SPPSFormer

This is the PyTorch implementation of this paper.

SPPSFormer: High-quality Superpoint-based Transformer for Roof Plane Instance Segmentation from Point Clouds

Cheng Zeng, Xiatian Qi, Huifan Wang, Kai Sun, Pengcheng Zhong, Qiao Xu, Yan Meng, Yangjie Sun and Yuxuan Liu

Dataset

The real building point clouds are selected from RoofN3D.

If you download the dataset directly from the RoofN3D official website, please process it following the method described in our paper and place the dataset in the following structure.

SPPSFormer-main
├── data
│   ├── rn3d
│   │   ├── train
│   │   ├── val

After placing the dataset, run the following commands in order:

python data/superpoint_stage1.py
python data/superpoint_stage2.py
python data/prepare_data.py

This will generate the superpoint-labeled files used in this paper and the .pth files required for model input.

Alternatively, you can use our pre-processed dataset (including .txt and .pth files), which is available for download at the following Baidu Yun link: Dataset . After downloading, please place the dataset in the following directory structure:

SPPSFormer-main
├── data
│   ├── rn3d
│   │   ├── train
│   │   ├── val

Environment

Requirements

• Python 3.x
• Pytorch 1.10
• CUDA 10.x or higher

Below is an example of how to create a Conda environment:

conda create -n sppsformer python=3.8
conda activate sppsformer

Install the dependencies:

conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=11.3 -c pytorch
conda install torch-scatter==2.0.9 -c conda-forge
conda install spconv-cu113==2.3.6 cumm-cu113==0.4.11 -c conda-forge

conda install cudatoolkit=11.7.0 cudnn=8.2.1.32 -y

pip install -r requirements.txt

Install pointgroup_ops:

cd spformer/lib
python setup.py develop

Training

python tools/train.py configs/rn3d.yaml

The resulting model weights will be stored in the exps folder.

Testing and visualization

For testing, you can also use our pre-trained weights, which are available for download at the following Baidu Yun link:Model weights

python tools/test.py configs/rn3d.yaml exps/rn3d/epoch_0100.pth --out out_rn3d
python tools/visualize_X.py

Finally, the txt files containing the ground-truth instance labels and predicted instance labels will be saved in the directory specified in the code. For visualization, please use software such as CloudCompare.

Post-processing

python tools/postprocessing_Batch_2.py
python tools/boundaryRefinement_Batch.py

Citation

If you find this work useful in your research, please cite: