Research map / Label-efficient and robust learning / Domain adaptation and generalization
Attention Decomposition for Cross-Domain Semantic Segmentation
Liqiang He, Sinisa Todorovic
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 |
|---|---|---|---|
| add pixel-wise segmentation loss Lseg (ablation) | Mask2Former baseline · GTA to Cityscapes | mIoU: 44.1 -> 51.4 | Improved |
| add domain-specific (DS) branch (ablation) | Mask2Former + Lseg · GTA to Cityscapes | mIoU: 51.4 -> 56.6 | Improved |
| add attention decomposition (DI cross-attention) without GRL (ablation) | DS branch model · GTA to Cityscapes | mIoU: 56.6 -> 57.5 | Improved |
| add GRL to DS branch (without AD) (ablation) | DS branch model · GTA to Cityscapes | mIoU: 56.6 -> 63.1 | Improved |
| combine attention decomposition with GRL (ablation) | ADFormer · GTA to Cityscapes | mIoU: 63.1 -> 69.2 | Improved |
| add thing-class ImageNet feature distance regularization (ablation) | ADFormer · GTA to Cityscapes | mIoU: 69.2 -> 66.3 | Hurt |
| increase GRL strength to 1e-2 from 1e-4 (ablation) | ADFormer · GTA to Cityscapes | mIoU: 69.2 -> 63.4 | Hurt |
| use larger Swin-B backbone instead of Swin-S (ablation) | ADFormer · GTA to Cityscapes | mIoU: 69.2 -> 70.4; params 64.3M -> 102.5M | Improved |
| increase number of decoder layers from 1+1 to 1+9 (ablation) | ADFormer · GTA to Cityscapes | mIoU: 66.7 -> 69.7 | Improved |
| vary loss weights lambda_set and lambda_seg away from 0.025 and 10.0 (ablation) | ADFormer · GTA to Cityscapes | mIoU: 69.2 -> 58.7 at worst | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- DAFormer — DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentation (CVPR 2022) · better on 4 of 4 shared results
- HRDA — HRDA: Context-Aware High-Resolution Domain-Adaptive Semantic Segmentation (ECCV 2022) · better on 0 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Rethinking Decoders for Transformer-based Semantic Segmentation: A Compression Perspective (NeurIPS 2024 · similarity 0.89)
- Segmenter: Transformer for Semantic Segmentation (ICCV 2021 · similarity 0.88)
- CDAC: Cross-domain Attention Consistency in Transformer for Domain Adaptive Semantic Segmentation (ICCV 2023 · similarity 0.88)
- Classes Matter: A Fine-grained Adversarial Approach to Cross-domain Semantic Segmentation (ECCV 2020 · similarity 0.88)
- Learning Content-Enhanced Mask Transformer for Domain Generalized Urban-Scene Segmentation (AAAI 2024 · similarity 0.88)
- 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.
