Synthesis and recognition of facial expressions in virtual 3D views

The human face exhibits complex and rich changes that are both unpredictable and varying in time. In this paper we present a novel method for synthesising and recognition of facial expression changes at extreme 3D views, based on images at near frontal views. Given a sequence of images of facial expressions at near frontal views, we automatically generate virtual expressions at extreme 3D views with corresponding semantic labelling of the expressions. This is accomplished by two components: a shape component where modelling of the shape changes is accomplished through the use of a mixture of probabilistic PCA (MPPCA) and a texture component where modelling of the semantic changes is performed through auto-clustering of facial expression subspaces in the MPPCA feature space.