Expressive Maps for 3D Facial Expression Recognition

Abstract

We present a semi-automatic 3D Facial Expression Recognition system based on geometric facial information. In this approach, the 3D facial meshes are first fitted to an Annotated Face Model (AFM). Then, the Expressive Maps are computed, which indicate the parts of the face that are most expressive according to a particular geometric feature (e.g., vertex coordinates, normals, and local curvature). The Expressive Maps provide a way to analyze the geometric features in terms of their discriminative information and their distribution along the face and allow the reduction of the dimensionality of the input space to 2:5% of the original size. Using the selected features a simple linear classifier was trained and yielded a very competitive average recognition rate of 90:4% when evaluated using ten-fold cross validation on the publicly available BU-3DFE database.