Luminous Landscape Forum
Equipment & Techniques => Digital Cameras & Shooting Techniques => Topic started by: Mark Regan on February 27, 2008, 12:49:25 pm
-
Hello,
I wonder why there is no 16 bits capture possible in Canons or Nikons?
It seems to be the main advantage in MF backs.
Why is there only 14 bits, which don't make a big difference to 12 bits?
-
The more bits you have the more data you have to move off the sensor and out to the buffer and memory card. Since the value of 14 bits over 12 bits is somewhat questionable there doesn't seem to be much incentive to have to move ~15% more data.
-
Hello,
I wonder why there is no 16 bits capture possible in Canons or Nikons?
It seems to be the main advantage in MF backs.
Why is there only 14 bits, which don't make a big difference to 12 bits?
[a href=\"index.php?act=findpost&pid=177743\"][{POST_SNAPBACK}][/a]
shoot raw and then export/save out of your raw app of your choice ( adobe raw, canon dpp, C1.. etc ) as a 16 bit tiff.
this is pretty much standard op even with MFDB's.
-
I wonder why there is no 16 bits capture possible in Canons or Nikons?
Let's turn around the question: what information should be stored in 16 bits?
Increasing the bit depth does not contribute to the increase of useful image data on its own.
It seems to be the main advantage in MF backs
MFDBs have higher dynamic range, which requires more gradations. So, it is logical, that their raw images have greater bit depth.
On the other hand, 16 bits are exaggerated. The MFDB owners should think,. that tehy received something valuable for that money, even though they don't utilize it.
Why is there only 14 bits, which don't make a big difference to 12 bits?
This is funny: you are asking, why not 16 bits instead of 14, while claiming, that 14 bits don't make much difference compared to 12 bits.
For your math: 14 bits are *four* times more (in terms of stored data) than 12 bits - just like 16 bits can hold four times more data than 14 bits.
Btw, the Nikon D300 and D3 allow recording 14 OR 12 bit raw data, and as far as I can see, most users stick to 12 bits, even though that numerical range is nor fully utilized.
I am happy with the Canon 40D's 14 bits; that is somewhat more than necessary, but you know, why there is no row 13 on airplanes, no floow nr 13 in many houses, etc?
-
The bit depth of of the AtoD needs to be high enough so that it has lower noise than the sensor. The newer sensors from Canon and Nikon are lower noise, hence they need a higher bit depth AtoD.
You can also, if you want, keep adding more bits on the AtoD, beyond what is necessary, and get no performance benefit. I don't know if anyone has done any real measurements on a MFDB, but I severely doubt them having a dynamic range that requires a 16 bit AtoD.
I do know of cameras in the digital cinema realm (no, not the one I work on) that record 16bits to file, but of those 2 are blank :-)
Graeme
-
It is interesting to note that the DSLR sensors are capable of more DR than MFDB sensors but either due to the A/D converters or circuitry the MFDB captures 2 stops more DR than the DSLR's?
Marc
http://www.clarkvision.com/imagedetail/dig...mary/index.html (http://www.clarkvision.com/imagedetail/digital.sensor.performance.summary/index.html)
-
It is interesting to note that the DSLR sensors are capable of more DR than MFDB sensors but either due to the A/D converters or circuitry the MFDB captures 2 stops more DR than the DSLR's?
Marc
http://www.clarkvision.com/imagedetail/dig...mary/index.html (http://www.clarkvision.com/imagedetail/digital.sensor.performance.summary/index.html)
[a href=\"index.php?act=findpost&pid=177767\"][{POST_SNAPBACK}][/a]
The medium format users brag about the dynamic range of their cameras, but this claimed DR does not show up in the specs. For example, Roger tabulates data for the KAF 18000CE, a Kodak chip targeted to medium format digital backs. It is a 4904 x 3678 chip with 9 micron pixels. The full well capacity is 100,000 electrons and the read noise is 18 electrons, giving an engineering DR of 12.4 f/stops (the Kodak spec is 74 db or 12.3 stops).
For comparison, the Canon 1DMIII has a full well of 70,200 electrons and a read noise of 4.0 electrons, giving a DR of 14.1 stops. Since the Kodak chip is higher resolution, the noise spectrum would be higher frequency (fine grained), which could affect the photographic DR. Any explanations? In the Kodak CCD chip, the ADC is external to the chip itself, but the DR of the sensor would be the limiting factor and a 16 bit ADC would not be able to increase the DR.
Bill
-
The bit depth of of the AtoD needs to be high enough so that it has lower noise than the sensor. The newer sensors from Canon and Nikon are lower noise, hence they need a higher bit depth AtoD.
