Link Search Menu Expand Document (external link)

Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields

[Project Page] [Paper] [Code]

Authors

  • Jonathan T. Barron
  • Ben Mildenhall
  • Matthew Tancik
  • Peter Hedman
  • Ricardo Martin-Brualla
  • Pratul P. Srinivasan

Table of contents


What’s Mip-NeRF?

Mip-NeRF

The NeRF-variants produce view-dependent color with a single novel view-point ray per pixel, which mainly generate blurry images. Multi-scale representation, ray casting with cone, not a single ray, prevent aliasing of generated images. Specifically, Mip-NeRF embeds the 3D position and volume of the conical frustum in the encoding process, which is conceptually equal to “mip-map”(pre-defined filter model in graphics fields.), and proposed representation overcome NeRF’s blurry result on the multi-scale dataset. Mip-NeRF is also 22x faster than the original NeRF.

mipnerf

Scores

Blender

ScenePSNRSSIMLPIPSCheckpoint
chair35.200.98090.0283link
drums25.530.93180.0795link
ficus33.230.97980.0258link
hotdog37.440.98140.0345link
lego35.800.97810.0265link
materials30.580.95820.0526link
mic36.410.99090.0126link
ship30.520.88460.1544link

MS-Blender

ScenePSNRSSIMLPIPSCheckpoint
chair37.360.98780.0161link
drums27.120.94580.0561link
ficus33.000.98310.0185link
hotdog39.360.98790.0178link
lego35.710.98400.0174link
materials32.630.97670.0260link
mic37.930.99260.0101link
ship33.240.92350.0818link

LLFF

ScenePSNRSSIMLPIPSCheckpoint
fern24.920.79570.2793link
flower27.800.84300.1937link
fortress31.730.89330.1492link
horns27.790.85710.2236link
leaves20.940.70380.3060link
orchids20.270.64020.3204link
room33.240.95520.1533link
trex27.690.90250.2220link

Tanks and Temples

ScenePSNRSSIMLPIPSCheckpoint
M6018.410.64350.4935link
Playground21.830.66410.4948link
Train17.870.57490.4964link
Truck21.710.69030.4562link

LF

ScenePSNRSSIMLPIPSCheckpoint
Africa28.650.86760.3050link
Basket21.980.81580.4056link
Ship26.430.79680.3650link
Statue29.860.88180.2973link
Torch23.290.75130.4012link

Shiny Blender

ScenePSNRSSIMLPIPSCheckpoint
ball27.290.93870.2006link
car26.720.92260.0690link
coffee30.830.96580.1297link
helmet27.790.94170.1391link
teapot45.500.99640.0122link
toaster22.520.88940.1645link

NeRF-360-v2

ScenePSNRSSIMLPIPSCheckpoint
bicycle21.720.42460.5880link
bonsai29.120.84430.3164link
counter26.770.76220.3666link
garden23.710.55440.4334link
kitchen27.980.80250.2570link
room30.230.85570.3342link
stump22.740.47220.5716link