We have presented a generative model of visual facial behaviour that is based on the assumption that people repeat facial expressions over time. It has been shown that this model is able to reproduce simple behaviours.
A novel decomposition of sequences into visual units allows a higher level of abstraction than other models that are based on sequences of frames. The use of variable length Markov model allows us to efficiently encode history of visual units in long sequences. Generation of new sequences can be done thanks to the combination of the generative features of both the variable length Markov model and the appearance model. Furthermore, the use of variable length Markov model for learning helps us to better understand what is going on with the model. There are no hidden variables or hidden states.
However, the transition between visual units has to be smoothed. For the time being, some jumps are perceptible in the generated sequences. Another improvement that can be done is to modify the model so that it handles outliers and timings in a better way.
In our future work, we plan to use two active appearance models with the same framework in order to model interactions between two persons speaking together in an interview type scenario in a similar manner to [6].