Outlier Rejection for Absolute Pose Estimation with Known Orientation

Estimating the pose of a camera is a core problem in many geometric vision applications. While there has been much progress in the last two decades, the main difficulty is still dealing with data contaminated by outliers. For many scenes, e.g. with poor lightning conditions or repetitive textures, it is common that most of the correspondences are outliers. For real applications it is therefore essential to have robust estimation methods. In this paper we present an outlier rejection method for absolute pose estimation. We focus on the special case when the orientation of the camera is known. The problem is solved by projecting to a lower dimensional subspace where we are able to efficiently compute upper bounds on the maximum number of inliers. The method guarantees that only correspondences which cannot belong to an optimal pose are removed. In a number of challenging experiments we evaluate our method on both real and synthetic data and show improved performance compared to competing methods.