Abstract
In research presented at last year's VSS, it was demonstrated that when humans label image contours according to the type or physical cause of the contour, they are heavily influenced by the available vertex information. The present study was designed to refine our understanding of how vertices influence observer labeling; different vertices may be more or less informative in narrowing down possible labels. In addition, observers may be using prior experience to reduce possible interpretations to those that are most likely in the environment. A key test of this hypothesis is to gradually reveal image information: if observers are biased towards particular interpretations, these labels will be assigned to ambiguous stimuli that may then be made unambiguous by increasing available vertex information. The study therefore used computer generated images of simple scenes, beginning with a small, circular cutout centered on a particular contour and gradually increasing in size to reveal additional vertices one at a time. The range of sizes was designed to span completely ambiguous to completely unambiguous scenes in no more than 5 steps. Observers were asked to use mouse controlled sliders to indicate their confidence in assigning five different labels (reflectance, illumination, orientation, orientation occlusion, and smooth occlusion) to the single central contour. The results suggest that different vertices are in fact different in how informative they are about constituent edges, with some vertices being sufficient alone and others requiring multiple additional vertices to produce correct labeling. Observers also assigned high confidence incorrect labels to contours when stimuli were moderately ambiguous, later switching to high confidence correct answers when additional vertices were added. This suggests that observer bias does play a significant role in the use of vertex information.
This research was supported by a grant from NSF (BCS-0546107).