OmniObject3D is a large-scale high-quality 3D dataset for real-world scanned objects of multiple categories. It supports the generalization and robustness of various research topics such as novel view synthesis, neural surface reconstruction, 3D generation, and point cloud analysis. Developers and researchers will benefit from using this dataset when understanding and constructing their 2D-3D-related algorithms.
6,000 high-quality 3D models
5 videos per object covering different trajectories and lighting conditions
More than 190 object categories
Mesh plus texture, point cloud, real-captured video, synthetic multi-view images, camera pose
Multiple down-stream tasks supported