Joanna Kaleta, Kacper Kania, Tomasz Trzcinski, Marek Kowalski
Separating lighting from geometry using unconstrained photo collections is a challenging task. Many previous attempts have either compromised the quality of the final output or suffered from slow training and inference speeds, raising doubts about their practicality. We introduce LumiGauss, a technique designed for 3D scene reconstruction using 2D Gaussian Splatting. Our approach achieves high-quality reconstructions while enabling the synthesis of realistic lighting under novel environmental maps. This solution would be highly advantageous, as it could drastically cut down the extensive manual effort required to create sophisticated 3D assets, which often takes several days.
LumiGauss leverages precomputed radiance transfer, allowing seamless integration with graphics engines. Our representation also enables accurate shadow modeling, significantly enhancing the realism of generated renderings. We validate our method on the NeRF-OSR dataset, demonstrating its accuracy. Additionally, LumiGauss exhibits improved efficiency compared to baseline approaches. Access our code here: https://github.com/joaxkal/lumigauss
Authors: Joanna Kaleta, Kacper Kania, Tomasz Trzcinski, Marek Kowalski
DOI: 10.48550/arXiv.2408.04474
Keywords: 3D reconstruction, Environmental lighting, Gaussian Splatting, Realistic lighting synthesis, Precomputed radiance transfer, NeRF-OSR dataset, LumiGauss