Rotation RetinaNet

The codes build R-RetinaNet for arbitrary-oriented object detection. It supports the following datasets: DOTA, HRSC2016, ICDAR2013, ICDAR2015, UCAS-AOD, NWPU VHR-10, VOC2007.

Performance

This implementation reaches 24 fps on RTX 2080 Ti.

The performance is shown as follow:

HRSC2016

Note that VOC07 metric is used for evaluation.

Dataset	Bbox	Backbone	Input size	mAP
HRSC2016	OBB	ResNet-50	416 x 416	80.81
UCAS-AOD	OBB	ResNet-50	800 x 800	87.57
ICDAR 2013	OBB	ResNet-50	800 x 800	77.20
ICDAR 2015	HBB	ResNet-50	800 x 800	77.50
NWPU VHR-10	HBB	ResNet-50	800 x 800	86.40

Note that VOC07 metric is used for HRSC2016, UCAS-AOD, NWPU VHR-10, F1 score for IC13 and IC13. 3 anchors are preset for each location in feature maps. All experiments are conducted with data augmentation including random flip, rotation and HSV color space transform.

Getting Started

Installation

Build the Cython and CUDA modules:

cd $ROOT/utils
sh make.sh

Install DotaDevKit:

sudo apt-get install swig
swig -c++ -python polyiou.i
python setup.py build_ext --inplace

Inference

python demo.py

Train

1. prepare dataset and move it into the $ROOT directory.
2. generate imageset files:

cd $ROOT/datasets
python generate_imageset.py

3. start training:

python train.py

Evaluation

prepare labels, take hrsc for example:

cd $ROOT/datasets/evaluate
python hrsc2gt.py

start evaluation:

python eval.py

Note that :

• the script needs to be executed only once.
• the imageset file used in hrsc2gt.py is generated from generate_imageset.py.

Detection Result