Salient-Boundary-Guided Pseudo-Pixel Supervision for Weakly-Supervised Semantic Segmentation

This letter presents an innovative approach for generating pixel-wise pseudo masks as supervision for image-level Weakly Supervised Semantic Segmentation (WSSS). This is achieved by leveraging abundant object boundaries extracted with the guidance of saliency maps (SMs). Initially, we synthesize the boundary labels by combining Class Activation Maps (CAMs) and SMs. Then, an elaborately-designed joint training strategy is employed to fully exploit the complementary relationship between the foreground of CAMs and the background and boundary of SMs to yield rich object boundaries. Finally, we refine the CAMs based on the constraints imposed by the extracted boundaries, leading to more accurate pixel-wise pseudo masks. We thoroughly evaluate the performance of our proposed pseudo masks through extensive experiments, demonstrating their effectiveness as the supervision for accurate semantic segmentation. Specifically, our method achieves 71.7% mIoU and 39.1% mIoU on the validation sets of PASCAL VOC 2012 and MS COCO 2014, respectively.