Simple Class Relation Helps Generalized Few-Shot Semantic Segmentation

Abstract

We propose a method for generalized few-shot semantic segmentation (GFSS). Unlike classic few-shot semantic segmentation (FSS), which focuses on recognizing novel-class objects only, GFSS aims to recognize both base and novel-class objects. Thus, GFSS is regarded as a more realistic setting than FSS. Our method finds simple relations between base and novel classes and then trains models to recognize novel classes based on related base classes. Through experiments, we demonstrated the superior performance of our method against other GFSS methods.