Purchase this article with an account.
Arvind Iyer, Johannes Burge; Predicting effects of natural depth variation on binocular disparity estimation. Journal of Vision 2017;17(10):1069. doi: 10.1167/17.10.1069.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Reliable estimation of binocular disparity is fundamental to our ability to estimate the relative depth of objects in natural scenes. Classical laboratory stimuli are not representative of the stimuli that disparity-processing mechanisms encounter in natural scenes. Here, we examine the impact of natural disparity variation (arising from natural depth variation) on disparity detection performance. First, we obtained a database of calibrated natural stereo-images with precisely co-registered laser measurements of groundtruth distance at each pixel. Second, we developed a procedure that uses the groundtruth distance data and has arcsec precision, to sample 50000 stereo-patches centered on binocular corresponding points. For each stereo-patch, we computed a ground-truth map of local absolute disparities in the foveal region. The disparity contrast (root-mean square contrast of the disparity map) quantifies local disparity variation around fixation. High disparity contrast hampers precise estimation of fixational disparity. An ideal observer for disparity estimation in natural scenes, and a local windowed cross-correlation routine both yield disparity estimates where reliability decreases with increases in disparity contrast. Thus, disparity contrast predicts the reliability of disparity estimates. Information about estimate reliability can aid subsequent visual processing and perceptual decisions. However, just like with fixational disparity, the visual system has no direct access to groundtruth disparity contrast, and must instead estimate it from the binocular luminance images. Interestingly, two simple statistics, (i) the binocular luminance difference, and (ii) the root-mean-square contrast of the binocular difference image (left-eye minus right-eye image) are both predictors of disparity contrast. The difference image contrast is a particularly strong predictor. The natural scene statistics indicate that the contributions of these two predictors to estimates of disparity contrast are multiplicative and separable. These findings suggest multiple computational and psychophysical investigations of mechanisms for estimating local disparity variation, and assessing its impact on human disparity estimation performance in natural scenes.
Meeting abstract presented at VSS 2017
This PDF is available to Subscribers Only