Estimating The VIS-NIR Spectral BRDF Of A Material From A Photograph

We propose a deep learning approach to estimate the VIS-NIR spectral BRDF of a material from a photograph. Materials transmit and reflect light emitted by light sources and have a major impact on the energy performance of buildings and their appearance as perceived by users. Obtaining the spectral BRDF of a material from an RGB photograph is a practical way of gathering the data needed to simulate and optimize light sources (including natural light) in buildings. We present the spectral BRDF model we used, the database we built, the neural network architectures involved, and the results obtained. In particular, spectral reflectance over the range 780 nm - 1000 nm is estimated with a relative error of 4.1%, making an RGB camera a promising tool for estimating spectral VIS-NIR quantities when it does not itself capture wavelengths above 780 nm.