Graphical Models and Visual SLAM
Visual SLAM, in one form, can be represented as a graphical model over camera poses and 3D point positions, where the observations are projections of the points on the cameras. To solve the graph efficiently, Sparse Bundle Adjustment exploits the sparse primay structure of the model, where cameras are connected only through 3D points. In many applications, there is also a sparse secondary structure, since the cameras are only locally connected by commonly-viewed points. We show how these systems, which arise naturally in Visual SLAM, can be solved very efficiently using a combination of sparse structure engineering and sparse linear solvers.