In this chapter, we have developed a measure of comparison of facial behaviour models to assess quantitatively our model of behaviour. This measure is based on a comparison of trajectories in the appearance parameter space. A Bhattacharyya overlap is used to compare histograms of densities of points in the appearance parameter space. This measure shows that our model outperforms an alternative on structured videos sequences while maintaining a good performance on unstructured video sequences.
A psychophysical experiment was performed to assess qualitatively our model of facial behaviour. It showed that our model is significantly better than the model based on an autoregressive process when it models structured video sequences. We do not have any significant results for unstructured video sequences. However, the figures suggest that our model does not perform worse than the autoregressive process in that case.
Agreements between the two methods of evaluation suggest that our quantitative measure of similarity between two facial behaviour models follows the natural choice of people.
Furthermore, people seem to have trouble to distinguish between our model with and without the linear model of residuals. However, in structured video sequences, the experiment shows that our model is better than the autoregressive process even if we do not use the residual model.