A Systematic Comparison of Depth Map Representations for Face Recognition

Nowadays, we are witnessing the wide diffusion of active depth sensors. However, their different building technologies and the small-scale single-sensor datasets negatively affect the generalization capabilities and performance of deep learning approaches based on depth data, especially when used for recognition purposes. In this paper, we present a systematic comparison on the use of depth data for the deep face recognition task, focusing the analysis on different data representations, pre-processing steps, and normalization techniques. Depth and normal images, voxels, and point clouds are computed from depth maps and tested with several well-known deep architectures. Extensive intra- and cross-dataset experiments, performed on four public databases, suggest that representations and methods based on normal images and point clouds perform and generalize better than other 2D and 3D alternatives. Moreover, we propose an extremely challenging dataset, namely MultiSFace, to specifically analyze the influence of the depth map quality and the acquisition distance on the face recognition accuracy.