Learning HMM action manifolds

Fabio Cuzzolin; Oxford-Brookes University

26th May, 2010 at 1.00 pm - EM1.27

Joint UoE/HWU videoconference seminar

In action recognition it is sometimes useful, rather than to extract some spatio-temporal features from the volumes representing actions,to encode action dynamics by means of dynamical systems. Hidden Markov models (HMMs) are a popular choice in that respect: actions can be then classified by measuring distances in the space of HMMs. However, using an arbitrary distance to classify dynamical models does not necessarily produce good classification results. In this talk we outline a framework based on pullback metrics which allows instead, given a training set of models, to learn an optimal pullback distance between HMMs tuned for that specific training set, starting from any base distance/divergence. We show results on the KTH and Weizmann dataset illustrating the gain in classification performance this method delivers.