When training an algorithm using video, seek the minimal overhead because you're going to be decompressing several billion frames.