This repo contains the official implementation of our paper:

DLow: Diversifying Latent Flows for Diverse Human Motion Prediction
Ye Yuan, Kris Kitani
ECCV 2020
[website] [paper] [talk] [summary] [demo]

• Please follow the data preprocessing steps (DATASETS.md) inside the VideoPose3D repo. Place the prepocessed data data_3d_h36m.npz (Human3.6M) and data_3d_humaneva15.npz (HumanEva-I) under the data folder.

• Tested OS: MacOS, Linux
• Packages:
  • Python >= 3.6
  • PyTorch >= 0.4
  • Tensorboard
• Note: All scripts should be run from the root of this repo to avoid path issues.

• Download our pretrained models from Google Drive (or BaiduYun, password: y9ph) and place the unzipped results folder inside the root of this repo.

We have provided 4 example YAML configs inside motion_pred/cfg:

• h36m_nsamp10.yml and h36m_nsamp50.yml for Human3.6M for number of samples 10 and 50 respectively.
• humaneva_nsamp10.yml and humaneva_nsamp50.yml for HumanEva-I for number of samples 10 and 50 respectively.
• These configs also have corresponding pretrained models inside results.

python motion_pred/exp_vae.py --cfg h36m_nsamp10

python motion_pred/exp_dlow.py --cfg h36m_nsamp10

python motion_pred/eval.py --cfg h36m_nsamp10 --mode vis

Useful keyboard shortcuts for the visualization GUI:

python motion_pred/eval.py --cfg h36m_nsamp50 --mode stats

If you find our work useful in your research, please cite our paper DLow:

@inproceedings{yuan2020dlow,
    title={Dlow: Diversifying latent flows for diverse human motion prediction},
    author={Yuan, Ye and Kitani, Kris},
    booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
    year={2020}
}

Part of the code is borrowed from the VideoPose3D repo.

The software in this repo is freely available for free non-commercial use. Please see the license for further details.