The volume of geospatial imagery has increased significantly in recent years through the access of easy-to-use collection devices such as cellphones and automated airborne platforms. Commercial drone technology allows users to collect high-resolution geo-tagged images of outdoor objects from multiple perspectives quickly and at a relatively low cost. A commonly desired goal within many industries is to rapidly convert these 2D images into 3D models that accurately represent a larger area of interest and can be used in existing visualization tools. Although semi-automated processes and photogrammetry methods have helped to reduce the time required to produce useful results compared to manual 3D modeling, they still often require significant processing time and do not scale well to large areas.

Generative AI, a modern branch of deep learning, focuses on producing synthetic output that conforms to the pattern of a training dataset. Within the domain of generative AI, neural radiance fields (NeRFs) enable the synthesis of novel views based on a collection of overlapping 2D images. In addition, the trained AI model can be utilized to produce extremely detailed 3D meshes through a process of neural surface reconstruction. The resulting mesh can then be converted into traditional open standards, enabling use of this data in a wide range of applications.

This paper describes a process that utilizes GPU-acceleration of traditionally CPU-bound algorithms combined with generative AI-based methods, enabling the rapid conversion of 2D images into geospatial 3D models which can then be represented in the 3D Tiles format. We will also demonstrate visualization in a real-time ray-tracing environment capable of interactively rendering extremely complex scenes. Finally, we will discuss modifications to the workflow that resulted in order-of-magnitude speedups, the challenges and limitations of processing geospatial imagery using these methods, and cover current developments likely to impact the future performance, fidelity, and scalability of this process.

Keywords
3D;AI;CONTENT GENERATION;DEEP LEARNING;GEOSPATIAL DATA ;OPEN STANDARDS;SYNTHETIC ENVIRONMENT;TERRAIN;VISUALIZATION

Additional Keywords
Generative AI, Neural Radiance Fields