[Boulder]

Originally Posted by HolyWu

It's the viewing device. See https://github.com/Netflix/vmaf/blob...ible-artifacts.

And this seems to confirm that upscaling to that resolution is the way to go, although getting a VMAF score from an originally 1080p source on a 4K display is not optimal.

If, say, for a distorted video of 480 resolution, we still want to predict its quality viewing from 3 times the height (not 6.75), how can this be achieved?

    If the 480 distorted video comes with a source (reference) video of 1080 resolution, then the right way to do it is to upsample the 480 video to 1080, and calculate the VMAF at 1080, together with its 1080 source.

    If the 480 distorted video has only a 480 reference, then you can still upsample both distorted/reference to 1080, and calculate VMAF. A caveat is, since the VMAF model was not trained with upsampled references,
the prediction would not be as accurate as 1).