Massachusetts Institute of Technology (MIT) professor Antonio Torralba and students from the school's Computer Science and Artificial Intelligence Lab (CSAIL) say they have developed an object recognition system that requires no training and can identify objects at least as well as and any other available program.
The CSAIL system uses a modified version of a motion estimation algorithm. Consecutive frames of video normally change very little, so data compression schemes often store the unchanging elements of a scene once and only update the positions of moving objects, with motion estimation algorithms determining what objects move.
The MIT system treats unrelated images as if they were consecutive frames in a video sequence. When the modified motion estimation algorithm attempts to determine what objects have "moved" between one image and the next, the system can identify objects of the same type.
"It's a real commonsense solution to a fundamental problem in computer vision," says University of Central Florida professor Marshall Tappen. "The results are great and better than you can get with much more complicated methods."
From MIT News
View Full Article
Abstracts Copyright © 2009 Information Inc., Bethesda, Maryland, USA
No entries found