image detail lose in general imaging chip

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 846 forks source link

image detail lose in general imaging chip #57

Open zkvz opened 5 years ago

zkvz commented 5 years ago

Hi, cchen156. Thank you for your contribution! we train and test SID in our own DataSet（raw data from general imaging chip）. we find serious detail lose phenomenon especially for text and grass. we tried different CNN structure.

cchen156 commented 5 years ago

This is normally because the input data is not good. Can you check your raw in Photoshop? Can you see the text and grass in Photoshop? You could adjust the exposure in Photoshop to make the content visible.

zkvz commented 5 years ago

Yes, we check our raw data in Photoshop. we find chip's ISP can get more details with little noise and CNN get less details with no noise (about 0.01 lux), CNN get a oversmoothed result. we find the same phenomenon in sony dataset. When the image signal strength is weak, CNN maybe smooth out the details. I don't have any good solution, could you give me some advice.Thank you !