Current Server Time: 01:13 (Central Europe)

#apertus IRC Channel Logs

2014/03/02

Timezone: UTC


23:11
danhanes
left the channel
23:20
sb0
left the channel
23:37
sb0
joined the channel
23:38
Bertl
troy_s: ping
23:49
sb0
left the channel
23:51
sb0
joined the channel
00:35
gwelkind
joined the channel
00:40
troy_s
Bertl: Greets Bertl you in?
00:40
troy_s
Bertl: Just did some profiling here at home to test the general RGBs... and it is entirely consistent with my estimations. (I also have a DE 1.3 here on a pretty lame ass flash shot)
00:40
troy_s
(Which is ... well a little better than we are getting with the default curves coming out of the CMV12000)
00:42
gwelkind
left the channel
00:56
gwelkind
joined the channel
01:13
Bertl
okay?
01:14
Bertl
guess I I'm off to bed ... pretty tired, seems I almost fell asleep
01:14
Bertl
let's postpone this a little ...
01:30
gwelkind
left the channel
01:35
gwelkind
joined the channel
01:56
gwelkind
left the channel
02:42
troy_s
Bertl: Ping me when you awake Herb.
03:16
gwelkind
joined the channel
04:31
sb0
left the channel
04:41
gwelkind
left the channel
06:48
jerknextdoor
left the channel
08:29
jucar
left the channel
08:50
jucar
joined the channel
09:01
jucar1
joined the channel
09:04
jucar
left the channel
09:44
jucar1
left the channel
11:44
intracube
left the channel
12:57
Bertl
morning folks!
12:57
Bertl
troy_s: ping
13:12
ThatCantBe
left the channel
14:09
se6astian
joined the channel
14:09
se6astian
good afternoon
14:12
Bertl
afternoon!
16:00
jucar
joined the channel
16:24
ThatCantBe
joined the channel
16:47
troy_s
Bertl: Greets!
16:47
troy_s
(wow our time zones are apart)
16:49
Bertl
at the moment, yes
16:50
troy_s
How goes the fight Bertl
16:52
Bertl
no fight today :)
16:55
troy_s
Bertl: Rest is for the weak!
16:56
troy_s
Bertl: Did you manage to tweak the offsets to get the blasted sensor to be more continuous?
16:57
Bertl
as I wrote, it looks fine to me if you are not at the lower or upper limits where clipping happens
16:58
Bertl
I think the main problem is that it doesn't seem to be easy/simple to adjust it in such a way that the entire 12bit range is covered
16:59
troy_s
Bertl: The upper limits? Uh that GS strip is only a couple of stops.
16:59
troy_s
Bertl: It is well out of whack.
16:59
troy_s
Bertl: As is evidenced by the error ratios.
16:59
Bertl
the range is 12bit, but with the default settings, the 'linear' range only covers like 10-11bit
17:00
troy_s
Bertl: Those errors shouldn't be nearly that high.
17:00
troy_s
Bertl: Even closer to middle ranges is whack. I am not sure why.
17:00
Bertl
i.e. the channels converge around 100-150 and they top around 2500-3000
17:01
Bertl
each channel is a little different, each column also because of the FPN
17:01
troy_s
Bertl: If that is "the best it can do" with rainbowing on the GS strips at 70% RGB, it is done.
17:01
troy_s
Bertl: I cannot think for a moment a single imager will find it tolerable.
17:02
Bertl
as long as you stay in a middle ray, let's say, 500-2500 the channels look fine, linear and consistent
17:02
Bertl
s/ray/range/
17:02
troy_s
Bertl: This is a cinema camera. To ask shooters to stay below 70RGB is uh...
17:02
troy_s
Bertl: And I am not at all certain it is behaving correctly.
17:03
Bertl
I'm not convinced yet that this sensor is the best choice, it also isn't the only one we are considering
17:03
troy_s
Bertl: Again, why on earth is the red photosite at equal to green in D60?!
17:03
Bertl
but the actual 'range' doesn't matter that much
17:03
troy_s
Bertl: So you do not think there is the chance of a setting or configuration issue?
17:04
troy_s
Bertl: Explain?
17:04
Bertl
everything is possible, we do not have much support from cmosis
17:04
troy_s
Bertl: Yes. Not surprising.
17:04
Bertl
but for now, I'm going to simplify the calibration part, or at least try to :)
17:05
troy_s
I really wonder about that red response. I mean I completely can see the double hum non narrow response of the filter.
