zwx8981/unique

The repository for 'Uncertainty-aware blind image quality assessment in the laboratory and wild' and 'Learning to blindly assess image quality in the laboratory and wild'

Weixia Zhang (张维夏)👋

I am an Associate Research Scientist at AI Institute, Shanghai Jiao Tong University. Currently, I work on perceptual quality evaluation and enhancement for visual content produced by various manners, i.e., PGC, UGC, and AIGC.

unique's Issues

Normalization for testing images

Hi,

I am wondering why you do normalization on testing images?
transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))])

Could you share me the result from matlab?

@zwx8981 Thanks for your sharing. I want to train the dataset, but I have no matlab, Could you share me the result from data_all.m . Thank you very much!!

about the BID datasets

Thanks for your great job! Could you please provide a download link for BID dataset? I cannot find it on the internet. Thanks very much

How are the cross-dataset experiments conducted?

Hi,

Thank you so much for the amazing work! I have a question regarding Table IV in the original paper.

In the table, baseline models are trained with some datasets and tested on the other datasets. How are the cross-dataset experiments conducted given that the MOS/DMOS ranges are different across datasets? Do you re-train the FC layers for each dataset during testing? Thanks.

About the LIVE dataset

Thanks for your great job! Could you please provide a download link of "dmos_realigned.mat" of LIVE? I cannot find it on the internet. Thanks very much

关于demo.py的测试问题

您好，非常感谢您的工作！

我想测试自己的数据，但没有mos和std的值，可以直接使用demo.py来得到测试结果吗？我是否需要重新训练模型？

十分期待您的回复。

About the dataset that is lack of the variance data

Hello, the model seems work very well on dataset such as tid2013, koniq10k etc. But there is a problem that what if the variance of MOS data is unknown?

bilinear pooling layer

Hi,
you proposed to use bilinear pooling instead of first-order average pooling in the original resnet, as mentioned in the paper.
However, I found the average pooling is kept in the implementation of resnet (BaseCNN.py). A bilinear pooling layer is added on top of that, no?

Negative predicted score

Hi, I simply train on single koniq-10k dataset with mos between 1 -5 , I got negative predicted score on some images. Do you think that is possible?

Model validation

Thanks for your great job!
Did you use validation data to obtain the best checkpoint (model)?

请问怎样测试单幅或多幅自己的图像质量？

怎样调用训练好的模型测试单幅或多幅自己的图像质量？貌似并没有可用的模型文件

Input scaling

I noticed that there is no scaling operation for input (./255) before applying mean-std normalization (mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)) . Why not firstlt scaling input?

[Feature requested] Pyhon3 script for generatin a training set

Thanks for the excellent work! Could you please provide a Python3 version of scrit to generate training set? Because Matlab is very expensive...

About the mean output

Hi, thanks for your significant work. In this work. a softplus function is applied to the std output. However, there is no constrains to
the mean output. Does the performance will be influenced If a sigmoid function is applied to the mean output?

about the usage of "detach()"

Hi, sorry to have interrupted again. I wonder why we need to add detach() when computing the probability and the hingeloss.

p = 0.5 * (1 + torch.erf(y_diff / torch.sqrt(2 * y_var.detach())))

            std_label = torch.sign((gstd1 - gstd2))                self.std_loss = self.std_loss_fn(y1_var, y2_var, std_label.detach())<!--EndFragment-->

Output Range for your model?

Hi, I am performing inference using your model for my dataset. For my dataset, the maximum score I get is 1.1834485530853271 and the minimum score is 0.5403027534484863. What is the output range for your model? Is it [0-1] or [1-5] or [1-10]?

Training UNIQUE on single dataset

Hi,
I have a question about Table VII in your recent arxiv paper.
I noticed that you did single-dataset training on baseline model (regression) to compare with UNIQIE model trained on multiple datasets, as shown in Table VII.
Why did not train UNIQUE model on one single dataset?
Thanks in advance!

Line70 in DBCNN.py

Hi, if std modeling is false, I think the outdim of features should be 1.
Could you please check line70 in DBCNN.py? if the dim of X is 2, then this case is only valid for std modeling mode, right?

About the input size in training and testing time

In the paper, it is said that during training the input size is set to 384 x 384 for all the images from all the databases, while that during the test, the network will inference on the original size. What if the test size is also 384 x 384? Will this affect the performance?

test new dataset

Hello, may I ask, if I want to test the quality of the image on the coco dataset,Can I test the coco data set directly with the pre-training model you gave，direct the path of the test picture in the demo to the coco dataset，or need to train the model from scratch?
Thank you very much for your reply