Surface Reconstruction in Photometric Stereo with Calibration Error

Abstract

A method is described for surface reconstruction that accounts for the calibration errors in photometric stereo. The angular errors in calibrated light directions due to noise cause errors in surface normal estimates. Investigation of the effect of these calibration errors on the surface normals revealed that errors in the estimated light directions and in the surface normal estimates follow a Fisher distribution. By accounting for the Fisher noise in surface normals, the proposed method reconstructs a surface using maximum likelihood estimation. Extensive comparison with previous methods using synthetic and real scenes demonstrated that the proposed method outperforms them in the presence of calibration errors.