#apertus IRC Channel Logs

yep

*this

I think that'd be a pretty good idea

But who would maintain this?

yeah, definitely

It's like with Gimp. I wish there was a page I could look at on their site that tracked the development progress they're currently making

doesn't have to contain much, shouldn't be longer than a few pages (3-4 at most)

but it should definitely be updated on a weekly basis (maybe sometimes twice a week if something interresting happens)

back to the noise/artefacts, can you try to explain to me what you are seeing in e.g. http://vserver.13thfloor.at/Stuff/AXIOM/RAW/colors_500ms.png ?

I also see vertical and horizontal lines

for me, actually the horizontal lines are more visible

yes, same here, it's just that my IRC window is blocking most of my web browser as I type this, and the vertical lines are easier for me to see

but it might be the result of the debayering, we should look at the raw grey images instead

I wasn't sure if they were hot or dead pixels on the defective sensor

I'll email Sebastian re: creating a development section on site, updated once to twice per week

ok, thanks. Have a great one!

FergusL: Greets. About to shower then I have a few...

evening Bertl

If someone can pastebin I will read when done.

troy_s: Bertl and I can easily sum up

and quickly we came to matters which I discussed with you : linearization, corrections made to the image

Bertl: did you continue to investigate about yesterday's issue ?

the fixed stuck pixels

wb

FergusL: What in particular regarding linearization?

(and have the primaries been mapped for the sensor yet?)

Bertl: Or pre?

troy_s: we have pre and post debayering data

the .raw8/.raw16 are pre debayering

please define what 'native primaries' means

Are you familiar with the CIE research?

Oh. Good time to learn a tad about it, but I can give you a condensed notes version if you like.

yes, that would be appreciated

In 1931 two scientists (Guild and Wright IIRC) concurrently were trying to map the range of average human vision.

With me so far?

Color is much more than wavelengths of light... it is known as a psychophysical phenomenon - it doesn't exist in the real world but is rather how our sensors interpret the data.

not just the sensors, at some point the brain too

The researchers collected data using a very simple test. They took the most saturated versions of tri color light in red, green, and blue and presented swatches in a 2 degree circle to observers

Strictly controlled such that the observers could dial in the three lights to match the swatches. The three primaries - or colors of the lights - were carefully selected such that equal parts added to an achromatic color (white if you will, not that white exists per se)

somewhere on the web, so that I get an idea how they looked like

okay, no problem

so basically they had to figure out the R,G and B values for a saturated color

The problem was that there were a good number of colors that the observers could not dial in with the lights - they were beyond the 'gamut' that the three lights could express.

(bear with me :))

Does this make sense so far?

So to make the swatches matchable, the researchers added quantities of the primaries to bring them into gamut, making them less saturated.

reading up on sRGB, this is what you are referring to, no?

It became known as the 1931 2° XYZ color space.

And if you see most horseshoe charts out there, that is a derivative of it known as the xyY chart.

That is the CIE xyY chart

okay, understood, so the primaries are 3 points in that space now, yes?

Exactly.

and how would I go about mapping them?

To "describe" a color, a simple xy pair can work.

Of known photographic densities

I see, well, we will hopefully get some charts in the near future

Is the data off the sensor curved? or lineae?

linear?

(you can use many charts to get a ballpark idea - IT8 or even a Gretag Macbeth chart can give you a rough idea)

(tone response curve) if it is baked into the data.

To make it more agreeable to LDR display.

s/>/?

SHOULD be estimates. Every sensor would deviate more or less depending on quality.

http://vserver.13thfloor.at/Stuff/AXIOM/RAW/colors_500ms.png

How are you debayering?

What algorithm?

Because the debayer is likely applying a TRC to the data.

by just mapping the channels manually

"applying a TRC", like "bending" it ?

Exactly.

Of course the bayered image is nothing more than a greyscale, but if you treat those values as raw rgb, the image is probably a linear one. Or it is encoded into a YCbCr signal etc.

I could sponsor one for the project probably. Wolf makes decent ones.

Ha, that could be great, we were considering ordering one in some way and sending it to Bertl

woh

C1

as there are only 24 swatches

IT8s are a reference standard. There is another with many more swatches.

Some known white point film lighting gear like a Kino would be needed as the white point must be known (and checked if possible using a color meter)

Expose middle geey

grey

largely because to compute the matrix (most sensors are uniform and fall apart near the high and low end)

you _must_ know the color of illuminant.

http://vserver.13thfloor.at/Stuff/AXIOM/RAW/colors_500ms_composed.png

this is after separating the channels, throwing away one green channel and combining them to RGB

Some processor level TRC

Odd as hell.

