Semi-Supervised Object Detection with Uncurated Unlabeled Data for Remote Sensing Images

💥 News

The paper is accepted by International Journal of Applied Earth Observation and Geoinformation (JAG) and is accessible here.

📖 Introduction

Semi-supervised object detection (SSOD) assumes that both labeled and unlabeled data originate from the same label space, constituting in-distribution (ID) samples. Open-set semi-supervised object detection (OSSOD) accommodates the existence of substantial out-of-distribution (OOD) samples, mirroring the complexities of real-world scenarios.

🎨 Overview

Overview of the proposed open-set semi-supersvised object detection (OSSOD) framework.

🎮 Getting Started

1. Install Environment

conda create --name ossod python=3.8
conda activate ossod

pip install torch===1.7.1+cu110 torchvision===0.8.2+cu110 torchaudio===0.7.2 -f https://download.pytorch.org/whl/torch_stable.html --no-cache

git clone https://github.com/Lans1ng/OSSOD.git
cd OSSOD
pip install -r requirements.txt

2. Prepare Dataset

Download the DIOR dataset (Google Drive or BaiduNetDisk)

Organize the dataset as follow:

dataset/
├──DIOR/
│   ├── annotations_json_split1
│   ├── annotations_json_split2
│   ├── JPEGImages
│   │   ├── 00001.jpg
│   │   ├── 00002.jpg
│   │   ├── ...
│   │   ├── 23463.jpg

JPEGImages represents all images in DIOR dataset
Annotations for split1 and split2 are already included in the repo.

3. Download Checkpoints

Before training，please download the pretrained backbone (ResNet-50) to pretrained_model/backbone.

4. Training

Take split1 of the DIOR dataset as examples:

Use in-distribution (ID) labeled data to train a baseline

python train.py --config configs/baseline/baseline_ssod_split1.py --work-dir work_dirs/split1/baseline_ssod --base_stage True

Use unlabeled data to train semi-supervised detector:

Train detectors with ID unlabeled data

#unbiased teacher
python train.py --config configs/unbiased_teacher/split1/unbiased_teacher_id.py --work-dir work_dirs/split1/unbiased_teacher_ID 

#our proposed ossod
python train.py --config configs/unbiased_teacher/split1/unbiased_teacher_ossod_id.py --work-dir work_dirs/split1/unbiased_teacher_ossod_ID

Train detectors with ID and mixed unlabeled data

#unbiased teacher
python train.py --config configs/unbiased_teacher/split1/unbiased_teacher_id_mix.py --work-dir work_dirs/split1/unbiased_teacher_ID_MIX 

#our proposed ossod
python train.py --config configs/unbiased_teacher/split1/unbiased_teacher_ossod_id_mix.py --work-dir work_dirs/split1/unbiased_teacher_ossod_ID_MIX

Train detectors with ID, mixed and out-of-distribution (OOD) unlabeled data

#unbiased teacher
python train.py --config configs/unbiased_teacher/split1/unbiased_teacher_id_mix_ood.py --work-dir work_dirs/split1/unbiased_teacher_ID_MIX_OOD 

#our proposed ossod
python train.py --config configs/unbiased_teacher/split1/unbiased_teacher_ossod_id_mix_ood.py --work-dir work_dirs/split1/unbiased_teacher_ossod_ID_MIX_OOD

5. Testing

Take split1 of the DIOR dataset as examples:

python eval.py --config configs/unbiased_teacher/split1/unbiased_teacher_ossod_id_mix_ood.py --checkpoint work_dirs/split1/unbiased_teacher_ossod_ID_MIX_OODlatest.pth --eval mAP  --show-dir results

💡 Acknowledgement

🖊️ Citation

If you find this project useful in your research, please consider cite:

@article{liu2024semi,
  title={Semi-Supervised Object Detection with Uncurated Unlabeled Data for Remote Sensing Images},
  author={Liu, Nanqing and Xu, Xun and Gao, Yingjie and Zhao, Yitao and Li, Heng-Chao},
  journal={International Journal of Applied Earth Observation and Geoinformation},
  year={2024}
}