Methodology and Performance Analysis of 3-D Facial Expression Recognition Using Statistical Shape Representation

This paper presents the methodology and performance of a statistical shape representation for automatic facial expression analysis in 3-D. The core of the method uses the statistical shape modelling technique with the deformable model-based surface matching process, which is capable of simulation and interpretation of 3-D human facial expressions. Using the proposed method, a 3-D face is represented by a low-dimensional shape space vector conveying information about face shape. Since the method relies only on the 3-D shape, it is inherently invariant to changes in the background, illumination, and viewing angle, which are the difficulties often suffered in 2-D facial expression analysis. Using 3-D static facial data from the BU-3DFE database as well as the 3-D dynamic facial expression database recently built by the authors in the ADSIP Research Centre, the paper also reports on the performance of the proposed facial expression representation. Furthermore, to demonstrate the effectiveness of the proposed facial expression representation, a comparison is made with human performance by involving a number of human participants to validate the facial expressions in the ADSIP database.