Preventing Catastrophic Forgetting in Generalized Few-Shot Semantic Segmentation

Abstract

The goal of generalized few-shot semantic segmentation (GFSS) is to recognize both base- and novel-class objects at inference, using a learned base-class model and few-shot data for novel classes. An issue is catastrophic forgetting of the learned base-class model when training with the novel-class data. This paper presents the method for GFSS and theoretically derives that the method prevents catastrophic forgetting of the base-class model.