(I am wondering why and how they would apply a TRC to the bayer sensel data.)

well, in some word the processor is kind of "ours"

(although it sure doesn't look like a log curve)

so whatever curve or transformation happens, it should be the ADC profile

nope, removed

Bertl: about the sensor doc, do you have a link to it handy

Sorry. Hotel wifi punted me.

hold on

Still needs to be accurately mapped so that you can invert it (yuck) to get to a scene referred image (or close to it)

yes

I still have hope for something else going on

FergusL: It is entirely possible there is a register to poke to get the raw data.

FergusL: My worry is that (and yes it has happened) that some sensor manufacturer out of China or wherever is applying a transform to yield colors they like or think looks "right"

FergusL: And to prove my depressing point... read this...

but I guess it should be trivial to check for linearity by taking samples with linear exposure times, no?

lear incresing exposure times that is

hm...

higher values ?

(my take on it)

and configure exposure times from e.g. 5ms to 500ms in 5 ms intervals

and then I plot the centers of the to be expected bell curve in a linear plot

that should show the response curve of the pixels

well in fact I think I see

well, I don't have any swatches atm :)

Light meter?

Balls.

yes, I juggle :)

Did you shoot the pattern?

so, any flaws you can find in my reasoning regarding detecting any non linear transfer function?

Hrm.

I am unsure.

what troy_s just said about the "target audience" is very important, I think we're atm discussing points that will have to be put forward in our documentation

hehe :)

well, that would be selecting the test to support your argument :)

If they are baking some nonstandard curve in merely to present the data in a "nice" fashion, it makes the data useless if we can't reverse it.

i.e. if there is a linear relation between light and read out values, then we should get a linear relation between the sampled/mean values for linear exposure steps, no?

and if it isn't linear, then we should see the curve

that is like when I say, let's plot this data to see if it is linear

and you say, well, if it fits a curve, your plot is wrong

a linear image is deadly easy to see

(but a face would be easier perhaps)

A linear representation in data looks wholly "crunched" down

Very "dark" to use a low level guess.

very dark with only the highly exposed areas of the image shining through

taht's roughly what a "linear image" looks like, even though that doesn't make much sense

FergusL: Can you quickly invert an srgb curve on that image and post it?

let me do this

Bertl: I have looked at a few images in my time, and that is how I can roughly guess what it likely looks like in the physical scene.

from bottom to top

FergusL: Agree?

so why not figure how we can 'measure' it with what we(I) currently have available?

If you place a light source close

the measurement of X distance (near left edge perhaps) should be 1/4 the rgb values at 2x

(assuming a simple pinpoint tungsten fixture, that SHOULD hold roughly true. EG No reflector or diffuse source)

okay, so I get a small lamp which can be considered a point source

the rgb values, if linear, should hold according to 1/d^2

Does that make sense?

raking along the paper

bright to dark

should hold up in data according to the inverse square law

so if say, the distance from left edge is 2x

(you can probably cheat the source to get you more easily calculated values too.)

let me see what lamps I can find

(that is effectively a form of compression)

Well it isn't unheard of. Most commercial cameras offer a log mode.

sure, but the specification of a sensor for medical imaging should at least give a hint

I am no sensor guru on that front however.

I glanced at the HDR line skipping technique etc, and I would have expected more documentation on how to interpret the blasted data.

on this sensor, that is

(I'm tempted to just try with nuke)

FergusL: It does a window mode and a line skip mode.

you consider it's not working ?

I was just looking at the data sheet for the HDR modes.

thanks for your help !

Bertl: G'nite.

going to sleep too, btw

thanks, I'll be there hopefully "earlier" tomorrow

yeah, I'll do some tests with the lamp

morning!

How are you feeling? Still stressed?

hi, a bit less stressed today :)

Bertl: FergusL: If I remember correctly, what the sensor documentation *claims* it does is provide up to 3 regions with different slope.

This could bear checking, however.

Thanks, I can see it

just about to reply back ;)

I'm not sure if this is being used

the name carries the answer: piecewise LINEAR response mode ;)

yes

troy_s claimed it appeared to have a TRC applied.

"TRC"?

hmm, I dont really follow

maybe we can get the new article done today

But I can try to have it done by the end of this night

can you give me what you have so far and I put it into the article layout for now

Give me fifteen minutes and I'll have it ready for you

Sebastian, here's the link to the google doc: https://docs.google.com/document/d/1AtexpX8avHEXORvTlnYrDMPn69FQv-uwM62DHqtSXok/edit

creating floorpln animation of the FPGA now

random names google generates

its me ^^

Sebsatian, what else would you like me to add to the document?

well now we need to add the first captured image

Just read the article, looking FANTASTIC!

