Mip-NeRF 360

Official Mip-NeRF 360 implementation addapted to handle different camera distortion/intrinsic parameters. It was designed for unbounded object-centric 360-degree capture and handles anti-aliasing well. It is, however slower to train and render compared to other approaches.

Web: https://jonbarron.info/mipnerf360/
Paper: Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields
Authors: Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, Peter Hedman

Mip-NeRF 360

Mip-NeRF 360 is a collection of four indoor and five outdoor object-centric scenes. The camera trajectory is an orbit around the object with fixed elevation and radius. The test set takes each n-th frame of the trajectory as test views.

Scene PSNR SSIM LPIPS (VGG) Time GPU mem.
garden 27.00
Paper's PSNR: 26.98
0.814
Paper's SSIM: 0.813
0.189
Paper's LPIPS (VGG): 0.17
30h 13m 41s 33.06 GB
bicycle 24.28
Paper's PSNR: 24.37
0.686
Paper's SSIM: 0.685
0.329
Paper's LPIPS (VGG): 0.301
30h 14m 19s 32.86 GB
flowers 21.74
Paper's PSNR: 21.73
0.583
Paper's SSIM: 0.583
0.371
Paper's LPIPS (VGG): 0.344
30h 13m 53s 32.78 GB
treehill 22.89
Paper's PSNR: 22.87
0.633
Paper's SSIM: 0.632
0.378
Paper's LPIPS (VGG): 0.339
30h 13m 9s 32.98 GB
stump 26.50
Paper's PSNR: 26.4
0.749
Paper's SSIM: 0.744
0.298
Paper's LPIPS (VGG): 0.261
30h 15m 5s 32.88 GB
kitchen 32.10
Paper's PSNR: 32.23
0.919
Paper's SSIM: 0.92
0.155
Paper's LPIPS (VGG): 0.127
30h 15m 15s 34.48 GB
bonsai 33.48
Paper's PSNR: 33.46
0.940
Paper's SSIM: 0.941
0.211
Paper's LPIPS (VGG): 0.176
30h 13m 6s 34.48 GB
counter 29.51
Paper's PSNR: 29.55
0.894
Paper's SSIM: 0.894
0.252
Paper's LPIPS (VGG): 0.204
30h 14m 4s 34.47 GB
room 31.64
Paper's PSNR: 31.63
0.912
Paper's SSIM: 0.913
0.267
Paper's LPIPS (VGG): 0.211
30h 18m 49s 34.47 GB
Average 27.68
Paper's PSNR: 27.69
0.792
Paper's SSIM: 0.792
0.272
Paper's LPIPS (VGG): 0.237
30h 14m 36s 33.61 GB

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 33.20 0.975 0.028 2h 43m 18s 126.91 GB
drums 24.36 0.923 0.083 8h 51m 32s 30.02 GB
ficus 26.66 0.952 0.048 2h 43m 27s 126.91 GB
hotdog 36.44 0.979 0.039 2h 43m 11s 126.91 GB
materials 27.91 0.944 0.067 2h 43m 22s 126.91 GB
mic 31.50 0.984 0.021 2h 43m 9s 126.91 GB
ship 28.66 0.875 0.164 2h 45m 55s 126.91 GB
chair 34.01 0.977 0.032 2h 43m 17s 126.91 GB
Average 30.34 0.951 0.060 3h 29m 39s 114.80 GB

LLFF

LLFF is a dataset of forward-facing scenes with a small variation in camera pose. NeRF methods usually use NDC-space parametrization for the scene representation.

Scene PSNR SSIM LPIPS (VGG) Time GPU mem.
Fern 24.59 0.820 0.210 7h 29m 8s 129.06 GB
Flower 27.56 0.867 0.140 14h 2m 26s 64.85 GB
Fortress 31.34 0.900 0.117 14h 2m 6s 64.85 GB
Horns 28.51 0.909 0.124 7h 26m 41s 129.06 GB
Leaves 19.84 0.721 0.231 7h 26m 52s 129.06 GB
Orchids 19.51 0.660 0.246 7h 27m 12s 129.06 GB
Room 33.49 0.965 0.115 7h 26m 22s 129.06 GB
Trex 27.86 0.927 0.158 14h 2m 15s 64.85 GB
Average 26.59 0.846 0.168 9h 55m 23s 104.98 GB