Scholarlink

<p>In continual semantic segmentation (CSS), models face challenges such as catastrophic forgetting and background shift, requiring them to learn novel classes while maintaining the ability to segment previously learned ones. An effective approach commonly used to tackle these challenges is Knowledge Distillation (KD), which encourages the new model’s output distribution to closely match that of the pre-trained model through regularization. However, since KD depends on the output of the old model, erroneous knowledge about the new classes is inevitably transferred, and these errors propagate from one stage to the next as learning progresses. To address this limitation, we propose Label-Guided Class Similarity Distillation (LCSKD), which extracts new class pixels using the new class labels, evaluates the similarity between new and old classes based on the prototype’s PCA feature mapping, and weights the old model’s output distributions after reconstructing class relationships. This approach mitigates catastrophic forgetting and background shift by timely correcting the old model’s erroneous knowledge about new classes, preventing the continuous accumulation of errors. Extensive experiments on two popular CSS benchmark datasets, Pascal-VOC 2012 and ADE20K, demonstrate the effectiveness of our proposed method. Code can be found at <a href="https://github.com/lizrz/LCKD">https://github.com/lizrz/LCKD</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Label-guided class similarity distillation for continual semantic segmentation

  • 期刊论文

  • 发表日期：

    2026年4月29日

查看全文
  • Bing Wang,
  • Yanchuan Li,
  • Shuo Yang,
  • Taifeng Huang,
  • Zutong Zhao,
  • Shiyin Zhang

摘要

In continual semantic segmentation (CSS), models face challenges such as catastrophic forgetting and background shift, requiring them to learn novel classes while maintaining the ability to segment previously learned ones. An effective approach commonly used to tackle these challenges is Knowledge Distillation (KD), which encourages the new model’s output distribution to closely match that of the pre-trained model through regularization. However, since KD depends on the output of the old model, erroneous knowledge about the new classes is inevitably transferred, and these errors propagate from one stage to the next as learning progresses. To address this limitation, we propose Label-Guided Class Similarity Distillation (LCSKD), which extracts new class pixels using the new class labels, evaluates the similarity between new and old classes based on the prototype’s PCA feature mapping, and weights the old model’s output distributions after reconstructing class relationships. This approach mitigates catastrophic forgetting and background shift by timely correcting the old model’s erroneous knowledge about new classes, preventing the continuous accumulation of errors. Extensive experiments on two popular CSS benchmark datasets, Pascal-VOC 2012 and ADE20K, demonstrate the effectiveness of our proposed method. Code can be found at https://github.com/lizrz/LCKD.