Differentiable Linearly Transformed Cosines for Inverse Area Light Shading

Abstract

Fully decomposing material, geometry, and lighting is particularly challenging. We propose an inverse rendering method that uses active area lighting, where inverse area light shading is calculated using differential linearly transformed cosines (LTC). Area lighting with LTC enables efficient rendering without Monte Carlo integration while still providing a wider range of BRDF sampling per shot than point lighting, which makes material reconstruction more accurate. Additionally, we introduce a visibility weighting scheme for shadows as traditional LTC methods cannot handle these. We integrate our approach into both mesh and 3D Gaussian splatting pipelines, where it improves BRDF reconstruction in both cases with a PSNR improvement of >3 dB for relighting. Further, in a relighting experiment, we need 1/4 of the input photos for the same quality than using point lights. Our experiments show that our method outperforms state-of-the-art environment light and point light approaches, providing superior fidelity, stability in material and geometry reconstruction, within 15 minutes on average.

Yaoan Gao 高耀安

M.Sc. Candidate

My research interests include inverse rendering and real-time rendering.