May 15th, 2012 @ 2:30pm Jean Ponce (Ecole Normale Superieure)

Graph Matching and Object Categorization

Abstract: Geometrically consistent correspondences between local image features are a key element of image retrieval and specific object detection technology, but they are seldom used in object categorization, where histograms of quantized features (“bags of features”), which discard all spatial information, and their variants have been dominant. We propose to revisit feature matching and represent each image by a graph whose nodes correspond to a dense set of regions, and edges reflect the underlying image grid structure, acting as springs to guarantee the geometric consistency of nearby regions during matching. A fast approximate algorithm for matching the graphs associated with two images is presented. This algorithm is used to construct a similarity measure (“kernel”) appropriate for image categorization using classifiers such as support vector machines. Experiments with standard benchmarks (Caltech 101, Caltech 256, and Scenes datasets) demonstrate performances that match or exceed the state of the art for methods using a single type of image features. I will conclude with a brief discussion of our current research on deformable part models for object detection.

Joint work with Olivier Duchenne and Armand Joulin.

seminars/seminaritems/2012-05-15.txt · Last modified: 2012/05/15 12:00 by silberman