#apertus IRC Channel Logs

intracube: it was me. I said junk.

I meant it

not sure how that would be quantified though

intracube: it is junk. It is guessed.

(You can see the SIGGRAPH papers from 2014)

No questions asked. You simply cannot know what was there nor implement anything into a pipeline.

It just cannot, will not, ever work.

No it isn't!

End.

It _is_ one variant on a visual effect

So yes great for a post production visual effect pass. Absolutely worth a try.

one is the simple case where one channel clips out

It isn't one channel.

possibly different for each channel

?

But I will say beyond certainty that the percentage of overlap with actual shooters trying to finish projects are about 5%, and if that 5%, 0.05% attempting a project with any degree of complexity in the post pipe.

troy_s: probably best I sit at the back and just listen for now :)

intracube: You know more about colour than many.

intracube: Also, wow... DJV is a lovely piece of work. Helluva impressed. Lovely code too.

news to me

looks interesting

troy_s: this is not missing data, this sensor is not clipping harshly to white; rather, it's nonlinear and noisy data

you are talking about something else - highlight recovery by guessing the clipped channels from nearby pixels (which is, indeed, guessing)

alexML: We had a similar issue with the Alpha with irregular data. It was a long path and after looking at the profiling output, it was clear that the (seductively looking) result was non-data and ended up just chopping it off.

Bertl: Yep

I am looking forward to seeing irieger's early PWL tests.

troy_s: by irregular you mean the repeatability wasn't good, or some other issue?

alexML: Seductively looked otherwise, but the delta Es didn't lie.

for the 20ms (correctly exposed) it8, I get Profile check complete, errors: max. = 22.797085, avg. = 3.219586, RMS = 3.991097

indeed, pure junk :P

You know you can spit out a DE per swatch right?

Remember that a delta E is an average across all of them in that case, so by suggesting that you are closer

Aka: even the dE is junk in that case.

You have to analyse each swatch against the value and the dE.

TL;DR of _course_ you can get a lower average dE by more greatly mangling a few swatches of data.

here you go: https://dl.dropboxusercontent.com/u/4124919/bleeding-edge/apertus/highlights/it8-argyll.log

alexML: That last slew of data is what is important, and that will reveal all.

Ah got em.

not sure I understood your question

That is pre massage. The post massage has a ∆e of 21

The bulk of those readings look junk.

Very, very bad.

Let alone >10 and >15

alexML: Do me a favor and dump a log from an untouched chart with synthetic versus your best massaged.

the second half is from *untouched* chart!

again, the 80ms one is overexposed and recovered by curves, the 20ms is exposed to the right and not touched in any way other than subtracting a dark frame

alexML: Ideally we have two full logs.

alexML: the log is in descending descending ∆e

yes, my log has two parts (second part starts at line 411)

so there are actually two logs, I just placed them in one file

gotta go

alexML: Where is the 00W swatch?

I removed it a while ago, because it was ending up with max delta E

It will end up with max delta E because the matrix doesn't fit.

Ugh.

In the case of the white swatch, it is an absolute baseline reference which is good in this case both as an indicator as to how out of whack the results are but also to keep things perfectly scaled. That is, the black and white swatches will help to anchor the chart absolutely.

Because the highest delta Es currently are on the mustard D08 swatch and the fully saturated L15 swatch, which at least hints that the most out of junk is the Red and Green photosites

So without looking at the data or knowing what you did

That's just going by the data.

But the values are roughly equally crappy on all fronts.

lower is better, right?

alexML: Because in this case we are taking a constellation of 3D points and trying to fit them to a constellation of other points.

alexML: So to draw an analogy, if we have three points that should fit in a line, where the first point is A, B, and C, we can of course cheat C to make A and B much better

So what worries me is how ghastly out of whack L15 is.

That's a really not super duper good value.

But my gut says something more funky at work.

Also what is worrying is D08

alexML: C08 then diverges on your chart, which, as hard as it is to believe, gets colossally more worse from nineish to twelveish (which is roughly like suggesting that going from the Titanic to the water is worse)

But again... not worth reading too much in. The results are pretty unfortunate (although vastly better than the Alpha it seems), and the tweaking isn't doing anything.

C08 is better in the massaged data, btw

Uh no.

to 12.73

12.73 is plain data

Right?

Peak error 22?

(already told you this 10 times)

It all looks the same.

