
February 23, 2011
Training Machines to Recognize People
We've established that a computer can beat the best humans at Jeopardy. But can it tell Alex Trebek from Chuck Woolery?
On Feb. 21, Deva Ramanan, Assistant Professor of Computer Science at UC Irvine, was at Duke talking about his work on enhancing computer vision to the point where machines might be able to recognize individual people.
He said it has been extremely difficult for a computer to recognize humans in different images or video sequences because of significant variations in every frame of reference. “Finding people in images is difficult [because of] variation in illumination, appearance, pose, viewpoint and background clutter. These are the classic nuisance factors for object recognition,” Ramanan said.
Ramanan and his team have worked on adaptive and dynamic algorithms to solve this problem by breaking down the entire image template into local, global and temporal models.
They also extend the template into semantic parts, which implies that the computer program is already told a priori the parts of a body that it needs to learn. The computer then has to learn the different appearances and the template based on the training data set.
“One can also apply suppression techniques—for example, two objects can never occupy the same 3D volume, people don’t stand on top of each other but stand next to each other, and bottles are often supported by tables.” Using a variety of models, the team has been able to build a relatively robust model of human-detection in images and movies.
Applications of such a people-detection system include video surveillance, autonomous vehicle navigation, healthcare, image and movie search and building smarter visual interfaces.
Ramanan is the recipient of the 2009 David Marr Prize, an NSF Career Award and the 2010 PASCAL Visual Object Class Challenge Lifetime Achievement Award.
Posted by vm19. 0 comments




