Facebook’s A.I. researchers have now released a new report and associated whitepaper showing how they managed to teach A.I. to map human pixels from a static RGB image to a 3D surface. That is, of course, the simplified explanation of the extremely complex process, which the team refers to as Dense Human Pose Estimation. In effect, what DensePose does is create a 3D wireframe of the human body in real-time using A.I. and working at several frames per second. The system can then use that data to coordinate the placement of a realistic “skin” over the top of the wireframe. It’s able to accomplish thanks to the creation of a database containing over 50,000 images. Human workers annotated those images with more than 5 million data points in order to teach the convolutional A.I. to see human bodies and movement in a way that is closer to the way humans do.
The technology could be used to great effect with regard to remapping human bodies with the attributes of a completely different person. For example, it could be used to superimpose an actor onto the body of a stuntman with far greater efficiency than previous methods. However, the implications do continue far beyond that. In fact, Facebook may seek to use its newly discovered methods in order to train its A.I. to be more aware for use in its security measures and in more humanitarian efforts. For example, it could augment the company’s proactive efforts for detecting when users are feeling suicidal which were revealed last year. It might also eventually find use in other, similarly safety-geared solutions such as being used to read body language to predict dangerous or abnormal behaviors out in the real world instead of just online. It goes without saying that none of those is necessarily what Facebook is aiming for but the possibilities presented by an A.I. that can read humans well enough to remap them in real time are essentially only limited by the scope of imagination.
Obviously, the research document associated with DensePose contains a staggering amount of added detail and information about this breakthrough and its implications. There is far more information included there than could ever be included in a single article. Having said that, anybody interested in checking it out for themselves can hit the button below for the full whitepaper as a PDF file.
DensePose: Dense Human Pose Estimation In The Wild PDF