Aleix Face
Recognition Page
In the past (and perhaps in the present too), researchers have been concerned on “how to recognize faces when large and –hopefully-- representative datasets are available to the system.” Two paradigms have been largely applied and tested for face recognition: the feature-based and the appearance-based (or texture-based) approaches.
Within the appearance-based approach to face recognition, several pattern recognition and machine learning techniques have been employed. Although arguable, one could say that the better a classifier separates the training data or discriminates between classes, the better our system will perform for the test data. Unfortunately, this is only true when the learning data successfully samples the underlying (and unknown) distribution. In many applications tough, the available training data does not properly sample the underlying distribution, because faces can undergo complex changes of shape and texture; e.g., (among others) lighting, orientation, facial expression and occlusions.
We have shown [A.M. Martinez and A.C. Kak, “PCA versus LDA,” Trans. PAMI-23(2): 228-233, 2001] that when small, non-representative datasets are used for training, theoretically “inferior” (in the sense of class separability) techniques can perform better than “superior” ones (more discriminant). (See below for details.)
How can we recognize faces when the training dataset is small then?
Some progress has been recently achieved to recognize faces that appear at different illumination conditions and/or orientations (from learning to testing). However, little (or no) work has been reported on how to compensate for partially occluded faces or how to compensate for the texture change due to the different facial expressions of a person’s face. Below we summarize our ongoing work on these two problems.
1.
PCA versus LDA
In the context of the appearance-based
paradigm for object recognition, it is generally believed that algorithms based
on LDA (Linear Discriminant Analysis) are superior to those based on PCA
(Principal Components Analysis). In
this research we show that this is not always the case. We present our case first by using
intuitively plausible arguments (as shown in Fig.1 below) and then by showing
actual results on a face database. Our
overall conclusion is that when the training dataset is small, PCA can
outperform LDA, and also that PCA is less sensitive to different training datasets.
Fig.1 There are two different classes embedded in two different
“Gaussian-like” distributions. However, only two sample per class are supplied
to the learning procedure (PCA or LDA). The classification result of the PCA
procedure (using only the first eigenvector) is more desirable than the result
of the LDA. $D_{PCA}$ and $D_{LDA}$ represent the decision thresholds obtained
by using nearest-neighbor classification.
Publications:
·
A.M Martínez
and A.C. Kak,
“PCA versus LDA,”
IEEE Transactions on
Pattern Analysis and Machine Intelligence, Vol. 23, No. 2, pp. 228-233,
February 2001.
2.
Recognizing imprecisely localized and partially occluded
faces
How can one build systems that recognize
this face 
given that only this image 
was used for
“learning” this person’s class?
New face recognition approaches are
needed, because although much progress has been recently achieved in the field
(e.g. within the eigenspace domain), still many problems are to be robustly
solved. Two of these problems are occlusions and the imprecise localization of
faces (which ultimately imply a failure in identification). While little has
been done to account for the first problem, almost nothing has been proposed to
account for the second. This paper presents a probabilistic approach that
attempts to solve both problems while using an eigenspace representation. To
resolve the localization problem, we need to find the subspace (within the
feature space, e.g. eigenspace) that represents this error for each of the
training image. To resolve the occlusion problem, each face is divided into n
local regions which are analyzed in isolation. In contrast with other previous
approaches, where a simple voting space is used, we present a probabilistic
method that analyzes how ``good" a local match is. Our method has proven
to be superior to a local voting PCA on a set of 2600 face images.
Publications:
·
A.M
Martínez,
“Recognizing Imprecisely Localized, Partially
Occluded and Expression Variant Faces from a Single Sample per Class,”
IEEE Transactions on
Pattern Analysis and Machine Intelligence, in press.
·
A.M.
Martínez
“Recognition of Partially Occluded and/or Imprecisely
Localized Faces Using a Probabilistic Approach,”
Proc. of IEEE Computer Vision and Pattern Recognition
(CVPR), Vol. I, pp. 712-717, 2000.
3.
Expression variant faces
How can we recognize the identity of the following
image 
if only this face 
was use to learn this class?
Expression variant faces: This problem can be formulated as follows:
“how can we robustly identify a person's face for whom the learning and
testing face images differ in facial expression?”. We say that a system is
expression-dependent if the image of a face displaying an emotion a
(say, happiness) is more difficult (or less difficult) to recognize than the
image of a face displaying an emotion b (say, sadness), given that the
image of a face that displays another emotion (say, neutral – no-expression)
was used for learning. An expression-invariant method would correspond to one
that equally identifies the identity of the subject independent of the facial
expression displayed on the test image. In an attempt to overcome this problem,
the PCA approach uses the second order statistics of the image set in the hope
that these features will be invariant to different facial expressions.
Unfortunately, it can be theoretically shown that for any given invariant
paradigm (the PCA approach in this case) there is always a set of (testing)
images for which the learned measure (or function) will not be optimal.
Our approach: In general, different emotions are
expressed as facial expressions with more emphasis on specific parts of the
face than others. Those parts that are expected to be less affected by the
current expressed emotion are expected to obtain better identification matches.
This knowledge can be added to the classifier as prior information by weighting
each of the above n local parts with values that vary depending on the
facial expression displayed on the testing image. The more the texture of a
local area of the testing images diverges from the texture displayed on the
same local area of the learning sample (due to differences in facial
expression), the lower the assigned weight should be and vice versa.
Publications:
·
A.M
Martínez,
“Recognizing Imprecisely Localized, Partially
Occluded and Expression Variant Faces from a Single Sample per Class,”
IEEE Transactions on
Pattern Analysis and Machine Intelligence, in press.
·
A.M
Martínez,
“Semantic Access of Frontal Face Images: The
expression-invariant problem,”
Proc. of IEEE Workshop on Content-Based Access of Images and
Video Libraries, pp. 55-59, 2000.
4.
The AR face database
The AR face database was
created by Aleix M. Martinez and Robert Benavente at the Computer Vision Center
(CVC). It contains over 4,000 color images corresponding to 126 people's faces
(70 men and 56 women). Images feature frontal view faces with different facial
expressions, illumination conditions, and occlusions (sun glasses and scarf).
The pictures were taken at the CVC under strictly controlled conditions. No
restrictions on wear (clothes, glasses, etc.), make-up, hair style, etc. were
imposed to participants. Each person participated in two sessions, separated by
two weeks (14 days) time. The same pictures were taken in both sessions. This face database is publicly available and
free to University researchers. Commercial distribution or any act
related to commercial use of this database is strictly prohibited.
The AR
face database is publicly available here.
Publications:
·
A.M Martínez
and R. Banavente,
“The AR face database,”
CVC Tech. Report #24, June 1998.