PolyGNN

---

PolyGNN is an implementation of the paper PolyGNN: Polyhedron-based Graph Neural Network for 3D Building Reconstruction from Point Clouds. PolyGNN learns a piecewise planar occupancy function, supported by polyhedral decomposition, for efficient and scalable 3D building reconstruction.

🛠️ Setup

Repository

Clone the repository:

git clone https://github.com/chenzhaiyu/polygnn && cd polygnn

All-in-one installation

Create a conda environment with all dependencies:

conda env create -f environment.yml && conda activate polygnn

Manual installation

Still easy! Create a conda environment and install mamba for faster parsing:

conda create --name polygnn python=3.10 && conda activate polygnn
conda install mamba -c conda-forge

Install the required dependencies:

mamba install pytorch torchvision sage=10.0 pytorch-cuda=11.7 pyg=2.3 pytorch-scatter pytorch-sparse pytorch-cluster torchmetrics rtree -c pyg -c pytorch -c nvidia -c conda-forge
pip install abspy==0.2.6 hydra-core hydra-colorlog omegaconf trimesh tqdm wandb plyfile

🚀 Usage

Quick start

Download the mini dataset and pretrained weights:

python download.py dataset=mini

In case you encounter issues (e.g., Google Drive limits), manually download the data and weights here, then extract them into ./checkpoints/mini and ./data/mini, respectively. The mini dataset contains 200 random instances (~0.07% of the full dataset).

Train PolyGNN on the mini dataset (provided for your reference and is not intended for full-scale training):

python train.py dataset=mini

The data will be automatically preprocessed the first time you initiate training.

Evaluate PolyGNN with option to save predictions:

python test.py dataset=mini evaluate.save=true

Generate meshes from predictions:

python reconstruct.py dataset=mini reconstruct.type=mesh

Remap meshes to their original CRS:

python remap.py dataset=mini

Generate reconstruction statistics:

python stats.py dataset=mini

Available configurations

# check available configurations for training
python train.py --cfg job

# check available configurations for evaluation
python test.py --cfg job

Alternatively, review the configuration file: conf/config.yaml.

Full dataset

The Munich dataset is available for download on Zenodo. Note that it requires 332 GB of storage when decompressed. Meshes for CRS remapping can be downloaded here.

Custom data

PolyGNN requires polyhedron-based graphs as input. To prepare this from your own point clouds:

1. Extract planar primitives using tools such as Easy3D or GoCoPP, preferably in VertexGroup format.
2. Build CellComplex from the primitives using abspy. Example code:

   from abspy import VertexGroup, CellComplex
   vertex_group = VertexGroup(vertex_group_path, quiet=True)
   cell_complex = CellComplex(vertex_group.planes, vertex_group.aabbs,
                              vertex_group.points_grouped, build_graph=True, quiet=True)
   cell_complex.prioritise_planes(prioritise_verticals=True)
   cell_complex.construct()
   cell_complex.save(complex_path)

   Alternatively, you can modify CityDataset or TestOnlyDataset to accept inputs directly from VertexGroup or VertexGroupReference.
3. Structure your dataset similarly to the provided mini dataset:

   YOUR_DATASET_NAME
   └── raw
       ├── 03_meshes
       │   ├── DEBY_LOD2_104572462.obj
       │   ├── DEBY_LOD2_104575306.obj
       │   └── DEBY_LOD2_104575493.obj
       ├── 04_pts
       │   ├── DEBY_LOD2_104572462.npy
       │   ├── DEBY_LOD2_104575306.npy
       │   └── DEBY_LOD2_104575493.npy
       ├── 05_complexes
       │   ├── DEBY_LOD2_104572462.cc
       │   ├── DEBY_LOD2_104575306.cc
       │   └── DEBY_LOD2_104575493.cc
       ├── testset.txt
       └── trainset.txt

4. To train or evaluate PolyGNN using your dataset, run the following commands:

   # start training
   python train.py dataset=YOUR_DATASET_NAME
   
   # start evaluation
   python test.py dataset=YOUR_DATASET_NAME

   For evaluation only, you can instantiate your dataset as a TestOnlyDataset, as in this line.

👷 TODOs

• Demo with mini data and pretrained weights
• Short tutorial for getting started
• Host the full dataset

🎓 Citation

If you use PolyGNN in a scientific work, please consider citing the paper:

@article{chen2024polygnn,
title = {PolyGNN: Polyhedron-based graph neural network for 3D building reconstruction from point clouds},
journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
volume = {218},
pages = {693-706},
year = {2024},
issn = {0924-2716},
doi = {https://doi.org/10.1016/j.isprsjprs.2024.09.031},
url = {https://www.sciencedirect.com/science/article/pii/S0924271624003691},
author = {Zhaiyu Chen and Yilei Shi and Liangliang Nan and Zhitong Xiong and Xiao Xiang Zhu},
}

You might also want to check out abspy for 3D adaptive binary space partitioning and Points2Poly for reconstruction with deep implicit fields.