Identify computer generated characters by analysing facial expressions variation

Abstract

Significant improvements have been recently achieved in both quality and realism of computer generated characters, which are nowadays often very difficult to be distinguished from real ones. However, generating highly realistic facial expressions is still a challenging issue, since synthetic expressions usually follow a repetitive pattern, while in natural faces the same expression is usually produced in similar but not equal ways. In this paper, we propose a method to distinguish between computer generated and natural faces based on facial expressions analysis. In particular, small variations of the facial shape models corresponding to the same expression are used as evidence of synthetic characters.

Full text

Identify computer generated characters
by analysing facial expressions variation
Duc-Tien Dang-Nguyen, Giulia Boato, Francesco G.B. De Natale
Department of Information and Computer Science - University of Trento
via Sommarive 5, 38123 Trento - Italy
dangnguyen@disi.unitn.it
boato@disi.unitn.it
denatale@ing.unitn.it
Abstract—Significant improvements have been recently
achieved in both quality and realism of computer generated
characters, which are nowadays often very difficult to be distin-
guished from real ones. However, generating highly realistic facial
expressions is still a challenging issue, since synthetic expressions
usually follow a repetitive pattern, while in natural faces the
same expression is usually produced in similar but not equal
ways. In this paper, we propose a method to distinguish between
computer generated and natural faces based on facial expressions
analysis. In particular, small variations of the facial shape models
corresponding to the same expression are used as evidence of
synthetic characters.
I. INTRODUCTION
Digital graphics tools are nowadays widespread and ex-
ploited to create realistic computer media data both from
professional and non-professional users. In particular, com-
puter generated (CG) characters are increasingly used in many
applications such as talking-faces, e-learning, virtual meeting
and especially video games. Since the first virtual newsreader
Ananova
1
introduced in 2000, significant improvements have
been achieved in both quality and realism of CG characters,
which are nowadays often very difficult to be distinguished
from real ones.
At one hand, these results open a new area for advance
human-computer interaction. On the other hand, non existing
subjects or situations can be generated leading to the need
of techniques assessing data trustability and authenticity with
sufficient confidence. Therefore, the research community has
recently focused on the development of tools supporting the
discrimination between natural and CG multimedia content in
an accurate and reliable way.
In multimedia forensics, approaches distinguishing between
CG and natural data have been developed since 2005. Most of
them focus on still images, by estimating statistical differences
in wavelet-based decomposition [1][2]; by modelling physical
differences like local patch statistics, fractal and quadratic
geometry, and gradient on surface [3]; by evaluating the
1
http://news.bbc.co.uk/2/hi/entertainment/718327.stm
WIFS‘2012, December, 2-5, 2012, Tenerife, Spain. 978-1-
4244-9080-6/10/$26.00
c
2012 IEEE.
noise of the recording device [4]; by combining different
informations like the hybrid approach in [5]. Recently, a
geometric approach supporting the distinction of CG and real
human faces has been presented in [6], which exploits face
asymmetry as a discriminative feature. However, to the best
of our knowledge, there is no multimedia forensics approach
that aims at discriminating between CG versus natural objects
or subjects in video sequences. Such a goal requires different
techniques with respect to the state-of-the-art.
In this paper we propose a method to distinguish between
CG and real characters by analysing facial expressions. Repro-
ducing facial expressions is one of the most challenging issues
in creating virtual characters [7] and there are studies back to
1971 that analyse this problem (see for instance [8]). Most of
the algorithms generate synthetic facial expressions following
the Facial Action Coding System (FACS) by Ekman [9][10]
or MPEG-4 standard [11]. In FACS, the muscles on face
are coded as Action Units (AUs), and an expression is then
represented as a combination of AUs. In MPEG-4, explicit
movements for each point on the face is defined by Facial
Animation Parameters (FAPs). Based on these parameters
(FACS or FAPs), a physically-based model is applied to make
it more realistic. However, when CG contents become very
realistic they often become also unfamiliar (so called Uncanny
valley [12]) and recently some approaches attempt to overcome
such a problem [7][13]. Here, we propose to exploit this gap
to differentiate between computer generated and natural faces.
The underlying idea is that facial expressions in CG charac-
ters follow a repetitive pattern, while in natural faces the same
expression is usually produced in similar but not equal ways
(e.g., human beings do not always smile in the same way). Our
forensic technique take as input various instances of the same
character expression (extracting corresponding frames of the
video sequences) and determine whether the character is CG or
natural based on the analysis of the corresponding variations.
