Ad
Ad
Ad
Pages: « 1 ... 6 7 [8] 9 »   Bottom of Page
Print
Author Topic: ETTR vs ISO  (Read 16962 times)
jrsforums
Sr. Member
****
Offline Offline

Posts: 722


« Reply #140 on: May 28, 2013, 10:03:35 AM »
ReplyReply

Good show, thanks Jim.  Makes one wonder whether to save space one should stick to 12-bit mode.

Unlike Nikon, Canon does nt have a switch to 12bit mode.
Logged

John
Jim Kasson
Sr. Member
****
Offline Offline

Posts: 744


WWW
« Reply #141 on: May 28, 2013, 10:15:39 AM »
ReplyReply

The same is true of the D800E, although to a lesser extent:



Jim
Logged

Jack Hogan
Full Member
***
Offline Offline

Posts: 207


« Reply #142 on: May 28, 2013, 01:26:12 PM »
ReplyReply

Here are the simulation results, with the curves being the 14-bit SNRs less the 12-bit SNRs:



I don't know quite what to make of it. It shows that 14-bits gives an improvement in SNR for shadows at low ISO, but actually hurts a tiny bit for shadows at high ISO. In the middle tones and above, it makes no difference.

Hmmm, weird.  Did you just truncate the 12 bits?  What % of full scale is your Series1 at?
Logged
Jim Kasson
Sr. Member
****
Offline Offline

Posts: 744


WWW
« Reply #143 on: May 28, 2013, 01:52:37 PM »
ReplyReply

Did you just truncate the 12 bits?

Rounded the analog value, just like an ADC would be designed to do. You want the transitions to occur half an LSB to either side of the nominal.

What % of full scale is your Series1 at?

1/40,000 at  ISO 12800, and 128/40,000 at base ISO. [Clairification: that's with respect to FWC, with respect to the maximum output of the ADC, it's 128/40000 for all ISOs.] The ISO 12800 exposures are on average about 1.5 electrons/plane for the Canon, and close to one electron per plane for the Nikon.

Jim
« Last Edit: May 28, 2013, 04:24:54 PM by Jim Kasson » Logged

Jack Hogan
Full Member
***
Offline Offline

Posts: 207


« Reply #144 on: May 29, 2013, 02:12:33 AM »
ReplyReply

Rounded the analog value, just like an ADC would be designed to do. You want the transitions to occur half an LSB to either side of the nominal.

1/40,000 at  ISO 12800, and 128/40,000 at base ISO. [Clairification: that's with respect to FWC, with respect to the maximum output of the ADC, it's 128/40000 for all ISOs.] The ISO 12800 exposures are on average about 1.5 electrons/plane for the Canon, and close to one electron per plane for the Nikon.

Jim

Intuitively one would expect virtually no difference between the two given the 8 ADU noise 'floor'.  I wonder if what we are seeing is actually quantization noise.  How large of a sample do you use?  If too small the higher bit depth could appear to produce a better apparent SNR in the deep shadows - but it might simply be better defined noise.

Jack
« Last Edit: May 29, 2013, 02:16:36 AM by Jack Hogan » Logged
Jim Kasson
Sr. Member
****
Offline Offline

Posts: 744


WWW
« Reply #145 on: May 29, 2013, 09:19:15 AM »
ReplyReply

 How large of a sample do you use?

I used a 1600x1600 sample, so 800x800 = 640,000 green pixels. I used a separate sample for each data point.

Jim
« Last Edit: May 29, 2013, 10:10:36 AM by Jim Kasson » Logged

Jim Kasson
Sr. Member
****
Offline Offline

Posts: 744


WWW
« Reply #146 on: May 29, 2013, 10:23:12 AM »
ReplyReply

Intuitively one would expect virtually no difference between the two given the 8 ADU noise 'floor'. 

128/40,000 at base ISO is 52 14-bit ADUs, so it's quite a bit above the noise floor.  If you're talking about the negative values at high ISOs, I agree. Upon thinking about it, I think the small SNR reduction in those regions for the 12-bit resolution is due to the ADC digitizing analog values that would have been different in the 14-bit case into the same value, thus lowering standard deviation. I can run some tests that isolate that effect and report back.

Jim
Logged

bjanes
Sr. Member
****
Offline Offline

Posts: 2756



« Reply #147 on: May 29, 2013, 10:50:01 AM »
ReplyReply

Hi TJ,

I don't shoot Canon, but from what I can tell the 5DIII uses post-ADC digital gain only after ISO 12800, so no point raising ISO beyond that unless you need 'properly' bright images SOOC. Looking at Bill Claff's graph, it seems that you do get a benefit in shadow SNR by raising ISO in-camera from 100 all the way up to 12800, although the law of diminishing returns needs to be considered after about ISO 2500: raising it from 2500 to 12800 only nets you a 0.2 stop improvement in shadow SNR at the expense of 2.3 stops of reduced DR: a decision best evaluated each time based on the situation (blips in the curve above 12800 are suspicious, as the data up there is often 'cooked').

