Abstract
Observers are adept at estimating texture statistics such as mean element-orientation, a process that can be modeled using population coding of responses from orientation-selective neurons in V1. Here we consider how observers average the size of objects, given that (a) the neural substrate for object-size is less clear, and (b) limitations of previous paradigms used to explore size-averaging have sparked debate as to whether observers can average size at all. We used a noise paradigm: observers reported which of two sets of 16 Gabor elements had the greater mean element-size in the presence of different levels of element-size variability. We randomized the spacing of the Gabors (thus minimizing any cue from element-“coverage”) and both the contrast and orientation of elements (minimizing any cue from global statistics). In the first condition (scale averaging) the envelope-size and carrier spatial frequency (SF) of elements co-varied, so that all elements were scaled/rotated versions of one another. Under these conditions observers averaged ∼50% of the elements, effectively estimating the scale of each with a precision (σ) of ∼25%. This unequivocally indicates that observers can average element-scale. Fixing carrier spatial frequency (SF) forces observers to use envelopes (size averaging) and produced near-identical performance. Fixing envelope size forces subjects to use carrier-SF (SF-averaging) and produced moderately poorer performance. Thus observers must, at least in part, be using envelope size when scale-averaging. Critically, adding independent noise to the SF and the envelopes of elements substantially increases the number of elements that are averaged, indicating that observers can exploit independent statistical properties of both the envelope-size and SF of elements to make perceptual discriminations. We consider it likely that cues from feature (e.g. edge) density drive both these and a range of related tasks (e.g. judgment of number and density).