AutoClick: Auto Seed Selection for Interactive Segmentation

The existing interactive segmentation methods require the user to mark one or more seeds for later segmentation. However, when dealing with some special tasks like aeroengine blade instance segmentation, previous interactive segmentation methods have high interactive costs and are easily affected by the click position. In this paper, we propose a new paradigm for segmenting instances with less user interaction. We first use the key points of each instance as seeds to generate an initial segmentation mask. Then we update the segmentation mask iteratively. For the instance without key point, we manually click a point in the instance as the seed. By using this paradigm, we can achieve high segmentation precision with fewer interactions, making the process more intelligent. We have conducted our method on the task of aeroengine blade instance segmentation. Extensive experiments demonstrate that our method achieves superior instance segmentation performance than SOTA instance segmentation methods.