noise_flow icon indicating copy to clipboard operation
noise_flow copied to clipboard

Published 20 hours ago • verified publisher icon BorealisAI

Sample noise flow add to SIDD raw image problem using the pre-train model provided

Open circlehy opened this issue 5 years ago • 12 comments

Sample noise flow added to SIDD clean image in raw format, then transfer to RGB format. The added noise seems 'larger' than expected since the parameters were known. I am not sure how to solve the problem. cam: 0, iso: 100 clean 0051_002_S6_00100_00060_5500_N0051_GT_RAW_010_add_0_100 0_0_100_gt_sRGB

noisy 0051_002_S6_00100_00060_5500_N0051_GT_RAW_010_add_0_100 0_0_100_sRGB

circlehy avatar Oct 10 '19 12:10 circlehy

@circlehyhello, how do you transfer Raw format to RGB format?

jmu201521121021 avatar Nov 06 '19 09:11 jmu201521121021

@circlehyhello, how do you transfer Raw format to RGB format?

They provided the transfer code on the SIDD website.

circlehy avatar Nov 06 '19 09:11 circlehy

@circlehy thank you!

jmu201521121021 avatar Nov 06 '19 09:11 jmu201521121021

@circlehy "The added noise seems 'larger' than expected since the parameters were known",I also encountered this problem, do you solve the problem?

jmu201521121021 avatar Nov 06 '19 13:11 jmu201521121021

I also encountered this problem. I'll appreciate it that the authors could answer our questions. cam: GP iso: 3200 clean: 0 noise: 0 generate: 0

Fanbenchao avatar Nov 08 '19 11:11 Fanbenchao

@circlehy "The added noise seems 'larger' than expected since the parameters were known",I also encountered this problem, do you solve the problem?

No, the released model seems not work.

circlehy avatar Nov 12 '19 06:11 circlehy

Hi,

There was a bug in the sampling code, it was not using the trained batch norm layers, this is fixed now. Also, I have added code for rendering the sampled noisy images into sRGB.

AbdoKamel avatar Jan 13 '20 01:01 AbdoKamel

Hi, I still find some problems in the sampling code. I just use the pre-train model provided to generate some samples. The added noise seems different than expected: (The left is clean image, the center is original noisy image and the right is the generated noisy image) 10_03_0800 10_02_0800

Then I use the generated noisy images to finetune a denoising model pre-trained on SIDD dataset, the PSNR on SIDD validation dataset is decreasing significantly (from 39 down to 37). I am not sure if there are still bugs in the code.

JimmyChame avatar Feb 28 '20 08:02 JimmyChame

Hi, I still find some problems in the sampling code. I just use the pre-train model provided to generate some samples. The added noise seems different than expected: (The left is clean image, the center is original noisy image and the right is the generated noisy image) 10_03_0800 10_02_0800

Then I use the generated noisy images to finetune a denoising model pre-trained on SIDD dataset, the PSNR on SIDD validation dataset is decreasing significantly (from 39 down to 37). I am not sure if there are still bugs in the code.

Hi, have you solved this problem？ The same results appear on my trained model.

SSRheart avatar Mar 27 '20 05:03 SSRheart

Any update on this problem? Actually, I find the provided pretrained model is a simplified version of the one described on the paper, since the width parameter is set to 4 instead of 32 mentioned by authors

Vandermode avatar Jun 22 '20 08:06 Vandermode

10_02_0800 The same problem when run sample_noise_flow.py with "patch_size" set to 480. Have you solved this problem?

ymcasky avatar Jun 24 '20 03:06 ymcasky

Please check out this potential fix: https://github.com/BorealisAI/noise_flow/commit/f9e2d29f350f8d59fd6be9af58ab41eeaafeb1db

AbdoKamel avatar Oct 20 '20 16:10 AbdoKamel