Three-dimensional face pose estimation based on novel nonlinear discriminant representation

We investigate the appearance manifold of different face poses using manifold learning. The pose estimation problem is, however, exacerbated by changes in illumination, spatial scale, etc. In addition, manifold learning has some disadvantages. First, the discriminant ability of the low-dimensional subspaces obtained by manifold learning often is lower than traditional dimesionality reduction approaches. Second, manifold learning methods fail to remove the redundancy, such as high-order correlation, among original feature vectors. In this work, we propose a novel approach to address these problems. First, face images are transformed by Gabor filters to obtain a set of overcompleted feature vectors, which can remove intrinsic redundancies within images and provide orientation-selective properties to enhance differences among face poses as well. Second, supervised locality preserving projections (SLPPs) are proposed to reduce dimensionality and obtain the low-dimensional subspace, which has the ability to maximize between-class distance and minimize within-class distance. Finally, the support vector machine (SVM) classifier is applied to estimate face poses. The experimental results show that the proposed approach is effective and efficient. (c) 2006 Society of Photo-Optical Instrumentation Engineers.