Face Alignment Robust to Occlusion

Myung-Cheol Roh, Takaharu Oguri, and Takeo Kanade

Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, USA
{myungroh, oguri, takeo.kanade}@cs.cmu.edu
AbstractIn this paper we present an approach to robustly
align facial features to a face image even when the face is
partially occluded. Previous methods are vulnerable to partial
occlusion of the face, since it is assumed, explicitly or implicitly,
that there is no signicant occlusion. In order to cope with this
difculty, our approach relies on two schemes: one is explicit
multi-modal representation of the response from each of the
face feature detectors, and the other is RANSAC hypothesize-
and-test search for the correct alignment over subset samplings
of those in the feature response modes. We evaluated the
proposed method on a large number of facial images, occluded
and non-occluded. The results demonstrated that the alignment
is accurate and stable over a wide range of degrees and
variations of occlusion.
I. INTRODUCTION
In this paper we present a new approach to robustly
align facial features to a face image even when the face
is partially occluded. Previous face alignment methods[1],
[2], [3], [4], [5] nd the optimal t of a regularized face-
shape model (such as a PCA-regularized model) using all of
the observable local features, typically by iterative gradient
search around the initial position obtained by a face detector.
Although certain safety measures are often combined, such
as adjustable weight on different features depending on
features strength, these methods are vulnerable to partial
occlusion of the face, since it is assumed, explicitly or
implicitly, that there is no signicant occlusion and that the
given initial position is close to the correct position.
Fig. 1 shows examples of face alignments on non-occluded
and occluded facial images: Fig. 1(a) shows face detec-
tion results by Haar-like feature-based face detector[6], and
Fig. 1(b) shows alignment results by Bayesian Tangent Shape
Model (BTSM)[2]. For the left image in Fig. 1(b), facial
features in the upper part of the face were aligned, but those
in the lower face, including the mouth and jaw, were not
perfectly aligned because of confusion by a mustache and
beard. In the right image of Fig. 1(b), which contains heavy
occlusion, many of the visible features were not aligned
correctly. It is worthy of noting that alignments of the
features of upper face, though completely visible themselves,
are affected by the lower face occlusion.
In fact, the difculty that partial occlusion causes is that of
cyclic dilemma; occluded or erroneous features should not
participate in alignment, but one cannot tell whether a feature
is occluded unless the correct alignment is known. In order to
cope with this difculty, our approach relies on two schemes:
one is explicit multi-modal representation of the response
from each of the face feature detectors, and the other is
(a) Face detection (b) Face alignment
Fig. 1. Face detections by Haar-like feature based face detector[6], and
face alignment results by BTSM [2].
RANSAC (Random sampling consensus) hypothesize-and-
test search for the correct alignment over subset samplings
of those response modes.
Fig. 2 shows an overview of the proposed method. Given
an image, a large number of detectors, each trained for an
individual facial landmark feature, run and produce their
response map. Each of the red boxes in Fig. 2(a) repre-
sent the region where each feature detector is applied
1
. In
each response map, all the response peaks are identied
and localized by the Mean-Shift method, and each peak is
described by a 2-D Gaussian model (Fig. 2(b)). The correct
alignment we seek is the combination of response peaks,
one selected from each of visible facial features and none
from occluded features, whose locations match well with
the face shape model. Since we dont know which features
are visible beforehand, we employ the RANSAC strategy.
A subset of features is randomly chosen and assumed to be
inlier (visible), a hypothesis of face alignment is generated
from them, and its agreement with other remaining features is
tested in terms of the median of mismatch degrees (Fig. 2(c)).
This is repeated until an alignment with a better acceptable
degree of mismatch is found. The alignment is slightly
adjusted by using only the responses of those features whose
degree of mismatch is less than the median (that is, those that
are identied as visible). Finally, the shape is rened with ad-
ditional inliers which are also estimated as visible (Fig. 2(d)).
There are the two ideas in the RANSAC-based strategy. The
rst is the use of feature subsets for hypothesis generation. It
increases a chance that the generated hypothesis will include
the correct one. The second is the use of median (instead of
mean) of the degree of mismatch for evaluating hypotheses.
It favors a hypothesis in which a majority of features match
very well and some do not, instead of one in which all of
the features match relatively well on average.
1
Although to set search regions is not necessary in our method, for the
sake of efciency, whose size and location is found by a face detector[6].
239
(a) (b) (c) (d)
Fig. 2. An overview of the proposed method. (a) Search regions of three
examples in red boxes. (b) Feature response maps of right eye corner (top), a
nose part (middle), and cheek part (bottom) detectors. Red circles represent
estimated distribution. In the top feature map, any feature points was not
detected. (c) Hypotheses of face alignments by blue lines. (d) Red line shows
a nal alignment after rening. Yellow circles represent feature points which
are identied as visible.
We evaluated the proposed method on a large number
of facial images, occluded and non-occluded. The results
have demonstrated that the method produces accurate and
stable copying with a wide range of degrees and variations
of occlusion and confusing appearances.
II. FACE MODEL
We dene two terminologies as follows: Feature ID is
a name of a point in PDM (Point Distribution Model) and
Feature point is an object of a feature ID where the object
is a response peak which is described by a 2-D Gaussian
Model.
In this paper, shape variation of a face is represented by
the PDM in Constrained Local Model (CLM) framework.
The non-rigid shape function can be described as,
S = X+p (1)
where X = [x
t
1
, , x
t
N
I
], p is a shape parameter vector,
and is the matrix of concatenated eigenvectors. The rst
4 eigenvectors of are forced to correspond to similarity
variations. In order to get p, we employ the CLM as convex
quadratic tting in [1].
A. Learning Local Feature Detectors
In the CLM framework, the local feature detector for
each feature ID should be trained. In this paper, the linear
Support Vector Machines (SVMs) are employed for local
feature detectors because of its computational efciency[7].
To learn local patches, the images in training data are
normalized based on ground-truth. Rotations are normalized
by Generalized Procrustes Analysis (GPA)[8], and the scales
are normalized on the basis of the size of face region
given by a face detector. Positive patches are obtained by
centering around the ground-truth, while negative examples
are obtained by sampling patches shifted away from the
ground-truth. The output of the feature detector is obtained
(a) (b)
Fig. 3. Distributions of feature IDs in the training data: (a) Feature
distributions of all feature IDs. (b) Example of feature distributions: left
eye corner (top), right nostril (middle), and left chin (bottom).
by tting a logistic regression function to the score of the
SVM and the label {not aligned (-1), aligned (1) }[1].
B. Learning Property of Feature ID
In [1], for the sake of complexity, diagonal matrixes were
used to represent feature distributions. Fig. 3 shows the
distributions of feature responses for feature IDs. This image
shows that the distribution shapes of feature IDs in the nostril
and eye corner (top and middle images in Fig. 3(b)) are very
different from that of the feature ID in the cheek region
(bottom image in Fig. 3(b)). The eye corner and nostril
can be localized accurately since they have clear peaks in
the feature responses. However, the feature ID in the cheek
region cannot be localized accurately since they have a long
elliptical distribution. In order to reect the property, we use
full matrix to represent each distribution. This property is
also used in shape renement stage later.
To learn the property of kth feature ID, feature response
maps of local images around ground-truth are obtained in
training data, the mean of the feature response maps are
calculated, and distribution of the mean is described by a
full covariance matrix as follows:

