De-lighting a High-resolution Picture for Material Acquisition

Abstract

We propose a deep-learning based method for the removal of shades, projected shadows and highlights from a single picture of a quasi-planar surface captured in natural lighting conditions with any kind of camera device. To achieve this, we train an encoder-decoder to process physically based materials, rendered under various lighting conditions, to infer its spatially-varying albedo. Our network processes relatively small image tiles (512x512 pixels) and we propose a solution to handle larger image resolutions by solving a Poisson system across these tiles.