Abstract
To identify the sources of information used to group across occlusions, we propose a contour grouping task based directly on natural images. Prior to measuring performance, edge elements are extracted automatically from natural images and observers assign each edge element to a physical contour using the full image context [Vis. Res., 41, 711–724], providing an approximate “ground truth” of which edge elements derive from the same contour. On each trial of the experiment, a pair of edge elements is randomly selected from an image. In the stimulus display, an occluder is placed so its diameter just touches the center point of the two elements. The subject's task is to decide whether these two edge elements belong to the same or different contours. In the restricted case, only the two edge elements and occluder are displayed. In the unrestricted case, the full gray-scale image plus occluder is displayed. Results: (a) For the restricted case, naïve observers approach ideal performance as determined by the pair-wise statistics of edge elements in natural images. (b) For the unrestricted case, there is little improvement over the restricted case for small occluder diameters and modest improvement for large occluder diameters (here ideal performance is unknown). (c) Observers do not improve after practice with feedback. (d) Most of the information that humans use in performing contour grouping across occlusions in natural images is captured by the pair-wise statistics of the edge elements at the point where the contour becomes occluded.
Supported by NIH grant EY11247.