k
=

20
/
00

11
/
00

11
/
00

02
/
00

(2)
where
pq
=

y
(x x)
p
(y y)
q
F
k
(x, y),F
k
(x, y)
is feature response value at (x, y) and (x,y) is a centroid
position of feature response.
III. ALIGNMENT FACE WITH OCCLUSION
A. Feature Response Extraction
Unlike the conventional alignment methods which search
small areas based on a given initial shape[1], [2], [3], initial
shape is not required in the proposed method. However, for
the sake of efciency, a search window for each feature ID
is set according to a face position given by a face detector.
Within the search windows, feature responses are calculated
and candidate feature points are obtained.
To obtain multiple candidate feature points from a multi-
modal distribution of feature response, the feature response
240
is partitioned into multiple segment regions by the Mean-
Shift segmentation algorithm[9]. Then the distributions in
the segmented regions are approximated through convex
quadratic functions as follows:
argmin
A
l
k
,b
l
k
,c
l
k

x
l
k
||E
k
{Y (x
l
k
+ x)} x
T
A
l
k
x
+2b
lT
k
x c
l
k
||
2
subject to A
l
k
> 0 (3)
where
l
k
is a lth segment region in kth feature ID, x
l
k
is
the centroid of
l
k
, E
k
() is the inverted match-score function
obtained by applying the kth feature detector to the source
image Y , A
l
k
=

a
11
a
12
a
21
a
22

, and b
l
k
=

b
1
b
2

.
B. Shape Hallucination
In this paper, it is assumed that at least half of feature
IDs are not occluded. Let the number of candidate feature
points for ith feature ID be N
P
i
(1 < i < N
I
) where N
I
is number of the feature IDs. Then, among total feature
points

