Overview

• We trained generative adversarial networks (GANs) with conditions for synthetic gait data generation (marker, joint angles and ground reaction forces/moments) and used the generated data to augment the existing experimental datasets.

• Check out the web app where we deploy our generative models to generate, animate and download synthetic data.

• Please refer to our approach explained in the following papers:

@article{Bicer2022Generative,
  title={Generative Deep Learning Applied to Biomechanics: A New Augmentation Technique for Motion Capture Datasets},
  author={Metin Bicer, Andrew TM Phillips, Alessandro Melis, Alison H McGregor, Luca Modenese},
  journal={Journal of Biomechanics},
  volume = {144},
  pages = {111301},
  year = {2022},
  issn = {0021-9290},
  doi={https://doi.org/10.1016/j.jbiomech.2022.111301},
  url = {https://www.sciencedirect.com/science/article/pii/S0021929022003426},
  keywords = {Gait, Deep Learning, Generative Adversarial Networks}
}

@article{Bicer2024cGAN,
  title={Generative adversarial networks to create synthetic motion capture datasets including subject and gait characteristics},
  author={Metin Bicer, Andrew TM Phillips, Alessandro Melis, Alison H McGregor, Luca Modenese},
  journal={Journal of Biomechanics},
  volume = {##},
  pages = {112358},
  year = {2024},
  issn = {0021-9290},
  doi={https://doi.org/10.1016/j.jbiomech.2024.112358},
  url = {https://www.sciencedirect.com/science/article/pii/S0021929024004366},
  keywords = {Gait, Synthetic Mocap Dataset, Conditional Generative Adversarial Networks}
}

Our Approach

Here is an example of a real and a synthetic trial from the first paper:

Synthetic data were validated by means of statistical comparisons to the experimental data, as presented below from the second paper. This analysis was was carried out using dimensionless walking speed and averaging the sagittal plane kinematics and vertical ground reaction force by dividing trials into walking speed groups.

Requirements

1. create and activate a virtual environment (on Anaconda)
   using the environment.yml file

   conda env create --file environment.yml
   conda activate cgan
   

   OR directly from Anaconda by installing the following packages

   conda create --name cgan
   conda activate cgan
   

   • torch
   • pandas
   • nump
   • scipy
   • matplotlib
   • ezc3d
   • spm1d
   • sklearn
   • statsmodels
   • spm1d
2. With the environment activated, navigate to the OpenSim installation directory and install OpenSim API

   cd OPENSIM_INSTALL_DIR/sdk/Python
   python setup.py install	
   

Run

1. train_cgans.py: contains all codes required to train a conditional GAN and generate data with condition(s). Examples are given on lines 675-974. Training data is saved under data folder.
2. compare_train_data.py: reproduces training dataset characteristics from saved synthetic datasets (under Results folder). Condition ranges are given on lines 18-34.
3. dimensionless_walking_speed_analysis.py: carries out dimensionless walking speed analysis on the experimental training data and the synthetic dataset generated by the Multi-cGAN. Plots joint angles and vertical ground reaction forces, and compares the two set of data.
4. dynamical_consistency.py: reproduces dynamical inconsistency analysis for the WS-cGAN and Multi-cGAN. It can be used for other conditional models if inverse dynamics results are saved under Results/opensim folder.
5. previous_papers.py: reproduces results from previous papers by the synthetic datasets saved under Results/papers folder.
6. compare_test_data.py: reproduces plots, error metrics for comparing test experimental data to its synthetic replica generated by all conditional GANs.