The other take-away from the curve is that if you are after best SNR/DR it appears that you are better off increasing ISO one stop at a time (as opposed to in 1/3 or 1/2 stop increments) in correspondence of the steps in the graph: IOW if you are sitting at ISO 300 and you feel the urge to increase ISO a little, there is no point going to 400 or 500.  Just leave it there with the benefit of slightly better DR until you feel the need to go to 600.

Jack

PS the 5DIII appears to have an effective QE of about 14% and PRNU of about 0.4%, plus a fairly noisy analog amplifier contributing about 8 ADU random noise throughout the range, so the last three bits never contain much information.



Jack,

Your graph is interesting and can serve as a topic for discussion. It fits Emil's simplified read noise formula R2 = (G R0)2 + R2, where R is the total read noise, G is the ISO, R1 is noise upstream of the ISO amplifier (presumably mainly the sensor itself) and R2 is the noise downstream of the amplifier (presumably with the ADC representing a major component). Roger Clark in his sensor performance analysis (Fig 8a) implicates the ADC as a main contributor of read noise for the Canon camera. The amp noise itself is a tweener in Emil's analysis.

If I interpret it correctly, your graph shows a constant sensor read noise component which is multiplied by the ISO amplifier. What you term the amp read noise could include the ADC noise which is constant and adds in quadrature to the sensor noise to give the up-sloping total read noise curve. Bill Claff states that a curved plot of the read noise expressed in ADU numbers indicates that the ADC contributes significantly to the read noise whereas a straight line plot indicates that the sensor is the main source of the read noise.

Bill's plots for the Nikon D800e are combined and shown below for ISO up to 1600. The read noise in electrons is relatively constant but the values are affected by the ADC readout and the noise in ADUs is linear indicating the source lies mainly in the sensor. The curves cross at approximately the unity gain of the sensor (ca 320), but I don't know if this is fortuitous or significant.



What do you think?

Bill

« Last Edit: May 29, 2013, 10:58:34 AM by bjanes » Logged
Jim Kasson
Sr. Member
****
Offline Offline

Posts: 744


WWW
« Reply #148 on: May 29, 2013, 11:16:02 AM »
ReplyReply



What do you think?

The data point at ISO 1600 looks suspicious to me. In my tests on a real-world D800E, I see the read noise referred to the sensor at ISO 1600 as just slightly worse than at ISO 800. I also don't know of a mechanism to have the read noise drop suddenly like that. Otherwise, this is generally close to what I see in my testing.

Jim
Logged

bjanes
Sr. Member
****
Offline Offline

Posts: 2756



« Reply #149 on: May 29, 2013, 11:26:49 AM »
ReplyReply

The data point at ISO 1600 looks suspicious to me. In my tests on a real-world D800E, I see the read noise referred to the sensor at ISO 1600 as just slightly worse than at ISO 800. I also don't know of a mechanism to have the read noise drop suddenly like that. Otherwise, this is generally close to what I see in my testing.

Jim

Jim,

Bill does note that at ISO 1600 and above, the D800e appears to manipulate the raw data prior to writing it out to the memory card. Others have noted this by looking at the histograms in Rawdigger.

Bill
Logged
Jim Kasson
Sr. Member
****
Offline Offline

Posts: 744


WWW
« Reply #150 on: May 29, 2013, 11:55:23 AM »
ReplyReply

Bill does note that at ISO 1600 and above, the D800e appears to manipulate the raw data prior to writing it out to the memory card. Others have noted this by looking at the histograms in Rawdigger.

Good point. I have not noticed this effect in my measurements, but it might be the result of the digital gain that I've noticed the camera applying at ISO 3200 and above (leftward bit shifts of the digital information).

If that's the case, then the ADC is effectively progressively losing resolution, with possible effects on the noise statistics like I was struggling with a few posts above in another context.

Now I really want to run some tests.

Jim
Logged

Jim Kasson
Sr. Member
****
Offline Offline

Posts: 744


WWW
« Reply #151 on: May 29, 2013, 12:36:18 PM »
ReplyReply

I've run some tests on quantizing signals plus Gaussian noise at varying resolutions, both with and without the quantizing of the sensor values to integer numbers of electrons. The results are to a great extent independent of resolution, at least in the 12 to 14 bit range.

So that's not it. I wonder if the effect of decreasing resolution for shadows at high ISOs slightly improving the SNR can be seen in a real-world camera that has both 12 and 14 bit options. The effect in the simulated D800E is so slight that it would be really hard to observe with a real camera.

It's a mystery to me.

Jim
« Last Edit: May 29, 2013, 12:51:14 PM by Jim Kasson » Logged

Jack Hogan
Full Member
***
Offline Offline

Posts: 207


« Reply #152 on: May 29, 2013, 01:58:19 PM »
ReplyReply

If I interpret it correctly, your graph shows a constant sensor read noise component which is multiplied by the ISO amplifier. What you term the amp read noise could include the ADC noise which is constant and adds in quadrature to the sensor noise to give the up-sloping total read noise curve. Bill Claff states that a curved plot of the read noise expressed in ADU numbers indicates that the ADC contributes significantly to the read noise whereas a straight line plot indicates that the sensor is the main source of the read noise.

