Dynamic Anchor Learning for Arbitrary-Oriented Object Detection

Abstract

To be updated.

Basic Information

The codes build RetinaNet with the proposed DAL method for rotation object detection. The supported datasets include:

• DOTA
• HRSC
• ICDAR2013
• ICDAR2015
• UCAS-AOD
• NWPU VHR-10
• VOC

Performance

Note that we use only 3 horizontal perset anchors at each location on feature map for rotation detection(while 5 for IC15) . This implementation reaches 24 fps on RTX 2080 Ti.

HRSC2016

Note that VOC07 metric is used for evaluation.

Method	Backbone	Input Size	mAP
RetinaNet	ResNet-50	416 x 416	80.81
RetinaNet + DAL	ResNet-50	416 x 416	88.60
RetinaNet + DAL	ResNet-101	416 x 416	88.95
RetinaNet + DAL	ResNet-101	800 x 800	89.77

UCAS-AOD

Refer to this repo.

ICDAR 2015

The performance for long text detection is not good enough. The submissions are shown in the official website: Incidental Scene Text 2015.

Method	$P$	$R$	$F_1$
RetinaNet	77.2	77.8	77.5
RetinaNet + DAL	83.7	79.5	81.5
RetinaNet + DAL(ms)	84.4	80.5	82.4

DOTA

Method	Backbone	mAP
RetinaNet	ResNet-50	68.43
RetinaNet + DAL	ResNet-50	71.44
RetinaNet + DAL	ResNet-101	71.78
S2A-Net	ResNet-50	74.12
S2A-Net + DAL	ResNet-50	76.95

Experiments on DOTA are implemented based on mmdetection, since the recognition ability of some classes(TC, BC, GTF) is abnormally poor. Trained models are available here.

Getting Started

Installation

Build the Cython and CUDA modules:

cd $ROOT/utils
sh make.sh
cd $ROOT/utils/overlaps_cuda
python setup.py build_ext --inplace

Installation for DOTA_devkit:

cd $ROOT/DOTA_devkit
sudo apt-get install swig
swig -c++ -python polyiou.i
python setup.py build_ext --inplace

Inference

You can use the following command to test a dataset. Note that weight, img_dir, dataset,hyp should be modified as appropriate.

python demo.py

Train

1. Move the dataset to the $ROOT directory.
2. Generate imageset files for daatset division via:

cd $ROOT/datasets
python generate_imageset.py

3. Modify the configuration file hyp.py and arguments in train.py, then start training:

python train.py

Evaluation

Different datasets use different test methods. For UCAS-AOD/HRSC2016/VOC/NWPU VHR-10, you need to prepare labels in the appropriate format in advance. Take evaluation on HRSC2016 for example:

cd $ROOT/datasets/evaluate
python hrsc2gt.py

then you can conduct evaluation:

python eval.py

Note that :

• the script needs to be executed only once, but testing on different datasets needs to be executed again.
• the imageset file used in hrsc2gt.py is generated from generate_imageset.py.

Detection Results

Citations

To be updated.