Interactive Image Segmentation with Superpixel Propagation

Interactive tools play an important role in solving image segmentation problems. In this paper, we present a new interactive segmentation framework for high-accuracy segmentation. The main interaction of the user is to provide clicks inside the object of interest and control the mask-growing process with a slider. We use propagation on superpixels for region growth. To do large-scale evaluation we automate the user interactions and compare our method with state-of-the-art approaches on a few datasets with detailed annotations. Our method consistently outperforms the competitors on high accuracies and certain classes of images. We also do experiments with human annotators to show how more time-consuming naive approaches are compared to our method. Moreover, in contrast to state-of-the-art deep learning methods that stop improving segmentation accuracy beyond 20–100 clicks, our algorithm guarantees accuracy improvement after every iteration.