They have not improved beyond what Canon has been doing for years, though, in terms of read noise at base ISO. The bottom line has been 1.25 12-bit ADUs (or equivalent), for which 12 bits is still satisfactory. You need to get down around 1.1 before you can see quantization effects. The Pentax K10D, AFAIK, is the only camera that really needs a 13th bit, at ISO 100, with a 12-bit read noise of 0.9 ADU.
You can also, if you want, keep adding more bits on the AtoD, beyond what is necessary, and get no performance benefit. I don't know if anyone has done any real measurements on a MFDB, but I severely doubt them having a dynamic range that requires a 16 bit AtoD.
The best I've seen there (P30) is about 1.25 12-bit ADU, also. No real need for more than 12 bits.
I do know of cameras in the digital cinema realm (no, not the one I work on) that record 16bits to file, but of those 2 are blank :-)
[a href=\"index.php?act=findpost&pid=177758\"][{POST_SNAPBACK}][/a]
If they compress those two bits away, then that shouldn't do any harm. A little LSB padding always affords the possibility of better handling by the converter.
-
It is interesting to note that the DSLR sensors are capable of more DR than MFDB sensors but either due to the A/D converters or circuitry the MFDB captures 2 stops more DR than the DSLR's?
http://www.clarkvision.com/imagedetail/dig...mary/index.html (http://www.clarkvision.com/imagedetail/digital.sensor.performance.summary/index.html)
[a href=\"index.php?act=findpost&pid=177767\"][{POST_SNAPBACK}][/a]
A quick look through the page, I can't find exactly what you're referring to, but he may be using sensor specs for cameras he doesn't have access to, or their RAWs, and mfr specs are not to be trusted, or considered relevant in a full camera system.
Generally speaking, MFDBs perform similarly to smaller-sensor cameras at the pixel level; they have more pixels which lowers the noise further, and increases the DR, at the image level.
-
For comparison, the Canon 1DMIII has a full well of 70,200 electrons and a read noise of 4.0 electrons, giving a DR of 14.1 stops.[a href=\"index.php?act=findpost&pid=177940\"][{POST_SNAPBACK}][/a]
That never actually happens. The read noise you quote is for ISO 3200, and the full-well is for ISO 50.
-
That never actually happens. The read noise you quote is for ISO 3200, and the full-well is for ISO 50.
[a href=\"index.php?act=findpost&pid=177950\"][{POST_SNAPBACK}][/a]
just curious
why would full well vary by ISO?
And does read noise vary much by ISO?
bob
-
That never actually happens. The read noise you quote is for ISO 3200, and the full-well is for ISO 50.
[a href=\"index.php?act=findpost&pid=177950\"][{POST_SNAPBACK}][/a]
Looking more closely at Roger's data, what you point out seems to be true. It does not make sense to use the read noise at ISO 3200 to calculate the DR at full well. Roger should revise his tables.
Bill
-
just curious
why would full well very by ISO?
And does read noise very much by ISO?
bob
[{POST_SNAPBACK}][/a] (http://index.php?act=findpost&pid=177971\")
The answer is that the camera gain (electrons per 12 bit data number) varies with ISO. Look at Table 1a on the [a href=\"http://www.clarkvision.com/imagedetail/evaluation-1d2/index.html]Clark[/url] web site. For the referenced camera at ISO 50 each camera data number represents 30.62 electrons and at ISO 3200 the gain is 3.93 electrons/12 bit data number.
Therefore, a least significant bit error at ISO 50 represents more electrons.
Bill
-
I'm not sure what it means but if I do a RAW conversion of a 5D file and a P30 file the P30 fills the histogram (16 bits) the 5D does not (12 bits)
I used ACR as a common converter with the same settings then I adjusted the whites to match.
Marc
5D
[attachment=5320:attachment]
P30
[attachment=5321:attachment]
-
I'm not sure what it means but if I do a RAW conversion of a 5D file and a P30 file the P30 fills the histogram (16 bits) the 5D does not (12 bits)
It means nothing on its own. If you post the raw images, we can talk about *those* histograms.
-
The bit depth of of the AtoD needs to be high enough so that it has lower noise than the sensor. The newer sensors from Canon and Nikon are lower noise, hence they need a higher bit depth AtoD.
You can also, if you want, keep adding more bits on the AtoD, beyond what is necessary, and get no performance benefit. I don't know if anyone has done any real measurements on a MFDB, but I severely doubt them having a dynamic range that requires a 16 bit AtoD.
