INSTALL.md

Set up the python environment

# python 3.7 or 3.8 are both fine. 
conda create -n nhp python=3.x
conda activate nhp

# make sure that the pytorch cuda is consistent with the system cuda
# install pytorch via conda
https://pytorch.org/get-started/locally/

# install additional requirements
pip install -r requirements.txt

# install requirement for mesh rendering
pip install PyOpenGL PyOpenGL_accelerate
sudo apt-get install freeglut3-dev

# install spconv
# Sparse Convolution now provides easy installation via pip
# Go to the SpConv Github and install the proper version
https://github.com/traveller59/spconv

Set up datasets

ZJU-Mocap dataset

1. To request the ZJU-Mocap dataset download link, please fill in the agreement, and email to the Neural Body author ([email protected]) and cc Xiaowei Zhou ([email protected]).
2. Create a soft link:

   ROOT=/path/to/Neural_Human_Performer
   cd $ROOT/data
   ln -s /path/to/zju_mocap zju_mocap
   

3. Our framework uses the newly fitted SMPL parameters and vertices of the Neural Body.
   Rename the new_params folder into params and the new_vertices folder into vertices.
4. Download the visibility maps at here.
5. Create a soft link:

   ROOT=/path/to/Neural_Human_Performer
   cd $ROOT/data
   ln -s /path/to/zju_rasterization zju_rasterization
   

6. Change the file name format of the subject 313 and 315:

   python lib/utils/modify_313_315_filename.py
   

Download SMPL

1. Go to SMPL website and sign up.
2. Download the SMPL 1.0.0
3. Create a directory with the following structure:

data
└── smplx
   ├── smpl
   │   ├── SMPL_FEMALE.pkl
   │   └── SMPL_MALE.pkl
   │   └── SMPL_NEUTRAL.pkl
   └── J_regressor_body25.npy