This paper presents the first fully convolutional end-to-end solution for instance-aware semantic segmentation. The proposed method, called FCIS, extends the approach in [5] by introducing position-sensitive inside/outside score maps, which allow the network to jointly perform object segmentation and detection. The underlying convolutional representation and score maps are fully shared between the two sub-tasks, enabling efficient and accurate instance-aware segmentation. The network is highly integrated and achieves state-of-the-art performance in both accuracy and efficiency. It wins the COCO 2016 segmentation competition by a large margin. The method is efficient, with inference taking 0.24 seconds per image using a ResNet-101 model on a Nvidia K40 GPU, which is 6× faster than the previous winning method MNC. The method is also effective on the COCO dataset, achieving significantly higher accuracy than MNC on the large-scale COCO dataset. The approach is fully convolutional, with no extra parameters, and is highly efficient, with negligible per-ROI computation cost. The method is evaluated on the PASCAL VOC and COCO datasets, and shows superior performance compared to alternative methods. The method is also effective for object detection, achieving a high accuracy on the COCO test-dev set. The paper also discusses related work, including semantic image segmentation, object segment proposal, and instance-aware semantic segmentation. The method is compared with other approaches, including MNC, and is shown to be superior in terms of accuracy and efficiency. The method is also evaluated on different network depths and multi-scale testing, showing that it benefits from these techniques. The method is also compared with other entries in the COCO segmentation challenge, and is shown to be the first-place winner in 2016. The paper concludes that the proposed method is a significant advancement in instance-aware semantic segmentation, achieving state-of-the-art performance in both accuracy and efficiency.This paper presents the first fully convolutional end-to-end solution for instance-aware semantic segmentation. The proposed method, called FCIS, extends the approach in [5] by introducing position-sensitive inside/outside score maps, which allow the network to jointly perform object segmentation and detection. The underlying convolutional representation and score maps are fully shared between the two sub-tasks, enabling efficient and accurate instance-aware segmentation. The network is highly integrated and achieves state-of-the-art performance in both accuracy and efficiency. It wins the COCO 2016 segmentation competition by a large margin. The method is efficient, with inference taking 0.24 seconds per image using a ResNet-101 model on a Nvidia K40 GPU, which is 6× faster than the previous winning method MNC. The method is also effective on the COCO dataset, achieving significantly higher accuracy than MNC on the large-scale COCO dataset. The approach is fully convolutional, with no extra parameters, and is highly efficient, with negligible per-ROI computation cost. The method is evaluated on the PASCAL VOC and COCO datasets, and shows superior performance compared to alternative methods. The method is also effective for object detection, achieving a high accuracy on the COCO test-dev set. The paper also discusses related work, including semantic image segmentation, object segment proposal, and instance-aware semantic segmentation. The method is compared with other approaches, including MNC, and is shown to be superior in terms of accuracy and efficiency. The method is also evaluated on different network depths and multi-scale testing, showing that it benefits from these techniques. The method is also compared with other entries in the COCO segmentation challenge, and is shown to be the first-place winner in 2016. The paper concludes that the proposed method is a significant advancement in instance-aware semantic segmentation, achieving state-of-the-art performance in both accuracy and efficiency.