An improved feature matching algorithm for reef surface reconstruction

Abstract

We present a new technique which accurately removes outliers in stereo correspondences and increases the number of matching features to obtain a dense 3D stereometric reconstruction. In comparison to standard approaches which make use of Random Sample Consensus (RANSAC) in removing outlying features, we introduce Subimage Template Matching (STM) using correlation by fast Fourier transform. This resulted to an increased number of matching features with much realistic pairings. To improve the density of matching features, the stereo pairs were further processed by creating new bounds with the initial computed matching features as vertices.