We show that CG faces often replicate the same expression
exactly in the same way, i.e., the variations is smaller than the
natural ones, and can therefore be automatically detected.
The rest of this paper is organized as follows: the proposed
method is described in section II, experimental results are re-
ported in section III, while section IV draws some conclusions.

II. PROPOSED METHOD
Our method contains five steps as detailed in Figure 1. From
a given video sequence, frames that contain human faces are
extracted in the first step A. Then, in step B, facial expression
recognition is applied in order to recognize the expressions of
the faces. Six types of facial expressions are used in this step,
following the six universal expressions of Ekman (happiness,
sadness, disgust, surprise, anger, and fear) [9] plus a ‘neutral’
one. Based on the recognition results, faces corresponding to a
particular expression (e.g., happiness) are selected for the next
steps. Notice that the ‘neutral’ expressions are not considered,
i.e., faces showing no expression are not taken into account for
further processing. In the next step C the Active Shape Model
(ASM), which represents the shape of a face, is extracted from
each face. In order to measure their variations, all shapes have
to be comparable. Thus, in step D, each extracted ASM is
then normalized to a standard shape. After this step, all ASM
shapes are normalized and are comparable. Finally, in step
E, differences between normalized shapes are analysed, and
based on the variation analysis results, the given sequence is
confirmed to be CG or natural.
The right part of Figure 1 shows an illustration of the
analysis procedure on happiness expression. Seven frames that
contain faces are extracted in step A. Then, facial expression
recognition is applied in step B and three happy faces are
kept. For each face, the corresponding ASM model, which is
represented by a set of reference points, is extracted in step
C. Then, each model is normalized to a standard shape, step
D. All normalized shapes are then compared together in step
E, and based on the analysis results the given character is
confirm as computer generated since the differences between
the normalized shapes are small (details about the variation
analysis are given in the following Subsection II-E).
A. Human faces extraction
Face detection problem has been solved with the Viola-
Jones method [14], which can be applied in real-time ap-
plications with a high accuracy. In this step, we reuse this
approach to detect faces from video frames, and frames that
contain faces are extracted. More details about this well-known
method can be found in [15] and [14]. It is worth mentioning
that in this first work we do not face the problem of face
recognition, thus assuming to have just a single person per
video sequence (the analysed character).
B. Facial expression recognition
Facial expression recognition is a nontrivial problem in
facial analysis. In this study, we applied an EigenFaces-
based application [16] developed by Rosa for facial expression
recognition. The goal of this step is to filter out the outlier
expressions and keep the recognized ones for further steps.
Notice that this application associates an expression to a given
face without requiring any detection of reference points. In
Figure 1 an example of results of this application is shown
with 7 faces (3 happy, 2 disgust, 1 surprise and 1 neutral).
Fig. 2. The 87 points of Active Shape Model (ASM). Source: Microsoft
Research Face SDK.
C. Active Shape Model Extraction
Input images for this step are confirmed to have the same
facial expression of the same person, thanks to the prepro-
cessing in the first two steps. In order to extract face shapes,
which are used in our analysis, an alignment method is applied.
In this step, we follow the Component-based Discriminative
Search approach [17], proposed by Liang et al. The general
idea of this approach is to find the best matching from the
mode candidates, where modes are important predefined points
on face images (e.g., eyes, nose, mouth) and are detected from
multiple component positions [17]. Given a face image, the
result of this step is a set of reference points, representing the
detected face. In Figure 3 (a) an example of this step is shown,
where the right image shows reference points representing the
face in the left image. In this method, the authors exploit
the so called ASM, which contains 87 reference points as
shown in Figure 2. Another example of this step on a CG
face is also reported in Figure 3 (c), where the left image
shows the synthetic facial image and the right one shows the
corresponding ASM.
