Hajimiri Sina ; Boudiaf Malik ; Ayed Ismail Ben ; Dolz Jose

Abstract

This paper introduces a generalized few-shot segmentation framework with astraightforward training process and an easy-to-optimize inference phase. Inparticular, we propose a simple yet effective model based on the well-knownInfoMax principle, where the Mutual Information (MI) between the learnedfeature representations and their corresponding predictions is maximized. Inaddition, the terms derived from our MI-based formulation are coupled with aknowledge distillation term to retain the knowledge on base classes. With asimple training process, our inference model can be applied on top of anysegmentation network trained on base classes. The proposed inference yieldssubstantial improvements on the popular few-shot segmentation benchmarks,PASCAL-$5^i$ and COCO-$20^i$. Particularly, for novel classes, the improvementgains range from 7% to 26% (PASCAL-$5^i$) and from 3% to 12% (COCO-$20^i$) inthe 1-shot and 5-shot scenarios, respectively. Furthermore, we propose a morechallenging setting, where performance gaps are further exacerbated. Our codeis publicly available at https://github.com/sinahmr/DIaM.