YOLOX-OBB

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

Installation

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

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

2. Install pycocotools

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

3. Install swig

   sudo apt-get install swig

4. Create the c++ extension for python

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

5. Install apex

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

   Data Preparation

6. Split images and annotations(You can refer to DOTA_devkit_YOLO)

   python DOTA_devkit_YOLO/ImgSplit_multi_process.py

   2. Transform annotations into voc-like format
   • This is a object in voc-like format annotation:

   python custom tools/DOTA2VOC_obb.py

   3. 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

7. Modify configs

　change the data path with yours in yolox_dota_s_obb_kld.py

data_dir = 'your_data_path'

2. 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

3. 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.

4. Merge results(You can refer to DOTA_devkit_YOLO)

   python DOTA_devkit_YOLO/ResultMerge.py

5. 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 mAP@0.5 on DOTA v1.0.

Reference

YOLOX

DOTA_devkit_YOLO

YOLOv5_DOTA_OBB

RotationDetection