Research map / Label-efficient and robust learning / Domain adaptation and generalization
Open-Set Domain Adaptation for 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.
Introduces open-set domain adaptation for semantic segmentation (OSDA-SS) and proposes BUS, combining a dilation-erosion contrastive loss (DECON) and a domain/size-mixing augmentation (OpenReMix) on a head-expanded DAFormer/MIC self-training baseline.
Problem
Unsupervised domain adaptation for semantic segmentation assumes a shared label space, so target-private classes are misclassified as known classes; the paper targets rejecting unknown classes and predicting their boundaries and shapes in the target domain.
Approach
Extends the classifier head by one unknown class in a self-training UDA pipeline, adds a contrastive loss that uses morphological dilation/erosion masks of pseudo-labelled private regions to sharpen known/unknown boundaries, and a mixing augmentation (OpenReMix) that resizes source objects and pastes target-private regions into source images for size-invariant shape learning.
Key results
- On GTA5 to Cityscapes, BUS reaches H-Score 62.81 vs. 23.36 for DAFormer (+39.45 points), with private IoU 55.42.
- On SYNTHIA to Cityscapes, BUS reaches H-Score 44.01 vs. 20.82 for HRDA (+23.19 points).
- Ablation on GTA5 to Cityscapes: head-expansion baseline H-Score 43.79; adding DECON gives 57.57 and OpenReMix gives 58.34; both together give 62.81.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Expand the classifier head from C to C+1 classes (ablation) | MIC on GTA5 → Cityscapes · GTA5 → Cityscapes | H-Score: 19.71 -> 43.79 | Helped |
| Add DECON loss (ablation) | Head-expansion baseline on GTA5 → Cityscapes · GTA5 → Cityscapes | H-Score: 43.79 -> 57.57 | Helped |
| Add OpenReMix (ablation) | Head-expansion baseline on GTA5 → Cityscapes · GTA5 → Cityscapes | H-Score: 43.79 -> 58.34 | Helped |
| Add both DECON loss and OpenReMix (ablation) | Head-expansion baseline on GTA5 → Cityscapes · GTA5 → Cityscapes | H-Score: 43.79 -> 62.81 | Helped |
Limitations
- Primarily based on pseudo-labeling, so poorly calibrated models may fail to assign private pixels to the unknown class.
- Evaluated only on two synthetic-to-real scenarios (GTA5 and SYNTHIA to Cityscapes) with a single unknown class merged across six or fewer thing classes.
Experimental setup
Datasets: GTA5 SYNTHIA Cityscapes
Benchmarks: GTA5 to Cityscapes SYNTHIA to Cityscapes
Metrics: mIoU H-Score common mIoU private IoU
Baselines: OSBP UAN UniOT AdaptSegNet PixMatch DAFormer HRDA MIC
Builds on: DAFormer MIC ClassMix DACS OSBP
Closest papers by meaning
- Universal Domain Adaptation for Semantic Segmentation (CVPR 2025 · similarity 0.93)
- Source-Free Domain Adaptation for Semantic Segmentation (CVPR 2021 · similarity 0.92)
- Discover, Hallucinate, and Adapt: Open Compound Domain Adaptation for Semantic Segmentation (NeurIPS 2020 · similarity 0.91)
- Balanced Learning for Domain Adaptive Semantic Segmentation (ICML 2025 · similarity 0.91)
- Towards Unsupervised Domain Bridging via Image Degradation in Semantic Segmentation (NeurIPS 2025 · similarity 0.90)
- SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation (CVPR 2025 · similarity 0.89)
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.