Research map / Label-efficient and robust learning / Domain adaptation and generalization
Dual-Agent Optimization framework 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 DATO, a dual-agent framework for cross-domain few-shot segmentation that uses adversarially trained frequency-domain agents to make features domain-invariant and agent-based correlations to rectify support-query matching.
Problem
Cross-domain few-shot segmentation suffers performance drops because of feature domain sensitivity and support-to-query matching domain sensitivity when a model trained on a source domain is transferred to unseen target domains.
Approach
Jointly optimizes features in the frequency domain via adversarially trained agents (perturbing low-frequency components with ALOFT) and rectifies matching in the spatial domain by correlating pixels with the agent-aggregated domain-invariant feature set, so the two procedures are coupled through the shared agents.
Key results
- Achieves 70.31% 1-shot and 73.82% 5-shot average mIoU on the CD-FSS benchmark with ResNet-50, surpassing IFA by 2.51% and 2.42%.
- Reaches 76.49% (1-shot) and 78.27% (5-shot) FB-IoU, outperforming the previous best IFA by 3.11% and 2.06%.
- Ablation: CMA adds 5.45% 1-shot mIoU over the 60.00% baseline; CRS adds a further 2.6% 1-shot; plugging into FPTrans raises 1-shot average mIoU from 56.69 to 63.13.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DATO instead of DMTNet | ResNet-50 CD-FSS segmentation · Chest X-ray 1-shot and 5-shot | mIoU: 73.74 -> 79.58; 77.30 -> 81.07 | Helped |
| use DATO instead of IFA | ResNet-50 CD-FSS segmentation · ISIC, FSS-1000, and Deepglobe (1-shot and 5-shot) | mIoU: 66.30 -> 68.76; 69.80 -> 70.30; 80.10 -> 81.79; 82.40 -> 84.61; 50.60 -> 51.10; 58.80 -> 59.30 | Helped |
| use DATO instead of IFA | ResNet-50 CD-FSS segmentation · CD-FSS average, 1-shot and 5-shot | mIoU: 67.80 -> 70.31; 71.40 -> 73.82 | Helped |
| add agent enhancement (ablation) | DATO · CD-FSS benchmark, 1-shot | mIoU: 60.00 -> 65.45 | Helped |
| add agent enhancement (ablation) | DATO · CD-FSS benchmark, 5-shot | mIoU: 68.00 -> 70.24 | Helped |
| add diversity loss (ablation) | DATO with CMA enhancement · CD-FSS benchmark, 1-shot | mIoU: 65.45 -> 67.71 | Helped |
| add diversity loss (ablation) | DATO with CMA enhancement · CD-FSS benchmark, 5-shot | mIoU: 70.24 -> 71.31 | Helped |
| add correlation rectification strategy (ablation) | DATO with CMA and diversity loss · CD-FSS benchmark, 1-shot | mIoU: 67.71 -> 70.31 | Helped |
| add correlation rectification strategy (ablation) | DATO with CMA and diversity loss · CD-FSS benchmark, 5-shot | mIoU: 71.31 -> 73.82 | Helped |
| vary the number of agents from 20 to 60 (ablation) | DATO · CD-FSS benchmark, 1-shot | mIoU: 69.93; 69.98; 70.15; 70.17; 70.22; 70.31; 70.23; 70.20; 70.22 | Mixed |
| vary diversity loss weight (ablation) | DATO · CD-FSS benchmark, 1-shot | mIoU: 70.03; 70.11; 70.31; 70.08 | Mixed |
Surprising findings
- Agent-enhanced target-domain support and query features show higher similarity, and the method improves HDMNet and FPTrans substantially, showing architecture-agnostic gains.
Limitations
- Excessive number of agents introduces fragmented disturbances and reduces performance (optimal Na=45).
- Relies on target-domain support-set fine-tuning stage for adaptation.
Experimental setup
Datasets: PASCAL VOC 2012 SBD Chest X-ray ISIC FSS-1000 Deepglobe
Benchmarks: Chest X-ray ISIC FSS-1000 Deepglobe
Metrics: mIoU FB-IoU
Baselines: PATNet RestNet DMTNet APSeg ABCDFSS DR-Adapter APM IFA PerSAM HDMNet FPTrans
Builds on: SSP ALOFT PATNet IFA
Closest papers by meaning
- Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.92)
- The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.90)
- Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic Segmentation (NeurIPS 2024 · similarity 0.90)
- Cross-Domain Few-Shot Segmentation via Iterative Support-Query Correspondence Mining (CVPR 2024 · similarity 0.90)
- Domain-Rectifying Adapter for Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.89)
- Self-Disentanglement and Re-Composition for Cross-Domain Few-Shot 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.