Current Server Time: 17:03 (Central Europe)

#apertus IRC Channel Logs

2016/01/05

Timezone: UTC


23:07
intracube
left the channel
23:29
intracube
joined the channel
23:41
aombk2
joined the channel
23:50
pozitrono
joined the channel
23:52
pozitron
left the channel
00:18
aombk2
left the channel
00:19
davidak
left the channel
00:24
slikdigit
left the channel
00:36
troy_s
left the channel
02:58
troy_s
joined the channel
04:12
jucar
joined the channel
04:23
jucar
left the channel
04:28
jucar
joined the channel
04:39
jucar
left the channel
04:52
jucar
joined the channel
05:11
jucar
left the channel
06:45
Bertl_zZ
changed nick to: Bertl
06:45
Bertl
morning folks!
08:10
pozitrono
left the channel
08:16
troy_s
left the channel
09:23
se6astian|away
changed nick to: se6astian
10:01
arpu
left the channel
11:04
se6astian
changed nick to: se6astian|away
11:08
Bertl
off for now ... bbl
11:08
Bertl
changed nick to: Bertl_oO
11:48
jucar
joined the channel
12:04
jucar
left the channel
13:39
se6astian|away
changed nick to: se6astian
14:13
troy_s
joined the channel
15:54
arpu
joined the channel
16:56
slikdigit
joined the channel
16:57
arpu
left the channel
17:00
slikdigit
left the channel
17:00
slikdigit
joined the channel
17:19
jucar
joined the channel
17:35
slikdigit
left the channel
18:02
arpu
joined the channel
18:06
slikdigit
joined the channel
18:07
slikdigit
left the channel
18:07
slikdigit
joined the channel
18:28
davidak
joined the channel
18:36
wakalixes
joined the channel
18:38
Bertl_oO
welcome wakalixes!
18:39
wakalixes
hi Bertl
18:41
wakalixes
the little demo of the axiom beta at the 32C3 was really nice
18:41
Bertl_oO
glad you liked it
18:42
wakalixes
playing around with some config registers of the sensor found in the datasheet, that sound interesting
18:43
wakalixes
parasew said you are going to deliver the first 20 prototypes these days?
18:45
Bertl_oO
we are working on it right now, yes
18:53
Bertl_oO
off for a nap ... bbl
18:53
Bertl_oO
changed nick to: Bertl_zZ
19:02
se6astian
the prototypes are being hand assembled, lots of electronic components to be placed
19:02
se6astian
so it takes a bit of time
19:02
se6astian
but for 20 units automated assembly is just too expensive per unit
19:03
se6astian
but if those 20 units work reliably and we find no bugs with the electronics we will of course do automated assembly for a larger first real production batch
19:03
se6astian
depending on how many units crowdfunding backers/voucher holders want from this hardware generation maybe around 100 units already
19:03
se6astian
interested in getting one? :)
19:37
philippej_
joined the channel
19:42
wakalixes
indeed, I would be interested in getting one
19:42
wakalixes
where do I need to sign up :-)
19:43
wakalixes
what are the goals which need to be achieved for testing, hardware and software
19:47
se6astian
there is a preorder notification list to sign up for here: https://www.apertus.org/axiom-beta
19:48
se6astian
check out the phased payments details below as well please
19:48
philippej_
and an explanation of phased paiements here : https://apertus.org/axiom-beta-phased-payments
19:48
se6astian
:)
19:48
philippej_
hehe :-)
19:48
wakalixes
hehe
19:49
se6astian
what do you mean with "goals which need to be achieved for testing, hardware and software" in particular?
19:49
se6astian
what things need to be tested?
