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

Self-Disentanglement and Re-Composition for Cross-Domain Few-Shot Segmentation

ICML 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Cross-domain few-shot with transformers · 0 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.

Proposes SDRC, which self-disentangles ViT layer features with an orthogonal-space decoupling module and re-composes them via cross-pattern comparisons with learned fusion weights, reaching new state-of-the-art on cross-domain few-shot segmentation.

Problem

Cross-domain few-shot segmentation relies on distance comparison of encoder features, and feature entanglement binds source-domain patterns together so they transfer poorly to unseen target domains.

Approach

Decompose the ViT output into per-layer components (natural residual decomposition), decouple them with an orthogonal-regularized low-rank module, then compare components across layers (cross-pattern comparison) and learn re-composition weights, with the weights adapted only during target fine-tuning.

Key results

• Achieves 63.22% average mIoU in 1-shot and 67.30% in 5-shot on the CD-FSS benchmark, improving over the prior best APSeg/DRA by about 1.92 and 1.88 points.
• Cross-pattern comparison alone raises 1-shot average mIoU from 49.88 to 59.50 in the ablation.
• Adaptive fusion weights trained on the source domain perform worse (61.01) than when learned only on the target (63.22) in 1-shot.

Measured effects reported in the paper

Surprising findings

• Simply shifting cross-matched ViT layers yields higher domain similarity (CKA) than layer-wise matching, suggesting learnable cross-layer matching is preferable.
• Training the AFW weights on the source domain hurts performance, as they become biased toward source data.

Limitations

• Relies on the assumption that ViT layers share a feature space via residual connections; Swin requires extra mapping layers.
• Multi-background prototype gains (63.22 to 63.59) are slight and not worth the clustering overhead, per the authors.

Experimental setup

Datasets: PASCAL VOC 2012 SBD FSS-1000 DeepGlobe ISIC2018 Chest X-ray

Benchmarks: FSS-1000 DeepGlobe ISIC2018 Chest X-ray

Metrics: mIoU

Baselines: PANet RPMMs PFENet RePRI HSNet PATNet PerSAM APM ABCDFSS DRA APSeg

Builds on: PATNet ViT mask average pooling prototypes CKA similarity

Closest papers by meaning

• Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.89)
• Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic Segmentation (NeurIPS 2024 · similarity 0.89)
• Dual-Agent Optimization framework for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.89)
• The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation (CVPR 2025 · similarity 0.89)
• Cross-Domain Few-Shot Segmentation via Iterative Support-Query Correspondence Mining (CVPR 2024 · similarity 0.87)
• Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation (ICML 2025 · similarity 0.87)