Moon Maid's sensor comparison (PART 2)

Started Mar 1, 2012 | Discussions
Absolutic
Absolutic Veteran Member • Posts: 5,360
Moon Maid's sensor comparison (PART 2)

Dear Marienne. Thank you for your test. Now reading the Nikon Engineers blog on D800
http://imaging.nikon.com/history/scenes/32/index.htm

These Nikon D800 developers suggest shooting at Medium size (appx. 20MP) claiming it is a healthy equilibrium, in terms of general shooting. My understanding from reading their blog is when picking medium, the camera's EXPEED3 chip is involved in processing from 36 to 20MP. I see no problem for myself shooting at 20MP in FX mode for general photography, assuming there is some advantage in terms of low light performance. Do you happen to also have charts for 20MP medium size?

-- hide signature --

 Absolutic's gear list:Absolutic's gear list
Nikon Coolpix P7700 Sony RX1 Sony Cyber-shot DSC-RX10 Sony Alpha a7R III Olympus PEN-F +21 more
Nikon D800
If you believe there are incorrect tags, please send us this post using our feedback form.
AgentHEX Contributing Member • Posts: 676
Re: Moon Maid's sensor comparison (PART 2)

Thanks for doing this. Does anyone know the reason why dpreview cuts off at 150 msgs? That's really annoying for longer discussions.

Absolutic wrote:

Dear Marienne. Thank you for your test. Now reading the Nikon Engineers blog on D800
http://imaging.nikon.com/history/scenes/32/index.htm

These Nikon D800 developers suggest shooting at Medium size (appx. 20MP) claiming it is a healthy equilibrium, in terms of general shooting. My understanding from reading their blog is when picking medium, the camera's EXPEED3 chip is involved in processing from 36 to 20MP. I see no problem for myself shooting at 20MP in FX mode for general photography, assuming there is some advantage in terms of low light performance. Do you happen to also have charts for 20MP medium size?

-- hide signature --

Marius D New Member • Posts: 10
D800 read noise climbing more slowly than ISO

Marianne Oelund wrote:

I haven't included this effect in the posted curves because I need more data before I'll accept it, but it does appear from the samples I have, that the D800 read noise (in ADU) is climbing more slowly than proportionality to ISO. On the surface, one wouldn't expect this, since ISO sensitivities above ISO 1000 (if it follows the D7000 example) should be only by numerical scaling.

I can see two possible explanations: 1) Some gain sharing at high ISO, such that some of the boost is still provided by the analog electronics; 2) A slight amount of NR is being applied to the RAW data.

There is another possible explanation: 3) reducing ADC reference voltage (V.ref) to match the input analog domain as ISO raises. This should appear like a numerical scaling, except for reduced quantization errors. My guess is they are using both analog amplification and V.ref scaling. From around 1000 ISO up to 6400 there is no analog amplification, only V.ref scaling with a small but measurable contribution in reducing read noise due to the optimized quantization and other factors like better reset voltage reading or optical-black photo-detectors reading. Reset voltage and optical-black are both very low, reducing ADC's V.ref make sense for a precise digital conversion of these values.

bobn2
bobn2 Forum Pro • Posts: 56,299
Re: D800 read noise climbing more slowly than ISO

Marius D wrote:

Marianne Oelund wrote:

I haven't included this effect in the posted curves because I need more data before I'll accept it, but it does appear from the samples I have, that the D800 read noise (in ADU) is climbing more slowly than proportionality to ISO. On the surface, one wouldn't expect this, since ISO sensitivities above ISO 1000 (if it follows the D7000 example) should be only by numerical scaling.

I can see two possible explanations: 1) Some gain sharing at high ISO, such that some of the boost is still provided by the analog electronics; 2) A slight amount of NR is being applied to the RAW data.

There is another possible explanation: 3) reducing ADC reference voltage (V.ref) to match the input analog domain as ISO raises. This should appear like a numerical scaling, except for reduced quantization errors. My guess is they are using both analog amplification and V.ref scaling. From around 1000 ISO up to 6400 there is no analog amplification, only V.ref scaling with a small but measurable contribution in reducing read noise due to the optimized quantization and other factors like better reset voltage reading or optical-black photo-detectors reading. Reset voltage and optical-black are both very low, reducing ADC's V.ref make sense for a precise digital conversion of these values.