19:49
wakalixes
what are you going to test with the firts prototypes
19:50
se6astian
lots of things
19:50
wakalixes
is it hardware issues, thermal managment
19:50
se6astian
exactly
19:50
se6astian
we know that the hardware works
19:50
se6astian
but we did not test every component in every possible configuration
19:50
se6astian
so that we will do
19:51
se6astian
and then we will need to create lots of software
19:51
se6astian
for color calibration
19:51
se6astian
fixed pattern noise calibration and real time compensation
19:51
wakalixes
and hdr
19:51
se6astian
more nosie related compensations
19:51
se6astian
dead pixel compensation
19:51
se6astian
hdr tuning as well yes :)
19:51
se6astian
designing a proper log curve
19:52
se6astian
3d luts
19:52
se6astian
etc.
19:52
se6astian
experimental raw output modes
19:52
se6astian
https://www.apertus.org/axiom-beta-roadmap
19:52
wakalixes
ok, sounds like quite some work
19:52
se6astian
camera hosted websites to control the camera
19:52
se6astian
and visualize stuff
19:53
se6astian
histograms, curves, heat and power consumption related guis
19:53
se6astian
etc. etc. :)
19:54
wakalixes
looking at https://www.apertus.org/sites/default/files/dr-illustration-v03.png
19:55
wakalixes
why not trying to achieve a similar curve to the film negative?
19:55
wakalixes
which is symmetric with respect to dark and bright colors?
19:57
wakalixes
is it because of the the pixel reset of the multiple slope feature?
19:57
philippej_
I guess it will be possible to more closely match the response of film
19:58
philippej_
it depends on the latitude we'll have regarding the analog stage on the sensor
19:58
wakalixes
because you can only reset bright pixels you need to have a large exposure time for all, so you can not bend the curve upwards (make it concave)
19:59
alexML
wakalixes: that's how the raw sensor output looks like response; from there, you can undo the piecewise curve and get a linear response, and from there you can apply any curves you wish
20:01
wakalixes
sure, you can mimic any curve in post
20:02
alexML
if you ask me, I would convert the piecewise 12-bit raw data to 16-bit linear (DNG or EXR, depending on how much DR you want), then edit it
20:02
wakalixes
but for the shown piecewise linear curve one has high dynamic for dark pixels but low dynamic for bright pixels
20:04
wakalixes
it is not symmetric, this might be only good for dark areas, but you loose details in the bright
20:04
alexML
I highly doubt the detail loss will be noticeable
20:04
alexML
just apply log2 to that curve
20:05
troy_s
wakalixes: The output of the camera curve should attempt to distribute equal bits per stop as a general decent concept.
20:05
wakalixes
ok, but what about the ampified noise in the bright?
20:05
alexML
yes, and the piecewise curve approximates exactly that
20:05
troy_s
alexML: Is there a defined one now?
20:05
troy_s
wakalixes: What do you mean amplified noise?
20:06
troy_s
wakalixes: Last I looked, I haven't seen any piecewise generated images (well beyond the Alpha stuff that was ages ago and simply prototypical)
20:06
troy_s
alexML: Have you seen any PW imagery yet?
20:07
alexML
troy_s: I've only seen the same samples as you, and the documentation in the datasheet
20:07
alexML
you can fine-tune the two knots on the piecewise curve
20:07
alexML
that's it
20:07
troy_s
wakalixes: The long and short of it is that in an 'ideal' world, of course we would only have a linear signal that grabbed _all_ the data for a 50 stop range, but that isn't realistic.
20:07
wakalixes
applying log2 to the curve will make the response of the camera be somewhat linear, right?
20:08
troy_s
wakalixes: So the question is, what is a reasonable trade off? We obviously desire the latitude, but we also want to keep a decent degree of proper perceptual adjustment in post. That implies attempting to balance out the bits with "what we see", hence the log2 concept.
20:08
troy_s
But not a pure log2.
20:09
troy_s
Because that wouldn't result in a very balanced perceptual equivalence.
20:09
wakalixes
and because the slope is much less steep for the bright areas, you will amplify the quantization noise by trying to get a linear response
20:09
troy_s
Thinking.
