Analysis of Representation and Feature Extraction Schemes for 3D Face Recognition

In this paper, we design various 3D face recognizers that are based on different representation and feature extraction schemes. For each scheme, an extensive range of possible approaches are realized, and we also propose the use of novel methods such as principal curvature directions, several subspace techniques e.g., Discrete Fourier/Cosine Transforms and Non-negative Matrix Factorization (NMF). Identification experiments performed on the largest available 3D face database (Face Recognition Grand Challenge) reveals that 1) representation schemes play more important role than the feature extraction methods, and 2) principal curvature directions outperform other shape-based descriptors, and 3) given enough training measurements, subspace methods such as NMF, and Independent Component Analysis could also perform well.