f64051041/saras-net

SARAS-Net: Scale And Relation Aware Siamese Network for Change Detection

Target: Change detection aims to find the difference between two images at different times and output a change map.

This paper has been accepted in AAAI-23.

For more information, please see our paper at AAAI or arxiv.

Overview of SARAS-Net:

Visualize each module by Gradcam:

Requirements

cuda: 11.0  
python: 3.6.9  
pytorch: 1.7.0  
torchvision: 0.8.1 

Installation

git clone https://github.com/f64051041/SARAS-Net.git  
cd SARAS-Net  

Quick start

Download LEVIR-CD weight : https://drive.google.com/file/d/1Gs6iYQcZI1Jm4NhthCwWI2olbO-bpvTd/view?usp=share_link
After downloaded the model weight, you can put it in SARAS-Net/.
Then, run a demo to get started as follows:

python demo.py

After that, you can find the prediction results in SARAS-Net/samples/

Train

You can find SARAS-Net/cfgs/config.py to set the training parameter.

python train.py

Test

After training, you can put weight in SARAS-Net/.
Then, run a cal_acc.py to get started as follows:

python cal_acc.py

You can set show_result = True in cal_acc.py to show the result for each pairs.

Data structure

Train Data Path

train_dataset  
  |- train_dataset 
      |- image1, image2, gt  
  |- val_dataset  
      |- image1, image2, gt  
  |- train.txt
  |- val.txt

The format of train.txt and val.txt please refer to SARAS-Net/train_dataset/train.txt and SARAS-Net/train_dataset/val.txt

Test Data Path

test_dataset  
  |- A 
      |- image1 
  |- B  
      |- image2 
  |- label
      |- gt 

Data Download

LEVIR-CD: https://justchenhao.github.io/LEVIR/

WHU-CD: https://study.rsgis.whu.edu.cn/pages/download/building_dataset.html

DSIFN-CD: https://github.com/GeoZcx/A-deeply-supervised-image-fusion-network-for-change-detection-in-remote-sensing-images/tree/master/dataset

Quick train on LEVIR-CD : https://drive.google.com/file/d/1DAlxuqalNIPopt-WgtDmCYO98_jWM3ER/view?usp=share_link

Quick test on LEVIR-CD : https://drive.google.com/file/d/1Bj5GQ3hZcDVSpFGZKxm7zIuCBP5XEr6x/view?usp=share_link

Result

Dataset	Pre.	Rec.	F1-score	IoU	OA
LEVIR-CD	91.97%	91.85%	91.91%	84.95%	99.10%
CCD-CD	97.76%	97.23%	97.49%	95.11%	99.35%
WHU-CD	88.41%	85.81%	87.09%	77.14%	98.89%
DSIFN-CD	67.65%	67.51%	67.58%	51.04%	89.01%

saras-net's Issues

Why the count of parameters and flops differ from that shown in the paper?

Here are my codes.

model = Change_detection()
input1 = torch.rand((1, 3, 256, 256))
input2 = torch.rand((1, 3, 256, 256))
from fvcore.nn import FlopCountAnalysis, flop_count_table
model.eval()
flops = FlopCountAnalysis(model, (input1, input2))
print(flop_count_table(flops))

Thank you for help.

Missing SOTA comparisions and paper references to add

Hi,

I like your work, and thanks for making the code available to the researchers.

However, your method does not compare with recent state-of-the-art change detection methods like ChangeFormerand DDPM-CD.

At least you do not mention there exists such works in the related works.

Any reason for this?

visualization by Gradcam

Thank you for open-sourcing the code, but I didn't find the code for visualization by Gradcam in the code, can you post it? thank you very much

I didn't know it was wrong

Paper ref to add

https://arxiv.org/abs/2212.01287

BasePath、DataPath、SavePath meaning

hi, i change my own training dataset name, but i got error as follows:

AttributeError: 'Dataset' object has no attribute 'img_pth'

my setting is like this :

BASE_PATH ='/data/chengxi.han/Sigma122/SARAS-Net' (Project location )
DATA_PATH ='/data/chengxi.han/data/LEVIR-CD-Dataset256' (Data location )
SAVE_PATH ='/data/chengxi.han/Sigma122/SARAS-Net/Results/LEVIR' (Project location )

hope for your reply, many thanks.

關於多類別數據集的訓練和測試

你好作者，首先感謝您的貢獻。我想問一下這個模型可以訓練多類別的數據集嗎，不是只有change和unchange兩個類別，我將變化的類別分爲”從沙漠到河流“→1、”從沙漠到建築“→2，等等。用SARAS模型怎麽修改呢，期待您的回復

Can you provide the clipped data set or the code for clipping data set

Can you provide the clipped data set or the code for clipping data set

Weights of DSFIN dataset?

Hello,

Thank you for sharing your code.

Can you share the weights of the DSFIN dataset?

Thanks in advance.