From: ejmartin on 19 Jul 2008 15:36
On Jul 19, 1:44 pm, "Roger N. Clark (change username to rnclark)"
> ejmartin wrote:
> > Well, that's the tradeoff isn't it? At least, that's the tradeoff for
> > current implementations at high ISO, where it's the sensor properties
> > that control DR -- big photosites for more sensitivity in shadows, or
> > small photosites for more resolution. At low ISO, due to the
> > limitations of current implementations in the ISO amplifier/ADC, one
> > actually does better per unit area with small photosites, because they
> > place less demands on the dynamic range of that downstream
> > electronics. Small photosites will be better at low ISO until camera
> > companies start delivering all the DR that the sensor is offering.
> Actually, its not simply low iso, its the lowest ISO, e.g. ISO 100.
> Try again at ISO200 and most modern DSLRs have about the same dynamic
> range at ISO200 as at ISO 100, and nearly as much at 400. So this
> biased test with small sensor P&S cameras is only valid at the lowest
> ISO where other camera electronics can't match the huge dynamic range of
> the large pixels.
> You can see some of these effects in Figures 4 and 5 at:http://www.clarkvision.com/imagedetail/digital.sensor.performance.sum...
> In the future I expect this low iso limitation to be corrected; consumer
> astrophoto cameras do not have this problem, but they use 16-bit converters.
I would have said that the problem of the system not delivering what
the sensor is offering is a problem any time the read noise for a
given ISO (referred to electron equivalents) is higher than the read
noise at high ISO (referred to electron equivalents). Then the
downstream electronics is a limiting factor in the delivered DR at
that ISO, rather than the sensor read noise being that limiting
factor. And that's pretty much true up to ISO 1600 for Canon DSLR's
(it's less true for Nikons, but then again their high ISO read noises
aren't yet as good as Canon's) since system read noise continues to
drop until that point -- though it's getting pretty close on most
Canons by ISO 800, with only a slight further drop between 800 and
1600; I am also ignoring for this purpose the intermediate ISO's on
Canons, which are completely bungled (they are fine on Nikons). You
can some of these effects in figures 14-15 and the table just below
It's interesting that astrophoto cameras are not so limited; can you
point me to some example spec sheets?
I think there are also simple low-tech fixes for this issue in current
DSLR's. The problem is the limited DR of the ISO amplifier and ADC.
So one could run two amplifier/ADC processing chains in parallel,
using the same sensor signal. One amplifier would be set to ISO 100
and the other to 1600 (or better yet, one should simply optimize two
fixed amplifications). The ISO 100 path captures all the highlights
(assuming proper exposure), but has high read noise in the shadows.
The ISO 1600 path oversaturates the top four stops of highlights, but
captures much cleaner shadows since the system's ISO 1600 read noise
in electrons is so much lower than that of the ISO 100 path. One
could then combine the digitized output of these two paths in firmware
ala HDR before writing out the data to RAW. To get an idea of the
improvement to be had, here is a plot of the SNR vs exposure of a 1D3
using the standard noise model (read plus photon shot) for both ISO
100 and 1600, normalized to the same absolute EV:
The blue curve is for ISO 100, the red curve (which stops where the
ISO 1600 ADC saturates, four stops down relative to the ISO 100 one)
is for ISO 1600. Between 9 and 10 on the horizontal axis one would
probably want to ramp between the two outputs to avoid artifacts. The
resulting SNR vs exposure would be the envelope of the two curves, and
note that the total range between saturation and S/N=1 (0 on the
vertical axis) is a DR of 14 stops. Note also the very long range of
exposure levels where the signal is shot noise dominated, vs the blue
ISO 100 curve which basically shows read noise contamination at the
point where it starts pulling away from the ISO 1600 curve.
Two final points. First, since the DR demands on any one of the
amplifier/ADC paths is somewhat lifted, one can get away with using
cheaper 12-bit converters and still cover the range (there is about
one stop overlap between where the ISO 1600 path is available, and
where the ISO 100 path runs out of bit depth). Second, one can run
the two amplifiers at these fixed amplification settings, and simply
record all the image information the sensor was able to record. There
is no reason any more to have variable gain amplification (ISO), since
all sensor info was recorded with adequate bit depth (16 total using
12-bit ADC's and the HDR blending). ISO becomes metadata like white
balance -- a suggestion to the raw converter as to what EV shift
should be applied at the conversion stage to normalize the histogram.
I'm told that this sort of dual amplification scheme has been used in
the past in certain scientific applications in CCD imaging. I don't
see any obvious technological obstacle to implementing it today for
DSLR's, for an easy two-stop improvement in DR.
From: John Sheehy on 19 Jul 2008 15:43
ASAAR <caught(a)22.com> wrote in
>> If I did it, you would say I'm lying or rigged something if you
>> didn't like the results.
>> . . .
>> I remember looking at the RAW data from the Nikons a while back, and
>> I remember that the two cameras had very different QEs, with the D40X
>> being more sensitive, photon-wise, but the D40 having lower read
>> noise at higher ISOs.
>> . . .
>> Well, can you get someone to provide the RAWs (not RAW conversions!)
>> from the two Nikons, same scene and lens, and make sure the lighting
>> and the exposure is the same?
