Code for TGRS article 'ISNet: Towards Improving Separability for Remote Sensing Image Change Detection'.

---

Here I provide PyTorch implementations for ISNet and ISNet-lw.

Requirements

TITAN X
python 3.6.5
PyTorch 1.7.0

Installation

Clone this repo:

git clone https://github.com/xingronaldo/ISNet.git
cd ISNet/ISNet

• Install DCNv2

cd DCNv2
python setup.py build develop
cd ..

Attention: GTX/RTX series GPUs may fail to compile DCNv2. TITAN/Tesla series GPUs are recommended.

• Install other dependencies

All other dependencies can be installed via 'pip'.

Dataset Preparation

Download data and add them to ./datasets.

Note that

1. The data structure for the Season-Varying dataset has been already given in that folder. The LEVIR-CD dataset and the SYSU-CD dataset share the same data structure.

2. The instances in original LEVIR-CD dataset are cropped from 1024×1024 to 256×256.

Test

You can download our pretrained models for LEVIR-CD, SYSU-CD, and Season-Varying from Baidu Netdisk, code: tgrs, Baidu Netdisk, code: tgrs, and Baidu Netdisk, code: tgrs, respectively.

Then put them in ./checkpoints/LEVIR-CD/trained_models, ./checkpoints/SYSU-CD/trained_models, and ./checkpoints/SV/trained_models, separately.

• Test on the LEVIR-CD dataset

python test.py --dataset LEVIR-CD --name LEVIR-CD --load_pretrain True --which_epoch 255

• Test on the SYSU-CD dataset

python test.py --dataset SYSU-CD --name SYSU-CD --load_pretrain True --which_epoch 57

• Test on the Season-Varying dataset

python test.py --dataset SV --name SV --load_pretrain True --which_epoch 194

Train & Validation

python trainval.py --dataset SV --name SV 

All the hyperparameters can be adjusted in ./option.

logs:

1. During training, the occupied GPU memory is around 3357MB when batch size is 8, and around 4101MB when batch size is 16, on single TITAN X.

2. Time comparison for ISNet and ISNet-lw is given below.

Time Comparison

Supplement

You can download all predictions (in the form of the middle, below) of our ISNet for LEVIR-CD, SYSU-CD, and Season-Varying test sets from Baidu Netdisk, code: tgrs, Baidu Netdisk, code: tgrs, and Baidu Netdisk, code: tgrs, respectively.

To obtain marked predictions (in the form of the right, below) , use the code in ./ISNet/util/mark_prediction.py.

Label

Prediction

Marked Prediction

Citation

@article{Cheng2022ISNet,
title={ISNet: Towards Improving Separability for Remote Sensing Image Change Detection},
author={Cheng, Gong and Wang, Guangxing and Han, Junwei},
journal={IEEE Transactions on Geoscience and Remote Sensing},
volume={60},
number={},
pages={},
doi={10.1109/TGRS.2022.3174276},
year={2022}
}

Contact

Don't hesitate to contact me if you have any question.

Email: guangxingwang@mail.nwpu.edu.cn