Peaks that far out of whack just stinks of junk.

which means - that thing you called junk isn't really worse than the unmassaged data

It means it is fricking bad in the unprocessed.

And if you want to draw dickle lines between a dE of 18 versus a dE of 22 go nuts. I won't stop you.

Which means values that high are are just whack.

and look at say, only values below 6 (caveats above)

charts even

Just a random value.. I yanked H11

to a nasty ass 2.99

in the case of the former, it is an actual useful value

let me look at another

26 swatches deviate below 3.0

145 swatches above dE 3.0 on one chart, versus 119

We can weight those dEs.

irieger: G13 (Cyan) - dE 1.27, G14 (Magenta) - dE 5.36, G15 (Yellow) - dE 2.05, G16 (Key) - dE 0.57, G17 (Red) - dE 0.63, G18 (Green) - dE 3.07, G19 (Blue) - dE 5.57

(Granted I am sort of looking at the data blind in reference to the exposure etc.)

Almost as though the B photosite is giving bad values, which of course could be due to the whacky lighting used for the chart.

off for a nap ... bbl

irieger: I mean look at those dEs on that middle ish row.

irieger: Do you, in your educated position, think those two swatches hitting five plus are reasonable?

irieger: With red being almost perfect (so whatever band crosses that combination of inks/paper/gloss and the Apertus RGB sensels, has no problem there.

So it seems to be having problems fitting green and blue.

irieger: See my point?

irieger: Look at my line above.

irieger: At "middleish" perceptual values I'm guessing.

Well three...

(just because it seems not quite as obvious and glaring as the other three)

Magenta, and Blue.

So now look at Magenta.

Is that the blue's photosite contribution that is driving it to a not quite super fit?

(I suppose optionally we could say that the B setting is fine and that G and R are out of whack progressively more as well.)

I've got some graphs btw

https://dl.dropboxusercontent.com/u/4124919/bleeding-edge/apertus/color-calibration/it8-20ms-my-matrix.png

alexML: And what are you using to generate the xyY map? colourscience's python library?

I used octave, and took the data from here: CIE 1931 2-deg chromaticity coordinates from http://cvrl.ioo.ucl.ac.uk/

(I plotted this in python as well, but for matrix stuff I prefer octave)

Lessens the chances of mistakes.

second image shows both matrices in xyY

It only shows two.

Would be interesting to see how shapers influence the dE on the brown gold and yellow, as well as the blue ramp.

Hrm. Interestingly on the axis of the green primary (fictional obviously) argyll nails most values except for some of those golden ones

these are the residuals - in camera raw space - after going from xyz back to cam (with the inverted matrix)

well, the graph shows recorded raw values vs raw values computed from ref xyz via inverted matrix

alexML: What do the two axes represent?

Y: recorded raw values

both are log2 graphs

Sort of a RMSE chart then?

So no relation to saturation or empty wells.

no trickery on this one

on the unmangled data

Let Argyll do its best work with a shaper and a matrix

Then examine the dEs based on swatch positions.

(80% linear)

But I find it quite interesting that the pure blue ramp on the IT8 and the magenta (granted it is an ink based magenta, not a light based one of course) are quite whacky relative to the other values.

Shaper _will_ help the blue.

I assure you that you will see much better dEs.

in the big logs I sent you back then, shapers were enabled

Ok. So that double log comparison is _with_ shaper?

you have the full commands listed

Oh god.

Also adds code for error. Bertl's skip command is probably much more safe on the whole.

Egads.

this?

that chart looks way underexposed unless I am drunk.

log2(rgb values)

linearized RGB

I mean white is coming in at 0.36ish right?

where are you reading the 0.36?

Argyll dumps the RGB input values so you can immediately check for clips and such.

with very low exposure from what I can see there.

(obviously what many folks who expose 'typical' shoot the chart to slap middle grey in the middle, when what we are really wanting is the data to all land in the most linear region of the chart, which is easiest to peg if we plop the diffuse white about 80% up the well fill zone roughly.)

not sure what units are those, but the second chart is really exposed to the right

Hrm. Need to check the output log for the output there but I seem to recall that one of the tools dumps RGB

(or xyY, etc.)

Columns even.

Look at the actual values in the Argyll output, and specifically GS00

Easiest method to tell where the ideal exposure is.

If those are linear RGB

(Format tag)

Hrm. So I had it backwards.

