Research map / Label-efficient and robust learning / Domain adaptation and generalization

The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation

CVPR 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Cross-domain few-shot with transformers · 6 citations (OpenAlex)

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.

Explains why cross-domain few-shot segmentation performance peaks early in source training by attributing it to low-level features' sharp loss landscapes, and proposes a random-convolution/FFT sharpness-aware low-level enhancing module plus a test-time low-level calibration module.

Problem

Cross-domain few-shot segmentation (CDFSS): models transferred from PASCAL VOC to distant target domains reach peak performance in very early epochs and then degrade sharply as source-domain training proceeds.

Approach

Low-level features absorb domain-specific information and become sharp-minimum sensitive to domain shift. A Low-level Enhancing Module perturbs low-level support features via random convolution with FFT amplitude/phase recombination for sharpness-aware training, and a training-free Low-level Calibration Module uses top-K confident query patches' low-level features to refine the coarse mask at test

Key results

• Proposed method reaches 60.46% average MIoU in 1-shot with ResNet-50 and 65.01% in 5-shot, versus APSeg's prior state of the art (61.30% 1-shot / 65.09% 5-shot) per the table parsed.
• With ViT-B/16 backbone the method reports 62.43% average 1-shot...
• Ablation on ResNet-50: LEM adds 1.14 average MIoU points and LCM adds 2.11 points in 1-shot over the 57.21 baseline.

Measured effects reported in the paper

Surprising findings

• Fixing shallow layers (stages 1-2) to ImageNet weights gives better target performance than training them, suggesting shallow layers absorb domain-specific information.

Limitations

• Gains on the ViT backbone are reported with an ambiguous table layout in the parsed text.
• Method relies on comparison to a limited set of CDFSS baselines and hyperparameters K, w, beta tuned on target-domain results.

Experimental setup

Datasets: PASCAL VOC 2012 SBD

Benchmarks: FSS-1000 Deepglobe ISIC Chest X-ray

Metrics: MIoU

Baselines: RePRI HSNet SSP FPTrans PerSAM PATNet APM ABCDFSS DR-Adapter APSeg

Builds on: PATNet SSP FPTrans Sharpness-Aware Minimization

Closest papers by meaning

• Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.93)
• Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic Segmentation (NeurIPS 2024 · similarity 0.90)
• 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)
• Cross-Domain Few-Shot Segmentation via Iterative Support-Query Correspondence Mining (CVPR 2024 · similarity 0.89)
• Domain-Rectifying Adapter for Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.89)