OmniLife360
收藏资源简介:
OmniLife360 是一个专为从野外捕获的 360 度全景视频进行 3D 重建而设计的基准数据集,旨在推动全景计算机视觉和三维重建方法的研究与评估。数据集包含 2,407 个全景视频序列,总时长约 66 小时,划分为 1,937 个训练序列和 470 个测试序列,覆盖 64 个场景类别和 31 个动作类别。数据来源包括 2,177 个来自网络的公开链接视频序列和 230 个由作者团队捕获的未公开链接序列。每个序列通过唯一路径标识,并提供丰富的元数据,如源视频标识符、URL、时间片段、场景/动作/区域标签、媒体属性和文字描述。数据以提取的帧图像形式组织,预处理为每秒 1 帧采样,尺寸调整为 1920x960 像素。此外,数据集提供基于 COLMAP 的稀疏 3D 重建模型,支持算法训练和基准测试,适用于 360 度全景场景下的三维重建、神经渲染和场景理解等任务。数据集经过隐私审查,不包含个人身份信息,使用受 CC BY-NC 4.0 许可证约束,仅用于学术研究。
OmniLife360 is a benchmark dataset designed for 3D reconstruction from 360-degree panoramic videos captured in the wild. It aims to advance research and evaluation in panoramic computer vision and 3D reconstruction methods. The dataset includes 2,407 panoramic video sequences, totaling approximately 66 hours, divided into 1,937 training sequences and 470 test sequences, covering 64 scene categories and 31 action categories. Data sources consist of 2,177 publicly linked video sequences from the web and 230 non-publicly linked sequences captured by the author team. Each sequence is identified by a unique path and provides rich metadata, such as source video identifiers, URLs, temporal segments, scene/action/region labels, media properties, and textual descriptions. The data is organized as extracted frame images, preprocessed by sampling at 1 frame per second and resizing to 1920x960 pixels. Additionally, the dataset offers sparse 3D reconstruction models based on COLMAP, supporting algorithm training and benchmarking for tasks like 3D reconstruction, neural rendering, and scene understanding in 360-degree panoramic scenarios. It has undergone privacy review, excludes personal identifiable information, and is licensed under CC BY-NC 4.0 for academic use only.




