PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction

Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, Guofeng Zhang

Project Page | arXiv

We present a Planar-based Gaussian Splatting Reconstruction representation for efficient and high-fidelity surface reconstruction from multi-view RGB images without any geometric prior (depth or normal from pre-trained model).

Updates

• [2024.07.18]: We fine-tuned the hyperparameters based on the original paper. The Chamfer Distance on the DTU dataset decreased to 0.47.

The Chamfer Distance↓ on the DTU dataset

	24	37	40	55	63	65	69	83	97	105	106	110	114	118	122	Mean	Time
PGSR(Paper)	0.34	0.58	0.29	0.29	0.78	0.58	0.54	1.01	0.73	0.51	0.49	0.69	0.31	0.37	0.38	0.53	0.6h
PGSR(Code_V1.0)	0.33	0.51	0.29	0.28	0.75	0.53	0.46	0.92	0.62	0.48	0.45	0.55	0.29	0.33	0.31	0.47	0.5h

The F1 Score↑ on the TnT dataset

	PGSR(Paper)	PGSR(Code_V1.0)
Barn	0.66	0.65
Caterpillar	0.41	0.44
Courthouse	0.21	0.20
Ignatius	0.80	0.81
Meetingroom	0.29	0.32
Truck	0.60	0.66
Mean	0.50	0.51
Time	1.2h	45m

Installation

The repository contains submodules, thus please check it out with

# SSH
git clone git@github.com:zju3dv/PGSR.git
cd PGSR

conda create -n pgsr python=3.8
conda activate pgsr

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 #replace your cuda version
pip install -r requirements.txt
pip install submodules/diff-plane-rasterization
pip install submodules/simple-knn

Dataset Preprocess

Please download the DTU dataset from Neus, the Tanks and Temples dataset from official webiste, the Mip-NeRF 360 dataset from the official webiste. You need to download the ground truth point clouds from the DTU dataset. For the Tanks and Temples dataset, you need to download the reconstruction, alignment and cropfiles from the official webiste.

The data folder should like this:

data
├── dtu_dataset
│   ├── dtu
│   │   ├── scan24
│   │   │   ├── image
│   │   │   ├── mask
│   │   │   ├── cameras_large.npz
│   │   │   └── cameras_sphere.npz
│   │   └── ...
│   ├── dtu_eval
│   │   ├── Points
│   │   │   └── stl
│   │   └── ObsMask
├── tnt_dataset
│   ├── tnt
│   │   ├── Ignatius
│   │   │   ├── images_raw
│   │   │   ├── Ignatius_COLMAP_SfM.log
│   │   │   ├── Ignatius_trans.txt
│   │   │   ├── Ignatius.json
│   │   │   ├── Ignatius_mapping_reference.txt
│   │   │   └── Ignatius.ply
│   │   └── ...
└── MipNeRF360
    ├── bicycle
    └── ...

Then run the scripts to preprocess dataset:

# Install COLMAP
Refer to https://colmap.github.io/install.html

# DTU dataset
python scripts/preprocess/convert_dtu.py --dtu_path your_dtu_path

# Tanks and Temples dataset
python scripts/preprocess/convert_tnt.py --tnt_path your_tnt_path

Training and Evaluation

# Fill in the relevant parameters in the script, then run it.

# DTU dataset
python scripts/run_dtu.py

# Tanks and Temples dataset
python scripts/run_tnt.py

# Mip360 dataset
python scripts/run_mip360.py

Custom Dataset

The data folder should like this:

data
├── data_name1
│   └── input
│       ├── *.jpg/*.png
│       └── ...
├── data_name2
└── ...

Then run the following script to preprocess the dataset and to train and test:

# Preprocess dataset
python scripts/preprocess/convert.py --data_path your_data_path

Some Suggestions:

• Adjust the threshold for selecting the nearest frame in ModelParams based on the dataset;
• -r n: Downsample the images by a factor of n to accelerate the training speed;
• --max_abs_split_points 0: For weakly textured scenes, to prevent overfitting in areas with weak textures, we recommend disabling this splitting strategy by setting it to 0;
• --opacity_cull_threshold 0.05: To reduce the number of Gaussian point clouds in a simple way, you can set this threshold.

# Training
python train.py -s data_path -m out_path --max_abs_split_points 0 --opacity_cull_threshold 0.05

Some Suggestions:

• Adjust max_depth and voxel_size based on the dataset;
• --use_depth_filter: Enable depth filtering to remove potentially inaccurate depth points using single-view and multi-view techniques. For scenes with floating points or insufficient viewpoints, it is recommended to turn this on.

# Rendering and Extract Mesh
python render.py -m out_path --max_depth 10.0 --voxel_size 0.01

Acknowledgements

This project is built upon 3DGS. Densify is based on AbsGau and GOF. DTU and Tanks and Temples dataset preprocess are based on Neuralangelo scripts. Evaluation scripts for DTU and Tanks and Temples dataset are based on DTUeval-python and TanksAndTemples respectively. We thank all the authors for their great work and repos.

Citation

If you find this code useful for your research, please use the following BibTeX entry.

@article{chen2024pgsr,
  title={PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction},
  author={Chen, Danpeng and Li, Hai and Ye, Weicai and Wang, Yifan and Xie, Weijian and Zhai, Shangjin and Wang, Nan and Liu, Haomin and Bao, Hujun and Zhang, Guofeng},
  journal={arXiv preprint arXiv:2406.06521},
  year={2024}
}