English | 简体中文
The repository is the code implementation of the paper Time Travelling Pixels: Bitemporal Features Integration with Foundation Model for Remote Sensing Image Change Detection, based on MMSegmentation and Open-CD projects.
The current branch has been tested under PyTorch 2.x and CUDA 12.1, supports Python 3.7+, and is compatible with most CUDA versions.
If you find this project helpful, please give us a star ⭐️, your support is our greatest motivation.
Main Features
- Consistent API interface and usage with MMSegmentation
- Open sourced the TTP model in the paper
- Tested with AMP training method
- Supports multiple dataset extensions
🌟 2023.12.23 Released the TTP project code, which is completely consistent with the API interface and usage of MMSegmentation.
🌟 2023.12.30 Released the model trained on Levir-CD.
🌟 2024.02.10 This project has been included in the Open-CD project.
- Introduction
- Update Log
- Table of Contents
- Installation
- Dataset Preparation
- Model Training
- Model Testing
- Image Prediction
- FAQ
- Acknowledgements
- Citation
- License
- Contact Us
- Linux or Windows
- Python 3.7+, recommended 3.10
- PyTorch 2.0 or higher, recommended 2.1
- CUDA 11.7 or higher, recommended 12.1
- MMCV 2.0 or higher, recommended 2.1
We recommend using Miniconda for installation. The following command will create a virtual environment named ttp
and install PyTorch and MMCV.
Note: If you have experience with PyTorch and have already installed it, you can skip to the next section. Otherwise, you can follow these steps to prepare.
Step 0: Install Miniconda.
Step 1: Create a virtual environment named ttp
and activate it.
conda create -n ttp python=3.10 -y
conda activate ttp
Step 2: Install PyTorch2.1.x.
Linux/Windows:
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu121
Or
conda install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda=12.1 -c pytorch -c nvidia
Step 3: Install MMCV2.1.x.
pip install -U openmim
mim install mmcv==2.1.0
Step 4: Install other dependencies.
pip install -U wandb einops importlib peft==0.8.2 scipy ftfy prettytable torchmetrics==1.3.1 transformers==4.38.1
Download or clone the TTP repository.
git clone git@github.com:KyanChen/TTP.git
cd TTP
- Image and label download address: Levir-CD.
You can also choose other sources to download the data, but you need to organize the dataset in the following format:
${DATASET_ROOT} # Dataset root directory, for example: /home/username/data/levir-cd
├── train
│ ├── A
│ ├── B
│ └── label
├── val
│ ├── A
│ ├── B
│ └── label
└── test
├── A
├── B
└── label
Note: In the project folder, we provide a folder named data
, which contains an example of the organization method of the above dataset.
If you want to use other datasets, you can refer to MMSegmentation documentation to prepare the datasets.
We provide the configuration files of the TTP model used in the paper, which can be found in the configs/TTP
folder. The Config file is completely consistent with the API interface and usage of MMSegmentation. Below we provide an analysis of some of the main parameters. If you want to know more about the meaning of the parameters, you can refer to MMSegmentation documentation.
Parameter Parsing:
work_dir
: The output path of the model training, which generally does not need to be modified.default_hooks-CheckpointHook
: Checkpoint saving configuration during model training, which generally does not need to be modified.default_hooks-visualization
: Visualization configuration during model training, comment out during training and uncomment during testing.vis_backends-WandbVisBackend
: Configuration of network-side visualization tools, after opening the comment, you need to register an account on thewandb
official website, and you can view the visualization results during the training process in the network browser.sam_pretrain_ckpt_path
: The checkpoint path of the SAM backbone provided by MMPretrain, refer to download address.model-backbone-peft_cfg
: Whether to introduce fine-tuning parameters, which generally does not need to be modified.dataset_type
: The type of dataset, needs to be modified according to the type of dataset.data_root
: Dataset root directory, modify to the absolute path of the dataset root directory.batch_size_per_gpu
: The batch size of a single card, needs to be modified according to the memory size.resume
: Whether to resume training, which generally does not need to be modified.load_from
: The checkpoint path of the model's pre-training, which generally does not need to be modified.max_epochs
: The maximum number of training rounds, which generally does not need to be modified.
