An Unsupervised Network for Stereo Matching of Very High Resolution Satellite Imagery

Abstract

Convolutional Neural Networks (CNN) were recently applied to stereo matching, which is one of the important steps in 3D reconstruction. In remote sensing using Very High Resolution (VHR) satellite imagery, CNN-based stereo matching is employed in Digital Surface Models (DSM) generation, achieving higher accuracy than conventional methods through supervised learning. Since ground-truth disparity maps calculated from reliable elevation data are hard to collect, unsupervised stereo matching methods are as significant as supervised ones. However, there are few studies on stereo matching of VHR satellite imagery with unsupervised learning. In this work, we propose an unsupervised stereo matching network for DSM generation. We also introduce a criterion to select the best epoch without using ground-truth data for validation in a training strategy of gradually increasing the weight of a smoothness loss. Experimental results show that our network performs better in the average endpoint error and the fraction of erroneous pixels than the baseline method of the used dataset without ground-truth data.