SuperpixelGraph: Semi-automatic generation of building footprint through semantic-sensitive superpixel and neural graph networks

---

This repository provides a PyTorch implementation of semantic-sensitive superpixel generation network in our work SuperpixelGraph.

Requirements

---

The codes were developed and tested mainly with the following dependencies:

gdal=3.3.2
opencv=4.5.3
python=3.9
pytorch=1.9.1
pytorch-lightning=1.6.0
scikit-image=1.0.2

Running the code

---

Training

Use the following command to train model on WHU dataset:

python spn_train.py --data_dir /path/to/WHU/ --output_dir /path/to/outputs/ --dataset WHU --num_epochs 300

Testing and evaluation

Use the following command to run model on test set of WHU:

python spn_test.py --model_path ./ckpt/superpixel_net_downsize16_whu.ckpt --data_dir /path/to/WHU/ --output_dir /path/to/outputs/

or run model on a single image:

python spn_demo.py --model_path /path/to/model --image_path /path/to/image --output_dir /path/to/output/

If you want to adjust the number of superpixels, use options -H and -W to set the height and width of images inputted to the model. Larger height and width you set, more superpixels will be generated. Note that the height or width should be a multiple of 16.

Tools provided by superpixel-benchmark are used to evaluate the testing results.