Myung-Joon Kwon

Thanks for your good analysis. However, please keep in mind that CAT-Net targeted splicing and copy-move forgeries. You know, CAT-Net mainly uses compression artifacts to detect splicings or copy-moves since...

For the first question, you are right. Just use any binary classification code to train the DCT Stream. I used a learning rate of 0.05 which decayed every 10 epochs...

I used this repo for pretraining: https://github.com/HRNet/HRNet-Image-Classification

> And are the weights available for downloading (DCT_only_v2.pth.tar) from the pretrained classification model with Park? If not, could you make them public?, so that I can compare it with...

Hi, full CAT-Net takes 24-channel input (3 for RGB and 21 for DCT) and the DCT stream takes 21-channel input. Due to the input difference, you should uncomment this line:...

Glad it helped ☺️

@Vadim2S It's hard to debug your personalized code, especially when you didn't specify your changes. As far as I can tell, from the message `=> Cannot load pretrained RGB` and...

Hi, the problem was caused because Google changed its policy not to give infinite space for university accounts. So I changed it to Baiduyun. See the front page now. Thanks.

Hi, Actually, tampCOCO is unnecessarily large for training CAT-Net. You might find out that one epoch does not use all training images. Instead, the maximum number of images used in...

Yeah that’s right.