Abstract
Performance in any natural task is limited by at least two sources of uncertainty: natural stimulus variability and internal noise. Processing may also be limited by fixed sub-optimal computations. Here, using a double-pass experiment, we determine the absolute and relative importance of these three performance-limiting factors in a speed discrimination task with natural image movies. First, we develop an image-computable ideal observer that is constrained by the front-end properties of the human visual system. Second, we show that the ideal observer predicts the pattern of human discrimination thresholds. Third, we show that the absolute performance levels can be accounted for by degrading the ideal with a single free parameter (efficiency). Fourth, we show that human efficiency predicts human response repeatability in a double-pass experiment without additional free parameters. Theory and simulation indicate this last result should hold only if internal noise (and not sub-optimal computation) is the sole source of human inefficiency. In sum, this set of results suggests that humans use the optimal computations to estimate and discriminate speed with natural stimuli, that humans underperform the ideal only because of internal noise, and that natural stimulus variability and internal noise are equally important determinants of human performance limits. The results demonstrate the value of task-specific analyses of natural scene statistics, and encourage similar analyses in other domains.