Abstract
Natural scenes consist of objects of various physical properties that are arranged in the natural environment in a variety of ways. The higher-order statistics of natural scenes are crucial for our understanding of scene perception, space navigation, space memory, and the underlying neural mechanisms. We propose natural scene structures (NSS), i.e., topology-conserving, multi-size, multi-scale concatenations of features, as the basic features of natural scenes that are at a higher level of abstraction than pixels, edges, junctions, and textures. We took five steps to compile NSS: 1) sample a large number of circular patches in hexagonal configurations at multiple spatial scales; 2) perform independent component analysis on the circular patches and obtain independent components (ICs) at each spatial scale; 3) fit Gabor functions to the ICs and classify the ICs at multiple spatial scales into a set of clusters (referred to as IC clusters) using the parameters of the fitted Gabor functions as features; 4) project the circular patches to the IC clusters, compute the features of the circular patches, and pool the features of the patches in the hexagonal configuration at multiple spatial scales; and 5) partition the space of feature vectors into a set of NSS. The NSS obtained in this way provide a classification of natural scene patches of large sizes (1-10 degrees of visual angle) that include all concatenations of visual features and encode local topology, scaling-invariance, and scaling-variance. We examined the correlation among the NSS and found a variety of correlational patterns that are very different from the 1/f correlation at the pixel level in natural scenes. For example, the correlation may increase with spatial separation or have peaks at various spatial separations. We speculated the implications of these correlational structures on scene perception and the underlying neural mechanisms
Meeting abstract presented at VSS 2014