Modified neural models based on the bidirectional reflectance distribution functions

Abstract

Bidirectional reflectance distribution functions are used to model light reflection from object surfaces. Empirical and microfacet analytical bidirectional reflectance distribution functions are most often used, but they do not always provide sufficiently accurate reproduction of reflection features. Tables of measured values provide highly realistic representation of material reflective properties but require significant computational resources. Therefore, neural bidirectional reflectance distribution functions are relevant for dense representation of measured values. The study is aimed at developing modified neural models for more efficient and accurate modeling of light reflection. Machine learning and computer modeling research methods were used. The results of the study include the usage of physically meaningful material features together with a single block of weight parameters for compact representation of a fixed material set. It provides accurate reflectance value prediction without additional computational units or recalculation of material features. Neural bidirectional reflectance distribution functions based on material clusters are also proposed, enabling the use of a relatively small number of neural network weight files and minimal material features, thereby increasing reflectance values prediction performance. A combined process of material cluster formation is described. The adaptive combining of transformation functions for high-dynamic-range of reflectance values at the stage of neural network training is considered. Improved prediction accuracy for new test materials is demonstrated. The proposed models are implemented in the 3D-rendering tool Idx3d. In conclusion, the developed models enable the use of a single set of weight parameters, more productive reflectance prediction, and more efficient processing of training values.