Does Higher Bit-depth Matter?

[Panopeeper]

(I believe) the resulting RAW files do not have the same tone curve.  (I don't think they *are* 100% linear in RAW space...)

1. The raw data does not have any tone curves,

2. The individual color channels are originally linear. The interpretation by the raw processor can (and sometimes does) introduce some small degree of non-linearity, but you have no way of compensating for that, i.e.  you have no choice but regard the result as linear and curse if it is not.

In order to verify this, shoot different shades of the *very same gray* (like on a color checker), pick WB on the white and check how gray the darker ones are.

[Mort54]

Dear Mort and friends,
One should do side by side comparison between Canon 1dsm3 14 BIT and 16 bit Phase One P30+ of the same scene containing bright highlights and deep shadows with lots of colors (fruit and vegetables), in 100 ASA, in raw, same development, and should show enlargments 100% of side by side same area in order to show and convince which has better DR, lower noise and clearer distinct details in shadows and highlights, instead of bla, bla, bla.....and make unpopular the one that wants to know what camera to choose for his own jobs. so simple and elementary.
Declarations, assumptions and theories of self assurance, one sided photographers, are not a substitute of scientific real tests that everyone can see and judge.

I could care less what camera someone wants to use. That's a personal choice based on their own particular needs. If you think your 1DsIII is better than a MFDB, then great, enjoy your camera. I personally have both a DSLR and a MFDB, and enjoy using both. They each have their strengths and weaknesses.

As for proving anything to you or anyone else, why should we? You're the one coming to the MF forum and making these claims. The burden of proof is on you. Why don't YOU put together side by side shots, of fruit and such, and show 100% crops, bla, bla, bla. Most of the people on here who have MFDBs also have DSLRs, so we already know what out backs and !dsIII's and D3's will and won't do, and have a pretty good feel for how they compare. So the ball is in your court. If you think DSLRs are as good as or better than MFDBs, feel free to prove it. Or better yet, just put what you have to good use and don't dwell on which is better at this or that.

Regards,
Mort.

[ejmartin]

Again if you want to be scientifically precise and not comparing photographers ability,  light difference, dof and lens difference, than:
One should do side by side comparison between Canon 1dsm3 14 BIT and 16 bit Phase One P30+ of the same scene containing bright highlights and deep shadows with lots of colors (fruit and vegetables), in 100 ASA, in raw, same development, and should show enlargments 100% of side by side same area in order to show and convince which has better DR, lower noise and clearer distinct details in shadows and highlights, instead of blah, blah, blah.....and make unpopular the one that wants to know what camera to choose for his own jobs. so simple and elementary.
I agree to add to the side by side set: a colorchecker chart, GM is fine. three stops overexposed and three stops underexposed will add to see DR
One can shoot the  Phase One P30+ in f: 4 and the 1dsm3 in F:2.8 (compensating for exposure) in one test in order to see 3D effect on both cameras (putting distant and suitable, nice background full of out of focus items).
Declarations, assumptions and theories of self assurance, one sided photographers, are not a substitute of scientific real tests that everyone can see and judge.
Thanks,
Menachem Reiss. www.reiss.co.il
[a href=\"index.php?act=findpost&pid=205667\"][{POST_SNAPBACK}][/a]

I am after measuring the pixel "gain" (photons/raw level) and read noise.  From these, S/N characteristics as a function of exposure and dynamic range are entirely determined.  One does not need to have the cameras side by side in order to measure these quantitative properties.  I spelled out the test conditions so that photographic ability is not part of the equation -- unless someone has trouble getting an out-of-focus image of a GMCC chart ;-)

As I mentioned, this exercise has already been done for the relevant DSLR's (Canon 1 series and Nikon D3).   Perhaps it has been done as well for some MFDB's, but if so I have not seen the relevant data reported anywhere.  

As for the supposed "3D effect", I'm not trying to isolate that.

