Abstract
Estimating the absolute pose of a camera relative to a 3D representation of a scene is a fundamental step in many geometric Com- puter Vision applications. When the camera is calibrated, the pose can be computed very efficiently. If the calibration is unknown, the prob- lem becomes much harder, resulting in slower solvers or solvers requiring more samples and thus significantly longer run-times for RANSAC. In this paper, we challenge the notion that using minimal solvers is always optimal and propose to compute the pose for a camera with unknown focal length by randomly sampling a focal length value and using an ef- ficient pose solver for the now calibrated camera. Our main contribution is a novel sampling scheme that enables us to guide the sampling process towards promising focal length values and avoids considering all possi- ble values once a good pose is found. The resulting RANSAC variant is significantly faster than current state-of-the-art pose solvers, especially for low inlier ratios, while achieving a similar or better pose accuracy.