YOLOX-OBB

YOLOX in DOTA with KLD loss. (Oriented Object Detection)（Rotated BBox）基于YOLOX的旋转目标检测

Installation

Install YOLOX-OBB(You can refer to the installation of YOLOX)

cd YOLOX-OBB
pip3 install -r requirements.txt
pip3 install -v -e

Install pycocotools

pip3 install cython; pip3 install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

Install swig

sudo apt-get install swig

Create the c++ extension for python

cd DOTA_devkit_YOLO
swig -c++ -python polyiou.i
python setup.py build_ext --inplace
cd -

Install apex

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --no-cache-dir ./
cd -

Data Preparation

Split images and annotations(You can refer to DOTA_devkit_YOLO)

 python DOTA_devkit_YOLO/ImgSplit_multi_process.py

Transform annotations into voc-like format

This is a object in voc-like format annotation:

python custom tools/DOTA2VOC_obb.py

Organize Directories(All annotations of train-images and val-images must be put into Annotations folder)

|--your_data
    |--VOC2012
        |--Annotations
            |-- xxx.xml
                ... 
        |--ImageSets
            |--Main
                |--train.txt
                |--val.txt
                |--test.txt
        |--JPEGImages
        |--JPEGImages-val
        |-JPEGImages-test

Train

Modify configs

　change the data path with yours in yolox_dota_s_obb_kld.py

data_dir = 'your_data_path'

Train

CUDA_VISIBLE_DEVICES=0,1 python3 tools/train.py -f exps/example/yolox_voc/yolox_dota_s_obb_kld.py -d 2 -b 16 --fp16 -c weights/yolox_s.pth.tar

Val

get results

CUDA_VISIBLE_DEVICES=0,1 python tools/eval.py -f exps/example/yolox_voc/yolox_dota_s_obb_kld.py -d 2 -b 16 -c YOLOX_outputs/yolox_dota_s_obb_kld/latest_ckpt.pth

　Results will be save to your_data/results/VOC2012/Main

If test, you must comment line 151 'target = self.load_anno(index)' and uncomment line 152 'target = []' in dota_obb.py before run the above instruction. Because test-set has no annotations.

Merge results(You can refer to DOTA_devkit_YOLO)

python DOTA_devkit_YOLO/ResultMerge.py

Evaluation

python DOTA_devkit_YOLO/dota_v1.5_evaluation_task1.py(You can refer to [DOTA_devkit_YOLO](https://github.com/hukaixuan19970627/DOTA_devkit_YOLO))

If test, you should upload your results to DOTA Evaluation Server.

Unfortunately

This implementation only get 0.712 [email protected] on DOTA v1.0.

Reference

Cheap and reliable Node.js hosting starts at $3/month, and $1/month static HTML hosting

buzhidaoshenme / YOLOX-OBB