Research map / Detection and segmentation / Weakly supervised detection and segmentation
CorrMatch: Label Propagation via Correlation Matching 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 CorrMatch, a single-stage semi-supervised semantic segmentation framework that uses correlation maps for pixel and region label propagation on top of weak-to-strong consistency with DeepLabV3+.
Problem
Semi-supervised semantic segmentation struggles to use unlabeled data efficiently because fixed confidence thresholds trade off pseudo-label quantity against accuracy, and prior methods rely on multiple networks, training stages, or multiple strong-augmentation streams.
Approach
Uses pixel-pixel correlation maps computed from encoder features to spread high-confidence predictions (pixel propagation) and to transfer class labels onto class-agnostic shape masks extracted from correlation rows (region propagation), expanding high-confidence pseudo-label regions without extra inference cost.
Key results
- On Pascal VOC 2012 classic setting with 92 labeled images (1/16), CorrMatch reaches 76.4% mIoU, versus 75.2% for UniMatch.
- With 1464 labels on classic VOC 2012, CorrMatch achieves 81.8% mIoU, outperforming UniMatch (81.2%).
- On Cityscapes 1/16 split, CorrMatch achieves 77.3% mIoU, +0.7% over UniMatch.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add pixel propagation to the no-propagation method (ablation) | CorrMatch · Pascal VOC 2012 val, original 92-label split | mIoU (%): 74.4 -> 75.8 | Helped |
| Add region propagation after pixel propagation (ablation) | CorrMatch · Pascal VOC 2012 val, original 92-label split | mIoU (%): 75.8 -> 76.4 | Helped |
| Add pixel propagation to the no-propagation method (ablation) | CorrMatch · Pascal VOC 2012 val, original 366-label split | mIoU (%): 78.5 -> 78.9 | Helped |
| Add region propagation after pixel propagation (ablation) | CorrMatch · Pascal VOC 2012 val, original 366-label split | mIoU (%): 78.9 -> 79.4 | Helped |
| Add pixel propagation to the no-propagation method (ablation) | CorrMatch · Pascal VOC 2012 val, original 1464-label split | mIoU (%): 80.5 -> 81.3 | Helped |
| Add region propagation after pixel propagation (ablation) | CorrMatch · Pascal VOC 2012 val, original 1464-label split | mIoU (%): 81.3 -> 81.8 | Helped |
| Add soft unsupervised loss to hard-loss training (ablation) | CorrMatch · Pascal VOC 2012 val, original 92-label split | mIoU (%): 73.6 -> 74.4 | Helped |
| Replace hard unsupervised loss with soft unsupervised loss (ablation) | CorrMatch · Pascal VOC 2012 val, original | mIoU (%) | Hurt |
| Use the dynamic threshold instead of fixed threshold 0.95 (ablation) | CorrMatch baseline · Pascal VOC 2012 val, original 92-label split | mIoU (%): 73.1 -> 73.6 | Helped |
| Extract correlation features from the backbone rather than ASPP, Fusion, or Classifier (ablation) | CorrMatch · Pascal VOC 2012 val, original 1464-label split | mIoU (%): 81.8; 80.6; 80.1; 80.8 | Helped |
| Use random sampling rather than uniform sampling for region shapes (ablation) | CorrMatch · Pascal VOC 2012 val, original 1464-label split | mIoU (%): 81.1 vs 81.0; 81.2 vs 81.1; 81.4 vs 81.2; 81.8 vs 81.4; 81.7 vs 81.0 | Helped |
| Increase random shape samples from 16 to 128 (ablation) | CorrMatch · Pascal VOC 2012 val, original 1464-label split | mIoU (%): 81.1 -> 81.8 | Helped |
Surprising findings
- Fixed-threshold ablation shows dynamic thresholding alone gives only +0.5% and +0.1%, suggesting the gains mainly come from label propagation.
Limitations
- Region propagation requires computing correlation maps and sampling shapes, adding training-time cost relative to plain thresholding.
- Gains over UniMatch are modest (often about 0.5-1.3% mIoU) on most splits.
Experimental setup
Datasets: Pascal VOC 2012 SBD augmented VOC 2012 Cityscapes
Benchmarks: Pascal VOC 2012 Cityscapes
Metrics: mIoU
Baselines: UniMatch ST++ PS-MT PCR GTA U2PL AugSeg CCVC CPS AEL
Builds on: FixMatch UniMatch DeepLabV3+ FreeMatch
Closest papers by meaning
- AllSpark: Reborn Labeled Features from Unlabeled in Transformer for Semi-Supervised Semantic Segmentation (CVPR 2024 · similarity 0.87)
- ScaleMatch: Multi-scale Consistency Enhancement for Semi-supervised Semantic Segmentation (AAAI 2025 · similarity 0.87)
- Separate and Conquer: Decoupling Co-occurrence via Decomposition and Representation for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.87)
- LPOSS: Label Propagation Over Patches and Pixels for Open-vocabulary Semantic Segmentation (CVPR 2025 · similarity 0.86)
- Connectivity-Driven Pseudo-Labeling Makes Stronger Cross-Domain Segmenters (NeurIPS 2024 · similarity 0.86)
- RankMatch: Exploring the Better Consistency Regularization for Semi-supervised Semantic Segmentation (CVPR 2024 · 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.