[bjanes]

I am after measuring the pixel "gain" (photons/raw level) and read noise.  From these, S/N characteristics as a function of exposure and dynamic range are entirely determined.  One does not need to have the cameras side by side in order to measure these quantitative properties.  I spelled out the test conditions so that photographic ability is not part of the equation -- unless someone has trouble getting an out-of-focus image of a GMCC chart ;-)

As I mentioned, this exercise has already been done for the relevant DSLR's (Canon 1 series and Nikon D3).   Perhaps it has been done as well for some MFDB's, but if so I have not seen the relevant data reported anywhere. 

[a href=\"index.php?act=findpost&pid=205767\"][{POST_SNAPBACK}][/a]

Kodak and Dalsa do publish the specs for many of the sensors used in MFDBs. For example, the Kodak KAF 39000 has a pixel spacing of 6.8 microns, full well capacity of 60K electrons, and a read noise of 16 electrons at a certain reading speed and unspecified ISO. The quoted read noise is not particularly impressive and it would be interesting to see the KAF 39000 tested under the same conditions as the Nikon and Canon sensors you mention. The KAF 39000 has more pixels than Canon or Nikon, but it remains to be seen if the great dynamic range boasted by the MFDB crowd is real or imagined.

Of course the ADC is external to the sensor in a CCD design, and the digital back maker could use a low read speed and a high bit, high quality ADC to improve performance. One doesn't shoot at 10 frames/second with a MFDB.

Bill

[ejmartin]

Kodak and Dalsa do publish the specs for many of the sensors used in MFDBs. For example, the Kodak KAF 39000 has a pixel spacing of 6.8 microns, full well capacity of 60K electrons, and a read noise of 16 electrons at a certain reading speed and unspecified ISO. The quoted read noise is not particularly impressive and it would be interesting to see the KAF 39000 tested under the same conditions as the Nikon and Canon sensors you mention. The KAF 39000 has more pixels than Canon or Nikon, but it remains to be seen if the great dynamic range boasted by the MFDB crowd is real or imagined.

Of course the ADC is external to the sensor in a CCD design, and the digital back maker could use a low read speed and a high bit, high quality ADC to improve performance. One doesn't shoot at 10 frames/second with a MFDB.

Bill
[{POST_SNAPBACK}][/a]

It would indeed be interesting if someone were to do the legwork on the sensors (hint     ).  Who uses the KAF 39000?  The sensor DR bounds the overall DR; already with this sensor 39000/16=3750~11.9 stops, realized DR after amplification and ADC quantization will surely be less.  16 bits is really of no help there, except insofar as one wants to cut down on quantization error when all ISO is implemented in software during post-processing.

BTW, if we are talking purely about sensor properties, according to Bart van der Wolf
[a href=\"http://www.openphotographyforums.com/forums/showthread.php?t=4784]http://www.openphotographyforums.com/forum...read.php?t=4784[/url]
http://www.openphotographyforums.com/forum...read.php?t=4771
the 1Ds3 sensor has a full well capacity of about 57000 electrons, and a read noise of about 5 electrons (best measured at ISO 1600, where components downstream of the sensor contribute negligibly to read noise).  Thus the 1D3 sensor's DR is 57000/5=11400~13.5 stops.  

Now, to be fair, the noise in the downstream components is such that Canon is unable to realize this DR at base ISO in practice, losing about two stops in read noise due to amplifier and ADC noise at ISO 100 (though I think they could if they were a little smarter about the design, by running two amplifiers in parallel and then combining the results, as I have described elsewhere).

There is also the issue of how quickly the sensor climbs up out of the read noise dominated territory, which bears on shadow IQ, and that is where larger formats really come into their own.

[bjanes]

the 1Ds3 sensor has a full well capacity of about 57000 electrons, and a read noise of about 5 electrons (best measured at ISO 1600, where components downstream of the sensor contribute negligibly to read noise).  Thus the 1D3 sensor's DR is 57000/5=11400~13.5 stops. 

