Constructing the cyclopean view

Abstract

A new neural network approach to stereovision is presented, which is able to fuse the left and right stereo images into the Cyclopean view of the scene during disparity calculations. Fusion and disparity calculations are achieved within a single network structure by utilizing coherence detection among independently working, simple disparity estimators; the resulting dense disparity maps display hyper-acuity in accordance with human stereo vision. Since the proposed algorithm is non-iterative, it is much faster than classical cooperative approaches to stereo vision.