Adaptive Sampling for BRDF Acquisition

Abstract The bidirectional reflectance distribution function (BRDF) describes the ratio of incoming radiance to outgoing radiance for all possible pairs of incoming and outgoing directions, defined over a spatial point. BRDF plays a key role in appearance modelling in computer graphics. Precise BRDF representation typically involves collecting millions of samples from incoming and outgoing directions, taking several hours or days of measurement using a gonioreflectometer. In this paper, we present an adaptive sampling framework for fast and accurate acquisition of BRDFs, where the number of measurements adapts to the complexity of the underlying BRDF function. We enhance the sampling efficiency of existing BRDF sampling techniques by accounting for the diverse reflectance properties of different materials. To achieve this, we categorise BRDFs in measured datasets into distinct clusters based on their sparsity and extract the necessary number of measurements for faithful reconstruction. Using a lightweight neural network, we predict the material's cluster from a single image, which allows us to determine the optimal sample count and sampling pattern, that is, the light/camera configuration. Our evaluation and analysis, compared to state‐of‐the‐art methods, demonstrate a notable performance boost, particularly for challenging materials like specular BRDFs.