17:05
troy_s
Bertl: You will likely fail. Not easy to "remove" colors.
17:05
Bertl
not planning to remove anything :)
17:05
troy_s
Bertl: You can try a colorchecker.
17:06
troy_s
Bertl: Just going to hide the warts?
17:06
Bertl
but as first step, I'll eliminate the illumination from the equation
17:06
troy_s
My larger concern is that the sensor cannot deliver a matrix with a smaller error than about 9, no matter how it has tried.
17:06
Bertl
doesn't need to be a matrix
17:06
troy_s
I know this
17:06
Bertl
we have many options there
17:07
troy_s
But that is symptomatic.
17:07
troy_s
Even junky attempts should deliver a matrix.
17:07
Bertl
I blame the calibration process
17:07
troy_s
Example: I did a _quickie_ flash shot and got de 1.4 (de2k)
17:07
troy_s
How so?
17:07
troy_s
You can look at those strips Bertl
17:08
troy_s
achromatic strips with rainbow firing
17:08
troy_s
Sensor is behaving exceptionally poorly.
17:08
Bertl
because it doesn't account for the to-be-corrected effects currently plaquing the calibration attempts
17:08
troy_s
What is going to be corrected exactly?
17:09
troy_s
Is that GS strip going to be homogenous?
17:09
Bertl
first, we have to make sure to only use the 'linear' range (i.e. not the areas where the clipping happens)
17:10
troy_s
So tell me why every other sensor ever profiled has never had these constraints?
17:10
Bertl
it doesn't matter what the actual range there is, as long as it is consistent
17:10
Bertl
which sensor have you calibrated yet?
17:10
Bertl
I mean, starting with real raw data
17:10
troy_s
Normally you can push white achromatic patches to 94+ RGB
17:10
Bertl
not the preprocessed data called 'raw'
17:10
troy_s
You can process any DSLR with real raw data.
17:10
Bertl
show me
17:11
troy_s
They all start raw Bertl. dcraw peels off the bayer
17:11
troy_s
using AHD or VNG.
17:11
Bertl
those are already pre processed
17:11
troy_s
So again
17:11
Bertl
i.e. not the data acquired from the sensor
17:11
troy_s
if you are suggesting pre-processing
17:11
troy_s
What precisely do you propose to pre-process that signal?
17:11
troy_s
You intend to clamp the sensor range?
17:12
troy_s
(asking so I can understand)
17:12
Bertl
first, noise correction (we are already able to do that, it's just not properly calibrated)
17:12
troy_s
(But again, from what I see, if the raw output is symptomatic of the full latitude, you would have to very much constrict that pipe.)
17:12
Bertl
then range clipping and stretching (can be 11 or 10 bit)
17:12
troy_s
Ok
17:12
troy_s
I understand then.
17:12
Bertl
maybe we can figure out how to do that in the sensor properly
17:13
troy_s
Yes.
17:13
Bertl
then maybe (we have to check that) linearity corrections
17:13
troy_s
Yes.
17:13
Bertl
if there is some kind of subtle non linearity
17:13
troy_s
Is there a place for linearity corrections?
17:13
Bertl
we can do that with a LUT
17:13
troy_s
It seems there might be a chance from the curves I looked at.
17:13
Bertl
(per channel or one in general, we'll see)
17:14
troy_s
The curves fan out quite smoothly
17:14
Bertl
after that, it's the 'normal' matrix ICC
17:14
troy_s
which probably exacerbates the edges badly
17:14
Bertl
i.e. the data you get from a DLSR as raw
17:14
troy_s
So... is there a way to offset the individual sensels?
17:14
Bertl
not in the sensor, but we already do that in the FPN corrections
17:15
Bertl
we can also do that in a separate LUT if we want
17:15
troy_s
(was thinking it might be an interesting thing if it can twiddle the second green sensor for HDR - undergain it ex)
17:15
troy_s
So perhaps we should peel apart the gamma response tables and try one of them?
17:16
troy_s
(Namely to drag that red way the hell down)
17:17
Bertl
as cmosis obviously isn't doing their homework, we have to do that if we want to use the sensor
17:17
troy_s
Bertl: Possible to scale the channel down and offset?
