Current Server Time: 06:26 (Central Europe)

#apertus IRC Channel Logs

2016/01/21

Timezone: UTC


00:11
darthrak_
joined the channel
00:25
darthrak_
left the channel
01:26
darthrak_
joined the channel
01:30
darthrak_
left the channel
01:51
slikdigit
joined the channel
01:52
slikdigit
left the channel
01:52
slikdigit
joined the channel
03:27
darthrak_
joined the channel
03:32
darthrak_
left the channel
05:35
hozer
left the channel
07:28
darthrak_
joined the channel
07:33
darthrak_
left the channel
07:42
jucar
joined the channel
07:48
se6astian|away
changed nick to: se6astian
07:49
se6astian
good morning
08:07
jucar
left the channel
08:39
darthrak_
joined the channel
08:46
intracube
so that bottom right discolouration here: http://i.imgur.com/j23B4rt.png
08:47
intracube
is visible in the nonlinearity images after a contrast stretch: http://i.imgur.com/S8g32Eo.jpg
08:47
intracube
only in the green channel, though
08:48
intracube
a diff of green with either red or blue channel + further contrast stretch: http://i.imgur.com/3Qvbd1x.jpg
08:48
intracube
^ alexML
08:54
alexML
yeah, I need a better clip frame (I'm experimenting with something that accounts for nonuniform illumination of the wall, but no meaningful results yet)
08:55
intracube
isn't that green a sensor issue rather than scene illumination?
08:56
alexML
I'm also trying to recover the response curve with other methods; unfortunately, with existing programs like mkhdr or pfshdrcalibrate, I've got only bogus results
08:56
alexML
so I'm reinventing the wheel :P
08:56
intracube
heh
08:56
alexML
it's not a sensor issue, it's just the calibration frame
08:56
intracube
oh ok
08:57
alexML
(I'm working on existing samples, remember)
08:57
alexML
(I don't have a beta yet)
08:57
intracube
^ soon though :)
08:58
alexML
I also thought at something that would flip the exposure back and forth near the clipping point - if I let that run for a while, I guess it should give pretty exact per-pixel curves
08:59
alexML
(there is plenty of noise in that region)
08:59
intracube
you mean the granularity in the highlights (particularly saturated areas)?: http://i.imgur.com/qVfdkGs.jpg
09:00
alexML
yeah
09:00
alexML
that throws off the calibration, so if I want to get per-pixel curves from the existing dataset, there's a ton of noise in the data
09:01
alexML
but if you repeat that experiment 16 times, that noise would be reduced by 2 stops
09:01
alexML
if you repeat it 64 times => noise reduced by 3 stops
09:01
alexML
and so on
09:01
intracube
so in theory, it could be reduced a fair bit
09:02
alexML
well, I want to get the exact (unbiased) curves first
09:02
alexML
so the noise would have zero mean (that is, so it won't introduce color casts)
09:02
alexML
once you have good average colors, you can just denoise those highlights a bit
09:03
alexML
and then there's the repeatability issue
09:03
alexML
that needs to be checked, how it varies with exposure, temperature, whatever
09:04
intracube
and this profiling would have to be done for each individual sensor
09:04
alexML
and if it's just a simple offset, maybe it can be detected on the fly
09:04
alexML
yeah
09:04
intracube
and also possible changes as the sensors age
09:04
alexML
same for dark frames
09:04
Bertl_oO
changed nick to: Bertl
09:04
alexML
yeah, I want to have these calibration routines in the camera
09:05
intracube
is this extreme non-linearity in the highlights typical for cmos?
