View Single Post
Old 2nd March 2019, 02:11   #75  |  Link
poisondeathray
Registered User
 
Join Date: Sep 2007
Posts: 5,346
Quote:
Originally Posted by Iron_Mike View Post
well "joe public" ultimately watches the content... I can tell the diff between CRF 10 and CRF 18 (frame per frame pixel peeping), but I also evaluate content on fully calibrated screens...
It's an assumption . Audience might videophiles, or doom9ers, or you might be doing these tests for you

Quote:
problem is w/ "scientific metrics" is that they often not relate a lot to the HVS (Human Vision System), which is the only thing that matters when humans watch the streamed content...

Yes, pros/cons to every measure , but there are other HVS modelled metrics.


Quote:
VMAF attempts to address that with their sample data... question always are if enough people were sampled, what kind of people (gender/age/race/ethicity - diff between European and Asian samples etc) and the sample procedure was done as best as possible...
It's just that the RD curve characteristics limit VMAF's usefulness in some situations , since it's trained on higher CRF ranges.

So another way to phrase it - is the data set is not valid at higher bitrates. You cannot apply VMAF at higher bitrates because it was trained at CRF 22-28
poisondeathray is offline