TGRS-HRRSD-Dataset: High Resolution Remote Sensing Detection (HRRSD)

I. NOTE: JPEG files are available on BaiduCloud, GoogleDrive, and an ipv6 site bt.byr.cn

• HRRSD contains 21,761 images acquired from Google Earth and Baidu Map with the spatial resolution from 0.15-m to 1.2-m.

• There are 55,740 object instances in HRRSD.

• HRRSD contains 13 categories of RSI objects.

Moreover, this dataset is divided as several subsets, image numbers in each subset are 5401 for ‘train’, 5417 for ‘val’, and 10943 for ‘test’. And ‘train-val’ subset is a merge of ‘train’ and ‘val’.

II. Mean and Std

In most current object detection systems, means and std values of datasets are required.

You may refer to:

mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375]

Moreover, you can compute the values by yourself with file OPT2017/avr_std_detection_sets.py:

$ cd your_HRRSD_path/OPT2017
$ python avr_std_detection_sets.py 500

III. Folders

Labels

• /OPT2017/Annotations: *.xml
• /OPT2017/labels: *.txt with the form of (class x y width height)

Images

• /OPT2017/JPEGImages: *.jpg

Dataset Division

• /OPT2017/ImageSets/Main: Division of the dataset.

IV. Statistics and Benchmark

Statistics

In this table, N_* refers to numbers of objects. 'Train', 'Val', 'Test' are three subsets of the dataset. 'Mean Resized Scale' shows average scale of each category. 'Resized Scale Std' is the standard deviation of category scale.

Benchmark

GACL-Net \cite{lu2020gated} is a method proposed to improve object localization performance. Title of this paper is "Gated and Axis-Concentrated Localization Network for Remote Sensing Object Detection". The suffix "r50" denotes using 50 layer Resnet.

V. FAQ

If any question is met, please contanct me with the e-mail: [email protected].

Qestion 1: AP for the "T junction" class is always NAN or 0, why?

Anwser Q1: In some object detection frameworks, there may be a piece of code like "cls_names = lower( cls_names )". This will set class names to lower case, but class names in xml files contain "T junction" where "T" is uppercase. This actually will cause several problems. The solution is using debug sofwares to find the code of changing word cases and correct it. For the dataset, I won't change the "T junction" labels in xmls currently for lacking time.

VI. Citation

If you find HRRSD dataset useful in your research, please consider citing:

@article{zhang2019hierarchical,
  title={Hierarchical and Robust Convolutional Neural Network for Very High-Resolution Remote Sensing Object Detection},
  author={Zhang, Yuanlin and Yuan, Yuan and Feng, Yachuang and Lu, Xiaoqiang},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  volume={57},
  number={8},
  pages={5535--5548},
  year={2019},
  publisher={IEEE}
}

For more comparative experimental results, please refer to:

@article{lu2020gated,
  title={Gated and Axis-Concentrated Localization Network for Remote Sensing Object Detection},
  author={Lu, Xiaoqiang and Zhang, Yuanlin and Yuan, Yuan and Feng, Yachuang},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  volume={58},
  number={1},
  pages={179--192},
  year={2020},
  publisher={IEEE}
}