1492: "Dress color"

This forum is for the individual discussion thread that goes with each new comic.

Moderators: Moderators General, Prelates, Magistrates

FancyHat
Posts: 341
Joined: Thu Oct 25, 2012 7:28 pm UTC

Re: 1492: "Dress color"

Postby FancyHat » Sat Feb 28, 2015 8:14 pm UTC

sevenperforce wrote:I always tell people that moonlight is the actual color of sunlight, but sunlight is so much stronger that our view of it is dominated by the rayleigh scattering that knocks out the blue photons in favor of the yellow ones.

Uh?

Surely the same proportions of the same frequencies of moonlight and sunlight get scattered equally, so that couldn't itself be the explanation for the apparent colour differences.
I am male, I am 'him'.

User avatar
sevenperforce
Posts: 658
Joined: Wed Feb 04, 2015 8:01 am UTC

Re: 1492: "Dress color"

Postby sevenperforce » Sat Feb 28, 2015 8:20 pm UTC

FancyHat wrote:
sevenperforce wrote:I always tell people that moonlight is the actual color of sunlight, but sunlight is so much stronger that our view of it is dominated by the rayleigh scattering that knocks out the blue photons in favor of the yellow ones.

Uh?

Surely the same proportions of the same frequencies of moonlight and sunlight get scattered equally, so that couldn't itself be the explanation for the apparent colour differences.

Indeed, but like gmal and ucim pointed out, dim light is perceived as bluer because blue cones are more sensitive. Sunlight saturates our rod cells because it is far more intense than moonlight and so it inhibits that response. The end result is that we don't perceive the Rayleigh scattering effects on moonlight, whereas the Rayleigh scattering effects are ALL we perceive when we look at sunlight. So our perception of moonlight is much closer to the actual color of sunlight than our perception of sunlight is.

User avatar
djn
Posts: 610
Joined: Mon May 07, 2007 1:33 pm UTC
Location: Oslo, Norway

Re: 1492: "Dress color"

Postby djn » Sat Feb 28, 2015 9:04 pm UTC

orthogon wrote: But some people seem to be saying that they do more than this: they seem to automatically compensate for the settings of somebody else's camera: they "see it as overexposed" in the same way as I see the chessboard square as being in shadow. Apparently a lifetime of looking at photographs (and, especially, bad photographs) has made these corrections second nature, so they don't really even perceive the uncorrected image at all. This, I find surprising, to say the least.

Edit: an unmatched parenthesis can etc. etc.



I think that might be part of the reason I see it as a bright dress in darkness, yeah - I've taken more than enough pictures with badly lit foregrounds but bright backgrounds (e.g. if the fill flash doesn't go off), and that exactly what this looks like (but isn't). The black/white thing at the bottom/left would then be in the brightly-lit background, and the dress would appear much darker than it really is because the camera exposed for the total scene, not just the focus point. The color balance in a scene like that could easily be way off, too.

Still, it looks lavender, not white.

User avatar
gmalivuk
GNU Terry Pratchett
Posts: 26822
Joined: Wed Feb 28, 2007 6:02 pm UTC
Location: Here and There
Contact:

Re: 1492: "Dress color"

Postby gmalivuk » Sat Feb 28, 2015 9:25 pm UTC

orthogon wrote:But some people seem to be saying that they do more than this: they seem to automatically compensate for the settings of somebody else's camera: they "see it as overexposed" in the same way as I see the chessboard square as being in shadow. Apparently a lifetime of looking at photographs (and, especially, bad photographs) has made these corrections second nature, so they don't really even perceive the uncorrected image at all.
The vast majority of pictures taken these days are with phone cameras, so "the settings of someone else's camera" are not really distinct from the settings of my own, in that both cameras see the world through a tiny little lens and tend not to leave a whole lot of room for manual adjustment on the part of the picture-taker. But even back when most people were still taking pictures with their own dedicated photographic devices, surely everyone living in a technological society here at the beginning of the 21st century has seen their share of overexposed pictures, right?

I absolutely perceive the uncorrected image. I perceive it as being a bad photograph of what is most likely a black and blue dress. The brightened version I've seen passed around as an attempt to justify the "white and gold" interpretation just looks like a worse photograph of what is most likely a black and blue dress. (It looks bad enough, in fact, that it's outside the range of what even fairly shitty modern cameras are likely to do on their own, so I think even in isolation I would be pretty confident that it had been manipulated on a computer after being taken.)
Unless stated otherwise, I do not care whether a statement, by itself, constitutes a persuasive political argument. I care whether it's true.
---
If this post has math that doesn't work for you, use TeX the World for Firefox or Chrome

(he/him/his)

User avatar
Klear
Posts: 1965
Joined: Sun Jun 13, 2010 8:43 am UTC
Location: Prague

Re: 1492: "Dress color"

Postby Klear » Sat Feb 28, 2015 10:18 pm UTC

This gif helped me immensely to understand the white/gold view:

Image

After a while, I began to see even the first frame as blue and black, but at least now I get it.

Pseudoedit: Actually, I think I've already forgotten the feeling I got when I saw it first. Only the memory of there being a difference of perception remains. Damn my brain!

User avatar
Angua
Don't call her Delphine.
Posts: 5941
Joined: Tue Sep 16, 2008 12:42 pm UTC
Location: UK/[St. Kitts and] Nevis Occasionally, I migrate to the US for a bit

Re: 1492: "Dress color"

Postby Angua » Sat Feb 28, 2015 10:47 pm UTC

That's a cool gif. I can't see the start of it as being other than white and gold though.
Crabtree's bludgeon: “no set of mutually inconsistent observations can exist for which some human intellect cannot conceive a coherent explanation, however complicated”
GNU Terry Pratchett

FancyHat
Posts: 341
Joined: Thu Oct 25, 2012 7:28 pm UTC

Re: 1492: "Dress color"

Postby FancyHat » Sat Feb 28, 2015 11:24 pm UTC

Angua wrote:That's a cool gif. I can't see the start of it as being other than white and gold though.

