GuidedBox: Weakly Supervised Road Segmentation

🚀 Abstract

Road segmentation in remote sensing is essential for applications such as urban planning, traffic monitoring, and autonomous driving. However, obtaining pixel-wise segmentation labels is labor-intensive. GuidedBox addresses this challenge with a novel weakly supervised approach that leverages segmentation-guided box annotations.

Using a teacher-student framework, the teacher generates high-quality pseudo masks, while a noise-aware confidence scoring mechanism filters low-quality masks to optimize training dynamically. Our method achieves state-of-the-art performance with an AP50 score of 0.9231 on the Massachusetts Roads Dataset, surpassing existing methods such as SOLOv2, CondInst, and Mask R-CNN.

👥 Author

Teerapong Panboonyuen (Kao Panboonyuen)
Laboratory of Mapping and Positioning from Space (MAPS) Technology Research Center,
Department of Survey Engineering, Faculty of Engineering, Chulalongkorn University

📊 Results and Achievements

Achieved a 0.9231 AP50 score on the Massachusetts Roads Dataset with Swin-Base-FPN.
Outperformed state-of-the-art methods like SOLOv2, CondInst, and Mask R-CNN in weakly supervised road segmentation.

⚙️ Quick Start

Requirements

Python 3.8+
PyTorch 1.8+
CUDA 11.1+
Other dependencies listed in requirements.txt

Installation

git clone https://github.com/kaopanboonyuen/GuidedBox.git
cd GuidedBox

python3 -m venv guidedbox-env
source guidedbox-env/bin/activate  # Windows: guidedbox-env\Scripts\activate

pip install -r requirements.txt

Download Dataset

Download the Massachusetts Roads Dataset and place it inside the data/ directory.

📈 How to Use

Train the Model

python train.py --config configs/guidedbox_config.yaml

Evaluate the Model

python evaluate.py --checkpoint checkpoints/guidedbox_best_model.pth --data data/test/

Run Inference

python inference.py --image_path images/sample.jpg --output_dir results/

🌍 Live Demo

Experience GuidedBox online: GuidedBox Demo

📂 Datasets

- Public Dataset: Massachusetts Roads Dataset

🔍 Citations

If you find GuidedBox useful, please cite:

@article{panboonyuen2025guidedbox,
  title={GuidedBox: A segmentation-guided box teacher-student approach for weakly supervised road segmentation},
  author={Panboonyuen, Teerapong},
  journal={European Journal of Remote Sensing},
  year={2025}
}

And for the dataset:

@phdthesis{MnihThesis,
  author = {Volodymyr Mnih},
  title = {Machine Learning for Aerial Image Labeling},
  school = {University of Toronto},
  year = {2013}
}

📜 License

This project is licensed under the MIT License.

📧 Contact

Teerapong Panboonyuen
https://kaopanboonyuen.github.io
panboonyuen.kao@gmail.com