SASA method uses semantic anchors to stop class overwriting in incremental segmentation
Learnable 'semantic anchors' prevent feature drift as models learn new classes over time.
Weakly Incremental Learning for Semantic Segmentation (WILSS) faces a core challenge: noisy supervision progressively corrupts class-level representations, causing feature drift and overwriting of previously learned categories. Researchers from Beijing University of Posts and Telecommunications (Zhonggai Wang, Kai Fang, Guangyu Gao) introduce SASA (Semantic Anchors and Spatial Arbitration) to stabilize learning without expensive pixel-level labels.
SASA operates on two fronts. At the representation level, it introduces semantic anchors — learnable tokens that serve as rigid, class-specific references to preserve long-term semantic identity, while an elastic residual adaptation allows controlled instance-specific refinement. At the supervision level, a Spatial Label Arbitration mechanism performs geometry-aware decisions to filter unreliable signals and enforce a strict 'one object, one class' constraint. Extensive experiments on standard benchmarks show SASA consistently outperforms prior state-of-the-art, particularly in challenging multi-step incremental settings. The work has been accepted at ICME2026 and the code is publicly available.
- Semantic anchors (learnable tokens) act as rigid class-level references to prevent feature drift from noisy supervision.
- Spatial Label Arbitration enforces geometry-aware 'one object, one class' filtering to improve supervision reliability.
- Outperforms existing methods on benchmarks, with code released and accepted at ICME2026.
Why It Matters
Enables vision models to learn new classes incrementally from weak labels without catastrophic forgetting.