Convert human motion from video to .bvh with Google Colab

Note: This version is a fork from Dene33. I changed some parts of code to make the project work on Colab of Python 2.7 and tf 1.15

1. Go to https://colab.research.google.com
2. File > Upload notebook... > GitHub > Paste this link: https://github.com/Dene33/video_to_bvh/blob/master/video_to_bvh.ipynb
3. Ensure that Runtime > Change runtime type is Python 2!!! with GPU

Second step is to install all the required dependencies. Select the first code cell and push shift+enter. You'll see running lines of executing code. Wait until it's done (1-2 minutes).

1. Select the code cell and push shift+enter
2. Push select files button
3. Select the video you want to process (it should contain only one person, all body parts in frame, long videos will take a lot of time to process. IMPORTANT: Make sure your video's name contains numbers only.)

1. Specify desired fps rate at which you want to convert video to images. Lower fps = faster processing
2. Select the code cell and push shift+enter

This step does all the job:

1. Convertion of video to images (images are required for pose estimation to work)
2. 2d pose estimation. For each image creates corresponding .json file with 2djoints with format similar to output .json format of original openpose. Fork of keras_Realtime_Multi-Person_Pose_Estimation is used.
3. 3d pose estimation. Creates .csv file of all the frames of video with 3d joints coordinates. Fork of End-to-end Recovery of Human Shape and Pose
4. Convertion of estimated .csv files to .bvh with help of custom script with .blend file.

1. Select the code cell and push shift+enter .bvh will be saved to your PC.
2. If you want preview it, run Blender on your PC. File > Import > Motion Capture (.bvh) > alt+a

1. Select the code cell and push shift+enter.