KITTI-360

Submit

Method

Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields [mip-NeRF]
https://github.com/google/mipnerf

Submitted on 6 Feb. 2022 15:41 by
Yiyi Liao (MPI)

Running time:		10 s
Environment:		1 core @ 2.5 Ghz (Python)

Method Description:

mip-NeRF extends NeRF to represent the scene at a continuously-valued
scale. By efficiently rendering anti-aliased conical frustums instead of
rays, mip-NeRF reduces objectionable aliasing artifacts and significantly
improves NeRF's ability to represent fine details.

Parameters:

We train mip-NeRF following the official implementation and hyper-
parameters.

Latex Bibtex:

@article{barron2021mipnerf,
title={Mip-NeRF: A Multiscale Representation
for Anti-Aliasing Neural Radiance Fields},
author={Jonathan T. Barron and Ben Mildenhall and
Matthew Tancik and Peter Hedman and
Ricardo Martin-Brualla and Pratul P. Srinivasan},
journal={ICCV},
year={2021}
}

Detailed Results

This page provides detailed results for the method(s) selected. For the first 10 test images, we display the synthesized image and an error image. The error image visualizes the SSIM at every pixel. As the range of SSIM is within [-1,1] with 1 indicating the best performance and -1 indicating the worst, we visualize 1 - (1 + SSIM) / 2 such that bright region means large error and dark means low error.

Test Set Average

PSNR	SSIM	LPIPS
21.54	0.778	0.365

Error