Scholarlink

<p>Knowledge distillation (KD) is an attractive technique for attaining lightweight models in semantic segmentation. However, the existing methods directly emulating the dense logit map and affinity map only yield trivial improvements. This paper empirically imputes the failure to two fundamental issues: (i) The pre-trained teachers elicit excessive but useless overconfident logit maps. (ii) The difference in affinity map between the teacher and student is uneven among intra-class and inter-class parts. Based on these insightful observations, we propose Uncertainty-Aware and Decoupled Distillation (UADD), a novel KD framework for segmentation. Uncertainty-Aware Distillation (UAD) localizes the beneficial areas in the logit map with training-free uncertainty metrics and then regulates students to take more effort into imitating them. Decoupled Affinity Distillation (DAD) splits the affinity map into intra-class and inter-class parts and transfers them in a balanced manner. Extensive experiments on Cityscapes, PASCAL VOC, ADE20K, and CamVid demonstrate the effectiveness of our method. For example, DeepLabV3-ResNet18 achieves 77.00% mIoU with UADD on Cityscapes, which gives notable 2.78% gains and outperforms the state-of-the-art distillation methods. On top of that, the superior performance on MS-COCO of more challenging tasks, i.e., object detection and instance segmentation, also manifests its generalization and versatility. Code is available at: <a href="https://github.com/swift1988">https://github.com/swift1988</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

Uncertainty-Aware and Decoupled Distillation for Semantic Segmentation

  • 期刊论文

  • 发表日期：

    2025年12月18日

查看全文
  • Cong Li,
  • Gong Cheng,
  • Junwei Han

摘要

Knowledge distillation (KD) is an attractive technique for attaining lightweight models in semantic segmentation. However, the existing methods directly emulating the dense logit map and affinity map only yield trivial improvements. This paper empirically imputes the failure to two fundamental issues: (i) The pre-trained teachers elicit excessive but useless overconfident logit maps. (ii) The difference in affinity map between the teacher and student is uneven among intra-class and inter-class parts. Based on these insightful observations, we propose Uncertainty-Aware and Decoupled Distillation (UADD), a novel KD framework for segmentation. Uncertainty-Aware Distillation (UAD) localizes the beneficial areas in the logit map with training-free uncertainty metrics and then regulates students to take more effort into imitating them. Decoupled Affinity Distillation (DAD) splits the affinity map into intra-class and inter-class parts and transfers them in a balanced manner. Extensive experiments on Cityscapes, PASCAL VOC, ADE20K, and CamVid demonstrate the effectiveness of our method. For example, DeepLabV3-ResNet18 achieves 77.00% mIoU with UADD on Cityscapes, which gives notable 2.78% gains and outperforms the state-of-the-art distillation methods. On top of that, the superior performance on MS-COCO of more challenging tasks, i.e., object detection and instance segmentation, also manifests its generalization and versatility. Code is available at: https://github.com/swift1988.