Nonparametric Correction of Distortion 1 Nonparametric Correction of Distortion 1

Previous work on camera calibration, and particularly work on self-calibration, has assumed that the camera is well modeled by an ideal, relatively simple, parametric model. Real camera systems, however, diverge signiicantly from a simple pinhole perspective imaging model. Therefore, parametric calibration models are typically extended to include parameters which model distortions from perspective. However, tting these extra parameters greatly complicates calibration algorithms. This paper presents a new approach to camera calibration, which avoids elaborate parametric models. The imaging geometry of a camera is modeled by a 3 degree-of-freedom stereographic imaging model, together with a smooth distortion function D. D maps the raw camera output to the ideal model, subsuming the eeects of traditional parametric distortion parameters. The calibration algorithm reconstructs D from images of spheres, using the fact that spheres must appear circular in ideal stereographic projection. Accuracy and stability results for this algorithm are presented. The remaining free parameters (image center and focal length) can be estimated using several standard methods. New methods are presented for estimating the focal length (given the image center) using a single unknown image rotation and for pasting several images, taken by rotating the camera, into a panoramic image.