Research map / Detection and segmentation / Weakly supervised detection and segmentation
RankMatch: Exploring the Better Consistency Regularization 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 RankMatch, which adds agent-level inter-pixel correlation consistency and rank-aware (permutation-probability) agent consistency on top of UniMatch for semi-supervised semantic segmentation.
Problem
Semi-supervised semantic segmentation needs stronger unlabeled supervision than pixel-wise consistency, but contrastive methods are complex, memory-heavy, and prone to confirmation bias.
Approach
Select representative agents via orthogonal selection, compute pixel-to-agent correlations, and impose consistency on agent-ranking permutation probability distributions (top-4 agents) between teacher and student views.
Key results
- On PASCAL classic with ResNet-50 at 1/16 (92 labels), RankMatch reaches 71.6 mIoU vs 67.4 for UniMatch.
- On PASCAL classic full set with ResNet-101, RankMatch reaches 82.2 mIoU vs 81.9 for UniMatch.
- On Lucchi mitochondria at 1/32 labels, RankMatch achieves 86.9 mIoU, above DualRel (85.6).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add RankMatch rank-aware correlation consistency to UniMatch | UniMatch with ResNet-50 · PASCAL classic 1/16 (92) | mIoU: 67.4 -> 71.6 | Helped |
| add RankMatch rank-aware correlation consistency to UniMatch | UniMatch with ResNet-50 · PASCAL classic Full (1464) | mIoU: 79.3 -> 80.0 | Helped |
| add RankMatch rank-aware correlation consistency to UniMatch | UniMatch with ResNet-50 · PASCAL blender 1/16 (662) | mIoU: 75.8 -> 76.6 | Helped |
| add RankMatch rank-aware correlation consistency to UniMatch | UniMatch with ResNet-50 · PASCAL blender 1/4 (2646) | mIoU: 76.8 -> 78.3 | Helped |
| add RankMatch rank-aware correlation consistency to UniMatch | UniMatch with ResNet-50 · Cityscapes 1/16 (186) | mIoU: 75.0 -> 75.4 | Helped |
| add RankMatch rank-aware correlation consistency to UniMatch | UniMatch with ResNet-50 · Cityscapes 1/2 (1488) | mIoU: 78.6 -> 79.5 | Helped |
| add contrastive learning (ablation) | UniMatch with ResNet-50 · PASCAL classic 1/16 (92) | mIoU: 67.4 -> 68.6 | Helped |
| add agent-level correlation consistency without rank-aware modeling (ablation) | UniMatch with ResNet-50 · PASCAL classic 1/16 (92) | mIoU: 67.4 -> 70.3 | Helped |
| replace agent-independent correlation consistency with rank-aware correlation consistency (ablation) | Correlation consistency with ResNet-50 · PASCAL classic 1/16 (92) | mIoU: 70.3 -> 71.6 | Helped |
| select agents by orthogonal selection instead of using all pixels (ablation) | RankMatch with ResNet-50 · PASCAL classic 1/16 (92) | mIoU: 69.8 -> 71.6 | Helped |
| select agents by orthogonal selection instead of random selection (ablation) | RankMatch with ResNet-50 · PASCAL classic 1/16 (92) | mIoU: 70.2 -> 71.6 | Helped |
| replace rank-aware consistency with L2 consistency (ablation) | RankMatch with ResNet-50 · PASCAL classic 1/16 (92) | mIoU: 71.6 -> 70.2 | Hurt |
Surprising findings
- Naive agent-independent correlation consistency (KL) gives only 70.3 vs 71.6 for rank-aware, and L2/CE variants are similarly weaker.
- Contrastive learning (U2PL reproduced on UniMatch) gives a smaller gain (68.6) than correlation consistency alone (70.3) at 1/16 on PASCAL.
Limitations
- Full permutation computation over all agents is intractable, so only top-4 agent permutations are used.
- Agent count N and trade-off weight lambda require tuning (N=128, lambda=0.1).
Experimental setup
Datasets: PASCAL VOC 2012 SBD Cityscapes Lucchi
Benchmarks: PASCAL VOC 2012 classic PASCAL VOC 2012 blender Cityscapes Lucchi
Metrics: mIoU
Baselines: FixMatch UniMatch U2PL AugSeg iMAS DAW ESL DualRel CPS CCT
Builds on: UniMatch FixMatch DeepLabv3+
Closest papers by meaning
- ScaleMatch: Multi-scale Consistency Enhancement for Semi-supervised Semantic Segmentation (AAAI 2025 · similarity 0.89)
- Beyond Confidence: Exploiting Homogeneous Pattern for Semi-Supervised Semantic Segmentation (ICML 2025 · similarity 0.88)
- Relaxed Class-consensus Consistency for Semi-supervised Semantic Segmentation (AAAI 2025 · similarity 0.87)
- On the Relaxation of Conditional Independence Assumption for Image Segmentation (NeurIPS 2026 · similarity 0.87)
- Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · similarity 0.87)
- DuSSS: Dual Semantic Similarity-Supervised Vision-Language Model for Semi-Supervised Medical Image Segmentation (AAAI 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.