redundant face
descriptors, anchored at facial "parts" such as the mouth, nose, and
eyes that are automatically located using a constellation model. Metric
learning is used to obtain a representation that is compact and robust to
intra-person expression and pose changes, while being sensitive to inter-person
appearance variations. Experimentally we evaluate this approach for face
verification, and its impact on caption based recognition. In the former task,
we are interested in determining whether two face images represent the same
person or not. In the latter task the goal is to automatically associate faces
in an image to names found in the caption.
b.
Jakob Verbeek is a researcher in the LEAR (learning and recognition in
vision) team at INRIA Rhone-Alpes. His research focuses on machine learning
approaches to solve computer vision problems.
Recent work includes methods for automatic image annotation, face recognition,
and semi-supervised image categorization.
14)
16h40-17h15. “Person
recognition using a bag of soft biometrics (BoSB)”, by Antitza DANTCHEVA
and Prof. Jean-Luc DUGELAY. (antitza.dantcheva@eurecom.fr)
a.
We introduce and examine the novel idea of using a bag of soft
biometrics for person recognition. This novel tool inherits the
non-intrusiveness and computational efficiency of soft biometrics, which allow
for fast and enrollment-free biometric analysis, even in the absence of consent
and cooperation of the surveillance subject.
In this work we provide insight on general design aspects in soft-biometric
systems, and different aspects regarding capabilities, challenges and efficient
resource allocation. Moreover we propose a specific soft biometric system
including traits like hair, eye and skin color, as well as the presence of
beard, moustache and glasses. In conjunction with the system design and
detection algorithms, we also proceed to shed some light on the statistical characterization
of different parameters that are pertinent to the proposed system.
b. Antitza Dantcheva is a Ph.D. student at EURECOM, Sophia
Antipolis under the supervision of Prof. Jean-Luc Dugelay. Her current research
interests are in soft biometrics: algorithms and methods. Recent work include
studies on eye color as a soft biometric trait and on the reliability of soft
biometric systems.
15)
17h15-17h30.
“Conclusion & discussion”, Prof. L.Chen
Contact : Isabelle Dominique, isabelle.dominique@ec-lyon.fr, tél : +33 472 18 64 42
Pr.Liming Chen, liming.chen@ec-lyon.fr, tél : +33 472 18 65 76