Hi, I'm very interested in using your code in my project. I was able to run demo scripts. Now I want to use your code for my own data, but unfortunately I do not know how I can do this on my data. I put them in files A, B But I think I need guidance to test Please tell me how I can find a difference for my images (I am a newcomer, thank you)

Hi I really appreciate your work.Now I want to train on changesim，a dataset with 4 classes label ，such as "missing","new","ratation","replaced object". I tried change n_class , but it didnt work. what should I do to train on a dataset which is more than 2 classes? thx~

Hi , I found when i load the pretrained model（trained on ade160k dataset）, the keys of checkpoint are not matched. The pretrained model: The self.net_G:

So the keys of pretrained model are all missing keys.

Hi, I found a new dataset called Landsat-SCD, which came from this paperA transformer-based Siamese network and an open optical dataset for semantic change detection of remote sensing images This dataset can be downloaded at figshare Its structure is similar to LEVIR-CD, while the size is 416*416 and the label includes 10 classes, I wondered what should I do to run your model on this new dataset. Thank you for your work.

Hi wgcban, I notice that DSIFN-CD dataset has much higher accuracy than BIT[4] (IoU from BIT 52.97% to ChangeFormer 76.48%). On another dataset, LEVIR-CD, the difference is not as large as DSIFN-CD. Could you please explain the main source of the large improvement on DSIFN-CD? e.g. training strategy, data augmentation, model structure... Thanks Wesley

Hi~ Thx for your great work, :clap: it's really inspired a lot in siamese Transformer network realizing.

However, I still have some questions about the code implementation and the details of the paper.

1. In the code, the implementation of Sequence Reduction was completed through the Conv2d non-overlapping cutting feature map before MHSA. https://github.com/wgcban/ChangeFormer/blob/9025e26417cf8f10f29a48f34a05758498216465/models/ChangeFormer.py#L316 The effect of the implementation is similar to that of the first shape and then linear projection in the paper, but this code implementation will result in a reduction of the sequence length R^2 times (similar to the idea of cutting the image into 16 * 16 patches at the beginning of the ViT). However, the formula 2 in the paper shows that the reduction is times. Is there an error here?

2. In this code:https://github.com/wgcban/ChangeFormer/blob/9025e26417cf8f10f29a48f34a05758498216465/models/ChangeFormer.py#L507 the actual code implements two skips connected. In the pink block diagram of Transformer Block explained in the upper right corner of Fig1 in the paper, do you need to draw two skips connected?（Add skip bypass connecting Sequence Reduction input an MHSA output）

3. About Depth-wise Conv in Transformer Block as PE.Why you do this? How to realize position coding(Can you explain it)? Why is this effective?

4. patch_block1 is not used in the code. What is this module used for? Why is it inconsistent with the previous block1 dimension? （dim=embed_dims[1] and dim=embed_dims[0] respectively）https://github.com/wgcban/ChangeFormer/blob/9025e26417cf8f10f29a48f34a05758498216465/models/ChangeFormer.py#L52

Looking forward to your early reply!:smiley:

hi, I tried to train my personal data with 4 classes = {0,1,2,3}

pixels are like

0000000002220000 0000000002200000 0000000002000000 0010000000000000 0111110000000000 1111000000000000

grayscale.

when I train this data, the accuracy converges to 0.5 and never changes. Is there any problem that I miss?

what I changed is only n_classes = 4

thanks.

Hello, it's a nice code. It takes me a lot of time to run the program using the LEVIR-CD-256 dataset you have processed. Is this normal? How long will it take you to train the model？ Looking forward to your early reply!

Hi I really appreciate your work. I have a few questions about the model. First of all is it possible to modify the size of the input images? Then how can we retrain the model with our data? I noticed that the model detects the changes appeared in the image B. How can we generate a map for the disappearance of elements in A? Thanks. I remain open to your answers and suggestions.

HI~，after reading your paper, I still can't understand your design of Difference Module which consists of Conv2D, ReLU and BatchNorm2d,What is the reason for this design? in the eqn: Fidiff = BN(ReLU(Conv2D3×3(Cat(Fipre, Fipost)))) in the code： nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1), nn.ReLU(), nn.BatchNorm2d(out_channels), nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1), nn.ReLU() Looking forward to your early reply!

Hello, I am trying to reproduce the BIT-CD model code, and it looks incorrect when using WHU-CD dataset as well as DSIFN-CD dataset, and the result appears too high than your result. But it looks normal when using LEVID-CD dataset. Do I need to change the pre-training?

Hello again! I'm training the net on the former dataset,a problem happened when the trainer saves visualization, the vis_pred, vis_gt of saved images do not look normal. Though I found your visualization method right here https://github.com/wgcban/ChangeFormer/blob/9025e26417cf8f10f29a48f34a05758498216465/models/trainer.py#L220-L232 still don't know how to make it fit to save multiple classes gt and pred images.

You want to perform multiclass classification. Even if I change the code to args.n_class = 9, the class is predicted to be 2. What should I do? sry im korean It shouldn't be possible to modify only n_class, but should there be multiple classes of labels?