Deformable Part Models for the PASCAL Visual Object Detection Challenge
Speaker: David McAllester , Toyota Technological Institute at Chicago
We consider the problem of detecting and locating objects of a given class, perhaps people or cars, in static images. There are various approaches to this problem such as bag-of-feature models, fixed filters, and Viola-Jones detectors. Here we focus on deformable part models in which part location is related to, but not determined by, the overall or "average" location of the detected object. Deformable part models naturally generalize to grammar models in which parts are composed of deformable subparts and so on. Although deformable part models are very natural, it has been difficult to establish their value in practice. Using a deformable part model we have achieved a 70% improvement in average precision over the winning system (Dalal and Triggs) in the 2006 PASCAL person detection challenge. We are submitting detection results to the 2008 PASCAL challenge in all 20 categories. This talk will discuss the details of our detector and possible future directions.