I do know of cameras in the digital cinema realm (no, not the one I work on) that record 16bits to file, but of those 2 are blank :-)
Graeme
[a href=\"index.php?act=findpost&pid=177758\"][{POST_SNAPBACK}][/a]
Are you saying 12 bits can have more noise than 14 on higher ISO or lower light? If true that would be a reason for choosing 14bits on my D300.
-
Thanks for answering, even if some are way too complex for me, I think I get the idea.
-
Hello,
I wonder why there is no 16 bits capture possible in Canons or Nikons?
It seems to be the main advantage in MF backs.
Why is there only 14 bits, which don't make a big difference to 12 bits?
[a href=\"index.php?act=findpost&pid=177743\"][{POST_SNAPBACK}][/a]
Ignoring the noise versus signal debate. One of the greatest problems with Analogue to Digital converters is settling time of the input signal and also the amount of time it takes for the converter to actually 'compute' the digital value of the signal - The more bits of precision required then the longer it takes to do a conversion; the greater the precision required the longer you have to wait before the output of the channel amplifiers settle at the actual value of the input (pixel output) signal (slew rate limitation of the amplifiers).
For MF backs with typically low frame rates they can allow longer for the signal to settle and for the A2D converter to compute the code i.e. slowing the process down allows a greater bit depth. Nikon and Canon are pushing both sensor size (megapixels) and frame rate so, in order to live within the constraints of technology, need to limit the bit depth of the measured signal. If they reduced frame rate or sensor size then bit depth could be increased.
The other alternative is to run more A2D converters in parallel, which Canon and Nikon do to a certain extent. The problem then becomes ensuring consistency across the conversion pipeline (calibration of each converter) and cost (available silicon for implementation of the A2D converters).
Irrespective of noise if Canon/Nikon could make a marketing claim for 16-bit conversion without it costing too much they would do it, even if it didn't improve image quality.
-
It means nothing on its own. If you post the raw images, we can talk about *those* histograms.
[{POST_SNAPBACK}][/a] (http://index.php?act=findpost&pid=178000\")
Give me an hour and I'll post the raws through usendit
Marc
P30: [a href=\"http://download.yousendit.com/C223886E0467675D]http://download.yousendit.com/C223886E0467675D[/url]
5D: http://download.yousendit.com/3D785E3F6389F29D (http://download.yousendit.com/3D785E3F6389F29D)
-
My un-substantiated guess is that it has little to do with technology in fact.
The main reason is the existence of a carefully thought out release roadmap in which manufacturers have basically agreed to progress a a reasonnably slow pace in order to make sure that they will have appealing new features to deliver in their next generation bodies.
For high end users, bit rate is a characteristics that has marketing value, and manufacturers have IMHO understood that it was in their best shared interest to keep some bullets like this one availabe for the next round of fight.
What are the odds that Canon and Nikon, working on totally different technological/supplier communities, come up precisely at the same time with 14 bits sensors, nearly within days of each others?
This is the same reason why we are progressing mostly from 10MP to 12MP instead of going from 10MP to 14 or 16MP.
Cheers,
Bernard
-
just curious
why would full well very by ISO?
No. "Full well" is a feature of the sensor's photosites. It is how many electrons one has captured when it is full. A camera may or may not use the full range of the sensor at its lowest ISO.
And does read noise very much by ISO?
[a href=\"index.php?act=findpost&pid=177971\"][{POST_SNAPBACK}][/a]
Read noise as quoted in electrons may or may not decrease at higher ISOs, depending on the particular design. Most digital cameras have almost the same read noise in electrons at all ISOs, varying sometimes slightly by the contribution of the analog-to-digital converter and second-stage amplifier noises. Canon DSLRs and recent Nikon DSLRs have read noises that decrease rapidly, as measured in electrons, from the lowest ISO to about 4x to 8x that ISO, and decreasing more slowly, if at all, after 1600 or 3200.
Read noise as quoted in a percentage of maximum signal always increases with increasing ISO. With the cameras that have almost the same read noise in electrons at all ISOs, the read noise relative to maximum signal at any ISO is proportional to the ISO. For the Canon DSLRs and recent Nikons DSLRs, this read noise increases slowly at the low ISOs, and starts to almost double with a doubling of ISO between the higher ISOs.
-
What are the odds that Canon and Nikon, working on totally different technological/supplier communities, come up precisely at the same time with 14 bits sensors, nearly within days of each others?
Cheers,
Bernard
[a href=\"index.php?act=findpost&pid=178092\"][{POST_SNAPBACK}][/a]
Probably the same as Mclaren coming out with an F1 car using surprisingly similar technology to Ferrari within the same season
But then, Nasa didn't go to the moon either
Looks like we need a photo conspiracy blog. Manufacturers...they're just out to get you!
