Figure 2.5 shows how our method [6,5] compares to the literature. By focusing on the last three branches corresponding to the variable length Markov model branch, a progression can be seen. Galata et al. first used a variable length Markov model directly on the clustered prototypes of their system. However, in order to model longer-term behaviours, they had to introduce an atomic behaviour level to their model (which consists of a sequence of a few prototypes). Our approach is to model atomic behaviours without the drawback of using prototypes. Indeed, a prototype raises problems when we want to synthesise new behaviours. Each time we generate a prototype vector, the synthesis corresponds exactly to the prototype. So the same frame is generated many times in the output video sequence. This also highlights the need of a large number of prototypes to model a trajectory in the appearance parameter space.
Our alternative is inspired by the work of Walter et al. . They represent trajectories of hands by atomic trajectories and model those using a principle component analysis. Our pathlet group model is the transposition of this model to the appearance parameter space without, however, using the same algorithms to build the model.