Area matching based on belief propagation with applications to face modeling

Abstract

In the past few years several systems for face recognition and identification based on the analysis of 3D data have been proposed. In order for such systems to be of practical use, the 3D data extraction process is expected to be fast and reliable. In this paper we propose a general approach to multi-view 3D surface reconstruction, specialized for the specific problem of face modeling. Our method determines correspondences between surface patches on different views through a modeling of disparity maps based on Markov random fields (MRFs). In order to reduce the occurrence of outliers the MRF-based modeling is bound to satisfy the epipolar constraint. We apply the belief propagation algorithm to the MRF model in order to perform a maximum-a-posteriori estimation of such correspondences.