Adaptively Denoising Proposal Collection forWeakly Supervised Object Localization

In this paper, we address the problem of weakly supervisedobject localization (WSL), which trains a detection network on the datasetwith only image-level annotations. The proposed approach is built on theobservation that the proposal set from the training dataset is a collectionof background, object parts, and objects. Several strategies are taken toadaptively eliminate the noisy proposals and generate pseudo object-levelannotations for the weakly labeled dataset. A multiple instance learning(MIL) algorithm enhanced by mask-out strategy is adopted to collect theclass-specific object proposals, which are then utilized to adapt a pre-trained classification network to a detection network. In addition, thedetection results from the detection network are re-weighted by jointlyconsidering the detection scores and the overlap ratio of proposals in aproposal subset optimization framework. The optimal proposals work asobject-level labels that enable a pseudo-strongly supervised dataset fortraining the detection network. Consequently, we establish a fully adap-tive detection network. Extensive evaluations on the PASCAL VOC 2007and 2012 datasets demonstrate a significant improvement compared withthe state-of-the-art methods.