A matrix decomposition perspective on calibrated photometric stereo

Abstract

Leveraging on recent advances in robust matrix decomposition, we revisit Lambertian photometric stereo as a robust low-rank matrix recovery problem with both missing and corrupted entries, tailoring Grasta and R-GoDec to normal surface estimation. A method to automatically detect shadows is proposed. The performance of different robust matrix completion techniques are analyzed on the challenging DiLiGenT datasets.