Fig 1.
A) Face stimulus creation. The face image was iteratively phase-scrambled, re-combined with non-scrambled face pixels, and scrambled again (number of iterations was 500). B) Matching of face and scene images. The relative amplitude at vertical, 45° oblique, horizontal, and 135° oblique angles was compared across the face and scene categories. Through method of least squares the scene image with the best-matching amplitude profile was selected to be included in the stimulus set. Left panel: example of matched face and scene stimulus. Right panel: the amplitude profile plots (in arbitrary units) for these two images. The red line represents the face image, whereas the blue line represents the scene image.
Fig 2.
Stimuli of each of the five different stimulus categories were first made isotropic, i.e. all orientations carry equal amplitude. An orientation increment was created by increasing the relative amplitude in a 45°-wide band centered on one of four possible orientations: vertical, 45° oblique, horizontal, and 135° oblique. For illustration purposes, the intensity of the increments in the figure is magnified. Participants were to indicate on each trial whether the presented stimulus was isotropic or contained an orientation increment.
Fig 3.
A) Contrast detection task. Each trial started with a 500 ms fixation period. A target stimulus (here: upright face) was presented for 400 ms followed by a 500 ms random noise mask. The participants indicated by button press whether they perceived an increment of a particular orientation in the target image (here: present). B) High-level face identification task. Each trial started with a 500 ms fixation period. An initial face stimulus, filtered to contain information in a limited orientation band (here: centered on 135° oblique) was presented for 300 ms. After a brief noise mask was shown for 200 ms, a second face stimulus was presented for 300 ms. The participants indicated by button press whether the identity of the two faces was identical or different (here: different). The location of the second stimulus was varied slightly relative to the first one, in order to reduce the influence of local low-level visual properties.
Fig 4.
Performance difference matrix calculation and analyses.
The left matrix indicates how cell values are computed. Equations in black represent values in cells that are included in subsequent correlation analyses. In these equations, the difference of performance between orientations is computed. Equations in grey represent values that were not included in subsequent analyses because they either replicate values in one the other half of the matrix or because they by definition will have a value of zero (diagonal). Right panel. Individual difference matrix of a given participant in a particular condition was either correlated with their matrix for another condition, or with matrices constructed based on a priori models of orientation selectivity. For each a priori model, the line plot insets illustrate the predicted performance as a function of orientation.
Fig 5.
A) Accuracy scores for the contrast detection task with the face conditions (i.e. upright faces, inverted faces). B) Accuracy scores of face identification task across filter orientations. The black lines represents the data for upright face stimuli, the gray lines represents the data for inverted face stimuli. Error bars indicate within-subject-corrected 95% confidence intervals [50,54].
Fig 6.
Comparing individual difference vectors to a priori models.
Box plots of the Spearman rank correlations between individual difference vectors and a priori models split for different levels of task and stimulus category. The left column shows the model correlations for upright faces, whereas the right column shows the model correlations for inverted faces. The upper row shows the model correlations for the contrast detection task, whereas the lower row shows the model correlations for the face identification task. Model 1 represents the cardinal effect, Model 2 the horizontal effect, and Model 3 the idea that horizontal orientations lead to better performance from the other orientations (‘horizontal is special’). Horizontal lines within the boxes represents the median values, Xs represent the condition mean, and circles are the individual data points. Higher and lower edges of the boxes represent the borders of the third and first quartile, respectively.
Table 1.
Bayesian factors for the a priori model correlations.
BF10 values represent the strength of the evidence in favor of a positive model correlation over the alternative of no correlation.
Fig 7.
Comparing individual difference vectors across experimental conditions.
A) Spearman rank correlation of individuals’ difference vectors for upright and inverted faces. The left bar depicts the correlation for the contrast detection task. The right bar depicts the correlation for the high-level identification task. B) Spearman rank correlation of individuals’ difference vectors for contrast detection and face identification tasks. The left bar depicts the correlation for the upright face condition. The right bar depicts the correlation for the inverted face condition. Black horizontal lines within the boxes represents the median values, Xs represents the condition mean. Higher and lower edges of the boxes represent the borders of the third and first quartile, respectively.
Table 2.
Bayesian factors for the difference matrix correlations across tasks and across stimulus categories.
BF10 values represent the strength of the evidence in favor of a correlation between the tested patterns of orientation selectivity over the alternative of no correlation.
Table 3.
Noise ceilings for the a priori model correlations.
Noise ceilings indicate the maximal correlation achievable by a given model with the experimental data, given the data variability across participants (see Analyses section for details on calculation).