FACIAL LANDMARKING LOCALIZATION FOR EMOTION RECOGNITION USING BAYESIAN SHAPE MODELS

Jan Zizka (Eds) : CCSIT, SIPP, AISC, PDCTA - 2013
pp. 525–531, 2013. © CS & IT-CSCP 2013 DOI : 10.5121/csit.2013.3657
FACIAL LANDMARKING LOCALIZATION
FOR EMOTION RECOGNITION USING
BAYESIAN SHAPE MODELS
Hernan F. Garcia1
, Alejandro T. Valencia1
and Alvaro A. Orozco1
Technological University of Pereira, La Julita Periera, Risaralda, Colombia,
hernan.garcia@utp.edu.co,cristian.torres@utp.edu.co,
aaog@utp.edu.co
ABSTRACT
This work presents a framework for emotion recognition, based in facial expression analysis
using Bayesian Shape Models (BSM) for facial landmarking localization. The Facial Action
Coding System (FACS) compliant facial feature tracking based on Bayesian Shape Model. The
BSM estimate the parameters of the model with an implementation of the EM algorithm. We
describe the characterization methodology from parametric model and evaluated the accuracy
for feature detection and estimation of the parameters associated with facial expressions,
analyzing its robustness in pose and local variations. Then, a methodology for emotion
characterization is introduced to perform the recognition. The experimental results show that
the proposed model can effectively detect the different facial expressions. Outperforming
conventional approaches for emotion recognition obtaining high performance results in the
estimation of emotion present in a determined subject. The model used and characterization
methodology showed efficient to detect the emotion type in 95.6% of the cases.
KEYWORDS
Facial landmarking, Bayesian framework, Fitting, Action Unit
1. INTRODUCTION
This work is focus to the problem of emotion recognition in images sequences with varying facial
expression, ilumination and change pose in a fully automatic way. Although some progress has
already been made in emotion recognition, several unsolved issues still exist. For example, it is
still an open problem which features are the most important for emotion recognition. It is a
subject that was seldom studied in computer science. In order to define which facial features are
most important to recognize the emotion type, we rely on Ekman study to develop a methodology
which be able to perform a emotion recognition trough a robust facial expression analysis [1].
Although there have been a variety of works in the emotion recognition field, the results are not
yet optimal, because the techniques used for facial expressions analysis are not robust enough.
Busso et al. [2], presents a system for recognizing emotions through facial expressions displayed
in live video streams and video sequences. However, works such as those presented in [3] suggest

526 Computer Science & Information Technology (CS & IT)
that developing a good methodology of emotional states characterization based on facial
expressions, leads to more robust recognition systems. Facial Action Coding System (FACS)
proposed by Ekman et al. [1], is a comprehensive and anatomically based system that is used to
measure all visually discernible facial movements in terms of atomic facial actions called Action
Units (AUs). As AUs are independent of interpretation, they can be used for any high-level
decision-making process, including the recognition of basic emotions according to Emotional
FACS (EM-FACS), the recognition of various affective states according to the FACS Affect
Interpretation Database (FACSAID) introduced by Ekman et al. [4], [5], [6], [7]. From the
detected features, it is possible to estimate the emotion present in a particular subject, based on an
analysis of estimated facial expression shape in comparision to a set of facial expressions of each
emotion [3], [8], [9].
In this work, we develop a novel technique for emotion recognition based on computer vision
techniques by analysing the visual answers of the facial expression, from the variations that
present the facial features, specially those in regions of FACS. Those features are detected by
using bayesian estimation techniques as Bayesian Shape Model (BSM) proposed on [10], which
from the a-priori object knowledge (face to be analyzed) and being help by a parametric model
(Candide3 in this case) [11], allow to estimate the object shape with a high accuracy level. The
motivation for its conception, is an extension of the Active Shape Model (ASM) and Active
Appeareance Model (AAM) which predominantly considers the shape of an object class [12]
[13]. Although facial feature detection is widely used in facial recognition tasks, databases
searching, image restoration and other models based on coding sequences of images containing
human faces. The accuracy which these features are detected does not reach high percentages of
performance due some issues present at the scene as the lighting changes, variations in pose of
the face in the scene or elements that produce some kind of occlusion. Therefore, it is important
to analyze how useful would be the image processing techniques developed for this work. The
rest of the paper is arranged as follows. Section 2 provides a detailed discussion of model-based
facial feature extraction. Section 3 presents our emotion characterization method. Sections 4 and
5 discuss the experimental setup and results respectively. The paper concludes in Section 6, with
a summary and discussion for future research.
2. A BAYESIAN FORMULATION TO SHAPE REGISTRATION
The probabilistic formulation of shape registration problem contains two models: one denotes the
prior shape distribution in tangent shape space and the other is a likelihood model in image shape
space. Based on these two models we derive the posterior distribution of model parameters.
2.1 Tangent Space Approximation

