2D Gaussian Splatting

2DGS adopts 2D oriented disks as surface elements and allows high-quality rendering with Gaussian Splatting. In NerfBaselines, we fixed bug with cx,cy, added appearance embedding optimization, and added support for sampling masks.

Web: https://surfsplatting.github.io/
Paper: 2D Gaussian Splatting for Geometrically Accurate Radiance Fields
Authors: Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, Shenghua Gao

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 26.69
Paper's PSNR: 26.95

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.843
Paper's SSIM: 0.852

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.166 27m 56s 8.58 GB
bicycle 24.77
Paper's PSNR: 24.87

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.733
Paper's SSIM: 0.752

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.302 33m 18s 11.96 GB
flowers 21.14
Paper's PSNR: 21.15

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.572
Paper's SSIM: 0.588

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.403 26m 31s 7.64 GB
treehill 22.36
Paper's PSNR: 22.27

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.616
Paper's SSIM: 0.627

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.433 28m 44s 8.81 GB
stump 26.20
Paper's PSNR: 26.47

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.758
Paper's SSIM: 0.765

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.299 29m 12s 28.56 GB
kitchen 30.41
Paper's PSNR: 30.5

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.916
Paper's SSIM: 0.919

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.179 35m 31s 15.79 GB
bonsai 31.30
Paper's PSNR: 31.52

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.931
Paper's SSIM: 0.933

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.280 33m 16s 15.39 GB
counter 28.10
Paper's PSNR: 28.55

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.892
Paper's SSIM: 0.9

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.292 32m 38s 9.95 GB
room 30.37
Paper's PSNR: 31.06

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.906
Paper's SSIM: 0.912

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.317 33m 20s 11.74 GB
Average 26.81
Paper's PSNR: 27.04

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.796
Paper's SSIM: 0.805

Authors evaluated on larger images which were downscaled to the target size (avoiding JPEG compression artifacts) instead of using the official provided downscaled images. As mentioned in the 3DGS paper, this increases results slightly ~0.5 dB PSNR.

 0.297 31m 10s 13.16 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 22.86 0.835 0.304 13m 45s 6.55 GB
ballroom 23.40 0.808 0.120 19m 53s 6.12 GB
courtroom 18.92 0.660 0.422 15m 12s 9.39 GB
museum 21.11 0.764 0.170 26m 10s 8.09 GB
palace 16.91 0.668 0.512 13m 40s 5.50 GB
temple 19.74 0.763 0.305 13m 46s 4.06 GB
family 24.14 0.854 0.111 16m 16s 4.32 GB
francis 26.50 0.884 0.202 14m 56s 9.51 GB
horse 23.21 0.867 0.122 15m 15s 7.43 GB
lighthouse 18.00 0.707 0.438 13m 3s 8.47 GB
m60 22.83 0.825 0.249 15m 2s 8.98 GB
panther 21.70 0.771 0.358 13m 40s 6.39 GB
playground 21.45 0.745 0.316 16m 51s 7.16 GB
train 16.41 0.583 0.529 11m 43s 4.77 GB
barn 25.42
Paper's PSNR: 28.79

2DGS used different data pre-processing and train/test split for Tanks and Temples. It sets specific hyperparameters for each scene which may not be suitable with the public Tanks and Temples released by NerfBaselines. The results are not directly comparable and a hyperparameter tuning is needed to improve the results.

 0.816 0.213 13m 36s 6.29 GB
caterpillar 22.24
Paper's PSNR: 24.23

2DGS used different data pre-processing and train/test split for Tanks and Temples. It sets specific hyperparameters for each scene which may not be suitable with the public Tanks and Temples released by NerfBaselines. The results are not directly comparable and a hyperparameter tuning is needed to improve the results.

 0.761 0.227 14m 3s 7.15 GB
church 18.50 0.658 0.443 16m 11s 14.63 GB
courthouse 19.48
Paper's PSNR: 23.51

2DGS used different data pre-processing and train/test split for Tanks and Temples. It sets specific hyperparameters for each scene which may not be suitable with the public Tanks and Temples released by NerfBaselines. The results are not directly comparable and a hyperparameter tuning is needed to improve the results.

 0.707 0.398 13m 18s 8.98 GB
ignatius 21.34
Paper's PSNR: 23.82

2DGS used different data pre-processing and train/test split for Tanks and Temples. It sets specific hyperparameters for each scene which may not be suitable with the public Tanks and Temples released by NerfBaselines. The results are not directly comparable and a hyperparameter tuning is needed to improve the results.

 0.762 0.176 18m 57s 5.54 GB
meetingroom 24.26
Paper's PSNR: 26.15

2DGS used different data pre-processing and train/test split for Tanks and Temples. It sets specific hyperparameters for each scene which may not be suitable with the public Tanks and Temples released by NerfBaselines. The results are not directly comparable and a hyperparameter tuning is needed to improve the results.

 0.849 0.168 20m 5s 7.41 GB
truck 23.83
Paper's PSNR: 26.85

2DGS used different data pre-processing and train/test split for Tanks and Temples. It sets specific hyperparameters for each scene which may not be suitable with the public Tanks and Temples released by NerfBaselines. The results are not directly comparable and a hyperparameter tuning is needed to improve the results.

 0.843 0.123 15m 54s 4.99 GB
Average 21.54 0.768 0.281 15m 47s 7.23 GB