Analyzing the Impact of Low-Rank Adaptation for Cross-Domain Few-Shot Object Detection in Aerial Images

This repository contains the official implementation of our preprint paper:
📄 Analyzing the Impact of Low-Rank Adaptation for Cross-Domain Few-Shot Object Detection in Aerial Images

We implement LoRA (Low-Rank Adaptation) with DiffusionDet for cross-domain few-shot object detection, using the Hugging Face Transformers framework. A custom library, fsdetection, adapts HuggingFace's trainer and dataset components to better support cross-domain object detection.

📂 Project Structure

├── configs/
│   ├── dataset/                # Template and metadata for your dataset
│   │   └── template.py         # Fill this to describe your dataset structure
│   ├── models/                 
│   │   └── template.json       # Example config for LoRA/DiffusionDet
│   └── dataset/
│       └── coco_format.py      # Script to convert COCO-format dataset to Hugging Face format
├── launch_experiments.py       # Multi-training launcher
├── run_object_detection.py     # Single training run
├── upload_to_hub.py            # Push datasets to the Hugging Face Hub
├── ...

📦 Dataset Preparation & Hugging Face Integration

To use your own COCO-format dataset:

• Describe your dataset
  Fill out the file configs/dataset/template.py with information specific to your dataset (e.g., path, split keys, label mapping, etc.).

• Convert your dataset (and optionally push to the Hub)
  You don't run the script coco_format.py directly. Instead, use the COCO class defined inside it to convert your dataset to Hugging Face format.

To simplify the process, you can just run:

python upload_to_hub.py --hf_repo your_hf_repo --dataset dataset_name

• Using the dataset
  After pushing, your dataset is available to any script via datasets.load_dataset(...), or you can save and load it locally using Hugging Face’s dataset API.

🏁 Training

🔹Single Training

Use run_object_detection.py with arguments or a config file:

python run_object_detection.py ...

🔸Multiple Trainings

Use launch_experiments.py with a YAML file defining your sweep:

python launch_experiments.py --config-file configs/models/template.json ...

🌿 Branches

• main: Focused on cross-domain few-shot object detection.

• classic-det: Classic object detection use cases. (coming soon)

📝 Note: Both branches use similar launch commands, but configuration files may differ slightly.

🔧 Requirements

Install dependencies and local packages:

pip install -r requirements.txt

📄 Citation

If you use this code, please consider citing our preprint:

@misc{talaoubrid2025analyzingimpactlowrankadaptation,
  title={Analyzing the Impact of Low-Rank Adaptation for Cross-Domain Few-Shot Object Detection in Aerial Images}, 
  author={Hicham Talaoubrid and Anissa Mokraoui and Ismail Ben Ayed and Axel Prouvost and Sonimith Hang and Monit Korn and Rémi Harvey},
  year={2025},
  eprint={2504.06330},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2504.06330}, 
}

📨 Contact

For questions, feel free to open an issue or contact: [email protected]