[CVPR 2025] ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects

Woojin Lee^1*    Hyugjae Chang^1*    Jaeho Moon¹    Jaehyup Lee^2†    Munchurl Kim^1†

^*Co-first authors    ^†Co-corresponding authors

¹KAIST (Korea Advanced Institute of Science and Technology), South Korea
²KNU (Kyungpook National University), South Korea

---

This repository is the official PyTorch implementation of "ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects". ABBSPO is a weakly supervised oriented object detection (WS-OOD) framework that systematically identifies and addresses the scale and angle supervision mismatch arising from the use of tight horizontal bounding box annotations and minimum circumscribed rectangle operations for rotated box generation.

---

📧 News

• Dec 15, 2025: Initial code release
• Feb 27, 2025: Paper accepted to CVPR 2025

---

🔧 Tested Environment

• OS: Ubuntu 22.04
• Python: 3.8
• PyTorch: 1.13.1
• CUDA: 11.7
• GPU: NVIDIA RTX 3090

⚙️ Environment Setup

Install Environment (Recommended)

git clone https://github.com/KAIST-VICLab/ABBSPO.git
cd ABBSPO

conda create -n abbspo python=3.8
conda activate abbspo

Step 1. Install PyTorch (CUDA 11.7)

pip install torch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117

Step 2. Install MMCV with CUDA extensions (IMPORTANT)

pip install mmcv==2.0.0rc4 -f https://download.openmmlab.com/mmcv/dist/cu117/torch1.13/index.html

⚠️ Important Note: Installing mmcv without the OpenMMLab wheel URL will result in a CPU-only build and cause runtime errors such as:

ModuleNotFoundError: No module named 'mmcv._ext'

Step 3. Install remaining dependencies

pip install -r requirements.txt
pip install -v -e .

(Optional) Installation Sanity Check

python - <<EOF
from mmcv.ops import batched_nms
print("MMCV CUDA ops are correctly installed.")
EOF

📁 Data Preparation

Please refer to tools/data/README.md for dataset preparation.

📦 Pretrained Models

All checkpoints are trained for 12 epochs using the configuration files provided in configs/abbspo/.

Note: The DOTA-v1.0 checkpoint is trained using only the training split (without validation data).

Dataset	Model	Training Log
DIOR	model	log
DOTA-v1.0	model	log
SIMD	model	log

🚀 Get Started

Training

ABBSPO follows the standard MMRotate training pipeline.

Example training command on DIOR:

python tools/train.py \
configs/abbspo/abbspo-le90_r50_fpn-1x_dior.py

To train on other datasets, simply change the configuration file:

DOTA-v1.0: configs/abbspo/abbspo-le90_r50_fpn-1x_dota.py

SIMD: configs/abbspo/abbspo-le90_r50_fpn-1x_simd.py

Testing

Example evaluation command on DIOR:

python tools/test.py \
configs/abbspo/abbspo-le90_r50_fpn-1x_dior.py \
work_dirs/abbspo-le90_r50_fpn-1x_dior/epoch_12.pth 

To visualize detection results:

python tools/test.py \
configs/abbspo/abbspo-le90_r50_fpn-1x_dior.py \
work_dirs/abbspo-le90_r50_fpn-1x_dior/epoch_12.pth \
--show-dir visual_results/abbspo-le90_r50_fpn-1x_dior \
--show-score-thr 0.3

For more detailed configuration options and advanced usage, please refer to the MMRotate User Guide

Project Page

Please visit our project page for more experimental results.

Citation

If the content is useful, please cite our paper:

@inproceedings{lee2025abbspo,
  title={ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects},
  author={Lee, Woojin and Chang, Hyugjae and Moon, Jaeho and Lee, Jaehyup and Kim, Munchurl},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={8848--8858},
  year={2025}
}

Acknowledgement

This repository is built upon FMA-Net, One Look is Enough, and MMRotate. We gratefully thank the MMRotate team and H2RBox-v2 authors for their excellent open-source contributions, which made our implementation and experiments much easier.