N
I
i=1
N
P
i
, at least N
I
/2 feature points and at most
N
I
feature points exist that support an input facial image.
Therefore, the correct combination of feature points should
be found. In order to nd the correct combination of feature
points, RANSAC[10], [11] is employed.
1) Selecting proposal feature points: In using RANSAC
strategy, there are two steps of random samplings in our
method. The rst one is to select M feature IDs. This
is to propose a set of non-occluded feature IDs. The M
is a minimum number of feature IDs that is required to
hallucinate a face shape. Second we select a feature point
among multiple feature points for the each selected feature
ID. This is to propose true feature points among multiple
feature responses for the feature IDs.
The sampling is iterated until all the samples are inliers,
and the maximum number of iterations k in RANSAC can
be determined as follows[10]:
k =
log(1 p)
log(1 w)
(4)
where p is the probability that the RANSAC selects only
inliers in k iteration and w is the probability that all M
points are inliers. In our case, w is

M1
i=0

N
I
/2i
N
I
i
1
N
P
j

where N
P
j
is the number of feature points for jth feature
ID. For a simple example, let N
P
j
for all j are 2 and p, M,
and N
I
be 0.99, 10, and 83. Then, the maximum number of
iterations k is larger than 0.9 10
6
, which is not feasible.
In our experiments, a proper set of feature points could be
selected in 400 iterations in many cases, but in some other
cases it could not be done even after 2, 000 iterations.
In order to reduce the number of iterations, we propose
a method of coarse ltering of non-candidate feature points.
The method is using RANSAC based on the concept of the
generalized Hough transform. Two feature points which do
not belong to same feature IDs are randomly selected, and
scale and rotation are calculated. Then, a centroid of face
shape is estimated from the two points and agreement with
centroids obtained from other remaining feature points with
respect to the scale and the rotation is tested. The details of
the proposed ltering method is as follows:
1 Select two random feature points y
i
, y
j
which do not
belong to same feature ID.
2 Estimate scale S
l
and rotation R
l
parameters from two
feature points y
i
, y
j
in lth iteration.
3 Calculate distributions of estimated centroid of shape
for all feature points: F(x) =

T
k=1
f(x|y
k
, S
l
, R
l
)
where for given S
l
and R
l
the f is a normal distribution
function of centroid for feature point y
k
and T =

N
I
i=1
N
P
i
. x represents a position in an image.
4 Get maximum value C
l
S
l
,R
l
= max F(x) and the posi-
tion x
l
max
= arg max
X
F(x).
5 Repeat 1 to 4 until the number of iterations reaches
the given number or the C
l
S
l
,R
l
is larger than a given
threshold.
6 Get S
L
, R
L
, and x
L
max
where L = arg max
l
C
l
S
l
,R
l
.
7 Given S
L
and R
L
, calculate Mahalanobis distance be-
tween x
L
max
and each feature point with distribution
f(x
L
max
|y
k
, S
l
, R
l
).
8 Take out a feature point that has minimum distance.
9 Repeat 6-8 until we get a given number of feature IDs,
so that the number is at least larger than M.
As a result of the coarse ltering, almost all selected feature
points are inliers. In our experiments, we could get a proper
inlier set within 5 iterations. Thus, the number of iterations
for selecting M correct feature points could be reduced from
several hundred to less than ve.
2) Hallucinating shapes: From the coarsely ltered fea-
ture points and feature IDs, M feature IDs are selected and
then M feature points, which are associated to the selected
feature IDs, are selected. Then the parameter p is calculated
explicitly as follows:
p = (
t
A)
1
b (5)
where
A =

