A model of facial behaviour

We wish to model the way in which faces move in video sequences. We represent facial behaviour as a sequence of short actions. Each action is a sample from a statistical model representing the variability in the way it is performed. The ordering of actions is defined using a variable length Markov model. Action models and variable length Markov model are trained from a long (20000 frames) video sequence of a talking face. We propose a novel method of quantitatively evaluating the quality of the synthesis by measuring overlaps of parameter histograms. We apply this method to compare our technique with an alternative model that uses an autoregressive process.