Change Agent Download - Change Agent Source code download

Change Agent

Other source code

Download

Change-Agent: Toward Interactive Comprehensive Remote Sensing Change Interpretation and Analysis

Chenyang Liu, Keyan Chen, Haotian Zhang, Zipeng Qi, Zhengxia Zou, and Zhenwei Shi*✉

Share us a if you're interested in this repo

Official PyTorch implementation of the paper: "Change-Agent: Toward Interactive Comprehensive Remote Sensing Change Interpretation and Analysis" in [IEEE] (Accepted by IEEE TGRS 2024)

News

2024-06: The code is available.
2024-03: The paper is available.

LEVIR-MCI dataset

Download the LEVIR_MCI dataset: LEVIR-MCI (Available Now!).
This dataset is an extension of our previously established LEVIR-CC dataset. It contains bi-temporal images as well as diverse change detection masks and descriptive sentences. It provides a crucial data foundation for exploring multi-task learning for change detection and change captioning.

Training of the multi-level change interpretation model

The overview of the MCI model:

Preparation

Environment Installation:

Step 1: Create a virtual environment named Multi_change_env and activate it.

conda create -n Multi_change_env python=3.9
conda activate Multi_change_env

Step 2: Download or clone the repository.

git clone https://github.com/Chen-Yang-Liu/Change-Agent.git
cd ./Change-Agent/Multi_change

Step 3: Install dependencies.

pip install -r requirements.txt

Download Dataset:

Link: LEVIR-MCI. The data structure of LEVIR-MCI is organized as follows:

├─/DATA_PATH_ROOT/Levir-MCI-dataset/
        ├─LevirCCcaptions.json
        ├─images
             ├─train
             │  ├─A
             │  ├─B
             │  ├─label
             ├─val
             │  ├─A
             │  ├─B
             │  ├─label
             ├─test
             │  ├─A
             │  ├─B
             │  ├─label

where folder A contains pre-phase images, folder B contains post-phase images, and folder label contains the change detection masks.

Extract text files for the descriptions of each image pair in LEVIR-MCI:
```
python preprocess_data.py
```
After that, you can find some generated files in ./data/LEVIR_MCI/.

Train

Make sure you performed the data preparation above. Then, start training as follows:

python train.py --train_goal 2 --data_folder /DATA_PATH_ROOT/Levir-MCI-dataset/images --savepath ./models_ckpt/

Evaluate

python test.py --data_folder /DATA_PATH_ROOT/Levir-MCI-dataset/images --checkpoint {checkpoint_PATH}

We recommend training the model 5 times to get an average score.

Inference

Run inference to get started as follows:

python predict.py --imgA_path {imgA_path} --imgB_path {imgA_path} --mask_save_path ./CDmask.png

You can modify --checkpoint of Change_Perception.define_args() in predict.py. Then you can use your own model, of course, you also can download our pretrained model MCI_model.pth here: [Hugging face]. After that, put it in ./models_ckpt/.

Construction of Change-Agent

Agent Installation:

cd ./Change-Agent/lagent-main
pip install -e .[all]

Run Agent:

cd into the Multi_change folder:

cd ./Change-Agent/Multi_change

(1) Run Agent Cli Demo:

# You need to install streamlit first
# pip install streamlit
python try_chat.py

(2) Run Agent Web Demo:

# You need to install streamlit first
# pip install streamlit
streamlit run react_web_demo.py

Citation

If you find this paper useful in your research, please consider citing:

@ARTICLE{Liu_Change_Agent,
  author={Liu, Chenyang and Chen, Keyan and Zhang, Haotian and Qi, Zipeng and Zou, Zhengxia and Shi, Zhenwei},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Change-Agent: Toward Interactive Comprehensive Remote Sensing Change Interpretation and Analysis}, 
  year={2024},
  volume={},
  number={},
  pages={1-1},
  keywords={Remote sensing;Feature extraction;Semantics;Transformers;Roads;Earth;Task analysis;Interactive Change-Agent;change captioning;change detection;multi-task learning;large language model},
  doi={10.1109/TGRS.2024.3425815}}