Enhancing satellite semantic maps with ground-level imagery

The paper at hand introduces a novel system for producing an enhanced semantic map that leverages a reconstruction approach of street-view scenes using computer vision and machine learning techniques. Focusing on the recognition and localization of objects/entities, the composed map combines semantic information from publicly available, yet of lower accuracy, satellite images, with more detailed data from ground-level camera measurements. This merging is achieved by utilizing odometry information from a street-moving vehicle and the 3D reconstruction of its recorded view. Then, the 3D semantic segmentation results are georeferenced and superimposed on the semantic map from the satellite images. In such a way, areas that require fine semantic accuracy can be improved, while the rest are left with the segmentation results of the satellite information. Every part of the proposed system is individually evaluated. We additionally test the overall approach on a case-study of georeferencing new labels of traffic signs, which are detected through a specifically designed classification network over a publicly available dataset collected around the city of Berlin. (C) 2021 Elsevier B.V. All rights reserved.