python tools/train.py configs/TTP/xxx.py # xxx.py is the configuration file you want to use
sh ./tools/dist_train.sh configs/TTP/xxx.py ${GPU_NUM} # xxx.py is the configuration file you want to use, GPU_NUM is the number of GPUs used
If you want to use other change detection models, you can refer to Open-CD to train the models, or you can put their Config files into the configs
folder of this project, and then train them according to the above method.
python tools/test.py configs/TTP/xxx.py ${CHECKPOINT_FILE} # xxx.py is the configuration file you want to use, CHECKPOINT_FILE is the checkpoint file you want to use
python tools/test.py configs/TTP/xxx.py ${CHECKPOINT_FILE} --show_dir ${SAVE_OUTPUT_DIR} --draw_gt # xxx.py is the configuration file you want to use, CHECKPOINT_FILE is the checkpoint file you want to use, SAVE_OUTPUT_DIR is the directory of prediction label, draw_gt allows to draw ground ture label compared with prediction label
sh ./tools/dist_test.sh configs/TTP/xxx.py ${CHECKPOINT_FILE} ${GPU_NUM} # xxx.py is the configuration file you want to use, CHECKPOINT_FILE is the checkpoint file you want to use, GPU_NUM is the number of GPUs used
Note: If you need to get the visualization results, you can uncomment default_hooks-visualization
in the Config file.
python demo/image_demo_with_cdinferencer.py ${IMAGE_FILE1} ${IMAGE_FILE2} configs/TTP/ttp_sam_large_levircd_infer.py --checkpoint ${CHECKPOINT_FILE} --out-dir ${OUTPUT_DIR} # IMAGE_FILE is the image file you want to predict, xxx.py is the configuration file, CHECKPOINT_FILE is the checkpoint file you want to use, OUTPUT_DIR is the output path of the prediction result
We have listed some common problems and their corresponding solutions here. If you find that some problems are missing, please feel free to provide a PR to enrich this list. If you cannot get help here, please use issue to seek help. Please fill in all the required information in the template, which will help us locate the problem faster.
We recommend that you do not install them, because we have partially modified their code, which may cause errors in the code if you install them. If you get an error that the module has not been registered, please check:
- Whether these libraries are installed, if so, uninstall them
- Whether
@MODELS.register_module()
is added in front of the class name, if not, add it - Whether
from .xxx import xxx
is added in__init__.py
, if not, add it - Whether
custom_imports = dict(imports=['mmseg.ttp'], allow_failed_imports=False)
is added in the Config file, if not, add it
Here we list the resource consumption of using different training methods for your reference.
Model Name | Backbone Type | Image Size | GPU | Batch Size | Acceleration Strategy | Single Card Memory Usage | Training Time |
---|---|---|---|---|---|---|---|
TTP | ViT-L/16 | 512x512 | 4x RTX 4090 24G | 2 | FP32 | 14 GB | 3H |
TTP | ViT-L/16 | 512x512 | 4x RTX 4090 24G | 2 | FP16 | 12 GB | 2H |
If you get a Bad substitution
error when running dist_train.sh
, use bash dist_train.sh
to run the script.
5. You should set PYTHONPATH
to make sys.path
include the directory which contains your custom module
Please check the detailed error message, generally some dependent packages are not installed, please use pip install
to install the dependent packages.
The repository is the code implementation of the paper Time Travelling Pixels: Bitemporal Features Integration with Foundation Model for Remote Sensing Image Change Detection, based on MMSegmentation and Open-CD projects.
If you use the code or performance benchmarks of this project in your research, please refer to the following bibtex to cite TTP.
@misc{chen2023time,
title={Time Travelling Pixels: Bitemporal Features Integration with Foundation Model for Remote Sensing Image Change Detection},
author={Keyan Chen and Chengyang Liu and Wenyuan Li and Zili Liu and Hao Chen and Haotian Zhang and Zhengxia Zou and Zhenwei Shi},
year={2023},
eprint={2312.16202},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
The repository is licensed under the Apache 2.0 license.
If you have other questions❓, please contact us in time 👬