Photogrammetry: Building a 3D City Digital Twin Under a 16 GB Memory Limit

Building digital 3D models of architectural landmarks and entire cities is a large-scale and computationally demanding problem. In this talk, we will discuss the engineering aspects of an algorithm published at ICCV 2021 that reconstructs a surface as a triangular 3D mesh from depth maps and/or LiDAR scans.

We will cover what it takes to make such an algorithm practical for large-scale data processing:

• out-of-core processing for datasets that do not fit into memory;
• running on a cluster;
• sorting terabyte-scale data arrays;
• optimizing reads from disks and NAS;
• GPU acceleration using OpenCL;
• invariants and fast diagnosis of user crash reports: is it an algorithmic bug or unstable hardware?