Discriminative Hough-Voting for Object Detection with Parts

many detection models rely on sliding-window methods to search all possible candidates and then localize true instances by linear classifiers. Such exhaustive search involves massive computation. Hough voting scheme provides an alternative way to localize objects. Typical voting-based approaches, casting independent votes for a hypothesis, ignore the mutual relevance of features. The weights of the features for voting are learnt in a simple way. These two weaknesses limit the detection performance of the voting scheme. This paper introduces a novel voting-based model. We group the model features into parts. The features in one part are dependent and can cast consistent votes for a given hypothesis. For a given hypothesis we introduce an overall score function whose parameters can be optimized in a discriminative way. We apply a latent learning framework to deal with part-level weak supervision. The experiments evaluate the proposed model on two standard datasets. We demonstrate significant improvements in detection performance comparing the start-of-the-art detection models.