Originally Posted by RLBURNSIDE
I need to do some more reading up on Rec.2020 and Rec.709, because if all you do is increase each colour channel from 8 bits to 12, increasing the steps from 256 to 4096, then why does the green expand out so rapidly, and the others don't. I guess that's because the human visual system is more sensitive to green?
The green doesn't expand out just by changing from 8 to 12 bits. It expands out if you increase the size of your color gamut.
It just so happens that the red and blue primaries chosen for SD and HD were much more saturated than their respective green primaries. Interestingly, back in the NTSC days, the green primary was more saturated than it is in HD (in fact it was as saturated as Adobe RGB is).
So, if anything, one could argue that green has been chronically undersaturated, and we are finally on the threshold of restoring balance.
You do touch upon an interesting point, however.
Take a look at the colors associated with different wavelengths in the visible spectrum:
You can see that green and blue take up quite a bit less of the spectrum than does red.
Yet when we look at the plot of the spectral locus in the CIE xy color space, we see something interesting:
Notice how when you go from 600 nm to 620 nm, there is only a change of about 0.05 in y
. However, when you look at the change between 480 nm and 500 nm, there's a difference of about 0.4 in y
This is what accounts for the fact that green appears to occupy more real estate.
However, it's important to remember that the CIE xy color space is a very abstract space to begin with. It's a projection of the three dimensional XYZ color space, and the X Y Z primaries are themselves transformed from a set of real R G B primaries used in color matching experiments back in the early parts of last century.
Now despite the abstraction involved in the CIE xy plot, there may be some deep perceptual significance behind the asymmetries. My best guess is that it reflects the fact that we experience rather sudden perceptual changes in color at certain wavelength transitions. For example, when you go from 480 nm to 500 nm (look at the first image), notice how there's a sudden change from blue to green. Yet, when you go from 600 to 620, you're still within the reddish region. Accordingly, in a color space whose axes reflect changes in hue, you're gonna accelerate pretty fast along the spectral locus as you traverse the 480-500 nm wavelength transition. You can probably trace these sharp transitions back to the cone fundamentals
, by considering how the cone responses differ between those wavelength transitions, and then taking into account color opponent mechanisms.
Now it's important not to conflate the issue of color change as a function of wavelength
with the issue of perceptual resolution. Just because going from 480 to 500 nm results in a sudden change in color doesn't mean we can't see a huge number of gradations between those colors. It just means that we might have to quantize the wavelength that much more finely compared to when we go from 600 to 620 nm.
As for the question of whether we need more bits to encode a particular color channel, it is instructive to look at a more perceptually uniform color space, such as CIE LUV, where physical distance within the color space more closely matches perceptual distance:
The CIE LUV color space suggests that to get from white to fully saturated blue actually requires a taller perceptual "ladder" than red or green. That is, you would need to encode more individual steps to represent all the visible gradations from white to blue.
As for exploiting this perceptual asymmetry by more efficiently assigning bits to different color channels, I'm not in a position to comment. I can see how the scheme might work if you're encoding in RGB space, but I'm not sure how that would translate into other encoded spaces.
btw, Craig Blackwell has an excellent four
nine! part series on youtube that's a good starting point for those interested in colorimetry and color vision. Here's Part 1
.Edited by spacediver - 2/9/14 at 5:15pm