Can you see the end of it as being white and gold?

It looks like it's kind of the opposite of the dress photo, in a way. Whereas the dress photo has the dress overexposed and often perceived to be as if it's a white and gold dress in relative shadow instead, the end of Klear's excellent GIF seems to be a white and gold thing that really is in relative shadow, but which can be perceived as blue and black instead.
I am male, I am 'him'.

User avatar
Klear
Posts: 1965
Joined: Sun Jun 13, 2010 8:43 am UTC
Location: Prague

Re: 1492: "Dress color"

Postby Klear » Sat Feb 28, 2015 11:36 pm UTC

When it starts looping, my brain already considers the black/blue option as true and compensates from the first frame. I did see gold/white for a while reviewing the animation now that some time has passed.

I should make one that goes backwards...

User avatar
jovialbard
Posts: 613
Joined: Thu Apr 18, 2013 11:08 pm UTC

Re: 1492: "Dress color"

Postby jovialbard » Sun Mar 01, 2015 12:22 am UTC

The garmet in that gif really is blue/black right? If that's the case, then that doesn't help me understand anything. I already understand how a blue/black garmet can look white/gold when exposed to harsh yellow light, but the infamous dress doesn't have nearly the same rgb values as the first frame of that gif, not even close. So while the first frame of the gif looks white and gold, that's not really a result of perceptual correction is it? It really is pretty damn close to white and gold, unlike the dress, which is a marginally light blue and a very dark greyish brown. So I still don't really see how people could see that marginally pale blue and very dark greyish brown as white and gold... I think it comes down be being able to perceive the black as anything but black subjected to harsh yellow light. I feel like perceiving the black snaps the blue into "focus" and the secret to seeing white/gold must be seeing the gold first... but I just can't seem to see it.

If the garment in the gif is actually white and gold, then I'm utterly baffled as to how it looks blue and black at the end, given that there is no obvious light source other than the yellowish one that clearly still appears to be in front of the garment.
I wrote a book exploring the nature of identity and loss of self! And another book about babies!
Also, vlog and short-story blog

User avatar
Klear
Posts: 1965
Joined: Sun Jun 13, 2010 8:43 am UTC
Location: Prague

Re: 1492: "Dress color"

Postby Klear » Sun Mar 01, 2015 12:45 am UTC

jovialbard wrote:The garmet in that gif really is blue/black right? If that's the case, then that doesn't help me understand anything. I already understand how a blue/black garmet can look white/gold when exposed to harsh yellow light, but the infamous dress doesn't have nearly the same rgb values as the first frame of that gif, not even close. So while the first frame of the gif looks white and gold, that's not really a result of perceptual correction is it? It really is pretty damn close to white and gold, unlike the dress, which is a marginally light blue and a very dark greyish brown. So I still don't really see how people could see that marginally pale blue and very dark greyish brown as white and gold... I think it comes down be being able to perceive the black as anything but black subjected to harsh yellow light. I feel like perceiving the black snaps the blue into "focus" and the secret to seeing white/gold must be seeing the gold first... but I just can't seem to see it.

If the garment in the gif is actually white and gold, then I'm utterly baffled as to how it looks blue and black at the end, given that there is no obvious light source other than the yellowish one that clearly still appears to be in front of the garment.


It's because everybody has these "margins" set up differently.

I had trouble picturing how they saw the dress in white and gold. There have been some colour-shifted edits of the original pic, but none of them looked very good to me. After seeing this I can finally understand what the gold/white people actually see.

fluffysheap
Posts: 45
Joined: Sat Sep 28, 2013 7:53 am UTC

Re: 1492: "Dress color"

Postby fluffysheap » Sun Mar 01, 2015 1:33 am UTC

The GIF is interesting, but the white balance is so far off it's hard to tell much from it. Look at the walls or the light coming in from the window, later in the sequence. Those are probably white or nearly white, but they look very yellow. As previously discussed, when the white balance is too yellow, white things look blue. Even the early part of the GIF, when the cloth is up close and illuminated, it still has a distinct blue tint, albeit pale. However, unlike the original demon dress, there is no other color reference in the frame from which to determine ambient lighting/white balance information.

Edit: It might be that the camera in question is adjusting itself on-the-fly while the dress is in motion, in which case the GIF really proves nothing at all except that different camera settings look different, which everybody already knows.

User avatar
gmalivuk
GNU Terry Pratchett
Posts: 26822
Joined: Wed Feb 28, 2007 6:02 pm UTC
Location: Here and There
Contact:

Re: 1492: "Dress color"

Postby gmalivuk » Sun Mar 01, 2015 2:14 am UTC

fluffysheap wrote:Edit: It might be that the camera in question is adjusting itself on-the-fly while the dress is in motion, in which case the GIF really proves nothing at all except that different camera settings look different, which everybody already knows.
It's doing more than that, though. It's proving in particular that "naturally" adjusting cameras can make a black and blue garment look that entire range of shades, including the first frame that could easily be mistaken for a blue-white and gold-ish combination.
Unless stated otherwise, I do not care whether a statement, by itself, constitutes a persuasive political argument. I care whether it's true.
---
If this post has math that doesn't work for you, use TeX the World for Firefox or Chrome

(he/him/his)

B-Fisch
Posts: 18
Joined: Tue Jan 25, 2011 5:05 am UTC
Location: Lafayette, LA

Re: 1492: "Dress color"

Postby B-Fisch » Sun Mar 01, 2015 10:27 am UTC

I'm surprised no one's pointed out that this comic, about a blue (and black) object is number 1492.

