In this chapter we have explained how to generate new video sequences of facial behaviour using our model. First, new sequences of pathlet states are sampled from the VLMM. Then pathlets are sampled from each corresponding pathlet model. The pathlets are concatenated together with an optional smoothing performed by using the residual component of the pathlet model. This gives us a sequence of appearance parameters. Finally, the new video sequence is generated by synthesising each frame in turn using the appearance model.
Using the linear residual model avoids jumps in generated trajectories. But, as we can see by comparing figures 7.7 and 7.10, it also seems to generate less likely trajectories. We need to be able to assess how good the generated trajectories are. In the next chapter, we develop a measure of comparison between generated trajectories in the appearance parameter space and use it to assess our model by comparing it to an alternative. A psychophysical experiment is also used to assess how people judge the generated video sequences.