A comprehensive study of Shonan Rotation Averaging algorithm for solving rotation averaging problem in structure from motion system
MetadataShow full item record
The object of the proposed work is to further understand Shonan Rotation Averaging algorithm. Rotation averaging problem is to recover the absolute camera orientations given a set of relative camera rotations. The difficulty in rotation averaging algorithm is the high dimension and non-convexity caused by large amount of camera pose number and orthogonal constraints inside rotation matrix. Shonan Averaging algorithm applies the convex relaxation to the original problem and use the duality theory to prove the approximation will generate the global optimal solution. Additionally, Shonan Averaging is able to give a global optimal certification while provides fast and accurate result. This thesis studies Shonan Averaging from several dimensions: the different initialization method, several hyper parameters inside the algorithm and comparison with other optimization methods. At the end of the thesis, I discuss some possible directions that Shonan Averaging algorithm can improve.