Object and Place Recognition: Appearance vs Geometry

Stay connected



Share on facebook
Share on twitter
Share on linkedin

CIS Colloquium, Feb 24, 2010, 11:00AM – 12:00PM, Wachman 447

Object and Place Recognition: Appearance vs Geometry

Kostas Daniilidis, University of Pennsylvania

A common approach for recognizing scenes is comparing constellations and histograms of appearance features. These image matching measures are limited with respect to illumination and pose variations. a weakness mitigated by the abundance of training data. We take here an orthogonal approach based on shape and geometry. We built a correspondence free but geometry preserving system in 2006, by formulating the image matching problem as a filtering problem in the space of possible rigid geometries. To relax the requirement of rigid structure of the scene we introduced in 2007 the notion of co-saliency in order to simultaneously segment and correspond image segments. When retrieving objects we realized that most of the information in an object lies in its silhouette, while appearance of the interior of the object depends on the illumination, viewpoint, and the particular object instance. We applied the co-saliency framework to extract object silhouettes from video that are matched with a large set of silhouettes representing a dense sampling of all possible views of the models. We will end this talk with a new approach to the extraction of topological maps from panoramic video, relying on a soft alignmnent of scenes for loop closure.

Joint work with Ameesh Makadia, Alexander Toshev, Roy Anati, and Jianbo Shi.

Kostas Daniilidis is Professor of Computer and Information Science at the University of Pennsylvania where he has been faculty since 1998. He is the director of the interdisciplinary GRASP laboratory, which hosts 15 faculty and more than 70 doctoral students and postdocs, as well as the director of the Master’s in Robotics program. He obtained his undergraduate degree in Electrical Engineering from the National Technical University of Athens, 1986, and his PhD (Dr.rer.nat.) in Computer Science from the University of Karlsruhe, 1992. His research interests are on visual motion and navigation, image matching, stereo vision, object and place recognition, and camera design. He was Associate Editor of IEEE Transactions on Pattern Analysis and Machine Intelligence from 2003 to 2007. He co-chaired with Pollefeys 3DPVT 2006, and he is Program co-chair of ECCV 2010.