Interactive vision-language understanding: from question answering to guided segmentation

Interactive image segmentation enables users to interact minimally with a machine, facilitating the gradual refinement of the segmentation mask for a target of interest. Previous studies have demonstrated impressive performance in extracting a single target mask through interactive segmentation. However, existing methods often overlook the information cues from previously interacted objects, despite realworld scenarios frequently involving multiple objects in the same category, leading to increased redundant user effort. To this end, this chapter introduces a new interactive image segmentation framework for multiple objects within the same category in a single image. Specifically, this work leverages Transformers to extract interactionfocused visual features from the image and the interactions, thereby obtaining a satisfactory mask of the target as an exemplar. For multiple objects, an exemplar -informed module is proposed to enhance the learning of similarities among the objects of the target category. To combine features from different modules, cross-attention blocks are incorporated, followed by a feature fusion module. Extensive experiments conducted on different datasets demonstrate that the proposed framework achieves superior performance compared to previous methods. Particularly, it reduces users’ labor by around 15%, requiring two fewer clicks to achieve target IoUs of 85% and 90% for multiple objects. These results further highlight the proposed method’s potential as a flexible and practical annotation tool.