Abstract

It has been suggested that image representation in the visual system is matched to the spatial statistics of natural imagery.¹ The extension of this idea to include time requires the development of a reasonable model of the spatiotemporal spectrum of real-world time-varying imagery.² We have calculated the spatiotemporal power spectrum of 14 image sequences (256 × 256 × 64 at 30 frames/s with no scene cuts). The sequences include scenes with camera pans, object tracking, and object motion with no camera motion. They represent a small ensemble of time-varying imagery the visual system might encounter in the real world. The spectrum of each sequence was found to be spatiotemporally separable to a large degree and was well fit by a model of the form S(|k|,f) = [(2π |k|/a)² + 1]⁻ n1 [(2πf/b)² + 1]⁻n2,where k and f are spatial and temporal frequencies and a, b, n₁, and n₂ are parameters computed separately for each sequence. Reasonable fits to all spectra (in the mean-squared-error sense) were obtained with n₁ = 1.5 and n₂ = 1.0, which suggests that the image sequences can also be modeled reasonably well by an exponential autocorrelation function that is spatiotemporally separable.

© 1990 Optical Society of America