Source code for testing the detection performance of DAGNet on infrared UAV targets. The codes are validated on a server with NVIDIA RTX 3090.

Software installation

Basic environmental setups:

• Python 3.9
• CUDA 11.1
• cuDNN 8.0.5

Python environmental setups:

• lxml==4.6.3
• matplotlib==3.4.1
• mmcv==1.7.0
• mmcv_full==1.6.0
• numpy==1.20.1
• opencv_python==4.5.4.60
• torch==1.8.1+cu111
• torchvision==0.9.1+cu111
• tqdm==4.59.0

For Python environmental setups, run pip install -r requirements.txt. Note that the mmcv and some other packages have been updated, as well as the CUDA and cuDNN. If you run into some environmental problems, make sure the environment are correctly configured.

Weight file

The pretrained backbone weight file for finetuning can be downloaded from Google Drive / Baidu Drive (passcode: 2pyz).

The testing weight file can be downloaded from Google Drive / Baidu Drive (passcode: code).

The weight file should be placed in the weights directory, and it should be created manually.

Open-access datasets

In our paper we utilize two infrared image sequences that are online available, which can be downloaded from Google Drive / BaiduYun (password: code). The file name indicates its corresponding sequence number. We organize these sequences of images into two folders, which are JPEGImages (contains image files in BMP format) and Annotations (contains annotation files in XML format).

Test images

The images for test purposes are given in the folder named test_imgs

Training

In train.py, you need to determine the following key configurations:

• dataset_root: Where your VOC2007 directory is. Give the full absolute path if necessary.
• basenet: The path to the pretrained backbone weight file. Should be placed in the weights directory.
• batch_size: Batch size for training. Could be any number that is a power of 2, based on GPU usage.
• resume: Whether to resume the training. If True, then the training code will automatically find the latest available intermediate weight file to resume.
• cuda: Whether to use GPU for training.
• learning_rate: Learning rate for training. Default is 1e-3. Could smaller it if NaN occurs.
• save_folder: The directory to save all weight files during training.

There are two places you need to pay attention to:

• data/config.py: There's one dictionary variable called voc_dagnet, and the key num_classes controls the number of classes, while lr_steps controls when the learning rate drops according to the iterations.
• data/voc0712.py: The cuple variable VOC_CLASSES defines what class names are. If there's only one class to detect, make sure you add a comma after the first string.

Note that, if you run into some problems with multi-GPU training, it might probably caused by torch.nn.DataParallel. You can use torch.distributed.launch or torch.distributed.run instead and try.

Testing

Simply run the following code in the console.

python detect.py --weights ./weights/test.pth --image_dir ./test_imgs/ --output ./detection/ --cuda True

If you wanna run the script with CUDA device, set --cuda True, otherwise --cuda False.

By running the script, the detection results of the sample images will be saved in the detection directory that will appear after finishing running the script. In each detection result image, the red box indicates the detected UAV target.

Citation

@article{fang2023dagnet,
  author={Fang, Houzhang and Liao, Zikai and Wang, Xuhua and Chang, Yi and Yan, Luxin},
  journal={IEEE Transactions on Industrial Informatics}, 
  title={Differentiated Attention Guided Network Over Hierarchical and Aggregated Features for Intelligent UAV Surveillance}, 
  year={2023},
  volume={19},
  number={9},
  pages={9909-9920},
  doi={10.1109/TII.2022.3232777}
}