H3DGS

H3DGS extends 3DGS with LOD rendering strategy based on hierarchical representation of the scene. For large scenes it splits it into chunks, optimize each separatedly and merge them into single model.

Web: https://repo-sam.inria.fr/fungraph/hierarchical-3d-gaussians/
Paper: A Hierarchical 3D Gaussian Representation for Real-Time Rendering of Very Large Datasets
Authors: Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, George Drettakis

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 26.44 0.820 0.191 52m 35s 8.97 GB
bicycle 24.03 0.717 0.284 57m 19s 11.46 GB
flowers 20.53 0.571 0.344 59m 11s 11.99 GB
treehill 22.09 0.598 0.384 54m 52s 10.48 GB
stump 25.37 0.728 0.287 50m 50s 9.07 GB
kitchen 31.06 0.918 0.163 53m 54s 6.70 GB
bonsai 31.86 0.937 0.243 56m 6s 6.97 GB
counter 29.07 0.906 0.243 58m 13s 7.64 GB
room 31.89 0.918 0.284 56m 22s 8.41 GB
Average 26.93 0.790 0.269 55m 29s 9.08 GB

Hierarchical 3DGS

Hierarchical 3DGS is a dataset released with H3DGS paper. We implement the two public single-chunks scenes (SmallCity, Campus) used for evaluation. To collect the dataset, authors used a bicycle helmet on which they mounted 6 GoPro HERO6 Black cameras (5 for the Campus scene). They collected SmallCity and BigCity captures on a bicycle, riding at around 6–7km/h, while Campus was captured on foot wearing the helmet. Poses were estimated using COLMAP with custom parameters and hierarchical mapper. Additinal per-chunk bundle adjustment was performed. It is recommended to use exposure modeling with this dataset

Scene PSNR SSIM LPIPS (VGG) Time GPU mem.
SmallCity 26.42
Paper's PSNR: 26.29

Results in the paper were evaluated using different tau (0, 3, 6, 15), where tau=0 is the slowest, but highest quality. We choose tau=6 - consistent with most experiments in the paper, providing a good trade-off between quality and speed.

 0.807
Paper's SSIM: 0.81

Results in the paper were evaluated using different tau (0, 3, 6, 15), where tau=0 is the slowest, but highest quality. We choose tau=6 - consistent with most experiments in the paper, providing a good trade-off between quality and speed.

 0.331 1h 16m 1s 15.85 GB
Campus 24.63
Paper's PSNR: 24.5

Results in the paper were evaluated using different tau (0, 3, 6, 15), where tau=0 is the slowest, but highest quality. We choose tau=6 - consistent with most experiments in the paper, providing a good trade-off between quality and speed.

 0.798
Paper's SSIM: 0.801

Results in the paper were evaluated using different tau (0, 3, 6, 15), where tau=0 is the slowest, but highest quality. We choose tau=6 - consistent with most experiments in the paper, providing a good trade-off between quality and speed.

 0.396 1h 30m 7s 14.04 GB
Average 25.53
Paper's PSNR: 25.39

Results in the paper were evaluated using different tau (0, 3, 6, 15), where tau=0 is the slowest, but highest quality. We choose tau=6 - consistent with most experiments in the paper, providing a good trade-off between quality and speed.

 0.803
Paper's SSIM: 0.806

Results in the paper were evaluated using different tau (0, 3, 6, 15), where tau=0 is the slowest, but highest quality. We choose tau=6 - consistent with most experiments in the paper, providing a good trade-off between quality and speed.

 0.364 1h 23m 4s 14.94 GB