Structure from Motion (SfM) is a problem which asks: given photos of an object from different angles, can we reconstruct the object in 3D? This problem is important in computer vision, with applications including urban planning and autonomous navigation. A key part of SfM is bundle adjustment, where initial estimates of 3D points and camera locations are refined to match the images. This results in a high-dimensional nonlinear least-squares problem, which is typically solved using the Gauss-Newton method. In this talk, I will discuss how dimensionality reduction methods such as block coordinates and randomised sketching can be used to improve the scalability of Gauss-Newton for bundle adjustment problems.
- Numerical Analysis Group Internal Seminar