Research map / Label-efficient and robust learning / Semi-supervised learning
RWMS: Reliable Weighted Multi-Phase for Semi-supervised 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 RWMS, a dual-teacher self-training method for semi-supervised semantic segmentation that uses image-level (MIoU-based) reliability for multi-phase retraining and pixel-level confidence-rank reweighting, plus strong augmentation on unlabeled data.
Problem
Semi-supervised semantic segmentation suffers from noisy pseudo labels in self-training, and single models cannot reliably correct their own errors; mutual-training alternatives double training cost.
Approach
Use two teacher models (DeepLabv3+ and PSPNet) only to judge reliability: image-level MIoU between their pseudo labels selects reliable images for stage-1 retraining, and pixel-level weights based on the rank of teacher 2's confidence for teacher 1's predicted class reweight the loss, while only the main student is trained.
Key results
- On Pascal VOC 2012 with DeepLabv3+ ResNet50 at 1/8 labels, RWMS reaches 74.9 MIoU vs 71.9 for the self-training baseline and 72.4 for DCC at 1/4 (75.1 vs 72.4 baseline at 1/4).
- On Cityscapes with DeepLabv3+ ResNet50 at 1/8 labels, RWMS gets 71.9 MIoU vs 68.7 baseline and 69.7 DCC.
- Ablation: selective multi-phase retraining (74.9) beats one-phase (72.1) and random multi-phase (72.3) at 1/8 on VOC; removing pixel reweighting drops 1/8 performance from 74.9 to 73.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply RWMS instead of the strongest listed baseline | DeepLabv3+ ResNet50 · Pascal VOC 2012, 1/16 | MIoU: 70.1 -> 72.2 | Helped |
| Apply RWMS instead of the strongest listed baseline | DeepLabv3+ ResNet50 · Pascal VOC 2012, 1/8 | MIoU: 73.2 -> 74.9 | Helped |
| Apply RWMS instead of the strongest listed baseline | DeepLabv3+ ResNet50 · Pascal VOC 2012, 1/4 | MIoU: 74.6 -> 75.1 | Helped |
| Apply RWMS instead of the strongest listed baseline | DeepLabv3+ ResNet101 · Pascal VOC 2012, 1/16 | MIoU: 72.4 -> 74.0 | Helped |
| Apply RWMS instead of the strongest listed baseline | DeepLabv3+ ResNet101 · Pascal VOC 2012, 1/8 | MIoU: 75.1 -> 75.7 | Helped |
| Apply RWMS instead of the strongest listed baseline | DeepLabv3+ ResNet101 · Pascal VOC 2012, 1/4 | MIoU: 76.6 -> 76.8 | Helped |
| Apply RWMS instead of the strongest listed baseline | DeepLabv3+ ResNet50 · Cityscapes, 1/30 | MIoU: 55.7 -> 60.8 | Helped |
| Apply RWMS instead of the strongest listed baseline | DeepLabv3+ ResNet50 · Cityscapes, 1/8 | MIoU: 69.7 -> 71.9 | Helped |
| Apply RWMS instead of the strongest listed baseline | DeepLabv3+ ResNet50 · Cityscapes, 1/4 | MIoU: 72.7 -> 73.2 | Helped |
| Add selective reliability-based multi-phase training instead of one-phase retraining (ablation) | RWMS · Pascal VOC 2012, 1/8 | MIoU: 72.1 -> 74.9 | Helped |
| Use random image grouping instead of reliability-based selection (ablation) | RWMS · Pascal VOC 2012, 1/4 | MIoU: 75.1 -> 72.8 | Hurt |
| Replace the PSPNet with ResNet50 auxiliary teacher with DeepLabv2 with ResNet101 (ablation) | RWMS · Pascal VOC 2012, 1/8 | MIoU: 74.9 -> 74.3 | Hurt |
Surprising findings
- Arbitrarily increasing the number of retraining stages without reliability-based selection does not improve and can underperform single-phase retraining.
- Greater disagreement between the two teacher models (lower pseudo-label MIoU) was associated with better semi-supervised performance (PSPNet teacher 2 gave the highest 74.9 MIoU).
Limitations
- Reliance on a fixed top-50% split for image reliability, with no sensitivity analysis on the split ratio.
- Only evaluated on Pascal VOC 2012 and Cityscapes with ResNet backbones and DeepLab/PSPNet architectures.
- Hyperparameters a and b of the pixel weighting function are described as predefined but their values are not reported in the text.
Experimental setup
Datasets: Pascal VOC 2012 SBD Cityscapes
Benchmarks: Pascal VOC 2012 Cityscapes
Metrics: MIoU
Baselines: SupOnly Baseline self-training ECS GCT DCC ELN SOSA DMT CutMix ClassMix
Builds on: self-training ST++ DMT FixMatch
Closest papers by meaning
- Semi-Supervised Semantic Segmentation Using Unreliable Pseudo-Labels (CVPR 2022 · similarity 0.90)
- DAW: Exploring the Better Weighting Function for Semi-supervised Semantic Segmentation (NeurIPS 2023 · similarity 0.89)
- Single-Stage Semantic Segmentation From Image Labels (CVPR 2020 · similarity 0.89)
- Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · similarity 0.88)
- ScaleMatch: Multi-scale Consistency Enhancement for Semi-supervised Semantic Segmentation (AAAI 2025 · similarity 0.87)
- AllSpark: Reborn Labeled Features from Unlabeled in Transformer for Semi-Supervised Semantic Segmentation (CVPR 2024 · 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.