Research map / Label-efficient and robust learning / Domain adaptation and generalization
DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentation
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DAFormer instead of the strongest prior method | UDA semantic segmentation · GTA5 → Cityscapes | mIoU: 57.5 -> 68.3 | Helped |
| Use DAFormer instead of the strongest prior method | UDA semantic segmentation · Synthia → Cityscapes | mIoU: 55.5 -> 60.9 | Helped |
| Use a MiT Transformer encoder instead of ResNet-101 (ablation) | SegFormer · GTA → Cityscapes | UDA mIoU: 50.9 -> 58.2 | Helped |
| Add learning-rate warmup (ablation) | SegFormer · GTA → Cityscapes | UDA mIoU: 51.8 -> 58.2 | Helped |
| Add rare class sampling (ablation) | SegFormer · GTA → Cityscapes | UDA mIoU: +5.8 mIoU | Helped |
| Use rare class sampling instead of uniform class-balanced sampling (ablation) | SegFormer · GTA → Cityscapes | UDA mIoU: 62.0 -> 64.0 | Helped |
| Add thing-class ImageNet feature distance (ablation) | SegFormer · GTA → Cityscapes | UDA mIoU: +3.5 mIoU | Helped |
| Restrict feature distance to thing-classes rather than all classes (ablation) | SegFormer · GTA → Cityscapes | UDA mIoU: 58.8 -> 61.7 | Helped |
| Combine RCS and feature distance (ablation) | SegFormer · GTA → Cityscapes | UDA mIoU: 64.0 -> 66.2 | Helped |
| Ignore pseudo-label regions at image top and bottom and increase teacher EMA factor (ablation) | SegFormer with RCS and FD · GTA → Cityscapes | UDA mIoU: 66.2 -> 67.0 | Helped |
| Use context-aware fusion across multiple encoder levels rather than context only at the bottleneck (ablation) | DAFormer decoder · GTA → Cityscapes | UDA mIoU: 67.0 -> 68.3 | Helped |
| Use depthwise separable convolutions in the decoder (ablation) | DAFormer decoder · GTA → Cityscapes | UDA mIoU: 67.0 -> 68.3 | Helped |
Closest papers by meaning
- Universal Domain Adaptation for Semantic Segmentation (CVPR 2025 · similarity 0.90)
- Balanced Learning for Domain Adaptive Semantic Segmentation (ICML 2025 · similarity 0.90)
- Category Anchor-Guided Unsupervised Domain Adaptation for Semantic Segmentation (NeurIPS 2019 · similarity 0.90)
- Towards Unsupervised Domain Bridging via Image Degradation in Semantic Segmentation (NeurIPS 2025 · similarity 0.90)
- Open-Set Domain Adaptation for Semantic Segmentation (CVPR 2024 · similarity 0.89)
- Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation (ICML 2025 · similarity 0.88)
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.