Selecting models from videos for appearance-based face recognition

We propose an unsupervised approach to select representative face samples (models) from raw videos and build an appearance-based face recognition system. The approach is based on representing the face manifold in a low-dimensional space using the locally linear embedding (LLE) algorithm and then performing K-means clustering. We define the face models as the cluster centers. Our strategy is motivated by the efficiency of LLE to recover meaningful low-dimensional structures hidden in complex and high dimensional data such as face images. Two other well-known unsupervised learning algorithms (Isomap and SOM) are also considered. We compare and assess the efficiency of these different schemes on the CMU MoBo database which contains 96 face sequences of 24 subjects. The results clearly show significant performance enhancements over traditional methods such as the PCA-based one.