This repository contains the code to tackle the HUMBI challenge.
It is based on the paper on Vunets for the pose synthesis: A Variational U-Net for Conditional Appearance and Shape Generation; source code for the same.
We added an additional GAN network layer. Below are some results on the training data of HUMBI.
We gained over 63% accuracy on the testing data.
The code was developed with Python 3. Dependencies can be installed with
pip install -r requirements.txt
Generate pickle files for training data paths and joint order match using:
python pickle_file.py and update the path in the config file ownconfig.yaml accordingly.
For running with the baseline (Vunet)
python main.py --config owndata.yaml --mode test --checkpoint <path to checkpoint of first round>
For running the proposed model
python main_new.py --config owndata.yaml --mode test --checkpoint <path to checkpoint of first round>
Tony Liu liux4408@umn.edu
Prithvi Raj Botcha botch025@umn.edu
Jiaqi Liu liu00687@umn.edu