Reconstruction of radiance field with neural network for real-time rendering

Abstract

Fast global illumination methods has been one of the most exiting area of rendering because it could solve many industrial problems of achieving balance between realistic rendering and stable frame rate. So this study employs a neural network to learn a radiance field from the distribution of a particular group of objects by reconstructing a 3D radiance field using spherical harmonics and neural networks. The network learned the estimated radiance distribution in space by voxel data and creating a hash grid containing information from spherical harmonics. This research also provided a new simplified light transition function that emulated the behavior of light in a specific scene, as global illumination could be broken down into a sequence of the light transport progress. It may be possible to see how the light changes throughout a 3D environment by computing the global GI using this multiple-ray bounce model.