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

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Data scarcity, privacy, and security challenges limit data-driven AI.
• Few-shot semantic segmentation (FSS) methods struggle with cross-domain shifts.
• Cross-domain FSS (CD-FSS) aims to segment novel classes across domains with minimal supervision.

Purpose of the Study:

• To address limitations in feature representation and transmission for CD-FSS.
• To overcome the challenges posed by domain shifts in CD-FSS.
• To propose a novel network architecture for effective CD-FSS.

Main Methods:

• Analysis of existing CD-FSS networks to identify limitations of convolutional inductive bias and single attention mechanisms.
• Introduction of the "focus on your focus" (FOYF) network, integrating multiple attention specialization modules based on Swin Transformer.
• Hierarchical fusion of multi-granularity feature patterns and dynamic activation of cross-scene correspondence links.

Main Results:

• FOYF demonstrates superior performance on CD-FSS benchmarks.
• The proposed method effectively overcomes domain shift bottlenecks by propagating meta-learned scene-level priors.
• Experimental validation confirms the network's ability to mine intra-object attributes and inter-object associations.

Conclusions:

• The FOYF network offers a novel technological pathway for cost-effective learning in CD-FSS.
• This work advances the development of CD-FSS in data-constrained environments.
• Specialized attention mechanisms are crucial for handling complex cross-scene information interactions in CD-FSS.