Thesis Advisor Author Detecting Junctions in Photographs of Objects

Line drawings are inherently simple, yet retain most of the information of a full image. This suggest that line drawings, also called contour images, could be useful as an intermediate representation for computer understanding of images. However, extracting physically meaningful contours from photographs has proved to be difficult. Recently, there has been significant progress in detecting contours. As one example, the gPb algorithm (Arbeláez et al., 2010) uses local cues such as brightness, color, and texture, along with global cues that interpret the image as a whole in order to achieve high-performance contour detection. However, contour detectors perform poorly near junctions. Junctions are points where two or more contours meet, creating two or more regions near a single point. Contour detectors, including gPb, assume that contours occur at the meeting of only two regions. We find that this assumption breaks down near junctions, resulting in poor performance. This is unfortunate because the location of junctions and the orientation of the incident edges at a junction are useful in reasoning about the geometric relationships between contours, and thus reasoning about the shape of the object. These two observations, that junctions carry significant information and that con-