D. Normalized Face Computation
ASM models precisely and suitably represent faces, but
they are incomparable since faces could be different in sizes
or orientations. They need to be normalized in order to be
comparable. In this step, we apply the traditional approach
from [18] to normalize a shape of a face in order to have
a common coordinate system. This normalization is an affine
transformation used to transform the reference points into fixed
positions. Since eye inner corners and the philtrum are stable
under different expressions, these points have been chosen
as reference points. Shown in Figure 2, the reference points
number 0 and 8 are two inner eye corners. The last reference
point, the philtrum, can be computed via the top point of outer
lip and two nostrils (point 51 and 41, 42 on the ASM model,
respectively), as follows:
p
philtrum
=
p
41
+p
42
2
+ p
51
2
(1)

Disgust Happy Surprise Happy Disgust Happy Neutral
Human Faces
Extraction
Facial Expression
Recognition
Active Shape Model
Extraction
Variation Analysis
Normalized
Face Computation
RESULT
INPUT
?
A
B
C
D
E
Fig. 1. Schema of the proposed method: A. Human faces are extracted from the video sequence(s). B. Facial expressions are recognized (in the example 3
happy, 2 disgust, 1 surprise and 1 neutral). C. Faces with the same expression are selected (in this example only happy faces) and their active shape models
are extracted. D. The extracted models are normalized. E. Differences on the normalized models are analysed to determine whether the character is CG.
where p
41
, p
42
, and p
51
are the reference points on the
extracted ASM.
After computing the three reference points, each ASM
model is normalized by moving {p
41
, p
42
, p
phitrum
} into
their normalized positions, as follows: (i) rotate the seg-
ment [p
41
, p
42
] into an horizontal line segment; (ii) shear the
philtrum to be on the perpendicular line through the middle
point of [p
41
, p
42
]; and finally (iii) scale the image so that the
length of segment [p
41
, p
42
] and the distance from p
philtrum
to [p
41
, p
42
] have predefined fixed values (see [18] for more
details).
Shown in Figure 3 (b) and (d) are examples of the normal-
ized faces after Face Normalization step. The left images show
the normalized faces and the right ones show the normalized
reference points.
E. Variation Analysis
In this step, differences among normalized ASM models are
analysed in order to determine if a given character (and there-
fore the corresponding set of faces) is CG or real. We analyse
the differences as described in the following paragraphs.
First, the distance d
i,p
of each reference point p on a model
Fig. 3. ASM and normalized ASM: (a) and (c) show a photographic and
a computer generated happy face, respectively, and their corresponding ASM
points; (b) and (d) show the normalized images of (a) and (c), respectively,
and their corresponding normalized points.

i to the average of all points p of all models is calculated as:
d
i,p
= k(x, y)
i,p
− (x, y)
p
k (2)
where (x, y)
i,p
is the position of the reference point p on the
model i; (x, y)
p
=
1
N
N
P
i=1
(x, y)
i,p
, where N is the number of
normalized ASM models; and k·k is Euclidean distance.
Depending on the facial expression ξ (among six universal
expressions), a subset S
ξ
of reference points (not all 87 points)
are selected for the analysis. For example, with the happy
facial expression (ξ = 1) only reference points from 0 to 15
and from 48 to 67, which represent the eyes and the mouth, are
considered, i.e., S
1
= {0, 1, 2..15, 48, 49, .., 67}. The subsets
are selected based on our experiments and suggestions from
EMFACS [9], in which a facial expression is represented by a
combination of AUs codes. Shown in Table I are the reference
points selected in our method and the correspondent AUs
codes from EMFACS. Some explanations of the AUs codes
are also listed in Table II. Full codes in EMFACS could be
seen in [9].
TABLE I
EXPRESSIONS WITH ACTION UNITS AND CORRESPONDENT ASM POINTS
ξ Expression Action Units (AUs) Reference Points (S
ξ
)
1 Happiness 6+12 S
1
= {0 − 15, 48 − 67}
2 Sadness 1+4+15 S
2
= {0 − 35, 48 − 57}
3 Surprise 1+2+5B+26 S
3
= {16 − 35, 48 − 67}
4 Fear 1+2+4+5+20+26 S
4
= {16 − 35, 48 − 57}
5 Anger 4+5+7+23 S
5
= {0 − 64}
6 Disgust 9+15+16 S
6
= {0 − 15, 48 − 67}
TABLE II
EXAMPLE OF SOME FACIAL ACTIONS [9]
AU Number FACS name
1 Inner Brow Raiser
4 Brow Lowerer
6 Cheek Raiser
12 Lip Corner Puller
15 Lip Corner Depressor
.. ..
