Cross-Layer Feature Pyramid Transformer for Small Object Detection in Aerial Images

This repository provides the official PyTorch implementation of CFPT.

In this paper, we propose the cross-layer feature pyramid transformer designed for small object detection in aerial images.

Below is the performance comparison with other feature pyramid networks based on RetinaNet on the VisDrone-2019 DET dataset.

The architecture of CFPT is as described below.

Weights

Due to the accidental deletion of the model weights prepared for this paper, we retrained the entire network, resulting in slight differences in performance metrics compared to the original study.

Installation

Our experiments are based on torch 1.10+cu113, mmdet 2.24.1 and mmcv-full 1.6.0.

Please see get_started.md for the basic usage of MMDetection.

1. Install PyTorch.

2. Install mmcv-full and MMDetection toolbox.

   pip install openmim
   mim install mmcv-full==1.6.0
   

3. Install albumentations and other packages.

   pip install einops
   pip install timm
   pip install yapf==0.40.1
   pip install albumentations==1.1.0
   

4. Clone and install this repository.

   git clone https://github.com/duzw9311/CFPT.git
   cd ./CFPT
   pip install -e .
   

Usage

Data Preparation

Download the VisDrone2019-DET dataset converted to COCO annotation format. You can download it from this link.

Training

python tools/train.py configs/CFPT/retinanet_r18_cfpt_1x_visdrone.py 

Testing

python tools/test.py configs/CFPT/retinanet_r18_cfpt_1x_visdrone.py work_dirs/retinanet_r18_cfpt_1x_visdrone/latest.pth --eval bbox

Acknowledgement

This repository is built upon the MMDetection library. Thanks to the authors of CEASC and other researchers in the field of object detection for their open-source code.

Citation

If you find this paper helpful for your project, we'd appreciate it if you could cite it.

@article{du2024cross,
  title={Cross-Layer Feature Pyramid Transformer for Small Object Detection in Aerial Images},
  author={Du, Zewen and Hu, Zhenjiang and Zhao, Guiyu and Jin, Ying and Ma, Hongbin},
  journal={arXiv preprint arXiv:2407.19696},
  year={2024}
}