#apertus IRC Channel Logs

off for a nap ... bbl

Bertl: FANTASTIC Bertl !!!

Bertl: 12 mean error is crap. You can use profcheck to get DE values

Bertl: Use DE2000

(as in crap)

so you've been giving me crap all the time :)

appreciate it, but seriously, any idea what the problem with the calibration is?

Bertl: To be fair, all Gabe and I do is take the crap you two give us and reshape it into different crap. Lulz.

Bertl: Which is inconsistently lit (top to bottom) and sub exposed.

Bertl: Se6's are very exhaustive

and as you say, uneven lighting, could this be the main issue?

Bertl: It is also a scanner image.

okay, we haven't found a 'camera' chart yet

IIRC, somebody suggested to send us one, but then didn't

(I think Se6 bought a new one)

3) I haven't twiddled enough yet.

*AFAIK

maybe but please double check with sebastian, I had a quick look at the ITU docu, and it doesn't really differentiate between scanner or camera

http://www.targets.coloraid.de

Yep

I am unsure. I suspect the camera charts have some anti-yellowing twinges to pull the illuminant to more neutral

Well a glossy chart has a wider gamut, but more difficult to shoot.

Bertl: But again, we are miles away from a high grade calibration

Bertl: I just want to get to a 2de zone

Which is more than satisfactory for me.

Bertl: If you can tiff 16 that link se6 sent us

I will do some more tests tomorrow

that looks somewhat fishy to me with 'red' around 2 or so

Which is bad

Basically our chart shot stinks

exhaustive

With well noted color temp etc. I want to generate a neutral (as in color temp is achromatic equal channel)

Relatively smooth, but wonk

Linear 16 bit TIFFs of the new batch would be excellent

(much dickery to get things working for forward thinking items)

I have about 98% of that done, just need to cooy / paste some code.

A viewer app that will spit out TIFFs and EXRs and DPXs with full color management.

convert \( -size 4096x3072 -depth 16 gray:IT8_incand.raw16 \) \( -clone 0 -crop -1-1 \) \( -clone 0 -crop -1+0 \) \( -clone 0 -crop +0-1 \) -sample 2048x1536 \( -clone 0,1 -average \) -delete 0,1 +swap -combine IT8_incand.tiff

What does sample do?

reduce the pixels to separate the bayer pattern

this is a simple/primitive debayer, we have it on the wiki since I came up with that

Genius boy.

there is a small variation if you have flipped images

convert \( -size 4096x3072 -depth 16 gray:colors_500ms.raw16 \) \( -clone 0 -crop -1-1 \) \( -clone 0 -crop -1+0 \) \( -clone 0 -crop +0-1 \) -sample 2048x1536 \( -clone 2,3 -average \) -delete 2,3 -swap 0,1 +swap -combine colors_500ms.tiff

i.e. if the first one looks wrong, take the second one :)

And why?

the sensor can do image flipping, horizontal and vertical

in the beginning, I turned that off, so the early images are not flipped

Need a bash script to do that.

yes, the pattern is flipped as well, i.e. RGGB becomes GBRG

Is there a simple bash script on the wiki?

bash script for what exactly?

To convert the images.

they are listed here for reference: https://wiki.apertus.org/index.php?title=Axiom_Alpha_Software#Simple_debayer_with_ImageMagick:

(Not sure the difference in batches.)

and AFAIK, imagemagic doesn't apply anything unless you specify extra spaces

(e.g. with -colorspace)

a good way to check is with the test pattern

let me see if I have that uploaded somewhere

If you can do a quick test that'd be awesome.

(The PNGs I have _no_ way of knowing, but I do know that PNGs are enforced sRGB so any tool should actually be assuming a transfer curve.

With the test pattern raw data sample?

Where's the tester?

(By the way, OIIO has some image buffer algos that are at least as effective, plus OIIO is specifically crafted for scene referred images - a great toolset for this project.)

I see the hist command on the wiki

if you like, I can also create an artificial one with full gray range

Damn. Looks full frame

(Slightly rotated, but Argyll can deal with that)

probably, try the other line

That's it.

Shuffle does what Bertl?

@shuffle, do you want a technical explanation or a simple one?

Bertl: A technical explanation for someone with an IQ of about 40.

so for example, 0123456789ABCDEF -> 01245689ACDE37BF

so for example, swapping one green with the red channel gives really cool images :)

Damn. Having an error with scanin

Rotate is sucking

is giving me tif_dirwrite.c TIFFWriteDirectoryTagCheckedRational: Assertion 'value>=0.0' failed.

Again, PNG is junk. Argyll won't read it and it cannot be linear.

try with .png, then convert it to tiff

And what's betting Imagemagick will convert the transfer curve?

Huh?

