Abstract
Junctions, formed at the intersection of image contours, are thought to play an important and early role in vision. The interest in junctions can be attributed in part to the notion that they are local image features that are easy to detect but that nonetheless provide valuable information about important events in the world, such as occlusion and transparency. These assumptions generally hold in the synthetic stimuli often used in the lab. Here we test the notion that there are locally defined junctions in real images that might be detected with simple, early visual mechanisms. Our approach was to use human observers as a tool to measure the visual information available in local image regions.
We had one set of observers label all the points in a set of real images where one edge occluded another. A second set of observers was presented with variable-size circular subregions of these images, and was asked to judge whether the regions were centered on an occlusion point. Control experiments with synthetic stimuli confirmed that if locally defined junctions are present at the occlusion points, this task is easy given a small region surrounding the point. With real image patches the results were markedly different. Performance was poor for small regions and did not approach ceiling levels until observers were given fairly large (∼50 pixels in diameter) regions over which to make the judgement. Performance also tended to be poor if these large regions were blurred and subsampled, ruling out the possibility that the effects are just due to junctions at coarser scales.
Our experiments suggest that although some junctions in real images are locally defined and can be detected with simple mechanisms, a substantial fraction necessitate the use of more complex and global processes. This raises the possibility that junctions may not play the bottom-up role in vision which they have often been ascribed.