guptapraful / niqe

NIQE for IQA in python..
223 stars 43 forks source link

NIQE result may be wrong #4

Open ZhangYuef opened 3 years ago

ZhangYuef commented 3 years ago

Thanks for sharing the code! But when I checked the same picture's niqe result by this repo and also Matlab niqe() function, I got different results: 30+ for this repo and 3+ for Matlab.

As this repo declares it uses Matlab pretrained model parameters, so this result gap is strange. I suspect there are some implementation bugs inside this repo. Please double check. Thanks in advance.

mlkk518 commented 3 years ago

Thanks for your sharing. I also have a testing about the Matlab version. Unfortunately, the same images but the big gaps.

qlhai commented 3 years ago

I get a test result. Maybe there exist some errors. image

OldBoysmh commented 2 years ago

Do I need to retrain the multivariate Gaussian model before using it? The original text used 125 images from pristinedata for training, do we need more data to retrain the multivariate Gaussian model. Because I see this function: get_patches_train_features

mlkk518 commented 2 years ago

这是来自QQ邮箱的自动回复邮件。你好,我常不进邮箱有时QQ联系哦。