search

cchen156 / Learning-to-See-in-the-Dark

Learning to See in the Dark. CVPR 2018
http://cchen156.web.engr.illinois.edu/SID.html
MIT License
5.44k stars 845 forks source link

Why not learn the amplification ratio? #89

Open manfred-git opened 5 years ago

manfred-git commented 5 years ago

Why not set the ratio to 1 and let the network learn a ratio when training the model?

I noticed that the author set the ratio to "gt exposure time / short exposure time" when training the model, and we need to set a ratio externally when testing the model.

However, after trying different ratios, I found that the result images had different qualities besides different exposure, and there exists a perfect ratio, which we don't know before testing.

I assume that the ratio we choose affects the network performance, since the ratio is applied to the image before it is put into the network. So why not let the network learn a ratio?

Garfield-kh commented 4 years ago

Same question here, have you see any reasearches work on this problem?

suhas1999 commented 2 years ago

there are few works that explored this. Here I am tagging one of those "Restoring Extremely Dark Images in Real-Time" CVPR2021