Abstract
The shape of the human spatial contrast sensitivity function (CSF) above 2–4 cpd is similar to that computed using ideal observers that incorporate models of the optics and cone photon absorption [1]. These ideal observers, however, are 20–30 times more sensitive than humans. It is thus likely that pre-neural factors determine the shape of the high-frequency limb of the CSF, but that other factors produce an overall sensitivity attenuation. We use the ISETBio framework [2] to derive computational observer CSFs that incorporate current models of optics, photon absorption, fixational eye movements and outer-segment photocurrent generation. Moreover, we examine the effect of assessing performance using support vector machines (SVM), without explicit knowledge of response dynamics. We first demonstrate that ISETBio accurately replicates previously reported [1] ideal observer CSFs. Next, we show that incorporating current models of human optics and photon absorption lead to only modest changes. Modeling stimulus detection using SVM classifiers reduces performance by a factor of ~4–5 relative to the signal-known-exactly ideal observer. Inclusion of fixational eye movements results in no significant performance attenuation when responses are pooled using quadrature-pair energy mechanisms. Finally, performance at the photocurrent stage is reduced overall by an additional factor of ~2. We conclude that the optics and cone photon absorption determine the shape of the spatial CSF above 2–4 cpd, and photocurrent generation and decision mechanisms bridge much of the remaining gap between ideal and human observer performance.