Abstract
The environments in which we live and the tasks we must perform to survive and reproduce have shaped the design of our perceptual systems through evolution and experience. Therefore, direct measurement of the statistical regularities in natural scenes has great potential value for advancing our understanding of visual perception. To study the perception of motion, Dong and Atick (1995) measured the spatiotemporal Fourier power spectrum of image sequences sampled from commercial and homemade movies. They found that the power spectrum was well fit by modeling the world as a collection of surfaces of 1/f noise translating relative to the image plane with a speed drawn from a power-law distribution. Unfortunately, their measurements were not specific to any perceptual task.
In order to study the natural scene statistics of motion in the context of a specific perceptual task, heading estimation, we measured scene statistics from movies of first-person translation through a forest environment. Image sequences were gathered in natural scenes using a calibrated camera mounted on a custom built sliding rail. Additionally, artificial image sequences were generated using a raytracer with a model of forest scenes based on measured natural scene statistics.
The power spectrum of the natural and artificial movies resemble each other qualitatively. At low temporal frequencies or at high spatial frequencies the power spectrum behaves as one would expect from Dong and Atick's translational motion model. At higher temporal or lower spatial frequencies, however, the power spectrum differs dramatically from the translational model's predictions. A likely explanation is that the expansion motion in the scene breaks the model's translational assumption. These results imply that it is important to sample natural scene statistics in a task specific manner. Also, the results suggest that using artificially generated scenes statistics can be a valuable supplement to natural scene measurements.
This research was supported by NIH grant R01-EY11747.