Salient-aware multiple instance learning optimized network for weakly supervised object detection

In recent years, weakly supervised object detection network has achieved great development. However, due to the lack of bounding box supervision, the framework based on multiple instance learning tends to activate a part of the object rather than the whole object, which severely affects the detection performance for nonrigid objects. To solve this problem, this paper uses traditional features and sample weighting to guide the network to focus on the whole rather than the part of the object. Especially, salient priors are introduced to provide coarse pseudo bounding boxes to assist network initialization and to explore more accurate features in conjunction with the multi-object search strategy. In addition, we design an area-guided sample weighting algorithm to optimize the network to search for objects from larger areas, which avoids local domination. Experiments on public datasets (PASCAL VOC2007, PASCAL VOC2012) show that the proposed algorithm outperforms several state-of-the-art models.