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
918 stars 200 forks source link

About your NIQE result #15

Closed UtopiaHu closed 5 years ago

UtopiaHu commented 5 years ago

The test dataset you provided includes NPE dataset and its 3 extensional datasets (data 11, data12, data 13). Are the NIQE scores of NPE dataset (Table 1, column 4) tested only on NPE dataset? or on NPE dataset together with its 3 extensions as a whole ?

yifanjiang19 commented 5 years ago

This experiment was conducted so long ago and I need some time to check the detail. But you could compare the results with a recent work https://arxiv.org/abs/1905.04161. We got nearly the same results which means the experiments setting is also the same.

UtopiaHu commented 5 years ago

@yueruchen Thanks for your patient reply. I have 2 more questions.

  1. Is your network capable of dealing with images with arbitrary size ?
  2. In your code, what is the difference between 'self.fake_patch' and 'self.fake_patch_1' ? Why not combine them ?
yifanjiang19 commented 5 years ago
  1. Yes, it is fully convolutional.
  2. In the original setting we only random crop 1 patch in a large image. Later we choose to select multiple patches so add another self.fake_patch_1, it is a historical problem, sorry.
UtopiaHu commented 5 years ago

I see. Thank you

UtopiaHu commented 5 years ago

So, if I wanna use the pretrained model to test my own image, I have to feed the image along with its corresponding gray image (self regularization). Am I right ?

yifanjiang19 commented 5 years ago

No, it will synthesize the gray image automatically.

UtopiaHu commented 5 years ago

I have tested your provided model on all subsets. The NIQE score on VV subset differs greatly from what you have reported in Table 1 (4.799 vs 2.581). Is it normal ?

yifanjiang19 commented 5 years ago

I just tested it again and got same results, please check your own problem, probably about 1) NIQE implementation (I use the official version). 2) do not change the resolution of test data.

UtopiaHu commented 5 years ago

I also used the official code of NIQE. When I tested on VV subset, some images raised the error of insufficient GPU memory. So, did you use at least 2 gpu cards during testing process ?

yifanjiang19 commented 5 years ago

@UtopiaHu I used V100 which has 32gb memory. If you want to reproduce our results you need to try a large memory gpu. Or you cloud try to retrain the network with a smaller model (decrese channel number)

UtopiaHu commented 5 years ago

I see, thanks

LIXINJJ commented 3 weeks ago

您好，我运行了模型，但是不知道评价指标如何体现出来，请问一下评价指标是将测试完成的图片，单独用代码测试出来的吗？还是将NIQE的评价指标代码加到模型里面。麻烦您解答一下。