Stochastic models for face image analysis

This study continues our work on using stochastic models for image analysis in the context of video indexing. Pseudo-two dimensional Hidden Markov Models (P2DHMM) were shown to be efficient and flexible tools for performing human face localisation in colour images from video sequences. In this context, little constraints can be applied on face images for their identification in view of indexing. Here, we present a technique based on P2DHMM for recovering face orientation in colour images cropped from video sequences. Such a procedure will ease identification by either direct comparison or clustering. Results are presented which show the accuracy of our technique and confirm the capabilities of such stochastic models in the task of video indexing.