In Fourteen Hundred Ninety-two,
Columbus Randall said the dress was blue.

sekito34
Posts: 2
Joined: Fri Feb 27, 2015 3:26 pm UTC

Re: 1492: "Dress color"

Postby sekito34 » Sun Mar 01, 2015 12:19 pm UTC

I have to say it surprised me how many people are asking the wrong questions.
Sure color constancy and different lighting can make colors look different, but that is not the point;
the point is, discounting monitor differences and lighting conditions, why do some people see blue/black while some see white/gold.

After much research into color perception, I am going to propose it has nothing to do with color constancy but everything to do with background variance.
Simple put, when the background has great variance in color (such as containing many different colors), perception of color at the foreground becomes less salient(that is, less accurate).
In the dress photo, the background is mostly white and yellow, therefore the variance is low, but the exact variance perceived depends from person to person.
For those who are more sensitive to light, the variance will be larger than those who are less sensitive to light (for example, people who sit it front of a monitor all day or people whose eyes are tired)
Those whose perceived great variance will be less able to detect the faint blue color of the dress, and will assume it as white. At this point, color constancy applies, making brownish stripes look gold.
For those who perceived less variance, they will be likely to detect the faint blue color, and color constancy causes the brownish stripes seem black.

Note: this also explains why tilting certain LED displays, which decrease brightness thus background variance, will make the dress look more blue/black

User avatar
gmalivuk
GNU Terry Pratchett
Posts: 26822
Joined: Wed Feb 28, 2007 6:02 pm UTC
Location: Here and There
Contact:

Re: 1492: "Dress color"

Postby gmalivuk » Sun Mar 01, 2015 2:33 pm UTC

Tilting monitors doesn't just change brightness, though. It also changes other things about colors and can greatly enhance or reduce contrast.
Unless stated otherwise, I do not care whether a statement, by itself, constitutes a persuasive political argument. I care whether it's true.
---
If this post has math that doesn't work for you, use TeX the World for Firefox or Chrome

(he/him/his)

jjcote
Posts: 32
Joined: Sat Apr 16, 2011 2:16 am UTC

Re: 1492: "Dress color"

Postby jjcote » Sun Mar 01, 2015 3:42 pm UTC

I'd like to point out that this dress would look great in white/gold, and I totally hope they start selling it that way. And how cool would it be for identical twins to show up at a party together wearing the two color combinations?

FancyHat
Posts: 341
Joined: Thu Oct 25, 2012 7:28 pm UTC

Re: 1492: "Dress color"

Postby FancyHat » Sun Mar 01, 2015 4:54 pm UTC

This USA Today article includes a useful image to illustrate what seems to be the two main ways people's brains are interpreting or misinterpreting the image. The image is credited to Claire Hummel, and seems to do a better job than the XKCD image.
I am male, I am 'him'.

User avatar
Quercus
Posts: 1810
Joined: Thu Sep 19, 2013 12:22 pm UTC
Location: London, UK
Contact:

Re: 1492: "Dress color"

Postby Quercus » Sun Mar 01, 2015 5:35 pm UTC

FancyHat wrote:This USA Today article includes a useful image to illustrate what seems to be the two main ways people's brains are interpreting or misinterpreting the image. The image is credited to Claire Hummel, and seems to do a better job than the XKCD image.


Okay, reading that article lead me to try to imagine the dress under a yellow light, and it started to shift for me - I originally saw gold/white, and now was starting to see black/blue (admittedly never a deep blue).

The strange thing was that the shift made me quite uncomfortable - it felt a bit like wearing glasses that are too strong, so much so that I had to close my eyes and shake my head to dispel the feeling. Strange.

LockeZ
Posts: 51
Joined: Mon Jan 19, 2009 8:30 am UTC

Re: 1492: "Dress color"

Postby LockeZ » Sun Mar 01, 2015 6:04 pm UTC

Look at the cow print to the left of the dress. Now look at the dress. Now look at the cow print. Now back at the dress.

Sadly, the dress isn't the cow print. But if it stopped using lady-scented laundry detergent and switched to Old SpiceTM, it could smell like the cow print.

Look at the washed-out background. Now look back. Where are you? You're on a boat. With the dress your dress could smell like. What are you wearing? I have it. It's that outfit you love, that I just made you slip right out of. Look again. The outfit is now diamonds.

Anything is possible when your dress smells like a cow print. I'm on a horse.

User avatar
Flumble
Yes Man
Posts: 2263
Joined: Sun Aug 05, 2012 9:35 pm UTC

Re: 1492: "Dress color"

Postby Flumble » Sun Mar 01, 2015 7:19 pm UTC

LockeZ wrote:Look at the cow print to the left of the dress. Now look at the dress. Now look at the cow print. Now back at the dress.

Sadly, the dress isn't the cow print. But if it stopped using lady-scented laundry detergent and switched to Old SpiceTM, it could smell like the cow print.

Look at the washed-out background. Now look back. Where are you? You're on a boat. With the dress your dress could smell like. What are you wearing? I have it. It's that outfit you love, that I just made you slip right out of. Look again. The outfit is now diamonds.

Anything is possible when your dress smells like a cow print. I'm on a horse.

It's too bad there's no such things as quantum internet, otherwise I'd give you an internet.

