Normalization for Unconstrained Pose-Invariant 3D Face Recognition

Xun Gong, Jun Luo, Zehua Fu

Abstract: This paper presents a framework for 3D face representation, including pose and depth image normalization. Different than a 2D image, a 3D face itself contains sufficient discriminant information. We propose to map the original 3D coordinates to a depth image using a specific resolution, hence, we can remain the original information in 3D space. 1) Posture correction, we propose 2 simple but effective methods to standardize a face model that is appropriate to handle in following steps; 2) create depth image which remain original measurement information. Tests on a large 3D face dataset containing 2700 3D faces from 450 subjects show that, the proposed normalization provides higher recognition accuracies over other representations.