Data-driven Reflectance Acquisition and Modeling for Predictive Rendering

The first part of the thesis focuses on acquisition of Bidirectional Reflectance Distribution Function (BRDF) and Spatially Varying BRDF (SVBRDF)—functions that describe light-surface interactions at each point based on incoming and reflected light directions. Lightweight setups are initially explored to enable efficient SVBRDF capture; however, their accuracy falls short for predictive rendering applications, motivating the adoption of gonioreflectometer-based setups. To improve measurement efficiency of such setups, a compressed sensing framework is introduced, which incorporates a deterministic sampling strategy. Additionally, a unified formulation for sparse BRDF acquisition is presented, allowing for the adaptation of sampling patterns and sample counts to the unique properties of each material. This approach significantly enhances reconstruction quality while preserving the same sampling budget.

The second part of the thesis addresses modeling of reflectance measurements, particularly the Bidirectional Texture Function (BTF) and BRDF. Sparse representation techniques applied to existing BTF datasets prove effective in compressing texture data while enabling real-time rendering of the measured BTFs. Despite these advances, a discrepancy often arises between model-space errors introduced during approximation and the image-space errors perceived in rendered outputs. To bridge this gap, a systematic psychophysical experiment is performed to analyze the impact of BRDF modeling techniques on rendered material quality. Building on these findings, a neural metric is developed to evaluate perceptual accuracy directly in BRDF-space. This metric exhibits strong correlation with subjective human evaluations and presents the potential to guide BRDF fitting algorithms toward solutions that produce visually accurate and compelling renderings of real-world materials.