By Yi Jie WONG & Yin-Loon Khor et al
This code is part of our solution for 2024 IEEE BigData Cup: Building Extraction Generalization Challenge (IEEE BEGC2024). Specifically, our solutions involved 2 methods:
- Additional dataset ➡️ extract additional building footprint data from the Microsoft Building Footprint (BF) dataset for Redmond, Washington, and Las Vegas, Nevada.
- Diffusion augmentation ➡️ using segmentation-guided diffusion model to transform land cover segmentation masks + building box labels (polygon) into realistic synthetic images
We find that our YOLO model trained using both the original and synthetic dataset generated by our diffusion model is better than the YOLO model trained with the original dataset alone. We use the extracted dataset to train our YOLOv8-based instance segmentation model, along with the training set provided by the IEEE BEGC2024 dataset. Results show that YOLOv8 trained on BEGC2024 with the additional dataset achieves a significant F1-score improvement compared to training on the BEGC2024 training set alone. Our approach ranked 1st globally in the IEEE Big Data Cup 2024 - BEGC2024 challenge! 🏅🎉🥳 Feel free to check the summary of our work here.
Please visit our codes in:
- YOLO-based Building Instance Segmentation here
- Segmentation Guided Diffusion Model for Diffusion Augmentation here
Our paper has been accepted by IEEE BigData 2024! Please cite our paper if this repo helps your research. The preprint is available here
@InProceedings{Wong2024,
title = {Cross-City Building Instance Segmentation: From More Data to Diffusion-Augmentation},
author = {Yi Jie Wong and Yin-Loon Khor and Mau-Luen Tham and Ban-Hoe Kwan and Anissa Mokraoui and Yoong Choon Chang},
booktitle={2024 IEEE International Conference on Big Data (Big Data)},
year={2024}}