Monday November 24 4:00 - 4:50 PM Kelley 1001
Sinisa Todorovic Assistant Professor School of EECS Oregon State University WHAT MAKES GOOD IMAGE GRAMMAR? Can we define a general purpose image grammar that would meet diverse needs of image understanding? How do we evaluate it? In this talk, we present partial answers to these and related questions. Hierarchical models, semantic contexts, taxonomy of visual categories, visual event ontology, stochastic graph matching, and bottom-up/top-down inference are popular research topics in computer vision. They can be viewed as different aspects of stochastic image grammars. The virtue of image grammars lies in their expressive power to represent an exponentially large number of object spatiotemporal configurations by using a relatively much smaller vocabulary and a few compositional rules. In addition to objects, in fact, various semantic contexts can be associated with all levels of hierarchical descriptions in grammars, which allows for rich image interpretations. We will examine a general purpose image grammar, developed during my postdoctoral work at UIUC in collaboration with Prof. Narendra Ahuja. The grammar uses syntax called Connected Segmentation Tree (CST), defined in terms of image regions, or segments. It captures the recursive embedding of all regions, their geometric and photometric properties, and their spatial layout. The CST is invariant to changes in imaging conditions (e.g., lighting, scale, orientation), and facilitates inference of semantics. Specifically, we will demonstrate that the CST allows efficient inference of grammar rules for representing a large set of object categories, under varying levels of supervision, including the complete absence of supervision. The grammar captures the recursive definitions of categories in terms of their (simpler) subcategories. Efficiency is achieved by allowing subcategories to be shared by multiple parent categories. The general nature of the proposed image grammar allows addressing other vision problems, including that of identifying and segmenting stochastically repeating parts of visual textures, commonly called texture elements, even when they partially occlude one another. This work has produced a number of fundamental contributions in vision, including: (1) The first operative definition of an object category; and (2) The first approach to show that human supervision is not necessary to communicate the nature of categories to a computer. Biography: Dr. Sinisa Todorovic received his B.S. degree in electrical engineering at the University of Belgrade, Serbia, in 1994, and earned his M.S. and Ph.D. degrees in electrical and computer engineering at the University of Florida, in 2002, and 2005, respectively. He was Postdoctoral Research Associate in the Beckman Institute at the University of Illinois Urbana-Champaign, between 2005-2008, where he collaborated with Prof. Narendra Ahuja. Currently, Dr. Todorovic is Assistant Professor in the School of Electrical Engineering and Computer Science at the Oregon State University. His research mostly focuses on computer vision and machine learning problems, including image segmentation, object/scene/activity recognition, and texture analysis/synthesis. He has published more than 30 journal and refereed conference papers. For a paper published in IEEE Transactions on Vehicular Technology, he was awarded 2004 Jack Neubauer Best Paper Award by the IEEE Vehicular Technology Society.
_______________________________________________ Colloquium mailing list [email protected] https://secure.engr.oregonstate.edu/mailman/listinfo/colloquium
