This paper introduces Copy-Paste as a simple yet effective data augmentation method for instance segmentation. The method involves randomly pasting objects from one image onto another, which generates new training data without requiring complex modeling of visual context. The authors demonstrate that this simple approach can significantly improve performance on standard instance segmentation benchmarks, such as COCO and LVIS, and is additive with semi-supervised learning techniques like self-training. The study shows that Copy-Paste achieves a mask AP of 49.1 and box AP of 57.3 on COCO, surpassing previous state-of-the-art results. It also improves performance on the LVIS benchmark, outperforming the 2020 challenge winner by +3.6 mask AP on rare categories. The method is shown to be robust across different training configurations, including varying backbone architectures, image sizes, and training schedules. It is also effective when combined with large-scale jittering and self-training, leading to further improvements in performance. Copy-Paste is easy to implement and does not increase training or inference costs. It is particularly effective for data-efficient training and helps improve performance on rare object categories. The method is shown to be effective across a wide range of model architectures and image resolutions, and it provides significant improvements in both single-stage and two-stage training settings. The paper also demonstrates that Copy-Paste can be used to improve transfer learning performance on other datasets, such as PASCAL VOC. The method is shown to be effective in both detection and segmentation tasks, and it provides better representations for transfer learning. Overall, the study shows that Copy-Paste is a simple, effective, and robust data augmentation method that can significantly improve instance segmentation performance.This paper introduces Copy-Paste as a simple yet effective data augmentation method for instance segmentation. The method involves randomly pasting objects from one image onto another, which generates new training data without requiring complex modeling of visual context. The authors demonstrate that this simple approach can significantly improve performance on standard instance segmentation benchmarks, such as COCO and LVIS, and is additive with semi-supervised learning techniques like self-training. The study shows that Copy-Paste achieves a mask AP of 49.1 and box AP of 57.3 on COCO, surpassing previous state-of-the-art results. It also improves performance on the LVIS benchmark, outperforming the 2020 challenge winner by +3.6 mask AP on rare categories. The method is shown to be robust across different training configurations, including varying backbone architectures, image sizes, and training schedules. It is also effective when combined with large-scale jittering and self-training, leading to further improvements in performance. Copy-Paste is easy to implement and does not increase training or inference costs. It is particularly effective for data-efficient training and helps improve performance on rare object categories. The method is shown to be effective across a wide range of model architectures and image resolutions, and it provides significant improvements in both single-stage and two-stage training settings. The paper also demonstrates that Copy-Paste can be used to improve transfer learning performance on other datasets, such as PASCAL VOC. The method is shown to be effective in both detection and segmentation tasks, and it provides better representations for transfer learning. Overall, the study shows that Copy-Paste is a simple, effective, and robust data augmentation method that can significantly improve instance segmentation performance.