A
sel

1
sel1
0
.
.
.
.
.
.
.
.
.
0 A
sel

M
sel
M

, b =

b
sel

1
sel1
.
.
.
b
sel

M
sel
M

(6)
The A
sel

j
sel
k
and b
sel

j
sel
k
are obtained by (3) where sel
k
and sel

j
are a kth element and a jth element among selected feature
points and feature IDs above and 1 r N
P
sel
k
.
In [1], they applied this tting method iteratively. Since
a distribution of a feature response is represented by one
quadratic tting function (unimodal distribution) even where
there could be multiple strong feature responses (multi-
modal distributions), it is required to reduce search region
iteratively until all quadratic tting functions localize correct
positions with smaller search regions for each iteration.
However, in our method, a feature response for each feature
ID is explicitly described by multi-model distributions. Thus,
given a set of selected feature points, a facial shape can be
hallucinated by a closed form.
241
3) Testing Hallucinated Shape: Since we are assuming
that at least half of all feature IDs are not occluded, we can
always accept N
I
/2 feature points in which the distances
between the positions of the hallucinated feature points and
of feature responses are smaller than the median of all the
residuals.
The residual between a position of a feature ID of the
shape, x and corresponding feature response of the feature
ID x can be calculated as follows:
R(x, x) = (x x)
t
[G(
m
) +
f
]
1
(x x) (7)
where
m
and
f
are covariance matrixes of shape model
and feature response. The G is a geometric transformation
function that is estimated from the hallucination shape. N
I
/2
feature points where the residual R(x
i
, x
i
) are smaller than
the median value of the residuals for all feature points x
i
are
accepted as inliers. Using the inliers, the shape is updated.
The shape is tested in terms of the median of the
residuals[12]. Repeat the proposal feature points selection,
the hallucination and the test until the median is smaller
than a given threshold or the number of the iteration reaches
a given number. Then, a hallucinated shape, with minimum
error, results in an aligned shape with only inlier feature
points.
C. Shape renement
In the previous step, at least N
I
/2 feature points were
selected safely. However, there can be more unoccluded
feature points. For example, in the case of a non-occluded
facial image, almost of all feature IDs are visible, but
only half of the feature points will be selected through
previous steps. There can be more visible inliers than half.
In the shape renement, more feature points are chosen as
inliers by means of measuring Mahalanobis distance between
the hallucinated feature point and a peak point of feature
response around the hallucinated feature point.
In order to pick up a peak point around the hallucinated
feature point, strong feature responses on perpendicular lines
to tangent lines at hallucinated feature points are searched.
Then the picked up position, which has strong feature
responses, can be accepted as an inlier. There are two ways
of searching strong feature points along the perpendicular
lines.
The rst is to search the strongest feature response along
the perpendicular lines as it is used in [2]. The feature
responses on the perpendicular lines are searched. However,
if a true strong feature response point is slightly off from the
perpendicular line, this fails to pick up the point.
The second is not to see the feature response only on
the perpendicular line but also to see neighbors according
to properties of feature IDs. As mentioned, each feature ID
has a different property of feature response distributions. As
shown in the Fig. 2(b) and Fig. 3, it is clear that the feature
response of the cheek region is distributed along the edge of
the cheek which is almost a linear shape, while that of the
nostril is distributed as almost a circular shape. By seeing
its neighboring responses, although a feature response on a
(a) (b)
Fig. 4. Distance metrics for feature IDs on eye corner and cheek. The 3
red dots in each gure represent current and its neighboring feature points.
perpendicular line is weak, this gives a chance to pick up the
point. If the picked up point has a strong feature response,
the point is selected as an inlier and contributes to rene the
shape.
The strength of the feature response S along the perpen-
dicular line is dened as follows:
S
i
(k) =

x
N
i
(x|k, )(x) (8)
where the (x) and the N
i
(x) are a feature response at x
and the normal probability density function. The i and k
are corresponding ith feature IDs to this feature point and a
position on the perpendicular line, respectively.
IV. ERROR MEASUREMENT
The properties of feature IDs shown in Fig. 3 give us
a very interesting point of view on how to measure errors
between aligned and groundtruth shapes. Conventionally, the
error was measured by Euclidean distance. This measurement
calculates distance without considering the feature properties.
Unlike the feature points of eye corners and lip corners, the
feature points of cheek regions and middle of eye-brows are
the hard to point out exact positions in an image. Many
of the positions on the edge can be answers. However, the
conventional Euclidean distance cannot handle with these
answers.
We propose a new metric based on shape property which
considers geometrical properties of feature IDs as follows:
1) If a feature ID is positioned on a sharp point such as
corners of eyes, the metric is likely to be a Euclidian
distance between a target point and associated ground-
truth point.
2) If a feature ID is positioned on a middle line such as
a cheek, the metric is a distance between the edge line
and a target point.
It is dened as follows:
D(i, x) = min(d(y
i1
, v
x
), d(y
i+1
, v
x
)) (9)
d(y
k
, v
x
) =

