search

BorealisAI / noise_flow

Noise Flow: Noise Modeling with Conditional Normalizing Flows
Other
150 stars 20 forks source link

Sample noise flow add to SIDD raw image problem using the pre-train model provided #3

Open circlehy opened 5 years ago

circlehy commented 5 years ago

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

jmu201521121021 commented 4 years ago

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

circlehy commented 4 years ago

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

They provided the transfer code on the SIDD website.

jmu201521121021 commented 4 years ago

@circlehy thank you!

jmu201521121021 commented 4 years ago

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

Fanbenchao commented 4 years ago

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

circlehy commented 4 years ago

@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.

AbdoKamel commented 4 years ago

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.

JimmyChame commented 4 years ago

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.

SSRheart commented 4 years ago

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.

Vandermode commented 4 years ago

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

ymcasky commented 4 years ago

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

AbdoKamel commented 4 years ago

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