WriteDirectoryTag is an odd one.

and so, imagemagick being lazy does not want to rotate it at all, just sets a flag, this image needs to be rotated by N degree

it should be simple to check the transfer curve theory, by converting a tiff to png and back to tiff, then comparing the two tiff images (original and converted one)

Grr.

Bertl: I can't seem to crop the blasted images down to the chart only and get the rotate decent

what version/OS do you use?

And god I hate it.

(I've come to hate Imagemagick. So crap)

But the crop function is crap

-crop 1841x1305+149+116

Crop should give you a window (if you follow it with a !) that is widthxheight+xoffset+yoffset

http://www.imagemagick.org/Usage/crop/

http://www.imagemagick.org/Usage/crop/#crop_repage

Keeps telling me no image is left

okay, can you give me the complete command you are trying?

check that with 'file <image>'

no, sorry, check with:

it might show you an offset +x+y on the input image

Hold.

Then I cropped using:

and that shits the bed

convert ./out.png -crop 1841x1305+149+116\! +repage ./test-crop.png

So frustrating.

Class: DirectClass

So the resolution is larger than my crop window.

Yay bugs

What junk

well, I can't complain here, it works just fine for me, it is not one of the most intuitive tools out there but in my experience it gets the job done

And ARGYLL BUG

Good luck... Here's what I've tried to do:

btw, if you just want to cut out the chart, it is simple to do with the perspective transform

just find the 4 corner points and use them as input

2) Rotate it and crop it or use the perspective tool.

(linear)

can you do an md5sum <image> for me?

7ffcf417c906fccaf180d739a9d484af

scanin will either segfault (using repository Argyll) or bail out.

7ffcf417c906fccaf180d739a9d484af IT8-tungsten-halogen-Run03-9.1.2014/IT8-tungsten-halogen-Run03-9.1.2014-15ms.raw16

15ms.

yes

nope, sorry

Why not just scale it to original size?

because that gives the correct source for the perfect chart :)

yes, I wanted a border reference of 48 pixels

ARRGGGHHHHHHHFFHFHFHFHFFUFUUUU

'��������������������������������¿����������ü��������������������������������������������������������������������������������������������������������

���������������������������������������������������������������������������������������������������������������������������������������������������ï¿

¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½

½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½' that's too long

convert \( -size 4096x3072 -depth 16 gray:IT8-tungsten-halogen-Run03-9.1.2014/IT8-tungsten-halogen-Run03-9.1.2014-15ms.raw16 \) \( -clone 0 -crop -1-1 \) \( -clone 0 -crop -1+0 \) \( -clone 0 -crop +0-1 \) -sample 2048x1536 \( -clone 0,1 -average \) -delete 0,1 +swap -combine +repage -distort Perspective '254,144 48,48 1876,188 2000,48 1854,1275 2000,1488 233,1242 48,1488' IT8-tungsten-halogen-Run03-9.1.2014/IT8-tungsten-halogen-Run03-9.1.2014-15ms_c

apparently a bugger on larger images

Still bombing.

Good god.

All of my Argyll's die

http://pastebin.com/wnnYJs94

Matrix = 0.107625 1.044939 -0.850347 -0.150982 1.577119 -1.062272 -0.256328 0.420066 1.000908

Error

I'd need to check on how underexposed it is.

If you batch that on all of the images

Bertl: This is the command to generate meaningful delta e's - which again, 1.0 or less is the discernable to average standard observer

and you can plot the curves using

-U in the colprof will apply a linear multiplier to get to 100 L

a near perfect profile will show you three curves

Odd how fanned out the blasted curves are (Got mine to work)

Average DE 10.5

That's ... uh... not good.

Bertl: That will give you some idea as to the readings.

so you already picked the best one :)

Egads.

(Properly exposed)

if that chart made a curve in blue or what i have noticed you talked about, then with linear matrix you are getting big errors, or does the error relate to a 3d lut ?

Did you plot them using xicclu?

I suspect our lighting sucks ass.

Bertl: Again, use profcheck

Bertl: DE is a known idea of value - where < 1.0 is not discernable. > 1.0 is.

troy_s: but the number is linear (e.g. %) or logarithmic (like bits / dB)

guest: Google DE2000 or DE76

ah, it needs the icc, where do I get the icc from?

Bertl: Look in the directory

:)

colprof creates the color profile

xicclu is an xwindows icc look up

try -v2

nice

yep, that does it

Si?

I _think_ that is largely due to the Tungsten

red starts on 200 wossname, and goes down to zero on whatnot

But that tungsten is pretty damn warm

And a well behaved profile should have smooth curves like that

but closer together

and about 100 to 100

guest: Meaning that the errors are biased toward perception and particular hue ranges

That the curves are decent

okay, so that is the profile curve, yes?

