search

HeinzJiao / PolyR-CNN

Official repository for the ISPRS paper "PolyR-CNN: R-CNN for End-to-End Polygonal Building Outline Extraction"
https://doi.org/10.1016/j.isprsjprs.2024.10.006
Other
7 stars 0 forks source link

readme

PolyR-CNN: R-CNN for End-to-End Polygonal Building Outline Extraction

This is the official repository for PolyR-CNN, a model designed for end-to-end polygonal building outline extraction.

Data Preprocessing

To minimize training time, we have preprocessed the CrowdAI dataset by localizing the polygon padding process to ensure a fixed number of vertices per polygon, reducing the need for on-the-fly computation during training. For detailed information on the polygon padding process, please refer to the article.

The preprocessed CrowdAI dataset can be downloaded here (Google Drive link to be provided).

To run the preprocessing yourself, you can execute the following command:

python3 preprocess_annotation.py --json_path /path/to/annotations.json --save_path /path/to/save/annotation_preprocessed.json --is_training True --num_corners 96

Environment Setup

Training

To start training the PolyR-CNN model on the CrowdAI dataset, you can use the following command:

python3 train_net.py --num-gpus <number_of_gpus> --config-file configs/polyrcnn.res50.100pro.aicrowd.yaml

You can use different configuration files from the configs/ folder depending on your requirements (e.g., switching to ResNet101 or Swin Transformer backbones).

Status

The remaining code is currently being organized and will be uploaded soon. Stay tuned for updates!

Coming Soon