:)

lunchtime

back

slightly updated: https://www.apertus.org/axiom-alpha-opens-eyes

ok I have the basic layout finished

I'm back

https://www.apertus.org/axiom-alpha-first-images

next steps maybe

Sebastian, I just want to edit one sentence in the first paragraph. It'll be a minor change.

hi Bertl

Hi Bertl

http://vserver.13thfloor.at/Stuff/AXIOM/RAW/linear_2.svg

does this look linear to you ?

FergusL: yes, the curves are what we expect assuming a linear sensor

(linear is not always what we think it is)

yes

here is measured divided by expected

i.e. a horizontal line means constant factor

but hm...

like, physically rigorous

basically 1/squar-distance corrected by lambert

I'm still missing some points of the theory, even after troy_s explained it to me countless times

linear as in 1 + 1 = 2

it would clearly target an educated audience but could also sum up about the theory lying behind at the same tie

CC se6astian ^

is there anything else you think should be added

and add a picture of the prototype assembly which I have on my home pc

yes, and don't forget, it is not a precise measurement or calibration, it is just a quick test to check the data

from what I understood, we could get an IT8 chart thanks to troy_s

Good decision. And on that note, I have to go get some sleep so I can wake up on time for work tomorrow :(

Sasha_C: make that! thanks again and have a good night!

cinematography mailing-list, I've already talked about it, I think

I think cml rules include that mails are not to be published outside of the mailing list but basically there were some doubts expressed about the modular desing

that's a valid point I believe

absolutely, I'm still monitoring cml for more mails about Apertus, I'll make sure to report to the community

about the power lines we already had a discussion in Brussels when we were in the train back to the airport

Bertl will surely remember more of the technical details

I see

worst case scenario is a primitive power plug on every module, directly connected to a battery pack :)

http://www.afcinema.com/-2014-.html ever heard about the AFC ? it's basically BSC or ASC in France, they hold a show every year

OTOH, when we transport 5V over 5 modules, we will get like 3V at the end, if all those modules consume a certain amount of power

(unles we use super conductors :)

it's the same effect that makes looong unshielded cables weak ?

yup

http://en.wikipedia.org/wiki/Electrical_resistance

not just distance, it also depends on the power

http://en.wikipedia.org/wiki/Ohm%27s_law

this equals a 'resistance' of 10 Ohm

I'm more used to using Ohm's law as U = RI, because most of the time R is known

and they build a voltage divider

where the voltage at the consumer is Rconsumer/R

how are connectors treated in this matter ? like wires touching each other

they have a resistance, yes

they also cause strange effects because different metals are involved

mostly thermoelectric effects

http://en.wikipedia.org/wiki/Seebeck_effect#Seebeck_effect

somewhat

back to the topic, Bertl, your experiment would suggest that the images coming from the sensor are linear ? did you test pre or post debayer ?

everything on the raw data, so no debayering or similar

(4 channels, 1 red, 1 blue, 2 green)

do you have the code somewhere for that "SIMPLE" debayer algorithm ?

sure, but it wasn't used

am just curious of what it does, especially regarding tone curve correction

https://github.com/jdthomas/bayer2rgb/blob/master/bayer.c line 569

https://github.com/jdthomas/bayer2rgb/blob/master/bayer.c#L569 <- this works :)

well, AFAICT, it's a perfectly fine integer division with rounding

yes yes, it's just an old fear of mine

*gets

apparently, no, it's not processing in any way

any thoughts on the apertus° association accepting bitcoins for donations in the future?

I'm all for it

agreed

(I'm not saying we have an agressive business model)

hm, nevermind, that would be obvious if you didn't

it wouldn't hurt to write our own debayer

originally I wanted to use dcraw, because it is kind of famous in the community

but I didn't manage to get it working with real raw data (i.e. bayer mosaic data)

and the folks I asked told me, it would be best to write my own 'raw' importer for dcraw, to read the raw files :)

so it seems that dcraw cannot read 'raw' after all, it can only read the various proprietary formats the companies call 'raw'

what's the "format" you get ? is it a standard of some sort ?

the 'format' is the one I described on the dev mailing list

it's also rather trivial to write it for ImageMagick with Python

no headers, no additional metadata yet

i.e. just dump out all the sensor registers one after the other for later processing

yep, I hope it will become useful soon

linear formats would be cool be cool at some point too, dpx or exr

Bertl: I'm not sure I understand, how different would this dump be from the raw data you get ?

Blender should, if you mark the input colorspace as... hm.. what it it called ? "none" or something

ohhh, the registers, absolutely, that'd be metadata, right ?

and for analysis, just read the last 256 bytes and you know exactly what the sensor was configured for

np

it's like imagemagicks made particularly for our field

see you

Bertl: what was used for the export to .png ?

cool

I'm using ImageMagick for almost everything

I haven't uploaded it yet, but I will upload it soon

http://vserver.13thfloor.at/Stuff/AXIOM/cmv_snap.c

thanks

what are the mmap() for ?

though it's also writing to registers later

the sensor registers, the data collected from the sensor, etc

(so that it can be directly accessed)

this is the code to train the LVDS channels

like, otherwise, we're back to NTSC ?