-
I'm not sure what it means but if I do a RAW conversion of a 5D file and a P30 file the P30 fills the histogram (16 bits) the 5D does not (12 bits)
1. The appearance of the histogram does not depend the very least on the bit depth.
2. I displayed both images with ACR; after WB on a grey square, resetting brightness and contrast to 0 and increasing the "exposure" of the P30 image by one EV, both histograms are filled.
Note, that ACR makes an automatic -1 EV adjustment to P30 and P45 images @ ISO 100, +1 @ ISO 400 and +2 @ ISO 800 (I don't know how this is @ ISO 200). I don't know the reason, but if you want to see the shot with the original exposure, you need to compensate for the automatic adjustment (it is not indicated anywhere by ACR!).
3. Several squares show clipping in both images.
Now, what about the raw data?
The first two attached images show the raw histograms.
The 5D image shows the perfect exposure. The exposure is maximized, but nothing clipped (apart a few stray pixels). Note the saturation level: 3692.
The green in the P30 images is significantly clipped: the white square (the bottom left) is completely clipped, *but nothing else*. The saturation level is 65535.
So, ACR's showing several squares as clipped is plainly incorrect - but why?
Let's see, what makes the ACR histogram sop different from the raw histogram?
*Note*: the following does not include
a. demosaicing,
b. color transformation from the camera's color space in sRGB or aRGB.The missing steps would cause other differences as well.
The first step is white balancing, on the middle grey square. This yields the following coefficients:
5D: red = 2.04, blue = 1.35
P30: red = 2.48, blue = 1.29
Now, let's see the histograms after WB application and mapping ("gamma encoding").
The third image is of the 5D. It looks good, except that there is some red clipping (the rightmost bump comes from the ominous white square of the checker). This red ended originally somewhere at 1830; after WB application it became 3733, and that is over the current white point (which started out with the saturation level).
The solution is either decreasing the brightness, or increasing the white point. I chose the latter one; see the fifth image, showing the histogram with white point 3942 - and there is no clipping any more.
The same could be achieved in ACR by either reducing the "exposure", or, much better, increasing the "recovery" (which does not recover anything in this case, because there is nothing lost).
Now, to the P30 shot. The fourth image shows the histogram after WB application and mapping. The bump of the white square vanished completely, i.e. it is cipped; thisis due to the WB application. Again the red dictates the upper limit: it ends around 29000, that with the WB coefficient 2.48 yields 72000. After increasing the white point to 72000 the red and blue clipping vanishes (see the sixth attached image). However, the green remains clipped: that is truy lost.
Finally, one note to the P30: it has a (small) weekness in comparison of the 5D. The difference in response between the red and the green is much higher with the P30. This results in a somewhat reduced dynamic range *in this lighting*.
-
I have noticed sometimes the P 30 clips the green channel even if the displayed histogram or blinkies don't show it, A quirk of the P30? I guess I have to be more conservative when I ettr.
Marc
-
1. The appearance of the histogram does not depend the very least on the bit depth.
2. I displayed both images with ACR; after WB on a grey square, resetting brightness and contrast to 0 and increasing the "exposure" of the P30 image by one EV, both histograms are filled.
Note, that ACR makes an automatic -1 EV adjustment to P30 and P45 images @ ISO 100, +1 @ ISO 400 and +2 @ ISO 800 (I don't know how this is @ ISO 200). I don't know the reason, but if you want to see the shot with the original exposure, you need to compensate for the automatic adjustment (it is not indicated anywhere by ACR!).
3. Several squares show clipping in both images.
Now, what about the raw data?
The first two attached images show the raw histograms.
The 5D image shows the perfect exposure. The exposure is maximized, but nothing clipped (apart a few stray pixels). Note the saturation level: 3692.
The green in the P30 images is significantly clipped: the white square (the bottom left) is completely clipped, *but nothing else*. The saturation level is 65535.
So, ACR's showing several squares as clipped is plainly incorrect - but why?
Let's see, what makes the ACR histogram sop different from the raw histogram?
*Note*: the following does not include
a. demosaicing,
b. color transformation from the camera's color space in sRGB or aRGB.The missing steps would cause other differences as well.
The first step is white balancing, on the middle grey square. This yields the following coefficients:
5D: red = 2.04, blue = 1.35
P30: red = 2.48, blue = 1.29
Now, let's see the histograms after WB application and mapping ("gamma encoding").
