---
license: mit
tags:
- 4d-hof
- ho3d
- hand-object-reconstruction
library_name: pytorch
---

# 4D-HOF: HO3D release assets

Checkpoint and prepared initialization for **4D-HOF: Hand-Object Flow Matching for Feed-Forward 4D Interaction Reconstruction**.

[Project page](https://tamu-visual-ai.github.io/4D-HOF/) · [Paper](https://arxiv.org/abs/2610.08782)

## Contents

- `_DATA/ho3dv3_init.tar.gz`: initialization bundles for 18 HO3D sequences.
- `checkpoints/4dhof.pth`: full flow matching checkpoint, provided uncompressed.
- `_DATA/SHA256SUMS`: SHA-256 checksums for the archive and checkpoint files.
- `_DATA/ho3dv3_init.manifest.json`: initialization file hashes, sequence list, and checkpoint hashes.

## Download

Install the dependencies using `bash setup_env.sh` in the 4D-HOF code repository, then activate the environment with `conda activate 4dhof`. From the code repository root, run:

```bash
hf download eggplant71/4D-HOF --repo-type model \
  --include "_DATA/*" "checkpoints/*" --local-dir .
sha256sum -c _DATA/SHA256SUMS
tar -xzf _DATA/ho3dv3_init.tar.gz -C _DATA
```

For an existing environment, `python -m pip install -r requirements.txt` installs the Hugging Face download dependencies.

Extraction creates `_DATA/ho3dv3_init/<sequence>/`, containing `contact.pkl`, `hand.pkl`, `intrinsics.pkl`, `mask.pkl`, `mesh.pkl`, and `pose.pkl`. The checkpoint is downloaded directly to `checkpoints/4dhof.pth`.

Obtain MANO and HO3D separately. Place `MANO_RIGHT.pkl` at `_DATA/mano/MANO_RIGHT.pkl` as described in the code repository. Only load trusted pickle and checkpoint files.

## Run

From the code repository root:

```bash
python infer_ho3d.py --ho3d-root /path/to/HO3D_v3
python eval_ho3d.py --ho3d-root /path/to/HO3D_v3
```