iSEE is claimed to learn object permanence from unlabeled video

A post describes iSEE as learning object permanence from video without labels. Its author says the model follows an object while it is hidden and recognizes it when it returns, contrasting that with today’s world models, which the author says lose objects once they go out of view.