Local processes and spatial pooling in texture and symmetry detection

Abstract

We examined the ability of human observers to detect three kinds of statistical structure in binary arrays: first-order statistics (luminance), local fourth-order statistics (isodipole textures), and long-range statistics (bilateral symmetry). Performance was closest to ideal on the luminance task and furthest from ideal on the symmetry task. For each kind of statistic, the dependence of performance on the degree of structure was well described by a model consisting of an initial stage of multiple independent detectors, followed by a pooling stage. For the luminance task and the isodipole task, performance was well-modeled by local processing followed by extensive spatial pooling. For the symmetry task, limitations at the local detection stage and a near-absence of spatial pooling were needed to model for performance.