Research map / Label-efficient and robust learning / Domain adaptation and generalization
Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic 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 frequency masker (APM plus ACPA) inserted into a frozen source-trained FSS backbone to reduce inter-channel correlation and improve cross-domain few-shot segmentation on four target datasets.
Problem
Few-shot segmentation methods degrade sharply when transferred from a source domain (PASCAL-5i) to distant target domains (FSS-1000, Deepglobe, ISIC2018, Chest X-ray) in cross-domain few-shot segmentation.
Approach
Filtering frequency components (amplitude and phase) of target-domain features, learned adaptively on the support set, reduces inter-channel correlation and channel bias, enlarging activated regions; an amplitude-phase masker and a phase-driven channel attention (ACPA) implement this with very few extra parameters, without source-domain retraining.
Key results
- APM-M with ACPA reaches 60.03 average 1-shot mIoU versus 47.68 for the HSNet baseline and 56.06 for PATNet.
- APM-S adds only about 0.01% extra parameters and raises average 1-shot mIoU by about 10 points (47.68 to 57.77).
- Chest X-ray 1-shot mIoU improves from 47.34 (baseline) to 78.25 with APM-M and ACPA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add APM-S to the HSNet baseline (ablation) | HSNet* · CD-FSS average, 1-shot | mIoU: 47.68 -> 57.77 | Helped |
| Add APM-S to the HSNet baseline (ablation) | HSNet* · CD-FSS average, 5-shot | mIoU: 50.10 -> 61.39 | Helped |
| Add ACPA to APM-S (ablation) | APM-S · CD-FSS average, 1-shot | mIoU: +1.16% | Helped |
| Add ACPA to APM-S (ablation) | APM-S · CD-FSS average, 5-shot | mIoU: +1.47% | Helped |
| Use APM-M instead of APM-S, with ACPA (ablation) | Frequency masker · CD-FSS average, 1-shot | mIoU: 58.93 -> 60.03 | Helped |
| Use APM-M instead of APM-S, with ACPA (ablation) | Frequency masker · CD-FSS average, 5-shot | mIoU: 62.86 -> 65.18 | Helped |
| Apply coarse high-low frequency filtering at the input level | Source-trained model · FSS-1000, 1-shot | mIoU: 77.54 -> 64.34 | Hurt |
| Apply frequency filtering at the input level | Source-trained model · Deepglobe, 1-shot | mIoU: 33.19 -> 40.42 | Helped |
| Apply frequency filtering at the input level | Source-trained model · ISIC, 1-shot | mIoU: 32.65 -> 33.16 | Helped |
| Apply low-frequency amplitude filtering at the input level | Source-trained model · Chest X-ray, 1-shot | mIoU: 47.34 -> 61.58 | Helped |
Surprising findings
- Simply filtering frequency components of target-domain inputs at test time can improve mIoU, sometimes by up to 14%, on a model trained only on the source domain.
- Masking some frequency components degrades FSS-1000 performance, so the effect is domain-dependent.
Limitations
- Performance on FSS-1000 is not improved by the coarse frequency masking, and the finer APM is needed to handle it.
- Mask learning depends on the support set and per-task hyperparameters (learning rates differ per dataset), which may be sensitive to tuning.
Experimental setup
Datasets: PASCAL-5i SDS annotations FSS-1000 Deepglobe ISIC2018 Chest X-ray
Benchmarks: FSS-1000 Deepglobe ISIC2018 Chest X-ray
Metrics: mIoU mutual information CKA
Baselines: PGNet PANet CaNet RPMMs PFENet RePRI HSNet PATNet FPTrans PerSAM DFF GFNet
Builds on: HSNet SENet DFF DAC
Closest papers by meaning
- The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.90)
- Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.90)
- Dual-Agent Optimization framework for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.90)
- Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation (ICML 2025 · similarity 0.89)
- Self-Disentanglement and Re-Composition for Cross-Domain Few-Shot Segmentation (ICML 2025 · similarity 0.89)
- Domain-Rectifying Adapter for Cross-Domain Few-Shot Segmentation (CVPR 2024 · 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.