Global robust image rotation from combined weighted averaging

In this paper we present a novel rotation averaging scheme as part of our global image orientation model. This model is based on homologous points in overlapping images and is robust against outliers. It is applicable to various kinds of image data and provides accurate initializations for a subsequent bundle adjustment. The computation of global rotations is a combined optimization scheme: First, rotations are estimated in a convex relaxed semidefinite program. Rotations are required to be in the convex hull of the rotation group SO(3), which in most cases leads to correct rotations. Second, the estimation is improved in an iterative least squares optimization in the Lie algebra of . In order to deal with outliers in the relative rotations, we developed a sequential graph optimization algorithm that is able to detect and eliminate incorrect rotations. From the beginning, we propagate covariance information which allows for a weighting in the least squares estimation. We evaluate our approach using both synthetic and real image datasets. Compared to recent state-of-the-art rotation averaging and global image orientation algorithms, our proposed scheme reaches a high degree of robustness and accuracy. Moreover, it is also applicable to large Internet datasets, which shows its efficiency.