Experimental Data for the Paper 'Hierarchical Fusion and Divergent Activation Based Weakly Supervised Learning for Object Detection from Remote Sensing Images'

Experimental Data for the Paper 'Hierarchical Fusion and Divergent Activation Based Weakly Supervised Learning for Object Detection from Remote Sensing Images'

In this repository, we provide the implementation of the algorithms developed in the paper 'Hierarchical Fusion and Divergent Activation Based Weakly Supervised Learning for Object Detection from Remote Sensing Images' along with the experimental results, and the methods used for comparison.
The goal is to provide the elements needed to validate and reproduce our research work as well as all the tools needed to reach the same conclusions as we did.
The data used in our experiments that we have the copyright of [A],[B] is already published on zenodo.
The licences valid for the elements of this repository are discussed under point "3. Licenses" below.

1. Structure

The repository contains the following items:

1. "CODE_AND_RESULTS.zip" with the source codes and results of our method and the comparison methods,
2. "README" - this text here.
3. "LICENSE" - the MIT License

We now focus on the structure of the file CODE_AND_RESULTS.zip.
It contains the following items:

1. The directory "new_methods" contains the source code and results of the new methods proposed in our paper.
2. The directory "comparison" contains the source code of the two approaches used for comparison: ACoL [A] and DANet [B].
3. The folder "tools_and_metrics" holds additional libraries, software tools, and metrics using in our experiments. 
4. "README" - this text here.
5. "LICENSE" - the MIT License

Inside the folder "new_methods," the following sub-folders are provided:

1. "data" includes data loading code and code for how organizing the input data of the neural network.
2. "expr" includes training code.
3. "model" includes neural network model, basic network and additional modules, depending on the file name, including improved network, and comparison model.
4. "utils" includes some used library functions and test codes when testing, including image segmentation, searching for the largest connected area and data visualization, etc. Verification on the WSADD dataset is done via test_airplane.py and on the DIOR dataset via val_model.py.

In our experiments, we used two datasets:

"WSADD" [A],[B], which is already published on zenodo under the Creative Commons Attribution 4.0 International license.
The "DIOR" proposed in [C].

2. References

[A] Z.-Z. Wu, T. Weise, Y. Wang, Y. Wang, Convolutional neural network based weakly supervised learning for aircraft detection from remote sensing image, IEEE Access 8 (2020) 158097-158106. doi:10.1109/ACCESS.2020.3019956.   
[B] Z.-Z. Wu. Weakly Supervised Airplane Detection Dataset: WSADD. May 2020. zenodo.org. doi:10.5281/zenodo.3843229.
[C] K. Li, G. Wan, G. Cheng, L. Meng, J. Han, Object detection in optical remote sensing images: A survey and a new benchmark, ISPRS Journal of Photogrammetry and Remote Sensing 159 (2020) 296-307. doi:10.1016/j.isprsjprs.2019.11.023.   
[D] X. Zhang, Y. Wei, J. Feng, Y. Yang, T. S. Huang, Adversarial complementary learning for weakly supervised object localization, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR'18), Jun. 18-22, 2018, Salt Lake City, UT, USA, IEEE Computer Society, 2018, pp. 1325-1334. doi:10.1109/CVPR.2018.00144.   
[E] H. Xue, C. Liu, F. Wan, J. Jiao, X. Ji, Q. Ye, DANet: Divergent activation for weakly supervised object localization, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV'19), Oct. 27-Nov. 2, 2019, Seoul, Korea, IEEE, 2019, pp. 6588-6597. doi:10.1109/ICCV.2019.00669.

3. Licenses

The following licenses apply for the files and folders in the archive "CODE_AND_RESULTS.zip":

• The files in the folder `new_methods` are under the MIT License.
• The files in the folder `comparison/ACoL` have been obtained from https://github.com/xiaomengyc/ACoL, which is under the MIT License.
• We put our code and data under the 
• The files in the folder "comparison/DANet" have been obtained from https://github.com/xuehaolan/DANet, which is an open source project without associated license at the time of this writing. They will therefore remain under the copyright of the user https://github.com/xuehaolan/.
• The files in the folder "tools_and_metrics/detections_DIOR" are related to the repository https://github.com/rafaelpadilla/Object-Detection-Metrics, which is under the MIT License, and therefore are under the same license.
• The files in the folder "tools_and_metrics/Nest-pytorch" are based on the repository https://github.com/ZhouYanzhao/Nest, which is under the MIT License.
• The files in the folder "tools_and_metrics/PRM-pytorch" are based on the repository https://github.com/ZhouYanzhao/PRM, which is an open source project without associated license at the time of this writing. They will therefore remain under the copyright of the user https://github.com/ZhouYanzhao/.

The MIT License is included here as file "LICENSE".

4. Contact

1. Dr. Zhize WU, wuzz@hfuu.edu.cn
2. Dr. Thomas WEISE, tweise@hfuu.edu.cn, tweise@ustc.edu.cn

Institute of Applied Optimization,   
School of Artificial Intelligence and Big Data,   
Hefei University, South Campus 2, Jinxiu Dadao 99,   
Hefei Economic and Technological Development Area,   
Shushan District, Hefei 230601, Anhui, China