An evaluation of local interest regions for non-rigid object class recognition

Non-rigid object class recognition is a challenging computer vision problem. Using descriptors extracted from local interest regions has important advantages like robustness to occlusion and photometric effects. In this work we compare different local interest region detectors for non-rigid object class recognition through the success-rate of a Generalized Hough Transform based recognition system and a database of 29 non-rigid object classes. The results of the experiments show that the Edge–Laplace (Mikolajczyk, Leibe, & Schiele, 2006; Mikolajczyk, Zisserman, & Schmid, 2003) interest region detector leads. We also evaluate interest regions based on a novel discriminancy measure we introduce. This measure compares success-rates of detectors to success-rates of our novel random region generator, ExpRand. By this respect, ExpRand attain success-rate on par with best detector, and is more discriminant than most detectors.