Hypernetwork-Based Spectral BRDF Reconstruction from Sparse RGB and Spectral Measurements
AI · 14.11.2026 · 7 (11) · 487
DOI: 10.3390/ai7110487 ↗Kurz gefasst
Fachartikel in der Zeitschrift AI, online erschienen am 14.11.2026. Verfasst von Roman Rodionov und 4 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Russian Science Foundation.
Abstract (Originalsprache)
The optical properties of materials are essential for computer-based lighting simulation and computer vision tasks. They are commonly described by a bidirectional reflectance distribution function (BRDF). This paper proposes a hypernetwork-based architecture for reconstructing spectral BRDFs from sparse measurements acquired in either spectral or RGB modes. The proposed neural network encodes measured data into a common latent space, from which a neural representation of the spectral material properties can further be decoded. The network accepts RGB input data directly, without prior conversion to a spectral format. To address the limited availability of measured spectral materials, the method is trained jointly on measured and synthetic spectral BRDFs and supports optional test-time refinement of the inferred material code. For sparse spectral inputs, test-time refinement improves all reported image-based quality metrics (PSNR, SSIM, and spectral angle error), regardless of the number of sampled direction pairs. With this refinement, the proposed method achieves the highest PSNR and SSIM among the PCA-based and per-material neural baselines, and the lowest spectral angle error for sparse inputs (up to 400 direction pairs). For RGB inputs, the impact of refinement on spectral accuracy depends on the lighting conditions, a consequence of the metameric ambiguity inherent in the RGB representation. Each reconstructed material is compactly stored as a 160-byte latent code together with a shared decoder. We also demonstrate that smooth interpolation between the obtained material representations can be used to generate new materials.
Förderung
- Russian Science Foundation
Zitieren
APA
Rodionov, R., Prikhodko, E., Sanzharov, V., Mayorov, N., & Voloboy, A. (2026). Hypernetwork-Based Spectral BRDF Reconstruction from Sparse RGB and Spectral Measurements. AI, 7(11), 487. https://doi.org/10.3390/ai7110487
Vancouver
Rodionov R, Prikhodko E, Sanzharov V, Mayorov N, Voloboy A. Hypernetwork-Based Spectral BRDF Reconstruction from Sparse RGB and Spectral Measurements. AI. 2026;7(11):487. doi:10.3390/ai7110487
BibTeX
@article{Rodionov2026,
author = {Rodionov, Roman and Prikhodko, Egor and Sanzharov, Vadim and Mayorov, Nikolay and Voloboy, Alexey},
title = {Hypernetwork-Based Spectral BRDF Reconstruction from Sparse RGB and Spectral Measurements},
journal = {AI},
year = {2026},
volume = {7},
number = {11},
pages = {487},
doi = {10.3390/ai7110487}
}
EUVIMED