yzygit1230 / scd-sam

SCD-SAM

• The pytorch implementation for SCD-SAM in paper "SCD-SAM: Adapting Segment Anything Model for Semantic Change Detection in Remote Sensing Imagery".

Requirements

• Python 3.6
• Pytorch 1.7.0

Datasets Preparation

The path list in the datasest folder is as follows:

|—train

• ||—A

• ||—B

• ||—labelA

• ||—labelB

|—test

• ||—A

• ||—B

• ||—labelA

• ||—labelB

where A contains pre-temporal images, B contains post-temporal images, labelA contains pre-temporal ground truth images, and labelB contains post-temporal ground truth images.

Train

• python train.py --dataset-dir dataset-path

Test

• python eval.py --ckp-paths weight-path --dataset-dir dataset-path

Visualization

• python visualization visualization.py --ckp-paths weight-path --dataset-dir dataset-path (Note that batch-size must be 1 when using visualization.py)

scd-sam's Issues

A question about SAM encoder and decoder.

SCD-SAM is a great work. However, according to your code, I found the SAM decoder is not used.
Could you tell me why?

Thanks for your answers.

url to add

https://ieeexplore.ieee.org/document/10543161

请问moblisam中的adapter在代码的哪一部分

能否提供论文中最优的模型权重？

很感谢您能分享这份工作！请问能否提供最佳模型权重呢？

code

Hello, I am very interested in your work, but when I reviewed your code, I did not find how the contextual semantic discrimination (CSD) module uses the binary change detection results to guide SCD. How do you use the BCD results to guide SCD?