Deva Ramanan trains computers to identify people and even to understand what those people are doing.

Deva Ramanan clicks a button on his MacBook Air and a video begins to play: Michelle Kwan skating in the 1998 Nagano Olympics. Next to it, a computer program renders what it “sees” in the footage: Kwan’s head, legs, torso, upper arms, and forearms, all distinguished by different colors. Ramanan, a computer scientist at University of California at Irvine, trains computers to recognize three-dimensional humans in flat photography.

Face-recognition software, which pinpoints the classic eyes-nose-mouth configuration, has been in use for years. But detecting a human body—any human body—is much more challenging for computers due to the endless variety of possible poses, angles, sizes, and outfits. Most researchers will feed a program millions of images to memorize, building a vast database of people. Ramanan, instead, trained his computer program to identify body parts and match them to a flexible human template. “You can think of it as a divide-and-conquer approach,” he says. The software runs through a checklist: Arms, torso, legs? Check. Thus, a human. Ramanan’s method is much faster and uses less processing power than the traditional one.

Ramanan envisions many potential applications for his people-finding algorithms, including fast and accurate pedestrian-detection systems in self-driving cars and videogame systems that track full-body movements. In the meantime, he’s focusing on teaching computers how to read and understand context—in other words, to think. “What if you really want to understand what a person is doing,” Ramanan asks. “Not just [understand] ‘Here’s the arm,’ but ‘This person is waiting for a bus.’ ” If his future projects succeed, computers’ reasoning ability will keep inching closer to that of the human brain itself.

Written by Elizabeth Svoboda from Pop Sci. More Info