Random Features for Large-Scale Kernel Machines

Abstract

In this paper, we contributed a stereo face recognition formulation which combines appearance and disparity/depth at feature level. We showed that the present-day passive stereovision in combination with 2D appearance images can match up to other methods which rely on active depth data. A Reduced Multivariate Polynomial Model was adopted to fuse the appearance and disparity images. RMPM is extended so that the problem of newuser registration can be overcome. We evaluated the performance of such fusion on XM2VTS face database. The evaluation results, which included results from appearance alone, depth alone and fusion of them respectively, using XM2VTS database, showed improvement of recognition rate from combining 3D information and 2D information. The performance using fused depth and appearance was found to be the best among the three tests. Furthermore, we implemented the algorithm on a real-time stereo vision system where near-frontal views were selected from stereo sequence for recognition. The evaluation results, which included results from appearance alone, depth alone and fusion of them respectively, using a database collected by the stereo vision system also showed improvement of the recognition rate by combining 3D information and 2D information. The RMPM can yield comparable results with SVM while the computation load of the RMPM is much lower than SVM.