Research map / Label-efficient and robust learning / Domain adaptation and generalization
Stable Neighbor Denoising for Source-free Domain Adaptive 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 Stable Neighbor Denoising (SND), which splits target samples into stable and unstable sets by evolution stability, retrieves stable neighbors via style and layout proxies, and uses bi-level optimization with category copy-paste compensation to denoise pseudo-labels for source-free segmentation adaptation.
Problem
Source-free unsupervised domain adaptation for semantic segmentation suffers from pseudo-label noise, causing class, domain, and confirmation biases in uncertainty-based self-training.
Approach
Pseudo-label noise concentrates in unstable samples whose predictions change a lot during self-training; SND denoises them via bi-level optimization guided by retrieved stable neighbors matched on image style and spatial layout, with object copy-paste to compensate missing classes.
Key results
- SND achieves 55.6 mIoU on GTA to Cityscapes and 54.1 on SYNTHIA to Cityscapes with DeepLab-v2, and SND+DTST reaches 58.1 and 55.9.
- On Cityscapes to ACDC, SND reaches 48.1 mIoU, and SND+DTST reaches 49.6 mIoU with DeepLab-v2.
- In continual SFUDA on Cityscapes to ACDC with SegFormer MiT-B5, SND reaches 61.8 average mIoU versus 59.2 for CoTTA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SND instead of CrossMatch | SFUDA segmentation · GTA5 → Cityscapes (Val.) | mIoU: 56.4 -> 55.6 | Hurt |
| Use SND instead of CrossMatch | SFUDA segmentation · Synthia → Cityscapes (Val.) | mIoU: 55.6 -> 54.1 | Hurt |
| Use SND instead of SFDASEG | SFUDA segmentation · Cityscapes → ACDC (Test) | mIoU: 46.4 -> 48.1 | Helped |
| Use SND instead of DTST | SFUDA segmentation · GTA5 → BDD100K (Test) | mIoU: 45.2 -> 47.8 | Helped |
| Use SND instead of CoTTA | Continual SFUDA segmentation · Cityscapes → ACDC | mIoU: 59.2 -> 61.8 | Helped |
| Use entropy, image-level loss, or domain distance instead of stability to divide subsets (ablation) | SND · GTA5 → Cityscapes | mIoU: 55.6 vs 52.9, 51.7, 52.7, 51.6 | Hurt |
| Use entropy, image-level loss, or domain distance instead of stability to divide subsets (ablation) | SND · Cityscapes → ACDC | mIoU: 48.1 vs 45.1, 43.6, 44.1, 42.4 | Hurt |
| Retrieve stable neighbors using layout and style instead of random, layout-only, or style-only selection (ablation) | SND · GTA5 → Cityscapes | mIoU: 54.2, 55.3, 55.1 -> 55.6 | Helped |
| Retrieve stable neighbors using layout and style instead of random, layout-only, or style-only selection (ablation) | SND · Cityscapes → ACDC | mIoU: 44.7, 44.9, 47.5 -> 48.1 | Helped |
| Add category compensation using stable samples (ablation) | SND · GTA5 → Cityscapes | mIoU: 53.0 -> 54.2 | Helped |
| Use Dsec stable category sets rather than Dse for category compensation (ablation) | SND · GTA5 → Cityscapes | mIoU: 54.2 -> 55.6 | Helped |
| Add category compensation using stable samples (ablation) | SND · Cityscapes → ACDC | mIoU: 46.3 -> 47.1 | Helped |
Surprising findings
- Un-adapted source-trained models fail to capture cross-image feature similarity, so neighbors must be retrieved using domain factors rather than learned features.
- Several existing uncertainty estimators (e.g., prototype distance, cross-image association) lower performance relative to unweighted pseudo-labels on some tasks.
Limitations
- Retrieval and compensation add computational and memory overhead relative to plain self-training.
- Stable set is fixed to top 5% by evolution stability, which is a hand-chosen hyperparameter.
- Gains on the ACDC task are smaller when combined with DTST relative to synthetic settings.
Experimental setup
Datasets: GTA5 SYNTHIA Cityscapes ACDC BDD100K
Benchmarks: GTA5 to Cityscapes SYNTHIA to Cityscapes Cityscapes to ACDC GTA5 to BDD100K
Metrics: mIoU IoU
Baselines: HCL SFDASEG DTST CrossMatch CROTS URMDA CoTTA TENT
Builds on: DTST meta-learning based noisy label denoising copy-paste mean teacher
Closest papers by meaning
- Balanced Learning for Domain Adaptive Semantic Segmentation (ICML 2025 · similarity 0.88)
- Universal Domain Adaptation for Semantic Segmentation (CVPR 2025 · similarity 0.88)
- Open-Set Domain Adaptation for Semantic Segmentation (CVPR 2024 · similarity 0.88)
- Towards Unsupervised Domain Bridging via Image Degradation in Semantic Segmentation (NeurIPS 2025 · similarity 0.88)
- SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation (CVPR 2025 · similarity 0.87)
- Density Matters: Improved Core-Set for Active Domain Adaptive Segmentation (AAAI 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.