The advent of 3D Gaussian Splatting (3DGS) has recently brought about a revolution in the field of neural rendering, facilitating high-quality renderings at real-time speed. However, 3DGS heavily depends on the initialized point cloud produced by Structure-from-Motion (SfM) techniques. When tackling with large-scale scenes that unavoidably contain texture-less surfaces, the SfM techniques always fail to produce enough points in these surfaces and cannot provide good initialization for 3DGS. As a result, 3DGS suffers from difficult optimization and low-quality renderings. In this paper, inspired by classical multi-view stereo (MVS) techniques, we propose GaussianPro, a novel method that applies a progressive propagation strategy to guide the densification of the 3D Gaussians. Compared to the simple split and clone strategies used in 3DGS, our method leverages the priors of the existing reconstructed geometries of the scene and patch matching techniques to produce new Gaussians with accurate positions and orientations. Experiments on both large-scale and small-scale scenes validate the effectiveness of our method, where our method significantly surpasses 3DGS on the Waymo dataset, exhibiting an improvement of 1.15dB in terms of PSNR.
- [✔] Code pre-release -- Beta version.
- [✔] Demo Scenes.
- CUDA acceleration -- Official version.
- Support for unordered set of images.
Ubuntu 20.04.1 LTS, GeForce 3090, CUDA 11.3 (tested) / 11.7 (tested), C++17
git clone https://github.com/kcheng1021/GaussianPro.git --recursive
conda env create --file environment.yml
cd ./submodules/Propagation
# The gpu compute architecture is specified as sm_86 in CMakeLists.
# Please replace it with a version that is suitable for your GPU.
cmake .
make
wget https://drive.google.com/file/d/1DXQRBcUIrnIC33WNq8pVLKZ_W1VwON3k/view?usp=sharing
Besides the public datasets, we also test GaussianPro from random selected Youtube videos and find consistent improvement. The processed data is provided below.
#youtube01: Park.
wget https://drive.google.com/file/d/1iHYTnI76Zx9VTKbMu1zUE7gVKP4UpBan/view?usp=sharing
#youtube02: Church
wget https://drive.google.com/file/d/1i2ReAJYkeLHBBbs_8Zn560Tke2F8yR1X/view?usp=sharing
#youtube03: The forbidden city.
wget https://drive.google.com/file/d/1PZ_917Oq0Y45_5dJ504RxRmpUnewYmyn/view?usp=sharing
#youtube04: Eiffel tower.
wget https://drive.google.com/file/d/1JoYyfAu3RNnj12C2gPvfljHLUKlUsSr1/view?usp=sharing
# Run the 3DGS, we modify the defaulting parameters in 3DGS to better learn large scenes. The description of parameters in GaussianPro will come out later.
# To run the Waymo scenes (3DGS and GaussianPro)
bash scripts/waymo.sh
# Run the Youtube scenes above
bash scripts/demo.sh
To ensure the reproducibility, we present a reference of the results in the provided demo scenes based on the current code.
| Waymo-1002751 | Youtube-01 | Youtube-02 | Youtube-03 | Youtube-04 | |
|---|---|---|---|---|---|
| 3DGS | 35.22,0.950,0.234 | 34.40,0.964,0.092 | 34.67,0.954,0.072 | 37.81,0.971,0.081 | 33.05,0.950,0.079 |
| GaussianPro | 35.97,0.959,0.207 | 35.29,0.969,0.076 | 35.08,0.959,0.064 | 38.27,0.974,0.072 | 33.66,0.956,0.072 |
If you want to try your scenes, make sure your images are sorted in the time order, i.e. video data. The current version does not support unordered image sets, but it will be updated in the next version. Then you can try the commands in demo.sh to run your own scenes.
Please ensure that your neighboring images have sufficient overlap.
For non-commercial use, this code is released under the LICENSE. For commercial use, please contact Xuejin Chen.
This project largely references 3D Gaussian Splatting and ACMH/ACMM. Thanks for their amazing works!