Modeling and Recognition of Dynamic Visual Processes

Prof. Stefano Soatto, Department of Computer Science, UCLA

When watching the sea at sunset or foliage moving in wind, one can show that the sensory data gathered -- i.e. images -- are not sufficient to estimate a physically correct representation of the scene. Nevertheless, we have no trouble recognizing sea waves or trees when we see them. In this talk I will discuss how to identify a dynamical model of a visual process in such a way that the model -- albeit not physically correct -- can be used for recognition. Recognition entails characterizing evolving probability distributions on geometric spaces such as Stiefel manifolds; identification can be seen as a dimensionality reduction step where one wants to estimate the portion of a (stochastically balanced) state space that maximizes mutual information.