Research map / Label-efficient and robust learning / Domain adaptation and generalization
Connectivity-Driven Pseudo-Labeling Makes Stronger Cross-Domain Segmenters
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 SeCo, which lifts pseudo-labels to SAM-based semantic connectivity regions and corrects connectivity-level noise via a connectivity classifier with GMM loss-based noise detection, improving UDA, source-free, black-box, and domain generalization segmentation.
Problem
Pixel-level pseudo-labels in cross-domain semantic segmentation are speckled and noisy under severe domain shift and contain open-set noise from wild synthetic data, causing error accumulation in self-training.
Approach
Aggregate speckled pseudo-labels into semantic connectivity regions using SAM (stuff via semantic alignment, things via box and point prompts), then treat connectivities as classification objects and detect/correct noisy ones with early-learning loss distribution modeled by a two-component GMM.
Key results
- Integrating SeCo with AdvEnt raises GTA5→Cityscapes mIoU by 13.4 points (45.5 to 58.9) and with DAFormer by 5.3 points (68.2 to 73.4).
- On source-free GTA5→Cityscapes, SeCo improves HCL by 10.6 points and DTST by 8.4 points.
- On GTA5→BDD-100K open compound adaptation, SeCo improves ProDA by 7.8 points (41.7 to 49.5) and PyCDA by 11.4 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SeCo to HRDA | HRDA · GTA5 → Cityscapes UDA | mIoU: 73.8 -> 76.1 | Helped |
| Add SeCo to DTST | DTST · GTA5 → Cityscapes source-free UDA | mIoU: 52.1 -> 60.5 | Helped |
| Add SeCo to BiMem | BiMem · GTA5 → Cityscapes black-box UDA | mIoU: 48.2 -> 56.7 | Helped |
| Add SeCo to HRDA | HRDA · SYNTHIA → Cityscapes UDA | mIoU: 65.8 -> 68.5 | Helped |
| Add SeCo to DTST | DTST · SYNTHIA → Cityscapes source-free UDA | mIoU: 47.7 -> 53.9 | Helped |
| Add SeCo to BiMem | BiMem · SYNTHIA → Cityscapes black-box UDA | mIoU: 42.2 -> 48.4 | Helped |
| Add SeCo to ProDA | ProDA · GTA5 → BDD-100k UDA | mIoU: 41.7 -> 49.5 | Helped |
| Add semantic alignment PSA(b) (ablation) | ProDA · GTA5 → Cityscapes UDA | mIoU: 53.7 -> 56.1 | Helped |
| Add PSA without SCC (ablation) | ProDA · GTA5 → Cityscapes UDA | mIoU: 53.7 -> 62.1 | Helped |
| Add SCC to PSA (ablation) | ProDA + PSA · GTA5 → Cityscapes UDA | mIoU: 62.1 -> 64.1 | Helped |
| Add semantic alignment PSA(b) (ablation) | DAFormer · GTA5 → Cityscapes UDA | mIoU: 68.2 -> 69.7 | Helped |
| Add SCC to PSA (ablation) | DTST + PSA · GTA5 → Cityscapes source-free UDA | mIoU: 54.4 -> 60.5 | Helped |
Surprising findings
- Directly applying SAM as a backbone reduces UDA performance (DAFormer 68.2 to 64.1 with SAM ViT-B), attributed to poor rare-class performance.
- Vanilla SAM refinement of coarse pseudo-labels gives minimal or negative gains on strong UDA baselines (e.g., ProDA 53.7 to 50.1).
Limitations
- Lacks an explicitly designed category-balanced connectivity noise filtering method.
- Open-set noise effectiveness on outdoor data could not be directly quantitatively evaluated due to environment limitations.
- Depends on SAM quality and on the quality of initial pseudo-labels; reliance on SAM's semantic granularity.
Experimental setup
Datasets: GTA5 SYNTHIA Cityscapes BDD-100K Mapillary ADE20K Endovis17 Endovis18 Potsdam Vaihingen
Benchmarks: Cityscapes BDD-100K Mapillary ADE20K Endovis17-Endovis18 Potsdam-Vaihingen
Metrics: mIoU
Baselines: AdvEnt ProDA DAFormer HRDA HCL DTST DINE BiMem SHADE TLDR MoDify CLOUDS
Builds on: SAM ProDA DivideMix CLOUDS DAFormer
Closest papers by meaning
- Learning Beyond Vision: Vision-Language Distillation and Edge-Aware Mix Diffusion in Semi-Supervised Semantic Segmentation (AAAI 2026 · similarity 0.89)
- W2P: Switching from Weak Supervision to Partial Supervision for Semantic Segmentation (AAAI 2024 · similarity 0.88)
- Fully Data-Driven Pseudo Label Estimation for Pointly-Supervised Panoptic Segmentation (AAAI 2024 · similarity 0.87)
- SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation (CVPR 2025 · similarity 0.87)
- Towards Robust Pseudo-Label Learning in Semantic Segmentation: An Encoding Perspective (NeurIPS 2025 · similarity 0.87)
- Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation (NeurIPS 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.