Optimal Activation Function for Anisotropic BRDF Modeling

0569632 - ÚTIA 2024 RIV PT eng C - Konferenční příspěvek (zahraniční konf.)
Mikeš, Stanislav - Haindl, Michal
Optimal Activation Function for Anisotropic BRDF Modeling.
Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - GRAPP. Lisbon: SciTePress, 2023 - (Sousa, A.; Bashford-Rogers, T.; Bouatouch, K.), s. 162-169. ISBN 978-989-758-634-7. ISSN 2184-4321.
[International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - GRAPP 2023 /18./. Lisbon (PT), 19.02.2023-21.02.2023]
Grant CEP: GA ČR(CZ) GA19-12340S
Institucionální podpora: RVO:67985556
Klíčová slova: Anisotropic BRDF models * neural network * activation function * BTF
Obor OECD: Automation and control systems
http://library.utia.cas.cz/separaty/2023/RO/mikes-0569632.pdf

We present simple and fast neural anisotropic Bidirectional Reflectance Distribution Function (NN-BRDF) efficient models, capable of accurately estimating unmeasured combinations of illumination and viewing angles from sparse Bidirectional Texture Function (BTF) measurement of neighboring points in the illumination/viewing hemisphere. Our models are optimized for the best-performing activation function from nineteen widely used nonlinear functions and can be directly used in rendering. We demonstrate that the activation function significantly influences the modeling precision. The models enable us to reach significant time and cost-saving in not trivial and costly BTF measurements while maintaining acceptably low modeling error. The presented models learn well, even from only three percent of the original BTF measurements, and we can prove this by precise evaluation of the modeling error, which is smaller than the errors of alternative analytical BRDF models.
Trvalý link: https://hdl.handle.net/11104/0341255