Hacker News Reader

Comment by Syzygies | Hacker News Reader

As a mathematician I'm hypersensitive to people overreaching with theory. It's often the King's clothes. There are indisputable reasons why a musical fifth sounds good, but vision is more organic. Evolution has left us with many layers of color perception, like an out-of-control code base. Sure, anything can be fit to a model if we ignore the parts that don't fit; that's how theory intoxicates.

My father devised the "Bayer filter" for digital photography, and favored green for practical reasons that have stood the test of time. There is some theory that applies here, it's just not the same as harmonics and sound.

I'd love to code up a machine learning project that showed the user many color combinations, responding to feedback. Willem de Kooning for example painted with an extraordinarily original palette. Could we tailor our own color palettes, with an AI assist?

View on HN

All theory is overreaching. It's called the Problem of Induction. All you can do is hypothesize, test, and revise. Mathematicians will never be happy about it.

It will always be worse the more complicated something is. One particle is hard. A hydrogen atom is harder. Oxygen is just this side of impossible. Molecules are impossible and you just start throwing away vast numbers of terms. By the time you get to an eyeball, much less a brain, you are basically just guessing and hoping you can cover some of the cases some of the time.

> I'd love to code up a machine learning project that showed the user many color combinations.

I teach painting in an art school. The huge problem with almost all pallet choosing apps (e.g. Adobe's https://color.adobe.com/) is that they produce swatches: a small collection of discreet color values (e.g. red, green and yellow). These would present as peaks in a hue histogram. These swatches would be great for choosing a color scheme for your kitchen, but are not suitable for a painter.

I have looked at more painting hue histograms than any man alive, and I can tell you that almost all painters employ ranges of colors. The hue histograms present as 'pie slices' on the histogram, not peaks. Hence: a wide pie slice of greens (plural) and s narrow pie slice of reds. Show me a color picker like that and I would likely share it with my students. As it stands, I warn against them.

Most modern dedicated painting apps (Krita as an example) support continuous gamut masking. There are also several third-party pickers for non-dedicated apps (ex. Photoshop) that support it.

> I'd love to code up a machine learning project that showed the user many color combinations

That would be fine, so long as you don't use an RGB display to display those colors.

The RGB color space is severely limited when compared to the colors a person can see. This becomes obvious as one plots color spaces on a CIE 1931 diagram.

Here's a simple tool to see some of this (recommendation: move "system gamut contrast" slider to 50%). Select a working color space from the pulldown to see how much of the visible range it encloses (or discards).

https://demonstrations.wolfram.com/CIEChromaticityDiagram/

A typical REC709 color space throws away over half the colors a person can see. In other words, as I said, using a monitor to run color experiments can be a serious mistake. Color Science can be strangely complex.

There is a company trying to push the envelope in display technology and encourage industry to move primaries beyond RGB. The results, based on prototypes, can be impressive:

https://6pcolor.com/

Disclosure: We have done color science, software and hardware consulting for 6P.

If not under an NDA, what are they using to represent the super-bright reds and greens? Lasers? And, more interestingly, the super-greens and super-blues that are very dark and highly-intense?

(I worked for a printing house and I know that some colors are outside RGB and outside CMYK, so they are printed as spot colors. There is a ton of Pantone colors you cannot mix, and must add from special one-color cans, like the Reflex Blue. Some customers want them badly.)

RGB lasers are already a thing for high-end projection. They have one picture of projectors on their website[1], I'll guess they are modified off-the-shelf ones with custom lasers. You can just barely make out the brand Christie on them, and that is a manufacturer who have laser projectors[2].

Considering their main focus is in the new primaries, it wouldn't surprise me if they just used the standard lasers for red and green, maybe even blue; sourcing suitable high-power lasers in some specific wavelengths might not be that easy.

