Awesome

CEASC: Adaptive Sparse Convolutional Networks with Global Context Enhancement for Faster Object Detection on Drone Images

The repo is the official implementation of CEASC.

Our CEASC module is at mmdet/models/dense_heads

Our Sparse Convolution Implementation is at Sparse_conv

Our config file is at configs/UAV

Requirement

Please follow docs/en/get_started.md and install the mmdetection toolbox.

a. Install Pytorch 1.10.1

b. Install MMDetection toolbox, required mmdet >= 2.7.0, mmcv-full >= 1.4.2.

• Our project utilizes mmdet == 2.24.1, mmcv-full == 1.5.1

c. Install albumentations and other packages.

pip install nltk
pip install -r requirements/albu.txt

d. Install our Sparse Convolution Implementation

cd ./Sparse_conv
python setup.py install
cd ..

Usage

1. Data preparation

You could download VisDrone and UAVDT dataset (COCO Format) from official links or from other repositories like UFPMP-Det.

2. Training

% training on a single GPU
python tools/train.py /path/to/config-file --work-dir /path/to/work-dir

% training on multi GPUs
bash tools/dist_train.sh /path/to/config-file num-gpus --work-dir /path/to/work-dir

Checkpoints:

We provide the following checkpoints:

• GFL v1 baseline, corresponding to baseline_gfl_res18_visdrone: Google Drive
• GFL v1 CEASC, corresponding to dynamic_gfl_res18_visdrone: Google Drive
• RetinaNet baseline, corresponding to baseline_retinanet_res18_visdrone: Google Drive
• RetinaNet CEASC, corresponding to dynamic_retinanet_res18_visdrone: Google Drive

3. Test

python tools/test.py /path/to/config-file /path/to/work-dir/latest.pth --eval bbox

Citation

If you find our paper or this project helps your research, please kindly consider citing our paper in your publication.

@misc{ceasc,
      title={Adaptive Sparse Convolutional Networks with Global Context Enhancement for Faster Object Detection on Drone Images}, 
      author={Bowei Du and Yecheng Huang and Jiaxin Chen and Di Huang},
      year={2023},
      eprint={2303.14488},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}