A study on region-based recognition of 3D faces with expression variations

In this paper, we propose the application of masks as a means to mitigate expression-distortions on 3D faces and to enhance their recognition performance. Masking becomes necessary to de-emphasize the face regions that deform under expression. We have conducted experiments with various masks, namely, ellipse-shaped binary masks, Gaussian, super-Gaussian and raised-cosine masks. The design issues of the masks, such as the mask size, the centre, the support region, the decay rate of the tails, etc. are studied and adjusted with respect to their recognition performances. We show first that warping the depth values of corresponding face points onto the same spatial coordinates while obtaining the 2D depth images is beneficial, and second, that proper masking can add several percentage points to the recognition performance.