ngchc / camerasr Goto Github PK

Hi,
I have downloaded and opened the raw data (.MAT files). I found that the values of these .MAT files are within [0, 1]. Has the raw data already been normalized, such as "raw = (raw - black_level)/(white_level - black_level)"?

By the way, the paper said that white balance is fixed for each capture. Could you provide the white balance setting?

Thank you!

hi,@ngchc, thanks for you release your great work. I have one problem with the preprocessing train data.

what is this mean? and why do this step ?

Sorry, I don’t understand very well. Why are the LR and HR images the same size? If I want to use your data to train a super-resolution network, I should enter a small LR and a large HR. How can I get a small LR in the real world?

Thank you for your meaningful work.
I would like to ask, when training VDSR model, is ’City100_NikonD5500‘ and ’City100_iPhoneX‘ dataset mixed together as training dataset?

Thanks for your great work. I have downloaded your raw data, but there is no information about the raw file such as white balance , bayer patterns etc, and I can not recover the color images correctly, could you please release the original raw files or give us the relevant information ? Thank you very much .

能问下你们论文所说的R-V方法可不可以理解为就是真实拍摄的低清图像和高清图像对？
然后你们论文说所取得的高清图像是低清的2.9倍，但是这个训练集的高清低清图片大小是一样的，这个是低清图片经过差值的到的嘛

figure2(c)R-V degradation ，my test psnr 25.451540226734174,is different from paper
In Tabel1's five interpolated LR(s)，my test psnr is 2 db lower。
My test code is below:
`def compute_psnr(im1, im2):

if im1.size != im2.size:
    raise Exception('the shapes of two images are not equal')

rmse = np.sqrt(((np.asfarray(im1) - np.asfarray(im2)) ** 2).mean())
psnr = 20 * np.log10(255.0 / rmse)
return psnr

path='../City100/City100_NikonD5500/'#27.1413
#path='../City100/City100_iphoneX/'#24.333565211970758

lr_path=path+'001L.png'

hr_path=path+'001H.png'
print(compute_psnr(Image.open(lr_path),Image.open(hr_path)))#26.544952727983393

'''

psnr_list=[]
bar=tqdm(range(1,101),desc='the bar')
for i in bar:

lr_path=path+"{:0>3d}".format(i)+'L.png'
hr_path=path+"{:0>3d}".format(i)+'H.png'
lr=Image.open(lr_path)
hr=Image.open(hr_path)


hr=ToTensor()(hr)
lr=ToTensor()(lr)
mse = ((hr - lr) ** 2).mean()
#print(mse)
psnr = 10 * log10(1 / mse)

psnr_list.append(psnr)

PSNR=np.array(psnr_list)
print('average')
print(PSNR.mean())
`

Excuse Me,
I find there are three model.data-00000-of-00001 files, I use command "tar -zxvf model.data-00000-of-00001.tar.gz.0" ,but it can not untar it, how to use your Gan model ?
Thanks in advance.

After you align the interpolated LR image with the HR ground truth image, what method of downsampling do you use to get the low resolution inputs down to 1/4 the resolution of the high resolution outputs as required by the SRGAN architecture?

Hello,
Thanks for your paper https://openaccess.thecvf.com/content_CVPR_2019/papers/Chen_Camera_Lens_Super-Resolution_CVPR_2019_paper.pdf and sharing code.

I would like to ask you some questions

1. For the section 6.1 in your paper "Advanced digital zoom"
   Once the training done, we have 3 trained networks BicubicSR, GaussianSR, CameraSR.
   To get the figure 8, figure 9 and figure 6b, what is input LR image that you use to inference with 3 above trained networks ?
   a. Is it always the low resolution image obtained from the camera for all 3 networks ?
   b. Or downsampled bicubic image for BicubicSR network; downsampled gaussian image for GaussianSR and low resolution image obtained from the camera for the CameraSR ?

If it is (a) then the explication is a domain gap between the the low resolution image obtained from the camera and the one with simulated method (bicubic or Gaussian) (you already mention this observation in section 5.1 :-) )

If it is (b) : ouf, I don't understand the logic of your work. Please give more information in this case

2. If it is (a), hence for the section 6.2. Generalizability : if we observe this generalization, it means that there is no (or if exist, very small) difference in degradation pipeline between different devices. And the bicubic/gaussian degradation is a simple model but it is not realistic at all.

Thanks !

你好，我发些你的训练数据集里都是png 图片，这些图片是照相机或者手机拍出来的么，一般我们手机拍的都是jpg 的图输入压缩后的图，不是无损图，我知道的是GAN的方法对压缩的图无法消除振铃效应，请问你有遇到此类问题么？

why the city100 dataset both the LR image and HR image has the same resolution?Can I get the LR images not interpolated?

It's a meaningful work for real world super resolution. May you share your code of generating paired images? It will help us build our own datasets. Thanks.

Hello, I try to download the dataset you shared, but seems there is some problem, it shows 0 kb in baidu cloud, also it is fail to download. Plz help to check it.