[a href=\"index.php?act=findpost&pid=206037\"][{POST_SNAPBACK}][/a]

Emil,

When calculating the DR at full well, why do you use the read noise for ISO 1600? Or is the read noise at base ISO that low?

Bill

[ejmartin]

Emil,

When calculating the DR at full well, why do you use the read noise for ISO 1600? Or is the read noise at base ISO that low?

Bill
[{POST_SNAPBACK}][/a]

I used to think this was goofy, but now I know better (and finally understand what Roger Clark has been trying to say for a long time).  The read noise in absolute terms (eg in photoelectron equivalents) is much higher at base ISO, but this fact is due to the electronics downstream of the sensor (the variable gain amplifier that implements ISO, and the ADC) rather than the noise of the sensor capture itself.  One can see that from the plots of read noise vs ISO at

[a href=\"http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p2.html#read_vs_iso]http://theory.uchicago.edu/~ejm/pix/20d/te...tml#read_vs_iso[/url]

See in particular the little table below fig 15.  The sensor noise is amplified by the ISO amplification, which is why high ISO read noise is large in ADU, but low when referred back to photoelectrons.  The low ISO read noise is small in ADU, but large when referred back to photoelectrons; the amplifier/ADC noise is independent of ISO when denominated in ADU.  So the dominant contribution to read noise at low ISO comes from the ISO amplifier and ADC, while at high ISO it is almost entirely due to the photosite electronics.  What this means is that the read noise at low ISO is being limited by the amplifier/ADC components, since if they contributed less than the sensor itself, one would get the full DR of the sensor.

This is why Roger Clark is careful to distinguish what he calls "sensor DR" from "camera DR".  The former is about 14 stops, while the latter is a bit under 12 stops for the 1D3/1Ds3.  The difference is due to the quality/capability of the electronics downstream of the sensor in Canon cameras (and also the D3/D300, I should add).

One then might ask, so what?  Won't the electronics downstream of the sensor always limit the DR in this way?  I don't think so.  In principle one could take the sensor output and run it through two parallel signal processing routes -- one set to ISO 100 and the other set to ISO 1600 (think of it as "in-camera HDR").  The low ISO path captures the highlights, but has  four times the read noise in electrons.  The high ISO path has the highlights clipped, but that's OK we're going to take those from the other path anyway; instead, the high ISO path captures the shadows with much lower read noise in electrons.  After digitisation, the two paths can be combined in much the same way one combines two exposures in HDR, one for highlights and the other for shadows.  It would be trivial to code this, the only cost is the extra hardware for the dual amplifications.  And since one is always capturing the full DR that the sensor can deliver, there is no reason whatsoever to have variable gain; ISO can be implemented as it is in many MFDB's, as a metadata tag for the raw converter to apply.  The difference is that one would actually have about 14 stops of DR and high bit depth would actually be justified for a change.

[Panopeeper]

In principle one could take the sensor output and run it through two parallel signal processing routes -- one set to ISO 100 and the other set to ISO 1600 (think of it as "in-camera HDR

Don't you think this *is* the way most MFDBs work? (The Phase One P45+ is rather the exception.)

This is, what I meant earlier, when saying that the lack of different ISO gains is an advantage, although some MFDB owners are upset by the "accusation", that their ISO selection is placebo.

I want a cropping camera with "uni-ISO".

[ejmartin]

Don't you think this *is* the way most MFDBs work? (The Phase One P45+ is rather the exception.)

This is, what I meant earlier, when saying that the lack of different ISO gains is an advantage, although some MFDB owners are upset by the "accusation", that their ISO selection is placebo.

[a href=\"index.php?act=findpost&pid=206074\"][{POST_SNAPBACK}][/a]

I have no evidence one way or the other as to whether MFDB's use such dual processing.  It would be easy to diagnose in the raw data however, I just haven't had enough examples to examine as yet.  I would be surprised however if they used this sort of dual amplification scheme; it's more likely that they spend the money to have lower noise amplifiers, higher DR ADC's etc so as to make such a scheme superfluous, with the DR limited by the sensor.  It's not so hard if the sensor DR is less than twelve stops anyway.

