Incorporating Lambertian Priors into Surface Normals Measurement

The goal of photometric stereo is to measure theprecise surface normal of a 3D object from observations withvarious shading cues. However, non-Lambertian surfaces in-fluence the measurement accuracy due to irregular shadingcues. Despite deep neural networks have been employed tosimulate the performance of non-Lambertian surfaces, the errorin specularities, shadows, and crinkle regions is hard to bereduced. In order to address this challenge, we here propose aphotometric stereo network that incorporates Lambertian priorsto better measure the surface normal. In this paper, we usethe initial normal under the Lambertian assumption as theprior information to refine the normal measurement, insteadof solely applying the observed shading cues to deriving thesurface normal. Our method utilizes the Lambertian informationto reparameterize the network weights and the powerful fittingability of deep neural networks to correct these errors causedby general reflectance properties. Our explorations include: theLambertian priors (1) reduce the learning hypothesis space,making our method learn the mapping in the same surfacenormal space and improving the accuracy of learning, and(2) provides the differential features learning, improving thesurfaces reconstruction of details. –Extensive experiments verifythe effectiveness of the proposed Lambertian prior photometricstereo network in accurate surface normal measurement, on thechallenging benchmark dataset.