Integration of local and global geometrical cues for 3D face recognition

We present a unified feature representation of 2.5D pointclouds and apply it to face recognition. The representation integrates local and global geometrical cues in a single compact representation which makes matching a probe to a large database computationally efficient. The global cues provide geometrical coherence for the local cues resulting in better descriptiveness of the unified representation. Multiple rank-0 tensors (scalar features) are computed at each point from its local neighborhood and from the global structure of the 2.5D pointcloud, forming multiple rank-0 tensor fields. The pointcloud is then represented by the multiple rank-0 tensor fields which are invariant to rigid transformations. Each local tensor field is integrated with every global field in a 2D histogram which is indexed by a local field in one dimension and a global field in the other dimension. Finally, PCA coefficients of the 2D histograms are concatenated into a single feature vector. The representation was tested on FRGC V2.0 data set and achieved 93.78% identification and 95.37% verification rate at 0.1% FAR.