A Partial Context-Aware Structure for Few-Shot Semantic Segmentation

Abstract

In recent years, great progress has been made in few-shot semantic segmentation. However, the segmentation performance and feature analysis model still need to be improved in the existing models. In this paper, we mainly make two improvements. First, we propose a partial context-aware structure to enrich the background of the support set target and construct a self-guidance approach for few-shot segmentation. Thus, the guidance feature includes the partial context information of the query set and can be used to suppress the background semantic gap between the support and query sets. Second, to mine more comprehensive support information, we enhance the domain adaptation between the support and query sets by constructing an auxiliary branch that leverages overlap feature. Note that the proposed method can be used to improve many existing models. Experimental results on both PASCAL-5i and COCO-20i show that the proposed method achieves better performances than that of the comparison models. Code is available on https://github.com/ily666666/pcas.

Keywords: Semantic segmentation, few-shot segmentation, few-shot learning