Research map / Label-efficient and robust learning / Domain adaptation and generalization
D2ADA: Dynamic Density-Aware Active Domain Adaptation for Semantic Segmentation
Tsung-Han Wu, Yi-Syuan Liou, Shao-Ji Yuan, Hsin-Ying Lee, Tung-I Chen, Kuan-Chih Huang, Winston H. Hsu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 density-aware selection instead of uncertainty-only selection at the 1% label budget (ablation) | D2ADA · GTA5->Cityscapes val, DeepLabV3+ | mIoU: 59.97 -> 62.75 at 1%; 70.70 -> 69.82 at 5% | Mixed |
| add class-balanced selection to density-aware selection (ablation) | D2ADA · GTA5->Cityscapes val, DeepLabV3+ | mIoU: 62.75 -> 64.03 (1%), 68.81 -> 69.38 (3%), 69.82 -> 70.69 (5%) | Improved |
| use dynamic scheduling instead of equal budget split between density and uncertainty (ablation) | D2ADA · GTA5->Cityscapes val, DeepLabV3+ | mIoU: 63.30 -> 64.03 (1%), 69.66 -> 69.86 (3%), 71.15 -> 71.25 (5%) | Improved |
| add dynamic scheduling with density-aware selection to uncertainty-only selection (ablation) | uncertainty-only active learning · GTA5->Cityscapes val, DeepLabV3+ | mIoU: 59.97 -> 64.03 (1%), 70.70 -> 71.25 (5%) | Improved |
| add uncertainty selection with dynamic scheduling to density plus class-balanced selection (ablation) | D2ADA · GTA5->Cityscapes val, DeepLabV3+ | mIoU: 64.03 -> 64.03 (1%), 69.38 -> 69.86 (3%), 70.69 -> 71.25 (5%) | Improved |
| use D2ADA instead of uncertainty-based selection (ENT, CONF, MAR) in the first round | active learning on GTA5->Cityscapes · GTA5->Cityscapes val | mIoU: over 3% mIoU | Improved |
| use AADA domain-discriminator selection instead of random selection in early rounds | active domain adaptation · GTA5->Cityscapes and SYNTHIA->Cityscapes val | mIoU | Hurt |
| use D2ADA with 2% labels instead of CLUE with 5% labels | active domain adaptation · GTA5->Cityscapes and SYNTHIA->Cityscapes val | mIoU: 2% vs 5% annotations | Improved |
| use DeepLabV3+ instead of DeepLabV2 as segmentation network for D2ADA | D2ADA · GTA5->Cityscapes val, 5% labels | mIoU: 69.7 -> 71.3 | Improved |
Closest papers by meaning
- Density Matters: Improved Core-Set for Active Domain Adaptive Segmentation (AAAI 2024 · similarity 0.93)
- Generalize Then Adapt: Source-Free Domain Adaptive Semantic Segmentation (ICCV 2021 · similarity 0.91)
- On the Road to Online Adaptation for Semantic Image Segmentation (CVPR 2022 · similarity 0.91)
- Pulling Target to Source: A New Perspective on Domain Adaptive Semantic Segmentation (IJCV 2024 · similarity 0.91)
- Multi-Anchor Active Domain Adaptation for Semantic Segmentation (ICCV 2021 · similarity 0.91)
- Balanced Learning for Domain Adaptive Semantic Segmentation (ICML 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.
