/polycity

PolyCity

Introduction

This is the official code for the paper entitled 'Joint Semantic-Geometric Learning for Polygonal Building Segmentation from High-Resolution Remote Sensing Images'.

Environment

  • All the code has been run and tested on Ubuntu 18.04, Python 3.7.4, Pytorch 1.6.0, CUDA 10.1 and GTX 2080Ti GPUs.

  • Setup python environment：

pip install -r requirements.txt

Usage

Step1：Building-HRNet

Data Preparation

  • The dataset we use is SpaceNet AOI 2 – Las Vegas. If you want to use other datasets, you need to prepare the dataset following the format of example data in building_hrnet/data/example_data we prepared.

  • The Python script building_hrnet/BinaryAnnotationGeneration.py can convert the coco format file into the binary file.

  • The MATLAB script building_hrnet/AnnotationGenerationFromBinary.m can convert the binary file into annotation file in the required format and corresponding colormap.

  • Modify the data directory in building_hrnet/files/vegas_train(val)_list.txt for your machine.

Train & Test

Modify the parameters for your machine in building_hrnet/train(test)_vegas.sh, then run:

For train: sh train_vegas.sh
For test: sh test_vegas.sh

Post-processing

Modify the data directory in building_hrnet/generate_vertex_115/select-vertex.sh and then run it.

Step2：Polygon-rnnpp

Data Preparation

  • Download the pre-trained Pytorch Resnet-50 model from here.

  • Edit the experiment file at polygon-rnnpp/Experiments/spaceNet/building_ggnn_vegas_epsilon1.json & building_ggnn_vegas_epsilon1_0.6.json.

  • Please prepare the dataset following the format of example data in polygon-rnnpp/data/example_data we prepared.

Train

  • Edit the data folder name in polygon-rnnpp/Scripts/train/train_spNb_ggnn_epsilon1.py at line 39 & 40.

  • Run polygon-rnnpp/train_spNb_ggnn_epsilon1.sh.

Test

  • Edit the data folder name in polygon-rnnpp/Scripts/prediction/generate_annotation_ggnn_epsilon1_0.6.py at line 36.

  • Run polygon-rnnpp/test_spNb_ggnn_epsilon1_0.6.sh.

Citing

If you find our work useful in your research, please consider cite.

@article{LI202326,
    author = {Weijia Li and Wenqian Zhao and Jinhua Yu and Juepeng Zheng and Conghui He and Haohuan Fu and Dahua Lin},
    title = {Joint semantic–geometric learning for polygonal building segmentation from high-resolution remote sensing images},
    journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
    volume = {201},
    pages = {26-37},
    year = {2023},
    issn = {0924-2716},
    doi = {https://doi.org/10.1016/j.isprsjprs.2023.05.010}
}

Contact

If you need the complete code of this project, please send email to liweij29@mail.sysu.edu.cn.

(The code is for academic use only. Please indicate your affiliation and the purpose of use.)