KITTI-360

Submit

Method

Panoptic Neural Fields: A Semantic Object-Aware Neural Scene Representation [PNF]

Submitted on 19 Aug. 2022 00:42 by
Abhijit Kundu (Google)

Running time:		15 s
Environment:		GPU @ 2.5 Ghz (Python)

Method Description:

Rendered color images from our Panoptic Neural Fields
(PNF) model. The same model is also used to render
semantic images (see our results on novel view semantics
task).

Parameters:

Stuff MLP: 10 layers, 256 neurons
Object MLPs: 4 layers, 128 neurons

Latex Bibtex:

@inproceedings{pnf2022,
title={Panoptic Neural Fields: A Semantic Object-Aware
Neural Scene Representation},
author={Abhijit Kundu and Kyle Genova and Xiaoqi Yin
and Alireza Fathi and Caroline Pantofaru and Leonidas
Guibas and Andrea Tagliasacchi and Frank Dellaert and
Thomas Funkhouser},
booktitle={CVPR},
year={2022}
}

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
22.07	0.820	0.221

Error