RSMNet: A Regional Similar Module Network for Weakly Supervised Object Localization

Weakly-supervised object detection (WSOD) has attracted many people's attention because it can reduce labor and time-consuming, and some bias or errors caused by the subjectivity of annotators. As a key step of WSOD, weakly supervised object localization (WSOL) refers to identify object class and localization via image-level labels in a given image for further training a detector. However, WSOL is a challenging task only using image-level labels and some existing methods often suffer from incomplete activation or background activation for object localization. In this paper, we propose a novel regional similarity module network (RSMNet) for weakly supervised object localization, in which we present a regional similarity module (RSM) to calculate a similarity matrix between any two local pixels in the feature maps, and then integrate this similarity matrix with CAM to generate a similar activation map (SAM) for the extension of activation regions. Meanwhile, we put forward a background regularization loss and a difference regularization loss to suppress the background activation. Finally, in the inference stage, we present a feature map accumulation (FMA) method to efficiently collect more semantic features so that complete activation regions can be achieved. Experimental results show that RSMNet can achieve competitive performance compare with state-of-the-art.