PointOdyssey

PointOdyssey是由斯坦福大学创建的大规模合成数据集，专为长期细粒度跟踪算法的训练和评估设计。该数据集包含104个视频，平均每个视频2000帧，远超以往工作的注释量。数据集通过使用真实世界动作捕捉数据驱动可变形角色，构建与动作捕捉环境匹配的3D场景，并利用真实视频中的结构从运动中挖掘的相机视角进行渲染，以实现自然主义的强调。此外，通过随机化角色外观、运动配置文件、材料、照明、3D资产和大气效果，创建了组合多样性。PointOdyssey旨在通过提供真实世界视频的复杂性、多样性和自然主义，推动长期细粒度跟踪技术的发展，解决现有数据集在长期跟踪方面的不足。

PointOdyssey is a large-scale synthetic dataset developed by Stanford University, specifically tailored for the training and evaluation of long-term fine-grained tracking algorithms. This dataset comprises 104 videos, with an average of 2000 frames per video, far exceeding the annotation volume of prior research works. The dataset drives deformable characters using real-world motion capture data, constructs 3D scenes aligned with the motion capture environment, and renders using camera perspectives extracted from structure-from-motion (SfM) of real-world videos to achieve naturalistic rendering. Furthermore, combinatorial diversity is generated by randomizing character appearances, motion profiles, materials, lighting, 3D assets, and atmospheric effects. PointOdyssey aims to advance the development of long-term fine-grained tracking technologies by providing the complexity, diversity, and naturalistic characteristics of real-world videos, thus addressing the limitations of existing datasets in long-term tracking tasks.