Use camera positions to more accuratly compute camera rotations

We could basically use the image positions now, to know which images are close to each other. This should be

[ ] integrated in the Epipolar geometry (only compute R)
[ ] Instead of presorting, use x closest cameras to do reconstruction. So we loop through all the cameras and for each camera we compare them to the closest ones

lole-elol / tum_cv_challenge_SS23