Self-supervised finetuning of stereo matching algorithms

Abstract

Abstract— Stereo vision is a commonly applied method to achieve depth perception on Micro Air Vehicles (MAVs). Stereo matching algorithms are often optimized for specific environments and camera properties, using the ground truth error as a supervisor. However, in practical applications ground truth data is usually not available. Therefore, in this research, we finetune several conventional stereo matching algorithms (BM, SGBM, and ELAS) and a neural network (AnyNet) using self-supervision. The settings of the conventional algorithms are optimized with NSGA-II, using the reconstruction error and disparity density as objective functions. AnyNet is finetuned with the reconstruction error, as well as with the disparity map of conventional methods. We conclude that finetuning the parameters of conventional stereo algorithms using the reconstruction error can lead to a slight improvement in performance compared with the general settings, depending on the stereo algorithm. The performance of the conventional methods is comparable to that of AnyNet on a major portion of the image. However, removing the values with low confidence in the disparity map of ELAS and interpolating the missing disparities leads to an accuracy well above AnyNet.