uni'wissen 02-2012 ENG

with.” Instead, Brox wants to teach computers how to acquire new knowledge on their own in order to automate the learning process and re- duce the number of necessary annotations. Finding Points that Move in Similar Way In order to achieve this, he does not use static images but films. “On videos the computer has an easier time segmenting objects or separating them from the background,” says Brox. He devel- oped a method of analysis that has the computer find all of the points in the video that move in similar ways – for instance a rider on a horse. “That is a clear indication that these points differ from the rest of the pictures.” Humans, animals, and vehicles move on their own. In the case of stationary objects, such as a piece of furniture, Brox creates the illusion of movement by moving the video camera around the object, enabling a three-dimensional reconstruction. The computer then compresses the segmentation: “It also as- signs all points that it could not locate previously to the object or the background with the help of color information.” In this way, the computer can separate the object from the background without needing to be shown a frame telling it where something relevant is located. All Brox has to do is let the computer know what class the object shown on the video belongs to. In addition, a video is composed of two dozen images per second. Since the object, for exam- ple a car, looks slightly different on each image due to the movement, the computer already learns a large number of variations just from watching the film. “Furthermore, it’s also clear that the reason for this variety is not that the cars in question are different. That is important infor- mation for the learning,” says Brox. The comput- er scientist has solved the first part of the learning process: The computer recognizes vari- ations of, for instance, the same cars on the ba- The computer separates the rider on a video (top) from the background by assigning the points that move in a similar way to the same object (middle). It adds the remaining points with the help of color information. In this way, every point on the image is assigned either to the rider or the background (bottom, overlaid on original image) 14

uni'wissen 02-2012 ENG

Pages