This is a derivative of DL3DV-10K. The underlying data remains subject to the DL3DV-10K Terms of Use and CC BY-NC 4.0. You must accept those terms in full before downloading.

By requesting access you agree to the DL3DV-10K Terms of Use (https://github.com/DL3DV-10K/Dataset/blob/main/License.md), in particular: you will use the data only for non-commercial research and educational purposes; you accept full responsibility for your use of it; you will not redistribute it except to individuals who have themselves accepted these terms; and you will be able to delete person-specific data on request. If you are employed by a for-profit entity, your employer is also bound by these terms and you confirm you are authorised to accept on its behalf.

DL3DV-10K-Meshed

A derivative of DL3DV-10K providing undistorted 480P views together with ground-truth surface geometry, used to train Surflo.

This dataset is not a replacement for DL3DV-10K. It redistributes some of DL3DV-10K imagery, and remains subject to the DL3DV-10K Terms of Use. See Licensing and terms before using it.

What we changed

Relative to DL3DV-ALL-480P:

Undistorted images. All 480P frames were reprocessed with COLMAP to remove lens distortion. Cameras are correspondingly pinhole.
Reference surface meshes. A full-scene mesh was reconstructed for every scene with Gaussian Wrapping (hence the gw_output/ directory name).
Per-view depth. Depth maps of the reconstructed mesh, rendered from each camera.
Oriented surface samples. 10M points with normals sampled on the mesh surface, chunked across several files per scene.

Not included: 2K/4K frames, source videos.

Scene hashes match DL3DV-10K exactly, so a scene here can always be traced back to its upstream counterpart.

Repository layout

Scenes are packed one archive per scene, grouped into subsets mirroring DL3DV-10K:

1K/<scene_hash>.tar
2K/<scene_hash>.tar
...
10K/<scene_hash>.tar
11K/<scene_hash>.tar

Input scene layout

Each scene is a COLMAP-style dataset (posed RGB views) augmented with pseudo-GT surface labels, packed as <scene>.tar with the internal layout:

<data_dir>/<scene>/
├── images/                          # RGB views (jpg/png; extension auto-detected)
│   └── <name>.<ext>
├── sparse/                          # COLMAP output
│   └── ...
└── gw_output/
    ├── cameras.json                 # per-view cameras (3DGS / COLMAP format)
    ├── scene_extent.pth             # dict: {"scene_radius": float, "scene_center": (3,)}
    ├── point_cloud/
    │   └── iteration_30000/
    │       └── point_cloud.ply      # Gaussians optimized with Gaussian Wrapping
    ├── mesh_<...>.ply               # Mesh representing the full scene, obtained with Gaussian Wrapping
    ├── depth/
    │   └── <name>.pth               # per-view mesh-depth tensor (H, W), stem matches image
    ├── surface_point_labels/
    │   └── *.pth                    # mesh surface points  — chunked (N, 3) tensors
    └── surface_normal_labels/
        └── *.pth                    # mesh surface normals — chunked (N, 3) tensors

Each cameras.json entry holds img_name (stem, no extension), width, height, position (3,), rotation (3×3), fx and fy.

The surface_point_labels / surface_normal_labels are oriented points sampled from the scene's ground-truth mesh. We obtained the meshes by running Gaussian Wrapping on each COLMAP scene (hence the gw_output/ name).

Point and normal chunks are index-aligned: the i-th row of a normal chunk is the normal at the i-th row of the correspondingly named point chunk. Concatenating all chunks in sorted filename order yields the full 10M-point set.

Getting access

This dataset is gated. You must accept the DL3DV-10K Terms of Use before downloading; approval is automatic once you do. You will then need to be authenticated:

hf auth login

Download

The full dataset is several terabytes.

Full download:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="AntoineGuedon/DL3DV-10K-Meshed",
    repo_type="dataset",
    local_dir="dl3dv-meshed",
)

Download one subset at a time:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="AntoineGuedon/DL3DV-10K-Meshed",
    repo_type="dataset",
    allow_patterns="1K/*.tar",
    local_dir="dl3dv-meshed",
)

A single scene, by hash:

from huggingface_hub import hf_hub_download

hf_hub_download(
    repo_id="AntoineGuedon/DL3DV-10K-Meshed",
    repo_type="dataset",
    filename="1K/<scene_hash>.tar",
    local_dir="dl3dv-meshed",
)

Licensing and terms

The image data originates from DL3DV-10K and remains subject to the DL3DV-10K Terms of Use and CC BY-NC 4.0. This release is a derivative; We are not relicensing the original data, and nothing here grants rights beyond those in the original license.

Our contributions — the undistorted reprocessing, rendered depth maps, and sampled surface points and normals — are released under CC BY-NC 4.0.

In particular, you may use this dataset for non-commercial research and educational purposes only, you must not redistribute it to anyone who has not accepted the DL3DV-10K Terms of Use, and you must be able to delete person-specific data on request.

Citation

If you use this dataset, please cite both DL3DV-10K and Surflo. If you use the surface labels, please also cite the mesh reconstruction method.

@inproceedings{ling2024dl3dv,
  title={DL3DV-10K: A Large-Scale Scene Dataset for Deep Learning-based 3D Vision},
  author={Ling, Lu and Sheng, Yichen and Tu, Zhi and Zhao, Wentian and Xin, Cheng and
          Wan, Kun and Yu, Lantao and Guo, Qianyu and Yu, Zixun and Lu, Yawen and others},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={22160--22169},
  year={2024}
}

@article{guedon2026surflo,
  title={Surflo: Consistent 3D Surface Flow from a Global State},
  author={Gu\'edon, Antoine and Nakamura, Shu and Dufour, Nicolas and
          Lei, Jiahui and Nishino, Ko and Kanazawa, Angjoo},
  journal={arXiv preprint arXiv:2606.13644},
  year={2026}
}

@article{gomez2026blobsspokes,
  title={From Blobs to Spokes: High-Fidelity Surface Reconstruction via Oriented Gaussians},
  author={Gomez, Diego and Gu\'edon, Antoine and Maruani, Nissim and
          Gong, Bingchen and Ovsjanikov, Maks},
  journal={ECCV},
  year={2026}
}

Acknowledgements

Our thanks to Lu Ling and the DL3DV-10K team for building and sharing the original dataset.

Downloads last month: 654

Total file size:

7.93 TB

Paper for AntoineGuedon/DL3DV-10K-Meshed

Surflo: Consistent 3D Surface Flow Model with Global State

Paper • 2606.13644 • Published Jun 11 • 5