نبذة مختصرة : In this Master Thesis, we propose an efficient method for rendering complex luminaires based on neural networks. We reduce the geometric complexity of the luminaires by using a simple proxy geometry, and encode the visually-complex emitted light field by using a neural radiance field (NeRF). We tackle the multiple challenges of using NeRFs for representing luminaires, including their extreme dynamic range, their high-frequency content on the spatio temporal domain, and the spherical coverage, as well as the required modifications for seamlessly integrating our NeRF in synthetic enviroments. For that, we use a combination of non-exponential transmittance functions, and a novel loss that accounts for the HDR content as well as alpha blending for integration. We implement our model into a modern deep learning framework, and demonstrate high-quality neural rendering of such luminaires. Then, we integrate our model into the rendering software Mitsuba, and demonstrate renders with much less variance with a given sample count, simultaneously achieving a high visual quality. Finally, we propose several avenues for future work where our neural implicit luminaires could be used for importance sampling and drastically reduce rendering times.
No Comments.