This study introduces Neural Refractive-Reflective Fields (NeRRF) which improve the shortcomings of Neural Radiance Fields (NeRF) in dealing with transparent or specular objects. The method uses refractive-reflective field to model the refraction and reflection effects of the object in a unified framework. A virtual cone supersampling technique is proposed for anti-aliasing. The method is benchmarked on different shapes and backgrounds and on both real-world and synthetic datasets. It is also applied in various editing applications, such as material editing, object replacement/insertion, and environment illumination estimation.
Publication date: 25 Sep 2023
Project Page: NeRRF
Paper: https://arxiv.org/pdf/2309.13039
