Light Field Networks: Neural Scene Representations with Single-Evaluation Rendering

Inferring representations of 3D scenes from 2D observations is a fundamental of computer graphics, computer vision, and artificial intelligence. 3D-structured neural scene representations are a promising approach to3D scene understanding. In this work, we propose a novel neural scene, Light Field Networks or LFNs, which represent both geometry and of the underlying 3D scene in a 360-degree, four-dimensional light parameterized via a neural implicit representation. Rendering a ray from LFN requires only a single network evaluation, as opposed to hundreds of per ray for ray-marching or volumetric based renderers in3D-structured neural scene representations. In the setting of simple scenes, we meta-learning to learn a prior over LFNs that enables multi-view light field reconstruction from as little as a single image. This results in dramatic reductions in time and memory complexity, enables real-time rendering. The cost of storing a 360-degree light field an LFN is two orders of magnitude lower than conventional methods such as Lumigraph. Utilizing the analytical differentiability of neural implicit and a novel parameterization of light space, we further the extraction of sparse depth maps from LFNs.