Hybrid Suppression and Attention with Online Augmentation for Weakly Supervised Semantic Segmentation

The primary approach in weakly supervised semantic segmentation involves utilizing a classification network to classify input images, extracting the model's regions of interest with Class Activation Maps (CAM), and creating pseudo masks for training the segmentation network. However, the model may excessively focus on certain features and struggle with accurately segmenting object boundaries. We proposed a method to address these issues and experimental results demonstrate our method achieving a pseudo mask mIoU of 71.153 on the PASCAL VOC 2012 training set. By training the semantic segmentation network using these pseudo masks, it achieves mIoU scores of 72.46 and 72.81 on the PASCAL VOC 2012 validation and test sets, superior to the state-of-the-art methods.