Link Search Menu Expand Document (external link)

Mip-NeRF360: Unbounded Anti-Aliased 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 Mip-NeRF 360?

Mip-NeRF 360

As NeRF++ tackles the unbounded scene problem, they extend the Mip-NeRF to the unbounded scene to deal with the real-world arbitrary camera pose orientation. Proposal-based distillation method efficiently generates volumetric density. Volume representation with contracted Gaussian region accurately models the unbounded 360 scenes. A novel distortion-based regularizer removes the “floater” artifacts.

mipnerf360

Scores

LLFF

ScenePSNRSSIMLPIPSCheckpoint
fern24.580.80470.2378link
flower27.810.85380.1607link
fortress31.170.89500.1259link
horns28.030.89220.1592link
leaves20.280.70790.2989link
orchids19.740.64880.2868link
room33.550.96370.1209link
trex27.870.91360.1827link

Tanks and Temples

ScenePSNRSSIMLPIPSCheckpoint
M6020.090.71230.3991link
Playground24.270.74010.3551link
Train19.740.67050.3813link
Truck24.140.80040.2963link

LF

ScenePSNRSSIMLPIPSCheckpoint
Africa29.580.91680.1984link
Basket21.190.87140.2943link
Ship30.160.89920.2329link
Statue34.900.95210.1505link
Torch25.860.85710.2707link

360-v2

ScenePSNRSSIMLPIPSCheckpoint
bicycle22.860.54340.4553link
bonsai32.970.93750.1439link
counter29.290.87090.2118link
garden26.010.74260.2703link
kitchen31.990.92470.1090link
room32.680.92580.1797link
stump25.280.65220.4153link