Weakly Supervised Object Localization Using Long-Range Semantic Foreground Activation
Lianxing Wang, Huaxiong Li · 2022 26th International Conference on Pattern Recognition (ICPR) · 2022
Weakly supervised object localization (WSOL) is a challenging task to find the object location using only image-level supervision. Previous works using CNN architectures stack in finding only the most discriminative parts. Transformer-based methods expand the finding areas but they still fail to take full advantage of attention information. To address this problem, we propose the Long-range Semantic Foreground Activation (LSFA) method. We demonstrate that localization maps should be generated by attention parameters under the guidance of non-discriminative foreground features. In LSFA, a visual transformer model is first used to generate attention information. Then, a long-range dependency activation module is constructed to help the learned attention maps to focus more on global information by weighting them with different parameters. Finally, a semantic foreground activation module is built to use non-discriminative foreground regions learned from attention maps as indices to activate token semantic areas. Experiments on two benchmark datasets CUB-200-2011 and ILSVRC demonstrate the superiority of our LSFA method in comparison with other state-of-the-art WSOL approaches.