Internship 2009-2010
Robust face descriptors in uncontrolled settings
Keywords:
Computer vision, machine learning, face description
and recognition, missing data, metric learning.
Supervisors : Matthieu Guillaumin, Jakob Verbeek and Cordelia Schmid.
General information:
Minimum internship duration: 4 months. Salary depending on intern's status.
The internship will take place in the LEAR team at INRIA, Grenoble,
France.
2009/12/14 Update: This internship offer is
closed. The position is filled.
Context:
LEAR's main focus is to use machine learning approaches to tackle
computer vision related tasks. Among them, the detection and analysis of
human faces is still a major challenge despite the large research
community and literature. Significant advances are still necessary
to make face recognition robust in uncontrolled
environments. Compared to settings that deal only with varying lighting
conditions, and small expression or pose changes, uncontrolled environments
are subject to major pose and expression changes, as well as occlusions
(hats, hands, hair, other persons, ...), and changes in age, glasses,
hair and facial hair styles. To achieve recognition in this setting, we
need to design and develop robust face descriptors and associate
them with powerful machine learning and statistical modeling techniques.
Current processing pipeline for face
recognition in the LEAR team
 |
Goal and approach:
Following recent successes [2,3,4,5,6], we will adopt an approach based on
face description, which implies representing faces as vectors in a
high-dimensional space. The first goal of the internship will be to
perform an experimental comparative study of existing descriptors, most
of which will be provided [1,2,5,7] but some will have to be re-implemented
[3,4,6]. This study will be conducted using standard face data sets and
benchmarks [10,11]. Following the lessons drawn from this study, the
intern will build an efficient processing pipeline for extracting a
robust face descriptors from real-world images, as shown in the
illustration.
Another approach to consider, following [9], consists in building a
recognition system by combining several binary classifiers that give
specific information about
visual attributes of a face image: is
this a man or woman? is it a child or adult? does this person wear
glasses? does (s)he wear a hat? is (s)he blonde? ... The collected
(real-valued) answers to those questions can be used as inputs for other
machine learning techniques.
The second goal of the internship is to deal with occlusions.
Occlusions that affect face images translate into noise in the face
descriptors discussed above, and this noise is usually not random.
Therefore, these occlusions can be modeled, rejected and considered as
missing data in the descriptor at recognition step. We propose to adapt
metric learning frameworks, which have shown great performance in a
number of computer vision related tasks [2,13], with an
Expectation-Maximization derivation that can handle missing data in a
principled way [12].
Requirement:
Excellent academic records, applied mathematics and
scientific programming skills are essential. Previous experience in
image processing is an asset.
Application:
Applicants should send a CV and contact information of
a referee by email to
matthieu<dot>guillaumin<at>inria<dot>fr
References:
[1] A. Kläser,
Human detection and character recognition in
TV-style movies, In Informatiktage, 2007.
[2] M. Guillaumin, J.Verbeek, and
C.Schmid,
Is that you? Metric Learning Approaches for
Face Identification, Proceedings of the IEEE International Conference on Computer Vision 2009.
[3] J. Sivic, M. Everingham, and
A.Zisserman,
'Who are you?' - Learning person specific classifiers
from video, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2009.
[4] N. Pinto, J. DiCarlo, and D. Cox,
How far can
you get with a modern face recognition test set using only simple
features?, Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition 2009.
[5] Y. Taigman, L. Wolf, and T. Hassner,
Multiple One-Shots for Utilizing Class Label Information,
Proceedings of the British Machine Vision Conference
2009.
[6] N. Kumar, A. C. Berg, P. N. Belhumeur, and S. K. Nayar,
Attribute
and Simile Classifiers for Face Verification, Proceedings of the IEEE International Conference on Computer Vision 2009.
[7] M. Guillaumin, T. Mensink, J. Verbeek, and C. Schmid,
Automatic
Face Naming with Caption-based Supervision, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2008.
[8]
A comprehensive face recognition
reference site [9]
Columbia University face verification
project
[10]
Labeled Faces in the Wild dataset, protocol and
results
[11]
Face Recognition Grand
Challenge
[12] S. Roweis,
EM Algorithms for PCA and SPCA, Neural Information Processing Systems 1997.
[13] P. Jain, B. Kulis and K. Grauman,
Fast image search for learned
metrics, Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, 2008.
