TKRL: Targeted Knowledge Rectification Learning Against Teacher-Originated Defects in Domain Continual Segmentation

Le résumé fourni par la source

Knowledge distillation can mitigate catastrophic forgetting in domain continual segmentation by transferring knowledge from the older model to the newer model. However, existing distillation-based methods primarily emphasize knowledge retention while overlooking inherent defects in the older teacher models. As a result, these teacher-originated defects, such as knowledge gaps or biases, are propagated and exacerbate forgetting. To address this challenge, we propose a Targeted Knowledge Rectification Learning framework (TKRL) to probe and correct teacher-originated defects. TKRL consists of two modules: 1) Probe-augmented Class Distillation, which generates gradient-driven "probes" to uncover underrepresented features in the older model, thereby bridging knowledge gaps by distilling hidden information into the new model; 2) Variance-guided Masked Autoencoder, which selectively masks and reconstructs critical high-uncertainty patches across multi-level semantic regions, thereby correcting biases inherited from the older model. Our experimental results show that TKRL effectively rectifies knowledge gaps and biases, thereby mitigating catastrophic forgetting and enhancing performance in domain continual segmentation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.