Discriminating 3D faces by statistics of depth differences

In this paper, we propose an efficient 3D face recognition method based on statistics of range image differences. Each pixel value of range image represents normalized depth value of corresponding point on facial surface, and so depth differences between two range images' pixels of the same position on face can straightforwardly describe the differences between two faces' structures. Here, we propose to use histogram proportion of depth differences to discriminate intra and inter personal differences for 3D face recognition. Depth differences are computed from a neighbor district instead of direct subtraction to avoid the impact of non-precise registration. Furthermore, three schemes are proposed to combine the local rigid region(nose) and holistic face to over-come expression variation for robust recognition. Promising experimental results are achieved on the 3D dataset of FRGC2.0, which is the most challenging 3D database so far.