;)

35.2, 46.12, 31.28

The XYZ would be pretty darn close to ideal if they were the RGB values if memory serves (which it has done less and less so of late.)

let me fill an image with white to check

I'm betting they are _damn_ dark in the sensor wells yes?

Heavily in the nonlinear region.

It's really lovely. Also has modern calculations for dE etc.

yes, the darks are dark, but the highlights are right below the "clipping point" (where it starts being highly nonlinear)

Uh what?

The sensor starts clipping at 0.5?

if you don't believe me, download the raw12 and look at them yourself

It's not that I don't believe you.

Link me to the IT8 in question?

I cannot fathom how 20 gives us values at 0.4 and 25, which is 1/8th of a stop, blows out.

it8-gainx1-offset2047-20ms-02.raw12

alexML: Are these daylight?

GS00 is 2235/2908/1979 in raw

no

FFfffffffffffffffffffffffffffffffffffffffffff

Sort of moots the whole thing.

Is that irieger's work?

you can see them in the file name

Ugh.

I guess it's irieger's work, not 100% sure

don't forget this sensor doesn't clip at 4095

egads.

We have to chat about exposure on these damn test charts

correction: the values on 20ms GS00 are 1541/1975/1963/1382 (RGGB), and black level is 143

so looks like this one clipped only on background, but chart data is usable

Got it.

Odd.

for me it wasn't that obvious, especially if you actually look at the samples

if you expose just a little more to the right, you are going to get "non-data" in the white patches

alexML: It isn't obvious.

it was curved

alexML: See how the boxes line up

alexML: Simple blending perspectival should be fine.

do me a favor and tell me the command for this particular file: http://files.apertus.org/AXIOM-Beta/snapshots/HCT%20Nikon-08.01.2106/HTC-gainx1-offset2047-20ms-02.raw12

to fix that "perspective"

https://dl.dropboxusercontent.com/u/4124919/bleeding-edge/apertus/highlights/HTC-gainx1-offset2047-20ms-02.jpg

anyway, if argyll uses robust averaging, it should ignore these misalignment

I'm surprised the boxes escape on that. It isn't that horrible.

So 15

yes

Crack open GDocs and import the TI3 along with the profile check into pages.

kind of, but lines won't do in this case

probably the best way to map it to a plan surface is a Shepard's distortion

yup, you would need to map each cross point (or at least a fair number)

Ugh.

Not worth it.

(dE2K is a more modernized dE calculation.)

Bertl: ya figure?

how do I find that?

alexML: You can change dE output in profcheck. Let me find the switch.

dash k alexML

Q09,Q08,R08,H13,I16, lost patience

That's Hutch?

Sorry was referencing the IT8 values.

Damn. Magenta is peaking dE again.

=> nonuniform illumination?

X as in horizontal position on the chart

Then yes.

alexML: Also, if you look at your 20 and your 25, then look at your hutch, look at the magenta strips.

probably, I'll stop it for today

I'll correct for the nonuniform illumination first

alexML: Did you pop these shots????

take

Argyll will fit for that to a degree. No sense in trying to fudge it.

It fits it.

But ultimately, the issue would be that fudging the data would be defeating the purpose.

We kind of are doing the tests to see how the tests align with the known constants, at that particular illumination level because of all the nuances of the photosite's lenses etc.

1) Square flat chart. No glare.

That will solve everything.

The rest is all junk.

We need that D50 full spectrum illuminant. It's pretty darn critical given that our reference data is there.

Helps the fitting.

they are distributed around the chart

Yep.

their location does matter

The fitting will have the same spectro values

of course

Pretty hard to unevenly light with the sun or several kilometers of sky

I doubt it's that easy

That said, by and large, the most important thing is that the CIE values from the spectro match the illumination entry point. Otherwise all bets are off.

And of course, having one's own spectro makes all of this much better.

just for fun: http://www.brucelindbloom.com/index.html?ReferenceImages.html

for sure

after evaluations.

But I guess the PWL is sort of the "next" thing needed really.

So PWL, then daylight charts, then perhaps a twiddle session.

I think most spectral renderers rely on 40 or so.

Bertl: Spectral renderers don't.

you know any device/sensor which has more than let's say five filters per pixel?

Are they?

They are baked spectrals.

you could already do pretty good with interference filters today

(at least none I know of :)

Purely for textures and what not.

WETA's entire pipeline has been more or less replaced with spectral imaging rendering.