As some of the dependencies for this project are non-trivial to install, the installation instructions are provided in this document. It is not necessary to follow these instructions exactly as long as all the necessary packages are installed.
We recommend using a new conda environment.
conda create -n contactopt python=3.8
conda activate contactopt
Detailed installation instructions can be found on the project website. Note that Pytorch3D places restrictions on the versions of Python and PyTorch used.
conda install -c pytorch pytorch=1.7.1 torchvision cudatoolkit=10.2
conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install pytorch3d -c pytorch3d
Detailed installation instructions can be found on the project website.
To ensure that the package has been installed correctly, test that the following snippet runs without error:
python -c "import torch; assert torch.cuda.is_available(); import torch_geometric.data"
Other dependencies play nicely with pip and can be installed with:
pip install git+https://github.com/hassony2/manopth.git open3d tensorboardX pyquaternion trimesh transforms3d chumpy opencv-python
Download the Python 3-compatible MANO code from the manopth website. Copy the mano folder from the manopth project to the root of the ContactOpt folder.
Due to license restrictions, the MANO data files must be downloaded from the original project website. Create an account and download 'Models & Code'. Extract the models folder to the recently created mano directory. The directory structure should be arranged so the following files can be found:
mano/webuser/lbs.py
mano/models/MANO_RIGHT.pkl
This step is required to interact with the ContactPose or Perturbed ContactPose dataset. This is a requirement for retraining the DeepContact network. Note that the Perturbed ContactPose data file is large (~40 GB) and requires a computer with ~64 GB of RAM.
Installation instructions can be found on the project website.
Edit the line at the top of contactopt/create_dataset_contactpose.py to point to the recently installed ContactPose directory.
sys.path.append('../ContactPose') # Change this path to point to the ContactPose repo
To generate the dataset files for Perturbed ContactPose, execute the following script. This may take up to an hour to complete.
python contactopt/create_dataset_contactpose.py
To evaluate the performance of ContactOpt on the results of a RGB pose estimator on the HO-3D dataset, download the pre-generated pose estimates file [mirror 1] (6 GB). Place it in the data folder.
The dataset file can be generated with:
python contactopt/create_dataset_im.py