libBICOS – An Open Source GPU-Accelerated Library implementing BInary COrrespondence Search for 3D Reconstruction

Abstract. In this paper, we present an implementation and publish an open source library for binary correspondence search (BICOS), an efficient method for accurate 3D reconstruction from structured light stereo imagery. Starting with two stacks of stereo-rectified images of a scene illuminated by a statistical light pattern the proposed method solves the problem of a pixelwise correspondence search. Our GPU-accelerated implementation reduces the latency of disparity computation using 7MP images on modern hardware down to 20 milliseconds. Based on the algorithm introduced by Dietrich et al. (2019), we extend their approach by increasing the descriptor size while augmenting postprocessing to increase its applicability on other types of pattern projections. Lastly, we provide benchmarks and example reconstructions using a stereo camera setup combined with an off-the-shelf projector to validate the algorithm’s performance. While many state-of-the-art single-shot stereo implementations are included in common computer vision libraries, high performance multi-shot methods are not broadly available. By publishing this method as a freely available library, in both a CUDA and CPU implementation, we hope for others to quickly gain traction in this field. The source code with build instructions and command-line tooling is available at https://github.com/JMUWRobotics/libBICOS under the GNU LGPLv3.