Abstract

Creating virtual representations of real-world scenes has been a long-standing goal in photogrammetry and computer vision, and has high practical relevance in industries involved in creating intelligent urban solutions. This includes a wide range of applications such as urban and community planning, reconnaissance missions by the military and government, autonomous robotics, virtual reality, cultural heritage preservation, and many others.

Over the last decades, image-based modeling emerged as one of the most popular solutions. The objective is to extract metric information directly from images. Many procedural techniques achieve good results in terms of robustness, accuracy, completeness, and efficiency. More recently, deep-learning-based techniques were proposed to tackle this problem by training on vast amounts of data to learn to associate features between images through deep convolutional neural networks and were shown to outperform traditional procedural techniques. However, many of the key challenges such as large displacement and scalability still remain, especially when dealing with large-scale aerial imagery.

This thesis investigates image-based modeling and proposes robust and scalable methods for large-scale aerial imagery. First, we present a method for reconstructing large-scale areas from aerial imagery that formulates the solution as a single-step process, reducing the processing time considerably. Next, we address feature matching and propose a variational optical flow technique (HybridFlow) for dense feature matching that leverages the robustness of graph matching to large displacements. The proposed solution efficiently handles arbitrary-sized aerial images. Finally, for general-purpose image-based modeling, we propose a deep-learning-based approach, an end-to-end multi-view structure from motion employing hypercorrelation volumes for learning dense feature matches. We demonstrate the application of the proposed techniques on several applications and report on task-related measures.