LWN: Comments on "SCALE 8x: Color management for everyone"

Terminology

Ross — 2010-03-14T21:38:25+00:00

It's great to see color reproduction get some attention and to have LWN coverage in particular.

However, I think the article sometimes uses term gamut to refer to the entire set of characteristics of a device colorspace, but it should only refer to the range of colors which can be displayed. To be specific, the meaning of the numeric coordinates of colorspaces can be different even when they have the same gamut, and because of things like the gamma response curve the coordinates don't necessarily even have a linear relationship.

Also, both the gamma and the amount of quantization of the colorspace coordinates are affected by the video card and drivers as much as by the display device. For example 15 or 16 bit depth modes will result in only 5 or 6 bits of accuracy per channel no matter how nice the display. Hopefully these limitations are taken into account in the profiles somehow.

SCALE 8x: Color management for everyone

psyquark — 2010-03-08T03:50:44+00:00

Almost.

All color triplets have a color space, either explicit or implicit. On a stand alone uncalibrated machine the implicit colorspace is the colorspace of the display. The problems start when sharing files with other computers, each of which has its own implicit colorspaces. Known colorspace to known colorspace mapping is doable but unknown colorspace to any colorspace is not feasible. To fix that the sRGB colorspace was created. It was defined to be a decent approximation of the standard displays of the time. The web standard became "If you don't know the space and care about color, treat it as sRGB."

The naive solution is to have the client handle all color management and output pixel values in the display's colorspace. This can be useful in cases when a specific or controlled rendering intent is needed to accommodate special needs. The problem is that client side management does not handle multiple outputs well. In fact, it is not possible for a client to handle simultaneous display on multiple outputs such as from software mirror to a projector or a visual pager application showing a preview on a second display.

The best solution is to have the Compositor handle the final colorspace conversion. It knows exactly where each pixel is going to be displayed because it will be putting that pixel there. There has been some work towards this (but not mentioned in any slides or the writeup) from a Google Summer of Code in 2008. It was called "Color Management Near X" and I believe the "net-color" standard. The solution was not to send colorspace with each pixel, but rather to set window properties defining the icc profile for the window or regions of the window. I should note that specifying client's colorspace to be the same as one of the outputs results in a null transform. That gives the application the same power as before, but allows the Compositor to give sane colors on other displays. Sadly I can't find much information on it with google.

As it stands "color managed applications" can display correctly on my wide-gamut LCD or my normal gamut LCD but not both. The colors will be flat wrong when they appear on the "other" output.

SCALE 8x: Color management for everyone

roelofs — 2010-03-05T23:39:59+00:00

ok. so the application would just need to pass some colour space metadata along with the pixels.

No--any application that does its own alpha-blending, for example, needs to convert the image(s) and background from their native color space(s) to linear (gamma = 1.0), do the compositing, and (usually) convert back. Other transformations (lighting/shading calculations and whatnot in 3D, IIRC) also require linear gamma for correctness. And it's hard to imagine a driver complex enough to support multi-source display app like a browser ("this block of pixels is linear, this block uses this custom ICC profile, that block is SGI/gamma-1.7, and everything else is sRGB"). Conceivably it's doable, but I'm not sure a driver is the best place for it.

Greg

YUV and YCbCr irrelevant

farnz — 2010-03-05T10:02:20+00:00

Depends on how you define a colour space; on the one hand, they are just a reversible transform from a known RGB colour space. On the other hand, practical YUV colour spaces, while sharing primaries with practical RGB colour spaces, tend to have different gamuts.

Additionally, they only correspond perfectly under the assumption of infinite precision; in colour management situations, it's of interest to the output driver to know that when I ask for 8-bit BT601 RGB (236, 250, 255), it actually came from YUV (230, 140, 120), and thus if display RGB (237, 251, 255) corresponds to BT601 RGB (237, 251, 255) while display RGB (236, 250, 255) corresponds to BT601 RGB (235, 249, 255), the former is a closer approximation to the desired colour than the latter.

