PlaneTransformer

This is the PyToch implementation of the following manuscript:

A transformer-based roof plane segmentation approach for airborne LiDAR point clouds

Siyuan You, Guozheng Xu, Pengwei Zhou, Yubing Wei, Jian Yao, Li Li

This manuscript has been accepted by TGRS Journal.

Datasets

To evaluate the performance of the propsed PlaneTransformer, we train and test the model on a synthetic and a real-world dataset, which can be derived from DeepRoofPlane.

Usage

The environment requires torch1.12.1+cu113 with python=3.9. The requirements list is shown as follow:

• scikit-learn
• pyyaml
• libboost
• plyfile
• open3d
• scipy
• matplotlib
• trimesh
• tensorboardx

If the above requirements are incomplete, you can install the missed dependencies according to the compilation errors.

Data preparation

Before traning the model, you need to prepare the data files as follows:

1. Generate the train.txt and test.txt. The building3d_train.txt and building3d_test.txt are provided in ./building3d/meta_data/, The roofpc3d_train.txt and roofpc3d_test.txt are provided in ./roofpc3d/meta_data/.
2. copy the [scans_id].txt to scans floder below the corresponding datasets folder.
3. Generate the Groud Truth of the planar centers by running the command:

cd TransformerPlane
python cal_plane_center.py

The data files are organized as follows.

PlaneTransformer
├── building3d
│   ├── meta_data
│   │   ├── building3d_train.txt
|   |   ├── building3d_test.txt
|   ├── scans
|   |   ├── [scans_id].txt & [scans_id]_planes.txt
├── roofpc3d
│   ├── meta_data
│   │   ├── roofpc3d_train.txt
|   |   ├── roofpc3d_test.txt
|   ├── scans
|   |   ├── [scans_id].txt & [scans_id]_planes.txt
|   ├── scans_test
|   |   ├── [scans_id].txt & [scans_id]_planes.txt

Train and test

Then, you can run main.py for training and testing the model. Notably, the commands with parameter settings required by the model are stored in the ./scripts/scripts.sh file. The model checkpoints will be stored in the ./outputs folder. An example for training command:

CUDA_VISIBLE_DEVICES=0 python main.py --dataset_name roofpc3d --ngpus 1 --nqueries 16 --enc_nlayers 2 --dec_nlayers 8 --max_epoch 180 --batchsize_per_gpu 10 --save_separate_checkpoint_every_epoch 30 --checkpoint_dir outputs/roofpc3d_q_16_dec_8

An example for testing command (--test_ckpt $test_model_dir$):

CUDA_VISIBLE_DEVICES=0 python main.py --dataset_name roofpc3d --ngpus 1 --nqueries 16 --enc_nlayers 2 --dec_nlayers 8  --batchsize_per_gpu 1 --checkpoint_dir outputs/roofpc3d_q_16_dec_8 --test_ckpt checkpoints/checkpoint.pth --test_only

• If you want post-processing with different distance metrics for point assignment, add "--dist_cmd $norm_dist||center_dist||plane_dist$" to testing command.
• If you want post-processing with different k neareat numbers for plane estimation, add "--q_k_num $int$" to testing command.

Results

• The model outputs: pred_planar_centers, pred_point_offsets, plane_segments will be saved in ./result_pred_pl, ./result_pred_pt and ./result_plane_seg. Several examples are provided in these folders.

• For the quantative evaluation of the experimental results, you can check ./evaluation_res.

Citation

If you find our work useful for your research, please consider citing our paper.

A transformer-based roof plane segmentation approach for airborne LiDAR point clouds

Siyuan You, Guozheng Xu, Pengwei Zhou, Yubing Wei, Jian Yao, Li Li

@article{you2025transformer,
  title={A Transformer-based Roof Plane Segmentation Approach for Airborne LiDAR Point Clouds},
  author={You, Siyuan and Xu, Guozheng and Zhou, Pengwei and Wei, Yubing and Yao, Jian and Li, Li},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  year={2025},
  publisher={IEEE}
}

In addition, if you use the Roofpc3d and Building3d dataset, please also consider citing the following paper:

@article{li2024boundary,
  title={A boundary-aware point clustering approach in Euclidean and embedding spaces for roof plane segmentation},
  author={Li, Li and Li, Qingqing and Xu, Guozheng and Zhou, Pengwei and Tu, Jingmin and Li, Jie and Li, Mingming and Yao, Jian},
  journal={ISPRS Journal of Photogrammetry and Remote Sensing},
  volume={218},
  pages={518--530},
  year={2024},
  publisher={Elsevier}
}

Contact:

Li Li (li.li@whu.edu.cn)