The paper introduces the first fully convolutional end-to-end solution for instance-aware semantic segmentation, combining the strengths of fully convolutional networks (FCNs) for semantic segmentation and instance mask proposal generation. The proposed method, called FCIS, jointly detects and segments object instances using position-sensitive inside/outside score maps, which are shared between the two tasks. This approach avoids the limitations of traditional FCNs, which are translation-invariant and do not account for individual object instances. FCIS achieves state-of-the-art performance in both accuracy and efficiency, outperforming previous methods like MNC in the COCO 2016 segmentation competition by a significant margin. The network structure is highly integrated, with simple and fast per-ROI computation, and it operates on box proposals instead of sliding windows, leveraging recent advances in object detection. The method is demonstrated to be effective through extensive experiments on the PASCAL VOC and COCO datasets, showing superior accuracy and speed compared to competing methods.The paper introduces the first fully convolutional end-to-end solution for instance-aware semantic segmentation, combining the strengths of fully convolutional networks (FCNs) for semantic segmentation and instance mask proposal generation. The proposed method, called FCIS, jointly detects and segments object instances using position-sensitive inside/outside score maps, which are shared between the two tasks. This approach avoids the limitations of traditional FCNs, which are translation-invariant and do not account for individual object instances. FCIS achieves state-of-the-art performance in both accuracy and efficiency, outperforming previous methods like MNC in the COCO 2016 segmentation competition by a significant margin. The network structure is highly integrated, with simple and fast per-ROI computation, and it operates on box proposals instead of sliding windows, leveraging recent advances in object detection. The method is demonstrated to be effective through extensive experiments on the PASCAL VOC and COCO datasets, showing superior accuracy and speed compared to competing methods.