Using SSIM to predict viewer MOS

[HulkHoganRules]

I guess this question is slightly off topic from the discussion here but I really don't know where else to ask this.

I'm trying to come up with an easy (and very simple) way to predict viewer MOS from the average SSIM metric produced by MSU VQMT.

The original SSIM paper by Zhou Wang on page 610 figure (d) shows a logistic function and relationship between SSIM and viewer MOS: http://www.cns.nyu.edu/ftp/lcv/wang03-reprint.pdf

This is where I get confused as the maths part confuses me. The MSU site states a Pearson correlation coefficient of 0.937 for SSIM: http://www.compression.ru/video/codec_comparison/01_subject_codec_comparison/subject_codec_comparison_part3.htm

The simple approach I want to adopt is SSIM on a scale from 0 to 1 on the x axis and MOS values from 1 to 5 on the y axis. How do I go about using this coefficient to map the logistic function like Wang showed?

[Dark Shikari]

The p-value represents how well the logistic function models the data -- i.e., the error relative to the data, or, equally, how much of the variance in the data is accounted for by the function. I don't think it will help you better model the data; it just serves to measure how good the model is.

[HulkHoganRules]

That makes much more sense now. I did write a long question here but I found the SSIM mappings I was looking for in this paper (if anyone is wondering): Impact of Frame Rate and Resolution on Objective
QoE Metrics

[xooyoozoo]

I'm trying to come up with an easy (and very simple) way to predict viewer MOS from the average SSIM metric produced by MSU VQMT.

You really can't. There's a reason why most subjective quality papers come with extremely detailed descriptions of setup and most psychovisual models come with variables on cycles per degree and other viewing conditions: subjective opinions are dramatically dependent on environment.

There's a few resources (http://sse.tongji.edu.cn/linzhang/IQA/IQA.htm) that map metric values with MOS on a variety of image databases. You'll find that the relationship between opinion scores and metric values vary wildly between databases, and the only really good way to measure metric proficiency is correlation and error (which is what most researchers ever publish).

Transitioning from image stills to film sequences then adds another degree of uncertainty to something inherently uncertain, as almost every metric used in video analysis work solely in the spatial domain. The ones that account for motive distortions are bewilderingly slow, which makes them useless despite any potential abilities (http://vision.okstate.edu/stmad/).