Larry S. Davis’ talk will cover recent work on the acquisition and application of contextual models for object and action recognition. For object recognition, I will discuss how fine grained contextual models that capture geometric and topological relations between objects in pictures can lead to significantly improved object recognition capabilities. Then, for action recognition, I will discuss the use of storylines, structured as AND/OR graphs, as contextual models to improve human action recognition in videos of American baseball plays. The research was done by my former Ph. D. student, Abhinav Gupta; the work on action recognition was performed jointly with Prof. Jianbo Shi of the University of Pennsylvania.
Talk slides in pdf [~1,8MB]http://www.iit.demokritos.gr/docs/seminars/iptseminars-2010-05-davis.pdf