Iterative weakly supervised semantic segmentation network with fused superpixel clues

Abstract

Most weakly supervised semantic segmentation networks lack semantic information between pixels, resulting in incomplete segmentation of targets. In this paper, we propose an iterative weakly supervised semantic segmentation network with fused superpixel clues. First, we generate attention class activation maps from the intermediate layers of the classification network, then use superpixel clues to optimize the class attention region and generate pseudo-labels with pixel correlations. After training the semantic segmentation network on these pseudo-labels, we continue to input the predicted pixel map iteratively into the superpixel network and the semantic segmentation network until the entire process is complete and the final semantic segmentation map is obtained. Results on PASCAL VOC 2012 demonstrate that the proposed method has superior segmentation performance compared with other weakly supervised semantic segmentation methods.