That's the profile curves.

philippej: The sensor _always_ needs a profile

where is the picture (jpeg is enough) of the chart you try to use ?

philippej: What you are seeing is in fact random when you view the data dumped directly to a near-sRGB display.

philippej: unless the values are clipped, it is fine even if it is warm

guest: it is a normal IT8 chart

still it looks more correct (or needing less corection) when it's daylight. What you call "damn warmish" :-)

philippej: It's why I prefer to show people in XYZ

Bertl: i meant the shot you pass to the color detector.. export it please to quickly viewable form

philippej: Means nothing.

guest: Hold.

i know, i want to check for other things for which i have made my calibs to take a week (in terms of exposing the chart correctly , and I still have no perfection in that )

guest: It's not that hard.

guest: and the most ideal

I also uploaded the profile results

I can explain that if you'd like, but likely tomorrow

http://www.pasteall.org/pic/show.php?id=66837

in reds

Bertl: -i

colprof can take the illuminant

guest: scanin samples across mean

IIRC

those patches are clipped, so you cant do a calibration with that shot

(Aka 2848k)

guest: That's an 8 bit dump

That's also a linear dump.

guest: That's an 8 bit dump.

HUH?

Some patches suck ass

that's the ODS of the data

you can see the swatches and the readings

guest: pastelink.me/dl/a8778e

guest: YOU AREN'T A GUEST ANY MORE

:)

Bertl: Check out that ODS. Guest _does_ have a point with the heavy exposure on the red channel, but that said, the others worry me more (the higher DE versions)

hello :)

I have to leave for a little, grab my snail-mail, bbs

That's also a useful linky on normalizing.

In humans. :)

(Shooting a red swatch in a blue light for example)

i did my calibration on cloudy sunlight

danieeel: Too blue

danieeel: ICCs are all based around D50

my point was rather to get correct images under skylight

danieeel: You want to eliminate the blue sky

with that i can not yet find a location which is not colored - it would need to be some skate field or some grayins surroundins to do it correctly

good morning

we can ask the pilot of our flight to berlin if he would mind a little detour ;P

Night all.

danieeel: You basically take the XYZ values, figure out the multipliers and scale them all

Ok... nacht.

back now ...

wb se6astian!

off for a nap ... bbl

hey philippej

time to leave

Bertl: Back for monkey work for a bit.

danieeel: If you are around, I can try to explain how to normalize your whites (Essentially Wrong Von Kries AFAICT)

i do not need to (yet) - found another issue

anybody else interested? or we let them to read back from logs?

danieeel: I suspect very few would be.

(and logs are decent.)

i thought of simpler charts too.. but I ended up with HCT

HCT is ideal

But more swatches also will reveal more inconsistencies

(Making LUTs more considerate of inconsistencies, but also more prone to potentail errors based on photograph of chart for example)

RGB YCbCr isn't nearly as complex as this.

(Two relative color models. Dead simple if you focus on that. One with by-the-book scaling if you are going to call it 709)

Good times.

So the TI3

See that spreadsheet I posted?

column names pleas

no

oocalc

what is that patch/swatch difference?

danieeel: Because when you do a scanin, you get a ti3 file

no

ok

danieeel: And to be clear, you are aware of the differences between RGB and XYZ (relative vs absolute and unscaled)

(XYZ is technically a "scene referred" absolute color model. See J. Selan's PDF on cinematic color if you still don't quite understand scene referred versus display referred (http://cinematiccolor.com))

Taking values is meaningless.

If I give you any display referred RGB values, you cannot replicate the color without asking what the primaries are (which is part of defining a color space)

So when someone says "I have RGB values 1.0, 0.324, and 0.4333" it means _nothing_

(Similar to saying some other stupid statement like "This is in linear color space!" Linear is an _attribute_ of a color space, not a color space. RGB is an attribute of a color space (the _model_) not a color space.)

display referred = places special meaning on a 'maximum' value - generally 1.0 (or whatever the bit depth permits). Sometimes called output referred.

that is ok

display referred is before or after the transfer curve?

display referred is an attribute of the data

device referred. display referred. output referred.

they have a minimum and a maximum.

ok

yes

if you can set on it an linear input curve

Because you are shipping linear values and your display is LDR.

Low Dynamic Range.

So your eye won't see it as the linearized values were in 'the real world'.

