PseudoClick: Interactive Image Segmentation with Click Imitation

The goal of click-based interactive image segmentation is to obtain precise object masks with limited user interaction, i.e., by a minimal number clicks. Existing methods require users provide all the clicks: first inspecting mask and then providing points on mislabeled regions, iteratively. We ask question: can our model directly predict where click, so as further reduce interaction cost? To this end, we propose PseudoClick, generic framework that enables existing networks candidate next These automatically generated clicks, termed pseudo clicks in work, serve an imitation human refine mask. build PseudoClick backbones show how click prediction mechanism leads improved performance. evaluate 10 public datasets from different domains modalities, showing not only outperforms approaches but also demonstrates strong generalization capability cross-domain evaluation. new state-of-the-art results several popular benchmarks, e.g., Pascal dataset, significantly reducing 12.4% achieve 85% IoU.