ClickAttention: Click region similarity guided interactive segmentation

Abstract

Interactive segmentation algorithms based on click points have attracted significant attention from researchers in recent years. However, most existing methods rely on sparse click maps as model inputs to segment specific target objects. These clicks primarily affect local regions, limiting the model’s ability to focus on the entire target object and often resulting in a higher number of required clicks. Additionally, many current algorithms struggle to balance performance and efficiency effectively. To address these challenges, we propose a click attention algorithm that expands the influence of positive clicks by leveraging the similarity between positively-clicked regions and the entire input. We further introduce a discriminative affinity loss to reduce attention coupling between positive and negative click regions, minimizing accuracy degradation caused by mutual interference. On the DAVIS dataset, our method achieves a 2 % performance gain (NoC@90) over the state-of-the-art SimpleClick-ViT-L, while using only 15.6 % of its parameters. Extensive experiments demonstrate that our approach outperforms existing methods and achieves state-of-the-art performance with fewer parameters. Data and code are published.