Abstract
The perceptual organization of natural scenes requires a global evaluation of spatially separated retinal image features. In the encoding of border-ownership, for instance, information from remote image locations co-determines the belongingness of a boundary to one of the two abutting surfaces. In natural viewing conditions a stable perceptual organization is established in a split-second, despite the low quality and the inherent ambiguity of the retinal input. This quick inference seems only possible through the use of prior knowledge on the structure of both the world and the visual input resulting from that world. Over the past few decades, natural systems analysis has successfully identified local-scale structure in natural images, and has related this structure to locally defined principles of perceptual organization (e.g., good continuation, co-linearity). In the current study, we extract more global scene statistics related to contours, shapes and objects, and examine their role in global-scale organizational processes. First, we record the contrast-polarity of luminance-defined oriented edges in natural images. We then analyze the consistency of contrast-polarity within a pair of spatially separated edges. We find that this consistency statistically depends on the geometrical relationship between the edges. Specifically, the resulting pattern reveals a co-circular organization of edges in natural images. To investigate the source of the observed co-circularity, we apply the same analysis on a subset of edges that do or do not coincide with human-traced object boundaries. We show that co-circularity is mainly preserved when the analysis is constrained on edges that fall on object boundaries, while it is largely reduced for edges that do not coincide with object boundaries. We conclude that large-scale statistical regularities in the input can be exploited by the visual system in global processes of perceptual organization in general, and in the computation of border-ownership in particular.
Meeting abstract presented at VSS 2015