Weakly supervised instance segmentation using multi-stage erasing refinement and saliency-guided proposals ordering

摘要

Weakly supervised instance segmentation is a new research topic in the field of computer vision. Compared with fully supervised instance segmentation, weakly supervised methods use weaker data annotations such as points, scribbles or class labels which are easy to obtain. Among these annotations, image-level instance segmentation using only class labels as supervision is the most challenging task. In this paper, we propose a novel weakly supervised instance segmentation framework using a multi-stage erasing refinement method and a saliency-guided proposals ordering method. Firstly, the multi-stage erasing refinement method is exploited to enhance the instance representation by iteratively discovering separate object-related regions, so as to obtain more complete discriminative regions. Then, the saliency-guided proposals ordering method utilizes the saliency map to alleviate the background noise and better select the object proposals for generating the instance segmentation result. Experimental results on the PASCAL VOC 2012 dataset and the COCO dataset demonstrate that our framework achieves superior performance compared with the state-of-the-art weakly supervised instance segmentation models and the ablation study shows the effectiveness of the proposed two methods.