Continuous material reflectance map for deep photometric stereo.

Solving calibrated photometric stereo under a sparse set of lights is of great interest for real-world applications. Since neural networks show advantages in dealing with material appearance, this paper proposes a bidirectional reflectance distribution function (BRDF) representation, which is based on reflectance maps for a sparse set of lights and can handle various types of BRDFs. We discuss the optimal way to compute these BRDF-based photometric stereo maps regarding the shape, size, and resolution, and experimentally investigate the contribution of these maps to normal map estimation. The training dataset was analyzed to establish the BRDF data to use between the measured and parametric BRDFs. The proposed method was compared to state-of-the-art photometric stereo algorithms for different datasets from numerical rendering simulations, DiliGenT, and our two acquisition systems. The results show that our representation outperforms the observation maps as BRDF representation for a neural network for various surface appearances on specular and diffuse areas.