Performance Comparison of BRDF Data Interpolation Using Machine Learning

Reproducing surface reflection accurately is crucial for realistic rendering and quality assessment in product design. The bidirectional reflectance distribution function (BRDF) quantifies these properties, but high-resolution BRDF measurement is often time-consuming and costly. This study investigates interpolation techniques for reconstructing high-resolution BRDF data from low-resolution measurements. We compared conventional methods (linear, spline) with machine learning approaches: random forest (RF), multilayer perceptron (MLP), and extreme gradient boosting (XGB). A dataset of 149 diverse material samples was measured with a multi-angle spectrophotometer, and reconstruction accuracy was evaluated using root mean squared error (RMSE). On average, MLP achieved the lowest RMSE, though the overall performance differences across the methods were not statistically significant. However, a detailed analysis revealed that interpolation accuracy is highly dependent on material characteristics. Conventional methods were sufficient for diffuse materials, while MLP significantly outperformed all other methods for materials with sharp specular or retroreflective peaks. This suggests that the optimal interpolation strategy should be chosen based on the material’s properties. Overall, our findings demonstrate that machine learning models, especially MLP, are highly effective for interpolating materials with complex reflection behaviors.