UnCommon Objects in 3D (uCO3D)
收藏资源简介:
UnCommon Objects in 3D (uCO3D) 是由Meta AI等机构创建的一个高质量3D对象数据集,旨在为3D深度学习和生成式AI提供支持。该数据集包含超过170,000个场景,涵盖1,000多个对象类别,每个视频帧都带有精确的3D注释,包括SfM相机、点云和3D高斯溅射重建。数据集通过众包方式收集,视频分辨率高,且经过严格的质量检查,确保每个视频都能提供完整的360度视角覆盖。uCO3D在3D模型训练中表现出色,优于现有的MVImgNet和CO3Dv2数据集,适用于3D重建、新视角合成和文本到3D生成等应用领域。
UnCommon Objects in 3D (uCO3D) is a high-quality 3D object dataset created by Meta AI and other institutions, aiming to support 3D deep learning and generative AI. This dataset contains over 170,000 scenes spanning more than 1,000 object categories, with each video frame equipped with precise 3D annotations including SfM cameras, point clouds, and 3D Gaussian Splatting reconstructions. Collected via crowdsourcing, the dataset features high-resolution videos and has undergone rigorous quality checks to ensure that each video provides complete 360-degree view coverage. uCO3D delivers excellent performance in 3D model training, outperforming existing datasets such as MVImgNet and CO3Dv2, and is applicable to application fields including 3D reconstruction, novel view synthesis, and text-to-3D generation.



