Exploiting Local Shape and Material Similarity for Effective SV-BRDF Reconstruction from Sparse Multi-Light Image Collections

Abstract

We present a practical solution to create a relightable model from small Multi-light Image Collections (MLICs) acquired using standard acquisition pipelines. The approach targets the difficult but very common situation in which the optical behavior of a flat, but visually and geometrically rich object, such as a painting or a bas relief, is measured using a fixed camera taking a limited number of images with a different local illumination. By exploiting information from neighboring pixels through a carefully-crafted weighting and regularization scheme, we are able to efficiently infer subtle and visually pleasing per-pixel analytical Bidirectional Reflectance Distribution Functions (BRDFs) representations from few per-pixel samples. The method has a low memory footprint and is easily parallelizable. We qualitatively and quantitatively evaluated it on both synthetic and real data in the scope of image-based relighting applications.