• Windows/MacOS/Linux
• Docker client
• Optional: NVIDIA docker for GPU support (10x faster processing)

You don't need to download anything from this repository -- docker run command will automatically download OpenPose/VideoPose3D from the default repository.

OpenPose returns 2d keypoints for each frame in the video. See OpenPose repository for details.

Create a directory (say data as in the example below) and put your video (e.g. video.mp4) in that directory.

For CPU processing of your video run:

docker run -v $(pwd)/data:/openpose/data stanfordnmbl/openpose-cpu\
  /openpose/build/examples/openpose/openpose.bin\
  --video /openpose/data/video.mp4\
  --display 0\
  --write_json /openpose/data/keypoints\
  --render_pose 0

Note that CPU processing is very slow -- it'll take at least 15x the duration of the video to process.

GPU processing is much faster (around real time) but it requires a GPU and NVIDIA docker. For GPU processing run:

docker run --gpus=1 -v $(pwd)/data:/openpose/data stanfordnmbl/openpose-gpu\
  /openpose/build/examples/openpose/openpose.bin\
  --video /openpose/data/video.mp4\
  --display 0\
  --write_json /openpose/data/keypoints\
  --render_pose 0

Resultis will be store in data/keypoints directory.

In the scripts above $(pwd)/data corresponds to the path in which your data directory is stored.

VideoPose3D returns 3d keypoints for each frame in the video. See VideoPose3D repository for details.

GPU and NVIDIA docker are required to run this software

Create input, output, and models directories. Put your mp4 video in the input directory.

Run

docker run --gpus all --rm\
  -v $(pwd)/input:/data/input\
  -v $(pwd)/output:/data/output\
  -v $(pwd)/models:/data/models\
  -it stanfordnmbl/videopose3d

Results will be saved in output.

In the first run, deep learning models will be downloaded to models directory which can take some time. These models will be saved for all subsequent runs which will make it much faster.

In this repository we only provide instructions on how to run software. Corresponding licenses apply.

• OpenPose: See OpenPose repository
• VideoPose3D: See VideoPose3D repository