Generic camera modeling using raxels and associated methods was recently introduced in Computer Vision. The main advantage is the applicability for any camera model, which contrasts to the many specific methods designed for a single camera model. This paper introduces a bundle adjustment based on angular error for the generic structure from motion problem. Experiments include automatic, robust and optimal estimation of scene structure and camera motion from a long image sequence acquired by a handheld non-central (calibrated) catadioptric camera.