Abstract
Purpose: It has been proven that many visual processing characteristics are related to the statistics of natural inputs to the visual systems. But it is not clear if and how such statistics will change due to different eye movement patterns by different viewers for different scenes, for different illuminations and for scenes in color or in black and white.
Methods: We record gaze positions during free-viewing video segments of natural activities in four viewing conditions: low, medium, high illuminations in black and white, and color. We use the recorded eye positions to derive the gaze-centered time-varying images from the original video. The light intensities of the images are calibrated separately for the four conditions. We analyze the statistical properties of the resulting images, in particular, the temporal correlations of the input signals at a given retinal location.
Results: There are significant temporal correlations during fixations and smooth pursuits, but the macro saccades effectively remove the temporal correlation between two signals before and after a saccade. However, the magnitudes of temporal correlations during fixations and smooth pursuits depend on the visual scenes. For different viewers and for sessions of the four viewing conditions, the temporal correlation remains the same for any given scene, although different viewers/sessions have different gaze positions and different saccadic eye movement timings during their natural-viewing of the same scene.
Conclusion: Visual input statistics of natural images are independent of viewers, spectra and illuminations, but dependent on scenes and saccadic timings. This enables a common mechanism across viewers at different viewing conditions to improve the efficiency of visual information coding through temporal decorrelation of the input signal. Such a mechanism can be dynamically changed according to saccadic timing and adapted to different scenes to ensure high coding efficiency.