Figure 1. Stimuli, model, and methods. A, Examples of images used in the experiment. Images varied considerably in the degree to which they contained exclusively man-made (left) or natural (right) elements, and the set included images for which the distinction may be unclear (middle). B, A feedforward filtering model was used to derive two summary statistics: CE and SC. Three opponent (grayscale, blue-yellow, red-blue) contrast magnitude maps were computed by convolving the image with multiscale filters (black circles). For CE, a range of smaller filter sizes (σ = diameters in degrees) was used; for SC, a range of larger filter sizes. For each image location and parameter, a single filter response was selected (red circles) from each range using minimum reliable scale selection (see Materials and Methods). These responses were then pooled across a selection of the visual field (red dotted circles): for CE, the resulting responses were averaged; for SC, the coefficient of variation (COV) was computed. These values were averaged across the three color-opponent maps resulting in one CE and SC value per image. C, A subset of 160 images (10% of the whole stimulus set, randomly selected) plotted against their CE and SC values. CE (the approximation of the β parameter of the Weibull function) describes the scale of the contrast distribution: it varies with the distribution of local contrasts strengths. SC (the approximation of the γ parameter of the Weibull function) describes the shape of the contrast distribution: it varies with the amount of scene fragmentation (scene clutter). Four representative pictures are shown in each corner of the parameter space. Images that are highly structured (e.g., a street corner) are found on the left, whereas highly cluttered images (e.g., a forest) are on the right. Images with higher figure-ground separation (depth) are located on the top, whereas flat images are found at the bottom. D, sERPs to images presented at the center of the screen were computed for each subject. The resulting estimates of sERP amplitude were regressed on CE and SC at each time sample and electrode separately. The design matrix for the regression contained five columns: a constant term (c) for the intercept, two columns for CE and two for SC, each containing the same parameter values for the first and second image presentation (p1 and p2). These were modeled as separate predictors to examine reliability of the obtained effects across repetitions. The outcome of the analysis is a measure of model fit (explained variance or R2) separately over subjects, time (samples), and space (electrodes); an example is shown for one subject at electrode Oz.