tomtomtommi / deepmih Goto Github PK

Hi! Thanks for the code and I have a problem. When the number of pictures increases to 3 or 4, do the parameters of net3 share in the following training process? Or we need another network? Thanks.

Hi! Thank you for sharing your code! Are the results in HiNet and DeepMIH tested on RGB or Y,U,V?  When we reproduced HiNet, we found that the results in the paper could not be reproduced, and the PSNR in the log file given by HiNet's official code did not reach that in the paper.  When we use the provided DeepMIH model to test, PSNR also does not achieve the results in the paper.  Since we were all testing in RGB, and the code included testing on Y, this was questionable. 

Hi, thanks for your interest. We only polish and offer the code and the model for DeepMIH with 2 images hidden, which contains the main experiments in the paper. For DeepMIH-P, it was added in the Response process, and we only added a commonly used vgg_loss, which can be found in lots of low-level works. Specifically, we mainly referred to HiDDeN.

Originally posted by @TomTomTommi in #1 (comment)

is:issue is:open 博主你好，请问有没有带备注版的代码啊，我在运行您的代码时多出报错，经过处理后dataset无法导入数据

Hi! Thank you for sharing your code! I found that the training set of your model in your paper uses the DIV2k dataset, but the training set of this dataset has only 800 images. I want to determine whether the pre-trained model you provided only uses these 800 images for training?

Hi, it's a good job! I have a problem, I can't find the vgg_loss code defined in your code, could you share it? I just use it to learn。

Hello, I am very interested in your work. I noticed that your work is compared with the previous related work, but the ISN code is not open source. Do you reproduce it yourself or find the original author. If you reproduce it yourself, can you tell me what the reference code is? If the original author sent the code to you, can you share it?

Excuse me，why convolution cannot be replaced with self-attention in the coupling module？