Non-rectangular Part Discovery for Object Detection

The deformable part-based model (DPM) is one of the most influential models for generic object detection and many efforts have been made to improve the model. Despite previous work, the problem of how to identify discriminative parts for DPM still remains largely unexplored. Most DPM based methods rely on a fixed number of parts of rectangular shapes, which may not be optimal for some object categories. In this paper, we present a novel approach to discover parts which can be non-rectangular by exploiting object structures. Instead of performing greedy part search as in DPM, our part discovery approach is carried out by first solving a K-way normalized cuts problem and then applying local refinement. Generally, the parts obtained by the proposed approach can better fit the object structures. We demonstrate the effectiveness of our approach on PASCAL VOC2007 and VOC2010 datasets.