Repository files navigation

Pytorch codes ofAdapting Segment Anything Model for Change Detection in HR Remote Sensing Images [paper]

The SAM-CD adoptsFastSAM as the visual encoder with some modifications.

SAM-CD now supports access toefficientSAM. Check the updated model atmodels/effSAM_CD.py (prior installation of efficientSAM at the project folder is required). However, direct integration of efficientSAM may cause an accuracy drop, so there is space to further improve the SAM-CD architecture.

1. Installation

   • InstallFastSAM following the instructions.
   • Modify the Ultralytics source files following the instructions at:'SAM-CD/models/FastSAM/README.md'.

2. Dataset preparation.

   • Please split the data into training, validation and test sets and organize them as follows:

      YOUR_DATA_DIR      ├── ...      ├── train      │   ├── A      │   ├── B      │   ├── label      ├── val      │   ├── A      │   ├── B      │   ├── label      ├── test      │   ├── A      │   ├── B      │   ├── label

• Find change line 13 inSAM-CD/datasets/Levir_CD.py (or other data-loading .py files), change/YOUR_DATA_ROOT/ to your local dataset directory.

3. Training

   classic CD training:python train_CD.py

   training CD with the proposed task-agnostic semantic learning:python train_SAM_CD.py

   line 16-45 are the major training args, which can be changed to load different datasets, models and adjust the training settings.

4. Inference and evaluation

   inference on test sets: set the chkpt_path and run

   python pred_CD.py

   evaluation of accuracy: set the prediction dir and GT dir, and run

   python eval_CD.py

(More details to be added...)

In the following, we summarize links to some frequently used CD datasets:

• LEVIR-CD
• WHU-CD(baidu)
• CLCD (Baidu)
• S2Looking
• SYSU-CD

For readers to easily evaluate the accuracy, we provide the trained weights of the SAM-CD.

Drive
Baidu (pswd: SMCD)

If you find this work useful or interesting, please consider citing the following BibTeX entry.

@article{ding2024adapting,title={Adapting Segment Anything Model for Change Detection in HR Remote Sensing Images},author={Ding, Lei and Zhu, Kun and Peng, Daifeng and Tang, Hao and Yang, Kuiwu and Bruzzone, Lorenzo},journal={IEEE Transactions on Geoscience and Remote Sensing}, year={2024},volume={62},pages={1-11},doi={10.1109/TGRS.2024.3368168}}