Object Detection with Heuristic Coarse-to-Fine Search

Abstract We consider the task of localizing and labeling instances of a generic object class within real-world images. Our focus is on a generalized class of pictorial structure models that are defined in terms of visual gram- mars. In particular, we address the challenging problem of performing detection efficiently even as model complexity grows within this class. Our proposed solution is a blend of heuristic best-first search and a coarse-to-fine detection process. This paper demonstrates that our algorithm can be successfully applied to two special cases of visual grammars: multiscale star models and mixtures of multiscale star models. We show that for problems where the desired output is the local optima of a thresholded function, best-first search gives additional pruning power to coarse-to-fine processes. Unfortunately, admissible heuristics that also provide good best-first search behavior can be difficult or impossible to find in practice. To resolve this deficiency, we provide theoretical results demonstrating that inadmissible heuristics can be used to increase detection speed while only slightly increasing the likelihood of suffering mistakes. The theoretical results are bolstered by strong experimental evidence obtained by applying inadmissible heuristic coarse-to-fine detection to our object recognition system during both training and testing. We increase testing speed by a factor of 2-3 for some classes while maintaining comparable,average precision scores on the challenging PASCAL 2007 dataset. Ultimately we expect to see even more,significant speed gains when,we explore more complex,grammar,models in future work. Acknowledgements I would like to thank everyone who has been patient with me while I’ve absorbed myself in this paper. I owe