09:05
intracube
I thought it was generally linear
09:05
Bertl
it isn't a problem of the photo sites, it is a problem of the ADC and gain
09:06
alexML
on Canon, it seems to clip quite harshly
09:06
intracube
Bertl: ah
09:06
alexML
on 5D3, I turned down all the gains I could identify, so it no longer clips to white
09:06
Bertl
or to be precise there are several effects going on
09:06
alexML
and got some fixed-pattern noise in highlights
09:06
alexML
but after correcting that, there was no useful data to recover, like on the CMV
09:07
Bertl
first, the CMOS charges a capacitor to a certain level
09:07
Bertl
then the photo diode discharges it when light is received
09:07
Bertl
this process is not really linear per se
09:08
intracube
ah ok
09:08
Bertl
also, it has a roll-off when the capacitor reaches zero voltage
09:08
Bertl
because the diode adds a certain offset there as well
09:08
Bertl
assuming that the actual sensel is operated in the linear range
09:09
Bertl
(or to be precise, in the range where it acts as expected)
09:09
Bertl
you'll get a nice and proper voltage to digitize
09:09
Bertl
this is where the second stage brings in non linearities
09:10
Bertl
i.e. you apply analog gain, which has nonlinearities at the rails
09:10
Bertl
and you also run the ADCs, which again have (usually quite small) nonlinearities at the limits
09:11
Bertl
given that everything is optimal, you will only see a proper linear range
09:11
Bertl
the CMV is a little odd in this regard, as it allows for a number of analog gains, which are well below the ADC range
09:11
Bertl
this
09:11
Bertl
per se is not a bad thing
09:12
Bertl
it just means that the full 12bit range will not map to a linear range in all gain settings
09:16
intracube
interesting, thanks for the explanation :)
09:17
Bertl
you're welcome!
09:23
darthrak_
left the channel
13:31
darthrak_
joined the channel
13:33
darthrak_
left the channel
13:34
darthrak_
joined the channel
13:35
darthrak_
left the channel
13:39
ar
left the channel
14:33
jucar
joined the channel
15:46
jucar
left the channel
16:17
jucar
joined the channel
16:22
se6astian
changed nick to: se6astian|away
16:33
jucar
left the channel
16:42
Bertl
off for now ... bbl
16:42
Bertl
changed nick to: Bertl_oO
17:22
se6astian|away
changed nick to: se6astian
18:07
jucar
joined the channel
19:49
se6astian
anyone currently here speak spanish?
19:51
Bertl_oO
Lo sentimos, no hablo español! :)
19:52
Bertl_oO
or maybe better? ¡Lo siento, no hablo español!
19:55
intracube
dos cervezas, por favor
19:56
Bertl_oO
LOL
20:08
troy_s
Bertl_oO: Yo
20:09
troy_s
Bertl_oO: intracube noticed that there's some odd glow in the lower right corner. Any clue as to what this might be? Light leak / splash off of an LED? Heat? Electronics?
20:09
davidak
joined the channel
20:09
troy_s
Bertl_oO: http://i.imgur.com/S8g32Eo.jpg
20:10
jucar
left the channel
20:20
alexML
troy_s: that corner was not completely overexposed, but the rest of the image was (obvious if check the other images in the set)
20:21
troy_s
alexML: I don't mind. I'm more interested in what causes it.
20:21
alexML
once you go past some overexposure level, the sensor output starts to decrease
20:21
troy_s
I'm more interested in the fact that it is there at all.
20:21
troy_s
And why.
20:21
troy_s
Could be an LED leak like was had in the early Alpha.
20:22
troy_s
Or heat, or magnetic magic fubar or who knows.
20:22
alexML
check the nearby samples (render them in the same way and play the movie)
20:22
alexML
you will no longer blame cosmic radiation after you do that :P
20:24
intracube
http://i.imgur.com/F0oLpFD.png
20:25
alexML
check from 30 to 90ms for example
20:25
alexML
or, even better from 1ms
20:26
troy_s
alexML: I don't blame anything. I'm just leaving everything open. Those shots are hooped.
20:26
troy_s
It is upper right too.
20:26
troy_s
So could very well be an LED leak off of one of the boards.
20:26
troy_s
Or any other thing.
20:27
alexML
it's not visible in dark frames, so I doubt it's a LED leak
20:27
troy_s
Dunno.
20:27
troy_s
That's my point.
20:27
troy_s
Bertl will have good insight.
20:27
alexML
but please - see the entire sequence before judging
20:27
troy_s
How come?
20:28
troy_s
Does it flicker?
20:28
intracube
I thought the 'nonlinearity-gain1x-offset' images were dark frame?
20:28
alexML
no, they are not
20:28
troy_s
Anyways... off to pick up damn Cintiq number THREE for the love of god.
