[css-color-hdr] In what colorspace is `eps` = 0.001 one just noticeable difference?

[svgeesus]

Perhaps the answer is in [ISO_21496-1] but I have not purchased a copy.

(W3C has an informal liaison with ISO TC 42/WG 23 and is establishing a Class C liaison which includes document sharing, but that has not been completed yet).

[svgeesus]

@ccameron-chromium this eps is from your proposal, can you explain?

[svgeesus]

@ccameron-chromium your original proposal says:

Let eps=0.001 (one linear just noticeable difference)

Since linear-light is not perceptually uniform it seems off to claim a perceptual distance metric applies to it. I would solve this by deleting the parenthetical, but it would be good to explain what eps is actually for.

[ccameron-chromium]

Sorry, I had thought that I'd written this down, but I hadn't.

I also see that I wrote "let eps=0.001 (one linear just noticeable difference)" in the spec. And I don't know why I wrote that. Epsilon is unrelated to JND, sorry.

Background:

The epsilon parameter in color-hdr is to match a parameter in ST 2094-50 (in development, trying to get it out to the public ASAP, it overlaps a lot with this ICC proposal), which in turn is to match a parameter in ISO 21496-1 gain maps.

The idea with gain maps is that the gain map stores the value as:
G=log(HDR/SDR)
and then reconstructs HDR from SDR as
HDR=SDR*exp2(G)

If you're starting with an HDR and SDR representation and computing G, this can start getting poorly conditioned, so we let there be the eps parameter so one can store
G=log( (HDR+eps) / (SDR+eps) )
to avoid divide-by-zero issues. And so then you have the reconstruction is
HDR=(SDR+eps)*exp2(G)-eps

The most extreme end of this is a place where the HDR color is non-zero, but the SDR color was zero.

In this specification:

This parameter will only matter when a component needs to go from zero to non-zero.

This is an example of going from 1 to 2, and epsilon makes no difference.

This however is going from 0 to 1, and it makes a huge difference.

How often do we go from 0 to 1? Effectively never. But zero-or-near to non-zero of an individual component can indeed happen. I think a default of 0.001 is fine for those. A more impactful factor is the gain application color space. The spec currently says XYZ, but most images and videos will use sRGB, P3, or Rec2020, so we may want to look into that.

[svgeesus]

Thanks for the detailed response, makes sense.

[svgeesus]

The epsilon parameter in color-hdr is to match a parameter in ST 2094-50 (in development, trying to get it out to the public ASAP, it overlaps a lot with this ICC proposal), which in turn is to match a parameter in ISO 21496-1 gain maps.

Follow-on issue

• The eps = 0.001 factor, and SMPTE-ST-2094-50 #12873