There is another possibility. I suspect that Nikon is actually adding a tiny bit of pseudo-random noise to provide a little dither right at the noise floor, my guess is to hide some of the very low level artifacts of some of the digital signal processing that they do (remember the EXMOR sensors are direct digital out, so there is little scope for analog signal processing apart from manipulating the configuration options that Sony has given them). If you look at the D7000, the high ISO read noise is a bit higher than other cameras using similar sensors. Possibly in the D800 Nikon has graded the added dither a little.
--
Bob

cluna Senior Member • Posts: 1,403
Re: D800 read noise climbing more slowly than ISO

bobn2 wrote:

There is another possibility. I suspect that Nikon is actually adding a tiny bit of pseudo-random noise to provide a little dither right at the noise floor, my guess is to hide some of the very low level artifacts of some of the digital signal processing that they do ...cut

I would hope they are including dither in the decimation pipeline already. We've been doing it in the over-sampled audio realm for a long time now.

-C

Marianne Oelund Veteran Member • Posts: 7,162
Histogram analysis needed

Marius D wrote:

There is another possible explanation: 3) reducing ADC reference voltage (V.ref) to match the input analog domain as ISO raises. This should appear like a numerical scaling, except for reduced quantization errors. My guess is they are using both analog amplification and V.ref scaling. From around 1000 ISO up to 6400 there is no analog amplification, only V.ref scaling with a small but measurable contribution in reducing read noise due to the optimized quantization and other factors like better reset voltage reading or optical-black photo-detectors reading. Reset voltage and optical-black are both very low, reducing ADC's V.ref make sense for a precise digital conversion of these values.

The first step in determination will be to find the numerical scaling factors used for ISO settings above 1000. This will tell us how much, if any, of the gain is from the analog side of the ADC.

Bob's suggestion is interesting, although it would imply that without the added dithering, base ISO read noise is almost unbelievably low - around 0.7 ADU.

bobn2
bobn2 Forum Pro • Posts: 56,299
Re: D800 read noise climbing more slowly than ISO

cluna wrote:

bobn2 wrote:

There is another possibility. I suspect that Nikon is actually adding a tiny bit of pseudo-random noise to provide a little dither right at the noise floor, my guess is to hide some of the very low level artifacts of some of the digital signal processing that they do ...cut

I would hope they are including dither in the decimation pipeline already. We've been doing it in the over-sampled audio realm for a long time now.

Yes, but I don't think there is a 'decimation pipeline' in the raw output. The EXMOR uses quite a simple digital ramp ADC, not a delta sigma, and it isn't oversampled in any sense.

-- hide signature --

Bob

cluna Senior Member • Posts: 1,403
Re: Histogram analysis needed

Marianne Oelund wrote:

The first step in determination will be to find the numerical scaling factors used for ISO settings above 1000. This will tell us how much, if any, of the gain is from the analog side of the ADC.

Bob's suggestion is interesting, although it would imply that without the added dithering, base ISO read noise is almost unbelievably low - around 0.7 ADU.

Would you mind including a Lo 1 to compare with true base ISO and simulated? Ive seen better output (IE shadow and sky noise reduced) on the D300s when using Lo1.

Thanks,

-C

Flashlight Veteran Member • Posts: 7,332
Re: Histogram analysis needed

cluna wrote:

Ive seen better output (IE shadow and sky noise reduced) on the D300s when using Lo1.

Use it all the time.

-- hide signature --

Philip

Marianne Oelund Veteran Member • Posts: 7,162
D300s base ISO

cluna wrote:

Would you mind including a Lo 1 to compare with true base ISO and simulated? Ive seen better output (IE shadow and sky noise reduced) on the D300s when using Lo1.

In the case of the D300/D300s, there is a slight advantage to going below ISO 100, as the true base is at L0.3. However, L1.0 is no better than L0.3.

For the D7000 and likely for the D800 as well, true base ISO is 100, so one will not see the advantage at L1.0 which you are seeing with the D300/s.

Marianne Oelund Veteran Member • Posts: 7,162
Additional "downsampled" SNR curves for D800

Absolutic wrote:

These Nikon D800 developers suggest shooting at Medium size (appx. 20MP) claiming it is a healthy equilibrium, in terms of general shooting. . . . Do you happen to also have charts for 20MP medium size?

Since there is a rather large number of size selections available from the camera (not to mention all the ratios you can use in post), here is a procedure for you to arrive at the SNR curve for any downsampling ratio:

d = 10 * LOG(36/n), where
d is the number of decibels to add to the D800 full-resolution SNR curve,
n is the number of megapixels you are downsampling to,
and LOG is base 10 logarithm (not natural logarithm).

