Research map / Label-efficient and robust learning / Knowledge distillation
Channel-Wise Knowledge Distillation for Dense Prediction
Changyong Shu, Yifan Liu, Jianfei Gao, Zheng Yan, Chunhua Shen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add channel-wise distillation on logits | PSPNet-R18 student · Cityscapes val (logits-map distillation) | mIoU: 69.10 -> 74.87 | Improved |
| add channel-wise distillation on feature map | PSPNet-R18 student · Cityscapes val (feature-map distillation) | mIoU: 69.10 -> 74.27 | Improved |
| use channel-wise distillation instead of best spatial method AT | PSPNet-R18 student · Cityscapes val (logits-map distillation) | mIoU: 72.32 -> 74.87 | Improved |
| normalize each channel with softmax and use KL instead of L2 without normalization (ablation) | CD · Cityscapes val (logits-map distillation) | mIoU: 70.83 -> 74.87 | Improved |
| use asymmetric KL instead of MSE on channel-normalized maps (ablation) | CD · Cityscapes val (logits-map distillation) | mIoU: 71.60 -> 74.87 | Improved |
| use asymmetric KL instead of symmetric Bhattacharyya distance (ablation) | CD · Cityscapes val (logits-map distillation) | mIoU: 72.21 -> 74.87 | Improved |
| normalize over channels rather than spatial locations (PI) (ablation) | CD · Cityscapes val (feature-map distillation) | mIoU: 70.02 -> 74.27 | Improved |
| swap teacher and student order in KL divergence (ablation) | CD · Cityscapes val (logits-map distillation) | training convergence | Hurt |
| lower temperature T so only salient pixels are matched (ablation) | CD · Cityscapes val (logits-map distillation) | mIoU | Hurt |
| initialize student with ImageNet weights (ablation) | PSPNet-R18 + CD logits · Cityscapes val | mIoU: 71.03 -> 75.90 | Improved |
| use halved-width student instead of full-width R18 (ablation) | PSPNet-R18 + CD logits · Cityscapes val (from scratch) | mIoU: 71.03 -> 68.57 | Hurt |
| add channel-wise distillation to RetinaNet | RetinaNet R50 · COCO 2017 detection | AP: 37.4 -> 40.8 | Improved |
Closest papers by meaning
- Structured Knowledge Distillation for Dense Prediction (TPAMI 2020 · similarity 0.91)
- Exploring Inter-Channel Correlation for Diversity-Preserved Knowledge Distillation (ICCV 2021 · similarity 0.90)
- Prediction-Guided Distillation for Dense Object Detection (ECCV 2022 · similarity 0.90)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.88)
- FreeKD: Knowledge Distillation via Semantic Frequency Prompt (CVPR 2024 · similarity 0.88)
- Over-parameterized Student Model via Tensor Decomposition Boosted Knowledge Distillation (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
