The project aims to develop deep learning approaches to create 3-dimensional representations of buildings without any human interaction. The project's ambition is to use aerial images to generate 3-dimensional objects.
The Norwegian mapping authority maintains vast amounts of geographical data with great potential. The current approach for extracting information from and using the data is time-consuming and manual. This project will focus on generating building objects for FKB-Bygning. FKB-Bygning is a database with information about all registered buildings in Norway.
The Norwegian mapping authority wishes to automate large parts of the process regarding the construction of 3-dimensional building objects for FKB-Bygning. The project will explore the use of aerial images for creating and locating 3-dimensional building objects on the map. The first step is to create deep learning models able to locate buildings in aerial images, which has already been done but not with sufficient precision. Furthermore, we have to recognize the same building in two different aerial images and use these in combination with photogrammetric methods to generate a 3-dimensional representation of the building.
The research will use relevant data from various areas in Norway to ensure the generalizability of the models. The artificial intelligence models are tested on real-world data, and therefore the results will reflect the performance in a live production setting.
The results from the project will contribute to the path toward an automatic approach for the construction of FKB-Bygning objects.
The project focuses on techniques and data sources that will enable the realization of 2.5/3-dimensional map objects. The motivation behind the project is automating labor-intensive work related to creating maps and increasing the utilization rate of the data captured. The cost-benefit relationship is central to all mapping projects. As a result, mapping projects use strict requirements and assessments before ordering a project. Through this project, we hope to benefit from and contribute to the rapid advancements in the following computer vision subfields; segmentation and 3-dimensional object reconstruction using aerial multiview images.
AI can locate and annotate objects such as buildings using orthophotos with generally good results but not precise enough for map production. However, converting the segmentation results into production-ready and fully approved vectorized objects for "Sentral Felles Kartdatabase" (SFKB)  is regarded as difficult. One of the issues is that the product specifications and registration instructions are complex and designed for a manual approach . Another issue is that most research employs a 2D mosaic of corrected aerial images (orthophoto).
The project will not use orthophoto as the sole data source but will include and use data sources such as laser scanning, oblique, and vertical images. Images in flight and cross-flight directions overlap by approximately 80% and 30%, respectively. The overlap ensures the collection of at least three images from different camera positions for each object. The various viewpoints serve as the input foundation for generating 3-dimensional models for map objects such as buildings. The final step is to convert the objects in the image to a 3-dimensional representation and convert model coordinates to map coordinates. Furthermore, utilizing oblique images enables AI to find object attributes not present in vertical images, allowing for a richer 3-dimensional object model.