Focus on Your Focus (FOYF): Cross-Domain Few-Shot Semantic Segmentation by Attention Specialization

Due to data scarcity, privacy, and security, it is difficult to obtain sufficient data to support the research and application of data-driven artificial intelligence (AI) technology across various fields. Although current few-shot semantic segmentation (FSS) methods mitigate performance degradation caused by limited same-domain data, they struggle in cross-domain scenarios with significant domain shifts, limiting their practical utility. Building upon FSS, cross-domain FSS (CD-FSS) generalizes the task to segment novel classes across domains using minimal supervision. However, CD-FSS faces persistent challenges in effective feature pattern representation and transmission across domains. This article comprehensively analyzes existing CD-FSS networks and reveals two crucial insights. First, the inherent inductive bias of the convolutional mode restricts the scene-level contextual information capture. Second, reliance on a single attention inadequately supports complex cross-scene information interaction. To address these issues, we propose an attention specialization network, called "focus on your focus" (FOYF). It integrates multiple attention specialization modules based on Swin Transformer, which profoundly mine intraobject attributes and interobject associations within scenes, dynamically activate cross-scene correspondence links, and hierarchically fuse multigranularity feature patterns. Experimental validation demonstrates the superior performance of FOYF on CD-FSS benchmarks by propagating metalearned scene-level priors, effectively overcoming domain shift bottlenecks. This work establishes a novel technological pathway for cost-effective learning, advancing the development of CD-FSS in data-constrained environments.