To 'cheat' the values

IF we had some UberTV, and the level of luminance were Uber(tm), we could in theory ship linearized values directly to the display

we should make a real example - tell me if that is correct

You go out and manually make a real world image in greyscale

so you take a spot meter and meter every single 'pixel'

and we are going to use linearized values

and the grass is f2.4

(the grass is)

(arbitrary)

If we did this for every pixel in the scene, we'd have a greyscale image in theory.

yes of course

i have got my linear data, and used a lut to view them correctly, in terms of intensity

If we dump that file to the display

(when they are displayed) maintain the physical real-world ratios

danieeel: Clear?

of course, have experienced that

2) display referred versus scene referred

Clear on all those points?

display linear: trim/crop/clip that to some usable max value

danieeel: A scene referred image of a typical landscape with sun in the frame would have ranges from extreme low levels in shadows, to extremely MASSIVE numbers of the sun itself.)

You got it!

As you will hopefully see in hindsight ages later.

now.. the referred - it points to two sides of the transfer curve?

as a point of conversion between scene/linear to display/curved

B) scene referred (and hence scene linear)

Clear on XYZ?

in scene linear, the Z is any number

:)

Clear?

Whoa! DIFFERENT MODEL FRIEND!

"increase intensity of a color" is very unspecific what you mean - saturation / brightness?

The side of the barn

XYZ all grow

Clear?

okay, assume that. (not used to it yet) so continue

XYZ is a unique model

Using three lights.

a green light

with SPECIFIC colors

(That is, for RED for example, the color does not CHANGE as it increases in intensity.

Clear?

The CIE used three lights

observers tried to match the colors in the swatches

This forms a TRIANGLE at the base

SOME swatches however, could not be matched using the three lights.

And so the researchers added small amounts of the primaries to the swatch values (thereby decreasing their saturation INTO gamut)

By figuring out how much the observer added AND paying attention to how much THEY added to the swatches, they could fix the 'actual' color position relative to the smaller gamut three RGB primaries.

(the transform of color in the real world is a simple thing - it is just a transform because all the values are linear as per energy laws)

Or something not quite clear?

See a swatch in front of you

you can dial up the lights

When I show you some colors, you can't reach them!

So I can desaturate the swatch green

until you can match it.

so if i have a very green swatch, on a poor green light, it will be percieved poor

the color is not saturated

and it will be the same green

Clear?

So we need to figure out a way to replicate it and figure out what the green was

So I 'reel it in' by adding RED and BLUE

Ok

and we have our ORIGINAL three CIE RGB values

Clear on this?

Which extend beyond the legs of the triangle

The problem is, our RGB values go from 0 to .... HEY THAT'S REALLY BRIGHT

When we scale them in

Which would be again, our CIE RGB triangle

And that forms... something... that ... looks like...

org%25252Fwiki%25252FCIE_1931_color_space&source=iu&usg=__PVA0sud8qse4CMkY1BYQPrCrkdQ%3D&sa=X&ei=Ccj_UoWXHZDjoAToh4CgAw&ved=0CDAQ9QEwAQ&biw=1145&bih=786#facrc=_&imgdii=_&imgrc=EkYw_LbmADWVaM%253A%3BSKRQ9HbFToIhFM%3Bhttp%253A%252F%252Fupload.wikimedia.org%252Fwikipedia%252Fcommons%252Fthumb%252F6%252F60%252FCIE1931xy_CIERGB.svg%252F325px-CIE1931xy_CIERGB.svg.png%3Bhttp%253A%252F%252Fen.wikipedia.

Grr.

http://upload.wikimedia.org/wikipedia/commons/thumb/6/60/CIE1931xy_CIERGB.svg/325px-CIE1931xy_CIERGB.svg.png

is 'scaled' RGB values (actually another model we are almost at.)

and the stuff that escapes

So... see the bluey green area? That's escape

There is much more slight escape in magenta / orangey

if we were to take 3 (r/g/b) lasers, the triangle would extend to the side of the colored area

but once we use 3 not so saturated lights or filters, the triangle is all inside

So equal intensity GREEN and RED are the MIDDLE of that line

that is the full line

Sense?

Yes... but there's a tricky thing to that chart

the chart is like looking DOWN on that pyramid

like - more up you are, smaller the triange?

The chart is like grabbing it in a scaling image app

as we scale it upward, we approach infinite intensity

Just like scaling the image

the out of gamut is always outside it

danieeel: Clear?

okay, fine with that

They could use RGB and have negative R, G, and B values to represent human standard observer vision

or they could 'normalize' them

So they did this

would be at 0 0 in that chart

the 'RED' would be at way the hell off the chart

Forming a unique triangle

they don't exist in the real world

danieeel: Clear?

yes

Let's call it XYZ

ah

(But remember, it isn't that simple, all colors grow UP and OUTWARD)

Nope

B) the XYZ model is infinitely large

C) Each X, Y, and Z are TWO coordinates

We are inside a 3D space, NOT a 2D.

So we have a NEW system

how does it get mapped to 2D XY?

Now to TALk about color

wait

XYZ can represent any color of any intensity

Yes.

Absolutely.