a(y
k
, v
x
)
2
sin + b(y
k
, v
x
)
2
(10)
where
a(y
k
, v
x
) =
v
x
y
k
| y
k
|
,
b(y
k
, v
x
) =| c(y
k
, x) a(y
k
, x) |
c(y
k
, v
x
) =| v
x
y
k
|,
sin =

y
k
v
h
| y
k
|| v
h
|

2
.
242
The i is an index of ith feature point of ground-truth and v
h
is a tangent line at the ith feature point. The v
x
and y
k
are

g
i
x and

g
i
g
k
, respectively, where the g
k
is a ground-truth
position of kth feature IDs.
Fig. 4 shows examples of distance metrics for feature IDs
on an eye corner and cheek. This measurement allows a point
on the cheek to move along the lines.
V. EXPERIMENTS AND ANALYSIS
In order to show robustness of the proposed method, we
conducted two experiments in terms of accuracy and stability.
The accuracy represents how face alignment is localized
accurately and the stability represents how face alignment
is stable under occlusions.
For experiment and analysis, two data sets were used: AR
database (ARDB)[13], and images in which occlusions were
generated articially (AGO). The ARDB includes images
of non-occlusion and two types of occlusions: (1) non-
occlusion, (2) upper parts occlusion with sunglasses, and (3)
lower parts occlusion with a scarf. For the AGO images,
occlusions were articially generated from non-occluded
face images.
In order to measure error of an alignment shape, we
introduce two concepts of errors: (e
n
) errors between ground-
truth and associated alignment points in the non-occluded
part and (e
o
) errors between those in the occluded part. The
e
n
represents how well the shape is aligned regardless of
occlusions. The error e
o
represents how well the occluded
points are hallucinated. In this paper, e
n
is our main concern.
125 images in the ARDB were used for learning feature
detectors and PDM. For the measurement, each image was
scaled according to the ground-truth of the eye positions so
that the inter-ocular distance was 50 pixels.
A. Accuracy test
In an accuracy test, localization accuracy was measured
using e
n
on the ARDB. Note that the test images includes
occluded images as well as non-occluded images.
Fig. 5 shows the result of the accuracy test. The test result
using the BTSM[2] is also given for comparison. The x-axis
represents images and the y-axis represents mean errors of
the alignment shapes. The images in the bottom of Fig. 5
show face detection (Black line) and the alignment results by
the BTSM (Blue line) and the proposed method (Red line).
The proposed method (Red circles) outperformed the BTSM
(Blue rectangles) for almost all of the images regardless of
occlusions. As the errors of the alignment by the BTSM
are increasing signicantly from 27th image, that of the
alignment by the proposed method are not increasing. Many
alignment results by the BTSM had large errors due to the
initial position problem as well as occlusions.
More examples are shown in Fig. 6. The rst row shows
face detection results and the second and third rows show
alignment results by the BTSM and the proposed method.
Also, feature points identied as visible were shown as
yellow circles in the third row. The proposed method showed
Fig. 5. The accuracy test. The x-axis represents images and the y-axis
represents errors of the alignment shapes. The images are aligned by the
errors of the BTSM results. Blue rectangles and red circles represent the
BTSM and the proposed face alignment method, respectively.
very good performance on the occluded images as well as
non-occluded images.
B. Stability test
A stability test is to measure robustness of face alignment
where various partial occlusions exist. For this test, AGO
images were generated as follows: A non-occluded facial
image on the ARDB was partitioned so that each partition
can cover 6% to 7% of facial features in ground-truth. Up to
9 partitions were randomly chosen and the selected partition
regions were lled with black so, at most, 63% of facial
features were occluded. For each degree of occlusion (the
percentage of occluded facial features), 5 occlusions of 5
subjects were generated. Some images are shown in the
bottom of Fig. 7.
In this stability test, the alignment results on original non-
occluded images were used as ground-truth.
Fig. 7 shows test results. The y-axis represents errors mea-