Two main properties are taken into account in this analysis:
mean and variance, calculated as their traditional definitions:
µ
p
=
1
N
N
X
i=1
d
i,p
, and σ
p
=
1
N
N
X
i=1
||d
i,p
− µ
p
||
2
(3)
where µ
p
and σ
p
are the mean and variance of all distances
d
i,p
at reference point p over all models.
The given set of models on expression ξ is confirmed to be
CG or natural by comparing the Expression Variation Value
EV V
ξ
to the threshold τ
ξ
. The value of EV V
ξ
is computed
as follows:
EV V
ξ
= α
ξ
1
|S
ξ
|
P
p
µ
p
λ
1,ξ
+ (1 − α
ξ
)
max
p
{σ
p
}
λ
2,ξ
(4)
where α
ξ
is a weighted constant, α
ξ
∈ [0; 1]; λ
1,ξ
and λ
2,ξ
are the normalization values used to normalize the numerators
into [0; 1]. In our experiments α
ξ
are set to 0.7 for ξ = 1, ..., 6.
EV V
ξ
is then compared with τ
ξ
, recognizing the character
corresponding to the set of faces as CG if EV V
ξ
< τ
ξ
, natural
otherwise.
Shown in Figure 4 are the mean values, corresponding to all
87 ASM points, for the sadness expression (ξ = 2) analysed
on the two set of images shown in Figure 5 (a). The horizontal
axis represents p, from 1 to 87, while the vertical axis shows
the value of µ
p
. Since the facial expression is sadness (ξ =
2), only the values from µ
0
to µ
35
and from µ
48
to µ
57
are
considered (see the selected reference points in Table I). In this
example, the Expression Variation Value EV V
2
of the CG face
is 0.35 comparing to 0.74 of the natural one (τ
2
= 0.6).
0 10 20 30 40 50 60 70 80 90
0
0.5
1
1.5
2
2.5
3
3.5
Computer Generated
Natural
Fig. 4. Example of differences on the mean of ASM points between CG
and photographic sad faces of Figure 5 (b).
Values of the thresholds τ
ξ(ξ=1..6
) are manually set based on
experiments, with the goal of keeping the miss classification
as small as possible.
III. EXPERIMENTAL RESULTS
In our experiments, we use two public datasets:
• Bo
˘
gazic¸i University Head Motion Analysis Project
Database (BUHMAP-DB) [19], which contains 440
videos of 11 people (6 female, 5 male) performing
5 repetitions on 8 different gestures. We selected the
happiness and sadness from this database, since the other
six gestures are not related to our topic. Finally, we have
110 videos from this dataset. Each video lasts about 1 -
2 seconds.
• The Japanese Female Facial Expression (JAFFE)
Database [20], which contains 213 images of 7 facial
expressions posed by 10 Japanese female models.
The first experiment is performed on happiness and sadness
expressions from BUHMAP-DB videos. Starting from the 11
people of BUHMAP-DB, we created 11 CG characters by
using FaceGen [21] and morphed all of them into both happy
and sad faces. FaceGen is a powerful tool which can be used
in building complex face structures from one to three images.
In our case, we pass a ‘neutral’ image to FaceGen in order to

build the face structure, then we use Morph options to generate
happiness and sadness expressions on the new generated face.
Thus, we obtained 110 sets of happy and sad faces, where
each model has 5 sets corresponding to happiness and 5 sets
corresponding to sadness. Shown in Figure 5 are two examples
of the CG versions and the original faces from BUHMAP-DB.
(a)
(b)
Fig. 5. Examples of (a) happy, and (b) sad faces from BUHMAP-DB and
the corresponding CG faces generated via FaceGen.
(a) (b) (c)
Fig. 6. Examples of (a) happy, (b) sad, and (c) surprised faces from JAFFE
and the corresponding CG faces generated via FaceGen.
The goal of this experiment is to analyse the differences
from CG models with the natural faces in order to confirm
the idea of the proposed method. The analysis is performed
as follows: for each video sequence 10 frames are uniformly
extracted and similarly for each CG model 10 images are
selected. Then, the sets of images are analysed and the corre-
sponding Expression Variation Values computed as described
in Section II-E. In this case since the expressions are already
known, we implement the method from step C. In this step,
we use Microsoft Face SDK [22] to extract the ASM models.
Finally, we apply step D and E to get the results.
Shown in Figure 7 are EV V
1
values computed on the 55
sets of CG and the 55 sets of natural happy faces. These values
are well separated between CG and natural. There is only
one miss classification using the threshold τ
1
= 0.45. The
accuracy, therefore, is 99% (equals 109/110).
