Research map / Label-efficient and robust learning / Domain adaptation and generalization
Domain-Rectifying Adapter for Cross-Domain Few-Shot Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a lightweight domain-rectifying adapter, trained on local-global style-perturbed VOC features with a cyclic alignment loss, that maps target-domain features back to the source domain before a frozen-style SSP few-shot segmenter.
Problem
Few-shot segmentation models degrade under domain shift to unseen target domains (cross-domain few-shot segmentation), and full-model adaptation overfits with scarce few-shot data.
Approach
Instead of adapting the whole network, train a compact adapter that predicts channel-wise rectification factors to map perturbed features from simulated target styles back to the source domain; style perturbations combine per-image (local) and dataset-level momentum-averaged (global) channel statistics, and a cyclic alignment loss enforces reverse rectification.
Key results
- Average 1-shot mIoU across the four CD-FSS datasets reaches 60.86 with the method versus 57.20 for the SSP baseline.
- On Chest X-ray, 1-shot and 5-shot results exceed PATNet by 15.74% and 12.11%.
- On SUIM (RD setting), average 1-way 1-shot mIoU is 40.25 versus 34.70 for RD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the domain-rectifying adapter method | SSP, ResNet-50 · CD-FSS average 1-shot | mIoU: 57.20 -> 60.86 | Helped |
| Add the domain-rectifying adapter method | SSP, ResNet-50 · CD-FSS average 5-shot | mIoU: 63.92 -> 65.42 | Helped |
| Add the domain-rectifying adapter method | SSP, ResNet-50 · Chest X-ray 1-shot | mIoU: 74.23 -> 82.35 | Helped |
| Add the domain-rectifying adapter method | SSP, ResNet-50 · ISIC 1-shot | mIoU: 35.09 -> 40.77 | Mixed |
| Add the domain-rectifying adapter method | SSP, ResNet-50 · FSS-1000 5-shot | mIoU: 80.56 -> 80.40 | Hurt |
| Add the feature perturbation module (ablation) | SSP baseline · CD-FSS 1-shot average | mIoU: 57.20 -> 58.45 | Helped |
| Add the rectification module (ablation) | SSP baseline · CD-FSS 1-shot average | mIoU: 57.20 -> 57.80 | Helped |
| Add perturbation and rectification modules together (ablation) | SSP baseline · CD-FSS 1-shot average | mIoU: 57.20 -> 59.17 | Helped |
| Add cyclic alignment loss to perturbation and rectification (ablation) | SSP, ResNet-50 · CD-FSS 1-shot average | mIoU: 59.17 -> 60.86 | Helped |
| Add cyclic loss to BCE loss (ablation) | SSP, ResNet-50 · CD-FSS 1-shot average | mIoU: 57.65 -> 58.56 | Helped |
| Add alignment loss to BCE loss (ablation) | SSP, ResNet-50 · CD-FSS 1-shot average | mIoU: 57.65 -> 59.12 | Helped |
| Combine local and global style perturbation (ablation) | SSP, ResNet-50 · CD-FSS 1-shot average | mIoU: 59.81 -> 60.86 | Helped |
Surprising findings
- Solely using feature perturbation degrades the model to a domain-generalization approach similar to NP, without the rectification module.
- Noise type is not essential: Beta and Uniform noise perform nearly as well as Gaussian.
Limitations
- Gains on FSS-1000 are marginal because the domain gap to PASCAL VOC is small.
- Local perturbation degrades at higher noise levels, requiring a separate global perturbation for stability.
Experimental setup
Datasets: PASCAL VOC 2012 SBD FSS-1000 DeepGlobe ISIC2018 Chest X-ray SUIM
Benchmarks: FSS-1000 DeepGlobe ISIC2018 Chest X-ray SUIM
Metrics: mean IoU
Baselines: PATNet SSP RD PerSAM MixStyle DSU NP AdaIN FPTrans HSNet
Builds on: SSP AdaIN MixStyle DSU PATNet
Closest papers by meaning
- Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation (ICML 2025 · similarity 0.93)
- Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.92)
- Dual-Agent Optimization framework for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.89)
- The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.89)
- Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic Segmentation (NeurIPS 2024 · 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.