Enhanced Spatial Awareness for Deep Interactive Image Segmentation

Existing deep interactive segmentation approaches can extract the desired object for the user based on simple click interaction. However, the first click provided by the user on the full image space domain is generally too local to capture the global target object, which causes them to rely on a large number of subsequent click corrections for satisfactory results. This paper explores how to strengthen the spatial awareness of user interaction especially after the first click input and increase the stability during the continuous iterative correction process. We first design an interactive cascaded localization strategy to determine the spatial range of the potential target, and then integrate this space-aware prior into a dual-stream network structure as a soft constraint for the segmentation. The above operation can increase the network's attention to the target of interest under very limited user interaction. A new training and inference strategy is also developed to completely adapt the benefit from the space-aware guidance. Furthermore, an object shape related loss is designed to better supervise the network based on user-provided prior guidance. Explicit subject, controllable correction and flexible interaction can help to significantly boost the interactive segmentation performance. The proposed method achieves state-of-the-art performance on several popular benchmarks.