Weakly Supervised Semantic Segmentation via Progressive Patch Learning

Most of the existing semantic segmentation approaches with image-level class labels as supervision, highly rely on initial activation map (CAM) generated from standard classification network. In this paper, a novel ‘`Progressive Patch Learning’' approach is proposed to improve local details extraction classification, producing CAM better covering whole object rather than only most discriminative regions in CAMs obtained conventional models. ‘`Patch destructs feature maps into patches and independently processes each patch parallel before final aggregation. Such mechanism enforces network find weak information scattered parts, achieving enhanced sensitivity. further extends destruction learning multi-level granularities progressive manner. Cooperating multi-stage optimization strategy, such implicitly provides model ability across different locality-granularities. As an alternative implicit multi-granularity fusion approach, we additionally propose explicit method simultaneously fuse features granularites single model, enhancing quality full coverage. Our achieves outstanding performance PASCAL VOC 2012 dataset e.g., 69.6% mIoU test set), which surpasses weakly supervised methods.