Abstract
Because of crowding among adjacent contours, object identification in the peripheral visual field is worse than predicted by acuity alone. Natural visual environments are full of contours and must therefore be profoundly crowded. We examined sensitivity to peripheral spatial structure in natural scenes by asking observers to detect the presence of patches of “dead leaves”: ellipses of random size, aspect ratio and orientation superimposed onto greyscale natural images to produce naturalistic edge structure with the same average luminance and contrast as the image patch they replaced. Three observers identified the location of the dead leaves patch relative to fixation (N, S, E or W) at three eccentricities (2, 4 and 8°), with the size of the patch under the control of an adaptive staircase. Size thresholds increased with eccentricity, consistent with the eccentricity-dependence of crowding. A reverse correlation analysis was used to determine which properties of the underlying image were correlated with patch detection. For each of approximately 10000 image trials per observer, the local luminance, rms contrast, edge density, orientation, orientation variance and amplitude spectrum slope were computed at 8 spatial scales (Gaussian σ from 0.3 to 4 degrees). Differences of image statistics between spatial scales allowed the comparison of different center/surround combinations. These image statistics were then used as predictors in a logistic regression analysis of trial-to-trial performance. At eccentricities of 2 and 4 degrees, observers' performance was best predicted by the difference between the smallest center and the largest surround, whereas at 8 degrees the best-fitting models compared coarse spatial scales. Regression coefficients highlight the importance of contrast, edge density, orientation variance and amplitude spectrum slope in predicting performance, with luminance and orientation contributing little to the model. These models allow prediction of when crowding may occur in a given natural image.
This project was funded by NIH grants EY019281 and EY018664.