Abstract

A model is described for stereopsis. Two stereo images are bandpass filtered at multiple scales using a Laplacian pyramid data structure.¹ These bandpass image pairs are then matched in parallel at all scales and locations using an iterative, relaxation labelinglike process. At a given spatial location i and scale (or pyramid level) k, the state consists of a disparity estimate (the disparity mean M_i^k) and a confidence in that estimate (the disparity variance V_i^k). Disparity matches are only made near the disparity currently indicated, M_i^k. The update of M_i^k includes a data-driven term (from the matches), a local averaging term (from M_j^k, where j is a location near i in the image at scale k), a top-down term (from M_i^k+1), and a bottom-up term (from M_i^k−1). These terms are gated by the corresponding confidence terms (V_i^k, V_i^k+1, and V_i^k−1), which reflect the quality of the match. Simulations of the model are discussed both in terms of the model’s efficacy in computing stereoscopic depth, and in general concerning how computation may take place cooperatively across multiple scales.

© 1986 Optical Society of America