20:09
troy_s
You won't be amplifying it unless I am misunderstanding you.
20:10
alexML
wakalixes: yes, the quantization noise gets amplified, but I bet the shot noise is much higher, so you won't notice it
20:10
troy_s
You'd have much thinner data across those stops for certain.
20:10
troy_s
I am not quite seeing the 'amplification' issue though.
20:11
wakalixes
troy_s, ok, the data will be thinner, I see that
20:11
troy_s
wakalixes: Technically there won't be any multiplication of data, at all.
20:11
wakalixes
by amplifying the noise I mean that any signal to noise ratio will become smaller
20:11
troy_s
wakalixes: So let's assume we can get a decent PW curve that grabs 14 stops (optimistic perhaps)
20:12
wakalixes
ok
20:12
troy_s
wakalixes: What noise are you speaking of? I'm not quite following you on that point.
20:12
troy_s
I mean we have the linear response
20:12
troy_s
so if we look at the linear response, we can do the following in post:
20:12
troy_s
Adjust exposure up and down and suffer _no_ degradation in terms of sparseness of data
20:13
troy_s
So if we slide the exposure up, there isn't any more noise
20:13
alexML
troy_s: he's talking about quantization noise, from round-off errors (since we are storing the data as integer values)
20:13
troy_s
we just run out of head room.
20:13
troy_s
Oh... I see.
20:13
alexML
so yeah, this noise gets amplified when you go from piecewise to linear
20:13
wakalixes
quantization noise coming from the ADC
20:13
troy_s
I suppose you could say that we are losing accuracy
20:13
alexML
but, once you apply the gamma curve (which is somewhat like a log), that noise gets back down
20:13
troy_s
but I still don't necessarily see how we are generating noise.
20:14
wakalixes
no, not generating noise
20:14
troy_s
We are losing accuracy in the sense that the photosites might be forced to be quantized into a unit they don't belong, but in the end, such is the tradeoff.
20:14
troy_s
I'd like to think of it as strictly "thinning" the data.
20:14
wakalixes
i mean that the SNR will be dependent on the slope of the piecewise curve, i.e. for dark areas larger and for bright areas smaller
20:15
troy_s
In theory, once the 3D LUTs are crafted according to that 'settled' data, the colorimetry should still be accurate.
20:15
se6astian
afk for a bit
20:15
troy_s
wakalixes: But in terms of a photographic image and our perceptual systems, we are already sort of doing that similar compression.
20:16
troy_s
So I'm still not entirely sure of the concern I suppose.
20:16
wakalixes
true
20:16
troy_s
My point would be, YES we are trading off that post production versatility. That's a rough one. You can't overexpose your image or underexpose too badly.
20:16
troy_s
That's absolutely true.
20:16
wakalixes
so the ADC will give a response curve with steps of the size of the LSB
20:16
troy_s
A curve forces you to be a little more accurate.
20:17
troy_s
Ideally I'd like to see something specifically tuned to the number of stops.
20:17
troy_s
So if that sensor captures 14.3 stops, allocate equal bits accurately across that range.
20:17
wakalixes
and andoing the piecewise slope will decrease the step size for dark areas but increase them for bright
20:17
troy_s
Yep.
20:17
wakalixes
undoing
20:18
troy_s
Because basically our perceptual system will find that range most "observed" as it were, and therefore needs to be denser in that sweet spot.
20:18
wakalixes
hence the noise from the ADC several bits will be "amplified" in the bright
20:18
troy_s
So in post, you could take CustomApertusTransferCurve and invert it
20:18
alexML
wakalixes: FYI, Nikon compresses the highlights in the same way when saving NEF
20:18
alexML
https://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p3.html
20:18
alexML
look for " An aside on "lossy" NEF compression: "
20:19
troy_s
Why wouldn't you do that given the current limitations of limited latitude sensors?
