Research map / Label-efficient and robust learning / Semi-supervised learning
Relaxed Class-consensus Consistency for Semi-supervised Semantic Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Relaxed Class-consensus Consistency (RCC), which models inter-class consensus via Shapley interactions of prediction entropy and aligns teacher/student class-consensus rankings, added on top of UniMatch for semi-supervised semantic segmentation.
Problem
Semi-supervised semantic segmentation methods enforce consistency independently per class and ignore semantic dependencies among classes, which hurts exploitation of unlabeled data especially in low-data regimes.
Approach
Treat classes as players in a cooperative game with prediction entropy as the characteristic function, measure pairwise class consensus via Shapley interaction, and enforce consistency between teacher and student on the ranking (rather than values) of these consensus scores, without external knowledge.
Key results
- On PASCAL classic with ResNet-50 under 1/16 (92 labels), RCC reaches 71.8% mIoU vs. 67.4% for the UniMatch baseline.
- On Cityscapes with ResNet-50 under 1/16, RCC improves over the supervised-only model by 12.4 mIoU (75.7% vs. 63.3%).
- Ablation on PASCAL classic 1/16: CC adds 2.8 mIoU over baseline and relaxed ranking alignment adds a further 1.6 mIoU (71.8%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add class-consensus consistency to UniMatch (ablation) | UniMatch · PASCAL classic 1/16 (92) | mIoU: 67.4 -> 70.2 | Helped |
| relax direct class-consensus alignment to ranking alignment (ablation) | CC · PASCAL classic 1/16 (92) | mIoU: 70.2 -> 71.8 | Helped |
| use RCC instead of UniMatch | UniMatch · PASCAL classic 1/16 (92) | mIoU (%): +4.4% | Helped |
| use RCC instead of UniMatch | UniMatch · PASCAL blender 1/16 (662) | mIoU (%): 75.8 -> 76.9 | Helped |
| use RCC instead of UniMatch | UniMatch · Cityscapes 1/16 (186) | mIoU (%): 75.0 -> 75.7 | Helped |
| use relaxed ranking alignment rather than L2, CE, or KL for CC (ablation) | class-consensus consistency · PASCAL classic 1/16 (92) | mIoU: L2 70.2; CE 69.9; KL 70.0; Relaxed 71.8 | Helped |
| vary the number of top classes used for consensus from 3 to 6 (ablation) | RCC · PASCAL classic 1/16 (92) | mIoU: 3: 70.3; 4: 71.8; 5: 72.0; 6: 71.9 | Mixed |
| set RCC trade-off weight λ to 0.1, 0.5, or 1.0 instead of 0.2 (ablation) | RCC · PASCAL classic 1/16 (92) | mIoU: 0.1: 70.5; 0.2: 71.8; 0.5: 71.5; 1.0: 70.0 | Mixed |
Surprising findings
- Strict alignment losses (L2, CE, KL) on class-consensus values underperform the ranking-based relaxed loss (70.2/69.9/70.0 vs. 71.8 mIoU).
Limitations
- Computing consensus over all K classes is combinatorially expensive, so only the top-N (N=4) classes per pixel are used.
- Class-consensus values come from unsupervised optimization and are less reliable than ground-truth-supervised signals, motivating the ranking relaxation.
- Adds training overhead relative to the UniMatch baseline.
Experimental setup
Datasets: PASCAL VOC 2012 SBD Cityscapes
Benchmarks: PASCAL VOC 2012 classic PASCAL VOC 2012 blender Cityscapes
Metrics: mIoU FPS FLOPs
Baselines: FixMatch UniMatch AugSeg iMAS DAW U2PL CSS Switch NP-SemiSeg Co-Train
Builds on: UniMatch FixMatch DeepLabv3+ Shapley interaction
Closest papers by meaning
- RankMatch: Exploring the Better Consistency Regularization for Semi-supervised Semantic Segmentation (CVPR 2024 · similarity 0.87)
- Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · similarity 0.87)
- 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)
- DuPL: Dual Student with Trustworthy Progressive Learning for Robust Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.86)
- Beyond Confidence: Exploiting Homogeneous Pattern for Semi-Supervised Semantic Segmentation (ICML 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.