The writing is great, but I keep wondering: Why do so many physicists use Jet or other non-perceptually uniform rainbow-like colourmaps for their graphs? Shouldn't that be part of great science communication, given that about 1 in 12 men have some form of colour blindness?
There are so many other visually appealing options [0] that are attractive while still being perceptually uniform. There's even Google's Turbo [1] colourmap for those who refuse to give up rainbow-like colourmaps.
Short answer: because none of them is the default in standard plotting tool.
There is so many alternative that it is hard to know which replacement you should pick. If you pick the wrong one you are a moron, if you stick with the bad default it's way less noticeable, even if it is a strictly worse choice.
The default issue is certainly keenly felt. My first scientific publication graphs were done with Jet. But since switching to some of the perceptually uniform ones (most often Viridis), I've never had a complaint.
But your point "if you stick with the bad it's way less noticeable" is understandable.
Since it's showing a probability density, wouldn't it make most sense to just use a grayscale in this case? As the probability density approaches zero the visual would disappear as well. This is more intuitive, as the probability is zero off the edge of the graph/page
I think you only need this color vomit if you have a change between two states, or you have multiple intermediary states and want some visual "clustering".
You can actually see the problem very clearly b/c it ends up having a weird dark-blue to white cut is some of the plots.
The advantage of using a rainbow scale as opposed to a gray scale is that a rotation in hue space brings out more 'detail' than a simple black to white gradient.
I still think it's a bit visually confusing when you're representing the existence or non-existence of something (the planet is in this region, and not anywhere else)
You could in theory do both.. rotate in hue and fade to white. If you did it right, when if you printed it out in B&W you'd get the grayscale equivalent :)
That was a very interesting read. Thank you for sharing. I could really use some of the other color spaces presented too.. in some other situations.. I'll just need to figure out how to plug them into my workflow :)
It's strange but the first linear color space (CET-L03 red-yellow-white) seems to go to a lightness of 100. On the far right it looks white as expected. However the other one (CET-L20) looks very nice, but it doesn't seem to reach 100. Even though in the plots and explanation it suggests it does. Strange..
The more I look into this the more it seems this inevitably (accidentally or intentionally) misrepresents data. Even the author highlights how it does this. All the options are in some ways effectively thresholding/binning ... but in non-systematic ways through the vagaries of the human visual system. Unless this is an art project, there are more systematic and honest ways to bin things.
I can see how it can be useful or even inescapable in some situations.. The most obvious being maps with different content that needs to be distinct and of course these bimodal "divergent color maps". But maps are already half-art :) and bimodal data you're effectively vaguely "pre-binned" the data into two equal halves. There are also situations where the color is adding an extra dimension (like these "cyclic colour maps"). The result with the fingerprint is honestly hard to visually interpret, but I don't see any clear alternative (other than a grid of arrows I guess)
But reading all this stuff leaves me with the impression that colors are dangerous. They really should be last resort. In simple/common situation like the probability densities of the mystery planet - grayscale looks like it's the only really honest option. Side by side with the grayscale, the linear colors don't seem to be adding all that much and are creating artificial "islands" (not to mention the colors are a real blast to the eyeballs)
Comments
The writing is great, but I keep wondering: Why do so many physicists use Jet or other non-perceptually uniform rainbow-like colourmaps for their graphs? Shouldn't that be part of great science communication, given that about 1 in 12 men have some form of colour blindness?
There are so many other visually appealing options [0] that are attractive while still being perceptually uniform. There's even Google's Turbo [1] colourmap for those who refuse to give up rainbow-like colourmaps.
[0]: https://matplotlib.org/stable/tutorials/colors/colormaps.htm...
[1]: https://ai.googleblog.com/2019/08/turbo-improved-rainbow-col...
Short answer: because none of them is the default in standard plotting tool.
There is so many alternative that it is hard to know which replacement you should pick. If you pick the wrong one you are a moron, if you stick with the bad default it's way less noticeable, even if it is a strictly worse choice.
The default issue is certainly keenly felt. My first scientific publication graphs were done with Jet. But since switching to some of the perceptually uniform ones (most often Viridis), I've never had a complaint.
But your point "if you stick with the bad it's way less noticeable" is understandable.
Matplotlib 2.0 came out in January 2017 and has a perceptually uniform, colorblond friendly colormap as default and three others available.
There is no exclude for using Jet today and Journals should reject papers using bad colormaps
"...given that about 1 in 12 men have some form of colour blindness?"
I guess women are better suited to science work after all. ;)
This feels overengineered...
Since it's showing a probability density, wouldn't it make most sense to just use a grayscale in this case? As the probability density approaches zero the visual would disappear as well. This is more intuitive, as the probability is zero off the edge of the graph/page
I think you only need this color vomit if you have a change between two states, or you have multiple intermediary states and want some visual "clustering".
You can actually see the problem very clearly b/c it ends up having a weird dark-blue to white cut is some of the plots.
The advantage of using a rainbow scale as opposed to a gray scale is that a rotation in hue space brings out more 'detail' than a simple black to white gradient.
that's an interesting point
I still think it's a bit visually confusing when you're representing the existence or non-existence of something (the planet is in this region, and not anywhere else)
You could in theory do both.. rotate in hue and fade to white. If you did it right, when if you printed it out in B&W you'd get the grayscale equivalent :)
Maybe it'd look a bit silly
There are colormaps that are designed to do exactly this, for example the linear maps here: https://colorcet.com/userguide/index.html
That was a very interesting read. Thank you for sharing. I could really use some of the other color spaces presented too.. in some other situations.. I'll just need to figure out how to plug them into my workflow :)
It's strange but the first linear color space (CET-L03 red-yellow-white) seems to go to a lightness of 100. On the far right it looks white as expected. However the other one (CET-L20) looks very nice, but it doesn't seem to reach 100. Even though in the plots and explanation it suggests it does. Strange..
You're welcome! I've also found https://github.com/1313e/CMasher
The more I look into this the more it seems this inevitably (accidentally or intentionally) misrepresents data. Even the author highlights how it does this. All the options are in some ways effectively thresholding/binning ... but in non-systematic ways through the vagaries of the human visual system. Unless this is an art project, there are more systematic and honest ways to bin things.
I can see how it can be useful or even inescapable in some situations.. The most obvious being maps with different content that needs to be distinct and of course these bimodal "divergent color maps". But maps are already half-art :) and bimodal data you're effectively vaguely "pre-binned" the data into two equal halves. There are also situations where the color is adding an extra dimension (like these "cyclic colour maps"). The result with the fingerprint is honestly hard to visually interpret, but I don't see any clear alternative (other than a grid of arrows I guess)
But reading all this stuff leaves me with the impression that colors are dangerous. They really should be last resort. In simple/common situation like the probability densities of the mystery planet - grayscale looks like it's the only really honest option. Side by side with the grayscale, the linear colors don't seem to be adding all that much and are creating artificial "islands" (not to mention the colors are a real blast to the eyeballs)