Research map / Label-efficient and robust learning / Domain adaptation and generalization
Universal 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.
Proposes UniMAP for universal unsupervised domain adaptation for semantic segmentation without known category overlap, using domain-specific prototypes (DSPD) and target-based source image matching (TIM), and introduces OPDA-SS benchmarks.
Problem
Unsupervised domain adaptation for semantic segmentation assumes known source/target category sets; when source-private or target-private classes exist, common classes get low pseudo-label confidence and are misassigned, causing negative transfer.
Approach
Assign two domain-specific prototypes (source and target) per class in a fixed simplex ETF space, use their relative distances to weight pixels likely to be common classes in pseudo-labeling, and pair each target image with the source image having the most target-pseudo-label-weighted common-class pixels.
Key results
- On Pascal-Context to Cityscapes OPDA-SS, UniMAP reaches 60.94 common, 31.27 private and 41.33 H-score, versus 30.11 H-score for BUS.
- On GTA5 to IDD OPDA-SS, UniMAP reaches 45.51 H-score, versus 41.26 for BUS, with private IoU 34.78 vs 29.70.
- Ablation on Pascal-Context to Cityscapes: baseline H-score 36.03; DSPD alone 38.04; TIM alone 38.39; both 41.33.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DSPD to the BUS baseline (ablation) | UniMAP · Pascal-Context → Cityscapes | H-Score: 36.03 -> 38.04 | Helped |
| add TIM to the BUS baseline (ablation) | UniMAP · Pascal-Context → Cityscapes | H-Score: 36.03 -> 38.39 | Helped |
| add DSPD and TIM to the BUS baseline (ablation) | UniMAP · Pascal-Context → Cityscapes | H-Score: 36.03 -> 41.33 | Helped |
| use pixel-wise weight scaling w without Lproto (ablation) | DSPD · Pascal-Context → Cityscapes | H-Score: 36.03 -> 31.08 | Hurt |
| use Lproto without pixel-wise weight scaling w (ablation) | DSPD · Pascal-Context → Cityscapes | H-Score: 36.03 -> 36.96 | Helped |
| combine Lproto with pixel-wise weight scaling w (ablation) | DSPD · Pascal-Context → Cityscapes | H-Score: 36.03 -> 38.04 | Helped |
| replace BUS with UniMAP | OPDA-SS semantic segmentation · Pascal-Context → Cityscapes | H-score: 30.11 -> 41.33 | Helped |
| replace BUS with UniMAP | OPDA-SS semantic segmentation · GTA5 → IDD | H-score: 41.26 -> 45.51 | Helped |
| replace BUS with UniMAP | OPDA-SS semantic segmentation · GTA5 → IDD | Common: 65.47 -> 64.08 | Hurt |
Surprising findings
- Adding source-private classes degrades existing UDA-SS methods, with common classes confused with source-private or target-private classes.
- Using the prototype weight scaling w alone reduces H-score (31.08 vs 36.03 baseline), requiring the prototype loss to be effective.
Limitations
- On GTA5 to IDD, common-class mIoU is lower than BUS (64.08 vs 65.47).
- Closed-set and open-set settings are slightly better for specialized prior methods than UniMAP.
Experimental setup
Datasets: Pascal-Context Cityscapes GTA5 IDD ImageNet-1k
Benchmarks: Pascal-Context to Cityscapes (OPDA-SS) GTA5 to IDD (OPDA-SS) Closed/Open/Partial set DA on Pascal-Context to Cityscapes
Metrics: common mIoU private IoU H-score average common mIoU
Baselines: UAN UniOT MLNet DAFormer HRDA MIC BUS
Builds on: BUS MIC DACS ProtoSeg UniDA in classification
Closest papers by meaning
- Unsupervised Model Adaptation for Continual Semantic Segmentation (AAAI 2021 · similarity 0.94)
- Open-Set Domain Adaptation for Semantic Segmentation (CVPR 2024 · similarity 0.93)
- Balanced Learning for Domain Adaptive Semantic Segmentation (ICML 2025 · similarity 0.92)
- Towards Unsupervised Domain Bridging via Image Degradation in Semantic Segmentation (NeurIPS 2025 · similarity 0.91)
- SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation (CVPR 2025 · similarity 0.91)
- DynAlign: Unsupervised Dynamic Taxonomy Alignment for Cross-Domain Segmentation (ICLR 2025 · similarity 0.90)
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.