|
Highlights
A new transnational research project, Visen, is starting on January 2013 within the framework of Chist Era net.
The Visen project aims at mining automatically the semantic content of visual data, still images or videos, to enable machine reading of images. It gathers four research groups from University of Surrey (Surrey, UK), Institut de Robòtica i Informàtica Industrial (IRI, Spain) , Ecole Centrale de Lyon (ECL, France), and University of Sheffield (Sheffield, UK) having each well established and complementary expertise in their respective areas of research.
Opening of a Postdoc position in Statistical Learning with applications to Computer Vision with a preferred start on December 2012
We are seeking a highly motivated postdoctoral fellow in the field of statistical learning with applications to computer vision! Curiosity, open-mind, creativity, persistence, and collaborative-work ability are the key personal skills we target. The definition of the profile can be found here.
Liris honored with both the golden and silver medals (1st and 2nd best performance) at the ImageCLEF 2012 Photo Annotation Challenge
Photo annotation task is to automatically annotate a large number of consumer images with high-level semantic concepts which can be present in the images. In 2012, 94 high-level concepts have been defined and categorized into five different groups: natural elements (day, night, sunrise, etc..), Environment (desert, coast, landscape, etc..), Person (baby, child, teenager etc..), visual (in focus, city life, active, etc.)., human elements (rail vehicle, water vehicle, air vehicle, etc.). This is an extremely difficult task in the field of computer vision and pattern recognition due to large intra-class variations of objects or concepts (e.g., landscape), strong inter-concepts similarity (e.g., motorcycle vs. bike) and semantic gap between semantic concepts and low-level descriptors that one can extract from images.
Imagine team of Liris has innovated at several levels in their submissions. They have proposed novel visual descriptors (e.g., OC-LBP, dynamism and harmony) to better capture the visual content, novel text descriptors HTC (Histogram of Texual Concepts) that can take into account the semantic similarity between concepts, and a new fusion scheme, namely SWLF (SelectiveWeighted Late Fusion), that characterizes each high-level concept by a specific set of descriptors and weights through learning the decision of experts based on their intrinsic quality. Their multimodal submissions, fusing visual and textual descriptors through SWLF, won both the gold and silver medals in displaying the two best performances evaluated in terms of three different metrics, including in particular the Mean Inferred Average Precision (MiAP). Furthermore, their textual submission also displayed the best performance within the category of submissions only using textual user tags associated, again according to the three different evaluation metrics whereas their visual submission was also ranked the best performance according to the two main evaluation metrics, namely MiAP and GMiAP, within the category of submissions only using the visual data.
The results of all 18 participating teams with their 80 submissions can be found at here. The synthesis, analysis and comparison of all the submissions to this challenge can be found here whereas the working note giving insights to the Liris' submissions can be found here.
The participants of Liris in this competition in 2012 were Ningning Liu, Emmanuel Dellandrea, Liming Chen, Aliaksandr Trus, Chao Zhu, Yu Zhang, Edmond-Charles Bichot, Stéphane Bres, Bruno Tellez. The technical and administrative support were Mrs. Colette Vial, Mrs. Isabelle Dominique and Dr.Aliaksandr Paradzinets (Ghanni). Participation of Liris in the ImageCLEF challenge 2012 and that in 2011 was partly supported by the French Research Agency, l'Agence Nationale de Recherche (ANR) through the ANR Videosense project under the grant 2009 CORD 026 02.
I am co-organizing the workshop on 3D face biometrics in conjunction of the 10th IEEE International Conference on automatic Face and Gesture Recognition (FG 2013), Shanghai, China-April 22-26, 2013.
Advances during the last decades have made high quality acquisition of 3D faces a reality. 3D scanning devices are now available not only in the form of Hi-res scanner which are able to acquire registered texture and range data from static environments but also 4D scanning devices that are able to acquire range data over time so as to capture the dynamics of 3D. In the meantime, much progress has been achieved in 3D face analysis and recognition methods, albeit in very controlled settings. Some topics, however, have yet to be thoroughly explored. On the one hand, the availability of Hi-res data from the environment is expected to boost the accuracy of recognition systems. On the other hand, the availability of a broad range of devices, differing in terms of resolution, poses new challenges and research issues addressing not only the impact of low resolution data on the recognition accuracy but also the integration and combination of different modalities - 2D still, 2D video, 3D still, 3D video -to improve the recognition accuracy. Furthermore, it has been shown that "soft" biometrics (gender, age, ethnicity, etc.) can complement the identity information provided by other biometrics to improve the accuracy of recognition. This workshop focuses on 3D face analysis and recognition and is particularly aimed at exploring ways to extract and effectively exploit still as well as dynamic facial features for recognition. Areas of coverage include, but are not limited to, 3D face detection, analysis, and recognition; gender, ethnicity and age classification from 3D data; analysis of facial expressions; multimodal 2D/3D face recognition; super-resolution facial models.
