BONAI (LOFT) — Pretrained Weights

Official pretrained weights for LOFT, the model introduced in Learning to Extract Building Footprints from Off-Nadir Aerial Images (IEEE TPAMI 2022), trained on the BONAI dataset.

BONAI is a dataset for building footprint extraction (BFE) in off-nadir aerial images, containing 268,958 building instances across 3,300 aerial images from six Chinese cities (Shanghai, Beijing, Harbin, Jinan, Chengdu, Xi'an), with instance-level roof and footprint annotations and the corresponding offset vector for each building.

Files

File	Size	SHA256
BONAI-LOFT-ROA-F1Score-64.31.pth	651,220,024 bytes	2003f790d9a8a6e829cfc25b245a5db106abb0ca45b455a8910d699bec521e6b

The checkpoint filename is kept as originally released by the authors; it indicates an F1 score of 64.31 for this checkpoint.

Usage

The model is implemented in the official MMDetection-based codebase. For configs, inference and evaluation (including tools/bonai/bonai_evaluation.py), see the GitHub repository:

• Code: https://github.com/jwwangchn/BONAI
• Paper: https://arxiv.org/abs/2204.13637
• Dataset: https://huggingface.co/datasets/jwwangchn/BONAI

License

The BONAI code is released under the MIT License (see the GitHub repository). This checkpoint is distributed by the BONAI authors for research use with the official code.

Citation

If you use the BONAI dataset, codebase or models in your research, please consider citing:

@article{wang2022bonai,
  author  = {Wang, Jinwang and Meng, Lingxuan and Li, Weijia and Yang, Wen and Yu, Lei and Xia, Gui-Song},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  title   = {Learning to Extract Building Footprints from Off-Nadir Aerial Images},
  year    = {2022},
  doi     = {10.1109/TPAMI.2022.3162583}
}