search

VITA-Group / EnlightenGAN

[IEEE TIP] "EnlightenGAN: Deep Light Enhancement without Paired Supervision" by Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, Zhangyang Wang
Other
890 stars 198 forks source link

EnlightenGAN-N #55

Closed TaQuangTu closed 4 years ago

TaQuangTu commented 4 years ago

First of all, thank for your great work. When I use your model to enhance low light images in night sets of KAIST dataset before pushing them to detection model, detection accuracy becomes worse (compared with using origin images). From my perspective, I see the improved images contain more noise than its original. I think it is caused by your model does not adapt to the dataset. Also your paper says that EnlightenGAN-N is domain-adapted version of EnlightenGAN, how can I try that model? Thank you!

Updated:

KAIST is a pedestrian dataset contains more than 90000 images captured from moving car on streets at both day and night time. Image's resolution is 640x512.

yifanjiang19 commented 4 years ago

I think KAIST is an extremely dark dataset which contains much severe noise and is not suitable for our task. I suggest you add another denoising module such as DNCNN or BM3D. But if you insist to have a try, you can directly replace the dark dataset with KAIST data, and tune some hyperparameter like -vgg.