Computer Science & Information Technology (CS & IT) 527
Fig. 1. The mean face model and all the training sets normalized face models [16]
3. EMOTION CHARACTERIZATION
3.1 Facial Expression Analysis
Facial expressions are generated by contraction or relaxation of facial muscles or by other
physiological processes such as coloring of the skin, tears in the eyes or sweat on the skin. We
restrict ourselves to the contraction or relaxation of the muscles category. The contraction or
relaxation of muscles can be described by a system of 43 Action Units (AUs) [1]
Labeling of the Emotion Type After choosing that AUs are used to represent the facial
expressions that correspond to each emotion, we chose Candide3 vertexes used to define the

configuration of the face and the calculation of the FACS [11]. Figure 2 shows the shapes of the
Candide3 models playing each emotion.
Fig. 2. Shape of each emotion type
4. EXPERIMENTAL SETUP
In this work we used the MUCT database which consists of 3755 faces from 276 subjects with 76
manual landmarks [16].
4.1 Evaluation Metrics
The shape model will be estimated for every subject, contemplating 5 changes of pose exposure
(-350
and 350
in the axis X and (-250
and 250
) in the axis Y to simulate the nods) and 1 frontal,
with which 1124 images will be had for changes of pose exposure and 276 frontal images. The
first step is to compute the average error of the distance between the manually labeled landmark
points pi and points estimated by the model for all training and test images. We compute the
relative error between the manually labeled landmark points and landmark points estimated by
the model for the eyelids region [17].
4.2 Landmarking The Training Set
The initial step is the selection of images to incorporate into the training set. In order to decide
which images will be include in the training set, the desired variations must be considered. In
order to include the whole region of the labeled face, there will be used a set of landmarks, which
consist in a group of 76 landmark points that depict the whole face regions in detail (eyes, nose,
mouth, chin, etc.) [16].
Fig. 3. Samples of the landmarking process

After estimate the landmark points of BSM, procrustes analysis is performed to align an
estimated facial expression with respect to standard shape (Neutral emotion). Then, we extract all
the combinations of the estimated vertexes model and perform combinations of these sets
(complete set, set of eyes, eyebrows and mouth).
5. RESULTS
5.1 Bayesian shape Model Estimation Error
In Table 5.1, It can be seen that although the accuracy in the estimation of the points is greater for
images of the training set, the average error is also small for the test images. This is due to a
rigorous procedure in the training and model building in which there were considered to be the
biggest quantity of possible forms to estimate. Moreover, it is noted that although the average
error for images with changes in pose is a bit higher than in the case of frontal images, this
indicates that the accuracy in estimating the model is so high for images that show changes in the
pose of the face. Also, it is of highlighting that the times average of estimation of the model
BSM, they are relatively small which would help in applications on line.
5.2 Distribution on the Relative Error
Figure 4(a) shows the distribution function of the relative error against successful detection rate,
on which it is observed that for a relative error of 0:091 in the case of the adjustment of the right
eye, 0:084 for the left eye and 0:125 for the mouth region in frontal images, the detection rate is
100%, indicating that the accuracy in the BSM model fit is high. In addition to images with pose
variations in BSM is achieved 100% for adjusting the eye region for relative errors to 0:120,
0:123 and 0:13 for the mouth region; being this much lower than the established criterion of 0:25.
It is considered that the criterion Rerr < 0:25 is not adapted to establish a detection like correct
and can be not very suitable when it is desirable to realize facial features detection in images with
minor scale. That's why it is considered to be a successful detection for errors Rerr < 0:15 [18].

Fig. 4. Shape of each emotion type.
Finally a set of fitting images by BSM model is shown in Figure 4(b) evidencing the robustness
of the proposed model.
5.3 Emotion Recognition
Table 2 shows the average success rates in the emotion detection obtained for all sets used in this
analysis using the MUCT database proving the high accuracy of the RMSE metric. Another
important factor in this parsing is the fact that the sets wich provide more evidence on the
emotional estimation state are the sets of B;C (mouth and eyebrow separated) with 95:6% of
average success rate, B;C;O (mouth, eyebrows and eyes separate set) with 92:2%, CO;B
(Eyebrow, Eyes together and Mouth separately) with 91:1%, and where the first of these provides
more discriminant results for the estimation of the emotional detection rate.
Table 2. Successful rate emotion recognition for MUCT database
.
6. CONCLUSION AND FUTURE WORKS
This paper presents a bayesian shape model for emotion recognition using facial expression
analisys. By projecting shape to tangent shape, we have built the model describing the prior
distribution of face shapes and their likelihood. We have developed the BSM algorithm to
uncover the shape parameters and transformation parameters of an arbitrary figure. The analyzed
regions correspond to the eyes, eyebrows and mouth region on which detection and tracking of
features is done using BSM. The results show that the estimation of the points is exact and
complies with the requests for this type of systems. Through quantitative analysis, the robustness

