Multi-scale scene graph generation for remote sensing imagery

Abstract. The map, as a way of representing geospatial data, is designed to reflect important information about the Earth as deeply and accurately as possible. To meet this requirement, maps are produced in different scales and different types, depending on the task being solved. Created by highly educated specialists, the map contains not only raw geospatial data, but also some high-level knowledge accumulated by people during the exploration of the Earth. The introduction of deep learning into the data analysis process has allowed the development of neural network models that can solve complex aerial image processing tasks, such as semantic image segmentation, object detection and recognition, and retrieving of semantic relations between objects in a scene. These advances created the background for moving to image (scene) understanding as a higher level of image analysis. The current study addresses to a problem of multi-scale scene graph generation from aerial images, similarly to creating maps of different scales.