Hierarchical Motion Understanding via Motion Programs

motion2prog_release

Hierarchical Motion Understanding via Motion Programs (CVPR 2021)
This repository contains the official implementation of:

Running motion2prog

0. We start with video file and first prepare the input data

$ ffmpeg -i ${video_dir}/video.mp4 ${video_dir}/frames/%05d.jpg
$ python AlphaPose/scripts/demo_inference.py 
    --cfg AlphaPose/pretrained_models/256x192_res50_lr1e-3_1x.yaml 
    --checkpoint AlphaPose/pretrained_models/halpe26_fast_res50_256x192.pth 
    --indir ${video_dir}/frames --outdir ${video_dir}/pose_mpii_track 
    --pose_track --showbox --flip --qsize 256
$ mv ${video_dir}/pose_mpii_track/alphapose-results.json 
    ${video_dir}/alphapose-results-halpe26-posetrack.json

We packaged a demo video with necessary inputs for quickly testing our code

$ wget https://sumith1896.github.io/motion2prog/static/demo.zip
$ mv demo.zip data/  && cd data/ && unzip demo.zip && cd ..

We need 2D pose detection results & extracted frames of video (for visualization)
We support loading from different pose detector formats in the load function in lkeypoints.py.
We used AlphaPose with the above

To finish reading, please visit source site