Preprocessing the videos

First step is to rescale the videos. We will rescale the entire datasets to a resolution of 256x256 and a fps of 30. Such operation can be one using ffmpeg.

VLOG dataset

In the following I will assume that you have download the VLOG dataset and the dataset is located in the directory /data/vlog/videos. I have attached 3 videos of the dataset in the Github repo to make sure you can replicate the operation on your own dataset, below is the paths to the two videos:

• /data/vlog/videos/O/B/o/v_eAalQCMBOBo/049/clip.mp4
• /data/vlog/videosa/q/g/v_4jqvo3TQaqg/016/clip.mp4
• /data/vlog/videosa/q/g/v_4jqvo3TQaqg/016/clip.mp4

For applying the preprocessing on the full directory and create a new one you can run the following command:

cd preprocessing
python rescale.py --width 256 --height 256 --fps 30

It will create a new directory named /data/vlog/videos_256x256_30 with the same structure of the initial one. You will also get a pickle file called dict_id_length.pickle in your new directory, the length of each videos will be saved into this dictionnary (it will be used while training/testing). Feel free to check the rescale.py to check other possible arguments (e.g. crf, ffmpeg version, suffix). The generic command is:

python rescale.py \
--dir <LOC-VIDEO-DIR> \
--width <OUTPUT-W> \
--height <OUTPUT-H> \
--fps <OUTPUT-FPS> \
--crf <OUTPUT-QUALITY> \
--ffmpeg <YOUR-FFMPEG> \
--suffix <SUFFIX-OF-ALL-VIDEOS>

You can now open the original video and the rescaled one using the following command:

# Original one
open /data/vlog/videos/O/B/o/v_eAalQCMBOBo/049/clip.mp4
du -sh /data/vlog/videos/O/B/o/v_eAalQCMBOBo/049/clip.mp4

# Rescaled one
open /data/vlog/videos_256x256_30/O/B/o/v_eAalQCMBOBo/049/clip.mp4
du -sh /data/vlog/videos_256x256_30/O/B/o/v_eAalQCMBOBo/049/clip.mp4

This preprocessing step allows us to reduce the size of our video by a certain amount as you can see by running the above lines (du -sh ...).

On the real VLOG dataset it takes ~15 hours to rescale the entire dataset.

You are now ready to test the input pipeline -> README_DATALOADER.md