Provably Optimal Solutions to Geometric Vision Problems

Опубликовано 6 сентября 2016, 16:30

In the past, the main methods for solving problems in Multiview Vision Geometry have been iterative techniques, which may suffer from falling into local minima, and trouble with convergence.  Recent research has turned to finding guaranteed globally optimal solutions to such problems.  Techniques include quasi-convex optimization, Second Order Cone Programming, Branch-and-bound techniques and fractional programming, solving many of the common vision geometry problems.  In this talk, we address problems such as essential-matrix estimation, many-view triangulation and motion of a vehicle with rigidly placed cameras.  We provide optimal solutions, (in L2 or L-infinity norm) where no such solution existed previously (as of 2007). A further applications of such methods in tracking a deforming sheet of material is discussed.