Voxel51/uco3d

uCO3D（非常规物体3D）是一个大规模的高质量360°真实生活物体视频数据集，包含全面的3D注释。它包括来自LVIS分类法的超过1,000个不同物体类别的170,000个物体。每个物体都通过类似转盘的视频进行捕捉，实现全方位的360°覆盖，确保物体的所有侧面都能看到。数据集提供了准确的3D注释，包括相机参数、点云和完整的3D高斯泼洒重建。所有物体都标准化到一个标准的参考框架，方便训练生成模型。uCO3D通过其规模、多样性和质量，解决了现有数据集的限制，为3D深度学习研究和应用提供了优越的资源。

uCO3D (UnCommon Objects in 3D) is a large-scale high-quality 360° video dataset of real-life objects with comprehensive 3D annotations. It includes 170,000 objects across more than 1,000 diverse object categories derived from the LVIS taxonomy. Each object is captured in a turntable-like video with full 360° coverage, ensuring all sides of the object are visible. The dataset provides accurate 3D annotations including camera parameters, point clouds, and complete 3D Gaussian Splatting reconstructions. All objects are normalized to a canonical reference frame, facilitating the training of generative models. With its combination of scale, diversity, and quality, uCO3D addresses limitations in existing datasets, providing a superior resource for 3D deep learning research and applications.