Improving photometric stereo through per-pixel light vector calculation

Photometric Stereo can recover dense (at pixel-level) local surface orientation, but the subsequent reconstruction procedures used to obtain 3D shape of the scene are prone to low frequency geometric distortion. This geometric distortion is mainly due to assumption of collimated light source, to overcome the error due to collimated light source we propose a novel calibration process to dynamically calculate light vectors for each pixel with little additional computation cost. We calculate distance of object from camera using the Lambertian Diffused Maxima (LDM) region, from which the corrected light vector per-pixel is derived and the absolute dimensions of the object can subsequently be estimated. Experiments performed on synthetic as well as real data show the proposed approach offers improved performance, achieving a reduction in the estimated surface normal error of almost 2 degrees.