Abstract
Binocular disparity is a primary cue for depth perception. However, not all visible regions of natural scenes create binocularly paired retinal projections. Monocularly visible (half-occluded) regions occur at depth edges where foreground surfaces occlude backgrounds. Previous research has shown that, with simple stimuli, half-occluded targets are perceived as near as possible, given the constraints imposed by stereo-geometry. Here, we asked how depth is perceived in half-occluded regions of natural scenes. We sampled hundreds of stereoscopic image patches (3×1°) from a natural image database with co-registered distance measurements. Each patch had an extended half-occluded region that was ~10–30 arcmin wide; the set of patches contained a range of different depths. On a haploscope, human observers viewed patches through a stereoscopic window and estimated the depth of the depicted half-occluded surfaces; a monocular image probe indicated the target surface location. Observers reported depth estimates by adjusting the disparity of a binocular response probe in a neutral field, viewed through an adjacent window; the task was easy and responses were repeatable. In condition one, all image cues were available. Humans reported depth estimates that matched the actual surface depths. In the minority of patches that produced inaccurate estimates, we measured the perceived shape of half-occluded surfaces by systematically changing the location of the image probe. In condition two, the image probe stereo-geometry was identical, but all image cues were removed. Now, human observers underestimated the depth, as predicted by classic findings. Thus, natural images contain cues that increase the accuracy of depth estimates near depth edges. Finally, we manipulated image cues (e.g. contrast) in different regions (e.g. the binocular background), to determine which cues and regions are critical for accurate depth perception near depth edges. Together, our findings show how stereo-geometry and image cues determine depth percepts in half-occluded regions of natural scenes.
Acknowledgement: R01-EY028571, 5T32EY007035-40