Research map / Label-efficient and robust learning / Semi-supervised learning
C3-SemiSeg: Contrastive Semi-Supervised Segmentation via Cross-Set Learning and Dynamic Class-Balancing
Yanning Zhou, Hang Xu, Wei Zhang, Bin Gao, Pheng-Ann Heng
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 consistency loss with augmentation (versus supervised baseline plus contrast only) (ablation) | C3-SemiSeg · Cityscapes val (1/30, 1/8, 1/4 labeled) | mIoU: 48.8 -> 54.4, 60.2 -> 63.3, 64.4 -> 65.8 | Improved |
| add pixel-wise contrastive loss on top of consistency (ablation) | C3-SemiSeg · Cityscapes val (1/30, 1/8, 1/4 labeled) | mIoU: 54.4 -> 55.0, 63.3 -> 63.7, 65.8 -> 66.4 | Improved |
| apply contrastive loss only to labeled features DL (ablation) | C3-SemiSeg · Cityscapes val (1/30, 1/8, 1/4 labeled) | mIoU: 54.4 -> 53.5, 63.3 -> 61.9, 65.7 -> 65.8 | Hurt |
| apply contrastive loss to unlabeled features DU (ablation) | C3-SemiSeg · Cityscapes val (1/30, 1/8, 1/4 labeled) | mIoU: 54.4 -> 54.6, 63.3 -> 63.5, 65.7 -> 66.4 | Improved |
| apply contrastive loss across DL and DU with segmentation-aware sampling (ablation) | C3-SemiSeg · Cityscapes val (1/30, 1/8, 1/4 labeled) | mIoU: 54.6 -> 55.0, 64.2 -> 63.7, 65.8 -> 66.4 | Improved |
| use strong augmentation with RandAugment on both teacher and student (ablation) | consistency regularization · Cityscapes val (1/30, 1/8, 1/4 labeled) | mIoU: 45.0 -> 43.2, 55.7 -> 54.8, 60.8 -> 57.1 | Hurt |
| use weak augmentation on both networks (ablation) | consistency regularization · Cityscapes val (1/30, 1/8, 1/4 labeled) | mIoU: 45.0 -> 50.6, 55.7 -> 59.3, 60.8 -> 62.0 | Improved |
| asymmetric augmentation: weak for teacher, strong for student (ablation) | consistency regularization · Cityscapes val (1/30, 1/8, 1/4 labeled) | mIoU: 50.6 -> 53.9, 59.3 -> 63.2, 62.0 -> 64.8 | Improved |
| add RandAugment to the student's strong augmentation (ablation) | consistency regularization · Cityscapes val (1/30, 1/8, 1/4 labeled) | mIoU: 53.9 -> 54.4, 63.2 -> 63.3, 64.8 -> 65.8 | Improved |
| add intra-set region-level data mixing (ablation) | C3-SemiSeg · Cityscapes val (1/30, 1/8, 1/4 labeled) | mIoU: 51.1 -> 53.3, 60.7 -> 62.4, 63.3 -> 65.2 | Improved |
| mix across labeled and unlabeled sets instead of within sets (ablation) | C3-SemiSeg · Cityscapes val (1/30, 1/8, 1/4 labeled) | mIoU: 53.3 -> 54.4, 62.4 -> 63.3, 65.2 -> 65.8 | Improved |
| add DCRS (s=0.8 versus s=1.0, i.e. no selection) (ablation) | C3-SemiSeg · Cityscapes val (1/8 labeled) | mIoU: 62.8 -> 63.3 | Improved |
Closest papers by meaning
- Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · similarity 0.90)
- Semi-Supervised Semantic Segmentation With Cross-Consistency Training (CVPR 2020 · similarity 0.90)
- Semi-Supervised Semantic Segmentation With Pixel-Level Contrastive Learning From a Class-Wise Memory Bank (ICCV 2021 · similarity 0.89)
- Pixel Contrastive-Consistent Semi-Supervised Semantic Segmentation (ICCV 2021 · similarity 0.89)
- ScaleMatch: Multi-scale Consistency Enhancement for Semi-supervised Semantic Segmentation (AAAI 2025 · similarity 0.87)
- SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation (CVPR 2025 · similarity 0.86)
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.
