Dual-Det : A Fast Detector for Oriented Object Detection in Aerial Images

The code is useful for DOTA, HRSC2016 and UCAS-AOD

How to get dataset?

• Dota: Dota is a large-scale dataset for object detection in aerial images. It can be used to develop and evaluate object detectors in aerial images. You can get the dataset via their home page.
• HRSC2016
• UCAS-AOD

Installation

0. Create a new conda environment and install pytorch v1.0+ and torchvision
1. Clone code

    git clone https://github.com/gqy4166000/DASR.git

2. Install the requirements

    pip install -r requirements.txt

3. Compile polyiou

    cd src
    sudo apt-get install swig
    swig -c++ -python polyiou.i
    python setup.py build_ext --inplace
    cd ..

4. Compile deformable convolutional

    cd src/lib/models/networks/DCNv2
    ./make.sh

Usage

Download the dataset and copy the partitioned data to the \data folder in the following format. For DOTA, images and labels need to be splited for use(by ImgSplit.py or ImgSplit_multi_process.py).

.
├── src
└── data
    ├── Dota1.0*
        ├── train_sp*
            ├──images
            └──labelTxt
        └── val_sp*
            ├──images
            └──labelTxt

*mean that you can change the folder name and the path name in the DOTA file must also be changed.

• Train

    python main.py --dataset dota --exp_id dota_train --gpus 0,1 --batch_size 32

• Val

    python main.py --dataset dota --exp_id dota_val --gpus 0 --test

You can adjust learning parameters in opt.py, and select single Angle, double Angle, and other branches in cfg.py.