[MEncoder-users] Codec-independant objective quality measurement

[Loren Merritt]

On Sat, 5 Aug 2006, Reimar Döffinger wrote:

> chroma downsampling is not really an artefact, it is part of the
> compression that is based on the fact that the human eye is really bad
> at seeing precise colour structures. Based on experience with my own
> eyes, weighting chroma with 1/4 is still too much ;-) - though that's
> not entirely correct, a general bias in chroma is usually easy to spot,
> but even a heavy loss of details is hardly visible, whereas with luma a
> bias (i.e. everthing's a bit too bright) is hardly visible whereas loss
> of details is. So using different ways to measure quality might be worth
> thinking about...

SSIM does define distortion in terms of a mean-shift component and a 
detail component, which could be weighted separately.
It was developed based on a study of greyscale images, so it was never 
intrended to apply to chroma. And the computation of mean-shift is wrong 
for chroma: it currently assigns higher weight to a given amount of shift 
in dark areas (which is good), but low chroma value means green not dark.

--Loren Merritt