Structure from Motion is one of the core problems of computer vision. Given many photos of a scene, we try to recover both the scene geometry and camera poses. Recent techniques have scaled to thousands of images automatically collected from the Internet, allowing us to build 3D models of famous places in the world from our office. However, using such unconstrained data poses additional challenges for robustness.

In this talk, I will first pose the problem, briefly discussing a pipeline for the computation, and presenting the underlying camera geometry constraints. In the second part of the talk, I will discuss current methods for solving for a scene consistent to our geometric constraints. This will conclude with some work-in-progress remarks on promising directions for solving with improved robustness.