This thesis has described a practical approach to modelling the behaviour of faces by learning from video sequences. The assumption is that people repeat fragments of behaviour, and that it is possible to learn both the general form of the fragments and the order in which they tend to happen.
The fragments are represented using models of pathlets in the parameter space of a facial appearance model. A VLMM is trained to generate the order of transitions between pathlets over both short and long timescales.