The third image is of the 5D. It looks good, except that there is some red clipping (the rightmost bump comes from the ominous white square of the checker). This red ended originally somewhere at 1830; after WB application it became 3733, and that is over the current white point (which started out with the saturation level).
The solution is either decreasing the brightness, or increasing the white point. I chose the latter one; see the fifth image, showing the histogram with white point 3942 - and there is no clipping any more.
The same could be achieved in ACR by either reducing the "exposure", or, much better, increasing the "recovery" (which does not recover anything in this case, because there is nothing lost).
Now, to the P30 shot. The fourth image shows the histogram after WB application and mapping. The bump of the white square vanished completely, i.e. it is cipped; thisis due to the WB application. Again the red dictates the upper limit: it ends around 29000, that with the WB coefficient 2.48 yields 72000. After increasing the white point to 72000 the red and blue clipping vanishes (see the sixth attached image). However, the green remains clipped: that is truy lost.
Finally, one note to the P30: it has a (small) weekness in comparison of the 5D. The difference in response between the red and the green is much higher with the P30. This results in a somewhat reduced dynamic range *in this lighting*.
[{POST_SNAPBACK}][/a] (http://index.php?act=findpost&pid=178168\")
This is an excellent demonstration that can be supplemented by Guillermo Luijk's discussion of [a href=\"http://www.guillermoluijk.com/tutorial/dcraw/index_en.htm]DCRaw[/url], even though Panopeeper's RawAnalyze tool can largely supplant DCRaw for this type of analysis.
Particularly important are the -H options in DCRaw, which control clipping of highlights in white balancing. It the white balancing multipliers are greater than 1 (as is usually the case) and integer math is used, one can have clipping from overflow. If the multipliers are less than 1, then there will be no clipping but one may have color shifts.
-H 0 forces at least one multiplier be equal to 1, and the rest will be greater or equal to 1
-H [1-9] forces at least one multiplier be equal to 1, and the rest will be less or equal to 1
Also important is the saturation command, -S; the following is a quote from from Guillermo's essay:
"However the saturation level -S is closely related to the behaviour of the RAW developer in the highlights so it is very interesting to be able to set it as DCRAW like any other RAW developer will use a standard saturation values table for each camera model, but this value could not be adequate for our particular unit:
* If our camera saturates the RGB channels in a level lower to that considered by DCRAW we could experiencience magenta cast issues with effects similar to those found in Fig. 6 because of channel missalignment.
* On the other side if our camera would saturate at levels higher than the standard value used we would be unnecessarily losing highlights information in the development process.
That is why it is a good idea to know exactly the saturation level of our camera's RGB channels and apply it to optimise the development. We shall take an example from a Canon 40D RAW file with highlight areas shot at ISO100:"
The saturation table may also be off with ACR, and this may explain its behavior with the P30.
Bill
-
I have noticed sometimes the P 30 clips the green channel even if the displayed histogram or blinkies don't show it, A quirk of the P30? I guess I have to be more conservative when I ettr.
Marc
[a href=\"index.php?act=findpost&pid=178179\"][{POST_SNAPBACK}][/a]
IIRC there are no separate gains for different ISO with this camera, rather simply a metadata tag that tells the raw converter what factor of two to multiply the raw data by before conversion (1/2 for ISO 100, 1 for ISO 200, 2 for ISO 400, etc). This resonates with panopeeper's comment about EV compensation that is applied in ACR. If the native ISO of the camera were 200, then ISO 100 is an overexposure by +1EV which is then subsequently pulled during raw conversion, which would explain the lack of headroom in the dynamic range. Consequently it would be better to shoot at ISO 200.
-
The saturation table may also be off with ACR, and this may explain its behavior with the P30
ACR's understanding of the P30 saturation is correct (it can not be missed easily, it's 65535).
However, ACR's saturation setting is off for several other cameras.
IIRC there are no separate gains for different ISO with this camera, rather simply a metadata tag that tells the raw converter what factor of two to multiply the raw data by before conversion (1/2 for ISO 100, 1 for ISO 200, 2 for ISO 400, etc).
That would be a too easy explanation. Although I don't have suitable images with the P30, but I do some with the P45. The situation is the same: an auto adjustment will be applied by ACR. However, those images demonstrate clearly, that the ISO setting is effective on the raw data. I would not vouch on the analog origin of that data (the histograms are incredible), but it is not obviously "uoresed". And, anyway, if the pixel values are stretched, as some cameras are doing it, then there is no reason to adjust the brightness, in fact that must not be done.
On the other hand, it is possible that some earlier Phase One camera acted that way, and Adobe has not noticed yet, that the ISOs are now true, or at least they look like true.