Researchers at the University of Washington have used a machine learning neural network algorithm to render two-dimensional video clips in three dimensions. University of Washington researchers led by Konstantinos Rematas have taught a machine learning neural network algorithm to render two-dimensional (2D) video clips posted on YouTube as three-dimensional (3D) images. The researchers collected footage from the FIFA football video game as a training dataset, since the game estimates the position in three dimensions of each player, yielding data about their actual location as well as how they are displayed in 2D. Once that training was complete, the researchers were able to use the algorithm to transform YouTube clip imagery into three dimensions. Viewers using an augmented reality headset can see the enhanced versions of the clips as though it were positioned on a flat surface in front of them.

The football World Cup is almost upon us. Many millions will watch the tournament unfold on TV screens around the globe. But what if you could enjoy a mini virtual reconstruction of each match on your dining table instead? To create such an experience, Konstantinos Rematas and colleagues at the University of Washington trained a machine learning algorithm to convert 2D YouTube clips into 3D reconstructions. They began by gathering footage from the football videogame FIFA. "I had to play FIFA 2017 for, I don't know, … To continue reading this premium article, subscribe for unlimited access.