RGB Luminance revisited

One of the first things you learn in graphics is that green is subjectively brighter than red than is brighter than blue.  This goes back to the old school TV formula for converting color (RGB) to BW TV.  The wikipedia article on this is pretty good, and makes the important practical distinction between non-linear luma and linear luminace that all that gamma business.   But is getting that luminance formula important?  And if so, what is it for modern LCD screens?  I have had to pick a formula for my apps (I use an HSV inspired color space so need a formula for V) and it in practice doesn't seem to matter, but I thought I should dig into this a little now that the betas are done. 

First, always question all assumptions if easy to do so!  Are the primaries really different?  I find this question confusing: which of the colors below is brighter?

So we shouldn't expect an easy or clean answer here, and thinking there is an exact one is probably naive.  But that doesn't mean we shouldn't pick a standard if they are different.  To get at this I use a really cool indirect measure Gordon Kindlmann and collaborators used over a decade ago for designing pseudocolor maps.  See the paper for the roots of this idea.  I use the image from that paper with a few mods for easy photoshopping.  The basic idea is we see something as a face if light and dark are distributed "well".   It's faces and saturated colors and who knows what monitor you are looking at this on, so if you think this is science, you didn't read the disclaimer that is implicit here.

Here are images with all-on RGB (255, 0, 0), (0, 255, 0), (0, 0, 255), and black, white, and neutral grey (128, 128, 128).

Note some clearly look like faces, and some are more confusing.   Where the grey level crosses over so they are both confusing, the saturated color and the grey are about the same "brightness" under this measure.  Note with black, the right part of the pair always looks more natural, and the opposite with white.  If you look in the middle pairs, it does appear true that green is brighter than neutral grey, blue is darker, and red is more uncertain.  So the green > red > blue we always hear is still true.

Using my computer and my eye, here are the cross-over greys:

Cross-over greys are about 76/255, 128/255,  213/255.

This implies the primaries are very different.  For luma (not linear intensities)  and these 3 data points this implies

L = 0.29B + 0.5R +  0.83G

Those don't add up to one, so the non-linearities are in play for sure.   What if we assume gamma = 2.2, and then redo this.  Now the cross over points are: (0.07, 0.21, 0.66).   That's a lot closer to adding up to one, so more what we are used no in messing with linear gamma.  But is that a coincidence or is it useful?  Let's see if it is predictive.  What if we use pure magenta (255, 0, 255) and predict the cross-over.  If we use the luminance formula were we normalize the constants above to get to sum to 1.0:

L = 0.07B + 0.22R + 0.71G

We get for magenta 0.7+0.22 = 0.29.  If we gamma correct that at 2.2 and convert to 0..255 we get (147/255).  Trying it, before I knew this number, I got about 150 as my subjective crossover:

Greys 0, 128, 150, 255

So it looks like for this half-baked test, the luminance formula for sRGB is not far off:

L = 0.0722B + 0.2126R + 0.722G

Actually it's suspiciously close.  Either the sRGB standard is really making monitor calibration/standardization finally kind of happen, or I got lucky. 

Note that while this post was accidently a post endorsing doing gamma right, I actually believe that what matters in practice is not doing gamma wrong.  But doing it "right" is one way to make sure you don't do it wrong (thinking 128/255 is half as much physical intensity as 255/255).