Fig. 4.

Pairwise decoding performance as a function of orientation difference (all pairs from eight orientations), for (I) grating images (pixel intensities), (II) fMRI images (voxel intensities), and (III) transformed grating images. The gratings (I) were 20×20 pixel black/white images with 2–3 stripes. The fMRI images (II) were those obtained in the present study (responses to gratings of eight orientations; 400 voxels from V1/V2). The transformed images (III) were created by linear orientation filtering (Gabor-like filters for four orientations) of the grating images (I) followed by thresholding (non-linearity) and addition of noise. The orientations of these images were decoded for each pair of orientations (chance level, 50%). For (I) and (III), the average performance with five sets of phase randomized images is plotted (error bar, standard deviation). For (II), the average performance of four subjects is shown. The grating images (I) resulted in poor performance regardless of orientation difference. In contrast, the fMRI images (II) and the transformed grating images (III) both showed performance that improved with orientation difference, reaching near perfect levels at 90°.