Abstract
Recent work has demonstrated that deep learning approaches can successfullybe used to recover accurate estimates of the spatially-varying BRDF (SVBRDF) ofa surface from as little as a single image. Closer inspection reveals, however,that most approaches in the literature are trained purely on synthetic data,which, while diverse and realistic, is often not representative of the richnessof the real world. In this paper, we show that training such networksexclusively on synthetic data is insufficient to achieve adequate results whentested on real data. Our analysis leverages a new dataset of real materialsobtained with a novel portable multi-light capture apparatus. Through anextensive series of experiments and with the use of a novel deep learningarchitecture, we explore two strategies for improving results on real data:finetuning, and a per-material optimization procedure. We show that adaptingnetwork weights to real data is of critical importance, resulting in anapproach which significantly outperforms previous methods for SVBRDF estimationon real materials. Dataset and code are available athttps://lvsn.github.io/real-svbrdf