Abstract
Recently, semantic segmentation has rapidly achieved high performance. Because the goal of semantic segmentation is to densely label each pixel with the corresponding category label, more efforts are now devoted to further differentiate instances belonging to the same category label. To achieve instance-aware semantic segmentation, most of existing methods need to use sliding windows or object proposals to locate instances. In this thesis, we utilize the semantic segmentation results to further differentiate instances in the same category. We propose a pixel-wise instance feature to identify different instance. Then, we propose a Class-shared net to aggregate the semantic segmentation results using our pixel-wise instance feature to achieve instance segmentation. The proposed method needs no external object proposal generator or any sliding windows to achieve instance segmentation. Experimental results demonstrate that the proposed method is efficient and achieve comparable results on Pascal VOC 2012 and MSCOCO datasets.