Recognising facial expressions using spherical harmonics

Abstract

This paper explores whether facial expressions can be recognised by using the distribution of surface normal directions in the extended Gaussian image (EGI). We work with range images and extract surface normals using a mesh fitting technique. Our representation of the surface normals is based on the co-efficients of spherical harmonics extracted from the EGI. We explore whether the co-efficients can be used to construct shape-spaces that capture variations in facial expression using a number of manifold learning techniques. Based on a comparison of various alternatives, the best results are given by the diffusion map. © 2008 Springer Berlin Heidelberg.