Research map / Detection and segmentation / Weakly supervised detection and segmentation
ScaleMatch: Multi-scale Consistency Enhancement 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 ScaleMatch, a semi-supervised semantic segmentation framework that fuses dual-scale pseudo-labels via a cross-scale interaction fusion module and enforces image- and feature-level scale-variation consistency.
Problem
Semi-supervised semantic segmentation models struggle to learn scale variation across objects when labeled data is scarce, and existing S4 methods perturb only at the image level without exploiting multi-scale information.
Approach
Combine a VRWKV-based cross-scale feature interaction and learned scale-aware gates to fuse predictions from different input scales into better pseudo-labels, then supervise extra variable-scale branches at the image level (ISVC) and feature level (FSVC) to get scale-invariant features at lower memory cost than multi-resolution inputs.
Key results
- On Pascal VOC original (513x513) with 1/4 labels, ScaleMatch reaches 80.53 mIoU (82.02 at 1/2), vs. UniMatch-style baselines around 79-80; with multi-scale inference ScaleMatch† reaches 81.46 at 1/4.
- On Cityscapes 1/16 split, ScaleMatch reaches 77.83 mIoU and 78.70 with multi-scale inference, outperforming CorrMatch (77.30) and RankMatch (77.10).
- Ablation on Pascal 1/2: ISVC adds about 3.2 mIoU over FixMatch baseline (77.89 to 81.06); FSVC and CIF add further gains to 82.02.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add image-level scale variation consistency (ISVC) (ablation) | FixMatch baseline · PASCAL Original 1/4 and 1/2 | mIoU: 75.78 -> 78.72; 77.89 -> 81.06 | Helped |
| add feature-level scale variation consistency (FSVC) to ISVC (ablation) | ScaleMatch · PASCAL Original 1/4 and 1/2 | mIoU: 78.72 -> 79.73; 81.06 -> 81.57 | Helped |
| add cross-scale interaction fusion (CIF) (ablation) | ScaleMatch with ISVC and FSVC · PASCAL Original 1/4 and 1/2 | mIoU: 79.73 -> 80.53; 81.57 -> 82.02 | Helped |
| add a 0.5× scale branch (ablation) | ISVC baseline · PASCAL Original | mIoU: 77.98 -> 79.12 | Helped |
| add a 2.0× scale branch (ablation) | ISVC baseline · PASCAL Original | mIoU: 77.98 -> 78.93 | Helped |
| use 0.5× and 2.0× scale branches together (ablation) | ISVC baseline · PASCAL Original | mIoU: 77.98 -> 80.86 | Helped |
| expand the scale group to 0.25×, 0.5×, 1.5×, and 2.0× (ablation) | ISVC · PASCAL Original | mIoU: 80.86 -> 81.06 | Helped |
| randomly sample scale factors from [0.5, 2.0] (ablation) | ISVC · PASCAL Original | mIoU: 81.06 -> 79.66 | Hurt |
| replace CIF fusion with average fusion (ablation) | ScaleMatch · PASCAL Original 1/2 | mIoU: 82.02 -> 81.68 | Hurt |
| start CIF warm-up at epoch 20 rather than epoch 10 (ablation) | ScaleMatch · PASCAL Original 1/2 | mIoU: 82.02 -> 81.95 | Hurt |
| add CIF-based multi-scale inference (ablation) | ScaleMatch · PASCAL Original 1/16, 1/8, 1/4, 1/2, and full | mIoU: 76.71 -> 77.80; 78.64 -> 79.72; 80.53 -> 81.46; 82.02 -> 83.15; 83.01 -> 84.43 | Helped |
| use ScaleMatch instead of the strongest listed baseline | PASCAL Original · PASCAL Original splits | mIoU: +0.64%; +0.23%; +0.73%; +0.81%; +0.81% | Helped |
Surprising findings
- Using randomly sampled scale factors in [0.5, 2.0] degraded performance (79.66 vs 81.06 for a fixed set), suggesting excessive scale variation hurts learning.
- Early in training, average fusion and CIF pseudo-labels are worse than original pseudo-labels, requiring a warm-up before CIF is used.
Limitations
- Warm-up timing for CIF is a hyperparameter that trades off noise against underused pseudo-label quality.
- Excessively wide scale ranges degrade performance, so scale factor set must be hand-selected.
- Multi-scale inference (†) adds inference cost beyond the reported single-scale numbers.
Experimental setup
Datasets: Pascal VOC 2012 SBD augmented Pascal VOC Cityscapes
Benchmarks: Pascal VOC 2012 original Pascal VOC 2012 augmented Cityscapes
Metrics: mIoU
Baselines: UniMatch CorrMatch DDFP AllSpark RankMatch AugSeg ST++ U2PL DAW ESL iMAS CCVC
Builds on: FixMatch UniMatch DeepLabV3+ Vision-RWKV CorrMatch
Closest papers by meaning
- Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · similarity 0.89)
- RankMatch: Exploring the Better Consistency Regularization for Semi-supervised Semantic Segmentation (CVPR 2024 · similarity 0.89)
- RWMS: Reliable Weighted Multi-Phase for Semi-supervised Segmentation (AAAI 2024 · similarity 0.87)
- CorrMatch: Label Propagation via Correlation Matching for Semi-Supervised Semantic Segmentation (CVPR 2024 · similarity 0.87)
- Towards the Uncharted: Density-Descending Feature Perturbation for Semi-supervised Semantic Segmentation (CVPR 2024 · similarity 0.87)
- Relaxed Class-consensus Consistency for Semi-supervised Semantic Segmentation (AAAI 2025 · similarity 0.87)
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.