A Survey Paper on Learning Pullback HMM Distance for Recognition of Action

Recent work in action recognition has exposed the limitations of features extracted from spatiotemporal video volumes. Whereas, encoding the actions dynamics using generative dynamical models has a number of attractive features, in this respect Hidden Markov models (HMMs) is a popular choice. A general framework based on pullback metrics for learning distance functions of a given training set of labeled videos has been generated, The optimal distance function is selected among a family of pullback ones, which is generated by a parameterized automorphism of the space models. An experimental result shows that how pullback learning greatly improves action recognition performances with respect to base distances.