3DGS-MCMC

3DGS-MCMC reinterprets 3D Gaussian Splatting as MCMC sampling, introducing noise-based updates and removing heuristic cloning strategies, leading to improved rendering quality, efficient Gaussian use, and robustness to initialization. In NerfBaselines, we fixed bug with cx,cy, added appearance embedding optimization, and added support for sampling masks and web demos.

Web: https://ubc-vision.github.io/3dgs-mcmc/
Paper: 3D Gaussian Splatting as Markov Chain Monte Carlo
Authors: Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, Kwang Moo Yi

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 27.81
Paper's PSNR: 28.16

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.

 0.877
Paper's SSIM: 0.89

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.

 0.109 48m 5s 35.23 GB
bicycle 25.69
Paper's PSNR: 26.15

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.

 0.799
Paper's SSIM: 0.81

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.

 0.194 49m 28s 38.87 GB
flowers 22.05 0.644 0.316 37m 29s 38.42 GB
treehill 22.97 0.658 0.315 37m 33s 38.15 GB
stump 27.38
Paper's PSNR: 27.8

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.

 0.811
Paper's SSIM: 0.82

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.

 0.200 40m 48s 34.43 GB
kitchen 31.91
Paper's PSNR: 32.27

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.

 0.933
Paper's SSIM: 0.94

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.

 0.148 42m 32s 20.35 GB
bonsai 32.66
Paper's PSNR: 32.88

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.

 0.947
Paper's SSIM: 0.95

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.

 0.236 35m 14s 14.88 GB
counter 29.32
Paper's PSNR: 29.51

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: 0.92

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.

 0.238 38m 44s 20.42 GB
room 32.05
Paper's PSNR: 32.48

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.

 0.927
Paper's SSIM: 0.94

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.

 0.262 40m 43s 18.96 GB
Average 27.98 0.835 0.224 41m 11s 28.86 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 35.75
Paper's PSNR: 36.01

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.984
Paper's SSIM: 0.98

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.019 8m 51s 4.50 GB
drums 26.22
Paper's PSNR: 26.29

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.954
Paper's SSIM: 0.95

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.045 9m 3s 4.51 GB
ficus 34.88
Paper's PSNR: 35.07

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.988
Paper's SSIM: 0.99

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.013 8m 2s 4.54 GB
hotdog 37.34
Paper's PSNR: 37.82

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.986
Paper's SSIM: 0.99

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.025 9m 55s 4.50 GB
materials 30.56
Paper's PSNR: 30.59

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.964
Paper's SSIM: 0.96

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.039 9m 8s 4.35 GB
mic 37.28
Paper's PSNR: 37.29

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.994
Paper's SSIM: 0.99

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.007 9m 36s 4.47 GB
ship 30.76
Paper's PSNR: 30.82

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.907
Paper's SSIM: 0.91

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.124 10m 38s 4.87 GB
chair 36.32
Paper's PSNR: 36.51

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.989
Paper's SSIM: 0.99

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.016 7m 54s 4.30 GB
Average 33.64
Paper's PSNR: 33.80

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.971
Paper's SSIM: 0.970

Exact hyperparameters for Blender dataset are not provided in the released source code. The default parameters were used in NerfBaselines likely leading to worse results.

 0.036 9m 8s 4.51 GB