hvision-nku / camoformer Goto Github PK

I meet nan problem during training :(

Nice work! We are looking forward to the entire code including data processing and train.

This is a nice job to solve COD. But I have a puzzling question.
This is the code in the file which named decoder_p.py, it shows that the E2 & E4 will change shapes to match with
E3's shape(H/16, W/16, Cd).
But the Figure 2 in your paper, shows that E2(H/8,W/8,Cd),E4 (H/32,W/32,Cd). I want to know the truth.
Looking forward to your reply。 Thank you very much.
decoder_p.py

Figure 2

Dear Bowen,

 I found there is no pretrained model (CamoFormer-trained) both in [Baidu Netdisk](https://pan.baidu.com/s/1QyE_HkxCp9vWlhnJAx20dQ?pwd=f4bp), [One Drive](https://mailnankaieducn-my.sharepoint.com/:f:/g/personal/bowenyin_mail_nankai_edu_cn/Eg2fWTz1pjFBiJfAdj1XqGQBbQ2pOjpNq0AbIjseadTgSA?e=BFvBkU). Could you check and update it? Thanks a lot~ 

Hi，thanks for your job，please tell me the pretrain version of resnet and datasets? I'm looking forward to your reply and thanks very much.

Does "mask" really mask key and query ?

Hi, this is a nice work, but I have some questions about masks.

I noticed that the mask is a binary image which sets some parts of features to 0 after multiplying the key and query.
But it still generates attention-weights when calculating the similarity matrix, shouldn't the mask be set to negative infinity ?

Hi~

Your work is really great!

The paper mentions a version that uses ConvNext as the backbone, but I checked the code and there is only a model that uses PVTv2 and Swin Transformer, is it missing ConvNext? If it is convenient, can you provide the code for this part? It would help me a lot!

The shape of query Q in Figure 3 is HWxC, but in the code

Appreciate your work！ And may i ask about the details in your cosine learning rate strategy, like Tmax setting or some others needed attention?

Hello, I have read your paper and I am very interested in your ideas. It is very good work. Could you please update the training code?

hi, thanks for your great work. is there any plan to release the full code about train, data precessing and so on?

Thank you for your excellent work. I have some confusion about B-TA, the background map generated in the text using subtraction is multiplied with Q and K. So that the object in the background map is 0 and the object in QK is 1. Wouldn't the multiplication make the generated feature map all 0?

Nice work! In your paper, the R in CamoFormer-R is interpreted as ResNet. I would like to ask if it refers to use ResNet-50 as the backbone?

您好，请问论文中Figure 7是如何可视化特征的？

Nice work, but I have a question while reading, how did you make your PR curves and F-measure curves, looking forward to your reply!