CLVOS23

CLVOS23是由滑铁卢大学创建的一个专注于持续学习的长视频目标分割数据集，包含9个视频序列，总帧数达到7411帧，新增5951帧，注释帧数从63增加到284。该数据集通过捕捉视频序列中的分布漂移，特别适合测试在线视频目标分割方法在长视频上的性能。创建过程中，数据集的注释基于视频序列中的分布漂移，而非均匀选择，这使得数据集更加真实且具有挑战性。CLVOS23的应用领域主要集中在视频摘要、人机交互和自动驾驶等，旨在解决长视频中目标对象分割的持续学习问题。

CLVOS23 is a long-video object segmentation dataset dedicated to continual learning, developed by the University of Waterloo. It contains 9 video sequences, with a total of 7411 frames, including 5951 newly added frames, and the number of annotated frames has increased from 63 to 284. By capturing distribution shifts within video sequences, this dataset is particularly well-suited for evaluating the performance of online video object segmentation methods on long-form videos. During the dataset construction process, annotations were selected based on distribution shifts in the video sequences rather than uniform sampling, rendering the dataset more realistic and challenging. The primary application areas of CLVOS23 cover video summarization, human-computer interaction, autonomous driving and other domains, with the goal of solving the continual learning problem for object segmentation in long videos.