Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Cross-Domain Few-Shot Segmentation via Iterative Support-Query Correspondence Mining
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 Iterative Few-shot Adaptor (IFA), a fine-tuning strategy built on bi-directional support-query prediction (BFP) that iteratively mines support-query correspondence to tackle overfitting in cross-domain few-shot segmentation.
Problem
Cross-domain few-shot segmentation (CD-FSS): segmenting novel categories in a distinct target domain with only a few annotated support exemplars, where meta-learned models are biased to the source domain and naive fine-tuning overfits.
Approach
Bi-directional Few-shot Prediction establishes both support-to-query and support-to-query-to-support correspondence, so support masks supervise predictions and augment supervision; Iterative Few-shot Adaptor repeats this recursively in fine-tuning, adding supervision at every iteration to exploit scarce target samples.
Key results
- IFA (T=3) reaches 67.8% average 1-shot mIoU across four CD-FSS targets vs 60.0% for reproduced SSP and 56.1% for PATNet.
- On ISIC, IFA improves 1-shot mIoU by 17.7% over the previous best method; on Deepglobe 1-shot it gains 9.3%.
- On DS-FSS (COCO-20i to Pascal-5i) with Res-101, IFA reaches 79.6% 1-shot mIoU vs 66.6% for Meta-Memory.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add BFP to baseline SSP for source training and fine-tune without query labels (ablation) | CD-FSS with SSP · Deepglobe | mIoU (%): 41.3 -> 44.1 | Helped |
| Add BFP to baseline SSP for source training and fine-tune without query labels (ablation) | CD-FSS with SSP · ISIC | mIoU (%): 48.6 -> 62.3 | Helped |
| Add iterative BFP fine-tuning (IFA T=3) after BFP training (ablation) | CD-FSS with SSP · Deepglobe | mIoU (%): 44.1 -> 50.6 | Helped |
| Add iterative BFP fine-tuning (IFA T=3) after BFP training (ablation) | CD-FSS with SSP · ISIC | mIoU (%): 62.3 -> 66.3 | Helped |
| Use IFA T=3 rather than TAFT fine-tuning | CD-FSS with SSP · Deepglobe | mIoU (%): 43.2 -> 50.6 | Helped |
| Use IFA T=3 rather than TAFT fine-tuning | CD-FSS with SSP · ISIC | mIoU (%): 64.3 -> 66.3 | Helped |
| Use IFA T=3 rather than CPFT fine-tuning | CD-FSS with SSP · Deepglobe | mIoU (%): 44.6 -> 50.6 | Helped |
| Use IFA T=3 rather than CPFT fine-tuning | CD-FSS with SSP · ISIC | mIoU (%): 54.5 -> 66.3 | Helped |
| Increase BFP iteration count from 1 to 3 (ablation) | IFA · Deepglobe | mIoU (%): 46.6 -> 50.6 | Helped |
| Increase BFP iteration count beyond 3 (ablation) | IFA · Deepglobe | mIoU (%): 50.6 -> 50.8; 50.6 -> 50.9 | No clear effect |
| Apply IFA T=3 rather than Meta-Memory | Domain-Shift Few-Shot Segmentation · Pascal-5i 1-shot | mIoU (%): 65.6 -> 71.0 | Helped |
| Apply IFA T=3 rather than Meta-Memory | Domain-Shift Few-Shot Segmentation · Pascal-5i 5-shot | mIoU (%): 70.1 -> 80.9 | Helped |
Surprising findings
- Naive fine-tuning on novel-category exemplars brings only small gains, which the authors attribute to overfitting.
- Performance saturates beyond 3 iterations of BFP, and more iterations may add superfluous information.
Limitations
- Iterating too many times yields saturation or superfluous information
- Relies on a meta-learned source-domain model and few annotated target support images
Experimental setup
Datasets: PASCAL VOC 2012 SBD Deepglobe ISIC2018 Chest X-Ray FSS-1000 COCO-20i Pascal-5i
Benchmarks: Deepglobe ISIC Chest X-Ray FSS-1000 Pascal-5i (DS-FSS)
Metrics: mIoU
Baselines: AMP PATNet PGNet PANet CaNet RPMMs PFENet RePRI HSNet SSP SAM TAFT
Builds on: Self-Support Prototype (SSP) PATNet Task-Adaptive Fine-Tuning (TAFT) Copy-Paste Fine-Tuning
Closest papers by meaning
- Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.91)
- Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation (ICML 2025 · similarity 0.90)
- Dual-Agent Optimization framework for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.90)
- The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.89)
- Domain-Rectifying Adapter for Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.88)
- Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic Segmentation (NeurIPS 2024 · 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.