Abstract
Introduction: Theories of vision, including those for motion, are increasingly relying on statistics of natural scenes. Much work has been done for stationary natural images, but little on natural movies. In this work, we analyzed the distribution of instantaneous local velocities in the natural optic flow, and their relationship with the distributions of spatial and temporal contrasts. Methods: We obtained and calibrated nineteen 1-min-long gray-level movies shot at outdoors human settings. We measured directly the spatial and temporal gradients, and from them, local velocities on 3000 points randomly placed on the movie frames. Noise affecting velocity estimates due to CCD fluctuations was eliminated with robust, nonparametric statistics. Results: The distribution of speeds of velocities parallel to the spatial gradient fall with speed (Horn & Schunck, 1981). A similar result holds for the distribution of real speeds, but they fall more slowly. The distribution of directions of motion shows a tendency towards horizontal (and sometimes vertical); the coefficient of variation quantifying this anisotropy is between 0.01 and 0.6. We see no correlation in the joint distributions of speed and spatial contrast, and speed and temporal contrast. Discussion: That the distribution of local speeds fall as speed increases is consistent with the analysis of Yuille & Ullman (1989). They predicted such a distribution from images of 3D-rigid objects moving in the world. This distribution supports the assumptions of the Yuille & Grzywacz (1988) and the Weiss et al. (2002) models of motion perception. In turn, our finding that there is no correlation in the joint distribution of speed and contrast, rules out simple Bayesian-prior explanations for why perceived speed of motion varies with contrast (Thompson, 1982). The anisotropy of direction of motions might be due to objects moving mainly parallel to ground.
Funded by a Grant from the Programa Ramòn y Cajal of the Spanish Ministerio de Ciencia y Tecnología to R.M.B. and by NEI Grants EY08921 and EY11170 to N.M.G.