DeRF: Decomposed Radiance Fields

TL;DR: In this article, the authors propose a technique based on spatial decomposition capable of mitigating the limitations of neural networks in real-world scenarios, which is provably compatible with the Painter's algorithm for GPU-friendly rendering.

Abstract: With the advent of Neural Radiance Fields (NeRF), neural networks can now render novel views of a 3D scene with quality that fools the human eye. Yet, generating these images is very computationally intensive, limiting their applicability in practical scenarios. In this paper, we propose a technique based on spatial decomposition capable of mitigating this issue. Our key observation is that there are diminishing returns in employing larger (deeper and/or wider) networks. Hence, we propose to spatially decompose a scene and dedicate smaller networks for each decomposed part. When working together, these networks can render the whole scene. This allows us near-constant inference time regardless of the number of decomposed parts. Moreover, we show that a Voronoi spatial decomposition is preferable for this purpose, as it is provably compatible with the Painter’s Algorithm for efficient and GPU-friendly rendering. Our experiments show that for real-world scenes, our method provides up to 3× more efficient inference than NeRF (with the same rendering quality), or an improvement of up to 1.0 dB in PSNR (for the same inference cost).