R2CNN-Plus-Plus_Tensorflow

Abstract

This is a tensorflow implementation of R²CNN++: Multi-Dimensional Attention Based Rotation Invariant Detector with Robust Anchor Strategy.
This project is based on R²CNN, and completed by YangXue and YangJirui.

DOTA test results

Comparison

Part of the results are from DOTA paper.

Task1 - Oriented Leaderboard

Task2 - Horizontal Leaderboard

Requirements

1、tensorflow >= 1.2
2、cuda8.0
3、python2.7 (anaconda2 recommend)
4、opencv(cv2)
5、tfplot

Download Model

1、please download resnet50_v1、resnet101_v1 pre-trained models on Imagenet, put it to data/pretrained_weights.
2、please download mobilenet_v2 pre-trained model on Imagenet, put it to data/pretrained_weights/mobilenet.
3、please download trained model by this project, put it to output/trained_weights.

Data Prepare

1、please download DOTA
2、crop data, reference:

cd $PATH_ROOT/data/io/DOTA
python train_crop.py 
python val_crop.py

3、data format

├── VOCdevkit
│   ├── VOCdevkit_train
│       ├── Annotation
│       ├── JPEGImages
│    ├── VOCdevkit_test
│       ├── Annotation
│       ├── JPEGImages

Compile

cd $PATH_ROOT/libs/box_utils/
python setup.py build_ext --inplace

cd $PATH_ROOT/libs/box_utils/cython_utils
python setup.py build_ext --inplace

Demo

Select a configuration file in the folder (libs/configs/) and copy its contents into cfgs.py, then download the corresponding weights.

DOTA

python demo_rh.py (demo_rh_pyramid.py) --src_folder='/PATH/TO/DOTA/IMAGES_ORIGINAL/' 
                                       --image_ext='.png' 
                                       --des_folder='/PATH/TO/SAVE/RESULTS/' 
                                       --save_res=False
                                       --gpu='0'

Eval

python eval.py --img_dir='/PATH/TO/DOTA/IMAGES/' 
               --image_ext='.png' 
               --test_annotation_path='/PATH/TO/TEST/ANNOTATION/'
               --gpu='0'

Inference

python inference.py --data_dir='/PATH/TO/DOTA/IMAGES_CROP/'      
                    --gpu='0'

Train

1、If you want to train your own data, please note:

(1) Modify parameters (such as CLASS_NUM, DATASET_NAME, VERSION, etc.) in $PATH_ROOT/libs/configs/cfgs.py
(2) Add category information in $PATH_ROOT/libs/label_name_dict/lable_dict.py     
(3) Add data_name to $PATH_ROOT/data/io/read_tfrecord.py 

2、make tfrecord

cd $PATH_ROOT/data/io/  
python convert_data_to_tfrecord.py --VOC_dir='/PATH/TO/VOCdevkit/VOCdevkit_train/' 
                                   --xml_dir='Annotation'
                                   --image_dir='JPEGImages'
                                   --save_name='train' 
                                   --img_format='.png' 
                                   --dataset='DOTA'

3、train

cd $PATH_ROOT/tools
python train.py

Tensorboard

cd $PATH_ROOT/output/summary
tensorboard --logdir=.

Citation

Some relevant achievements based on this code.

@inproceedings{yang2019scrdet,
        title={SCRDet: Towards More Robust Detection for Small, Cluttered and Rotated Objects},
        author={Yang, Xue and Yang, Jirui and Yan, Junchi and Zhang, Yue and Zhan, Tengfei and Guo, Zhi and Xian, Sun and Fu, Kun},
        booktitle={Proc. ICCV},
        year={2019}
}

@article{yang2018position,
        title={Position Detection and Direction Prediction for Arbitrary-Oriented Ships via Multitask Rotation Region Convolutional Neural Network},
        author={Yang, Xue and Sun, Hao and Sun, Xian and  Yan, Menglong and Guo, Zhi and Fu, Kun},
        journal={IEEE Access},
        volume={6},
        pages={50839-50849},
        year={2018},
        publisher={IEEE}
}

@article{yang2018r-dfpn,
        title={Automatic ship detection in remote sensing images from google earth of complex scenes based on multiscale rotation dense feature pyramid networks},
        author={Yang, Xue and Sun, Hao and Fu, Kun and Yang, Jirui and Sun, Xian and Yan, Menglong and Guo, Zhi},
        journal={Remote Sensing},
        volume={10},
        number={1},
        pages={132},
        year={2018},
        publisher={Multidisciplinary Digital Publishing Institute}
}