17:17
troy_s
It would probably greatly crunch that irregularity
17:17
Bertl
with a 12->16 LUT per channel you can go crazy
17:18
Bertl
and that is easy to do in the FPGA
17:18
troy_s
Ok. Well I guess will need to think on it.
17:18
troy_s
Muck with this data and see if there is a way to massage some reasonable values.
17:19
Bertl
so in general, I'd stay away from saturation and underexposure
17:19
Bertl
mostly because both ends will give strange (non linear) results
17:19
Bertl
mostly because one channel saturates before the other
17:19
troy_s
and over
17:20
troy_s
only middle grey zone seems close
17:20
Bertl
yes, that's what I mean
17:21
troy_s
Bertl: Will be interesting to see how you work at it.
17:21
troy_s
Bertl: One would think the sensor would allow for channel swizzing at some damn level.
17:22
Bertl
have to try a few things, but I have a number of ideas
17:22
troy_s
Okie. Keep me in the loop.
17:22
Bertl
will do
17:28
sb0
joined the channel
17:33
troy_s
Bertl: I don't think many commercial sensors do clips as far as I have seen. Can deduce it from the highlight reconstruction etc.
17:33
troy_s
(the techniques to balance the areas that the sensels are full in a single channel)
17:34
troy_s
Bertl: We could do a LUT.
17:34
troy_s
Bertl: If you can install it. See if it makes a difference.
17:35
troy_s
(push the values to a threshold - not fully "corrected" but rather just enough to achieve linearity in the channels)
17:41
Bertl
I don't think you would know (if the sensor does or not)
17:42
Bertl
if 'black' is mapped to 0 and 'white' is mapped to max, how to tell if it was clipped/scaled or not?
17:44
troy_s
I would expect to see some degredation in the ranges.
17:45
Bertl
okay, how many bit does the DSLR raw give you?
17:45
troy_s
(Likely as irregular gradation)
17:45
troy_s
Bertl: Dumps same as what you dump.
17:45
Bertl
so 12bit, yes?
17:45
troy_s
12 bit in 16 wrappers normally
17:46
troy_s
(I am sure there are exceptions, but ballpark guess)
17:46
Bertl
okay, so let's assume that we can use 11 of the 12bit, half the range
17:46
troy_s
I am very interested to see where the linearity skew happens worst
17:46
Bertl
we still have information for interpolating brigtness across the sensels
17:46
troy_s
How to calculate that is tricky.
17:47
troy_s
yes
17:47
troy_s
true
17:47
Bertl
which will give us roughly one bit
17:47
troy_s
Can you truncate the red?
17:47
troy_s
Actually, I would be interested to see what happens if you only clip the red bit.
17:48
Bertl
sure, will take a little to add (code wise) and to test though
17:48
troy_s
I am sure I can give you an easy "shoot until this value is xxx"
17:48
troy_s
in fact, I would very much love to see you do a series of tests with the older chart and plot them
17:49
troy_s
With that Bertl magic
17:49
troy_s
(like the data charts you made earlier)
17:49
troy_s
it might help to diagnose the most surgical adjustments
17:49
Bertl
I'm not sure I want to spend much more time on the chart at this stage
17:50
Bertl
for one thing, as I wrote, I want to eliminate the illumination problem from the equation
17:50
troy_s
I might be WAY off course, but my gut is telling me that the red is the biggest culprit at the moment. Just scaling back its sensitivity somehow.
17:50
troy_s
The evenness?
17:51
troy_s
Or the skewed linearity portions?
17:51
Bertl
all of it, eveness, color temperature, etc
17:51
troy_s
How to accomplish that though?
17:51
Bertl
I think I can do perfectly fine calibration just with physics
17:52
troy_s
How?
17:53
troy_s
Bertl: Non narrow filters.
17:53
troy_s
Bertl: Nigh on impossible, but you are welcome to give it a kick at the can. LOL.
17:53
Bertl
first, we need a light source of known wavelength, more precisely several of them
17:54
troy_s
The issue is the sensor is not firing in a homgeneous fashion.
17:54
troy_s
All of the massaging is already happening.
17:54
troy_s
Argyll (or any other calibration system) bends and warps like mad.
17:55
Bertl
yes, but that's exactly what we do not want
17:55
troy_s
Color isn't all physical nor all psychological - psychophysical.
17:55
troy_s
It is.
17:55
troy_s
Because you are dealing with the standard observer. You cannot escape that.
