From still image to video-based face recognition: an experimental analysis

In this work, we analyze the effects of face sequence length and image quality on the performance of video-based face recognition systems which use a spatio-temporal representation instead of a still image-based one. We experiment with two different databases and consider the temporal hidden Markov model as a baseline method for the spatio-temporal representation and PCA and LDA for the image-based one. We show that the face sequence length affects the joint spatio-temporal representation more than the static-image-based methods. On the other hand, the experiments indicate that static image-based systems are more sensitive to image quality than their spatio-temporal representation-based counterpart. The second major contribution in this work is the use of an efficient method for extracting the representative frames (exemplars) from raw video. We build an appearance-based face recognition system which uses the probabilistic voting strategy to assess the efficiency of our approach.