For example, for 20Mpix, we have d = 10 * LOG(36/20) = 2.55dB. Then, just draw a new SNR curve parallel to the full-res SNR curve (which is the red curve on the plots that I posted), 2.55dB higher.

You can also use this procedure when downsampling from crops such as DX crop. Just replace the constant "36" with the number of Mpix in the cropped image that you start with.

As a final note, the 20MP option which the engineers mentioned, only applies to the camera's JPEG output. It is not available in RAW, where the only reduced options are the 1.2x (25Mpix), 5:4 (30Mpix) and DX (15Mpix) crops.

DSPographer Senior Member • Posts: 2,464
Re: Additional "downsampled" SNR curves for D800

Marianne Oelund wrote:

Absolutic wrote:

These Nikon D800 developers suggest shooting at Medium size (appx. 20MP) claiming it is a healthy equilibrium, in terms of general shooting. . . . Do you happen to also have charts for 20MP medium size?

Since there is a rather large number of size selections available from the camera (not to mention all the ratios you can use in post), here is a procedure for you to arrive at the SNR curve for any downsampling ratio:

d = 10 * LOG(36/n), where
d is the number of decibels to add to the D800 full-resolution SNR curve,
n is the number of megapixels you are downsampling to,
and LOG is base 10 logarithm (not natural logarithm).

This is correct if the noise is uncorrelated. This is the typical case, but fails for pattern noise and data that has been filtered to suppress noise. Examining the spatial frequency content of the noise can help identify these two cases.

 DSPographer's gear list:DSPographer's gear list
Canon PowerShot G7 X Canon EOS 5D Mark II Canon EF 24mm f/2.8 Canon EF 28-135mm F3.5-5.6 IS USM Canon EF 50mm f/1.8 II +4 more
Marianne Oelund Veteran Member • Posts: 7,162
We're not including pattern noise for the moment

DSPographer wrote:

This is correct if the noise is uncorrelated. This is the typical case, but fails for pattern noise and data that has been filtered to suppress noise. Examining the spatial frequency content of the noise can help identify these two cases.

Yes, my plots have excluded pattern noise, partly because I have no data for that yet (D800 and D4) and partly because it's only of practical significance at high signals and at low ISO.

So far, I've found no proof of filtering, which I'm quite pleased about.

bobn2
bobn2 Forum Pro • Posts: 56,299
Re: Histogram analysis needed

Marianne Oelund wrote:

Marius D wrote:

There is another possible explanation: 3) reducing ADC reference voltage (V.ref) to match the input analog domain as ISO raises. This should appear like a numerical scaling, except for reduced quantization errors. My guess is they are using both analog amplification and V.ref scaling. From around 1000 ISO up to 6400 there is no analog amplification, only V.ref scaling with a small but measurable contribution in reducing read noise due to the optimized quantization and other factors like better reset voltage reading or optical-black photo-detectors reading. Reset voltage and optical-black are both very low, reducing ADC's V.ref make sense for a precise digital conversion of these values.

The first step in determination will be to find the numerical scaling factors used for ISO settings above 1000. This will tell us how much, if any, of the gain is from the analog side of the ADC.

Bob's suggestion is interesting, although it would imply that without the added dithering, base ISO read noise is almost unbelievably low - around 0.7 ADU.

That is very low indeed, but not unbelievably, I would think. The best pixel designs (sat 1D IV) seem to be delivering 1.5 e- out of 50k, or about 0.5 14 bit ADU if the full range were used.
--
Bob

Marianne Oelund Veteran Member • Posts: 7,162
A pleasant surprise

Marius D wrote:

There is another possible explanation: 3) reducing ADC reference voltage (V.ref) to match the input analog domain as ISO raises. . . . Reset voltage and optical-black are both very low, reducing ADC's V.ref make sense for a precise digital conversion of these values.

I've completed the missing-code analysis for ISO 800, 1600, 3200 and 6400, so I now have the digital scaling values for those settings.

The surprise is that digital scaling at ISO 1600 is only 5.0%, which is much less than expected. Above 1600, the numerical scaling is precisely in proportion to the ISO value.

This clears up most of the mystery, leaving only a relatively small factor to explain at the highest ISO settings (6400+). I'm happy to leave that until later.

All of this implies that the D800 is truly exhibiting lower read noise than expected at ISO 1600 and up, and there is no indication of filtering, up to at least ISO 3200.