User avatar
pinch
Posts: 7
Joined: Sun Mar 01, 2015 7:04 pm UTC

Re: 1492: "Dress color"

Postby pinch » Sun Mar 01, 2015 8:06 pm UTC

I agree that there's a perceptual factor, but there's a simpler explanation.

For whatever reason, the image posted has an extreme gamma boost, which also raised the black level. The dress itself is not overexposed, only the background.
I suspect that the camera was on autoexposure and took a shot that made the dress look way too dark. It appears to have been modified to correct for that by boosting the gamma by about 3.0, an extreme amount. Simple overexposure does not account for the error.

The excess gamma boost can be reversed. With no white balance correction, here's the result of a 3.0 gamma reversal:

(Rats! No permissions to link an image.) Here's a shortcut: goo.gl/zY11YG

A similar effect to reversing the gamma happens when viewing a mobile device in bright daylight, or if the black level and contast are too harsh on an LCD. Even a properly adjusted display can have a different gamma. Apple displays historically have a gamma of 2.2 compared to generic PC displays. I'd be curious to see a scatterplot of initial color choice vs effective display gamma on first viewing.

For those who prefer an argument from authority, I hold a 1989 patent for an early digital camera that did subpixel sampling as just seen at 2x in the latest Olympus DSLR - but at 8x, multiexposure HDR fusion before it was called that, and it gets cited as early prior art for onboard jpeg. So, trust me on this. ;)

Please be gentle, though - it's my first xkcd forum post.

User avatar
ucim
Posts: 6888
Joined: Fri Sep 28, 2012 3:23 pm UTC
Location: The One True Thread

Re: 1492: "Dress color"

Postby ucim » Sun Mar 01, 2015 10:17 pm UTC

pinch wrote:For whatever reason, the image posted has an extreme gamma boost, which also raised the black level. The dress itself is not overexposed, only the background.
I suspect that the camera was on autoexposure and took a shot that made the dress look way too dark. It appears to have been modified to correct for that by boosting the gamma by about 3.0, an extreme amount. Simple overexposure does not account for the error.

pinch's image:
Spoiler:
Image
Brilliant! I took the original into gimp and did the same thing, with the same result.

You have saved the internet!

pinch wrote:For those who prefer an argument from authority...
Cool on the patent. But I have it on good authority that we don't work here based on authority. :)

Jose
(edit to put in pinch's image alluded to above)
Last edited by ucim on Sun Mar 01, 2015 11:03 pm UTC, edited 3 times in total.
Order of the Sillies, Honoris Causam - bestowed by charlie_grumbles on NP 859 * OTTscar winner: Wordsmith - bestowed by yappobiscuts and the OTT on NP 1832 * Ecclesiastical Calendar of the Order of the Holy Contradiction * Heartfelt thanks from addams and from me - you really made a difference.

FancyHat
Posts: 341
Joined: Thu Oct 25, 2012 7:28 pm UTC

Re: 1492: "Dress color"

Postby FancyHat » Sun Mar 01, 2015 10:46 pm UTC

pinch wrote:For whatever reason, the image posted has an extreme gamma boost, which also raised the black level. The dress itself is not overexposed, only the background.

Are you sure the dress isn't overexposed and washed-out as a result, and that the dress itself isn't actually a lot darker than it appears in the photo?

Keyman's comment earlier in this thread included a different photo of exactly the same dress, under different lighting conditions, and it looks like a different dress.

The website selling those dresses shows the dress, and the famous photo does look overexposed and washed-out in comparison.
I am male, I am 'him'.

User avatar
BlitzGirl
Posts: 9118
Joined: Mon Sep 20, 2010 11:48 am UTC
Location: Out of the basement for Yip 6! Schizoblitz: 115/2672 NP
Contact:

Re: 1492: "Dress color"

Postby BlitzGirl » Sun Mar 01, 2015 10:55 pm UTC

FancyHat wrote:This USA Today article includes a useful image to illustrate what seems to be the two main ways people's brains are interpreting or misinterpreting the image. The image is credited to Claire Hummel, and seems to do a better job than the XKCD image.

To have an in-thread view:

Image

I agree it seems to be more effective than the xkcd version.
Spoiler:
Image
Knight Temporal of the One True Comic
BlitzGirl the Pink, Mopey Molpy Mome
Spoiler:
Image
Image
Image<Profile
~.Image~.FAQ->Image

User avatar
Quercus
Posts: 1810
Joined: Thu Sep 19, 2013 12:22 pm UTC
Location: London, UK
Contact:

Re: 1492: "Dress color"

Postby Quercus » Sun Mar 01, 2015 11:25 pm UTC

FancyHat wrote:
pinch wrote:For whatever reason, the image posted has an extreme gamma boost, which also raised the black level. The dress itself is not overexposed, only the background.

Are you sure the dress isn't overexposed and washed-out as a result, and that the dress itself isn't actually a lot darker than it appears in the photo?

Keyman's comment earlier in this thread included a different photo of exactly the same dress, under different lighting conditions, and it looks like a different dress.

The website selling those dresses shows the dress, and the famous photo does look overexposed and washed-out in comparison.


That appears to be the effect of the gamma boost. This does look a bit like overexposure, but isn't. It is, speaking vaguely and non-technically, "underexposed" and then "overcorrected". Check out pinch's image (ucim has put it in his post) - with the gamma taken back down to a reasonable value it looks exactly like the dress in the two links you gave.

What I'm wondering now is how the entire internet, including major newspapers and magazines citing "experts" failed to mention the gamma issue.

Incidentally, for me on a monitor calibrated (with a colorimeter) to 5800K and a display gamma of 2.2, the perceptual switch between blue/black and white/gold happened at around 70% of the gamma of the original photo.

