Skip to content

XuPaya/Stoch3DGS

Repository files navigation

Stochastic Ray Tracing for the Reconstruction of 3D Gaussian Splatting

Peiyu Xu, Xin Sun, Krishna Mullia, Raymond Fei, Iliyan Georgiev, Shuang Zhao,

CVPR2026


Project Page | Paper

This repository provides an implementations of Stochastic Ray Tracing for the Reconstruction of 3D Gaussian Splatting (CVPR2026). We build on top of the official implementation of 3DGRUT and provide an accelerated algorithm. We also provide scripts to reproduce results reported in our paper on the MipNeRF360 dataset.

1. Dependencies, Installation and Usage

Please refer to the original 3DGRUT repository for installation instructions and usage details. We provide additional scripts to train and evaluate our method on the MipNeRF360 dataset.

2. Baselines and Methods

We provide the following algorithms in our codebase:

  • 3DGRT: The original 3D Gaussian Ray Tracing algorithm proposed by Moenne-Loccoz et al (https://arxiv.org/abs/2407.07090), which uses a deterministic sorted ray tracing algorithm.

  • Stochastic GRT: The stochastic ray tracing algorithm proposed in our paper, which uses a stochastic sampling strategy for both the forward rendering and the backward gradient computation.

  • Quasi-Stochastic GRT: A variant of the stochastic ray tracing algorithm that uses sorted ray tracing for the forward pass and the stochastic strategy for the backward pass.

3. Evaluations on MipNeRF360

We provide scripts to reproduce results reported in our publication, and report detailed per-scene benchmark results.

Results Produced on RTX 5090

Evaluations on MipNeRF360

3DGRT (Baseline) Results

bash ./benchmark/mipnerf360.sh apps/colmap_3dgrt.yaml
bash ./benchmark/mipnerf360_render.sh results/mipnerf360
PSNR SSIM Train (s)
Bicycle 24.78 0.746 1738
Bonsai 31.59 0.938 3519
Counter 28.46 0.903 3294
Garden 26.95 0.849 1923
Kitchen 29.61 0.916 5004
Room 30.39 0.909 2629
Stump 26.46 0.771 1577
Average 28.32 0.862 2812

Stochastic GRT Results

bash ./benchmark/mipnerf360_stoch.sh apps/colmap_3dgrt_stoch.yaml
bash ./benchmark/mipnerf360_render.sh results/mipnerf360_stoch/
PSNR SSIM Train (s)
Bicycle 24.51 0.725 1352
Bonsai 30.95 0.930 2335
Counter 28.20 0.893 2308
Garden 26.39 0.833 1409
Kitchen 29.45 0.903 2798
Room 29.89 0.905 1586
Stump 26.33 0.768 1258
Average 27.96 0.851 1864

Quasi-Stochastic GRT Results

bash ./benchmark/mipnerf360_quasistoch.sh apps/colmap_3dgrt_quasistoch.yaml
bash ./benchmark/mipnerf360_render.sh results/mipnerf360_quasistoch/
PSNR SSIM Train (s)
Bicycle 24.61 0.727 1029
Bonsai 31.27 0.934 1679
Counter 28.21 0.897 1465
Garden 26.73 0.840 1105
Kitchen 29.58 0.909 1846
Room 30.35 0.909 1184
Stump 26.49 0.774 946
Average 28.18 0.856 1322
Results Produced on RTX 5880 Ada Generation

3DGRT (Baseline) Results

bash ./benchmark/mipnerf360.sh apps/colmap_3dgrt.yaml
bash ./benchmark/mipnerf360_render.sh results/mipnerf360
PSNR SSIM Train (s)
Bicycle 24.81 0.744 2607
Bonsai 31.78 0.937 4971
Counter 28.33 0.900 4643
Garden 26.77 0.847 2955
Kitchen 30.39 0.920 7679
Room 30.78 0.914 3708
Stump 26.40 0.771 2272
Average 28.47 0.862 4119

Stochastic GRT Results

bash ./benchmark/mipnerf360_stoch.sh apps/colmap_3dgrt_stoch.yaml
bash ./benchmark/mipnerf360_render.sh results/mipnerf360_stoch/
PSNR SSIM Train (s)
Bicycle 24.58 0.724 2005
Bonsai 31.34 0.931 3379
Counter 28.41 0.894 3445
Garden 26.22 0.826 2135
Kitchen 29.76 0.903 4164
Room 30.41 0.908 2293
Stump 26.29 0.767 1830
Average 28.14 0.850 2750

Quasi-Stochastic GRT Results

bash ./benchmark/mipnerf360_quasistoch.sh apps/colmap_3dgrt_quasistoch.yaml
bash ./benchmark/mipnerf360_render.sh results/mipnerf360_quasistoch/
PSNR SSIM Train (s)
Bicycle 24.63 0.727 1570
Bonsai 31.59 0.936 2407
Counter 28.35 0.899 2223
Garden 26.74 0.840 1796
Kitchen 30.00 0.911 2883
Room 30.39 0.909 1765
Stump 26.50 0.774 1434
Average 28.31 0.857 2011

About

No description, website, or topics provided.

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors