Self Correspondence Distillation for End-to-End Weakly-Supervised Semantic Segmentation

Efficiently training accurate deep models for weakly supervised semantic segmentation (WSSS) with image-level labels is challenging and important. Recently, end-to-end WSSS methods have become the focus of research due to their high efficiency. However, current suffer from insufficient extraction comprehensive information, resulting in low-quality pseudo-labels sub-optimal solutions WSSS. To this end, we propose a simple novel Self Correspondence Distillation (SCD) method refine without introducing external supervision. Our SCD enables network utilize feature correspondence derived itself as distillation target, which can enhance network's learning process by complementing information. In addition, further improve accuracy, design Variation-aware Refine Module local consistency computing pixel-level variation. Finally, present an efficient Transformer-based framework (TSCD) via task. Extensive experiments on PASCAL VOC 2012 MS COCO 2014 datasets demonstrate that our significantly outperforms other state-of-the-art methods. code available at https://github.com/Rongtao-Xu/RepresentationLearning/tree/main/SCD-AAAI2023.