Instant NGP

Instant-NGP is a method that uses hash-grid and a shallow MLP to accelerate training and rendering. This method trains very fast (~6 min) and renders also fast ~3 FPS.

Web: https://nvlabs.github.io/instant-ngp/
Paper: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding
Authors: Thomas Müller, Alex Evans, Christoph Schied, Alexander Keller

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 24.78 0.654 0.346 3m 57s 6.86 GB
bicycle 23.00 0.526 0.489 3m 37s 6.29 GB
flowers 20.54 0.462 0.478 4m 2s 6.26 GB
treehill 22.36 0.528 0.526 3m 40s 6.51 GB
stump 23.84 0.590 0.439 3m 52s 6.07 GB
kitchen 29.02 0.844 0.255 4m 19s 9.52 GB
bonsai 30.30 0.890 0.295 3m 33s 9.61 GB
counter 26.56 0.812 0.373 4m 10s 9.25 GB
room 29.16 0.850 0.383 3m 59s 9.73 GB
Average 25.51 0.684 0.398 3m 54s 7.79 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 35.65
Paper's PSNR: 36.39

Instant-NGP trained and evaluated on black background instead of white.

 0.981 0.020 2m 41s 2.59 GB
drums 24.57
Paper's PSNR: 26.02

Instant-NGP trained and evaluated on black background instead of white.

 0.930 0.109 2m 6s 2.58 GB
ficus 30.29
Paper's PSNR: 33.51

Instant-NGP trained and evaluated on black background instead of white.

 0.972 0.031 1m 58s 2.58 GB
hotdog 37.02
Paper's PSNR: 37.4

Instant-NGP trained and evaluated on black background instead of white.

 0.982 0.037 2m 33s 2.58 GB
materials 28.96
Paper's PSNR: 29.78

Instant-NGP trained and evaluated on black background instead of white.

 0.944 0.069 2m 11s 2.58 GB
mic 35.41
Paper's PSNR: 36.22

Instant-NGP trained and evaluated on black background instead of white.

 0.989 0.016 2m 25s 2.47 GB
ship 30.61
Paper's PSNR: 31.1

Instant-NGP trained and evaluated on black background instead of white.

 0.892 0.136 2m 24s 2.58 GB
chair 35.07
Paper's PSNR: 35.0

Instant-NGP trained and evaluated on black background instead of white.

 0.984 0.023 2m 49s 2.57 GB
Average 32.20
Paper's PSNR: 33.18

Instant-NGP trained and evaluated on black background instead of white.

 0.959 0.055 2m 23s 2.57 GB

Tanks and Temples

Tanks and Temples is a benchmark for image-based 3D reconstruction. The benchmark sequences were acquired outside the lab, in realistic conditions. Ground-truth data was captured using an industrial laser scanner. The benchmark includes both outdoor scenes and indoor environments. The dataset is split into three subsets: training, intermediate, and advanced.

Scene PSNR SSIM LPIPS Time GPU mem.
auditorium 20.67 0.761 0.429 4m 31s 3.92 GB
ballroom 21.62 0.652 0.352 4m 5s 4.02 GB
courtroom 19.44 0.640 0.448 4m 35s 3.93 GB
museum 15.19 0.471 0.606 5m 53s 3.93 GB
palace 19.09 0.668 0.440 4m 46s 4.64 GB
temple 17.84 0.689 0.424 4m 40s 3.94 GB
family 22.59 0.761 0.235 3m 42s 3.42 GB
francis 24.38 0.824 0.265 4m 10s 3.94 GB
horse 21.82 0.784 0.225 3m 47s 3.42 GB
lighthouse 21.65 0.765 0.281 4m 35s 4.02 GB
m60 25.82 0.832 0.202 4m 34s 4.04 GB
panther 26.81 0.844 0.208 4m 29s 4.04 GB
playground 23.33 0.696 0.344 4m 14s 3.97 GB
train 20.01 0.658 0.334 4m 39s 3.94 GB
barn 25.90 0.772 0.271 4m 41s 4.31 GB
caterpillar 21.72 0.633 0.360 4m 30s 4.22 GB
church 19.92 0.650 0.419 4m 21s 4.64 GB
courthouse 20.80 0.681 0.414 4m 55s 6.72 GB
ignatius 19.40 0.613 0.343 3m 58s 3.81 GB
meetingroom 23.24 0.783 0.326 3m 59s 4.18 GB
truck 22.85 0.770 0.216 4m 20s 3.77 GB
Average 21.62 0.712 0.340 4m 27s 4.13 GB