Dynamic Prototype Optimization for Generalized Few-Shot Semantic Segmentation

Generalized Few-Shot Semantic Segmentation (GFSS) extends traditional few-shot segmentation by allowing the model to segment both base and novel classes in a single forward pass without prior knowledge of class presence. However, existing methods struggle with poor novel class performance due to unstable prototypes and confusion with base or background classes. In this paper, we propose a plug-and-play prototype enhancement architecture that dynamically enriches novel class representations by leveraging structural similarity with base class prototypes. We further introduce an adaptive parameter strategy to improve generalizability across different backbones. Experimental results demonstrate that our approach effectively improves novel class segmentation while maintaining base class performance, achieving a better balance between generalization and stability.