20:19
troy_s
I mean it isn't terribly practical to shoot motion pictures or still pictures with a nine stop latitude.
20:19
troy_s
That's the reality.
20:19
wakalixes
that is true
20:19
troy_s
That's almost as bad as slide reversal stock. LOL.
20:20
troy_s
(Not quite. Some of us still have emotional scarring from shooting extremely thin latitude reversal stocks.)
20:20
troy_s
Needless, that's the great trade off.
20:20
wakalixes
I am asking because I know the sCMOS sensor CIS2521 (I was mentioning this one recently)
20:20
troy_s
I mean at some point that PW curve will break down anyways
20:20
troy_s
wakalixes: What are the dimensions and linear latitude range?
20:20
wakalixes
and this sensor has two outputs, 11bit with low gain and 11bit with high gain
20:21
troy_s
Oh interesting.
20:21
troy_s
What are the technical differences between the two modes?
20:21
wakalixes
and on top of all an ultra-low-noise readout
20:21
wakalixes
of 2e-
20:21
troy_s
What are the dimensions?
20:21
troy_s
11 bit seems a tad on the thin side of things as a gut response, not knowing anything more.
20:22
wakalixes
http://fairchildimaging.com/category/product-category/focal-plane-arrays/scmos
20:22
troy_s
alexML: Have you been making any experiments or otherwise lately?
20:22
alexML
mostly looking at dark and bright frames
20:23
alexML
found out there is an offset between odd and even rows, not totally correctable with dark frames though
20:23
troy_s
wakalixes: No dimensions on that page.
20:23
wakalixes
I think sebastian mentioned, there might be the possibility of using these sensors from BAE systems as well at some point
20:23
troy_s
alexML: That's odd.
20:23
troy_s
alexML: Sensor or hardware related?
20:23
alexML
there seem to be two parallel readout circuits, with slightly different electrical characteristics
20:24
troy_s
wakalixes: I still don't know the dimensions. Doesn't sound at all viable for the primary audience if it isn't Super35 letterbox at a minimum.
20:24
alexML
so, many of those artifacts appear because of the differences between the two circuits
20:24
troy_s
alexML: Interesting. Any chance of balancing those circuits out?
20:25
alexML
I'll try doing that in software
20:25
wakalixes
the CIS2521 has 2560 (H) x 2160 (V)
20:26
wakalixes
but there is LTN4625A with 4608 (H) x 2592 (V) pixels (presented in Feb 2015)
20:26
alexML
2560 x 6.5um = 16.6mm
20:26
troy_s
alexML: Not sure that is ideal colorimetry wise. Would be mroe prudent if it can be equalized in hardware.
20:26
troy_s
wakalixes: _Dimensions_ not resolution.
20:27
troy_s
2560 Bayered????
20:27
troy_s
That's uh...
20:27
wakalixes
it is APS-C
20:28
wakalixes
sure it is not so interesting for cinematographic purposes, more scientific
20:30
alexML
troy_s: there is an offset that can be adjusted in hardware
20:30
troy_s
APS-C is ballpark same width as Super35.
20:31
troy_s
alexML: Interesting.
20:31
alexML
black offset can be specified for odd and even rows
20:31
alexML
but other than that, all the other parameters look common
20:33
alexML
(common to both odd/even rows)
20:33
alexML
what I don't like is that, on a test gray frame, after subtracting a dark frame, only half of that offset is corrected
20:34
alexML
so I don't think a plain offset would be enough (maybe gain is different too)
20:44
alexML
hm, histogram matching says a plain offset is enough, no difference in gains (that's good)
20:44
alexML
but why the dark frame doesn't fix it, then?!
20:50
philippej_
left the channel
21:23
se6astian
off to bed
21:23
se6astian
good night
21:23
se6astian
changed nick to: se6astian|away
21:23
davidak
left the channel
21:49
wakalixes
left the channel
22:01
pozitron
joined the channel
22:57
comradekingu
joined the channel