Unsatisfactory results

[lulujianjie]

input:

output:

[Li-Chongyi]

Such an image with obvious noise challenges all low-light image enhancement methods. Zero-DCE can only enhance low-light image without obvious artifact or noise. Unfortunately, it cannot handle artifacts, blocking, and noise well. 发件人：JackLu <notifications@github.com> 发送日期：2020-07-03 11:26:17 收件人：Li-Chongyi/Zero-DCE <Zero-DCE@noreply.github.com> 抄送人：Subscribed <subscribed@noreply.github.com> 主题：[Li-Chongyi/Zero-DCE] Unsatisfactory results (#4) input: output: — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[lulujianjie]

In fact, I have tried many images without obvious noise. And I compared the results generated from your model and from MEITU (only adjust in low level, not using deep model).

One of the examples shows as following:
input

your DCE output

MEITU output(only add 30 brightness)

My question is if using the deep model to learn a simple brightness transformation will be overfiting. Although your model is not very large, it also has excessive capacity to fit a simple transformation.

[Li-Chongyi]

MEITU is commercial software that contains a serial of SP operations, not just light enhancement. Moreover, I guess that it operates on raw data (not sure). So it is unfair to compare it with. There is no overfitting because Zero-DCE is an optimization-based method without ground truth. If the Zero-DCE is used in the product, it needs some modifications such as coupled with denoising and removing color band, etc. Anyway, thanks for providing the resutls. 发件人：JackLu <notifications@github.com> 发送日期：2020-07-03 12:43:11 收件人：Li-Chongyi/Zero-DCE <Zero-DCE@noreply.github.com> 抄送人：Chongyi Li <lichongyi25@gmail.com>,Comment <comment@noreply.github.com> 主题：Re: [Li-Chongyi/Zero-DCE] Unsatisfactory results (#4) In fact, I have tried many images without obvious noise. And I compared the results generated from your model and from MEITU (only adjust in low level, not using deep model). One of the examples shows as following: input your DCE output MEITU output(only add 30 brightness) My question is if using the deep model to learn a simple brightness transformation will be easy to be overfiting. Although your model is not very large, it also has excessive capacity to fit a simple transformation. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[maldeer]

Hi, @lulujianjie
This is Murtadha Aldeer, a PhD candidate at Rutgers University, USA. I apologize as I am posting something irrelevant to this topic but I sent you many emails and every time I get delivery failure.
I have questions regarding using your code you developed for "Robust Single Accelerometer-Based Activity Recognition Using Modified Recurrence Plot" as I am doing some related work.

I appreciate it if you provide me with a working email to contact you.
My email address is maldeer@winlab.rutgers.edu

Appreciating your reply

Thank you

[lulujianjie]

Hi, @lulujianjie
This is Murtadha Aldeer, a PhD candidate at Rutgers University, USA. I apologize as I am posting something irrelevant to this topic but I sent you many emails and every time I get delivery failure.
I have questions regarding using your code you developed for "Robust Single Accelerometer-Based Activity Recognition Using Modified Recurrence Plot" as I am doing some related work.

I appreciate it if you provide me with a working email to contact you.
My email address is maldeer@winlab.rutgers.edu

Appreciating your reply

Thank you

Hi, @lulujianjie
This is Murtadha Aldeer, a PhD candidate at Rutgers University, USA. I apologize as I am posting something irrelevant to this topic but I sent you many emails and every time I get delivery failure.
I have questions regarding using your code you developed for "Robust Single Accelerometer-Based Activity Recognition Using Modified Recurrence Plot" as I am doing some related work.

I appreciate it if you provide me with a working email to contact you.
My email address is maldeer@winlab.rutgers.edu

Appreciating your reply

Thank you

Ok, I will send an email to you.

[lulujianjie]

MEITU is commercial software that contains a serial of SP operations, not just light enhancement. Moreover, I guess that it operates on raw data (not sure). So it is unfair to compare it with. There is no overfitting because Zero-DCE is an optimization-based method without ground truth. If the Zero-DCE is used in the product, it needs some modifications such as coupled with denoising and removing color band, etc. Anyway, thanks for providing the resutls. 发件人：JackLu notifications@github.com 发送日期：2020-07-03 12:43:11 收件人：Li-Chongyi/Zero-DCE Zero-DCE@noreply.github.com 抄送人：Chongyi Li lichongyi25@gmail.com,Comment comment@noreply.github.com 主题：Re: [Li-Chongyi/Zero-DCE] Unsatisfactory results (#4) In fact, I have tried many images without obvious noise. And I compared the results generated from your model and from MEITU (only adjust in low level, not using deep model). One of the examples shows as following: input your DCE output MEITU output(only add 30 brightness) My question is if using the deep model to learn a simple brightness transformation will be easy to be overfiting. Although your model is not very large, it also has excessive capacity to fit a simple transformation. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

Thanks for your reply. However, this is not a convincing response. There are a lot of inexplicable noises in the night sky of genenrated result as shown in figures, which are obviously caused by the deep model. As far as I know, illumination enhancement is a very basic operation in Meitu, Lightroom, etc.

[lulujianjie]

@lulujianjie
Hi, I have a hard time reproducing the astonishing well results as in the ZeroDCE paper as well. However, I should say this non-reference model training idea is cool.

Based on what described in the ZeroDCE paper, I have my own implementation of ZeroDCE. I took a significant amount try/modify it, including the loss function and parameter setting.

I encounter problems like bright/white area getting darken, and borders getting messed. But I never encounter such 'noisy point' artifact. Would you mind tring my implementation(https://github.com/bsun0802/Zero-DCE) and my pre-trained model?

You may try the two above inputs, which are raw input images in my experiment.

[maldeer]

Thank you, I will be waiting for your email

[Amr-Mustafa]

@Li-Chongyi You said earlier in this thread that this method needs to be coupled with denoising techniques and removing color band if it is to be used in production, I want to ask if you have any starting points on these subjects ?

[Li-Chongyi]

Hi, Thanks for asking. The company has several methods to enhance lowlight image and denoising with Zero-DCE coupled with a denoising technique. However, I cannot provide the details because of the conflict of interest. Sorry. Amr Mustafa <notifications@github.com> 于2020年7月29日周三 上午2:54写道：

…

@Li-Chongyi <https://github.com/Li-Chongyi> You said earlier in this thread that this method needs to be coupled with denoising techniques and removing color band if it is to be used in production, I want to ask if you have any starting points on these subjects ? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AIEVQEMFI7PAHVT7NBQEQBLR54NHDANCNFSM4OPN2SJA> .

[zhuozhongshuo]

@lulujianjie I get the similar results with you by using the default parameters. Which pytorch version do you use? My environment is:

• pytorch 1.8.2
• torchvision 0.9.2
• cuda 10.2
  a little different with the recommend env.