YouTubeVIS is a large-scale dataset for video instance segmentation, collected from the YouTubeVOS dataset. The dataset is designed for tasks such as simultaneous detection, segmentation, and tracking of target instances in videos.

The dataset is divided into 2019, 2021 and 2022 versions. The latest version is the 2022 version, which is used for the 4th Large-Scale Video Object Segmentation Challenge of CVPR 2022:

• Added 71 long videos to the validation set and 50 long videos to the test set, with additional separate evaluations;
• Contains 259 additional video instances with 9,304 high-quality human annotations;
• The additional long video (L) and the previous video (S) are evaluated separately, and the final mAP is the average of mAP_L and mAP_S.