Weakly Supervised Incremental Segmentation via Semantic Anchors and Spatial Arbitration
Researchers have developed a new approach called SASA to improve weakly supervised incremental learning for semantic segmentation. This method uses learnable tokens as semantic anchors to maintain class identity and a spatial arbitration mechanism to filter unreliable supervision signals. SASA aims to prevent newly learned classes from overwriting older ones, demonstrating superior performance in multi-step incremental learning scenarios. AI
IMPACT Enhances the robustness of AI models in learning new visual categories over time without forgetting previous ones.