Octree-GS

An LOD-structured 3D Gaussian approach supporting level-of-detail decomposition for scene representation that contributes to the final rendering results.

Web: https://city-super.github.io/octree-gs/
Paper: Octree-GS: Towards Consistent Real-time Rendering with LOD-Structured 3D Gaussians
Authors: Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, Bo Dai

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 27.72 0.869 0.124 30m 15s 7.65 GB
bicycle 25.03 0.759 0.251 24m 55s 7.50 GB
flowers 21.43 0.600 0.374 22m 57s 6.53 GB
treehill 23.01 0.643 0.360 24m 41s 6.44 GB
stump 26.48 0.763 0.276 20m 38s 5.79 GB
kitchen 30.52 0.917 0.172 35m 35s 10.41 GB
bonsai 30.88 0.923 0.280 27m 5s 10.86 GB
counter 29.49 0.907 0.262 34m 58s 9.85 GB
room 32.01 0.922 0.275 31m 21s 11.66 GB
Average 27.40 0.811 0.264 28m 3s 8.52 GB

Zip-NeRF

ZipNeRF is a dataset with four large scenes: Berlin, Alameda, London, and NYC, (1000-2000 photos each) captured using fisheye cameras. This implementation uses undistorted images which are provided with the dataset and the downsampled resolutions are between 1392 × 793 and 2000 × 1140 depending on scene. It is recommended to use exposure modeling with this dataset if available.

Scene PSNR SSIM LPIPS (VGG) Time GPU mem.
Alameda 22.79 0.730 0.448 44m 25s 28.68 GB
Berlin 13.64 0.669 0.640 1h 6m 40s 34.85 GB
London 25.76 0.807 0.433 33m 44s 27.35 GB
NYC 27.13 0.841 0.372 34m 55s 19.63 GB
Average 22.33 0.762 0.473 44m 56s 27.63 GB