Research map / Label-efficient and robust learning / Domain adaptation and generalization
Contrastive Model Adaptation for Cross-Condition Robustness in Semantic Segmentation
David Brüggemann, Christos Sakaridis, Tim Broedermann, Luc Van Gool
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 cross-domain contrastive (CDC) loss with reference images (ablation) | CMA · ACDC val | mIoU: 60.1 -> 67.2 | Improved |
| global average pooled embeddings in contrastive loss (ablation) | CMA · ACDC val | mIoU: +4.6% mIoU | Improved |
| patch-level grouping on 7x7 grid with reference warping instead of global embeddings (ablation) | CMA · ACDC val | mIoU: +1.2% mIoU | Improved |
| use patches without spatial warping | CMA · ACDC val | mIoU | Hurt |
| add confidence modulation to patch forming (ablation) | CMA · ACDC val | mIoU: +1% mIoU | Improved |
| add reference images naively (mixed or curriculum) instead of contrastive loss (ablation) | URMA / CMA w/o CDC · ACDC val | mIoU: 63.2 -> 62.9 (URMA mixed); 60.1 -> 60.3 (CMA mixed) | No clear effect |
| use contrastive reference-image loss vs best naive strategy | CMA · ACDC val | mIoU: 64.1 -> 67.2 | Improved |
| swap InfoNCE for RINCE or debiased loss (ablation) | CMA · ACDC val | mIoU: 67.2 -> 67.4 (RINCE); 66.3 (debiased) | No clear effect |
| add self-training loss (ablation) | CMA · ACDC val | mIoU: 57.7 -> 67.2 | Improved |
| add entropy loss (ablation) | CMA · ACDC val | mIoU: 66.7 -> 67.2 | Improved |
| vary grid size, temperature, confidence threshold (ablation) | CMA · ACDC val | mIoU: 66.7 to 67.2 across thresholds | No clear effect |
| adapt with CMA instead of source model on DeepLabv2 vs SegFormer | CMA · Cityscapes to ACDC test | mIoU: SegFormer 59.4 -> 69.1; DeepLabv2 39.5 -> 50.4 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- TENT — Tent: Fully Test-Time Adaptation by Entropy Minimization (ICLR 2021) · better on 2 of 2 shared results
- HRDA — HRDA: Context-Aware High-Resolution Domain-Adaptive Semantic Segmentation (ECCV 2022) · better on 1 of 1 shared result
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
- Bi-directional Contrastive Learning for Domain Adaptive Semantic Segmentation (ECCV 2022 · similarity 0.91)
- Condition-Invariant Semantic Segmentation (TPAMI 2025 · similarity 0.90)
- Prototypical Contrast Adaptation for Domain Adaptive Semantic Segmentation (ECCV 2022 · similarity 0.90)
- Uncertainty Reduction for Model Adaptation in Semantic Segmentation (CVPR 2021 · similarity 0.89)
- Unsupervised Model Adaptation for Continual Semantic Segmentation (AAAI 2021 · similarity 0.89)
- Universal Domain Adaptation for Semantic Segmentation (CVPR 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.
