Dance Me If You Can - Motion Transference using AI
This week's Paper of the Week can be found here: Everybody Dance Now
I don't know about you, but I've never won a dance-off. I wouldn't consider myself a particularly bad dancer, but the fact of the matter is that there is almost always someone who shows up on the dance floor and starts busting out moves that most of us have only ever seen in choreographed music videos. Now, unlike many of my previous Papers of the Week, this week we are not talking about life or death medical scenarios. As much as we all hate to lose, by the next morning, we return to our normal lives, which presumably do not resemble Footloose.
Instead, we might assuage our dancing deficiencies by watching YouTube choreography or music videos with interesting dance scenes. We might even look into online dance competitions, but only as a spectator, never as a competitor. Until recently, I would wager that most people had come to terms with the current state of their dance skills and were not in a hurry to find a workaround to winning online dance competitions.
Fortunately for your future online dance competitions, researchers at the University of California, Berkeley, have your back. Using artificial intelligence, they've developed a way to transfer the motion of a dancer in a music video directly to you - well, a video of you.
Using video of a professional dancer, they trained an artificial intelligence model to detect the current post of the subject of the video and apply that pose to an image of another person. As the AI "watches" the video, it creates a parallel video of the target person performing the same dance as in the original video. It also compensates for differences in distance from the camera and facial expressions to create the most realistic video possible. You can see some of them here.
Obviously, this usage of artificial intelligence isn't going to cure cancer, save the economy, or end world hunger. However, it is a fun demonstration of a popular subfield of artificial intelligence, which focuses on how we can use algorithms to transfer motions from one video to another. It is important to remember that this kind of research can have some harrowing outcomes, as many people noticed earlier this year when a paper was published showing facial motion transference from random actors to former presidents. Ethical incorporation of artificial intelligence, including motion transference, has the potential to have really positive impacts on society at large. But for now, we'll just have to settle for winning the next dance-off (or actually learning to dance).