This repository is the official implementation of Learning Explicit Contact for Implicit Reconstruction of Hand-held Objects from Monocular Images (AAAI 2024).

• Python 3.8

conda create --no-default-packages -n choi python=3.8
conda activate choi

• PyTorch is tested on version 1.8.0

conda install pytorch==1.8.0 torchvision==0.9.0 cudatoolkit=11.1.1 -c pytorch -c conda-forge

• Other packages are listed in requirements.txt

pip install -r requirements.txt

• Unzip weights.zip and the pre-trained model is placed in the ./weights/ho3d/checkpoints directory

• Unzip data.zip and the processed data and corresponding SDF files are placed in the ./data directory

• Download the HO3D dataset and put it into the ./data/ho3d directory

• Download the MANO model MANO_RIGHT.pkl and put it into the ./externals/mano directory

• To evaluate my model on HO3D, run:

python -m models.choi --config-file experiments/ho3d.yaml --ckpt weights/ho3d/checkpoints/ho3d_weight.ckpt

• The resulting file is generated in the ./output directory

Our method achieves the following performance on the HO3D test set:

If you find our work useful in your research, please consider citing:

@inproceedings{hu2024learning,
  title={Learning Explicit Contact for Implicit Reconstruction of Hand-held Objects from Monocular Images},
  author={Hu, Junxing and Zhang, Hongwen and Chen, Zerui and Li, Mengcheng and Wang, Yunlong and Liu, Yebin and Sun, Zhenan},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2024}
}

Part of the code is borrowed from IHOI, Neural Body, and MeshGraphormer. Many thanks for their contributions.