Sparse-as-Possible SVBRDF Acquisition

摘要

We present a novel method for capturing real-world, spatiallyvaryingrnsurface reflectance from a small number of object viewsrn(k). Our key observation is that a specific target’s reflectance can bernrepresented by a small number of custom basis materials (N) convexlyrnblended by an even smaller number of non-zero weights atrneach point (n). Based on this sparse basis/sparser blend model, werndevelop an SVBRDF reconstruction algorithm that jointly solvesrnfor n, N, the basis BRDFs, and their spatial blend weights withrnan alternating iterative optimization, each step of which solves arnlinearly-constrained quadratic programming problem. We developrna numerical tool that lets us estimate the number of views requiredrnand analyze the effect of lighting and geometry on reconstructionrnquality. We validate our method with images rendered from syntheticrnBRDFs, and demonstrate convincing results on real objectsrnof pre-scanned shape and lit by uncontrolled natural illumination,rnfrom very few or even a single input image.