Context is ev … well, something, anyway
Today, computers can’t reliably identify the objects in digital images. But if they could, they could comb through hours of video for the two or three minutes that a viewer might be interested in, or perform web searches where the search term was an image, not a sequence of words. And of course, object recognition is a prerequisite for the kind of home assistance robot that could execute an order like “Bring me the stapler.” Now, MIT researchers have found a way to improve object recognition systems by using information about context. If the MIT system thinks it’s identified a chair, for instance, it becomes more confident that the rectangular thing nearby is a table.A typical object recognition system will scan a digital image for groups of pixels that differ from those around them; those pixels could define an edge, a corner or some other feature of an object. Usually,...