The paper introduces a novel architecture for object instance segmentation, which combines feedforward networks with a top-down refinement approach. The goal is to efficiently merge spatially rich information from low-level features with high-level object knowledge from upper layers. The proposed method, called SharpMask, enhances the accuracy of object masks by refining a coarse 'mask encoding' generated in a feedforward pass using features from successively lower layers. This approach is demonstrated to improve average recall by 10-20% on the COCO dataset and achieve a 50% speedup compared to the original DeepMask network, with an average inference time of 0.76 seconds per image. The paper also discusses the design of the feedforward network, including trunk and head architectures, and provides experimental results showing the effectiveness of SharpMask in object proposal generation and object detection tasks.The paper introduces a novel architecture for object instance segmentation, which combines feedforward networks with a top-down refinement approach. The goal is to efficiently merge spatially rich information from low-level features with high-level object knowledge from upper layers. The proposed method, called SharpMask, enhances the accuracy of object masks by refining a coarse 'mask encoding' generated in a feedforward pass using features from successively lower layers. This approach is demonstrated to improve average recall by 10-20% on the COCO dataset and achieve a 50% speedup compared to the original DeepMask network, with an average inference time of 0.76 seconds per image. The paper also discusses the design of the feedforward network, including trunk and head architectures, and provides experimental results showing the effectiveness of SharpMask in object proposal generation and object detection tasks.