Local stereoscopic depth estimation

摘要

A method is discussed as to how a fixating binocular observer can recover local depth information with a single step computation avoiding the correspondence problem motivated from recent findings about the architecture of biological visual systems. Visual information is represented in the primate visual cortex (area 17, layer 4B) in a peculiar structure of alternating bands of left and right eye dominance. Recently, a number of computational algorithms based on this ocular stripe map architecture have been proposed. We investigated the cepstral filtering method of Yeshurun and Schwartz1 for fast disparity computation because of its simplicity and robustness. Based on a systematic investigation and evaluation of the properties of the cepstral filter, some deficiencies are discussed, and improvements to the algorithm are presented. The introduction and brief review of the biological background may be skipped, if the reader is interested only in the technical aspects of the method. In summary, we consider Gaussian window functions for the extraction of local image patches superior to rectangular windows because specific configurations are avoided where additional maxima in the cepstral output may disturb the detection of the correct peak. We show experimentally that exact disparity estimates can still be obtained by the filter, even when one of the subsignals undergoes moderate rotation (3°) or scaling (4%). The discussed framework is a fairly robust, single-step method for local depth estimation. We present results for synthetic as well as real image pairs.