Abstract:
Deep convolutional neural networks are now mainstream for click-based interactive image segmentation. In majority of the frameworks, false negatives and false positive regions are refined via a succession of positive and negative clicks placed centrally in these regions. We propose a simple yet intuitive two-in-one refinement strategy by using clicks placed on the boundary of the object of interest. As boundary clicks are a very strong cue for extracting the object of interest and we find that they are much more effective in correcting wrong segmentation masks. In addition, we propose a boundary-aware loss which encourages segmentation masks to respect instance boundaries. We place our new refinement scheme and loss formulation within a task-specialized segmentation framework and achieve state-of-the-art performance on the standard datasets - Berkeley, Pascal VOC 2012, DAVIS and MS COCO. We exceed competing methods by 6.5 %, 9.4 %, 10.5 % and 2.5 % respectively.
