Abstract
Interactive segmentation aims to accurately segment target objects withminimal user interactions. However, current methods often fail to accuratelyseparate target objects from the background, due to a limited understanding oforder, the relative depth between objects in a scene. To address this issue, wepropose OIS: order-aware interactive segmentation, where we explicitly encodethe relative depth between objects into order maps. We introduce a novelorder-aware attention, where the order maps seamlessly guide the userinteractions (in the form of clicks) to attend to the image features. Wefurther present an object-aware attention module to incorporate a strongobject-level understanding to better differentiate objects with similar order.Our approach allows both dense and sparse integration of user clicks, enhancingboth accuracy and efficiency as compared to prior works. Experimental resultsdemonstrate that OIS achieves state-of-the-art performance, improving mIoUafter one click by 7.61 on the HQSeg44K dataset and 1.32 on the DAVIS datasetas compared to the previous state-of-the-art SegNext, while also doublinginference speed compared to current leading methods. The project page ishttps://ukaukaaaa.github.io/projects/OIS/index.html