KiloNeRF: Speeding up Neural Radiance Fields with Thousands of Tiny MLPs

  • 2021-03-25 10:53:05
  • Christian Reiser, Songyou Peng, Yiyi Liao, Andreas Geiger
  • 56

Abstract

NeRF synthesizes novel views of a scene with unprecedented quality by fittinga neural radiance field to RGB images. However, NeRF requires querying a deepMulti-Layer Perceptron (MLP) millions of times, leading to slow renderingtimes, even on modern GPUs. In this paper, we demonstrate that significantspeed-ups are possible by utilizing thousands of tiny MLPs instead of onesingle large MLP. In our setting, each individual MLP only needs to representparts of the scene, thus smaller and faster-to-evaluate MLPs can be used. Bycombining this divide-and-conquer strategy with further optimizations,rendering is accelerated by two orders of magnitude compared to the originalNeRF model without incurring high storage costs. Further, using teacher-studentdistillation for training, we show that this speed-up can be achieved withoutsacrificing visual quality.

 

Quick Read (beta)

loading the full paper ...