1. Multi-Eigenspace Learning for Video-Base Face Recognition Liang Liu 1, Yunhong Wang 2, Tieniu Tan 1 1 National Laboratory of Pattern Recognition, Institute of Automation, CAS, Beijing, China 2 School of Computer Science and Engineering, Beihang University, Beijing, China 模式识别国家重点实验室 中国科学院自动化研究所，北京， 100080 National Laboratory of Pattern Recognition Institute of Automation, CAS, Beijing, P. R. China ICB 2007 The 2 nd International Conference on Biometrics In this paper, we propose a novel online learning method called Multi-Eigenspace Learning which can learn appearance models incrementally from a given video stream. For each subject, we try to learn a few eigenspace models using IPCA (Incremental Principal Component Analysis). Then, these learnt eigenspace models are used for video-based face recognition. Experimental results show that the proposed method can achieve high recognition rate. Multi-Eigenspace Learning  Eigenspace model description For the training face video stream of each subject, we aim to construct up to K eigenspaces to approximately represent the appearance manifold of that subject. There are four parameters for each eigenspace, namely where is the center of the eigenspace. U is a matrix whose columns are orthonormal bases of the eigenspace, namely eigenvectors. is a diagonal matrix. Elements along the diagonal are variances for each principal axis, namely eigenvalues. They are arranged in descending order. N is the number of samples to construct the eigenspace.  The algorithm of Multi-Eigenspace Learning. Experimental Results Conclusions  Multi-Eigenspace Learning is proposed to learn face appearance manifold online without a pre-trained model.  Experimental results show that the proposed method gives better performance than that given by some other online learning methods.  Data set description  There are 36 video sequences from 36 persons respectively.  The number of frames ranges from 236 to 1270.  Use the first half of each sequence for online learning.  Use the second half of each sequence for recognition. Fig. 2. Typical samples of the videos. Fig. 1. The flowchart of Multi-Eigenspace Learning. MethodProbabilistic Manifold Multi-Eigenspace Learning Recognition rates92.4%96.2% Online learning34.3s7.1s Pre-training77.3s- K Method 6789 EMS + transition90.091.996.491.4 Multi-Eigenspace Learning 97.596.798.196.4  Experiment 1 We compare the proposed algorithm with our previous work called EMS + transition, which is also an online learning method. This method try to learn K eigenspace models for all subjects, with consideration of transition matrix. Abstract Table 1. Average recognition rates (%) of different methods when choosing K = 6, 7, 8, 9 respectively.  Experiment 2 We also implemented the Probabilistic Manifold online learning algorithm for comparison. In this method, an appearance model is incrementally learnt online using a pre-trained generic model and successive frames from the video. Table 2. Comparison of the Probabilistic Manifold algorithm and our proposed algorithm. For Multi-Eigenspace Learning, we choose K = 8.