Besides projectors, the other thing they seem to have done is a "dvled" (direct view) display, stuff thats maybe most familiar as those outdoor signage displays. This press release[3] has one blurry picture of their four primary led display

[1] https://6pcolor.com/about/

[2] https://www.christiedigital.com/about/display-technology/las...

[3] https://www.einnews.com/pr_news/571752479/6p-color-demonstra...

It's funny that they call their tech "Full Color Range™" while it still catches smaller part of gamut that some industry standard colorspaces (ACES 2065-1) do with three primaries :)

Of course improving capture or display tech is good still, I'm just less convinced of the idea of trying to upend the whole industry

> smaller part of gamut that some industry standard colorspaces (ACES 2065-1) do with three primaries

The ACES 2065-1 primaries are not realizable, they are imaginary. In other words, one cannot make a display capable of displaying these primaries. They are mathematical tools enclosing the entire CIE 1931 space.

Yes, but of course it is debatable how relevant realizability is in practice; we have long history of using displays that have significantly smaller gamuts than whatever nominal colorspace being used.

Was rec2020 realizable in practice when it was created? I'd bet no.

Sorry, I think you might be mixing up a couple of ideas. In this context, not-realizable means “not real”. Rec 2020 primaries have always been realizable. ACES AP0 primaries only exist mathematically for the purpose of enclosing the entire “wedge” of visible colors.

I though the green in the bayer-filter was over-represented because the human eye is more sensitive to green compared to other colours?

Exactly. It's the best proxy for grayscale detail.

One could have balanced RGB in a hexagonal grid. That's harder to build and slower to process.

He would have considered a hex grid, but our rectilinear cognitive bias was pretty entrenched. Before this work, he programmed an entire image processing system in the "ed" PDP editor, manipulating character arrays. That experience influenced both the Bayer filter and Bayer dithering. And it taught me to look for simple ways to do things.

Theres also the fact that humans have wide variety in vision

Even ignoring color blindness, there's the fact that some people see significantly more range of colours than others (I am one).

Women tend to be more sensitive to slight variations than men (surmised to be evolution ensuring they can better gather from plants , knowing ripeness etc). I see even more variety than that norm.

Human vision is funny, because the majority actually don't see much, with some seeing even less, and some seeing so much they can perceive parts of the UV and Infrared spectrums (not pleasant since our world is absolutely not designed for such people).

It seems unlikely that AI could actually do well accommodating all the variations and would almost certainly be focused on averages while ignoring the color blind and "advantaged".

> There are indisputable reasons why a musical fifth sounds good, but vision is more organic.

Once could argue that our eyes can't even "hear" a fifth. The amount of data our ears extract out of basically two points is amazing. Our eyes in the meantime take an entire spectrum and come up with "blue-ish".

I think you're pointing at the fact that there are only three kinds of cones, whereas the ear has … well honesty I'm not sure how many different kinds of hair cell there are. apparently [16,000](https://www.nih.gov/news-events/nih-research-matters/hearing...)!

the eyes are much better at directionality though; probably some kind of trade-off going on there.

Well it's position vs frequency, so a tradeoff is forced by the Heisenberg uncertainty principle. I don't know if either system is actually reaching this limit though.

I thought about that, but you could just cram more cells to get higher resolution. Reflecting a bit more, hair cells vary by length, whereas each new cone requires a unique protein, so I'd expect hearing to respond to more easily to evolutionary pressures.

"All models are wrong, but some are useful." One has to always remember about both of these properties.

> hypersensitive to people overreaching with theory

What’s left of the scientific method if people don’t? I think reproducibility and inertia („science progresses one death at a time“) are problems, but I hadn’t thought that induction could be one.

Maybe off topic, but I'd love to know more about those "practical" reasons for having two green bits in the Bayer system. I guess to your point, they needed four readings and might as well double up on the band humans have more cones to detect.

But also I wonder how those kinds of decisions end up coloring so to speak AI. Training a model on image compression designed to fool the human eye is basically forcing it to reverse engineer the human eye.

Reply on news.ycombinator.com