HODome Dataset (NeuralDome, CVPR 2023)

HODome is a dense 76-view dome capture of human–object interaction: 10 subjects interacting with 23 objects. This repository hosts the annotations (body-model reconstructions, object 6-DoF, calibration, scanned object templates). The 76-view 4K videos are large and distributed separately (see Videos below).

• Paper: NeuralDome (CVPR 2023)
• Project page: https://juzezhang.github.io/NeuralDome/
• Toolbox (download / process / visualize): https://github.com/Juzezhang/NeuralDome_Toolbox

Layout

HODome/
├─ smplx/{Seq}.npz          ★ RECOMMENDED — SMPL-X reconstruction (body+hands, valid from frame 0)
├─ mhr/{Seq}/mhr/*.json     ★ RECOMMENDED — MHR model parameters
├─ smplh/{Seq}.npz            legacy SMPL-H GT — lower accuracy, prefer smplx / mhr
├─ object/{Seq}.npz          object 6DoF (object_R / object_T), ground-aligned
├─ scaned_object/{obj}/      scanned object templates (.obj)
├─ calibration_ground/{date}/   ground-aligned camera calibration
└─ (videos / mask_refine distributed separately — see below)

Human pose sources — three body models cover the same actor. Use smplx (recommended) or mhr (modern reconstructions, full hands, consistent identity from frame 0). smplh is the older ground-aligned GT and is noticeably less accurate — kept for backward compatibility. Object pose is always object/{Seq}.npz; the object mesh is the mean-centered scaned_object/{obj}/{obj}_face1000.obj posed by object_R/object_T (verts0 @ object_R.T + object_T).

Usage

pip install -U "huggingface_hub[cli]"
huggingface-cli download JuzeZhang/HODome --repo-type dataset --local-dir ./HODome
# MHR ships as per-subject tars:
cd ./HODome/mhr && for f in *.tar*; do tar -xf "$f"; done

Then use the NeuralDome_Toolbox to visualize (scripts/hodome_visualize_pyrender.py --source smplx).

Videos

The 76-view 4K RGB videos and refined masks are hundreds of GB and are distributed via the project's Google Drive (see the toolbox README's one-click scripts/download_hodome.py).

Citation

@inproceedings{zhang2023neuraldome,
  title={NeuralDome: A Neural Modeling Pipeline on Multi-View Human-Object Interactions},
  author={Zhang, Juze and Luo, Haimin and Yang, Hongdi and Xu, Xinru and Wu, Qianyang and Shi, Ye and Yu, Jingyi and Xu, Lan and Wang, Jingya},
  booktitle={CVPR}, year={2023}}

License: CC BY-NC 4.0 (research use). Body models (SMPL-X/SMPL/MANO) are not redistributed — obtain them from the official sites per the toolbox instructions.