Tetra-NeRF

Tetra-NeRF is a method that represents the scene as tetrahedral mesh obtained using Delaunay tetrahedralization. The input point cloud has to be provided (for COLMAP datasets the point cloud is automatically extracted). This is the official implementation from the paper.

Web: https://jkulhanek.com/tetra-nerf
Paper: Tetra-NeRF: Representing Neural Radiance Fields Using Tetrahedra
Authors: Jonas Kulhanek, Torsten Sattler

Blender

Blender (nerf-synthetic) is a synthetic dataset used to benchmark NeRF methods. It consists of 8 scenes of an object placed on a white background. Cameras are placed on a semi-sphere around the object. Scenes are licensed under various CC licenses.

Scene PSNR SSIM LPIPS (VGG) Time GPU mem.
lego 33.93
Paper's PSNR: 34.75
0.972
Paper's SSIM: 0.987
0.036 7h 8m 12s 25.21 GB
drums 24.99
Paper's PSNR: 25.01
0.927
Paper's SSIM: 0.947
0.087 8h 34m 5s 36.26 GB
ficus 32.37
Paper's PSNR: 33.31
0.977
Paper's SSIM: 0.989
0.032 4h 19m 24s 21.84 GB
hotdog 35.80
Paper's PSNR: 36.16
0.978
Paper's SSIM: 0.989
0.040 6h 41m 21s 23.06 GB
materials 28.75
Paper's PSNR: 29.3
0.941
Paper's SSIM: 0.968
0.076 6h 23m 31s 37.98 GB
mic 34.54
Paper's PSNR: 35.49
0.987
Paper's SSIM: 0.993
0.022 6h 51m 58s 31.04 GB
ship 31.06
Paper's PSNR: 31.13
0.896
Paper's SSIM: 0.994
0.129 9h 4m 53s 27.19 GB
chair 34.17
Paper's PSNR: 35.05
0.977
Paper's SSIM: 0.99
0.029 6h 3m 14s 34.01 GB
Average 31.95
Paper's PSNR: 32.52
0.957
Paper's SSIM: 0.982
0.056 6h 53m 20s 29.57 GB