Non-overlapping crop

[angle011]

Hi wgcban,
I'm really interested in your project.I want to reproduce ChangeFormer in LEVIR-CD.There are two Simple questions
1)you want to put LEVIR img (1024x1024)into non- overlapping 256. Do you means divide one image(1024x1024) into 16 images(256x256)?In other issues,you say There has image perparation code in /data_preparation,but I dont find,(of course if there is no code provided I can write by myself)
2)In paper,you have 6096/762/762 train/val/test-------but my LEVIR have 445--train(1024x1024），if you divide one into 16 there are 445x16=7120 train dataset

[wgcban]

Hi @angle011,

Thank you for your interest in our work.

1. Yes, you are correct. What I meant by '' creating non-overlapping patches'' is that dividing original 1024x1024 images into 16 crops. You can find the dataset preparation code here ChangeFormer/data_preparation/levir_cd_256.m. You just need to run the code in MATLAB and it will generate the dataset in the required order.
2. You are correct and thank you for pointing out a typing mistake in the paper. Finally, we have 7120/1024/2048 for train/val/test , respectively. I will update this in the next version.

Also, I have uploaded my processed LEVIR-CD and DSIFN-CD datasets to the DropBox and you can download the final datasets here:

• LEVIR-CD-256: click here to download
• DSIFN-CD-256: click here to download

Accordingly, I have updated the readme file as follows:
https://github.com/wgcban/ChangeFormer/blob/main/README.md#links-to-download-processed-datsets-used-for-trainvaltest

Let me know if you have more questions!
Best,
Chaminda

[angle011]

Hi @wgcban ,
I reproduce changeformer in LEVIR-CD .but I got some funny problems,
I have trained the model with the script you provided /run_ChangeFormer_LEVIR.sh with 3090. But when I check the metrics ,I find the mIou is 0.90( after I trained 200 epoches) rather than 0.8248.It confuses me a lot.
I have check the num of train_dataset and val_dataset(7120/2048) . I want to know wheather I have forget something in your codes
my project name is CD_ChangeFormerV6_LEVIR_b20_lr0.0001_adamw_train_val_200_linear_ce_multi_train_True_multi_infer_False_shuffle_AB_False_embed_dim_256
my log file s here
log.txt

[wgcban]

Hi @angle011

I'm happy to hear that you were able to reproduce ChangeFormer results! Congrats!

As I can see from your log.txt the last time the best model was updated is at the 181-th epoch.

Here are the corresponding CD performance metrics copied from your log.txt:

---

acc: 0.99012 miou: 0.90476 mf1: 0.94791 iou_0: 0.98966 iou_1: 0.81987 F1_0: 0.99480 F1_1: 0.90102 precision_0: 0.99370 precision_1: 0.92051 recall_0: 0.99591 recall_1: 0.88234
Lastest model updated. Epoch_acc=0.9479, Historical_best_acc=0.9478 (at epoch 180)

---

Here, I have highlighted the performance metrics that I have reported in the paper.

As you can see, the results that you obtained are almost closer to what I have reported in my paper. I think you are confused with the F1 score and IOU.

In my paper, I have reported IoU = 82.48 (change class iou-iou_1) and F1=90.40 (change class F1 score-f1_1), and you obtained IoU = 81.98 and F1=90.10, which are almost close to each other.

I hope this answers your question!
Let me know if you have further questions.

Best,
Chaminda

[chilin0525]

Hi @wgcban ,
I also have question about non-overlapping crop,
Does cropping size related to detect some tiny change? e.g. using 128x128 as a patch better than 256x256