Sergei Gepshtein: Low-level Vision

Low-level Vision

My latest work in low-level vision concerns the spatial resolution of stereopsis, with Martin Banks of UC Berkeley and Michael Landy of NYU. We asked why the finest detectable modulation of stereoscopic depth (spatial stereoresolution) is much lower than the finest detectable variation in luminance. Using a combination of psychophysics and theoretical analysis of the computations underlying stereopsis, we proposed a simple answer.

First we examined how the following factors constrain stereoresolution:

the sampling properties of the stimulus,
the disparity gradient limit [1],
low-pass spatial filtering by mechanisms early in the visual process.

We found contributions of each of these factors. Yet, we discovered that the highest attainable stereoresolution is not determined by any of the above.

Then we examined the method by which the visual system computes binocular matches between the two retinae [2]. A recent study of disparity-selective cells in area V1 of the primate [3] indicated that these cells solve the matching problem by detecting small patches of constant binocular disparity. We simulated binocular matching by cross-correlation of two retinal images and implemented the local constant-disparity constraint by having the model seek identical patterns of luminance on the two retinae. Our results suggest that the highest stereoresolution is determined by the size of correlation window that captures enough luminance variability from each retina, limited by the low-level spatial filtering of optical information. When a correlation window is TOO SMALL it does not capture enough luminance variability to identify correct matches.

A full-length paper describing our results has been just published; it is cited below. I am preparing a web site illustrating our model. If you want to receive a note when the site is up, please speak up.

Banks, M. S., Gepshtein, S. & Landy, M. S. (2004). Why is spatial stereoresolution so low? Journal of Neuroscience 24 (9) p. 2077-2089. [PDF] [On-line Supplement]
Gepshtein, S., Banks, M. S. & Landy, M. S. (2003). Spatial Resolution of Stereopsis. Journal of Vision 3(9), 466a. [poster (2MB)]

[1] Disparity gradient limit, discovered by Christopher Tyler in 1973, is a limit on the highest perceptible rate of change of depth from binocular disparity.

[2] Early in stereoscopic processing the visual system must identify those regions of the two retinal images that correspond to the same source in the world. In doing so the system solves the binocular correspondence problem (also called binocular matching problem).

[3] This study is just published, in the same issue with Banks, Gepshtein, & Landy (2004):

Nienborg, H., Bridge, H., Parker, A. J., and Cumming, B. G. (2004). Receptive field size in V1 neurons may limit acuity for perceiving disparity modulation. Journal of Neuroscience. 24 (9), p. 2065-2076.

Link credits: The page on stereopsis is by Dr. James B. Calvert of the University of Colorado, USA. The page on cross-correlation is by Dr. Paul Bourke of the Swinburne University of Technology, Australia.

low-level vision