> Gee, what makes you think that anyone would think that you might
> lie about anything? Hmm.
Well, it's pretty obvious that when people hold false beliefs, that the
truth looks like lies to them.
> You admit being familiar with the D40,
> the D40x and their RAW data, which are very similar contemporary
> cameras ("from the same era") that have vastly different numbers of
> pixels on similar sensors and which use the same mount,
Well, they are actually quite different. They have completely different
types of circuitry, and
> yet as we
> see above, you insisted that :
>> This is the best possible way to run this experiment, as there are no
>> possible pairings out there of cameras with the same mount and the
>> same size sensors with vastly different pixel densities of the same
5:3 is *NOT* a vastly different pixel density; in fact, the truly
relevant value is the square root of that. 8.35:1 *is* vastly different,
and if it is true that higher pixel density means lower IQ, then it
should certainly show up there more easily against the "noise" of "other
things not being equal". I feel like a kindegarten teacher here, having
to spell all these things out.
> In fact as others have noted, there are many such pairings.
> Refusing to admit such an obvious mistake tells us much about
> whether we should entertain the notion that you're capable of lying
> or rigging something if *you* don't like what you see. Refusing to
> address Roger's points only adds to the suspicion that you can't be
I haven't gotten to Roger's posts yet. The flow has not stopped here,
and I'm saving him for last because his posts are more work to reply to
than those of others.
Delay is not refusal.
<>>< ><<> ><<> <>>< ><<> <>>< <>>< ><<>
John P Sheehy <JPS(a)no.komm>
><<> <>>< <>>< ><<> <>>< ><<> ><<> <>><
From: ejmartin on 19 Jul 2008 16:07
On Jul 19, 2:09 pm, "Roger N. Clark (change username to rnclark)"
> I use the dropping MTF as diffraction overtakes the pixel pitch
> and the image becomes softer so effective resolution drops.
> But I assume S/N per effective resolution unit is constant so
> you are just slicing up the same total photon count into more pixels.
> But at a lower point the DR of individual pixels becomes too small
> and limits dynamic range too much, dropping image quality faster
> as pixel size continues to drop. One can derive various subjective
> functions which change the exact shape of the curves but don't
> change the fact that there is an optimum.
I agree that there is an optimum, the question is where it lies. Why
do you take the S/N per effective resolution as constant in the
diffraction limited regime? As the diffraction spot grows with
decreased aperture (at fixed pixel pitch), it comprises more and more
pixels and so the S/N ratio per resolution size grows and partially
compensates the resolution decrease in your AIQ metric, due to pixel
binning effects the SNR increases as the number of pixels inside the
Note also, that if one decreases pixel pitch at fixed aperture,
resolution does not completely saturate at the Airy disk size, as I
pointed out in a couple of recent posts elsewhere:
There are still resolution gains in the diffraction limited regime,
they are just somewhat less than proportional to the reduction in
pixel pitch. Again, when figuring SNR at the resolution scale, one
should bin the pixels within that scale size in order to determine
> For example, tests of the Canon 1D Mark III initially showed it had
> anomalously different full well and gains, but then it was discovered that
> the lowest two ISOs were really the same gain and the higher ISO data
> were simply scaled by 2.
Yes, I was the one who pointed that out to you :)
> So let's see a real analysis with all the data presented before concluding
> the FZ50 is out of the box. Also, be aware that not all "raw" data from
> cameras is actually raw data from the sensor. Perhaps the raw FZ50
> data files have had processing.
I agree. It would be nice if John would publish in one place the
results of all the tests he has done. I should say that I get similar
numbers for my LX1: about 4100 electrons full well at ISO 100, with
2.2µ pixels. I haven't seen any evidence of NR of the raw data that
would contaminate the results, but then again I haven't done as
exhaustive analysis of that as I have on DSLR's such as the D300 and
From: Ray Fischer on 19 Jul 2008 16:11
John Sheehy <JPS(a)no.komm> wrote:
>I am showing the visual, and I hope that it also shows that standard
>methods of quantitative noise measurements are faulty.
It shows that your measurements are faulty and deceptive.
>You can clearly see that at 100% pixel view for both cameras, that the
>FZ50 has more pixel noise. It always has less visible noise, however,
>when both crops are scaled to the same size.
But why would anyone ever want to do that?
From: Steve on 19 Jul 2008 16:38
On Sat, 19 Jul 2008 15:02:55 GMT, Steve <steve(a)example.com> wrote:
>However, using John's test, you can show that 1000 megapixels for an
>APS-C sized sensor has better S/N and DR than the 10 megapixel sensor.
>Because he would only use the portion of the DR that the 1000 MP
>sensor can manage when comparing it to the 10MP sensor, which would
>have tons of wasted headroom in his test. And he would cut a tiny
>area of the focal plane image out of the 10MP sensor and blow it up to
>10,000% and compare that with a 100% crop of the 1000MP camera.
And on top of that, he would use the ISO setting that the 1000MP could
work at for the entire test for both sensors and not vary the range if
ISO to include values that normal photographers would want to use.
Limiting his test to ISO 100 for both cameras and then pushing it to
13,500 is also silly. Lets see his results for ISO 400, ISO 800 and
Of course, he won't show those results.