Abstract
Purpose:
Classification by cause of local contrast elements is a critical part of object recognition and other image interpretation processes. For example, occlusion boundaries, albedo discontinuities, specularity edges, shadow edges, etc. in the scene, are all possible causes. It is not known whether this classification takes place prior to object recognition (bottom up process) or if it takes place after object recognition (top down process). In order to better understand the role of contrast cause estimation, we studied estimation performance given local information only.
Methods:
Natural images were collected by calibrated camera and converted to luminance images. By changing the background, the experimenter determined the true object contour. By manipulation of illumination, the shadow and specularity edges shadow edges were distinguished from the albedo edges.
Subjects then viewed local regions of the object images thru an aperture and were asked to estimate the cause of contrast in a four alternative forced choice procedure (occlusion, albedo, specularity, shadow). Five aperture sizes, from 21.6 to 124 moa, were used.
Subject responses were collected and compared to ground truth.
Results:
A contingency table analysis showed that subjects were above chance in correctly estimating the true cause of contrast in an image from local information. Performance was at 47.5% correct for small apertures and 48.75% correct for large apertures. Chi-Square for small apertures and for large apertures were both significant at p[[lt]].001. The strength of the observers performance was determined by Cramer's V, which was .395 for the small apertures and .362 for the large apertures. Both indicate a strong association between predicted cause and ground truth.
Conclusions:
The results show that observers perform above chance with very little global information. Also surprisingly, aperture size and global information conferred little advantage over the large range tested.
This research supported by NIH Grant Number P20-RR-020-151