Different deep learning stereo matching methods for stereoscopic digital elevation model estimations

Abstract

One of the most important classical problems in computer vision is stereo vision. This is simply the depth estimation of objects within images. Throughout the years’ different implementations have been to solve this problem and have done so effectively. One of the more recent applications is Digital Elevation Models in short DEMs which use satellite images to map 3-D points within a disparity map to identify the highest and lowest points. This paper aims to look at some classical implementations of this stereo vision and work towards more recent implementations that adopt deep learning techniques to estimate depth. Moreover, we will look at different processing pipelines used for DEMs and attempt to create an effective pipeline using some of the discovered technologies. Processing pipelines for DEMs have improved recently. NASA is pushing for further research to get people more interested in the subject and improve this technology further. The created pipeline uses s2p for processing satellite images and a network called GC-NET for stereo matching. Unfortunately, the results of this new pipeline were not noteworthy for several reasons that will be explored in the results section. Overall, this paper highlights some of the vast progress of stereo vision throughout the years and explores one of its many implementations and new possible solutions for DEMs.