A guess would be more movie meta data than you can think of; - show me movies with Swiss Alps settings - show me movies with the quote "I'll be back" - tell me which part of Back to the Future shows me where they have hoverboards - actor/director/artist/crew: for my IMDB profile, tell me when my name is on screen (start credits, end credits)

I would guess it’s in order to build the hash/IDs/searchable dataset that can be used to match the movies. Think Shazam style fingerprints but for movies.

I was looking into this sort of thing recently for a project. Very interesting technology space with some genuine direct benefits coming out of the application of machine learning.

I'm guessing something along the lines of Morbotron [1]. Not the actual use case for the screens/subs/etc. but more the data required to do either of them: frame by frame images for analysis with timestamps and subtitles.

1: https://morbotron.com/

We needed high quality video files to create the metadata. Machine vision wasn’t good enough yet so we did it the old fashion way. Interns and lots of typing in real time