uni'wissen 02-2012 ENG

by Nicolas Scherger of humans: Our appearance changes depending on what clothes we’re wearing or what perspec- tive one views us from, and when several people are in the same picture the diversity of appear- ances becomes completely confusing. In order to recognize humans with any degree of reliabil- ity, the computer must be able to distinguish be- tween these variations. “It can’t memorize all of the possibilities, because this would involve too many training examples,” says Brox. Instead, the computer has to learn how to abstract from par- ticulars: “It has to find features that are common to all humans,” meaning typical structures and forms, such as the head and shoulders. The idea of teaching computers to recognize features that are characteristic of particular classes of objects – humans, cars, dogs, potted plants – is by no means new. Up until now, how- ever, scientists have trained computers by mark- ing objects on individual images with so-called annotations, for instance a frame and the remark that there is a dog inside it. The computer recog- nizes the forms and structures that describe the dog on the basis of the edges inside the frame, at which the color in the picture changes. This method has disadvantages, says Brox: “The manual effort for the annotations is high, and the computer learns only what it has been provided “On videos the computer has an easier time segmenting objects or separating them from the background” The Computer Scientist Thomas Brox Is Teaching Computers to Find Objects on Pictures with the Help of Videos

uni'wissen 02-2012 ENG

Pages