YUV colour spaces differ from JPEG in a very important manner; JPEG is lossy by design, whereas YUV isn't lossy until you subsample. Any loss between an RGB and YUV colour space occurs due to loss of precision

Image scaling problem

rgmoore — 2010-03-05T02:02:13+00:00

If as the original poster said, everybody could assume the image is sRGB, then the scaling algorithim could be designed to correctly scale sRGB. This is much easier than something that can scale "anything".

A) I'm not sure that it would be any easier to do everything on the raw sRGB data. It's not just image scaling but all aspects of image processing that are easier to do on linear data. It's likely to be easier to write one algorithm to convert sRGB to linear and one to convert it back than to include an implicit implicit conversion in every image processing algorithm. And if you care about correctness- which you obviously do if you're bothering to worry about the gamma applied data- it's going to be much easier to prove that you're doing everything correctly by working on explicitly gamma corrected data than to count on having the correction in every routine.

B) The "everything is sRGB" assumption is untrue. Real world programs have to deal with all kinds of color spaces. Once you have to deal with more than two color spaces (sRGB and linear), the need for real color management will be much more obvious.

YUV and YCbCr irrelevant

spitzak — 2010-03-04T23:07:56+00:00

Spaces such as YUV and YCbCr are not "color spaces" really. They are simply an encoding of an RGB space, and any definition of them will include the values of the R, G, and B primaries and the fact that they are YUV or YCbCr encoded. There is no such thing as a "Y primary" and certainly no such thing as a "Cb" primary (since there .5 is black!).

Claiming these are color spaces is equivalent to claiming that jpeg is a color space, rather than a compression algorithm.

Image scaling problem

spitzak — 2010-03-04T23:04:44+00:00

This has little, if anything, to do with "color management".

If as the original poster said, everybody could assume the image is sRGB, then the scaling algorithim could be designed to correctly scale sRGB. This is much easier than something that can scale "anything".

Also from everything I have learned about color management, there appears to be a need for a "blending space" that is controllable and that scaling and mixing is always done linearly in this blending space. If this blending space is sRGB then the scaling is in fact required to do the "wrong" result. You need to change the "blending space" to some linear color space for blending to be correct.

SCALE 8x: Color management for everyone

rgmoore — 2010-03-04T17:00:18+00:00

i have never understood why applications need to care about colour management, surely it could all be done by the graphics driver.

That works for most applications that are only interested in displaying graphics; they send the color information (possibly with an ICC profile) to the graphics server and don't worry about anything else. But it's obviously not enough for a program that's supposed to be editing the graphics information itself, e.g. GIMP, Inkscape, etc. Those programs need to understand the color information to be able to edit it properly.

A good example of this kind of problem is one that I saw recently about problems with image scaling; most image processing applications do it wrong. The problem is that their color information has a gamma of 2.2, meaning that displayed intensity is supposed to be value**2.2. The correct way to apply scaling is to convert the color information to a linear value, apply the scaling, and then convert back. Instead, most image processing applications use the values with gamma applied, which results in the scaled images being too dark. A properly color aware application wouldn't make that mistake.

SCALE 8x: Color management for everyone

ssam — 2010-03-04T15:02:26+00:00

ok. so the application would just need to pass some colour space metadata along with the pixels.

SCALE 8x: Color management for everyone

farnz — 2010-03-04T14:28:30+00:00

sRGB isn't the only standard RGB colourspace. Just off the top of my head, I can think of sRGB, Adobe RGB, BT.709 RGB, CIE 1931 RGB and I'm sure there are more out there. I can very easily find images in three of those colourspaces (my digital camera can produce two of them, I can get a third by grabbing a frame from BBC HD and converting from YUV to RGB).

Plus, many images aren't in an RGB colourspace to begin with; if the image is CMYK (common if it's being prepared for print), or some variation on YCbCr (common if it's from TV). The application needs to tell the graphics driver about all of this information if you're to get a precise match; what's more, not all colour values are in 8-bits per channel RGB; if I had 16-bit per channel CMYK, I'd like the graphics driver to know that I want a slightly darker red than the RGB tuple suggests, so that it knows which way to round error.

