Local Reconstruction Model for Non-person Specific Facial Expression

This paper presents a non-person specific appearance model for modeling nonlinear facial expression variations. A combined express ion-and-texture similarity measure (ETSM) is proposed to reliably find the nearest neighborhoods of the novel image. In the ETSM, we consider both shape and texture differences. The shape is coded using MPEG-4 facial animation parameters (FAP), which can provide a compact and non-person specific representation of facial expression. We optimize the object function using gradient descent search in the convex hull of the example expression and texture vectors. Preliminary experimental results show that our model significantly improves the performance of the fitting over a conventional AAM