500-camera dome trains computer to read body language

PITTSBURGH, PENNSYLVANIA (REUTERS) - Wave hi to the cameras!

A computer program working with a dome embedded with 500 video cameras has been able to analyse multiple individuals' hands and fingers in real-time.

"Panoptic Studio is a system to capture human motion without using any markers or special equipment," said Hanbyul Joo, a PhD student at Carnegie Mellon University. "And the goal of the system is to understand how we move our bodies during the communication. So during our communication, humans use specific channels - facial expressions, body gestures, and hand gestures. If we don't have markers on their bodies, the basic problem is it's very hard to track this part in 3D space. To solve the issue our basic approach is to have a lot of cameras. In most cases interacting people face each other, so we cannot put a single camera to capture both faces. Naturally the conclusion is that we need many views."

The sky's the limit with Panoptic Studio.

Graduate student Gines Hidalgo said, "Thanks to open pose, we can process all these sequences from the dome, and we can get all the people's location in the same for each view. And at the same time, thanks to all the information the dome has given us, we are able to train and better prepare open pose and thanks, for example, to the dome data we were able to create the first hands detector that actually works with normal RGB (red, green, blue) cameras. And that was thanks to all the data set, the huge data set that we have generated with the dome data set."

Being able to detect the nuances of non-verbal human communication could hasten the use of robots in social spaces, allowing them to perceive what people around them are doing.

Self-driving cars could potentially receive early warnings that a pedestrian was about to step into the street by the system monitoring their body language.

And in sports analytics, real-time pose detection might enable computers to track what individual players are doing with their arms, legs and heads at each point in the game.