GuidedBox: A Segmentation-Guided Box Teacher-Student Approach for 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)
Senior Research Scientist, MARSAIL (Motor AI Recognition Solution Artificial Intelligence Laboratory)
C2F High-Potential Postdoctoral Researcher, Chulalongkorn University

📄 Official Publication

Explore the comprehensive research article published in European Journal of Remote Sensing (Taylor & Francis, 2025), where GuidedBox unveils an innovative teacher-student framework for weakly supervised road segmentation. This breakthrough method delivers exceptional performance on multiple challenging datasets, pushing the boundaries of remote sensing technology.

Discover detailed methodology, insightful experimental results, and access the open-source code that powers this advancement. Whether you’re a researcher, practitioner, or enthusiast, this paper offers valuable insights into state-of-the-art segmentation techniques.

📊 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

See GuidedBox in action with our interactive online demo. Explore how the segmentation-guided box teacher-student framework performs real-time road segmentation with precision and efficiency.

📂 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    = {Teerapong Panboonyuen},
  journal   = {European Journal of Remote Sensing},
  volume    = {58},
  number    = {1},
  year      = {2025},
  doi       = {10.1080/22797254.2025.2540963},
  publisher = {Taylor & Francis},
  url       = {https://doi.org/10.1080/22797254.2025.2540963},
  note      = {Article 2540963, Received 13 May 2024, Accepted 24 Jul 2025, Published online: 01 Aug 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