search

arcchang1236 / CA-NoiseGAN

The official implementation of our ECCV 2020 paper "Learning Camera-Aware Noise Models".
https://arcchang1236.github.io/CA-NoiseGAN/
MIT License
73 stars 7 forks source link

Recreating paper results #6

Open tomk2005 opened 3 years ago

tomk2005 commented 3 years ago

Hello,

I am having difficulty recreating your paper results. I'm currently running your code on the data you provided with pre-trained models with the config values of: patch_size = [64, 64] sample_amount = 1191

I then average KLD and get KLD ~0.5 . Should running test noise models.py with these settings recreate the paper's KLD (~0.00159)?

If so, can you possibly upload a more detailed requirements files of your environment? maybe it is some package mismatch?

ZcsrenlongZ commented 3 years ago

I have the same problem as user tomk2005, can you upload more detailed requirements?

OrGreenberg commented 2 years ago

Hi! same for me.

Here are the results I achieved running the code over the provided test set, without any changes in the config file, except for the data_dir. I used the pretrained models provided.

--> Denoiser Test (PSNR)

average psnr ours: 47.841578781906506 average psnr g: 41.00051686254721 average psnr pg: 46.08441194932963 average psnr noiseflow: 47.147988764349506 average psnr ours real: 47.71135122092602 average psnr real: 46.7334418265441

--> Noise Models Test (KLD):

G: 0.6996052614889056 | PG: 0.06394353525639518 | Ours (CA-NoiseGAN): 0.03184607044417602
[** noise-flow results are not presented]

The denoiser's results make sense, yet not fully aligned with the results presented in the paper (Table 7). On the other hand, the noise model results seem completely wrong, particularly compared to the ones presented in the paper (Table 1).

Any help here? :)

happycaoyue commented 2 years ago

Hi! same for me.

Here are the results I achieved running the code over the provided test set, without any changes in the config file, except for the data_dir. I used the pretrained models provided.

--> Denoiser Test (PSNR)

average psnr ours: 47.841578781906506 average psnr g: 41.00051686254721 average psnr pg: 46.08441194932963 average psnr noiseflow: 47.147988764349506 average psnr ours real: 47.71135122092602 average psnr real: 46.7334418265441

--> Noise Models Test (KLD):

G: 0.6996052614889056 | PG: 0.06394353525639518 | Ours (CA-NoiseGAN): 0.03184607044417602 [** noise-flow results are not presented]

The denoiser's results make sense, yet not fully aligned with the results presented in the paper (Table 7). On the other hand, the noise model results seem completely wrong, particularly compared to the ones presented in the paper (Table 1).

Any help here? :)

Can you leave an email address for me? Let's talk about noise modelling?