Learning 3D Human Dynamics from Video

Angjoo Kanazawa*, Jason Zhang*, Panna Felsen*, Jitendra Malik

University of California, Berkeley (* Equal contribution)

Requirements

Python 3 (tested on version 3.5)
TensorFlow (tested on version 1.8)
PyTorch for AlphaPose, PoseFlow, and NMR (tested on version 0.4.0)
AlphaPose/PoseFlow
Neural Mesh Renderer for rendering results. See below.
CUDA (tested on CUDA 9.0 with Titan 1080 TI)
ffmpeg (tested on version 3.4.4)

There is currently no CPU-only support.

Installation

Setup virtualenv

virtualenv venv_hmmr -p python3
source venv_hmmr/bin/activate
pip install -U pip
pip install numpy  # Some of the required packages need numpy to already be installed.
deactivate
source venv_hmmr/bin/activate
pip install -r requirements.txt

Install External Dependencies.

Neural Mesh Renderer and AlphaPose for rendering results:

cd src/external
sh install_external.sh

The above script also clones my fork of AlphaPose/PoseFlow, which is necessary to run the demo to extract tracks of people in videos. Please follow the directions in the installation, in particular running pip install -r requirements.txt from src/external/AlphaPose and downloading the trained models.

If you have a pre-installed version of AlphaPose, symlink the directory in src/external. The only change that my fork has is a very minor modification in AlphaPose/pytorch branch's demo.py: see this commit, copy over the changes in demo.py.

Demo

Download the pre-trained models. Place the models folder as a top-level directory.

wget http://angjookanazawa.com/cachedir/hmmr/hmmr_models.tar.gz && tar -xf hmmr_models.tar.gz

Download the demo_data videos. Place the demo_data folder as a top-level directory.

wget http://angjookanazawa.com/cachedir/hmmr/hmmr_demo_data.tar.gz && tar -xf hmmr_demo_data.tar.gz

Run the demo. This code runs AlphaPose/PoseFlow for you. Please make sure AlphaPose can be run on a directory of images if you are having any issues.

Sample usage:

# Run on a single video:
python -m demo_video --vid_path demo_data/penn_action-2278.mp4

# If there are multiple people in the video, you can also pass a track index:
python -m demo_video --track_id 1 --vid_path demo_data/insta_variety-tabletennis_43078913_895055920883203_6720141320083472384_n_short.mp4

# Run on an entire directory of videos:
python -m demo_video --vid_dir demo_data/

This will make a directory demo_output/<video_name>, where intermediate tracking results and our results are saved as video, as well as a pkl file. Alternatively you can specify the output directory as well. See demo_video.py

Training code

Coming soon.

Data

InstaVariety

We provided the raw list of videos used for InstaVariety, as well as the pre-processed files in tfrecords. Please see doc/insta_variety.md for more details..

Citation

If you use this code for your research, please consider citing:

@InProceedings{humanMotionKZFM19,
  title={Learning 3D Human Dynamics from Video},
  author = {Angjoo Kanazawa and Jason Y. Zhang and Panna Felsen and Jitendra Malik},
  booktitle={Computer Vision and Pattern Recognition (CVPR)},
  year={2019}

iamsile / human_dynamics