Augmenting deformable part models with irregular-shaped object patches

The performance of part-based object detectors generally degrades for highly flexible objects. The limited topological structure of models and pre-specified part shapes are two main factors preventing these detectors from fully capturing large deformations. To better capture the deformations, we propose a novel approach to integrate the detections from a family of part-based detectors with patches of objects that have irregular shape. This integration is formulated as MAP inference in a Conditional Random Field (CRF). The energy function defined over the CRF takes into account the information provided by an object patch classifier and the object detector, and the goal is to augment the partial detections with missing patches, and also to refine the detections that include background clutter. The proposed method is evaluated on the object detection task of PASCAL VOC. Our experimental results show significant improvement over a base part-based detector (which is among the current state-of-the-art methods) especially for the deformable object classes.