Revisiting the PnP Problem with a GPS

This paper revisits the pose estimation from point correspondences problem to properly exploit data provided by a GPS. In practice, the location given by the GPS is only a noisy estimate, and some point correspondences may be erroneous. Our method therefore starts from the GPS location estimate to progressively refine the full pose estimate by hypothesizing correct correspondences. We show how the GPS location estimate and the choice of a first random correspondence dramatically reduce the possibility for a second correspondence, which in turn constrains even more the remaining possible correspondences. This results in an efficient sampling of the solution space. Experimental results on a large 3D scene show that our method outperforms standard approaches and a recent related method [1] in terms of accuracy and robustness.