Synthesizing 3D face shapes using tensor-based multivariate statistical discriminant methods

Abstract

We have implemented methods to reconstruct and model 3D face shapes and to synthesize facial expressions from a set of real human 3D face surface maps. The method employed tensor-based statistical shape modelling and statistical discriminant modelling methods. In the statistical shape modelling approach, new face shapes are created by moving the surface points along the appropriate expressive direction in the training set space. In the statistical discriminant model, new face shapes, such as facial expressions, can be synthesized by moving the surface points along the most discriminant direction found from the classes of expressions in the training set. The advantage of the tensor-based statistical discriminant analysis method is that face shapes of varying degrees can be generated from a small number of examples available in the 3D face shape datasets. The results of the reconstructions and synthesis of three-dimensional faces are illustrated in the paper. © 2011 Springer-Verlag.