UCAS-AOD-benchmark

A benchmark of UCAS-AOD dataset. (Only Oriented box is tested)

To be continued...

Introduction

There is no official division of the UCAS-AOD dataset, thus it's troublesome to compare the performance on different models. You can directly make comparison with our test results if you adopt the same division strategy.

Dataset repare

1. Download UCAS-AOD dataset .
2. Unzip dataset package into your root_dir, and rename the folder to UCAS_AOD.
3. Put our imageset files train.txt, val.txt and test.txt into ImageSets folder in UCAS_AOD.
4. Run data_prepare.py (modify the dataset dir to your own), and you will obtain directory as follow:

UCAS_AOD
└───AllImages
│   │   P0001.png
│   │   P0002.png
│   │	...
│   └───P1510.png
└───Annotations
│   │   P0001.txt
│   │   P0002.txt
│   │	...
│   └───P1510.txt       
└───ImageSets 
│   │   train.txt
│   │   val.txt
│   └───test.txt  
└───Test
│   │   P0003.png
│   │	...
│   └───P1508.txt 
└───CAR
└───PLANE
└───Neg

5. Train, eval and test you model according to ImageSets settings.

notes: The integrated dataset contains 1510 images, with train set 755, val set 302, test set 452(following DOTA division 5:2:3). Files are numbered from 1-1510, in which 1-510 are cars, 511-1510 are airplanes. Besides, classname is attached to label file in format of classname x1 y1 x2 y2 x3 y3 x4 y4 theta lx ly w h ,

for example:

car  2.763971e+02	9.125021e+01	2.911375e+02	3.823406e+01	3.308891e+02	4.928647e+01	3.161486e+02	1.023026e+02	1.055379e+02	2.787673e+02	3.876027e+01	4.975157e+01	6.301615e+01	
car  3.002141e+02	1.003123e+02	3.209637e+02	4.665470e+01	3.566901e+02	6.047021e+01	3.359405e+02	1.141279e+02	1.111416e+02	3.055889e+02	4.856245e+01	4.572642e+01	6.365764e+01	
...

Experiment

Environment

• NVIDIA 2080 Ti
• pytorch>1.1.0
• CUDA 10.0

Details

• Models are Trained on trainset , and test on testset, valset is used for parameter optimization.
• All models are available at Baidu Drive with passward sd4f.
• na denotes number of anchors preset at each location of feature maps.
• Data augment is adopted (random flip, hsv augment, translation, rotation).
• All models are evaluated via VOC07 metric.

Benchmark

model	backbone	input_size	na	car	airplane	mAP	paper link	remark
R-Yolov3	Darknet53	800*800	9	74.63	89.52	82.08	arxiv	code1, code2
R-RetinaNet	ResNet50	800*800	3	84.64	90.51	87.57	ICCV 2017	code
Faster RCNN	ResNet50	800*800	3	86.87	89.86	88.36	CVPR 2018	code
RoI Transformer	ResNet50	800*800	3	88.02	90.02	89.02	CVPR 2019	code
RIDet-Q	ResNet50	800*800	9	88.50	89.96	89.23	arxiv	code
CFC-Net	ResNet50	800*800	1	89.29	88.69	89.49	arxiv	code
RIDet-O	ResNet50	800*800	9	88.88	90.35	89.62	arxiv	code
DAL	ResNet50	800*800	3	89.25	90.49	89.87	AAAI 2021	code
S2ANet	ResNet50	800*800	1	89.56	90.42	89.99	TGRS	code

Some Results

---

Notes : More results and PRs are welcomed if you test with imagesets division here.