Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Adapt Before Comparison: A New Perspective on 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 ABCDFSS, which drops source-domain segmentation training and instead adapts a frozen ImageNet ResNet50 at test time with small per-layer adapters trained via consistency and contrastive losses on the support and query, then compares features densely.
Problem
Cross-domain few-shot segmentation (CD-FSS), where a model trained on a source domain (PASCAL VOC/COCO) must segment novel classes in very different target domains from a few labeled support images.
Approach
Replace the learned downstream segmentation network with test-time task adaptation: small adapters appended to each ResNet bottleneck are trained from scratch on the support and query using augmentation-based dense contrastive, statistics-matching, and prototype contrastive losses; the adapted multi-layer correlation maps are then averaged, thresholded via k-means on the histogram, and refined wit
Key results
- Achieves a CD-FSS benchmark average of 60.7 mIoU in 1-shot and 65.0 in 5-shot, surpassing prior SOTA PATNet by 4.6 (1-shot) and 3.0 (5-shot) mIoU.
- Reaches 35.1 mIoU on SUIM 1-shot, slightly above PMNet (34.8) and RtD (34.7).
- Without task adaptation (w/o TA), average 1-shot mIoU drops by 5.86, showing test-time adaptation is key.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adapt backbone features at test time before comparing query and support | ABCDFSS · CD-FSS benchmark average, 1-shot | mIoU: +4.6 | Helped |
| adapt backbone features at test time before comparing query and support | ABCDFSS · CD-FSS benchmark average, 5-shot | mIoU: +3.0 | Helped |
| use task adaptation before dense comparison (ablation) | ABCDFSS · CD-FSS average 1-shot | mIoU: +5.86 | Helped |
| use task adaptation before dense comparison (ablation) | ABCDFSS · CD-FSS average 1-shot | FB-IoU: +3.19 | Helped |
| add unsupervised consistency losses Lq and Ls (ablation) | ABCDFSS · CD-FSS average 1-shot | mIoU: +2.06 | Helped |
| add unsupervised consistency losses Lq and Ls (ablation) | ABCDFSS · CD-FSS average 1-shot | FB-IoU: +1.21 | Helped |
| add support-supervised class-alignment loss Lp (ablation) | ABCDFSS · CD-FSS average 1-shot | mIoU: +3.00 | Helped |
| add support-supervised class-alignment loss Lp (ablation) | ABCDFSS · CD-FSS average 1-shot | FB-IoU: +2.78 | Helped |
| apply post-processing refinement (ablation) | ABCDFSS · CD-FSS average 1-shot | mIoU: 58.3 -> 60.7 | Helped |
| reuse one adapted task model for subsequent query images instead of adapting per episode (ablation) | ABCDFSS · FSS-1000 | mIoU: -0.71 | Hurt |
| reuse one adapted task model for subsequent query images instead of adapting per episode (ablation) | ABCDFSS · SUIM | FB-IoU: -2.87 | Hurt |
| replace level-wise adapters with one adapter on concatenated hypercolumn features (ablation) | ABCDFSS · CD-FSS average 1-shot | mIoU: −12.6 | Hurt |
Surprising findings
- Fine-tuning on the support set does not overfit with the proposed consistency-based losses; query discriminability improves as well.
- A naive predictor labeling all pixels as foreground scores 43.0 mIoU on Deepglobe, exceeding previous SOTA mIoU, so FB-IoU should be reported.
Limitations
- Underperforms on FSS-1000 (about 10 mIoU below PMNet) because it learns no global semantic context or spatial decoder.
- Adapts features vector-wise without scene-level semantic context; relies on heuristic fusion and post-processing.
- Deepglobe ground-truth annotations are noisy, limiting benchmark expressiveness.
Experimental setup
Datasets: Deepglobe ISIC Chest X-ray (Lung) FSS-1000 SUIM
Benchmarks: CD-FSS benchmark (Deepglobe, ISIC, Chest X-ray, FSS-1000) SUIM
Metrics: mIoU FB-IoU
Baselines: PATNet PMNet HSNet RePRI HDMNet RestNet PANet RtD SCL LinearResNet LinearDeeplab
Builds on: PATNet PMNet HSNet RePRI SimCLR-style dense contrastive learning (PixPro/DenseCL)
Closest papers by meaning
- The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.93)
- Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation (ICML 2025 · similarity 0.93)
- Dual-Agent Optimization framework for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.92)
- Domain-Rectifying Adapter for Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.92)
- Cross-Domain Few-Shot Segmentation via Iterative Support-Query Correspondence Mining (CVPR 2024 · similarity 0.91)
- Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic Segmentation (NeurIPS 2024 · similarity 0.90)
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.