Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation

Building instance segmentation models that are data-efficient and can handle\nrare object categories is an important challenge in computer vision. Leveraging\ndata augmentations is a promising direction towards addressing this challenge.\nHere, we perform a systematic study of the Copy-Paste augmentation ([13, 12])\nfor instance segmentation where we randomly paste objects onto an image. Prior\nstudies on Copy-Paste relied on modeling the surrounding visual context for\npasting the objects. However, we find that the simple mechanism of pasting\nobjects randomly is good enough and can provide solid gains on top of strong\nbaselines. Furthermore, we show Copy-Paste is additive with semi-supervised\nmethods that leverage extra data through pseudo labeling (e.g. self-training).\nOn COCO instance segmentation, we achieve 49.1 mask AP and 57.3 box AP, an\nimprovement of +0.6 mask AP and +1.5 box AP over the previous state-of-the-art.\nWe further demonstrate that Copy-Paste can lead to significant improvements on\nthe LVIS benchmark. Our baseline model outperforms the LVIS 2020 Challenge\nwinning entry by +3.6 mask AP on rare categories.\n