Interactive Image Segmentation Based on Feature-Aware Attention

Interactive segmentation is a technique for picking objects of interest in images according to users’ input interactions. Some recent works take the interactive guide deep neural network training, where click information utilized as weak-supervised information. However, limited by learning capability model, this structure does not accurately represent user’s interaction intention. In work, we propose multi-click solution employing human intention refine results. We coarse extract semantic and generate rough Then, designed feature-aware attention module symmetry user image Finally, establish refinement combine results with masks precise intentional segmentation. Furthermore, trained plug-and-play tool, which can be embedded into most models exploiting training process. conduct experiments on five common datasets (SBD, GrabCut, DAVIS, Berkeley, MS COCO) prove our improve performance networks.