of the BSM model in feature detection is evaluated, and it is maintained in nominal pose for a
range between [-400
to 400
] on X and [-200
to 200
] for Y . The used model and the methodology
of characterization showed efficiency to detect the emotion type in 95:55% of the evaluated
cases, showing a high performance in the sets analysis for the emotion recognition.In addition,
due to high accuracy in detecting and characterizing features proposed to estimate parameters
associated with the emotion type, the designed system has great potential for emotion recognition,
being of great interest in human/computer research systems.
Due to the current problem of facial pose changes in the scene, it would be of great interest to
design a 3D model that will be robust to these variations. Also, Taking advantage of sets analysis
demonstrated in this work, it would be interesting to develop a system that in addition to use the
sets analysis, add an adaptive methodology to estimate the status of each facial action unit to
improve performance of the emotion recognition system.
REFERENCES
[1] Ekman, P.: Emotions Revealed: Recognizing Faces and Feelings to Improve Communication and
Emotional Life. 2nd edn. Owl Books, 175 Fifth Avenue, New York (2007)
[2] Carlos, B., Zhigang, D., Serdar, Y., Murtaza, B., Chul Min, L., Abe, K., Sungbok, L., Ulrich, N.,
Shrikanth, N.: Analysis of emotion recognition using facial expressions, speech and multimodal
information. In: in Sixth International Conference on Multimodal Interfaces ICMI 2004, ACM Press
(2004) 205{211
[3] Song, M., You, M., Li, N., Chen, C.: A robust multimodal approach for emotion recognition.
Neurocomput. 71 (2008) 1913{1920
[4] Ekman, P., Rosenberg, E.: What the Face Reveals: Basic and Applied Studies of Spontaneous Expression
Using the Facial Action Coding System (FACS). Oxford Univ. Press (2005)
[5] Griesser, R.T., Cunningham, D.W., Wallraven, C., B ultho, H.H.: Psychophysical investigation of facial
expressions using computer animated faces. In: Proceedings of the 4th symposium on Applied perception
in graphics and visualization. APGV '07, ACM (2007) 11{18
[6] Cheon, Y., Kim, D.: A natural facial expression recognition using differential aam and knns. In:
Proceedings of the 2008 Tenth IEEE International Symposium on Multimedia. ISM '08, Washington, DC,
USA, IEEE Computer Society (2008) 220{227
[7] Cheon, Y., Kim, D.: Natural facial expression recognition using differential-aam and manifold learning.
Pattern Recogn. 42 (2009) 1340{1350
[8] Ioannou, S., Raouzaiou, A., Tzouvaras, V., Mailis, T., Karpouzis, K., Kollias, S.: Emotion recognition
through facial expression analysis based on a neurofuzzy network. Elsevier 18 (2005) 423{435
[9] Cohen, I., Garg, A., Huang, T.S.: Emotion recognition from facial expressions using multilevel hmm. In: in
In Neural Information Processing Systems. (2000)
[10] Xue, Z., Li, S.Z., Teoh, E.K.: Bayesian shape model for facial feature extraction and recognition. Pattern
Recognition 36 (2004) 2819{2833
[11] Ahlberg, J.: Candide-3 - an updated parameterised face. Technical report, Report No. LiTH-ISY-R-2326,
Dept. of Electrical Engineering, Linkoping University, Sweden (2001)
[12] Cootes, T., Taylor, C., Pt, M.M.: Statistical models of appearance for computer vision (2004)
[13] Matthews, I., Baker, S.: Active appearance models revisited. Int. J. Comput. Vision 60 (2004) 135{164
[14] Gower, J.C.: Generalised Procrustes analysis. Psychometrika 40 (1975) 33{51
[15] Larsen, R.: Functional 2D procrustes shape analysis. In: 14th Scandinavian Conference on Image Analysis.
Volume 3540 of LNCS., Berlin Heidelberg, Springer Verlag (2005) 205{213
[16] Milborrow, S., Morkel, J., Nicolls, F.: The MUCT Landmarked Face Database. Pattern Recognition
Association of South Africa (2010) http://www.milbo.org/muct.
[17] Hassaballah M., I.S.: Eye detection using intensity and appearance information. In: IAPR Conference on
Machine Vision Applications. (2009)
[18] M., H., Ido, S.: Eye detection using intensity and appearance information. (2009) 801809

FACIAL LANDMARKING LOCALIZATION FOR EMOTION RECOGNITION USING BAYESIAN SHAPE MODELS

More Related Content

What's hot (16)

Similar to FACIAL LANDMARKING LOCALIZATION FOR EMOTION RECOGNITION USING BAYESIAN SHAPE MODELS (20)

More from cscpconf (20)

Recently uploaded (20)

FACIAL LANDMARKING LOCALIZATION FOR EMOTION RECOGNITION USING BAYESIAN SHAPE MODELS