Siamese Representation Learning for Weakly Supervised Object Localization

Weakly supervised object localization (WSOL) technology tries to use image-level annotations as supervision to learn object-level detectors, which can effectively reduce the labor and time cost of constructing object detection systems. Existing WSOL methods usually adopt a computational framework based on class activation mapping, which only supports a single input scale and lacks constraints in feature learning. This makes the existing WSOL methods unable to learn different scale features of the same object category, resulting in the model cannot capture the complete object region in complex scenarios, and making it hard to locate the object instances of multiple scales at the same time. This paper proposes a novel WSOL framework based on siamese representation learning, which includes a self-supervised semantic mask (SSSM) module and a cross-scale semantic consistency constraint (CsSCC) module. The framework enhances the perception ability of the model to the object features at different scales by jointly learning different scale transformations of input samples. Extensive experiments on typical WSOL datasets demonstrate the advantage of our method in improving multi-scale localization ability and improving localization integrity under the weak supervision setting.