The paper "Mask Scoring R-CNN" addresses the issue of scoring instance segmentation masks, which is crucial for improving the performance of instance segmentation models. The authors propose a novel method called Mask Scoring R-CNN (MS R-CNN) that learns a calibrated mask score based on the Intersection-over-Union (IoU) between the predicted mask and the ground truth mask. This approach aims to align the mask score with the actual quality of the mask, rather than relying solely on the classification score, which often does not correlate well with mask quality. The MS R-CNN framework includes a new network block, the MaskIoU head, which takes the instance feature and the predicted mask as input and regresses the IoU. This head is trained using a simple regression loss, and its output is multiplied with the classification score to form the final mask score. The authors demonstrate that this method consistently improves the performance of instance segmentation on the COCO dataset, achieving a notable gain of about 1.5% in AP (Average Precision) across different backbone networks. The paper also includes extensive evaluations and ablation studies to validate the effectiveness of the proposed method. The results show that the MS R-CNN outperforms the state-of-the-art Mask R-CNN and other instance segmentation methods, providing a new direction for improving instance segmentation performance. The source code for the method is available at <https://github.com/zjhuang22/maskscoring_rcnn>.The paper "Mask Scoring R-CNN" addresses the issue of scoring instance segmentation masks, which is crucial for improving the performance of instance segmentation models. The authors propose a novel method called Mask Scoring R-CNN (MS R-CNN) that learns a calibrated mask score based on the Intersection-over-Union (IoU) between the predicted mask and the ground truth mask. This approach aims to align the mask score with the actual quality of the mask, rather than relying solely on the classification score, which often does not correlate well with mask quality. The MS R-CNN framework includes a new network block, the MaskIoU head, which takes the instance feature and the predicted mask as input and regresses the IoU. This head is trained using a simple regression loss, and its output is multiplied with the classification score to form the final mask score. The authors demonstrate that this method consistently improves the performance of instance segmentation on the COCO dataset, achieving a notable gain of about 1.5% in AP (Average Precision) across different backbone networks. The paper also includes extensive evaluations and ablation studies to validate the effectiveness of the proposed method. The results show that the MS R-CNN outperforms the state-of-the-art Mask R-CNN and other instance segmentation methods, providing a new direction for improving instance segmentation performance. The source code for the method is available at <https://github.com/zjhuang22/maskscoring_rcnn>.