Meeting Abstract
While the characteristics of static images have been explored in large image repositories, much less is known or measured in the temporal domain. We describe a new database for natural motion, which is comprised of a large collection of fixed-camera, high resolution grayscale and color video from field recordings of moving objects. Videos include motion of both animate and inanimate objects from a variety of camera-to-subject distances. This database effort carefully selects natural subjects to maximize the quantity, quality and continuity of motion in each clip. These data are well-suited for use as stimuli in visual neuroscience experiments because of their variety of motion and the fine temporal and spatial information available in each recording. We also report on preliminary analysis of the motion content of these scenes. Static images strikingly exhibit a power-law distribution of spatial frequencies. Power-law behavior in the frequency distribution of temporal fluctuations in total scene luminance has also been observed in natural contexts, and we find that our scenes similarly display power-law behavior with a scaling exponent that depends on the particular motion content. To quantify the dynamics of motion in our scenes, we fit translation fields to pairs of stimuli and analyze the statistics of the resulting flow. These flow fields are then used to guide the trajectories of “tracer” particles released at various points in the frame and subjected to the inferred flow from the natural movie. One important aspect understanding of predictive computation in the brain of the observer of natural motion (a predator, for example) is characterizing the structure of predictable events in the world itself. These tracer particles allow us to define the predictive information content of naturalistic trajectories.
