New research teaches AI how people move with internet videos

New research at the University of Michigan can train neural network models to identify a person's position in videos where only a portion of their body is visible in the shot. This breakthrough opens up a huge library of video content to a new use--teaching machines the meaning behind people's poses, as well as the different ways they interact with their environment. When we see a picture of a pair of hands typing at a desk, we can infer there's a person attached to those somewhere out of sight. Depending on the angle of their wrists and how tall the desk is, we might even be able to tell whether they're sitting or standing. These are the kinds of inferences neural network models haven't been great at to this point.