Research map / Label-efficient and robust learning / Semi-supervised learning
Multi-Granularity Distillation Scheme towards Lightweight Semi-Supervised Semantic Segmentation
Jie Qin, Jie Wu, Ming Li, Xuefeng Xiao, Min Zheng, Xingang Wang
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 |
|---|---|---|---|
| replace CPS with MGD distillation | ResNet-18 student · VOC2012 val (1/16 labeled) | mIoU: 47.13 -> 66.86 | Improved |
| replace KD with MGD distillation | ResNet-18 student · VOC2012 val (1/16 labeled) | mIoU: 56.37 -> 66.86 | Improved |
| replace CPS with MGD | ResNet-18 student · Cityscapes val (1/16 labeled) | mIoU: 59.43 -> 70.91 | Improved |
| replace CPS with MGD | MobileNetv2 student · Cityscapes val (1/16 labeled) | mIoU: 53.34 -> 58.00 | Improved |
| add pixel-level consistency loss to supervised training (ablation) | MGD ResNet-18 · VOC2012 val (1/8 labeled) | mIoU: 55.65 -> 63.61 | Improved |
| add image-level semantic loss from T_D (ablation) | MGD ResNet-18 · VOC2012 val (1/8 labeled) | mIoU: 63.61 -> 66.01 | Improved |
| add region-level content-aware loss from T_W (ablation) | MGD ResNet-18 · VOC2012 val (1/8 labeled) | mIoU: 63.61 -> 66.02 | Improved |
| add both image-level loss on T_D and region-level loss on T_W (ablation) | MGD ResNet-18 · VOC2012 val (1/8 labeled) | mIoU: 63.61 -> 68.29 | Improved |
| apply image- and region-level losses to both teachers instead of one each (ablation) | MGD ResNet-18 · VOC2012 val (1/8 labeled) | mIoU: 68.29 -> 67.51 | Hurt |
| use complementary T_D plus T_W instead of two T_D teachers (ablation) | MGD ResNet-18 · VOC2012 val (1/8 labeled) | mIoU: 66.14 -> 68.29 | Improved |
| use two teachers instead of a single T_D (ablation) | MGD ResNet-18 · VOC2012 val (1/8 labeled) | mIoU: 64.85 -> 68.29 | Improved |
| add unlabeled data distillation losses (ablation) | MGD ResNet-18 · VOC2012 val (1/8 labeled) | mIoU: 58.24 -> 68.29 | Improved |
Closest papers by meaning
- Structured Knowledge Distillation for Semantic Segmentation (CVPR 2019 · similarity 0.89)
- Make a Strong Teacher with Label Assistance: A Novel Knowledge Distillation Approach for Semantic Segmentation (ECCV 2024 · similarity 0.89)
- Label-Guided Knowledge Distillation for Continual Semantic Segmentation on 2D Images and 3D Point Clouds (ICCV 2023 · similarity 0.89)
- Distilling Knowledge from Heterogeneous Architectures for Semantic Segmentation (AAAI 2025 · similarity 0.88)
- Adaptive Perspective Distillation for Semantic Segmentation (TPAMI 2022 · similarity 0.88)
- Pixel Contrastive-Consistent Semi-Supervised Semantic Segmentation (ICCV 2021 · 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.