sured by (9) and the x-axis represents degree of occlusions.
The center line represents the mean errors with maximum
(above) and minimum (below) errors. The images in the
bottom of Fig. 7 show examples of alignment results on
the image with 1, 3, 5, and 7 degrees of occlusions. Feature
points identied as visible ones were shown as yellow circles.
In the Figure, with up to 7 degrees of occlusions, the
error is not increased signicantly and it means the proposed
method is stable where at least half of facial features are
visible. After 8 degrees of occlusions, the error is increased
drastically because more than half of facial features are
occluded.
VI. CONCLUSIONS
In this paper we proposed an approach of face alignment
that is robust to partial occlusion. In the proposed approach,
the feature responses from face feature detectors are explic-
itly represented by multi-modal distributions, hallucinated
243
Fig. 6. Comparison of face alignment results. The rst row shows input images and its face detection result, the second row the alignment results by the
BTSM, and the third row the alignment results by the proposed method. The rst to fourth columns are images of ARDB and the fth and sixth are those
which occlusions were articially generated.
Fig. 7. The stability test. The x-axis represents deree of occlusions and
the x-axis represents errors. The center line represents the mean errors with
maximum and minimum errors. The images at the bottom show examples
of alignment results at several degrees of occlusion - 7, 22, 36 and 50%.
facial shapes are generated, and a correct alignment is
selected by RANSAC hypothesize-and-test. A coarse inlier
feature points selection method, which reduces the number
of iteration in the RANSAC, and a new error measurement
metric that is based on shape properties are also introduced.
We evaluated the proposed method on ARDB and im-
ages in which occlusions were generated articially. The
experimental results have demonstrated that the alignment
is accurate and stable over a wide range of degrees and
variations of occlusion. Also, it shows that it is better to
use partial information which is not corrupted (visible) than
to use all information that may be corrupted (invisible).
Acknowledgment This research was supported by ONR
MURI Grant N00014-07-1-0747.
REFERENCES
[1] Y. Wang, S. Lucey and J. F. Cohn, Enforcing Convexity for Improved
Alignment with Constrained Local Models, in Proc. of the IEEE Conf.
on Computer Vision and Pattern Recognition, 2008.
[2] Y. Zhou, L. Gu, and H. Zhang, Bayesian Tangent Shape Model:
Estimating Shape and Pose Parameters via Bayesian Inference, in
Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition,
vol. 1, Wisconsin, 2003, pp. 109-116.
[3] T.F. Cootes, C.J. Taylor and Manchester M. Pt, Statistical Models of
Appearance for Computer Vision, 2000.
[4] L. Gu and T. Kanade, A Generative Shape Regularization Model for
Robust Face Alignment, in Proc. of the 10th European Conference
on Computer Vision, 2008.
[5] D. Cristinacce, T.F. Cootes, Feature Detection and Tracking with
Constrained Local Models, in British Machine Vision Conference,
2006, pp. 929-938.
[6] P. Viola, M. Jones, Robust Real-Time Face Detection, Int. Journal of
Computer Vision, vol. 57, no. 2, 2004, pp. 137-154.
[7] C. M. Bishop, Pattern Recognition and Machine Learning. Springer,
2006.
[8] A. Ross, Procrustes analysis, Technical Report, Department of Com-
puter Science and Engineering, University of South Carolina, SC
29208.
[9] D. Comaniciu and P. Meer, Mean Shift: A Robust Approach Toward
Feature Space Analysis, The IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol 24, 2002, pp 603-619.
[10] M.A. Fischler and R.C. Bolles, Random Sample Consensus: A
Paradigm for Model Fitting with Applications to Image Analysis
and Automated Cartography, Comm. of the ACM, vol. 24, 1981, pp.
381395.
[11] Y. Li, L. Gu, and T. Kanade, A Robust Shape Model for Multi-View
Car Alignment, in Proc. of the IEEE Int. Conf. on Computer Vision
and Pattern Recognition, 2009.
[12] P.J. Rousseeuw. Robust regression and outlier detection, Wiley, New
York, 1987.
[13] A.M. Martinez and R. Benavente, The AR Face Database, CVC
Tech. Report #24, 1998.
244