wooseoklee4 / ap-bsn Goto Github PK

Hi, Thank you for releasing your codes.

May I ask several questions?

1. Based on my understanding, you crop the SIDD dataset into 512 * 512 patches with 128 overlappings. So if one patch is [100:100+512, 100:100+512], then the next patch will be [100:612, 100+512-128: 100+512-128+512].
2. The input tensor is the cropped patch, a 120 * 120 tensor will be generated by cropping the patch. So the actual input tensor will be 120 * 120
3. How do you normalise your input data? I can see somewhere there is a function called normalize(...) in denoise_dataset.py but where this function is called? It uses the std and mean to normalize data, how the std and mean are generated?

Thank you for your time!

Excellent work!

Focusing on similar papers of CVPR2023, and then pay attention to your work.

What steps did you do to quantify the correlation? are there codes provided?

Thanks

Sorry to bother you again, I had a memory issue when I tested a single image.Since my computer only has 8G of running memory, does this experiment require at least 16G and above of running memory?

Hi, I use the pre-trained model 'APBSN_SIDD.pth' and test it on the SIDD benchmark. I submit the results to the SIDD benchmark website and get

PSNR_Srgb,SSIM_Srgb,TimeSrgb(s)
36.73,0.932,-1

But model trained on SIDD Medium dataset gets 35.97/0.925 on the SIDD benchmark as it mentioned in the paper.

Thank you for the code, in addition to this, I would like to ask you, where are the three mat files under the SIDD file downloaded from, if you are convenient please reply or send to my mailbox: [email protected] Thank you

Hi, I use SIDD-Medium dataset and get 24,542 cropped images after prepare. Then I run python train.py -c APBSN_SIDD -g 0 without changing the code as well as the config file. I have trained the code by default setting for several times on RTX 3070Ti Laptop, GTX 1080Ti and Tesla A100. But the performance seems not as good as that mentioned in the paper.

Moreover, I download the SIDD pretrained model and change the setting in config file to test the model on SIDD_val. Then I run python test.py -c APBSN_SIDD -g 0 --pretrained APBSN_SIDD.pth. The result is showed bellow.

I'm confused and want to confirm the experiment setting. Thanks!

Hi,when run 'python test.py -c APBSN_SIDD -g 0 --pretrained APBSN_SIDD.pth --test_img ./00001.png',occur error 'FileNotFoundError: [Errno 2] No such file or directory: './dataset/SIDD/BenchmarkNoisyBlocksSrgb.mat'.What's the reason?Can test data only be SSID data sets?

Hello, I'm trying to denoise a 2D image. I add synthetic noise as follows:
np.array(np.random.poisson(0.01 * x / 0.1) / 0.01
But the model is unable to denoise the image. Can you give some suggestions?

Hello, I want to ask how to solve this problem.When I ran the train.py, an error is reported here.

Could you provide the dataset of DND. I am a freshman and I can't register because there is no relevant Google Scholar Profile,My email address is [email protected]

Thanks for your reply！
Furthermore, I train your code on SIDD Medium sRGB dataset, and run python test.py -c APBSN_SIDD -g 0 -e 20 .I pack the results into SubmitSrgb.mat and submit it to the SIDD benchmark website. The results are as follows.

But I notice that the model trained on SIDD Medium sRGB dataset gets 35.97/0.925 on SIDD benchmark. The model trained on SIDD benchmark in a fully self-supervised fashion gets 36.91/0.931 on SIDD benchmark in the paper.

I have no idea why there's 0.85/0.007 gap when it trained on SIDD Medium and infered on SIDD benchmark.
Thanks for your time and attention again!

Thank you for your great work!

In Figure 2, you provide the analysis of spatial correlation on real-world noise. I try to calculate the accurate value of the correlation and reproduce the plots in Figure 2. Specifically, I divide the neighbor noise values by the center noise value, while my results seem different from Figure 2:

Can you provide more details about how Figure 2 is generated? It would be amazing if the code snippet can be provided!! Thanks :)