Offcial implementation of our AGOS framework for remote sensing scene classification.
This work has been published in IEEE Transactions on Geoscience and Remote Sensing entitled as All Grains, One Scheme (AGOS): Learning Multi-grain Instance Representation for Aerial Scene Classification.
The key idea of this prject is to extend the deep multiple instance learning into a multi-grain form. The multi-grain multi-instance learning is suitable to describe the varied object sizes and scale differences in remote sensing scenes. An technique framework of the proposed method is attached as follows.
The code is implemented on top of the Python3, and there are only a few dependencies that a development need to config. Before starting the code, please ensure the below packages and the corresponding versions are available.
Python > 3.5
Tensorflow > 1.6
OpenCV > 3
Numpy > 1.16
The datasets this paper use are all publicly available, and can be found in AID, UCM, and NWPU, respectively.
The ResNet-50 pre-trained model can be downloaded from here and is supposed to put into the checkpoint file folder.
For training:
Step 1, run the tfdata.py file to transfer the data into the tf.record file format.
python tfdata.py
Step 2, run the training.py file to start training.
python training.py
For testing:
Run the test1.py file to test the performance of a single model.
python test1.py
Run the testall.py file to test the performance of all the models in the checkpoints file folder.
python testall.py
Please note, before using the testall.py script, please remember to delete a file named checkpoint in the checkpoints file folder.
If you find this project useful, please cite:
@ARTICLE{Bi2022AGOS,
author={Bi, Qi and Zhou, Beichen and Qin, Kun and Ye, Qinghao and Xia, Gui-Song},
journal={IEEE Transactions on Geoscience and Remote Sensing},
title={All Grains, One Scheme (AGOS): Learning Multigrain Instance Representation for Aerial Scene Classification},
year={2022},
volume={60},
number={},
pages={1-17},
doi={10.1109/TGRS.2022.3201755}}
Other our former works related to Deep MIL may also be cited as:
@ARTICLE{Bi2021MIDCNet,
author={Bi, Qi and Qin, Kun and Li, Zhili and Zhang, Han and Xu, Kai and Xia, Gui-Song},
journal={IEEE Transactions on Image Processing},
title={A Multiple-Instance Densely-Connected ConvNet for Aerial Scene Classification},
year={2020},
volume={29},
pages={4911-4926},
doi={10.1109/TIP.2020.2975718}}
Qi Bi