A Clicking Strategy Inspired by Inter-Individual Variation for Interactive Image Segmentation

Click-based interactive segmentation is a fundamental task in computer vision that allows user clicks to refine the results. Existing works typically focus on developing powerful segmentation models, yet sparsely treating the clicking method itself. In this study, we propose a novel clicking strategy that specifically aims to reflect the inter-individual variations of different humans during training to improve segmentation results. Our method consists of three steps. In particular, we first apply the erosion operation on the ground-truth segmentation mask with different parameter settings to generate multiple eroded masks. These eroded masks are then regarded as possible hypotheses of users' interested regions. Then, we randomly select one mask from the hypotheses to simulate an arbitrary users' behavior. By next solving the visual center of the selected mask, the training clicks are eventually obtained via randomly sampling from the visual center region. In essence, our key idea is to cast multiple eroded regions as the potentially diverse users' interests, and include the resulting stochasticity into the model training for better generality. We directly adopt an existing segmentation backbone and incorporate our clicking strategy in the training to show the effectiveness of our method. Experimental results on five datasets generally demonstrate that our method contributes to state-of-the-art segmentation performance.