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

Any dataset preprocess script？ #1

Open butterl opened 5 years ago

butterl commented 5 years ago

Hi，Thanks for sharing this project！ Any dataset preprocess script for this project ？ the RAW to jpeg preprocess way affect a lot of the the output.

what about the psnr in you training experiment , I only got 21db from the original google code， not sure if I got somthing wrong in data preprocess

Thanks

creotiv commented 5 years ago

This solution currently just for learning. Because Torch F.grid_smaple doesn't have triliniear interpolation that was used in original network(which is strange cause it can use 3D image as input), thus it doesnt learn enough to give good result. Hope they will add this soon. If you are interested in working solution please use this repo with TF implementation https://github.com/creotiv/hdrnet First 2 models works ok, pyramid model broken because of problem with shapes in newer version of TF.

As for dataset, it should have not big variation(due to not so big network), its good for one style from +/- similar input(i mean brightness, contrast, etc) Here u can see result of training this network on small dataset(about 50 photos) https://www.slideshare.net/anikishaev/photo-echance-problems-solutions-ideas (slides 25-26)

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 Wed, Sep 4, 2019 at 5:23 AM butterl notifications@github.com wrote:

Hi，Thanks for sharing this project！ Any dataset preprocess script for this project ？ the RAW to jpeg preprocess way affect a lot of the the output.

what about the psnr in you training experiment , I only got 21db from the original google code， not sure if I got somthing wrong in data preprocess

Thanks

— 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/1?email_source=notifications&email_token=AAB5CDK5CJQLDQRGXZLTOL3QH4L2TA5CNFSM4ITMTAK2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HJEDPUA, or mute the thread https://github.com/notifications/unsubscribe-auth/AAB5CDOXJTZVSIG7EZPKKI3QH4L2TANCNFSM4ITMTAKQ .

eduardinjo commented 4 years ago

@creotiv What kind of bug did you experience with pyramids on 1.13? I got it working for 1.14 without errors and even 2.0 worked to an extent although I have some bug with pretrained encoder I'm sure.

creotiv commented 4 years ago

@eduardinjo not with pyramids, but with interpolation on grid_sampler. We have 3d cube where we need to interpolate position, and for this we need trilinear interpolation. Without it it wouldn't work. Of course, we can simulate it, but it will run very slowly

creotiv commented 3 years ago

Ive just fixed this repo. PSNR should be around 33-35. for dimple correction u need luma bins 8 and spatial bins 16, for more harder 16 and 32