Hi, thanks for sharing this great work!

It is really an inspiring idea to construct a series of mixup-like images to evaluate the reliability of various NR-IQA metrics.

Although the rationale is clear, I am still confused about how the "reliability" in Figure 4 is computed. Specifically, how do you measure/quantify the degree of a certain NR-IQA metric follows the monotonicity law? Is it the ratio of the samples in the dataset that totally follow the monotonicity through a certain NR-IQA metric, or some more well-defined scores?

Would appreciate it if you could share your ideas/code on that! (sorry if I miss it in the repository)