You Only Look Clusters for Tiny Object Detection in Aerial Images

This is the implementation of "YOLC: You Only Look Clusters for Tiny Object Detection in Aerial Images".[Paper]

<p align="center"> <img src="framework.jpg"/> </p>

Requirement

This repo is implemented based on mmdetection. Please refer to get_started.md

- Python >= 3.8
- PyTorch >= 1.7.0
- mmdetection == 2.26.0 (>=2.17.0, <3.0.0)
- kornia == 0.6.9

Checkpoint

We provide YOLC's checkpoint file. Please download model weights at the Link.

Data preparation

Use gen_crop.py to generate VisDrone train dataset. You can directly use annotations/VisDrone2019-DET_val_coco.json for its val dataset.

python gen_crop.py

Train

Training on a single GPU

python train.py configs/yolc.py

Training on multiple GPUs

./dist_train.sh configs/yolc.py <your_gpu_num>

Eval

# set saved_crop to 0,1,2,3 corresponding to LSM (k=saved_crop)
python eval_yolc.py

Citation

If you find our paper is helpful, please consider citing our paper:

@article{liu2024yolc,
  title={YOLC: You Only Look Clusters for Tiny Object Detection in Aerial Images},
  author={Liu, Chenguang and Gao, Guangshuai and Huang, Ziyue and Hu, Zhenghui and Liu, Qingjie and Wang, Yunhong},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  volume={25},
  number={10},
  pages={13863-13875},
  year={2024},
  publisher={IEEE}
}