CoMusion

CoMusion: Towards Consistent Stochastic Human Motion Prediction via Motion Diffusion (ECCV'2024)

by Jiarui Sun, and Girish Chowdhary.

[Project Page] [Paper]

This repository contains the official PyTorch implementation of CoMusion.

We develop a single-stage diffusion-based stochastic HMP framework to predict accurate, realistic, and consistent motions with respect to motion history.

Installation

1. Environment

Python/conda/mamba environment

Coming Soon!

2. Datasets

> Human3.6M

We follow https://github.com/wei-mao-2019/gsps for Human3.6M dataset preparation. All data needed can be downloaded from Google Drive and place all the dataset in data folder inside the root of this repo.

> AMASS

We follow https://github.com/BarqueroGerman/BeLFusion for AMASS dataset preparation. Due to the distribution policy of AMASS dataset, we are not allowed to distribute the data directly. Please reach out if you have questions.

Evaluation

Run the following scripts to evaluate CoMusion.

Human3.6M:

python train.py --cfg h36m --test

AMASS:

python train.py --cfg amass --test

Training

Run the following scripts to train CoMusion.

Human3.6M:

python train.py --cfg h36m

AMASS:

python train.py --cfg amass

Citation

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

Coming Soon!

Note: We thank German Barquero for the BeLFusion code and his prompt QA and support. We also borrow parts from GSPS by Wei Mao.