Weakly supervised object detection with interactive edge attentive collaboration

Weakly supervised object detectors based on image-level annotation tend to overfit in the discriminative regions while ignoring the integrity of the object. In this paper, a novel weakly supervised object detection network with the interactive edge attentive collaboration module is proposed to alleviate the local optimal problem, in which edge attention is extracted as an object unity supervision for the detection, and a collaborative loss is introduced to enable VGG16 feature map with global attentive ability. The module can be detached from the network in the test period, which ensures the high efficiency of the detector without introducing any additional inference cost. Extensive experiments are carried out on the PASCAL VOC 2007 and VOC 2012 datasets, which reach 52.3% mAP, 67.7% CorLoc and 49.1% mAP, 68.0% CorLoc respectively, outperforming state-of-the-arts.