ICCV Daily 2021 - Thursday

“ We divide each of the scenes into shots, but these shots are already edited, so we only have the parts that were shown in the movie, not the parts that were cut out, ” Alejandro explains. “ To solve this, we create a proxy task . That’s a task which is similar to a real task, but not exactly the same. Let’s say we have a pair of shots. The left- hand side shot is joined with its last frame to the first frame of the right- hand side shot. We’re going to imagine that these shots were joined in different places and create artificial cuts. Then we’re going to say, okay, give me the real one and all the other ones are negatives. We do this using contrastive learning . We create a lot of negatives and only one positive and are trying to learn embedding space that allows us to divide positives and 17 DAILY ICCV Thursday Alejandro Pardo