Epiflow—A paradigm for tracking stereo correspondences

Abstract

We propose the epiflow as a novel framework to integrate the motion and stereo epipolar geometries for the tracking of stereo matches. For a given stereo pair, spatio-temporal correspondences in the next stereo pair are established from the solution of an optimization problem that incorporates the epiflow geometrical constraint with the brightness constancy assumption as the photometric similarity measure. Two different implementations are proposed, each utilizing the stereo RANSAC algorithm as a mechanism to gain robustness and efficiency in handling large temporal disparity ranges, identify outliers and estimate the motion of the stereo rig as complementary information. The performance of the proposed methods is demonstrated in experiments with various complex underwater images, and comparison with results from the SIFT method allows the assessment of our novel contributions.