movies are copyrighted, but metadata about movies doesn't have to be. the research dataset provided by david bamman contains face recognition data for almost 3.000 movies, giving us the locations and identities of actor's faces for every frame of a movie. we use this data to reconstruct the original movie, drawing the correct face at the correct position and time, 30 times a second, for the entire runtime of a movie. the result is a copyright-free version of the original movie, weighing in at only ~10 megabytes — orders of magnitude smaller than any conventional compression algorithm.
original dataset by david bamman (CC BY 4.0)
