Research map / Label-efficient and robust learning / Domain adaptation and generalization
On the Road to Online Adaptation for Semantic Image 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 the selected moderate domain-randomization level rather than ERM training (ablation) | DeepLab-V2 pretraining · SYNTHIA, ACDC, Cityscapes A.W., and Cityscapes O. | mIoU: 35.9 ± 2.5 -> 39.8 ± 3.0; 29.5 ± 2.5 -> 33.6 ± 2.5; 35.6 ± 1.9 -> 38.3 ± 2.6; 40.3 ± 0.9 -> 45.2 ± 1.0 | Helped |
| Increase randomization from DR↑ to DR↑↑ (ablation) | DeepLab-V2 pretraining · SYNTHIA, ACDC, Cityscapes A.W., and Cityscapes O. | mIoU: 34.3 ± 3.3 -> 39.8 ± 3.0; 29.5 ± 2.4 -> 33.6 ± 2.5; 36.2 ± 2.3 -> 38.3 ± 2.6; 41.2 ± 1.0 -> 45.2 ± 1.0 | Helped |
| Increase randomization further from DR↑↑ to DR↑↑↑ (ablation) | DeepLab-V2 pretraining · SYNTHIA, ACDC, Cityscapes A.W., and Cityscapes O. | mIoU: 39.8 ± 3.0 -> 31.9 ± 3.0; 33.6 ± 2.5 -> 26.7 ± 2.3; 38.3 ± 2.6 -> 33.2 ± 2.5; 45.2 ± 1.0 -> 37.7 ± 1.1 | Hurt |
| Adapt BN statistics per frame | DR↑↑ DeepLab-V2 · SYNTHIA, ACDC, Cityscapes A.W., and Cityscapes O. | mIoU improvement over NA: +2.7 %; + 2.4 %; + 1.9 %; + 1.2 % | Helped |
| Continuously adapt with pseudo-labels rather than restarting from the pretrained model | DeepLab-V2 · SYNTHIA, ACDC, Cityscapes A.W., and Cityscapes O. | mIoU improvement over NA: -19.9 %; -11.7 %; -9.4 %; -17.4 % | Hurt |
| Add source-data regularization to continual pseudo-label adaptation (ablation) | C-PL · SYNTHIA, ACDC, Cityscapes A.W., and Cityscapes O. | mIoU improvement over NA: -19.9 % -> +4.9 %; -11.7 % -> +2.8 %; -9.4 % -> +3.9 %; -17.4 % -> +0.5 % | Helped |
| Add a class-count-based reset to continual pseudo-label adaptation (ablation) | C-PL · SYNTHIA, ACDC, Cityscapes A.W., and Cityscapes O. | mIoU improvement over NA: -19.9 % -> +7.2 %; -11.7 % -> +8.2 % ±3.4; -9.4 % -> +9.0 % ±5.1; -17.4 % -> +0.0 % | Helped |
| Add a class-count-based reset to continual TENT (ablation) | C-TENT · SYNTHIA, ACDC, Cityscapes A.W., and Cityscapes O. | mIoU improvement over NA: +2.5 % -> +8.3 % ±4.2; +2.7 % -> +7.3 % ±3.9; +6.4 % -> +9.1 % ±4.9; -0.9 % -> +0.9 % ±1.3 | Helped |
| Use single-image style transfer from a random source image | DR↑↑ DeepLab-V2 · ACDC | mIoU improvement over NA: -7.4 % | Hurt |
| Increase N-PL adaptation iterations from 1 to 5 (ablation) | N-PL · SYNTHIA, ACDC, and Cityscapes A.W. | mIoU improvement over NA: + 3.5 % ± 1.0 -> + 9.5 % ± 3.7; + 2.9 % ± 0.6 -> + 8.4 % ± 2.7; + 2.4 % ± 1.0 -> + 6.5 % ± 3.7 | Helped |
| Increase N-TENT iterations from 1 to 5 at learning rate 1.0 (ablation) | N-TENT · SYNTHIA, ACDC, Cityscapes A.W., and Cityscapes O. | mIoU improvement over NA: + 8.5 % ± 3.1 -> + 6.8 % ± 4.6; + 4.9 % ± 2.0 -> + 2.9 % ± 3.5; + 3.1 % ± 3.6 -> + 2.8 % ± 5.9; -1.2 % ± 0.7 -> -4.5 % ± | Mixed |
| Increase N-TENT iterations from 1 to 5 while reducing learning rate to 0.1 (ablation) | N-TENT · SYNTHIA, ACDC, and Cityscapes A.W. | mIoU improvement over NA: + 4.0 % ± 1.1 -> + 10.7 % ± 3.8; + 3.1 % ± 0.7 -> + 8.5 % ± 2.9; + 2.6 % ± 1.2 -> + 6.4 % ± 4.2 | Helped |
Closest papers by meaning
- Unsupervised Model Adaptation for Continual Semantic Segmentation (AAAI 2021 · similarity 0.93)
- Towards Better Stability and Adaptability: Improve Online Self-Training for Model Adaptation in Semantic Segmentation (CVPR 2023 · similarity 0.92)
- Learning to Adapt via Latent Domains for Adaptive Semantic Segmentation (NeurIPS 2021 · similarity 0.92)
- Universal Domain Adaptation for Semantic Segmentation (CVPR 2025 · similarity 0.90)
- Open-Set Domain Adaptation for Semantic Segmentation (CVPR 2024 · 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.