Purchase this article with an account.
Deep Ganguli, Jeremy Freeman, Umesh Rajashekar, Eero Simoncelli; Orientation statistics at fixation. Journal of Vision 2010;10(7):533. doi: 10.1167/10.7.533.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Eye movements are not random. When viewing images, human observers tend to fixate on regions that, on average, have higher contrast than randomly selected regions (Reinagel & Zador, 1999). We extend this analysis to the study of local orientation statistics via the “orientation tensor” (Granlund & Knutsson, 1994), computed as the 2x2 covariance matrix of local horizontal and vertical derivatives (i.e., the gradient vector) within an image patch. This may be converted into three natural parameters: energy, orientedness, and orientation. Energy is the total variance in the gradients, and is related to contrast; orientedness indicates the strength of the dominant orientation; orientation indicates the predominant orientation. We use an eye movement database (van der Linde et al., 2009) to measure the orientation tensor within local 1 deg image patches that are either fixated by human observers (n=29), or selected at random (by using fixations for a different, randomly chosen image). We then obtain image-specific log distributions of the three parameters of the orientation tensor. Averaged across all images and subjects, energy is higher in fixated patches, consistent with similar reports on contrast, but we do not observe such differences for orientation or orientedness. However, when we compare fixated and random distributions of these parameters on an image-by-image basis, we observe systematic differences. In particular, for the majority of images, the distribution of fixated patches, when compared to that of random patches from that image, is closer to the generic distribution averaged over all images. We use multi-variate techniques to characterize this effect across the database. We find that fixated distributions shift towards the generic distribution by about 10 to 20%, and the trend is significant for all three parameters. Our results suggest that when viewing a particular image, observers fixations are biased towards locations that reflect the typical orientation statistics of natural scenes.
This PDF is available to Subscribers Only