Saliency Guided Inter- and Intra-Class Relation Constraints for Weakly Supervised Semantic Segmentation

Weakly supervised semantic segmentation with only image-level labels aims to reduce annotation costs for the task. Existing approaches generally leverage class activation maps (CAMs) locate object regions pseudo label generation. However, CAMs can discover most discriminative parts of objects, thus leading inferior pixel-level labels. To address this issue, we propose a saliency guided Inter- and Intra-Class Relation Constrained (I2CRC) framework assist expansion activated in CAMs. Specifically, class-agnostic distance module pull intra-category features closer by aligning their prototypes. Further, class-specific push inter-class apart encourage region have higher than background. Besides strengthening capability classification network activate more integral CAMs, also introduce an refinement take full use both prediction initial obtaining superior pseudo-labels. Extensive experiments on PASCAL VOC 2012 COCO datasets demonstrate well effectiveness I2CRC over other state-of-the-art counterparts.