AProNet: Detecting Objects with Precise Orientation from Aerial Images

This is the repository of paper "AProNet: Detecting Objects with Precise Orientation from Aerial Images"

Installation

Requirements

Python: 3.6  
PyTorch: 1.2.0
CUDA: 9.2    
CUDNN: 7.6.2  

Installation

a. Create a conda virtual environment and activate it.

conda create --name AProNet python=3.6 -y  
conda activate AProNet  

b. Install PyTorch.

conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=9.2 -c pytorch

c. Clone the geovision-AProNet repository

git clone https://github.com/ZhWL123456/geovision-AProNet.git
workdir=./geovision-AProNet

d. Compile the poly_nms:

cd ${workdir}/maskrcnn_benchmark/utils/poly_nms  
python setup.py build_ext --inplace  

d. Compile the DOTA-devikit dependency:

sudo apt-get install swig  
cd ${workdir}/maskrcnn_benchmark/DOTA_devkit/polyiou  
swig -c++ -python csrc/polyiou.i  
python setup.py build_ext --inplace  

Setting the datasets

a. Prepare your dataset as the status format.
This project use the json annotation file with COCO format. Make your directory layout like this:

${data-dir}
└── trainset
    ├── images
    │   ├── 1.png
    │   └── 2.png
    └── labelTxt
        ├── 1.txt
        └── 2.txt

A example of the *.txt files ('1' means the object is difficult):

x1 y1 x2 y2 x3 y3 x4 y4 plane 0
x1 y1 x2 y2 x3 y3 x4 y4 harbor 1

Run the following Python snippet, and it will generate the json annotation file:

from txt2json import collect_unaug_dataset, convert
img_dic = collect_unaug_dataset( os.path.join( "trainset", "labelTxt" ) )
convert( img_dic, "trainset",  os.path.join( "trainset", "train.json" ) )

b. Edit the file maskrcnn_benchmark/config/paths_catalog.py (from line7 to 17) to set the dir of datasets.

    DATA_DIR = "datasets" #need to change
    DATASETS = {
        "dota_trainval_cut": {
            "img_dir": "${dataset}/trainval_cut/images",             #need to change
            "ann_file": "${dataset}/trainval_cut/trainval_cut.json"  #need to change
        },
        "dota_test_cut": {
            "img_dir": "${dataset}/test_cut/images",                 #need to change
            "ann_file": "${dataset}/test_cut/test_cut.json"          #need to change
        },
    }

c. If your dataset is DOTA(options):
For DOTA, you need to run the scripts XX to split the original images into chip images (e.g., 1024*1024), and convert annotations to mmdet's format.