A Dual Temporal Transformer Network With Multi-scale Contour Guided Graph Interaction For Change Detection

Here, we provide the pytorch implementation of the paper: A Dual Temporal Transformer Network With Multi-scale Contour Guided Graph Interaction For Change Detection

For more ore information, please see our published paper at

Overall Architecture

Contour-Guided Graph Interaction Module

Requirements

albumentations==1.3.1
numpy==1.21.5
opencv_python==4.8.1.78
opencv_python_headless==4.8.0.76
Pillow==9.1.0
scikit_learn==1.0.2
torch==1.7.1+cu110
torchvision>=0.8.2+cu110
wandb==0.15.8

1. Installation

Clone this repo:

git clone https://github.com/WesternTrail/DTT_CGINet.git
cd DTT_CGINet

2. Download the Datasets

We provide processed data sets and divided the train/val/test. All data sets are divided into 256 x 256 patches according to the method mentioned in the paper.

• LEVIR-CD256: baidu drive, code: r2ex

• BCDD256: baidu drive, code: bo65

• CDD: baidu drive, code: w2wf

Directory Structure

File Structure is as follows:

./
├─data
└─DTT_CGINet

data:
├─BCDD256
│  ├─test
│  │  ├─A
│  │  ├─B
│  │  └─label
│  ├─train
│  │  ├─A
│  │  ├─B
│  │  └─label
│  └─val
│      ├─A
│      ├─B
│      └─label
├─LEVIR-CD
└─CDD

3. Download the models (loading models):

• models code: c3tp

and put them into checkpoints directory.

4. Train

• The training process uses wandb to record logs, so you may need to set a proxy for the cmd command line before training. Please modify the following shell command appropriately:

   set http_proxy=http://127.0.0.1:7890 
   set https_proxy=http://127.0.0.1:7890

• After the proxy is set up, Then start training by:

   python train.py --data_path ${train_data_path} --val_path ${val_data_path} --lr ${lr} --batch_size ${-batch_size} 

5. test

• We evaluate the test set to obtain precision, recall and F1. When testing, please ensure that the mean and std of the corresponding data set are used during training. We use calc_dataset_mean_std.py to calculate the mean and std of the training set. Start testing by:

   python test.py --ckpt_url ${model_path} --data_path ${test_data_path}

6. predict

• We use this command to generate a change map. Similar to test, predict.py also needs to modify the mean and std corresponding to the data set. Start predicting by:

   python predict.py --ckpt_url ${model_path} --data_path ${test_data_path}

7. Citation

If our repo is useful to you, please cite our published paper as follow: