Initial directory structure:

$ROOT/
  yt_bb_detection_train.csv
  yt_bb_detection_validation.csv
  train/
  validation/

Note: Use bash and xargs to do slow things in parallel, and use python to do more involved tasks.

The following command will try to download all videos. It will not re-download videos that have been succesfully downloaded. It will also resume partial downloads that were interrupted.

$ download.sh $ROOT/train/download/ $ROOT/yt_bb_detection_train.csv

This script will create:

$ROOT/train/download/
  videos.txt
  complete/
  partial/

The video and audio will be saved in complete/[VIDEO]/[VIDEO].f[FORMAT].[EXT].

The script download.sh invokes download-one.sh many times in parallel using xargs -P. To change the number of parallel processes, change $num_parallel in download.sh. To modify the resolution that is downloaded, change $format in download-one.sh.

It is not obvious which file is the video and which is the audio. The script make_video_links.py can be used to create a directory with a single file per video.

$ python make_video_links.py $ROOT/train/download/ $ROOT/train/download/video/