Global Selection and Local Attention Network for Referring Image Segmentation

Referring image segmentation (RIS) aims to segment the target object based on a natural language expression. The challenge lies in comprehending both the image and the referring expression simultaneously, while establishing the alignment between these two modalities. Recently, the visual-language large-scale pre-trained model CLIP can well align the modalities. However, the alignment in these models is based on the global image. And RIS requires aligning global text features with local visual features, rather than global visual features. To this end, features extracted by CLIP can not be directly applied to RIS. In this paper, we propose a novel framework called Global Selection and Local Attention Network (GLNet), which builds upon CLIP. GLNet comprises two modules: Global Selection and Fusion Module (GSFM) and Local Attention Module (LAM). GSFM utilizes text information to adaptively select and fuse visual features from low-level and middle-level. LAM leverages attention mechanisms on both local visual features and local text features to establish relationships between objects and text. Extensive experiments demonstrate the exceptional performance of our proposed method in referring image segmentation. On RefCOCO+, GLNet achieves significant performance gains of +2.38%, +2.78%, and +2.50% on the three splits compared to SADLR.