On sadness expression, the result is even better, with 100%
of accuracy using the threshold τ
2
= 0.6. The EV V
2
values
for CG and natural characters are perfectly separated, as shown
in Figure 8.
Our second experiment is performed on the JAFFE database,
which contains all six expressions. Also in this case we used
FaceGen [21] to create the CG models reproducing the JAFFE
models (see Figure 6 for some examples). For each model in
this database, we reproduced all 6 expressions. Therefore, we
perform the second test on 120 sets of images, 60 sets of
CG and 60 sets of JAFEE real faces. The complete proposed
approach described in Section II is applied as a classification
approach on these sets.
Shown in Figure 9 is the average EV V
ξ
for each expression
(ξ = 1, ..., 6). The inner blue boundary represents the EV V
ξ
5 10 15 20 25 30 35 40 45 50
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Computer Generated
Natural
Fig. 7. Facial Expression Values computed on happiness expression. The
threshold value τ
1
is 0.45. The separation between CG and natural EV V
1
is
clearly visualized with only one miss classification.
5 10 15 20 25 30 35 40 45 50
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Computer Generated
Natural
Fig. 8. Facial Expression Values computed on sadness expression. The
threshold value τ
2
is 0.6. CG and natural EV V
2
are clearly separated.
computed from CG sets of images, and the outer red boundary
represents the natural EV V
ξ
. Results show that CG and natural
Expression Variation Values can be differentiated by using and
comparing with a set of thresholds τ
ξ
, visualized by the green
boundary. The classification performance of this experiment is
in average 96.67%. Details for each expression are reported
in the confusion matrices, Table III.
TABLE III
CONFUSION MATRICES ON CG AND NATURAL FACES, COMPUTED ON
JAFFE DATABASE.
ξ Expression CG Natural
1 Happiness CG 100% 0%
Natural 0% 100%
2 Sadness CG 100% 0%
Natural 0% 100%
3 Surprise CG 100% 0%
Natural 0% 100%
4 Fear CG 90% 10%
Natural 0% 100%
5 Anger CG 100% 0%
Natural 0% 100%
6 Disgust CG 80% 20%
Natural 10% 90%

0"
0.2"
0.4"
0.6"
0.8"
1"
Happiness"
Sadness"
Surprise"
Anger"
Fear"
Disgust"
Computer<Generated"
Natural"
Threshold"
Fig. 9. Average of Expression Variation Values analysed for all expressions.
CG and natural EV V
ξ
are separated for all ξ = 1, ..., 6.
The last experiment is performed by comparing Star Trek
Aurora
2
movie, a fully-animated product, against Star Trek
Odyssey, a live action movie from Star Trek: Hidden Frontier
series
3
. In Star Trek Aurora, two graphics applications, namely
Poser and Cinema 4D, are used to create the entire 3D world
and characters. We extracted 4 female characters in each
movie and selected frames that contain happy expression of
those characters. Happy faces are then confirmed by using
Rosa application [16]. An illustration of two characters in
happy emotion is shown in Figure 10. Finally, EV V s are
computed and compared. Using the same threshold as in the
first experiment (τ
1
= 0.45), all EV V
1
calculated for the 4
characters of Star Trek Aurora are smaller than τ
1
while all of
the EV V
1
from Star Trek Odyssey are over τ
1
, i.e., the CG
characters can be recognized and separated from the natural
ones.
(a)
(b)
Fig. 10. Examples of happy faces extracted from (a) Star Trek Aurora, (b)
Star Trek Odyssey.
IV. CONCLUSIONS
In this study, we introduced a novel problem about dif-
ferentiating between CG and natural human faces in video
sequences and we presented a method that allows distinguish-
ing CG characters based on facial expression analysis. Indeed,
results show that CG persons usually present smaller differ-
ences in face shape changing among the same expression, in
2
http://auroratrek.com
3
http://www.hiddenfrontier.com
comparison with real persons. Although experimental results
are performed just on small datasets, we proved that the
method can be effective. Further work will be devoted to au-
tomatic selection of thresholds and exploitation of transitional
parameters of faces.
REFERENCES
[1] S. Lyu and H. Farid, “How realistic is photorealistic?” IEEE Transac-
tions on Signal Processing, vol. 53, no. 2, pp. 845–850, 2005.