User avatar
addams
Posts: 10331
Joined: Sun Sep 12, 2010 4:44 am UTC
Location: Oregon Coast: 97444

Re: 1492: "Dress color"

Postby addams » Sun Mar 01, 2015 11:59 pm UTC

This is the Internet.
Is this Week, Color Blind Awareness Week?

It could be.
http://www.colourblindawareness.org/col ... rience-it/

I know there is Wild Variation in Color Experience.
I've met people that shared my color experience.

How did we Know we were seeing The Same Color?
Maybe, we were Not seeing the same Color.

We were Expressing Similar or The Same Reaction to Color.
That was a Fun Day. We were Running from Color to Color.

Normal people were busy being Normal.
We were not bothering them; Much.
Life is, just, an exchange of electrons; It is up to us to give it meaning.

We are all in The Gutter.
Some of us see The Gutter.
Some of us see The Stars.
by mr. Oscar Wilde.

Those that want to Know; Know.
Those that do not Know; Don't tell them.
They do terrible things to people that Tell Them.

FancyHat
Posts: 341
Joined: Thu Oct 25, 2012 7:28 pm UTC

Re: 1492: "Dress color"

Postby FancyHat » Mon Mar 02, 2015 12:33 am UTC

Quercus wrote:That appears to be the effect of the gamma boost. This does look a bit like overexposure, but isn't. It is, speaking vaguely and non-technically, "underexposed" and then "overcorrected". Check out pinch's image (ucim has put it in his post) - with the gamma taken back down to a reasonable value it looks exactly like the dress in the two links you gave.

Well, I don't think that really works. If it really was underexposed originally, wouldn't it be darker than in the other photos?

Certainly with the gamma changed, it's got that 'underexposed' kind of look to it, but I don't think that shows it was originally underexposed and then overcorrected. Overexposing a relatively dark object against a much lighter background won't necessarily make the object have that 'overexposed' look, since that part of the image can still be some way from saturating the sensor. This could be the case if the camera is incorrectly compensating as if it's a scene of a poorly lit foreground object against a brightly lit background, when actually the foreground object really is relatively dark and is just as well illuminated as the background.

What I'm wondering now is how the entire internet, including major newspapers and magazines citing "experts" failed to mention the gamma issue.

I think it's enough to say that it's not a very good photo and that it was taken with a camera phone. The thing people have been interested in is why some people perceived it as blue and black, while others perceived it as white and gold, when looking at the exact same image on the exact same screen at the exact same time. And, in addition to that, some people perceived it one way and then the other, without anything actually changing to explain that. That's what people have been interested in, rather than how such a lousy photo came to be in the first place.
I am male, I am 'him'.

User avatar
pinch
Posts: 7
Joined: Sun Mar 01, 2015 7:04 pm UTC

Re: 1492: "Dress color"

Postby pinch » Mon Mar 02, 2015 5:32 am UTC

Well, I don't think that really works. If it really was underexposed originally, wouldn't it be darker than in the other photos?

Yes, that's what it would look like in the original. The original probably was shot on a cell phone. It would be higher resolution and have exif data, if it ever shows up.

Certainly with the gamma changed, it's got that 'underexposed' kind of look to it, but I don't think that shows it was originally underexposed and then overcorrected. Overexposing a relatively dark object against a much lighter background won't necessarily make the object have that 'overexposed' look, since that part of the image can still be some way from saturating the sensor. This could be the case if the camera is incorrectly compensating as if it's a scene of a poorly lit foreground object against a brightly lit background, when actually the foreground object really is relatively dark and is just as well illuminated as the background.


But, cameras don't do that. They tend toward exposing for the average of the scene, which would underexpose the dress. It has been manipulated.

I think it's enough to say that it's not a very good photo and that it was taken with a camera phone. The thing people have been interested in is why some people perceived it as blue and black, while others perceived it as white and gold, when looking at the exact same image on the exact same screen at the exact same time. And, in addition to that, some people perceived it one way and then the other, without anything actually changing to explain that. That's what people have been interested in, rather than how such a lousy photo came to be in the first place.


If people looking at the exact same screen at the same time are seeing different colors, then it's likely a result of perceptual bias. When I first saw the image on a high gamma monitor, it certainly looked like a washed out white and gold. Discovering that it was in fact a blue dress shifted my perception slightly. Now that I've seen the corrected image, I can't see the "original" any other way.

If the gamma corrected image looks too dark to you, then consider that you may be viewing it on a low gamma display or bright environment. Try correcting with gamma around 2, and you will see a milder but accurate rendition: The dress is in a shop that's probably lit by CFL lamps, but the overhead light is incandescent (black becoming gold when boosted) and the background out the window is daylight.

The variations among displays are huge. Ask any expert in color printing, whose job it is to correct for what someone saw on their uncalibrated monitor.

The human visual system is amazing, but it can be hoodwinked.

User avatar
nayhem
Posts: 7
Joined: Mon May 18, 2009 6:49 pm UTC

Re: 1492: "Dress color"

Postby nayhem » Mon Mar 02, 2015 9:15 am UTC

pinch wrote:But, cameras don't do that. They tend toward exposing for the average of the scene, which would underexpose the dress. It has been manipulated.

Cameras (especially in smartphones) more often use center-weighted metering. Exposure was adjusted for the darker dress in the center, causing the lighting along the right side to blow out. The auto white balance would also likely have done a better job on a light object than a dark one.

It is much more likely that the user didn't know how or when to adjust exposure compensation (I didn't know my own phone had it until a few weeks ago), didn't notice that the camera adjusted the way it did, and didn't feel compelled to "correct" it.

