Domain-invariant Progressive Knowledge Distillation for UAV-based Object Detection

Liang Yao (姚亮) ,     Fan Liu (刘凡) ✉ ,     Chuanyi Zhang (张传一) ,    

Zhiquan Ou (欧志权) ,     Ting Wu (吴婷) ,    

News

• 2024/10/29: Our paper has been accepted by IEEE Geoscience and Remote Sensing Letters (IEEE GRSL).
• 2024/07/29: We propose a Progressive and Domain-invariant Knowledge Distillation method for UAV-OD. Codes and models will be open-sourced at this repository.

Getting Started

Install

• Clone this repo:

  git clone https://github.com/1e12Leon/PDIKD.git

• Create a conda virtual environment and activate it:

  conda create -n PDIKD python=3.6 -y
  conda activate PDIKD

• Install CUDA Toolkit 11.3 (link) and cudnn==8.2.1 (link), then install PyTorch==1.10.1:

  conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch -y
  # if you prefer other cuda versions, please choose suitable pytorch versions
  # see: https://pytorch.org/get-started/locally/

Data Preparation

This code repository uses VOC format object detection data.

VOCdevkit
    ├───VOC2007
    │   ├───ImageSets
    │   |     ├───Main
    │   |            test.txt
    │   |            train.txt
    │   |            trainval.txt
    │   |            val.txt
    │   ├───JPEGImages
    │   │       xxx1.jpg
    │   │       xxx2.jpg
    │   │       ...
    │   └───Annotations
                xxx1.xml
                xxx2.xml
                ...

Training

python distill.py

Experimental Results

---

Dataset	Method	Publication	mAP	AP_0.5	AP_0.75	AP_S	AP_M	AP_L
	YOLOv7-L (T)	CVPR2023	6.94	30.26	16.89	8.17	26.90	42.41
	YOLOv7-Tiny (S)	CVPR2023	11.62	21.95	11.25	4.71	18.39	32.60
VisDrone	FitNets	ICLR2015	12.87	24.71	12.34	5.21	20.62	34.77
	BCKD	ICCV2023	16.08	30.71	15.81	8.81	24.90	26.60
	CrossKD	CVPR2024	14.96	29.22	14.11	8.79	23.77	24.54
	Ours	-	17.07	31.92	16.77	9.56	25.90	38.98
	YOLOv7-L (T)	CVPR2023	33.66	64.01	30.36	18.44	42.19	40.97
	YOLOv7-Tiny (S)	CVPR2023	30.21	58.41	27.78	15.66	37.88	36.68
SynDrone	FitNets	ICLR2015	32.12	61.73	29.32	16.82	40.21	39.38
	BCKD	ICCV2023	32.66	63.12	29.95	21.52	40.52	31.93
	CrossKD	CVPR2024	31.79	61.85	28.46	17.84	39.53	30.88
	Ours	-	35.12	65.09	33.30	22.27	43.08	36.51

Citation

@article{yao2024domain,
  title={Domain-invariant Progressive Knowledge Distillation for UAV-based Object Detection},
  author={Yao, Liang and Liu, Fan and Zhang, Chuanyi and Ou, Zhiquan and Wu, Ting},
  journal={arXiv preprint arXiv:2408.11407},
  year={2024}
}

Contact

Please Contact yaoliang@hhu.edu.cn