I do agree that, insofar as the signal processing chain is limited by the sensor's DR rather than that of the rest of the electronics, there is no need to have variable ISO gain as one does on DSLR's currently.  Variable gain is not needed if the raw data contains all the data that the sensor is recording, regardless of exposure.

[joofa]

there is no need to have variable ISO gain as one does on DSLR's currently.  Variable gain is not needed if the raw data contains all the data that the sensor is recording, regardless of exposure.
[a href=\"index.php?act=findpost&pid=206084\"][{POST_SNAPBACK}][/a]

Analog ISO gain is useful as it provides the "dithering" via noise amplification -- something similar to the reduction of perception of harmonic distortion in audio ADC is accomplished by adding random dither signal before digitization. OTOH, compensating a lack of analog gain by digital gain (say software) results in posterization in certain low light scenes.

[bjanes]

I used to think this was goofy, but now I know better (and finally understand what Roger Clark has been trying to say for a long time). 

This is why Roger Clark is careful to distinguish what he calls "sensor DR" from "camera DR".  The former is about 14 stops, while the latter is a bit under 12 stops for the 1D3/1Ds3.  The difference is due to the quality/capability of the electronics downstream of the sensor in Canon cameras (and also the D3/D300, I should add).

[a href=\"index.php?act=findpost&pid=206070\"][{POST_SNAPBACK}][/a]

Emil,

Thanks for the detailed explanation. It all makes good sense. However, when one is taking pictures, it is the camera DR that is most important. When Kodak rates the KAF 39000 DR at 71.4 dB or 11.9 stops, that is the sensor DR and the camera DR would be less. The question is: less by how much with the best possible electronics downstream from the sensor.

Bill

[ejmartin]

Analog ISO gain is useful as it provides the "dithering" via noise amplification -- something similar to the reduction of perception of harmonic distortion in audio ADC is accomplished by adding random dither signal before digitization. OTOH, compensating a lack of analog gain by digital gain (say software) results in posterization in certain low light scenes.
[{POST_SNAPBACK}][/a]

As long as the noise sufficiently exceeds the quantization step, posterization will not occur:

[a href=\"http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p3.html]http://theory.uchicago.edu/~ejm/pix/20d/te...e/noise-p3.html[/url]

The dithering provided by the sensor electronics will be quite enough, provided the bit depth of the ADC is chosen appropriately; there is no need to introduce additional noise downstream of the sensor, it just further degrades the signal.

[joofa]

As long as the noise sufficiently exceeds the quantization step, posterization will not occur:

http://theory.uchicago.edu/~ejm/pix/20d/te...e/noise-p3.html

The dithering provided by the sensor electronics will be quite enough, provided the bit depth of the ADC is chosen appropriately; there is no need to introduce additional noise downstream of the sensor, it just further degrades the signal.
[a href=\"index.php?act=findpost&pid=206172\"][{POST_SNAPBACK}][/a]

Thanks for the link. You have good information. The section entitled "An aside on "lossy" NEF compression" seems like the time-honored analog-domain well-known techniques of applying concave non-linearities to signal to boost signal to noise ratio over linear quantization.

Back to topic. Unfortunately, one does not always have the choice of matching bit depth of ADC to control the posterization. We have seen standard sensors do bad posterization under low-light condition and analog amplification of noise via ISO gain has been our good friend vs. digital ISO gain.

[Panopeeper]

Unfortunately, one does not always have the choice of matching bit depth of ADC to control the posterization

What about the new Nikons with selectable bit depth?

Though I have the feeling, that internally it is always 14bit.