User avatar
Quercus
Posts: 1810
Joined: Thu Sep 19, 2013 12:22 pm UTC
Location: London, UK
Contact:

Re: 1492: "Dress color"

Postby Quercus » Mon Mar 02, 2015 10:45 am UTC

nayhem wrote:Cameras (especially in smartphones) more often use center-weighted metering.


Hmm, I did not know that - no wonder I don't like the images my smartphone produces - I don't get along with centre weighted metering (probably because I often position my subjects fairly off-centre). I've just found out that my phone offers multi metering too (which is usually biased towards the autofocus point, rather than to the centre), and I've switched my phone to that now.

FancyHat
Posts: 341
Joined: Thu Oct 25, 2012 7:28 pm UTC

Re: 1492: "Dress color"

Postby FancyHat » Mon Mar 02, 2015 3:37 pm UTC

pinch wrote:But, cameras don't do that. They tend toward exposing for the average of the scene, which would underexpose the dress. It has been manipulated.

Really? Are you sure about that?

There's a major reason why many cameras, including cameras in phones, would be designed to try to adjust for centre-field foreground objects by default: most people don't know much about photography, and just point their cameras straight at whatever they're trying to photograph, without knowing they need to do anything about significant differences in foreground and background lighting, but they still want their photos to come out okay. So, cameras that, by default, try to adjust, focus, expose and compensate for centre-field foreground objects, regardless of what's going on in the background round the edges, will tend to satisfy a lot of customers.

In this case, we've got a photo that, it seems, the majority of people initially perceive to be a white and gold dress in relative shadow, illuminated by relatively bluish light, compared to the much more brightly, more yellowy-lit and obviously overexposed background. If the camera treated the dress as poorly-lit when it was brightly-lit - which it could easily do since the dominant foreground object actually is quite a lot darker than most of the background - and if the camera white-balanced for relatively blue illuminating light - which it could easily do since the dominant foreground object actually is very blue, unlike most of the background - then you'd get the dress overexposed and wrongly white-balanced in such a way that it looks lighter and less blue than it actually is, but without necessarily looking to everyone like it was overexposed or wrongly white-balanced.

Your hypothesis that someone did a bad job of trying to lighten the original photo, by changing the gamma, to produce the photo that the world has now seen, doesn't really explain anything that isn't already explained without anyone having to change the gamma. I don't think it's a terribly likely hypothesis, and I don't think it works very well.

If the gamma corrected image looks too dark to you, then consider that you may be viewing it on a low gamma display or bright environment.

When I said, "it's got that 'underexposed' kind of look to it," I meant I could see why you could conclude the original image was underexposed. It had that murky, details-getting-lost-in-murk kind of look to it.

As for my display's gamma, if it was too low, then I'd either routinely find images appearing "too dark", or I'd be used to it, and wouldn't find your "gamma corrected" image "too dark". Either way, I wouldn't find your "gamma corrected" image remarkably different to most other images I see on my display.

Try correcting with gamma around 2, and you will see a milder but accurate rendition: The dress is in a shop that's probably lit by CFL lamps, but the overhead light is incandescent (black becoming gold when boosted) and the background out the window is daylight.

Where did you learn about it being in a shop? I'm really not sure the scene is as you describe it.

The human visual system is amazing, but it can be hoodwinked.

Yes, I think there's a bit of irony there.

In your earlier post, you mentioned your patent, as if that was evidence of expertise or authority. Unfortunately, there are plenty who aren't nearly the experts they think they are who cite their patents as if those patents mean a lot more than they actually do. There have been plenty of cranks and crackpots who have expected others to bow to their supposed superiority on the basis of such patents. I take such 'evidence' of 'expertise' as a little warning flag instead.

So, oh expert, please tell me what's wrong with the following:-

"To correctly reduce the resolution of an ordinary, greyscale image, it's necessary to slightly blur the image first, so as to avoid visible artefacts of resolution reduction. Otherwise, when blocks of pixels are averaged together into single, lower-resolution pixels, artefacts will often be produced, such as sharp edges that are close to vertical and horizontal getting a sort of stair-stepped effect. Blurring the image first with a simple filter, so as to remove details at scales similar to or smaller than the new, lower-resolution pixels, will avoid such artefacts."

Please also tell me a correct method for reducing the resolution of such images, and please explain why it's correct.
I am male, I am 'him'.

User avatar
gmalivuk
GNU Terry Pratchett
Posts: 26822
Joined: Wed Feb 28, 2007 6:02 pm UTC
Location: Here and There
Contact:

Re: 1492: "Dress color"

Postby gmalivuk » Mon Mar 02, 2015 4:05 pm UTC

Who knows if it's in a shop, but it's clearly in a room with other clothes, so shop isn't unreasonable. (Knowing more makes it less likely, though: the dress is available for purchase online and doesn't come with the jacket shown, which suggests it may be in someone's house instead.)

My phone camera definitely doesn't adjust exposure for the center but for the overall average, regardless of what your marketing thought experiment would tend to suggest.
Unless stated otherwise, I do not care whether a statement, by itself, constitutes a persuasive political argument. I care whether it's true.
---
If this post has math that doesn't work for you, use TeX the World for Firefox or Chrome

(he/him/his)

User avatar
Quercus
Posts: 1810
Joined: Thu Sep 19, 2013 12:22 pm UTC
Location: London, UK
Contact:

Re: 1492: "Dress color"

Postby Quercus » Mon Mar 02, 2015 4:19 pm UTC

Oh, I misread - I thought that pinch was saying that the camera automatically made that gamma shift as some sort of automatic adjustment that went horribly wrong in this edge case.

