Rawmantic 3D Training Field

Official Website | Xiaohongshu | Demo Page | 1M Download

The first large-scale "ImageNet"-style dataset for 3D understanding.

• Dataset page: https://raw3d.xyz/dataset
• Purpose: A comprehensive 3D dataset for research and applications in computer vision, robotics, AR/VR, and scene understanding.
• Topics: It supports the training of Monocular Depth (e.g., DepthAnythingV2), Stereo Matching (e.g., FoundationStereo), Feed-Forward (e.g., VGGT, Pi3), MVS, SLAM and so on.

Download

Still uploading, 30/100 scenes are done (listed in scenes.txt), near 400K samples, eventually will be over 1M multiview samples!

• Baidu Yunpan: https://pan.baidu.com/s/1DzJvQu6dEhi7svqE1S-18Q?pwd=rock
• Google Drive: please contact us if anyone can help, super slow transfering speed.
• Hugging Face: due to the single file size (max 5G) limited by huggingface, we will not submit our data to huggingface.

Dataset Reader

Features may include (see reader.py):

• loading 3D depth, normal, trajectory, and multi-view images
• future feature: reading object labels, bounding boxes, and semantic annotations
• supporting common dataset splits and evaluation protocols
• benchmark your SLAM, 3D reconstruction and other 3D related tasks.

Data Structure Example:

Rawmantic
└── LivingSpace
    └── StreetBlock
        └── LivingSpace_StreetBlock_00000000_ICPark
            └── 00000000_ICPark_02
                └── data
                    ├── depth_camera_left:  left  camera, reference view
                    ├── depth_camera_right: right camera, a stereo view
                    ├── depth_camera_third: third camera, a random view, i.e., multi-view
                    ├── normal_camera_left: surface normal in camera view
                    ├── normal_camera_right
                    ├── normal_camera_third
                    ├── pose_camera_left:   given intrinsics and extrinsics
                    ├── pose_camera_right
                    ├── pose_camera_third
                    ├── rgb_camera_left
                    ├── rgb_camera_right
                    └── rgb_camera_third

Release Plan

Release	Description	Status
v1.0	Initial public release with demo dataset and annotations	2026.06
v1.1	Expanded scenes	2026.07
v2.0	Full dataset scale release, benchmark split, and evaluation toolkit	Planned

Citation

If you use this dataset in your research, please cite it as:

@article{rawmantic2026data,
  title={Rawmantic 3D Training Field: From The Real World, For The Real World},
  author={Rawmantic Team},
  journal={Rawmantic},
  year={2026}
}

License

Refer to our website: https://raw3d.xyz/dataset

Contact & Contribution

If you want to contribute, report issues, or ask questions, please open an issue in this repository or reach out through the project's official channels.