We have seen standard sensors do bad posterization under low-light condition and analog amplification of noise via ISO gain has been our good friend vs. digital ISO gain

Are you referring to the "plain" digital gain (multiplying the pixel values by some constant), or do you think that the amplified noise is inherently better than for example two randomly generated low order bits?

[joofa]

What about the new Nikons with selectable bit depth?

I have to look into the new Nikons that you mention.

Are you referring to the "plain" digital gain (multiplying the pixel values by some constant), or do you think that the amplified noise is inherently better than for example two randomly generated low order bits?
[a href=\"index.php?act=findpost&pid=206301\"][{POST_SNAPBACK}][/a]

Adding noise after quantization is not as effective as adding noise before quantization. Hence, randomly generated low order bits may not give as good a response as higher analog gain (for noise amplification) before quantization.

[Panopeeper]

Adding noise after quantization is not as effective as adding noise before quantization

Why not? What makes one random value more effective than another one?

Btw, what is the measure of "effectivity" in this regard?

[ejmartin]

Thanks for the link. You have good information. The section entitled "An aside on "lossy" NEF compression" seems like the time-honored analog-domain well-known techniques of applying concave non-linearities to signal to boost signal to noise ratio over linear quantization.

Back to topic. Unfortunately, one does not always have the choice of matching bit depth of ADC to control the posterization. We have seen standard sensors do bad posterization under low-light condition and analog amplification of noise via ISO gain has been our good friend vs. digital ISO gain.
[{POST_SNAPBACK}][/a]

Do you have an example to show?  It is possible that if there is posterization, it is being introduced by the algorithm of raw conversion rather than being intrinsic to the raw data itself.  For instance, raw data that is not posterized can be made posterized by median filtering, a common technique for reducing noise:

[a href=\"http://forums.dpreview.com/forums/read.asp?forum=1021&message=28139198]http://forums.dpreview.com/forums/read.asp...essage=28139198[/url]

[ejmartin]

Why not? What makes one random value more effective than another one?
[{POST_SNAPBACK}][/a]

Adding noise to posterized data makes noisy posterized data.  Noise present before quantization dithers the tonal transitions and eliminates the posterization.

It's easy enough to check this for yourself.  Make a tonal gradient in photoshop.  Add N levels of gaussian noise.  Now try to posterize it via two successive levels adjustments, the first reducing the maximum output level to 255/N, the second setting the max input level to 255/N (this has the effect of approximately dividing the levels by N and then multiplying them back up by N; the result in integer math that photoshop uses is to increase the gap in the level spacing by a factor N).  Then do the same exercise, adding the noise after the two levels adjustments.  Adding noise after the posterization is MUCH less effective in masking the posterization.

Why is this so?  Because adding the noise after quantization does not change the mean of the pixels in a given posterized region.  Adding the noise before quantization results in a smooth modulation of the percentage of pixels having one quantized level vs the next one; the eye averages over spatial regions and the average of the pixel values reflects the original smooth tonal variation over space.  And this is the way dithering works -- it trades spatial resolution for smoother tonal transitions.

[a href=\"http://en.wikipedia.org/wiki/Dithering]http://en.wikipedia.org/wiki/Dithering[/url]

[joofa]

Why not? What makes one random value more effective than another one?

In regions of low-luminance gradients or contours, the input random noise causes the pixels to go above or below the original decision level (via quantization). The effects is to "break" the contouring. However, the averaged value of the quantized pixels is about the same with and without the additive noise.

If noise is not added before quantization then higher levels of added noise after quantization may be required to hide the contouring resulting in overall worse picture.

Btw, what is the measure of "effectivity" in this regard?
[a href=\"index.php?act=findpost&pid=206314\"][{POST_SNAPBACK}][/a]

Measured as the trade off between the output variance in intensity from the mean value of the output in a portion of the picture which was of constant intensity at the input averaged over all inputs, and the deviation of mean output from the input.

[joofa]

Do you have an example to show? 

Unfortunately, I can't post them here.