Abstract
The efficient coding hypothesis posits that visual systems are adapted to the regularities they are exposed to. Many studies have measured the statistics of natural images and revealed biases in the distribution of orientations, such as a predominance of cardinal compared to oblique orientations as well as an increase of radial versus tangential orientations with increasing eccentricity. These phenomena have their correspondence both in psychophysics with the oblique and meridional effect, and in the overrepresentations of corresponding orientations in V1. However, the commonly used natural image databases have a rather small field of view. To really quantify the statistics of the natural input to the visual system, a large field of view as well as the physical properties of the visual system should be taken into account.
Here, images with a field of view of 120° were generated during exploration of a virtual forest environment from both human and cat perspective. Images were projected onto idealized retinas according to models of the eyes’ geometrical optics. Image statistics across the visual field were examined using power spectra and sparse coding.
For small eccentricities, the measured statistics matched those of photographic images. Confirming previous research, there was a bias towards cardinal orientations in both second-order and higher-order statistics. For larger eccentricities, we found an increasing bias towards radial orientations. In the images taken from the cat’s viewpoint, differences between lower and upper visual field were more pronounced compared to the human viewpoint. This lays the groundwork for quantitatively relating natural image statistics across the visual field with neural representations of orientation and psychophysical behavior in orientation discrimination and confirms that the input to the visual system is influenced by the structure of the environment, but also the physical properties of the visual system and the observer’s viewpoint.