That I can believe, but I think it unlikely that someone manipulated this image by hand, simply because that sort of manipulation doesn't fit with the provenance of the photo (who would take a quick snap of a dress on a phone, open it up on the computer and make a "correction" that makes the image look substantially worse). I suppose we must consider deliberate engineering of an "ambiguous photo" in a (successful) effort to make it go viral, but that seems too much like a conspiracy theory to me.

gmalivuk wrote:My phone camera definitely doesn't adjust exposure for the center but for the overall average, regardless of what your marketing thought experiment would tend to suggest.

My phone camera appears to expose for any faces it detects in the frame, and if it can't find any faces default to exposing for the center. You can manually set it to face, center, multi or spot metering though. I'm actually quite impressed by the level of manual control in my phone camera - I do wish it had manual focus though.

dawolf
Posts: 164
Joined: Thu Jan 07, 2010 9:21 pm UTC

Re: 1492: "Dress color"

Postby dawolf » Mon Mar 02, 2015 4:40 pm UTC

OK, I get that people who think the photo has a yellow tint adjust, and see black/blue: and that those who see a blue tint adjust, and see yellow/gold.

What I don't understand is WHY people automatically adjust and assume a colour change. I see it as blue/gold, absent any definitive information that tells my brain to adjust.

I am also very confused why 90% of people pick white or black as one of their choices: the photo doesn't show either (as can be confirmed by rgb values). So why do so many people see a colour which really isn't there?

dawolf
Posts: 164
Joined: Thu Jan 07, 2010 9:21 pm UTC

Re: 1492: "Dress color"

Postby dawolf » Mon Mar 02, 2015 4:48 pm UTC

Soft Hyphen wrote:My immediate reaction on seeing the picture was that the dress is light blue and dark bronze. After some time, I now see it as being a deeper, darker blue and bronzish-tinted black (yes, that is a thing to me).

I cannot, for the life of me, not even for a single instant, see any white in that dress. There is not white, there is not underexposed white, there is not light grey or even dark grey. There is only blue. You can argue about what precise shade of blue, but it is very solidly blue, always has been, and always will be. And it does kinda freak me out when people say they see white and I can't help but think "What the heck's wrong with your eyes?!".....................


Well said, this is basically my confusion as well.

FancyHat
Posts: 341
Joined: Thu Oct 25, 2012 7:28 pm UTC

Re: 1492: "Dress color"

Postby FancyHat » Mon Mar 02, 2015 5:52 pm UTC

I WAS WRONG!

gmalivuk wrote:My phone camera definitely doesn't adjust exposure for the center but for the overall average, regardless of what your marketing thought experiment would tend to suggest.

Quickly looking at the histograms in GIMP, the histograms look consistent with very simple, overall averaging, both for pixel "values" and colour channels. The mean "value" is 153.8, while red is 137.2, green is 135.5 and blue is 138.2, which are very similar. The histograms themselves also have that 'adjusted for overall average' look. So yeah, I think simple, overall averaging by the camera for both exposure and white-balance is likely to be enough to explain the photo itself. So I think my "marketing" hypothesis was probably wrong after all.

Also, using GIMP to automatically equalize the photo of the dress being worn at the wedding produces a similar shade of blue for much of the dress itself, but the image quality is far worse than in the famous photo. It ends up with a kind of lumpy graininess (I'm not sure how else to describe it). Changing the gamma doesn't work very well, either. Which makes me wonder: if pinch was right, how come the overcorrected image lacks the kind of exaggerated, lumpy graininess it ought to have as a result? I've manually lightened underexposed images of fairly dark objects, and it seems quite normal to get lumpy graininess as a result. The famous photo simply doesn't have that.

Edited to add the following:-

So, I think I need to return to something pinch wrote:-

pinch wrote:But, cameras don't do that. They tend toward exposing for the average of the scene, which would underexpose the dress. It has been manipulated.

Even averaging for the entire scene, the dress still dominates, filling most of the picture. And being relatively dark, being royal blue and black, exposing for the average of the entire scene can still overexpose the dress, rather than underexpose it. With this scene, you just wouldn't get an underexposed dress as a result of exposing for the average of the whole scene.
I am male, I am 'him'.

User avatar
gmalivuk
GNU Terry Pratchett
Posts: 26822
Joined: Wed Feb 28, 2007 6:02 pm UTC
Location: Here and There
Contact:

Re: 1492: "Dress color"

Postby gmalivuk » Mon Mar 02, 2015 7:08 pm UTC

dawolf wrote:OK, I get that people who think the photo has a yellow tint adjust, and see black/blue: and that those who see a blue tint adjust, and see yellow/gold.

What I don't understand is WHY people automatically adjust and assume a colour change. I see it as blue/gold, absent any definitive information that tells my brain to adjust.

I am also very confused why 90% of people pick white or black as one of their choices: the photo doesn't show either (as can be confirmed by rgb values). So why do so many people see a colour which really isn't there?
Because we are very used to seeing photographs taken in light that affects colors. Some lights make things look yellow in a photo, other lights make things look blue. No fabric reflects or absorbs 100% of visible light, and so even things that we call white or black really contain some color. Changing the exposure or gamma or white balance can therefore make that color stand out more.

