Hi,

Great work! Congratulations. I downloaded your ELD dataset but it doesn't contain the images processed by your noisy model. Is it possible to also upload the noisy images?

Thanks.

Hi,

I download your pretrained model and ELD dataset to know more details of your paper.

But I failed with run test_ELD.sh

I found in the ELD dataset, every scene in a specific camera there is 16 raw data and some have 16 corresponding RGB data.

Is this should be 3 (isos) * 2(lowlight ratio) = 6 examples?

And I find img_ids_set = [[4, 9, 14], [5, 10, 15]] in the the_ELD.py adn gt_ids = np.array([1, 6, 11, 16]) in sid_dataset.py.

I do not know exactly what this ids means.

Hello and sorry for bothering u again :stuck_out_tongue_winking_eye: I'm trying to reproduce the Paired-SID baseline according to the training settings from your paper:

Our implementation is based on PyTorch. We train the models with 200 epoch using L1 loss and Adam optimizer with 
batch size 1. The learning rate is initially set to 10−4, then halved at epoch 100, and finally reduced to 10−5 
at epoch 180.

I tested the checkpoint you provided and it achieved 38.87dB on SID-Sony(Raw), while the model trained by myself can only acquire around 36.80dB under the same test settings. I launched another experiment with a larger batch size(16), longer training procedure(about 600 epochs) and more coarse-to-fine lr settings(1e-3 to 1e-6) and ended up with a model which is still worse than the baseline(~0.4dB lower). I wonder if I missed any important settings, if possible, could you please share your training settings with me? Thanks a lot! 😄

Hello, I have used python to open the raw file. And I can use rawpy class to obtain some parameters in the ISP process, like white balance, CCM. However, it didn't show which the demosaic method that the camera use. I want to implement the ISP process step by step, just like what your codes have done.

Hi,

Congratulations for your paper. It presents interesting insights. Are you planning to release the code for noise modelling, and if yes, could you let us know a tentative time/date for the same?

Thanks!

Thanks for sharing the code of the great works!

Since I do not have exactly the same devices as the paper, could you share the ISO corresponding to the minimum and maximum of ”K"?

Thanks a lot

Hello and thanks for your amazing job! I got quite confused about the effect of IlluminanceCorrect func at https://github.com/Vandermode/ELD/blob/aa0edb44a8fc20e01f83c1f6e93ee70d3190e142/models/ELD_model.py#L139 It seems like some kind of normalization, yet I didn't figure out why it should be applied here, could you please explain it for me? Thanks a lot :)

In you dataset description, the reference frame with a proper exposure time, and have a f in [100, 200] to decrease exposure time to simulate low light situation.
But I do not find where to multify the f in your paper, because it's obviously have luma difference in the raw pair.

There are 18 raw images in the SonyA7S2 sub-dataset and 17 raw images in the Canon700D sub-dataset. Since the verification code you released selects the picture based on the number, this error will cause the data to be mismatched, and then the wrong verification result will be obtained. Specifically, since GT is corresponding to a very dark image, the corrector will align the brightness to black, resulting in the PSNR being pulled up to about 70dB. I hope you can update ELD.7z and explain which version of the data should be used as a benchmark for future works to compare.

The short exposure pictures of SID(see in the dark) you use in this paper is not the original images, have you multiple a parameter on the short exposure pictures to make it easy to show?

Hi, I got an error when running the command python test_ELD.py --name sid-paired -r -re 200 --no-verbose --chop Error shows: Through debugging, I think there must be some pre-trained model files like "model.pt" in the checkpoint folders, but I don't see them. Otherwise, we would need a script to download such pre-trained model, right?

i use rawpy to read raw data from Canon 70d and 700d, and find the max value are always less than 2^14=16383. may i know the white level of these two cameras?

hi， the color bias is related with the gain K ?, According to some experiments，As K gets bigger, that color deviation gets bigger。 And for camera params in this demo， as K gets bigger value, color bias only sample some values in given color bias,that is right? how to sample color bias for A bigger K ?

Hi I have some questions about released camera parameters For example: In NikonD850_params.npy, it has key name like Profile-1(G_scale, g_scale, R_scale), Profile-2, G_shape. I don't know they are corresponded to which parameter listed in paper table1 respectively.

Another question, are these noise parameters the same in R, G, B channel of raw? (like system gain K, tukey lambda shape and sigma, row noise sigma)

Hi, I got the artifacts when generating synthetic low light clean images. According to your paper, the fake low light clean images = long exposure images / ratio, while actually such operation(large integers are divided by ratio then make float to integers) squeezes the range of the values which loses the accuracy and generates the "non-continuous step" in the image, which feels like a HDR image displayed on an 8-bit screen. The result is as follows: original long exposure image synthesize low light clean image after auto-brightness for imshow original low light noisy image synthesize noise image based on the "non-continious" low light clean image

How do you fix the artifacts?

Hi, For the normal distribution, the intercept of the line in the Quantile-quantile plot is the mean, while the slope is the std of the distribution. However, the Tukey lambda is a prox distribution with a specific lambda and doesn't have a variance parameter as the closed-form like the normal distribution has. Thus, how can we derive the scale mathematically??