Learning to estimate surface normal via deep photometric stereo networks

Photometric stereo, aiming at estimating the surface normal of an object from a set of images under different illumination conditions, has gained a lot of attention recently. However, most existing state-of-the-art works of photometric stereo heavily rely on elaborately light calibration which limit the practical application of this technology. In this paper, we propose a self-calibrating photometric stereo method which could accurately reconstruction the surface normal. Specifically, a two-stage deep architecture is developed to perform light calibration and surface normal estimation simultaneously. Extensively experiment results on public available real datasets demonstrate that our model could estimate surface normal more accurately than most state-of-the-arts.