Edit:
FancyHat wrote:Also, using GIMP to automatically equalize the photo of the dress being worn at the wedding produces a similar shade of blue for much of the dress itself, but the image quality is far worse than in the famous photo. It ends up with a kind of lumpy graininess (I'm not sure how else to describe it). Changing the gamma doesn't work very well, either. Which makes me wonder: if pinch was right, how come the overcorrected image lacks the kind of exaggerated, lumpy graininess it ought to have as a result? I've manually lightened underexposed images of fairly dark objects, and it seems quite normal to get lumpy graininess as a result. The famous photo simply doesn't have that.
1) The photo at the wedding looks to be fairly grainy to start with, if we're talking about the same photo, so it makes sense that certain manipulations would exaggerate that.
2) If the brightening was done in the camera before the image was compressed as a jpeg, it also makes sense that the result wouldn't be as ugly as when you manually do something similar to the jpeg image.

pinch wrote:But, cameras don't do that. They tend toward exposing for the average of the scene, which would underexpose the dress. It has been manipulated.

Even averaging for the entire scene, the dress still dominates, filling most of the picture. And being relatively dark, being royal blue and black, exposing for the average of the entire scene can still overexpose the dress, rather than underexpose it. With this scene, you just wouldn't get an underexposed dress as a result of exposing for the average of the whole scene.
If the bright parts are bright enough, though, then even though it takes up 75% of this* image, couldn't the dress still end up underexposed depending on the algorithm used to adjust exposure? If it's a bit dumb, it would notice that there's still too much light coming in unless it underexposes the dress. And then someone afterwards might have brightened it back up before putting it out in the world. (The lack of any significant number of pixels below level 30 or so, plus the pile of them bunched up above 240, would seem to suggest some kind of inexpert brightening after the picture was taken.)

* Also, the image could be cropped quite a bit.
Unless stated otherwise, I do not care whether a statement, by itself, constitutes a persuasive political argument. I care whether it's true.
---
If this post has math that doesn't work for you, use TeX the World for Firefox or Chrome

(he/him/his)

User avatar
pinch
Posts: 7
Joined: Sun Mar 01, 2015 7:04 pm UTC

Re: 1492: "Dress color"

Postby pinch » Tue Mar 03, 2015 12:12 am UTC

"Oh Mr Expert"?
It's wonderful to see healthy skepticism and curiosity running rampant, especially over image processing and color perception - topics that never, ever make it into general conversation. Some people do defer all thinking to experts, like doctors. I don't recommend it. Though I specifically prefaced it as a specious argument from authority, it's a good way of letting you know that I've thought about these technical issues for decades, not just hours. It wasn't an offer of free advice, but here's a hint: The best algorithms for image reduction will most accurately reproduce the image when upscaled - plus a bit of twerking for compressing extra edge information into the reduced space.

"Why no blotchy noise?"
Giving a gamma boost can make huge jumps in the first few numbers if the image contains near-black pixels. (Jpeg images contain a gamma compression, but it's shifted to linear at the foot to avoid those artifacts). This particular image didn't have black pixels before the gamma boost, so the black areas of the dress are at nearly half brightness.
This might be caused by lens flare from a greasy thumbprint on the lens.

"Auto-Overcompensation?"
The animated gif of autoexposure when zooming on a dark dress is a nice demonstration. A camera would not typically take a scene that has very bright highlights and seriously overexpose it. If it was cropped from an even larger area with brighter pixels surrounding the dress, the effect on original exposure would have been to darken it further. That may be why it was boosted. I can't imagine any scenario, aside from spot metering on the black portion of the dress, where a camera would overcompensate to boost the exposure like that. The wider area probably did have daylight, which caused a slight color balance shift toward yelllow. A very similar result to the gamma correction can be accomplished by reducing the green and red channels while boosting the blue. A combo of gamma reversal, white balance, and black level correction would probably give the best reconstruction of the original scene given the image that we've got.

"Why do people see colors that aren't there?"
I am always amazed by the ability of our visual system to accomodate to different colored lighting environments. Even with a mix of daylight, fluorescent, and candlelight, we can usually sort it out in real life and accurately intuit what the color is. Even color blind people do that, maybe better than folks with all three color receptors. We take cues from our experience, and it changes our perception. My first view of the dress image was the cropped snippet from xkcd. (Yeah, yeah.) With no supporting context, it doesn't look blue at all. Googling for "dress color" was easy, and I was still swayed by my first impression. Knowing the truth wasn't enough to shift it blue. Once I saw it correctly reconstructed, I can no longer see it as before. Others will have a completely different experience. so their perception will be different.

That's my guess. I'm reluctant to even vouch that it's an educated one.

BTW, my favorite tool for grokking the color space of an image is the 3D Color Inspector plugin in Fiji. (Like Gnu, Fiji Is Not Imagej: an open source java medical image analysis tool from the NIH.

Edit - gotta love that xkcd translates t-w-e-a-k to twerk ;)

User avatar
pinch
Posts: 7
Joined: Sun Mar 01, 2015 7:04 pm UTC

Re: 1492: "Dress color"

Postby pinch » Tue Mar 03, 2015 8:11 am UTC

Here's one more bit of circumstantial evidence for the image being manipulated. Check out the selfies on the home page of woman who started the fracas:

http://swiked.tumblr.com/post/112073818575/
Image

Note that her own photos suffer the same washed-out gamma boost effect that was applied to the dress image. She probably adjusts the images on her own monitor so they look best, but her screen is misadjusted too dark.

User avatar
Quercus
Posts: 1810
Joined: Thu Sep 19, 2013 12:22 pm UTC
Location: London, UK
Contact:

Re: 1492: "Dress color"

Postby Quercus » Tue Mar 03, 2015 10:35 am UTC

Good sleuthing!

Also, seeing the image in the context of those other washed out photos it appears to me mid-blue/black (my mind corrects it) - but the image in isolation still appears pale-blue/gold.


Return to “Individual XKCD Comic Threads”

Who is online

Users browsing this forum: No registered users and 98 guests