DeepSimHO: Stable Pose Estimation for Hand-Object Interaction via Physics Simulation

NeurIPS, 2023
Rong Wang . Wei Mao · Hongdong Li

Installation

Environment

• Create a conda venv and install pytorch:

conda env create -f environment.yml
conda activate deepsimHO
pip install torch==1.10.1+cu111 torchvision==0.11.2+cu111 -f https://download.pytorch.org/whl/cu111/torch_stable.html

• Install base dependencies:

pip install -r requirements.txt

• Install additional packages for compatability issues:

pip install kaolin==0.13.0 -f https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-1.10.1_cu111.html
pip install jaxlib==0.1.65 -f https://storage.googleapis.com/jax-releases/jax_releases.html

• Follow the instructions to install dex-ycb-toolkit. Note you should clone it into main/thirdparty instead.

• [Optional] If you are running on a headless server, you should create virtual displays for MuJoCo:

Xvfb :1 &
export DISPLAY=:1

Data Preperation

• Follow the docs to prepare datasets for HO3D and DexYCB and download MANO assets, the datasets should be extracted into main/data and main/assets.

• We provide cached DexYCB and HO3D index for train and test set indicating samples whose ground truth annotations are stable in our settings for MuJoCo, download and extract the folder into main/common.

• We provide MeshLab resampled YCB object models to train object metrics, download it and extracted to main/data. MuJoCo related data can be found in MuJoCo_data.

• Your main folder should have the structure as below:

    ├── data
    │   ├── DexYCB
    │   ├── HO3D
    │   ├── models_resample_mashlab
    |   ├── YCB_models_process
    │   └── YCB_models_supp
    ├── assets
    │   ├── postprocess
    │   ├── mano_v1_2
    │   └── ... (download from ArtiBoost)
    ├── checkpoints
    │   ├── pretrained
    │   |   ├── dexycb.pth.tar
    |   |   └── ho3d.pth.tar
    ├── common
    |   ├── cache
    │   |   ├── DexYCB
    |   |   └── HO3D
    ├── thirdparty
    │   ├── dex-ycb-toolkit
    |   ...

Evaluation

HO3Dv2-

Download the pretrained model into main/checkpoints/pretrained and run the below command.

cd main
CUDA_VISIBLE_DEVICES=0 python scripts/main.py --cfg config/test/test_ho3dv2-_artiboost_pretrain.yaml --gpu_id 0 --evaluate'

DexYCB

Download the pretrained model into main/checkpoints/pretrained and run the below command.

cd main
CUDA_VISIBLE_DEVICES=0 python scripts/main.py --cfg config/test/test_dexycb_artiboost_pretrain.yaml --gpu_id 0 --evaluate'

If runs properly, you should find the results like below. Note the physics metrics may vary slightly due to the simulation precision.

HO3D:
test Epoch 0 | final_loss: 3.3620e-03 | joints_3d_abs_mepe: 97.3570 | corners_3d_abs_mepe: 52.7965 | CP: 0.9659 | PD: 0.0113 | SD: 0.0243 | SR: 0.1917: 100%|███| 48/48

DexYCB:
test Epoch 0 | final_loss: 1.7286e+00, sym_corners_3d_loss: 1.7285e+01 | joints_3d_abs_mepe: 11.2349 | CP: 0.9590 | PD: 0.0142 | SD: 0.0242 | SR: 0.3283: 100%|███| 50/50

Train

We take training on DexYCB as the example. First train the DeepSim MLP independently:

python scripts/main.py --cfg config/train/train_dexycb_artiboost_mlp.yaml --gpu_id 0

Then jointly train with a base network (We use pretrained Heatmap-based, Object symmetry model in ArtiBoost):

python scripts/main.py --cfg config/train/train_dexycb_artiboost_pretrain.yaml --gpu_id 0

Note that training with MuJoCo is mostly done in CPU, where we use multi-process simulation for speed up.

Citation

If you use the code for your research, please cite with the below:

@inproceedings{
    wang2023deepsimho,
    title={DeepSim{HO}: Stable Pose Estimation for Hand-Object Interaction via Physics Simulation},
    author={Rong Wang and Wei Mao and Hongdong Li},
    booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
    year={2023},
    url={https://openreview.net/forum?id=SxVHyYavHy}
}

Acknowledge

We use ArtiBoost as the code base and thank the authors for releasing the code.