Skip to main content

Research · Technical

NeRF: Representing Scenes as Neural Radiance Fields

ECCV 2020 · Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T., Ramamoorthi, R., Ng, R. · Mar 2020

Key points

  1. NeRF represents a scene as a continuous volumetric function optimised from posed photographs.
  2. Differentiable volume rendering produces photorealistic novel views without explicit geometry.
  3. It is the named origin of Instant-NGP, Nerfstudio and 3D Gaussian splatting.

Summary

Mildenhall et al. (2020) introduced Neural Radiance Fields, a method that represents a 3D scene as a continuous function mapping spatial coordinates and viewing direction to colour and volume density, parameterised by a multilayer perceptron optimised from a set of posed input photographs. Differentiable volume rendering integrates the field along camera rays to produce novel-view images that match the input photographs. NeRF received an ECCV 2020 Honourable Mention and spawned a large successor family including Instant-NGP (real-time training), Nerfstudio (production tooling) and 3D Gaussian Splatting (explicit primitive representation). These successors are relevant to VFX and virtual-production workflows; none of them are self-explanatory without NeRF.

Source

Source: ECCV 2020 ↗ (Research)

Cite this item

Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T., Ramamoorthi, R. & Ng, R. (2020). ‘NeRF: Representing Scenes as Neural Radiance Fields’, ECCV 2020. Available at: https://arxiv.org/abs/2003.08934

Your reference manager can also read this page directly: with the Zotero (or Mendeley) browser connector installed, save it straight to your library. Whole-collection exports: RIS, BibTeX, CSL-JSON.