Question on the meaning of the result.

[chenkang455]

Hi, thanks for your great work. I am currently working on low-level vision and trying to utilize LIQE for testing the quantitative performance the restored image. I use the following code:

metric = pyiqa.create_metric('liqe',as_loss=False).cuda() 
tensor([1.0016], device='cuda:0')

I am a little bit confused about this result, can you be kindly explain the meanning of this result like PSNR or SSIM ? (its range and higher better or lower better)

[zwx8981]

Thanks for your interest ! The higher LIQE output the better. It ranges in 1 to 5

[chenkang455]

Thanks for your interest ! The higher LIQE output the better. It ranges in 1 to 5

thanks for your response. I find that my result ranges from 1 to 1.01, is that normal?

[zwx8981]

@chenkang455 Not that normal. Can you send me some images with both low and high quality in your case? I will see what's the problem.

[chenkang455]

@zwx8981 , hi these are my test images.

[zwx8981]

@chenkang455 Thanks for your sharing ! Let me explain it step by step.

1, Although it is clear that the bottom image is of higher quality than the top one in your case, neither images are with high quality in a general context of "image quality". Even the bottom image is somewhat blurred. 2, I would suggest using liqe_mix instead of liqe by pyiqa.create_metric('liqe_mix', as_loss=False), which is trained on multiple image quality assessment datasets, leading to better quality prediction accuracy. 3, One last important thing, the resolution of these two images are quite different from the images used to train LIQE. I would suggest resize the image (keep aspect ratio) such that the short edge is 384.

Combing the practices of 2 and 3, liqe_mix would give about 1 to the top image and 2.4 to the bottom image.

Best

[chenkang455]

@chenkang455 Thanks for your sharing ! Let me explain it step by step.

1, Although it is clear that the bottom image is of higher quality than the top one in your case, neither images are with high quality in a general context of "image quality". Even the bottom image is somewhat blurred. 2, I would suggest using liqe_mix instead of liqe by pyiqa.create_metric('liqe_mix', as_loss=False), which is trained on multiple image quality assessment datasets, leading to better quality prediction accuracy. 3, One last important thing, the resolution of these two images are quite different from the images used to train LIQE. I would suggest resize the image (keep aspect ratio) such that the short edge is 384.

Combing the practices of 2 and 3, liqe_mix would give about 1 to the top image and 2.4 to the bottom image.

Best

I have solved this problem. Thanks for your advice and your work is solid indeed！

[zwx8981]

Good to know that !

[tuvovan]

@chenkang455 Thanks for your sharing ! Let me explain it step by step. 1, Although it is clear that the bottom image is of higher quality than the top one in your case, neither images are with high quality in a general context of "image quality". Even the bottom image is somewhat blurred. 2, I would suggest using liqe_mix instead of liqe by pyiqa.create_metric('liqe_mix', as_loss=False), which is trained on multiple image quality assessment datasets, leading to better quality prediction accuracy. 3, One last important thing, the resolution of these two images are quite different from the images used to train LIQE. I would suggest resize the image (keep aspect ratio) such that the short edge is 384. Combing the practices of 2 and 3, liqe_mix would give about 1 to the top image and 2.4 to the bottom image. Best

I have solved this problem. Thanks for your advice and your work is solid indeed！

hey @chenkang455 may I know how did you solve the problem? thanks!

[chenkang455]

Just follow the steps explained by authors.