The closer we look at this camera, the more impressive it becomes - just as with its images.

nadjap Regular Member • Posts: 162
Re: We're not including pattern noise for the moment

Dear Marianne, thank You for Your work - even if I don´t understand every little bit - caused by my bad english

Marianne Oelund Veteran Member • Posts: 7,162
Plots updated

Marianne Oelund wrote:

All of this implies that the D800 is truly exhibiting lower read noise than expected at ISO 1600 and up, and there is no indication of filtering, up to at least ISO 3200.

The closer we look at this camera, the more impressive it becomes - just as with its images.

The SNR plots have been updated to reflect the better D800 read noise figures at high ISO, correct a typo in the previous-version spreadsheet (which affected D800 ISO 6400 curves) and to improve the color scheme for readability.

See
http://forums.dpreview.com/forums/read.asp?forum=1021&message=40756917

Note: You may need to clear your browser's cache, to see the new versions.

rhlpetrus Forum Pro • Posts: 24,916
Re: A pleasant surprise

Marianne Oelund wrote:

Marius D wrote:

There is another possible explanation: 3) reducing ADC reference voltage (V.ref) to match the input analog domain as ISO raises. . . . Reset voltage and optical-black are both very low, reducing ADC's V.ref make sense for a precise digital conversion of these values.

I've completed the missing-code analysis for ISO 800, 1600, 3200 and 6400, so I now have the digital scaling values for those settings.

The surprise is that digital scaling at ISO 1600 is only 5.0%, which is much less than expected. Above 1600, the numerical scaling is precisely in proportion to the ISO value.

This clears up most of the mystery, leaving only a relatively small factor to explain at the highest ISO settings (6400+). I'm happy to leave that until later.

All of this implies that the D800 is truly exhibiting lower read noise than expected at ISO 1600 and up, and there is no indication of filtering, up to at least ISO 3200.

The closer we look at this camera, the more impressive it becomes - just as with its images.

Very nice, thanks again. My next camera :)!

-- hide signature --

Renato.
http://www.flickr.com/photos/rhlpedrosa/
OnExposure member
http://www.onexposure.net/

Good shooting and good luck
(after Ed Murrow)

 rhlpetrus's gear list:rhlpetrus's gear list
Leica D-Lux (Typ 109) Nikon 1 V1 Nikon D7000 Nikon 1 Nikkor 10mm f/2.8 Nikon AF Nikkor 35mm f/2D +4 more
Luke Kaven Veteran Member • Posts: 5,705
Re: A pleasant surprise

Marianne Oelund wrote:

Marius D wrote:

There is another possible explanation: 3) reducing ADC reference voltage (V.ref) to match the input analog domain as ISO raises. . . . Reset voltage and optical-black are both very low, reducing ADC's V.ref make sense for a precise digital conversion of these values.

I've completed the missing-code analysis for ISO 800, 1600, 3200 and 6400, so I now have the digital scaling values for those settings.

The surprise is that digital scaling at ISO 1600 is only 5.0%, which is much less than expected. Above 1600, the numerical scaling is precisely in proportion to the ISO value.

This clears up most of the mystery, leaving only a relatively small factor to explain at the highest ISO settings (6400+). I'm happy to leave that until later.

All of this implies that the D800 is truly exhibiting lower read noise than expected at ISO 1600 and up, and there is no indication of filtering, up to at least ISO 3200.

The closer we look at this camera, the more impressive it becomes - just as with its images.

I woke up today wondering if you'd gotten to the bottom of this. Definitely a very pleasant surprise.

So if I understand correctly, locating the missing codes gave you the signature of the digital scaling function. The missing codes helped to confirm that there is no additional filtering (e.g., soft NR) occurring thereafter. If there had been soft NR or other filtering, then the missing codes would almost certainly have been smoothed out, is that it?

So does the hypothesis narrow down to gain-sharing? I wonder what advances would have favored that strategy. Can't wait to see the revised plots.

Luke Kaven Veteran Member • Posts: 5,705
Re: Plots updated

Marianne Oelund wrote:

Marianne Oelund wrote:

All of this implies that the D800 is truly exhibiting lower read noise than expected at ISO 1600 and up, and there is no indication of filtering, up to at least ISO 3200.

The closer we look at this camera, the more impressive it becomes - just as with its images.

The SNR plots have been updated to reflect the better D800 read noise figures at high ISO, correct a typo in the previous-version spreadsheet (which affected D800 ISO 6400 curves) and to improve the color scheme for readability.

Oh my god, that's amazing...amazing. The parity is much closer now, really very close by just these measurements.

Keyboard shortcuts:
FForum MMy threads