Adversarial Parametric Pose Prior

This is a Pytorch implementation of the CVPR'22 paper "Adversarial Parametric Pose Prior".

You can find the paper here.

Examples of spherical interpolations when sampling from GAN-S pose prior

Installation

We provide the conda environment for Linux. To create and activate it, do:

conda env create -f environment.yaml
conda activate adv_prior

The main requirements are:

• python=3.8, numpy, matplotlib
• scipy, scikit-learn, jupyter
• easydict, pyyaml, tqdm
• pytorch=1.7, cudatoolkit=10.2
• pytorch3d
• smplx
• torchgeometry

As for SMPL mesh, put the basicModel_neutral_lbs_10_207_0_v1.0.0.pkl pickle file from SMPLIFY_CODE_V2.ZIP in data/. Rename the file, replacing basicModel with smpl.

Technical Details

Note: We make use of the torchgeometry library which contains an error. For more details please see this StackOverflow thread. Essentially, the file {ANACONDA_HOME}/envs/adv_prior/lib/python3.8/site-packages/torchgeometry/core/conversions.py must be updated in the function rotation_matrix_to_quaternion:

• line 302: mask_c1 = mask_d2 * ~(mask_d0_d1) ~~mask_d2 * (1 - mask_d0_d1)~~
• line 303: mask_c2 = ~(mask_d2) * mask_d0_nd1 ~~(1 - mask_d2) * mask_d0_nd1~~
• line 304: mask_c3 = (~(mask_d2)) * (~(mask_d0_nd1)) ~~(1 - mask_d2) * (1 - mask_d0_nd1)~~

Training

To launch the training of the GAN-S model on AMASS data, run:

python run/main.py --cfg experiments/train_gan.yaml

Demos

We provide two short demo jupyter notebooks for sampling from and interpolating in the latent space. All details can be found here.