Communications of the ACM
Context Is Ev . . . Well, Something, Anyway
Massachusetts Institute of Technology researchers have developed a way to improve object recognition systems by using information about their context. The researchers used a set of more than 4,000 images and 107 different types of objects and created an algorithm that sorted the images into a hierarchical map of the object categories. In the map, each object is connected to at most one object above it in the hierarchy, which greatly reduces the number of connections the system must consider.
The connection between two objects is given a weight that describes how often the objects appear together in the image set. When the system analyzes a new image, it uses object recognition algorithms to create a list of candidate objects and a confidence score.
In testing, the system had a success rate of about 25 percent. "We absolutely cannot afford to take our eye off the ball of the component recognition systems that need to feed these context engines," says University of California, San Diego professor Serge Belongie. To be useful, object recognition systems need to be much more precise than today's prototypes, Belongie says.
From MIT News
View Full Article
Abstracts Copyright © 2010 Information Inc., Bethesda, Maryland, USA 
No entries found