GraphMDN: Leveraging graph structure and deep learning to solve inverse problems

This code implements the work in our paper of the same title, which can be found on arXiv.

This code "repository" contains the code necessary to train and evaluate the GraphMDN model introduced in the accompanying text. The code will be made public at a later date via github repository.

This code extends the SemGCN model to incorporate MDN structured outputs, and uses data preprocessing code from VideoPose3d

Dataset setup

You can find the instructions for setting up the Human3.6M and results of 2D detections in data/README.md. The code for data preparation is borrowed from VideoPose3D.

Evaluating our pretrained model

First, download the checkpoint archive from our "Release" page, named graphmdn_ijcnn_release_checkpoint.zip. Extract this archive into the root folder of the GraphMDN repository; if done correctly, GraphMDN/checkpoint should exist as a directory. From here, the pretrained model can be evaluated.

To evaluate, run:

python main_eval.py --evaluate checkpoint/trained_model/ckpt_best.pth.tar

Training from scratch

If you want to reproduce the results of our pretrained models, run the following commands.

python main_mdn.py

Additionaly training parameters and arguments can be seen in the common/arguments.py file.