Link Search Menu Expand Document (external link)

Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields

[Project Page] [Paper] [Code]

Authors

  • Jonathan T. Barron
  • Ben Mildenhall
  • Dor Verbin
  • Pratul P. Srinivasan
  • Peter Hedman

Table of contents


What’s Ref-NeRF?

Ref-NeRF

Current NeRF-like methods fail to reconstruct the scenes, which contain glossy and reflective surfaces. To overcome these failure cases, Ref-NeRF proposes a new parameterization with reflected radiance producing spatially-varying material properties. They show that Ref-NeRF outperforms the accuracy of the normal map which synthesizes the realistic reflective surface like a chrome ball. With the re-parameterization and regularizer, Ref-NeRF renders the scene including specularities and reflections with the correct normal map crucial for generating a shiny material scene.

refnerf

Scores

Blender

ScenePSNRSSIMLPIPSCheckpoint
chair35.840.98420.0223link
drums25.520.93400.0735link
ficus31.320.97120.0407link
hotdog36.540.98060.0347link
lego35.790.97950.0240link
materials35.710.98480.0280link
mic35.960.99120.0104link
ship29.510.87070.1643link

Shiny Blender

ScenePSNRSSIMLPIPSCheckpoint
ball43.090.99330.0921link
car30.700.95540.0454link
coffee32.270.96680.1242link
helmet29.660.95850.1033link
teapot45.200.99600.0176link
toaster24.880.91410.1318link