Spatio-Semantic Expert Routing Architecture with Mixture-of-Experts for Referring Image Segmentation

Researchers propose SERA, a new architecture for referring image segmentation that uses mixture-of-experts and expression-aware routing to improve pixel-level mask generation from natural language descriptions. The system introduces lightweight expert refinement stages and parameter-efficient tuning that updates less than 1% of backbone parameters while achieving superior performance on spatial localization and boundary delineation tasks.

• →SERA introduces expression-aware expert routing for better referring image segmentation with improved spatial coherence and boundary precision.
• →The architecture uses parameter-efficient tuning that updates only normalization and bias terms, affecting less than 1% of backbone parameters.
• →SERA-Adapter and SERA-Fusion provide complementary refinement at different stages within the vision-language framework.
• →The system addresses limitations of uniform refinement strategies that cause fragmented regions and inaccurate boundaries.
• →Experimental results show consistent outperformance on standard benchmarks, especially for expressions requiring precise spatial localization.