Real-time RGBD-based Extended Body Pose Estimation

This repository is a real-time demo for our paper that was published at WACV 2021 conference

The output of our module is in SMPL-X parametric body mesh model:

• RNN estimates body pose from joints detected by Azure Kinect Body Tracking API
• For face (expression and jaw) and hand pose we crop from rgb image:
  • for hand model we use minimal-hand
  • our face NN takes media-pipe keypoints as input

Combined system runs at 30 fps on a 2080ti GPU and 8 core @ 4GHz CPU.

How to use

Build

• Prereqs: your nvidia driver should support cuda 10.2, Windows or Mac are not supported.
• Clone repo:
  • git clone https://github.com/rmbashirov/rgbd-kinect-pose.git
  • cd rgbd-kinect-pose
  • git submodule update --force --init --remote
• Docker setup:
  • Install docker engine
  • Install nvidia-docker
  • Set nvidia your default runtime for docker
  • Make docker run without sudo: create docker group and add current user to it:

    sudo groupadd docker
    sudo usermod -aG docker $USER
    

  • reboot
• Create docker image:
  • Build on your own: run 2 cmds
  • Or download docker image from baidu
• Attach your Azure Kinect camera
• Check your Azure Kinect camera is working inside Docker container:
  • Enter Docker container: ./run_local.sh from docker dir
  • Then run python -m pyk4a.viewer --vis_color --no_bt --no_depth inside docker container

Download data

• Download our data archive smplx_kinect_demo_data.tar.gz
• Unzip: mkdir /your/unpacked/dir, tar -zxf smplx_kinect_demo_data.tar.gz -C /your/unpacked/dir
• Download models for hand, see link in "Download models from here" line in our fork, put to /your/unpacked/dir/minimal_hand/model
• To download SMPL-X parametric body model go to this project website, register, go to the downloads section, download SMPL-X v1.1 model, put to /your/unpacked/dir/pykinect/body_models/smplx
• /your/unpacked/dir should look like this
• Set data_dirpath and output_dirpath variables in config file:
  • data_dirpath is a path to /your/unpacked/dir
  • output_dirpath is used to check timings or to store result images
  • ensure these paths are visible inside docker container, set VOLUMES variable here

Run

• Run demo: in src dir run ./run_server.sh, the latter will enter docker container and will use config file where shape of the person is loaded from an external file: in our work we did not focus on person's shape estimation

What else

Apart from our main body pose estimation contribution you can find this repository useful for:

• minimal_pytorch_rasterizer python package: CUDA non-differentiable mesh rasterization library for pytorch tensors with python bindings
• pyk4a python package: real-time streaming from Azure Kinect camera, this package also works in our provided docker environment
• multiprocessing_pipeline python package: set-up pipeline graph of python blocks running in parallel, see usage in server.py

Citation

If you find the project helpful, please consider citing us:

@inproceedings{bashirov2021real,
  title={Real-Time RGBD-Based Extended Body Pose Estimation},
  author={Bashirov, Renat and Ianina, Anastasia and Iskakov, Karim and Kononenko, Yevgeniy and Strizhkova, Valeriya and Lempitsky, Victor and Vakhitov, Alexander},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={2807--2816},
  year={2021}
}

Non-commercial use only