LEARNING EFFICIENT CODES FOR 3D FACE RECOGNITION

Face representation based on the Visual Codebook becomes popular because of its excellent recognition performance, in which the critical problem is how to learn the most efficient codes to represent the facial characteristics. In this paper, we introduce the Quadtree clustering algorithm to learn the facial-codes to boost 3D face recognition performance. The merits of Quadtree clustering come from: (1) It is robust to data noises; (2) It can adaptively assign clustering centers according to the density of data distribution. We make a comparison between Quadtree and some widely used clustering methods, such as G-means, K-means, Normalized-cut and Mean-shift. Experimental results show that using the facial-codes learned by Quadtree clustering gives the best performance for 3D face recognition.