[2] Y. Wang and P. Moulin, “On discrimination between photorealistic and
photographic images,” in IEEE International Conference on Acoustics,
Speech and Signal Processing, 2006, pp. II.161–II.164.
[3] T. T. Ng, S. F. Chang, J. Hsu, L. Xie, and M. P. Tsui, “Physics-motivated
features for distinguishing photographic images and computer graphics,”
in ACM Multimedia, 2005, pp. 239–248.
[4] N. Khanna, G. T. C. Chiu, J. P. Allebach, and E. J. Delp, “Forensic tech-
niques for classifying scanner, computer generated and digital camera
images,” in IEEE International Conference on Acoustics, Speech and
Signal Processing, 2008, pp. 1653–1656.
[5] V. Conotter and L. Cordin, “Detecting photographic and computer
generated composites,” in SPIE Symposium on Electronic Imaging, 2011.
[6] D.-T. Dang-Nguyen, G. Boato, and F. G. B. DeNatale, “Discrimina-
tion between computer generated and natural human faces based on
asymmetry information,” in European Signal Processing Conference
(EUSIPCO), Bucharest, 2012.
[7] A. Tinwell, M. Grimshaw, D. A. Nabi, and A. Williams, “Facial
expression of emotion and perception of the Uncanny Valley in virtual
characters,” Computers in Human Behavior, vol. 27, no. 2, pp. 741–749,
Mar. 2011.
[8] P. Ekman and W. V. Friesen, “Constants across cultures in the face and
emotion,” Journal of Personality and Social Psychology, vol. 17, no. 2,
pp. 124–129, 1971.
[9] P. Ekman and W. Friesen, Facial Action Coding System: A Technique
for the Measurement of Facial Movement. Palo Alto: Consulting
Psychologists Press, 1978.
[10] P. Ekman, W. Friesen, and J. Hager, Facial Action Coding System
(FACS): Manual. Salt Lake City (USA): A Human Face, 2002.
[11] ISO, ISO/IEC 14496-2:1999: Information technology — Coding of
audio-visual objects — Part 2:Visual, 1999.
[12] M. Mori, “Bukimi no tani (the uncanny valley),” Energy, vol. 7, no. 4,
pp. 33–35, 1970.
[13] K. F. MacDorman, R. D. Green, C.-C. Ho, and C. T. Koch, “Too real for
comfort? Uncanny responses to computer generated faces,” Computers
in Human Behavior, vol. 25, no. 3, pp. 695–710, May 2009.
[14] P. A. Viola and M. J. Jones, “Robust real-time face detection,” Interna-
tional Journal of Computer Vision, vol. 57, no. 2, pp. 137–154, 2004.
[15] ——, “Rapid object detection using a boosted cascade of simple
features,” in CVPR, 2001, pp. 511–518.
[16] L. Rosa, “EigenExpressions for Facial Expression Recognition,”
http://www.advancedsourcecode.com/facialexpression.asp, 2007.
[17] L. Liang, R. Xiao, F. Wen, and J. Sun, “Face alignment via component-
based discriminative search,” in Proceedings of the 10th European
Conference on Computer Vision: Part II, ser. ECCV ’08. Berlin,
Heidelberg: Springer-Verlag, 2008, pp. 72–85.
[18] Y. Liu, K. L. Schidt, J. F. Cohn, and S. Mitra, “Facial asymmetry
quantification for expression invariant human identification,” Computer
Vision and Image Understanding Journal, vol. 91, no. 1/2, pp. 138–159,
2003.
[19] O. Aran, I. Ari, M. A. G
¨
uvensan, H. Haberdar, Z. Kurt, H. . T
¨
urkmen,
A. Uyar, and L. Akarun, “A database of non-manual signs in Turkish
sign language,” in Signal Processing and Communications Applications
(SIU2007), Eskis¸ehir, 2007.
[20] M. K. J. G. Michael J. Lyons, Shigeru Akamatsu, “Coding facial
expressions with gabor wavelets,” in IEEE International Conference on
Automatic Face and Gesture Recognition, 1998, pp. 200–205.
[21] “FaceGen Modeller from Singular Inversions,” http://www.facegen.com,
2004.
[22] “Microsoft Research Face SDK,” http://research.microsoft.com/en-
us/projects/facesdk/, May, 2012.