Smile detectors correlation

Authors: Kivanc Yuksel, Xin Chang, Wladyslaw Skarbek

Face expression recognition plays important role in many application based on RGB camera image. In this paper we develop a novel smile recognition algorithm based on extraction of 68 facial salient points (fp68) using ensemble of regression trees. The fp68 sets are 2D discrete shapes which contain enough geometric information to be used for detection of smiling events. We train the smile detector using the Structural Support Vector Machine model (SSVM) getting two hyperplanes contrary to the single plane of the traditional SVM. We observed visually that such geometric detector strongly depends on mouth opening area. We support this observation by strict statistical data analysis. To this goal two Bayesian detectors were developed and compared with SSVM detector. The first uses the mouth area in 2D image, while the second refers to the mouth area in 3D animated face model. The 3D modeling is based on Candide-3 model and it is performed in real time along with three smile detectors and statistics collectors. The mouth area/Bayes detectors exhibit high correlation with fp68/SSVM detector in a range [0.8,1.0] what depends mainly on light conditions and individual features. The correlation for 3D area is slightly higher than for 2D area. The statistical dependences are verified using Anderson, Darling and chi2 null hypotheses tests defined for various setups of light conditions (30=15 * 2 lab seats combined with sunshine and electric light) and individuals (15 students). The null hypothesis that the data have the same statistical distribution for all setups is strongly rejected. The same refers to hypotheses on independence of detectors correlation with respect to individuals and light conditions.

keywords:facial salient points, smile detection, structural support vector machine, Bayes classifier, Candide 3D model, statistical analysis