SpatialVID
收藏资源简介:
SpatialVID是一个大规模视频数据集,包含显式空间注释,如相机姿态、深度图、结构化标题和序列化运动指令。该数据集由7089小时的动态场景组成,通过从超过21000小时的原始视频中提取和处理而来。数据集通过一个层次化的过滤流程,去除了质量缺陷、尺寸不正确或标题不相关的视频内容,然后通过美学指标、亮度、光学字符识别和运动值的分层评分策略进行排序。高评分的视频片段经过双重注释流程,捕捉空间结构和语义信息,最终形成了SpatialVID数据集。该数据集旨在连接原始像素与物理世界,为视频和3D视觉研究提供关键资产。
SpatialVID is a large-scale video dataset equipped with explicit spatial annotations, including camera poses, depth maps, structured captions, and sequential motion instructions. It consists of 7089 hours of dynamic scenes, extracted and processed from over 21,000 hours of raw video footage. The dataset first undergoes a hierarchical filtering pipeline to remove video content with quality defects, incorrect dimensions, or irrelevant captions, and then is ranked using a hierarchical scoring strategy covering aesthetic metrics, brightness, optical character recognition (OCR), and motion values. High-scoring video clips go through a dual-annotation workflow to capture spatial structure and semantic information, ultimately forming the final SpatialVID dataset. This dataset aims to bridge raw pixels and the physical world, serving as a critical asset for video and 3D vision research.

- 1通过南京大学 中国科学院自动化研究所 · 2025年



