Facial Expressions Recognition by Animated Motion of Candide 3D Model

Authors:
Xin Chang, prof. Wladyslaw Skarbek

Abstract:
This paper refers to facial multi-expression recognition as an element of human-computer interface. The discriminative features are extracted from each camera frame using Candide 3D model for human head. Namely, for the selected parts, including mouth, nose, eyes and eye brows areas, shape deformation units and animated motion units are specified. They are corrected versions of the units defined originally for Candide-3 model. By nonlinear least squared LM method scalar parameters for affine motion, the shape deformation, and the animated motion are identified. The error function is based on the orthographic projection of Candide 3D points corresponding to on-line detected 68 facial landmarks. The feature vectors are comprised of less than 10 coefficients for controlling the animated motion of Candide 3D model, only. The multiple expression classifier is based on the Structural Support Vector Machine linear model. The SSVM model is trained by few hundreds images for each of four expression classes: idle, smile, anger, and surprise. On-line experiments with web camera confirm high correlation of the proposed class with the subjective impression of face expression. Moreover, comparing to FP68/SSVM classifier our new proposal outperforms it by 3X15 rule: its success rate is about 15% higher while having more than 15 times shorter feature vector, and therefore 15 times shorter recognition time.

Key words:
Facial landmarks, multiple expression classifier, Candide 3D model, shape deformation units, animated motion unit, Structural Support Vector Machine.

session name:
"Algorithms for Image Processing and Analysis" chaired by prof. Władysław Skarbek