Abstract
The assessment of three dimensional shape perception from monocular images continues to be an open problem in the vision science community. What features and processes mediate shape perception in natural images? Previous studies have investigated the effects of shading, texture and contour in artificial stimuli. In this experiment, however, we choose to investigate how these cues interact in the context of natural images such as a vase, a human face, or a landscape scene.
We degraded a set of natural images to measure the influence of shading, texture and contour on shape perception. The first degradation is a gaussian blur kernel applied to the stimulus set. To create the second degradation, human subjects manually segment the images, and the distinct regions in the human segmentations are replaced with the region's average color and luminance. The first degradation intends to reduce the influence of shading and texture, while the second intends to completely eliminate any influence. Nine subjects were tested in this experiment, where each subject was presented each image in one of three different conditions: normal, degradation 1, degradation 2. Subjects made surface normal settings at a large number of sample points on each image using the elliptical gauge figure paradigm introduced by Koenderink et al. (1992). The subjects' task is to adjust the elliptical disks so that they appear maximally aligned with the local surface shape.
Subjects' responses display a surprising robustness to shading and texture degradations in natural images. Although images subjectively look flatter under the degradations, subjects are able to make surface normal settings by ignoring the fact that shading and texture indicate flatness. This is an example of cue integration where the false cue can be ignored, and these results indicate the primacy of the contour cue in shape perception of familiar object categories.