mtlab / rsgunet_image_enhance Goto Github PK

could supply the missing file，which is in /home/ustc-ee-huangjie/video-practice/hdrnet ，I want test run.py in my case

I downloaded the DPED data set directly, but I found that the data processing method in your code is not the data set processing method, can you provide your data set? Or can you tell me how to deal with the DPED data set to meet the input data processing method in your code? This is the reported error:
[2332] ERROR data_pipeline.py:51 | Training dir ./iphone/training_data does not containt 'input' folder
[2332] ERROR data_pipeline.py:53 | Training dir ./iphone/training_data does not containt 'output' folder

train.py 中swaps没有变化。应该会影响D网络的训练。这里是不是bug？

Hi @MTLab, I'm currently running the train, but when I decided to run the net against some other images I realized that there was some missing files: /home/ustc-ee-huangjie/video-practice/hdrnet. Could you please add the missing files in the repository? I think it's just missing models in order to run models.Resnet. Thanks!

@MTLab Hello, I found that the unet-like architecture is employed in your experiment. This structure adopts convolutions with the stride of 2 and pooling layer. If the input image with the size of the odd number, how does the output image remain the same size as the input?

Great job. Congrads.

One quick issue I found is this: Your paper states that you used the U-Net structure, which includes the max-pooling to make "down-sample".
I found that your UNet.py does not include this max-pooling. Instead, a convolution with stride of 2 is used to strink the tiles in steps.
Is there an explicit purpose of this modification?

Many thanks.

Hi thank you for your great work first!

I followed all your instructions in read me, using the DPED dataset, the iphone-cannon part (seems you are using the part only according to the dataset list provided)
I didn't change anything and the training goes well, I can get around 22 psnr on the training set. However I only get around 11, 12 maximum on the test set ...

I modified the provided run.py to fit the enhancement network, converting the input image according to the training procedure, the result for normal light condition images are ... I'd say so so, cause the test psnr I get is only 12, which a little color distortion, tends to be brown-gray.
However when the input image is in very low light condition, it become even worse, the output is entirely gray(and a little brown)

What could be wrong? The reported psnr on DPED benchmark is 21.89, how can I reproduce that result? How many iters are requested to get that result? Or can anyone else get a good psnr on the test set? I wonder if there is something wrong, cause the test psnr stay around 11.5 in very early iters.
I noticed that the provide weights for those loss proposed aren't configured exactly the same as in the paper. So I also tried to make them configured exactly the same, the result still stay the same ...

Thank you very much if you can give me any hints !