3D model based expression tracking in intrinsic expression space

A method of learning the intrinsic facial expression space for expression tracking is proposed. First, a partial 3D face model is constructed from a trinocular image and the expression space is parameterized using MPEG4 FAP. Then an algorithm of learning the intrinsic expression space from the parameterized FAP space is derived. The resulted intrinsic expression space reduces even to 5 dimensions. We will show that the obtained expression space is superior to the space obtained by PCA. Then the dynamical model is derived and trained on this intrinsic expression space. Finally, the learned tracker is developed in a particle-filter-style tracking framework. Experiments on both synthetic and real videos show that the learned tracker performs stably over a long sequence and the results are encouraging.