Pose2vec

This repository contains the following:

• Utilities for various human skeleton preprocessing steps in numpy and tensorflow.
• Tensorflow model to learn a continuous pose embedding space.

This code has been used to train the PoseGAN (or EnGAN) Model in the paper:
Maharshi Gor*, Jogendra Nath Kundu*, R Venkatesh Babu, "Unsupervised Feature Learning of Human Actions as Trajectories in Pose Embedding Manifold", IEEE Winter Conference on Applications of Computer Vision (WACV), 2019.

It is also used for training pose representations in the paper:
Maharshi Gor*, Jogendra Nath Kundu*, R Venkatesh Babu, "BiHMP-GAN: Bidirectional 3D Human Motion Prediction GAN", Thirty Third AAAI Conference on Artificial Intelligence, 2019.

Citing this work

If you find this work useful in your research, please consider citing:

@article{kundu2018unsupervised,
  title={Unsupervised Feature Learning of Human Actions as Trajectories in Pose Embedding Manifold},
  author={Kundu, Jogendra Nath and Gor, Maharshi and Uppala, Phani Krishna and Babu, R Venkatesh},
  journal={arXiv preprint arXiv:1812.02592},
  year={2018}
}

Data and Pretrained Weights

Use the following command to download the data and pretrained weights.

# For downloading the data. It will be saved in the data/ directory
python -m scripts.download_data

# For downloading the pretrained weights. It will be saved in the pretrained_weights/azimuth/ directory
python -m scripts.download_weights

Qualitative Results:

• Grid Interpolations
  

• Reconstructions (left: Ground Truth, right: Reconstruction)