Activation Extending Based on Long-range Dependencies for Weakly Supervised Semantic Segmentation

Weakly supervised semantic segmentation (WSSS) principally obtains pseudo-labels based on the class activation maps (CAM) to handle expensive annotation resources. However, CAM easily involves false and local activation due to the the lack of annotation information. This paper suggests weakly supervised learning as semantic information mining to extend object mask. We proposes a novel architecture to mining semantic information by modeling through long-range dependencies from in-sample and inter-sample. Considering the confusion caused by the long-range dependencies, the images are divided into blocks and carried out self-attention operation on the premise of fewer classes to obtain long-range dependencies, to reduce false predictions. Moreover, we perform global to local weighted self-supervised contrastive learning among image blocks, and the local activation of CAM is transferred to different foreground area. Experiments verified that superior semantic details and more reliable pseudo-labels are captured through these suggested modules. Experiments on PASCAL VOC 2012 demonstrated the proposed model achieves 65.2$\%$, 64.9$\%$, and 65.8$\%$ mIoU in train, val, and test sets, which is superior to the comparison baselines.