Publications by authors named "Hangzhou He"
Branches Mutual Promotion for End-to-End Weakly Supervised Semantic Segmentation.
IEEE Trans Neural Netw Learn Syst
October 2024
Article Synopsis
- End-to-end weakly supervised semantic segmentation (E2E-WSSS) optimizes segmentation models using only image annotations, relying on a classification branch for pseudo annotations.
- The current approach causes the classification branch to dominate the training, limiting cooperation between the segmentation and classification branches.
- The proposed method equalizes the roles of both branches, implementing a bidirectional supervision mechanism and interaction operations, resulting in improved performance over existing E2E-WSSS methods.