If you’ve been cursed with two left feet and have a hard time pulling off even the simplest dance moves, you might not have much longer to worry about not being able to impress your partner or your friends on the dance floor. A team of researchers from the University of California in Berkeley has come up with a new deep learning system, and as reports have suggested, it could make even the worst dancer appear capable once the technology has been refined.
According to Engadget, the deep learning system works by taking a source video with an experienced, skilled dancer and “translating” their moves so that it appears as if the less skilled target dancer is actually performing them with ease. The system makes use of three algorithms, including one that creates a virtual stick figure to map the moves and poses, and two others that help “create the full picture” and make bad dancers appear realistically competent.
While the target dancer would still need to move around for a few seconds so that the system has a point of reference, Engadget noted that preliminary results have been “realistic” enough in making unskilled dancers look more graceful.
As seen in a paper published earlier this week on arXIV, the UC Berkeley researchers seem to have made their artificial intelligence-based system work reasonably well for different types of dance, from classical ballet to modern styles. However, there’s more to the new technology than simply transferring the moves of a good dancer to those of someone less skilled. The Verge wrote that the algorithms have sub-routines that allow the stick figure’s motions to be further refined to reduce jerky motions, and also re-trace the face of the target dancer to make things look even more realistic.
An AI program will soon be here to help your deepfake dancing - just don't call it deepfake https://t.co/wRNnGsaGJ0— Franklin Ockey (@franklinoap) August 27, 2018
Although the deep learning network has the potential to help bad dancers “fake it till they make it” with reasonable accuracy, the technology has its share of limitations, as pointed out by both Engadget and The Verge. As the system still has difficulty modeling people wearing loose clothing, target dancers need to wear tight clothes in order for the technology to work. There are also some inconsistencies in the dancers’ movements, as the joints of both the source and target dancer don’t always match up on video. Furthermore, the software used in the AI system still has some difficulty recreating more complex moves and gestures.
In addition to the aforementioned limitations, Engadget wrote that there might be some ethical concerns involved when making bad dancers look good through the use of AI. There’s also the chance that the system could be abused, much like some previous applications of “deepfake” technology. However, the publication stressed that the new system has lots of potential, and could prove useful if it is utilized mainly for fun applications, such as making overly serious coworkers look goofier and more fun-loving than they usually are.