wgcban / semicd Goto Github PK

Wele Gedara Chaminda Bandara, and Vishal M. Patel

📖 📖 📖 📖 View paper here.

🔖 🔖 🔖 View project page here.

This repocitory contains the official implementation of our paper: Revisiting Consistency Regularization for Semi-supervised Change Detection in Remote Sensing Images.

This repo was tested with Ubuntu 18.04.3 LTS, Python 3.8, PyTorch 1.1.0, and CUDA 10.0. But it should be runnable with recent PyTorch versions >=1.1.0.

The required packages are pytorch and torchvision, together with PIL and opencv for data-preprocessing and tqdm for showing the training progress.

conda create -n SemiCD python=3.8

conda activate SemiCD

pip3 install -r requirements.txt

We use two publicly available, widely-used CD datasets for our experiments, namely LEVIR-CD and WHU-CD. Note that LEVIR-CD and WHU-CD are building CD datasets.

As we described in the paper, following previous works ChangeFormer and BIT-CD on supervised CD, we create non-overlapping patches of size 256x256 for the training. The dataset preparation codes are written in MATLAB and can be found in dataset_preperation folder. These scripts will also generate the supervised and unsupervised training scripts that we used to train the model under diffrent percentage of labeled data.

Instead, you can directely download the processed LEVIR-CD and WHU-CD through the following links. Save these datasets anywhere you want and change the data_dir to each dataset in the corresponding config file.

The processed LEVIR-CD dataset, and supervised-unsupervised splits can be downloaded here.

The processed WHU-CD dataset, and supervised-unsupervised splits can be downloaded here.

To train a model, first download processed dataset above and save them in any directory you want, then set data_dir to the dataset path in the config file in configs/config_LEVIR.json/configs/config_WHU.json and set the rest of the parameters, like experim_name, sup_percent, unsup_percent, supervised, semi, save_dir, log_dir ... etc., more details below.

Then simply run:

python train.py --config configs/config_LEVIR.json

The following table summarizes the required changes in config file to train a model supervised or unsupervised with different percentage of labeled data.

Please follow the same changes that we outlined above to WHU-CD dataset as well. Then simply run:

python train.py --config configs/config_WHU.json

In this case we use LEVIR-CD as the supervised dataset and WHU-CD as the unsupervised dataset. Therefore, you need to update the train_supervised data_dir as the path to LEVIR-CD dataset, and train_unsupervised data_dir as the path to WHU-CD dataset in config_LEVIR-sup_WHU-unsup.json. Then change the sup_percent in the config file as you want and then simply run:

python train.py --config configs/config_LEVIR-sup_WHU-unsup.json

The log files and the .pth checkpoints will be saved in saved\EXP_NAME, to monitor the training using tensorboard, please run:

tensorboard --logdir saved

To resume training using a saved .pth model:

python train.py --config configs/config_LEVIR.json --resume saved/SemiCD/checkpoint.pth

Results: The results will be saved in saved as an html file, containing the validation results, and the name it will take is experim_name specified in configs/config_LEVIR.json.

For inference, we need a pretrained model, the pre-chage and pos-change imags that we wouldlike to dtet changes and the config used in training (to load the correct model and other parameters),

python inference.py --config config_LEVIR.json --model best_model.pth --images images_folder

Here are the flags available for inference:

--images       Folder containing the jpg images to segment.
--model        Path to the trained pth model.
--config       The config file used for training the model.

Pre-trained models can be downloaded from the following links.

Pre-trained models on LEVIR-CD can be downloaded from here.

Pre-trained models on WHU-CD can be downloaded from here.

Pre-trained models for cross-dataset experiments can be downloaded from here.

If you find this repo useful for your research, please consider citing the paper as follows:

@misc{bandara2022revisiting,
      title={Revisiting Consistency Regularization for Semi-supervised Change Detection in Remote Sensing Images}, 
      author={Wele Gedara Chaminda Bandara and Vishal M. Patel},
      year={2022},
      eprint={2204.08454},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

• This code is based on CCT.
• Code structure was based on Pytorch-Template
• ResNet backbone was downloaded from torchcv