For further details, please visit the website of the workshop here.
Liris at Ecole Centrale de Lyon took part to the track on 3D face
retrieval and recognition at Shrec 2011 contest. 2 runs submitted by
Liris were ranked the first and the second performance in terms of rank
one recognition rate out of 14 runs submitted by four research groups.
They were ranked the second and the third performance in terms of recall
and precision.
Huibin Li and Liming Chen are the members of Liris at ECL taking part to
this contest.
The paper describing and comparing all the methods submitted to Shrec
2011 can be found here.
The Photo annotation challenge at ImageClef 2011 aims at the automatic
annotation of a large number of consumer photos with multiple semantic
concepts, including visual objects (car, anmal, people, etc.), scenes
(indoor, outdoor, city, etc.), events (voyage, working, etc.), and enven
sentiments (happy, scary, etc.). This year, 18 groups from 11 countries
participated with 79 runs. For their first participation, Liris
achieved a 43.7% MiAP using a multimodal model and was ranked the second
performance behind TUBFI, ajoint submission from TU Berlin and
Fraunhofer First, which achieved a 44,3% MiAP also with a multimodal
model.
The following people took part to this challenge with Liris: Ningning
Liu (ningning.liu@ec-lyon.fr), Yu Zhang (yu.zhang@ec-lyon.fr), Emmanuel
Dellandréa (emmanuel.dellandrea@ec-lyon.fr), Stéphane Brès
(stephane.bres@insa-lyon.fr) et Liming Chen (liming.chen@ec-lyon.fr).
The paper describing and comparing all the methods submitted to the
ImageClef photo annotation task can be found here.
The paper describing our methods submitted to the ImageClef photo annotation task can be found here.
Short Biography
Prof. Liming Chen was awarded a joint BSc degree in Mathematics and
Computer Science from the University of Nantes in 1984. He obtained a
Master degree in 1986 and a PhD in computer science from the University
of Paris 6 in 1989. He first served as associate professor at the
Université de Technologie de Compiègne, then joined Ecole Centrale de
Lyon as Professor in 1998, where he leads an advanced research team on
multimedia computing and pattern recognition. From 2001 to 2003, he also
served as Chief Scientific Officer in a Paris-based company, Avivias,
specialized in media asset management. In 2005, he served as Scientific
multimedia expert in France Telecom R&D China. He has been Head of
the department of Mathematics and Computer science from 2007. Prof.
Liming Chen has taken out 3 patents, authored more than 100 publications
and acted as chairman, PC member and reviewer in a number of high
profile journal and conferences since 1995. He has been a (co)-principal
investigator on a number of research grants from EU FP programme,
French research funding bodies and local government departments. He has
directed more than 15 PhD theses. His current research spans from 2D/3D
face analysis and recognition, image and video analysis and
categorization, to affect analysis both in image, audio and video.
Open Source software
- 3D MODELS PROCESSING TOOL developed by Dr. Przemyslaw Szeptycki within his Phd thesis
The application allows to process 2.5D/3D face models using pre-programmed algorithms and is distributed under GNU GENERAL PUBLIC LICENSE ver. 3 and it has been used in many universities around the world...
Workshop on 3D and 2D Face Analysis and Recognition at ECL, Jan. 28, 2011
Face plays prominent role in human communication and it is potentially
the best biometrics for people identification related applications. Over
the past three decades, face analysis and recognition has attracted
tremendous research effort from various disciplines and has witnessed
impressive progress in basic and applied research, product development
and applications.
This one day workshop focuses on 2D and 3D face analysis and
recognition. The workshop is aimed towards bringing together scientists
and patricians from a wide range of theoretical and application areas
whose work impacts 2D and 3D face analysis and recognition. Its goal is
to provide a state-of-the-art overview of paradigms and challenges on
this challenging topic.
For more information...
- Databases systems
- Design of information systems
- Computer vision
- Pattern recognition
- Wael Bensoltana Co-advised with Dr.Mohsen Ardabilian and Prof.Chokri Ben Amar at ENIS, Tunisia
- Huibin Li Co-advised with Prof. Jean-Marie Morvan at ICJ, UCBL
- Pierre Lemaire Co-advised with Prof. Mohamed Daoudi at LIFL, Telecom Lille 1
- Yu Zhang Co-advised with Dr.Stéphane Brès at Liris, Insa de Lyon
The 3D face analyzer project targets at reliable recognition of facial
attributes on 2.5D or 3D face models, thus making use of face shape,
texture and landmarks at the same time. While developing 3D
analysis-based techniques directly aiming at recognition of facial
attributes, we also want to make forward knowledge on some underlying
fundamental issues, e.g. stability of discrete geometric measures and
descriptions (curvature, distance, etc.) across variations in terms of
model resolution and precision, 3D non-rigid surface registration and
matching in the presence of noisy data. Another important aim of the
project is the collection of significantly representative datasets of 3D
face models in facial expressions, age and gender for the purpose of
training and testing.
The VideoSense project aims at automatic video tagging by high level
concepts, including static concepts (e.g. object, scene, people, etc.),
events, and emotions, while targeting two applications, namely video
recommendation and ads monetization, on the Ghanni’s media assess
management platform. The innovations targeted by the project include
video content description by low-level features, emotional video content
recognition, cross-concept detection and, multimodal fusion, and the
use of a pivot language for dealing with multilingual textual resources
associated with video data.
The Omnia project aims at filtering documents containing text and
images, in a context of data profusion, as they are found on intranets
and on Internet, and to present them to users in a content processing
tool such as DocuShare (Xerox). The originality of the project is to
work on 3 dimensions (image, text, emotion) and in a multilingual
context. Images and texts give rise to 2 categorizations, relative to
the informational aspects and to specific emotional aspects (coming
directly from the images, or relative to their perception as expressed
in the texts). These 2 types of content will be processed independently
(annotation followed by indexation and categorization), with learning
techniques, and will then be merged at the level of the filter and query
tool. Their "primitives" will be linked to an interlingual
representation of word senses based on English (UNL), that will open the
way to multilingualism at the level of "publishing" the document
categories, and of processing queries in natural languages equipped with
UNL dictionaries.
Face recognition from still images is an attractive biometrics for a
broad range of applications. Nevertheless, although numerous and
significant works on that domain, when dealing with still images only,
this modality provides low performances under difficult conditions (eg.
presence of facial expressions) in terms of authentication compared to
fingerprints for example.
In this project, we investigate the possible contribution of considering
an additional dimension in face recognition “3-D” to improve
performances of authentication while keeping existing advantages of face
recognition from still images like no contact, low cooperation from
user needed, well-accepted modality.
Several points will be studied in order to cover as well as possible the
domain: surface matching in 3D, asymmetrical protocol (ie. enrolments
in 3D but authentication in 2D) and combination between 3D (shape) with
2D (texture or appearance).
This industrial research is oriented towards new solutions in face
biometrics, allowing a leapfrog in performances compared to present
solutions, thanks to the use of 3D. These solutions must be realistic in
the sense of application: field usable thanks to the use of "light"
sensors, like, for instance, video surveillance cameras, even if the
enrollement phase is performed with more complex equipment ("asymetric"
approach). They must offer fairly acceptable performance (in the range
or better than fingerprints). Finally, they must be transferable to real
commercial applications in a short time range (4-6 years).
International publications, as well as our experience of the domain,
lead to the conclusion that such an ambitious objective cannot be
reached by a unique biometric "solution", but by the clever association
of several processes: multimodality, of course, as the 3D face sensors
deliver also generally an appearance (texture) image, but also multiple
3D algorithmics ("multimatcher"). Our teams have developed during years
different and complementary methods in 3D face recognition, methods that
we want to combine in order to make the resulting score better that the
one achieved by the best of the methods.
Finally, it is also important in such projects related to biometrics,
that all participants also worry about possible impacts of such
technologies on privacy.
The consortium is composed of USTL, Eurécom, Ecole Centrale de Lyon and
Thales. All partners have already some expertise on 3-D and/or Face
recognition with different but complementary backgrounds and technical
approaches. They know fairly well each other thanks to some past
projects (e.g. Semantic 3D for USTL/LIFL and Eurécom, or Technovision
IV2 for Thales, ECL and Eurécom) or miscellaneous collaborations (eg.
PhD. boards, lectures, etc.)
The project is organized as 5 workpackages. WP0 is classically dedicated
to the coordination and management of the project. WP1 is dedicated to
the study of asymmetrical protocols, ie enrollment in 3-D but
verification from video or images. WP2 focuses on facial deformations
and variability. It includes the following sub items: geometrical
approach, region-based facial surface matching and comparison, and
learning. WP3 is entitled “face recognition by the fusion of shape based
and texture based matching”. This work package includes the following
subworkpackages: fusion strategies combining shape and texture, and
multi matchers. Finally, the fourth work package is dedicated to the
evaluation that includes first the definition of the evaluation
framework and second criteria and the evaluation of algorithm
performances.
A selected publications can be found here.
A full listing of my publications is available on the Liris website.
My Google Scholar Profile can be found
here, and DBLP listing
here, which, funny enough, is a joint listing with my homonym Dr.Liming Chen at University of Ulster, UK in the field of AI, data and knowledge engineering.
|