Object Proposal via Partial Re-ranking

Object proposals are an ensemble of bounding boxes with high potential to contain objects. Usually, the ranking models are utilized in order to provide a manageable number of candidate boxes. To obtain the rank for each candidate, prior ranking models generally compare each pair of candidates. However, one may be interested in only the top-$k$ candidates rather than all ones. Thus, in this paper, we propose a new ranking model for object proposals, which aims to produce a reliable estimation for only the top-$k$ candidates. To this end, we compute the IoU for each candidate and split the candidates into two subsets consisting of the top-$k$ candidates and the others respectively. Partial ranking constraints are imposed on the two subsets: any candidate from the first subset is better than that from the second one. In this way, the constraints are reduced dramatically compared to the full ranking model, which further facilitates an efficient learning procedure. Moreover, we show that our partial ranking model can be reduced into the large margin based framework. Extensive experiments demonstrate that after a re-ranking step of our model, the top-$k$ detection rate can be significantly improved.