This becomes more important on 30-bit displays (where an RGB tuple is three ten-bit values, not three 8-bit values), as the conversion formulae are not precise, and it would be nice to use the extra accuracy to improve colour reproduction.

SCALE 8x: Color management for everyone

ssam — 2010-03-04T13:51:56+00:00

i have never understood why applications need to care about colour management, surely it could all be done by the graphics driver.

i have an image file where each pixel is given by a red, blue and green component, and a standard that says what a give combination or r, g and b should look like (sRGB). so if my pixel is 'lwn orange' #FFCE9C, it could be well defined what this should look like.

now my image viewer tells the X server to paint the pixel #FFCE9C, X passes this value the graphic driver, and the graphics driver figures out what signal need to be sent down my dvi/vga cable to show the colour.

if its the application doing the translation, then what happens if i take a screen shot? also what happens if the application window is spread across 2 monitors?

Thank you

ncm — 2010-03-04T06:20:52+00:00

Here is yet another article that makes me glad I subscribe to LWN.

SCALE 8x: Color management for everyone

joedrew — 2010-03-04T04:40:29+00:00

For Firefox's colour management, we initially used LCMS, but found that our needs and desires increasingly didn't align well with the LCMS developers' needs and desires, so one of my co-workers, Jeff Muizelaar, took parts of LCMS and made a similar project he called QCMS.

As of Firefox 3.5 (and continuing in Firefox 3.6), we colour-correct tagged images by default; users who want to change that default (so that we treat untagged images as if they're in the sRGB colour space) can do so easily; check out the MDC article on ICC color correction in Firefox.

SCALE 8x: Color management for everyone

azouhr — 2010-03-02T15:03:17+00:00

On software.opensuse.org, you can find Oyranos packages for

openSUSE 11.0, 11.1 and Factory
Mandriva 2007
Fedora 10

SCALE 8x: Color management for everyone

hughsient — 2010-03-02T12:22:47+00:00

Yes, and it can now also create profiles for cameras, scanners and printers, assuming you have the correct type of hardware. I released 2.29.4 yesterday, which is a much more complete project than any of the pre-release snapshots. I urge anyone interested in the GUI (pretty bits) to join the GNOME Color Manager mailing list and discuss with us future direction. Thanks.

SCALE 8x: Color management for everyone

lolando — 2010-03-02T09:34:22+00:00

GCM can use Argyll, if it's installed. That's only one part of the functionality though, another one being the installation of the profile (whether obtained from Argyll or from the monitor's driver CD) into the session/X server/etc.

SCALE 8x: Color management for everyone

boudewijn — 2010-03-02T07:38:22+00:00

There is a color management settings module for KDE already, but it depends on Oyranos, which isn't packaged by any distributions that I know, probably because of its dependency on the Elektra configuration system. So the kcm isn't in kde by default yet.

Krita is a Qt-based (well... KDE) that has used lcms (and therefore icc profiles) for color management since version 1.5 which was released in 2006. KDE-based Digikam has support for color management as well, I think also since 2006. Qt-based Scribus ditto.

SCALE 8x: Color management for everyone

cuboci — 2010-03-02T06:39:26+00:00

Doesn't Gnome Color Manager use Argyll behind the scenes to create ICC profiles for your
display? Last time I checked it was able to use my Spyder2 Express device to measure my
laptop's display.

SCALE 8x: Color management for everyone

nlucas — 2010-03-02T02:25:13+00:00

Just for curiosity tried to google a bit for ICC work on QT, but didn't find
nothing. Is this something only GTK+ devs are working on?

SCALE 8x: Color management for everyone

mattdm — 2010-03-02T00:46:14+00:00

It is promising that headway is being made on that front as well, with GCM and GTK+; perhaps in a few release cycles Linux will have full color management out-of-the-box.

See http://fedoraproject.org/wiki/Features/ColorManagement targeted for the upcoming Fedora 13. I'm not holding my breath for RHEL6, but hopefully this will make it in there too.