This paper introduces a multi-task network cascade (MNC) for instance-aware semantic segmentation. The model consists of three networks that differentiate instances, estimate masks, and categorize objects. These networks form a cascaded structure, sharing convolutional features to reduce computation and improve learning. The model is trained end-to-end, enabling efficient and accurate instance-aware segmentation. The method achieves state-of-the-art results on the PASCAL VOC dataset, with a mean Average Precision (mAP) of 63.5% and a test time of 360ms per image, which is two orders of magnitude faster than previous systems. The method also excels in object detection, surpassing Fast/Faster R-CNN systems. The MNC framework is extended to more stages, improving accuracy further. The method is also applied to the MS COCO dataset, achieving a 1st place in the segmentation track of the 2015 COCO competition. The approach is effective for instance-aware segmentation and can be generalized to other recognition tasks. The method is designed for fast inference and is orthogonal to other strategies for semantic segmentation. The paper also discusses related work, implementation details, and experiments on various datasets.This paper introduces a multi-task network cascade (MNC) for instance-aware semantic segmentation. The model consists of three networks that differentiate instances, estimate masks, and categorize objects. These networks form a cascaded structure, sharing convolutional features to reduce computation and improve learning. The model is trained end-to-end, enabling efficient and accurate instance-aware segmentation. The method achieves state-of-the-art results on the PASCAL VOC dataset, with a mean Average Precision (mAP) of 63.5% and a test time of 360ms per image, which is two orders of magnitude faster than previous systems. The method also excels in object detection, surpassing Fast/Faster R-CNN systems. The MNC framework is extended to more stages, improving accuracy further. The method is also applied to the MS COCO dataset, achieving a 1st place in the segmentation track of the 2015 COCO competition. The approach is effective for instance-aware segmentation and can be generalized to other recognition tasks. The method is designed for fast inference and is orthogonal to other strategies for semantic segmentation. The paper also discusses related work, implementation details, and experiments on various datasets.