STELLAR: A LARGE SATELLITE STEREO DATASET FOR DIGITAL SURFACE MODEL GENERATION

Abstract. Stellar is a large, satellite stereo dataset. It contains rectified stereo pairs of the terrain captured by the satellite image sensors and corresponding true disparity maps and semantic segmentation. Unlike stereo vision in autonomous driving and mobile imaging, a satellite stereo pair is not captured simultaneously. Thus, the same object in a satellite stereo pair is more likely to have a varied visual appearance. Stellar provides flexible access to such stereo pairs to train methods to be robust to such appearance variation. We use publicly available data sources, and invented several techniques to perform data registration, rectification, and semantic segmentation on the data to build Stellar. In our preliminary experiment, we fine-tuned two deep-learning stereo methods on Stellar. The result demonstrates that most of the time, these methods generate denser and more accurate disparity maps for satellite stereo by fine-tuning on Stellar, compared to without fine-tuning on satellite stereo datasets, or fine-tuning on previous, smaller satellite stereo datasets. Stellar is available to download at https://github.com/guo-research-group/Stellar.