search

creotiv / hdrnet-pytorch

Unofficial PyTorch implementation of 'Deep Bilateral Learning for Real-Time Image Enhancement', SIGGRAPH 2017 https://groups.csail.mit.edu/graphics/hdrnet/
228 stars 45 forks source link

Train this system on Adobe FiveK Dataset #2

Closed marcelsan closed 5 years ago

marcelsan commented 5 years ago

Hi,

Thank you for sharing this code. I am wondering how you trained this system using the Adobe FiveK dataset. Did you first converted the RAW images to JPGE? Would it be possible to train this system using RAW images as input? I think this preprocess step, even using a fast demosaicing algorithm, would potentially hurt the system performance. Am I missing anything? Does Google HDRNet original implementation also preprocess the RAW images to JPGE?

Thank you.

creotiv commented 5 years ago

You cant use RAW images, cause you can't see it, to get some visible result you need to preprocess raw, by setting bw balance, demosaic algo, contrasts, etc. After that you get prepossessed RAW. After you done this, you can save data in 8Bit precision JPEG,PNG etc.. or 16,32bit, TIFF for example. After that you can use tiff in your training and you dont need to run raw convertaion on each training step. No it will train at the same speed, as you anyway convert your 8bit image to the 32bit float, so no difference for network.

In my experience i realy didn't saw any visible improvement with using higher precision images, it's because we still doing interpolation, so getting +/- same results.

Also this code currently more for learning then real-life use, because sampling function doesn't have tri-linear interpolation, and without it network is not working properly. So if you need to train something, please use Tensorflow implementation of this network :(

Kindly yours, Andrey Nikishaev

Areas ML/DS/CV/Soft Dev/BizDev/Growth Hacking/Customer Rel/IT LinkedIn http://ua.linkedin.com/in/creotiv GitHub http://github.com/creotiv Slideshare https://www.slideshare.net/anikishaev/ Skype creotiv.in.ua Mobile +380632410666

On Tue, Oct 15, 2019 at 7:30 PM Marcel Santana notifications@github.com wrote:

Hi,

Thank you for sharing this code. I am wondering how you trained this system using the Adobe FiveK dataset. Did you first converted the RAW images to JPGE? Would it be possible to train this system using RAW images as input? I think this preprocess step, even using a fast demosaicing algorithm, would potentially hurt the system performance. Am I missing anything? Does Google HDRNet original implementation also preprocess the RAW images to JPGE?

Thank you.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/creotiv/hdrnet-pytorch/issues/2?email_source=notifications&email_token=AAB5CDMYSITPBD5AWM5MEC3QOXVY5A5CNFSM4JA7CPZKYY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HR5UAJA, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAB5CDJNLV2JCDIO67NSVTLQOXVY5ANCNFSM4JA7CPZA .

marcelsan commented 5 years ago

Sorry, I meant that this preprocessing could slow down the inference. I understand that for training we would run this preprocess only once.

Also, I imagined that we could have the RAW image as input to the network and the output be the final image with all the bw balance, demosaic algo, contrasts, etc. learned within the network.

Were you able to train the tensorflow implementation of HDRNet?

creotiv commented 5 years ago

Its impossible to push RAW into network, cause it have to many formats of data and other things, and also one raw photo is ~150-300MB, you anyway need to make preprocessing. The only thing you can do is to push preprocessed raw image into network and get more better preprocessing, or get relative fixes to bw, contrast,etc.

yeah, HDRNet working fine, the only problem ive got is pyramid network not working due to bug with shapes in TF1.13. As for TF2.0 i don't know never tried it under this version.

Kindly yours, Andrey Nikishaev

Areas ML/DS/CV/Soft Dev/BizDev/Growth Hacking/Customer Rel/IT LinkedIn http://ua.linkedin.com/in/creotiv GitHub http://github.com/creotiv Slideshare https://www.slideshare.net/anikishaev/ Skype creotiv.in.ua Mobile +380632410666

On Tue, Oct 15, 2019 at 7:53 PM Marcel Santana notifications@github.com wrote:

Sorry, I meant that this preprocessing could slow down the inference. I understand that for training we would run this preprocess only once.

Also, I imagined that we could have the RAW image as input to the network and the output be the final image with all the bw balance, demosaic algo, contrasts, etc. learned within the network.

Were you able to train the tensorflow implementation of HDRNet?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/creotiv/hdrnet-pytorch/issues/2?email_source=notifications&email_token=AAB5CDKGW5SIPW6GOI4MCW3QOXYQZA5CNFSM4JA7CPZKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBJPIAI#issuecomment-542307329, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAB5CDPCCCKXBBZPQVGYZQLQOXYQZANCNFSM4JA7CPZA .