20:28
intracube
ooh
20:28
alexML
they are a blank out of focus wall
20:29
troy_s
Dark frames are portal capped images then?
20:29
troy_s
(I take it)
20:29
intracube
ok, that does change things slightly. but this issue only seems to be on the green channel
20:29
intracube
nothing on red or blue
20:29
troy_s
That's odd.
20:29
troy_s
If that's true then it is a magical source.
20:29
intracube
will have a look at the darkframes
20:30
troy_s
To only impact one data channel would be pretty damn magical.
20:30
alexML
they all behave the same electrically, but green is usually receiving more light (hint: white balance)
20:30
troy_s
White balance?
20:30
troy_s
This is off of the raw data.
20:30
troy_s
So how would white balance even factor in pre-transform?
20:30
alexML
yes, if you render raw data without white balance, what color cast do you get?
20:30
troy_s
You don't get any colour cast
20:30
troy_s
There's no colour.
20:31
troy_s
You might as well be dumping a musical track to an sRGB screen.
20:32
intracube
just to confirm; I'm not seeing anything on a dark frame image
20:32
intracube
(darkframe-100ms-gainx4-10.raw12)
20:33
troy_s
intracube: Then it must be across all data channels if it is a leak
20:34
intracube
yep, unless it is a fairly pure green light from somewhere
20:34
intracube
but you'd expect to see something on the other channels, I guess
20:34
alexML
again, it's a blank wall
20:35
alexML
in the middle of the image, r=g=b (roughly, on raw channels, no colorimetry)
20:35
alexML
in that corner (not yet clipped), green starts to clip, red and blue are lower
20:35
troy_s
alexML: So it's a leak?
20:35
alexML
what color do you get when you render with wb multiplers 1,1,1
20:35
alexML
no
20:36
alexML
it's not a leak
20:36
Bertl_oO
well, I was wondering about the corner as well (on the image intracube "corrected")
20:36
troy_s
Bertl_oO: More like amplified.
20:36
alexML
that's how a white paper appears when it's close to overexposure, and uniwb
20:36
troy_s
It's got to be a leak.
20:36
alexML
try it on Canon, Nikon, whatever you want
20:36
troy_s
If it isn't on the dark frame, and it is on the other non-darks
20:36
troy_s
It is a leak of some sort.
20:37
alexML
it would look very similar
20:37
Bertl_oO
but then I saw the other image somebody visiting sebastian made from the setup
20:37
troy_s
Bertl_oO: And?
20:37
troy_s
Bertl_oO: Difference?
20:37
Bertl_oO
and in this image, there is a bright light source on the left, and a table lamp on the right, just illuminating that corner
20:38
troy_s
Flare leak.
20:38
troy_s
Or just a literal light
20:38
troy_s
I didn't think twice when I first saw it and just assumed it was a glow from a lamp.
20:38
troy_s
It's not shaped enough to be a flare, but it is also in the top right corner too.
20:38
troy_s
Slightly.
20:39
Bertl_oO
so, I would double check with sebastian, that the setup is not flawed here before discussing this effect further
20:39
troy_s
It has falloff like a lamp.
20:39
troy_s
I agree.
20:39
troy_s
I mean as far as I am concerned anyways, it's all moot until the next iteration comes in with proper latitude.
20:39
troy_s
Not much to be done until then.
20:40
troy_s
Bertl_oO: Would you be able to get a camera set up such that it could be remotely controlled?
20:40
troy_s
Just to tweak the PWLC etc?
20:40
troy_s
(God. Must go get Cintiq Blech.)
20:40
intracube
troy_s: I forgot to mention... I blurred the images to get rid of HF FPN too
20:40
intracube
so don't read too much into the shape of it
20:42
alexML
again - take an image of a blank wall, with any digital camera, overexpose it, leave just one corner a little unclipped, and render the raw file in the same way
20:42
alexML
upload the result here
20:43
Bertl_oO
troy_s: I'd say we already have part of that in place
20:44
Bertl_oO
with a bunch of remote controlled switches to change the lighting, it should probably work
20:44
Bertl_oO
the Beta itself is already remotely accessible
20:46
troy_s
Bertl_oO: should put a rotate data channels into the format so folks done erroneously call them colours
20:47
troy_s
Bertl_oO: with a calendar we could have a nice lab
20:48
troy_s
Very keen on PWLC
20:49
troy_s
Would be a terrific way to test calculated knee points
20:49
alexML
troy_s: we are looking at low-level sensor output, before the matrix gets applied
20:50
troy_s
(Reverse a log step per bit per stop optimization to knees)
20:50
troy_s
I know this
20:51
alexML
I doubt that :P
20:51
troy_s
Notice how I never say colour when looking at data?
