Now, one thing that humans have always wanted is something that can predict the future. Well, we haven’t been even remotely close to that dream. Still, intelligent minds across the globe have made efforts to create systems that can predict events and outcomes, for instance, the result of a game.
A team of researchers from Google, Adobe, and the University of Michigan have made such an effort. Their new AI can tell us what will happen next in a video. Yes, after feeding 3 to 4 frames, it can predict up to 25 future frames of the future.
They focused on three prediction categories namely, object interactions, structured motion, and partial observability. And for that, they trained different versions of their model using videos of robot arm interacting with towels, humans doing different actions, and KITTI driving data sourced from dashboard camera footage.
Their model grows on top of the stochastic video generation (SVG) and adds a component to model inherent uncertainty in future predictions. Furthermore, it doesn’t rely on prediction techniques used in previous methods, like optical flows.