Abstract
It is known that static the structure of the natural environment, as characterized by the distribution of spatial frequencies found from databases of natural images, displays a characteristic distribution of spatial frequencies. Here, we analyse video sequences generated by physically moving a camera through various visual scenes to examine the static structure present in sequences of frames as well as the dynamic structure present in the motion produced. We show that the motion signal maps generated by a two-dimensional array of correlation-based motion detectors (2DMD) vary systematically across space with changes in the motion field containing important cues to self motion as well as scene structure. For each movie sequence we filtered the motion signals with first-order differential of Gaussian filters to extract a measure of the continuity of the flow-field. We also extracted the power spectra of both the static image sequence and the dynamic motion outputs and examined to what extent the motion structure was dependent on the image structure. Despite the substantial levels of noise affecting the motion signal distributions - attributed to the sparse nature of optic flow and the presence of irregular camera jitter - we are able to observe characteristic patterns as well as differences between environments, such as indoor and outdoor scenes. Considering the information present in the spatial layout and variation of the motion signal distributions arising from real-life image sequences helps us to understand optic flow processing in the context of the constraints on possible motion patterns introduced by the structure of natural images.
Supported by EPSRC and The Leverhulme Foundation.