20:51
alexML
yeah, how should we call it?
20:51
troy_s
Data
20:51
intracube
red: http://i.imgur.com/lFH9rOq.png green: http://i.imgur.com/qAhCrX1.png blue: http://i.imgur.com/aun2ttc.png
20:52
intracube
troy_s: this is the data off the r, g, b sensels though. so you can legitimately call them that
20:52
intracube
those images contrast all contrast stretched the same amount
20:53
niemand
joined the channel
20:53
troy_s
Which has no reference so you may equally call them radish, turnip, potato
20:54
intracube
no reference? the sensels have colour filters, no? :)
20:54
troy_s
Yes but no reference
20:55
troy_s
If you rotate those channels the data is equivalent
20:55
alexML
troy_s: to be able to rotate the channels, you first need linear response (not clipped stuff)
20:55
alexML
we don't have that here, so we can't rotate yet
20:56
alexML
we are just looking at the raw data before rotating
20:56
alexML
that's it
20:58
alexML
intracube: you have applied white balance to the data
20:58
alexML
can you show the raw (unmodified) one?
20:58
alexML
maybe r,g1,g2,b
20:59
troy_s
Say what?
20:59
intracube
alexML: this was with Bertl's script. I didn't think it did any white balance
21:00
intracube
and just debayered
21:00
troy_s
If you rotate those channels and profile it will deliver guess what
21:02
intracube
convert \( -size 4096x3072 -depth 12 gray:input_raw_image.raw12 \) \( -clone 0 -crop -1-1 \) \( -clone 0 -crop -1+0 \) \( -clone 0 -crop +0-1 \) -sample 2048x1536 \( -clone 2,3 -average \) -delete 2,3 -swap 0,1 +swap -combine output.png
21:02
intracube
^ alexML
21:03
intracube
doesn't fully understand how it works :)
21:04
alexML
intracube: nevermind, can you do the same thing for the other images in the set?
21:05
alexML
let's say from 1 to 100ms
21:05
troy_s
Anyways had my fun in the asylum.
21:05
alexML
in particular, compare the 75ms one with the 88ms one (since I didn't download past that)
21:05
alexML
compare all 3 (or 4) channels
21:05
intracube
alexML: I'll have a go later, yep
21:06
intracube
need to get some food in a bit
21:10
troy_s
Did you rotate the channels and characterize yet? ;)
21:11
alexML
troy_s: this is an overexposed image, what do you expect?!
21:12
troy_s
I expect the exact same results actually.
21:12
troy_s
Rotated primaries ;)
21:12
troy_s
12:55 <alexML> troy_s: to be able to rotate the channels, you first need linear response (not clipped stuff)
21:12
troy_s
That is crap
21:12
alexML
proof please
21:13
alexML
(as in, apply the matrix on a nonlinear image and show that you get correct results)
21:13
troy_s
Not the damn matrix
21:13
troy_s
To characterize.
21:13
troy_s
Hence why I keep repeating they ain't colors.
21:14
troy_s
Anyways... Fun time is over. This is depressing.
21:34
darthrak_
joined the channel
21:44
Bertl_oO
intracube: do you want to know how it works? :)
21:47
Bertl_oO
if so, please remind me when I'm back from my nap ...
21:47
Bertl_oO
changed nick to: Bertl_zZ
22:00
darthrak_
left the channel
22:01
se6astian
off to bed
22:05
davidak
left the channel
22:06
se6astian
changed nick to: se6astian|away
22:19
davidak
joined the channel
22:27
intracube
left the channel
22:32
davidak
left the channel
22:42
niemand
left the channel
23:00
darthrak_
joined the channel
23:05
darthrak_
left the channel
23:37
SL2
joined the channel