I see

ding ding :)

if you can elaborate on the acronyms, maybe I can comment :)

open <foo> IO

Bertl: OCIO = Sony (Jeremy Selan's) OpenColorIO

Bertl: TRC = Tone Response Curve (aka Transfer Curve) (aka the wrong term Gamma)

okay, well, we probably can do that at some point when we have test charts

Bertl: I trust your math. Did you try the test we discussed via lamp?

Bertl: It is possible the tape was overexposed all to hell... but the image sure looked "normal", which suggests it wasn't linear. Could be subject matter of course.

troy_s: everything is possible regarding exposure

and yes, I did the discussed lamp test

http://vserver.13thfloor.at/Stuff/AXIOM/RAW/linear_2.svg

regarding generic scene, I can probably take photo or centerfold and capture that :)

correct

we will be able to do that once the case is finished though

I'm looking for an internship, actually

check with se6astian what might be possible and what not (wikipedia says, in france labor law requires a minimum payment)

well the company is in Belgium

if you just need a signed paper I guess that should be possible

that's the spirit!

I just need a signed paper, indeed, I'll look into the details

Uh. I'm going to Red then.

but if it does not mean additional work or expenses for him I guess it should be fine

possibly they will even let you create their entire next product on your own, I only heard that they managed to make the scarlet worse than the Red One :)

added prototype picture and oscars cava picture as link to the article: https://www.apertus.org/axiom-alpha-first-images

ah, cool!

yes but of the epic, not the red one

of course the scarlet changed so many times before it was released

besides the fact that it links to a 'Latest News' page which shows news from 3 weeks ago as 'latest' it looks really nice

about the prototype

a catastrophe!

hopefully! and not "failed"

currently its strictly seperated

people tend to mix up the two already so we have to label them quite cleverly

how about a logo/icon?

like for example an 'alpha' or the greek letter maybe?

might get us sued :)

I don't think that a letter can be registered or trademarked (yet)

also one can have an 'alpha' that doesn't look like the Sony one

what if we used the prototype assembly as a kind of "icon" (like on the left of: https://www.apertus.org/sites/default/files/axiom-alpha-progress-update-ad_0.jpg)

yeah, well, it is probably too delailed to work as icon

for the very basic functionality it has now

gotta go afk for a bit

dmj_nova: http://vserver.13thfloor.at/Stuff/AXIOM/utilization.rpt

Serdes?

Bertl: hmm...that's really not that much

haha

hehe, you don't say :)

*emailed

Reply-To: *email address removed*

you might want to add the reply-to header to the envelope

probably in the contact form, unless google groups strip that information

check if the google-groups settings force a reply to?

I just googled for the issue and it seems there is an option

I guess the last one would be the proper setting, but I have to test it

well, it is how the headers are formulated

where reply would reply to the sender

say that you are looking for a well payed job :)

as I said, let the user decide sounds correct to me

you can always add the group in the cc reply

Users decide where replies are sent - Replies to posted messages are decided on a per-member level.

lets give it a try

now just add the dev group to the cc, like the team one

hmm?

that is probably your mail client, but here it works fine

gmail webclient

checking...

but if it works for you now I will leave the setting and we will just see if the problem is gone for most people

great

afk<-

Bertl: Is there a public repo of the apertus mailing list?

I don't think so

Bertl: Is there a copy of your results from the low tech test?

give me you e-mail (pm) and I'll bounce it to you

se6astian: Sure.

I am not going to lie... that is an impressive test there Bertl. I don't quite understand the axes, but pretty impressive nonetheless.

i.e. the light is on the right side

in he curvy graph it is plotted as is, in the more horizontal graph one divides the other

I think the peaks at the left end are due to back level noise

well, not exactly, we 'forgot' the cosine law yesterday

i.e. the light hits the paper at an angle, which affects the intensity seen from the camera proporional to the cosine of the angle between light source and normal

think about it like this, if you hit a unit square from the top, it will be well illuminated

(Luminance strictly means the radiometrically correct value (Y), where luma is the TRC luminance (Y'))

Bertl: Gotcha.

Bertl: I usually am holding a diffuse meter when using the inverse square... not looking via a spot meter.

but for the given setup, it has to be accounted :)

so yes, I probably over exposed the color band image

no, but the DVD cover was without the IR/UV cutoff

Because there is no way in hell that a linear typical image will resolve correctly on an LDR display. (Granted, it is an LDR paper source and jacking the exposure would likely bring it into range)