17:55
Bertl
agreed, but at the sensor level, it's just physics
17:55
Bertl
i.e. no observer but the sensor itself
17:56
troy_s
Yes, but that portion of physics isn't all pure physics.
17:56
Bertl
an that's exactly what we want to measure first
17:56
troy_s
You cannot generate color purely in the physics domain.
17:56
Bertl
not talking color, just talking wavelength :)
17:56
troy_s
Ugh.
17:57
troy_s
All of this stuff has been rehashed over and over for many years now.
17:57
Bertl
note: I'm not trying to calibrate the color
17:57
troy_s
Seems like reinventing the wheel to work around and back to the same issue.
17:57
troy_s
Sure.
17:58
Bertl
that is something we can do once we know the sensor range/behaviour with the mechanisms we already have
17:58
Bertl
like argyll and friends
17:58
troy_s
That is sort of what I was suggesting - tweak the knobs based on data.
17:59
troy_s
If you chart how that sensor responds, you can dial it in with a decent dataset.
17:59
troy_s
Should be obvious where the issue appears.
17:59
troy_s
I mean it doesn't take Mr
17:59
troy_s
Mr. Fairchild to look at the results and see roughly where badness is ensuing.
18:00
troy_s
Heck, you can probably see the error shooting just a grey card.
18:00
troy_s
(at various exposures)
18:01
troy_s
And seeing how the curves from the channels unfold (likely R then G then B)
18:04
intracube_
joined the channel
18:05
troy_s
greets intracube_
18:07
intracube_
hey troy_s, hows things?
18:07
intracube_
changed nick to: intracube
18:07
troy_s
Holding.
18:08
troy_s
Bertl: There is _no_ disclosed gain controls on the channels at all eh?
18:08
troy_s
(Other than on a global level)
18:08
Bertl
there is gain, but not per channel
18:09
Bertl
you can check yourself, the datasheet is on github IIRC
18:09
troy_s
Bertl: There must be undocumented interfaces to that thing eh?
18:10
Bertl
undocumented = we don't know about
18:10
troy_s
Bertl: Abd out of curiosity, has undergaining been tried?
18:10
troy_s
(I know, was asking for specultation if you think that would be common)
18:11
Bertl
I think there is a lot of hidden knobs and switches
18:12
Bertl
you can see that by comparing different datasheet version
18:12
Bertl
*versions
18:12
Bertl
for example there is no mentioning of the test pattern in older data sheets
18:12
Bertl
and suddenly *bam* there is a test pattern :)
18:18
troy_s
Hrm. I wonder what they would do if you voiced a concern over the response for color.
18:18
troy_s
They might listen to you Bertl.
18:19
Bertl
we will try to improve the connection, but as I said, the sensor itself is just one of many
18:20
Bertl
of course, it would be a shame to just have wasted a lot of time on figuring out sensor details, but that's life
18:21
Bertl
anyway, nothing decided yet, and we'll know more soon
18:21
intracube
troy_s: holding?
18:46
gwelkind
joined the channel
18:57
gwelkind
left the channel
19:29
danieel
Bertl: how precisely you can set exposure time?
19:29
danieel
i was thinking if you can make a graph how values rise depending on time (should be linear)
19:31
Bertl
roughly in 10us steps
19:38
gwelkind
joined the channel
19:44
Bertl
yes and no (regarding linearity) it gives a good approximation in the typical range (i.e. 1ms - 1000ms) outside the exposure will become nonlinear because auf noise and decay
19:58
gwelkind
left the channel
20:19
danieel
i would wonder if you get a 4K BMCC and tap the sensor's spi if you discover access to undocumented registers
20:20
gwelkind
joined the channel
20:20
Bertl
there are 128 registers, so it's not a problem to 'discover' undocumented ones
20:20
danieel
(i did a 5.5 fhd lcd interface, and it did not worked according to datasheet, so I borrowed a competitive interface and tapped it... no surprise there - different data were used! )
20:20
Bertl
I don't think that they actively hide registers though
20:21
danieel
not hide, just enable bits which are not documented
20:21
danieel
or do you have the map of all 128x16 ?
20:21
Bertl
no, we do not know what each bit does
20:22
danieel
i would assume writing 1's to undocumented shall make one suspecting
20:22
Bertl
but logging SPI traffic won't help there, just narrow down the search space
20:38
intracube
left the channel
21:07
se6astian
time for bed
21:07
se6astian
good night!
21:07
se6astian
left the channel