DanceTrack

DanceTrack是一个多人类跟踪数据集，具有两个强调的属性 :( 1) 外观均匀: 人类外观高度相似且几乎没有区别; (2) 运动多样: 人类处于复杂的运动模式中，并且它们的相对位置经常交换。我们希望均匀的外观和复杂的运动模式相结合，使DanceTrack成为鼓励更全面，更智能的多目标跟踪算法的平台。
DanceTrack的注释根据知识共享署名4.0许可证获得许可。DanceTrack的数据集仅可用于非商业研究目的。DanceTrack的所有视频和图像都是从互联网上获得的，这些视频和图像不是HKU、CMU或ByteDance的财产。这三个组织不对这些视频和图像的内容或含义负责。

DanceTrack is a multi-human tracking dataset with two highlighted attributes: (1) Uniform Appearance: Human appearances are highly similar and barely distinguishable; (2) Diverse Motion: Humans exhibit complex motion patterns, and their relative positions frequently swap. We believe that the combination of uniform appearance and complex motion patterns makes DanceTrack a platform that encourages more comprehensive and intelligent multi-object tracking algorithms.
Annotations of DanceTrack are licensed under the Creative Commons Attribution 4.0 License. The DanceTrack dataset is only available for non-commercial research purposes. All videos and images in the DanceTrack dataset were obtained from the Internet, and these videos and images are not the property of HKU, CMU, or ByteDance. These three organizations bear no responsibility for the content or implication of these videos and images.

搜集汇总

背景与挑战

背景概述

DanceTrack是一个多人类跟踪数据集，专注于外观高度相似且运动模式复杂的人类目标，旨在推动更全面的多目标跟踪算法发展。该数据集由卡内基梅隆大学、字节跳动和香港大学于2022年联合发布，包含18.1GB的视频帧数据，适用于计算机视觉和目标跟踪的非商业研究。