Generating new sequences

A new video sequence of faces can be generated from the model as follows. First, given a history of generated sub-trajectory groups, one can find the longest memory encoded in the variable length Markov model tree. Thus the probability of generating a new group can be read directly from the tree, if it is encoded in the tree. The sequences not encoded in the tree have small probabilities, that we can approximate by an uniform distribution. After having fetched the probabilities of generation of each sub-trajectory groups, we sample from this set of probabilities to generate the next sub-trajectory group. We then generate new parameters by sampling from a Gaussian distribution. The new sub-trajectory can then be generated as described in section 3.4.1. A linear model is chosen for the residuals $R_{s,\delta}$ so that the beginning of the generated sub-trajectory matches the end of the previous generated sub-trajectory to avoid perceptible jumps in the generated video. All the sub-trajectories generated are then concatenated. This gives a sequence of appearance parameters. The video sequence can then be generated by synthesising those parameters into a face as described in section 3.1.

An example of generated trajectory is given by Figure 4(c). The three dimensions represent the three first modes of variation of the active appearance model used. Another example of generated video can be seen on Figure 3. The original video sequence is shown on Figure 2.

Figure 2: Frames taken every 4 seconds from the original long video sequence.

Figure 3: Frames taken every 4 seconds from the video sequence generated with our model.

Figure 4: Figure 4(a) represents points on the trajectory that correspond to the original video sequence. Figure 4(b) represents points on the trajectory generated by our model if the residuals used are set to zero. Figure 4(c) represents the equivalent with a linear model for residuals. Figure 4(d) represents points on the trajectory generated by Campbell's model.

[Training sequence of a face gesturing "no"] [Generated sequence with our model without using the residuals] [Generated sequence with our model] [Generated sequence with Campbell's model]