PointMotionBench
收藏资源简介:
PointMotionBench 是一个用于评估视频中3D点运动的基准数据集,整合了来自DAVIS、HOT3D和WorldTrack三个源数据集的样本,覆盖第一人称和第三人称两种视角的多样化场景。每个数据样本包含一个RGB视频剪辑、与之配对的人工验证自然语言描述,以及视频中物体的3D和2D跟踪表面点轨迹。数据集旨在为3D点运动预测、视频理解及相关多模态学习任务提供统一的评估基准。数据规模方面,共包含约2720个视频剪辑:DAVIS提供90个剪辑(24 fps,第三人称视角,涵盖多样室内外场景),HOT3D提供2475个剪辑(30 fps,第一人称视角,主要记录物体操纵场景),WorldTrack提供155个剪辑(30 fps,包含第一人称和演播室场景,并细分为adt_mini、ds_mini、po_mini和pstudio_mini四个子集,各子集剪辑长度在12到300帧之间)。数据集提供了2D和3D的点轨迹标注,主要用于研究、教育和基准测试目的。
PointMotionBench is a benchmark dataset for evaluating 3D point motion in videos. It integrates samples from three source datasets: DAVIS, HOT3D, and WorldTrack, covering diverse scenarios from both first-person and third-person perspectives. Each data sample includes an RGB video clip, paired with manually verified natural language descriptions, as well as 2D and 3D tracked surface point trajectories of objects in the video. This dataset aims to provide a unified evaluation benchmark for 3D point motion prediction, video understanding, and related multimodal learning tasks. In terms of scale, it contains approximately 2,720 video clips in total: DAVIS provides 90 clips (24 fps, third-person perspective, covering diverse indoor and outdoor scenarios); HOT3D provides 2,475 clips (30 fps, first-person perspective, mainly recording object manipulation scenarios); WorldTrack provides 155 clips (30 fps, including first-person and studio scenarios, which are further divided into four subsets: adt_mini, ds_mini, po_mini, and pstudio_mini, with clip lengths ranging from 12 to 300 frames per subset). The dataset provides 2D and 3D point trajectory annotations, and is primarily intended for research, education, and benchmarking purposes.




