DATA.md

Data Preparation

AVA

First, download AVA training and validation videos. An example script can be found at download_videos.sh.

Then, extract frames between 15th and 30th minutes. You may use our implementation at tools/extract_frames.py. To execute it, use the following command, which will save frames to the data folder:

python tools/extract_frames.py --video_dir YOUR_VIDEO_DIRECTORY --frame_dir data [--num_processes NUM_PROCESSES]

The final data folder should have the following structure:

data
|_ [AVA train_val video name 0]
|  |_ image_000001.jpg
|  |_ image_000002.jpg
|  |_ ...
|_ [AVA train_val video name 1]
|  |_ image_000001.jpg
|  |_ image_000002.jpg
|  |_ ...
|_ ...
|_ [AVA train_val video name 298]
   |_ image_000001.jpg
   |_ image_000002.jpg
   |_ ...