OmniObject3D is a large-scale 3D object dataset that includes over 6,000 high-quality real scanned 3D objects, covering 190 categories of everyday items, providing detailed mesh models, textures, and multi-view renderings, suitable for tasks such as 3D reconstruction, classification, and generation.
A large-scale real scanned 3D object dataset that provides a solid foundation for 3D vision research
Each 3D object provides a detailed triangular mesh model with high polygon density and realistic geometric details, suitable for rendering and analysis.
All objects are captured using professional 3D scanning equipment, not synthetically generated, ensuring the authenticity and accuracy of geometric shapes and surface textures.
Covers 190 categories of everyday items such as furniture, food, toys, tools, and electronic devices, with a wide distribution of categories that closely resemble real-world scenes.
Provides multi-angle rendered images for each object, supporting research tasks such as Novel View Synthesis and multi-view 3D reconstruction.
Includes high-resolution texture maps that realistically reproduce the surface materials, colors, and details of objects, suitable for realistic rendering and visual research.
Provides high-density point cloud data sampled from mesh models, which can be directly used for point cloud classification, segmentation, and 3D feature learning algorithms.
From academic research to industrial applications, covering core tasks of 3D vision
Utilize multi-view images and real geometric data to train and evaluate reconstruction algorithms such as NeRF and 3D Gaussian Splatting
Perform 3D object classification based on point clouds or mesh models, evaluating the performance of classification networks like PointNet and DGCNN
Provide high-quality training data for 3D diffusion models and generative adversarial networks, researching the generation of 3D objects from text/images
Train view synthesis models using multi-view rendered images to generate realistic images from sparse views at any angle
The following is a typical directory structure and metadata example of the OmniObject3D dataset
OmniObject3D/ ├── raw_scans/ │ ├── chair/ │ │ ├── chair_001/ │ │ │ ├── mesh.obj # Triangle mesh model │ │ │ ├── mesh.mtl # Material file │ │ │ ├── texture.png # Texture map │ │ │ └── pointcloud.ply # Point cloud data │ │ ├── chair_002/ │ │ └── ... │ ├── cup/ │ ├── shoe/ │ └── ... (190 categories) ├── renders/ │ ├── chair/ │ │ ├── chair_001/ │ │ │ ├── view_000.png # Multi-view rendered image │ │ │ ├── view_001.png │ │ │ └── ... │ │ └── ... │ └── ... └── metadata.json # Category and object metadata
From browsing to loading, you can start your 3D visual research in just a few minutes
View dataset details on the Ace Data Cloud platform to understand metadata such as category distribution, object count, and data format.
Download specific categories or the complete dataset as needed, including mesh models, texture maps, point cloud data, and multi-view renderings.
Use Open3D or trimesh to load 3D models and start research tasks such as reconstruction, classification, and generation.
A large-scale real scanned 3D object dataset with open licensing, available for immediate download. Whether you are a 3D vision researcher or a 3D generation developer, this dataset is worth trying.
