Location Matters: Frequency-Spatial Dual-Space Adaptation for Cross-Domain Few-Shot Segmentation

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Cross-domain few-shot semantic segmentation (CD-FSS) methods often neglect the inherent spatial correspondence between images.
• This spatial correlation, crucial for understanding object structure across domains, is driven by consistent foreground object layouts.

Purpose of the Study:

• To propose a novel frequency-spatial dual space adaptation (FDSA) framework for CD-FSS.
• To exploit the domain-agnostic spatial correspondence prior for improved segmentation accuracy.
• To learn domain-invariant structures and task-specific priors by integrating frequency and spatial domain adaptations.

Main Methods:

• The proposed FDSA framework comprises two modules: Frequency Structural Adapter (FSA) and Spatial Geometry Adapter (SGA).
• FSA modulates images in the frequency domain to preserve structural integrity by emphasizing low-frequency semantics and reducing high-frequency noise.
• SGA utilizes local descriptors to extract keypoints and generate Gaussian-based geometric priors for aligning regions, further enhanced by spatial-guided SAM refinement (SSR) for mask refinement.

Main Results:

• The FDSA framework achieves state-of-the-art performance on four standard CD-FSS benchmarks.
• The integration of frequency and spatial domain adaptations effectively captures domain-invariant structures and task-specific priors.
• Spatial-guided SAM refinement (SSR) enables high-quality segmentation mask refinement without manual intervention.

Conclusions:

• The proposed FDSA framework successfully addresses the limitations of existing CD-FSS methods by exploiting spatial correspondence.
• The dual-space adaptation approach significantly improves the learning of domain-invariant structures and task-specific priors.
• The method demonstrates superior performance and provides a robust solution for cross-domain few-shot semantic segmentation tasks.