lidq92 / cnniqa Goto Github PK

Validation Results - Epoch: 431 SROCC: 0.0865 KROCC: 0.0628 PLCC: 0.0553 RMSE: 27.1083 MAE: 22.8260 OR: 86.42%
Testing Results - Epoch: 431 SROCC: -0.0685 KROCC: -0.0521 PLCC: -0.0312 RMSE: 26.2257 MAE: 21.7633 OR: 86.88%
Why is my SROCC value so low

Hi! check out Pytorch Lightning as an option for your backend! We're looking for awesome project implemented in Lightning.

Your project will be really easy to maintain on Lightning!

https://github.com/williamFalcon/pytorch-lightning

我看生成对抗网络都是用于图像增强或者图像生成的，但是不知道标签变为数字（主观mos值），以及生成的东西也变为数字不知道是否可行

我用matlab改了一下就pycharm就调用不了了

Hello good author,
First Thank you so much for your amazing code, can I get some help please
i tried changing the dataset with another one of 3D sterio images ( that contains right and left images ) and also tried fitting the data like what you did in your code and i made datainfo.mat file.
But i still get compatibily issues between the code and the data i prepared, probably caused by a missundestanding of something on my behalf.
That's why i wish for you to answer some questions that i have in mind please.
Why exactly did you choose 1000 senarios? Why not another nuber?
And also what is the role of the variables exp_id ? How does it help the code's progress?

thank you.

FileNotFoundError: [Errno 2] No such file or directory: '/data/databaserelease2/jp2k/img2.bmp
请问这是什么原因呢？

为什么在进行model训练和测试的时候读取的图片都是distorted image呢？live里面一共982张图片，779张distorted image，其余混杂着reference image，为什么在训练测试model的时候要剔除这些reference image呢？这样对model训练是不是可以提升准确性呢？谢谢大佬回答。

你好，一般在计算srocc的时候，需要将客观数据进行线性映射，我在您代码里没有看到有映射的步骤，不知道是不是理解有误。

当做跨数据集测试的时候，在live上训练，在csiq和tid2013上测试srocc和plcc结果，plcc的结果需要进行非线性映射吧，真的求教一下这个非线性映射怎么做呢，不做的话对plcc的结果影响差异很大吗。有些论文要做1000次才得到最后的结果，那个这个过程到底怎么做呢，麻烦您了，感谢。

→← 新手看代码比较费劲
＾

请问，能不能把几个数据集也传上来啊，LIVE数据集的官网国内打不开

Dear Dingquan,

Thanks a lot for your wonderful work, and it helps a lot.

When I tried to execute the main.py file under the command of CUDA_VISIBLE_DEVICES=0 python main.py --exp_id=0 --database=LIVE, it seems that ignite module is missing.

So, I searched a lot to debug, like pip install pytorch-ignite, but it won't work.

Is there any way to install this module?

Thanks in advance and have a nice day!

Traceback (most recent call last):
  File "main.py", line 32, in <module>
    from ignite.engine import Events, create_supervised_trainer, create_supervised_evaluator
ModuleNotFoundError: No module named 'ignite.engine'

Best,

Bruce Shuyue Jia

如题，我看了很久，没想明白
TID2013fullinfo.mat 数据里index数组的意义，数组维度1000*25,

Hi there,

I guess that the for statement in line 81 should be changed from
for idx in range(len(self.index)):
to
for idx in self.index:
so that it does not iterate through the same images in training and testing processes anymore.

In the first case the len(self.index) is always smaller than the len(ref_ids), so the algorithm does not involve files with idx > len(self.index) here:
im = self.loader(os.path.join(im_dir, im_names[idx]))

Greetings,
Emilijus

您好，我是一名做图像质量评估的学生，刚刚入门，最开始就是使用您这个代码进行了测试，由于数据集不一样，我自己进行了主观打分实验制作了数据集，想要将数据集封装成和您这个一样的mat格式进行训练，我是使用matlab制作了mat文件。训练时出现的这样的错误，不知道是训练的问题还是数据集的问题。您有什么好的解决方法吗？

数据集mat文件如下：

如果我的数据集制作过程不对，请您给我指正一下可以吗？

I want to change the model, and have successfully implemented my model to run in main.py. But in test_demo.py and test_cross_dataset.py, file CNNIQA-LIVE is used. I wonder how can I turn my own models into such files to use them in test_demo.py and test_cross_dataset.py ?
Thanks a lot

Hi,

I can't seem to figure out the mapping of the output result.
Is the valid range 0 to 100?

I run the test_demo.py with some pictures and get some numbers between 20 and 50,
so what these numbers means?
they means the scores of image quality?
and the full mark is 100?

Hello author,
Thank you very much for your code, I have encountered a problem in use, I installed all the packages as required, the code can be trained normally, but the error is reported’Failed to save model graph:‘.I can't find a solution and hope the author can guide me。

hello,i encounter a question.
Can you tell me how to solve this question,thank you.
RuntimeError: input and target shapes do not match: input [1], target [1 x 1] at c:\programdata\miniconda3\conda-bld\pytorch_1524543037166\work\aten\src\thcunn\generic/AbsCriterion.cu:15

可以用什么打开呢

Hello! We kindly invite you to participate in our video quality metrics benchmark. You can submit CNNIQA to the benchmark, following the submission steps, described here. The dataset distortions refer to compression artifacts on professional and user-generated content. The full dataset is used to measure methods overall performance, so we do not share it to avoid overfitting. Nevertheless, we provided the open part of it (around 1,000 videos) within our paper "Video compression dataset and benchmark of learning-based video-quality metrics", accepted to NeurIPS 2022.

作者您好！看了您的CSIQ的mat文件，里面只有750张图片，不包括contrast损失类型的图片，这是为什么啊，谢谢您的解答。

Correct me if I'm wrong
The lower the score the higher the quality of the picture
Why is 2.jpg have better quality than 1.jpg?

1.jpg

2.jpg

The Result I get are:

只要改readme里面的对应的接口就可以了是吧？