Yes to all of the above, Bill.  The two read noise values (sensor and amp/ADC) in the simplified formula are found fitting the total read noise derivations.  [Please note that as mentioned earlier to Jim the inverse gain that is shown in my 5DIII chart is incorrect because I forgot that some Canons have severe zero offsets.  At ISO 100 it should read 5.11 instead of 4.13, with the other ISOs proportionately lower.]

Bill's plots for the Nikon D800e are combined and shown below for ISO up to 1600. The read noise in electrons is relatively constant but the values are affected by the ADC readout and the noise in ADUs is linear indicating the source lies mainly in the sensor. The curves cross at approximately the unity gain of the sensor (ca 320), but I don't know if this is fortuitous or significant.

Yes again.  Total read noise in e- and ADU are related by the inverse gain, which you can find simply by dividing the two together.  For instance for the D800e at ISO 100, from Bill's site, RN100(ADU)=1.261 and RN100(e-)=4.228, therefore igain according to Bill is 3.39 e-/ADU at this ISO.  When will the two read the same number?  At the ISO where gain is 1.  No luck involved this time Smiley

Quote


What do you think?

Looks good, it shows a nice, almost 'ISOless' system.  I think you may have a typo in the electrons curve at ISO 1600: it looks lower than 2.721 e-

Jack
Logged
Jack Hogan
Full Member
***
Offline Offline

Posts: 207


« Reply #153 on: May 29, 2013, 02:01:22 PM »
ReplyReply

Good point. I have not noticed this effect in my measurements, but it might be the result of the digital gain that I've noticed the camera applying at ISO 3200 and above (leftward bit shifts of the digital information).

As you can see in the image in Reply#80 a couple of pages back, the D800e appears to start applying ISO digitally just before ISO 1600 and thereafter.

Jack
Logged
Jim Kasson
Sr. Member
****
Offline Offline

Posts: 744


WWW
« Reply #154 on: May 29, 2013, 02:19:48 PM »
ReplyReply

As you can see in the image in Reply#80 a couple of pages back, the D800e appears to start applying ISO digitally just before ISO 1600 and thereafter.

Jack


You're right, Jack. I wrote about that a few months ago, but forgot the histo dropouts at 1600.  Of course, the D800 uses digital gain at all ISOs in the red and blue channels.

Jim
Logged

Jack Hogan
Full Member
***
Offline Offline

Posts: 207


« Reply #155 on: May 29, 2013, 03:07:37 PM »
ReplyReply

I've run some tests on quantizing signals plus Gaussian noise at varying resolutions, both with and without the quantizing of the sensor values to integer numbers of electrons. The results are to a great extent independent of resolution, at least in the 12 to 14 bit range.

OK, that makes sense, especially for as large a sample as you are using.

So that's not it. I wonder if the effect of decreasing resolution for shadows at high ISOs slightly improving the SNR can be seen in a real-world camera that has both 12 and 14 bit options. The effect in the simulated D800E is so slight that it would be really hard to observe with a real camera.

It's a mystery to me.

Jim

Can't think of anything else myself.

Jack
Logged
Jack Hogan
Full Member
***
Offline Offline

Posts: 207


« Reply #156 on: May 29, 2013, 03:10:56 PM »
ReplyReply


Of course, the D800 uses digital gain at all ISOs in the red and blue channels.

Jim

Yes, it's what Iliah Borg calls WB preconditioning.

Jack
Logged
Jim Kasson
Sr. Member
****
Offline Offline

Posts: 744


WWW
« Reply #157 on: May 29, 2013, 03:39:41 PM »
ReplyReply

Yes, it's what Iliah Borg calls WB preconditioning.

I don't know why any digital multiplication of unsigned integers isn't better done in the raw converter , but there it is.

Jim
Logged

Jim Kasson
Sr. Member
****
Offline Offline

Posts: 744


WWW
« Reply #158 on: May 29, 2013, 03:43:38 PM »
ReplyReply



By the way, if you use dark-field exposures to calculate read noise, you get values that are quite different from these, probably because of the D800E's truncation of negative values. I understand that Canon has a built-in offset to keep this from happening.

Jim
Logged

bjanes
Sr. Member
****
Offline Offline

Posts: 2756



« Reply #159 on: May 29, 2013, 04:52:31 PM »
ReplyReply

By the way, if you use dark-field exposures to calculate read noise, you get values that are quite different from these, probably because of the D800E's truncation of negative values. I understand that Canon has a built-in offset to keep this from happening.

Jim,

Those read noises were from Bill Claff's data. He took the read noise from the optical black region whose data are not truncated. He uses his own proprietary software, but one can use RawDigger to examine the optical black area.

Here is the optical black area of the D800e at ISO 100 with the optical black area selected by hand (one could use selection by the numbers to get a bigger sample). The read noise for the green channels is 1.1 ADU. The histogram is approximately gaussian and an offset of 600 ADU has been applied to prevent truncation.

Bill

Logged
Pages: « 1 ... 6 7 [8] 9 »   Top of Page
Print
Jump to:  

Ad
Ad
Ad