HiSup: Accurate polygonal mapping of buildings in satellite imagery with hierarchical supervision

Bowen Xu*, Jiakun Xu*, Nan Xue† and Gui-Song Xia†

(* indicates equal contributions, and † indicates the corresponding authors)

---

Highlights

• We address the mask reversibility issue of polygonal mapping of buildings by proposing the HiSup, which takes the hierarchical supervision signals of vertices, boundaries and masks to guide the convolutional neural network to learn in a shape-aware fashion.
• We present a key component of aggregating the embedding of attraction fields into the learning of high-level masks and the bottom-level vertices.
• We set several new state-of-the-art performances on the challenging AICrowd benchmark in terms of AP and Boundary AP by learning semantically-accurate and geometrically-precise masks.

Installation

Ubuntu-18.04, CUDA 11.0, pytorch1.7/1.8, GCC 7.3

conda create -n hisup python=3.7
conda activate hisup
conda install pytorch==1.7.0 torchvision==0.8.0 cudatoolkit=11.0 -c pytorch

cd HiSup
conda develop .
pip install -r requirements.txt

cd hisup/csrc/lib
make

For evaluation with boundary IoU, please install boundary IoU API following the installation instruction.

git clone [email protected]:bowenc0221/boundary-iou-api.git
cd boundary_iou_api
pip install -e .

Quickstart with the pretrained model

You can run the following command to get quickstart.

python scripts/demo.py --dataset crowdai --img [YOUR_IMAGE_PATH]

--dataset crowdai means load the model pretrained on AICrowd dataset, if you want to load the model pretrained on Inria dataset, simply change crowdai into inria.

You can also run our demo using Colab:

Training & Testing

Data prepare

• Download the train.tar.gz and val.tar.gz from AICrowd dataset
• Download the Inria dataset and put data in the inria/raw files
• Run the inria_to_coco.py from the tools file to get training data in COCO format for Inria dataset. After the generation, the training data should be in inria/train file.

The structure of the data file should be like:

/data # AICrowd dataset downloaded from website
|-- crowdai
    |-- train
    |   |-- images
    |   |-- annotation.json
    |   |-- annotation-small.json
    |-- val
    |   |-- images
    |   |-- annotation.json
    |   |-- annotation-small.json
|-- inria
    |-- raw
        |-- train
        |   |-- images
        |   |-- gt
        |-- test
        |   |-- images
    |-- train
        |   |-- images
        |   |-- annotation.json

Training

The model with HRNetV2 as backbone are initialized with imagenet pretrained parameters. You could download them from https://github.com/HRNet/HRNet-Image-Classification and put them in the path of ./hisup/backbones/hrnet_imagenet.

Single GPU training

python scripts/train.py --config-file config-files/crowdai-small_hrnet48.yaml 

Multiple GPUs training

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 scripts/multi_train.py --config-file config-files/crowdai_hrnet48.yaml

Testing

After training, a file defined by the "OUTPUT_DIR" in the config file will appear in the catalog, which contains the trained parameters. For the "crowdai-small_hrnet48.yaml", the trained parameters are like:

/outputs/crowdai_hrnet48
-- config.yml        # saved hyper-parameters setting
-- log.txt           # saved experimental log
-- train.log         # saved training log
-- model_00030.pth   # parameters
-- last_checkpoint   # directory of the parameters' file

We also provide the pretrained models with HRNetV2-W48 as backbone on AICrowd dataset and Inria dataset. You can download the pretrained models, and put them in the right directory according to configuration files.

python scripts/test.py --config-file config-files/crowdai_hrnet48.yaml

Evaluation

We provide implementation of different metrics for evaluation. You can run the following command to evaluate the test results in MS-COCO format. The prediction in json format corresponding to the validation set of AICrowd dataset is provided.

python tools/evaluation.py --gt-file [GT_ANNOTATION_FILE] --dt-file [PREDICT_ANNOTATION_FILE] --eval-type boundary_iou

Citation

If you find our work useful in your research, please consider citing:

@article{XU2023284,
    author = {Bowen Xu and Jiakun Xu and Nan Xue and Gui-Song Xia},
    title = {HiSup: Accurate polygonal mapping of buildings in satellite imagery with hierarchical supervision},
    journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
    volume = {198},
    pages = {284-296},
    year = {2023},
    issn = {0924-2716},
    doi = {https://doi.org/10.1016/j.isprsjprs.2023.03.006},
}

Acknowledgement

This repo benefits from hawp, ECA-Net, HR-Net, boundary iou api, frame-field, polymapper, polyworld. We thank the authors for their great work.