EOS-1D X Mark II Claims of 15 Stops of DR [CR3]

[3kramd5]

it's only 2/14ths larger.

So a 22MPixel 16 bit file will look much like a 19MPixel 14 bit file, assuming the same number of bits are thrashing at the bottom.

Forgive my ignorance. Ya lost me. You're saying a 16bit RAW file from 22MP sensor will have a similar file size to a 14bit RAW file from a 19MP sensor?? I'm missing something. How could that be smaller/same size?

16bits per pixel * 19,000,000pixel = 304,000,000bits
14bits per pixel * 22,000,000pixel = 308,000,000bits

 

[MrToes]

I'll believe it when I see it!

 

[heptagon]

I am not up on the binary bits part of it all. Is that how it works, you need 16 bit raw to record 16 stops dr?

Some enlightening information about Signal-to-Noise Ratio (SNR) and the Dynamic Range (DR):

1) The SNR is usually defined as the square of the signal amplitude (e.g. the number of photons detected in a pixel) divided by the square of the standard deviation of the noise on that pixel.

The following factors add up into that pixel noise:

* The poisson distribution of the number of detected photons.

* The random electronic thermal noise.

* Signal interference with other electronic components.

* Quantization errors.

2) To get the DR you ask the question: What is the highest amplitude that I can get and what is the amplitude with an SNR of 1? Then you divide them and get the DR.

Why is the SNR not equal to the DR?

The reason is the poisson noise. While the other noise factors mostly stay at the same level for dark and bright pixels on the same image (but change with different selected ISO values), the noise of a bright pixel is higher than the noise of a dark pixel. When you have a high amplitude, e.g. a high number of photons, there is a high variation in the actual number of photons in that pixel. This is simple poisson statistics and cannot be avoided.

The bright parts of your image are dominated by the poisson noise and all sensors are electronically very close to perfection there. Look up the SNR values at DXO and they are the same for all sensors with the same size.

Why is the DR of Sony sensors better than the DR of Canon sensors?

Sony has a better A/D-converter which reduces the electronic interference noise. Therefore dark pixels have lower noise. Bright pixels are virtually identical.

Why is the DR of Sony and Canon sensors equal at high ISO?

Canon amplifies the signal on the sensor before sending it to the A/D-converter, therefore the electronic interference is not as high in comparison. Here, the maximum brightness is limited.


Now to answer your question:

In a linear system, yes, you need 16 bits to represent the values needed for 16 stops of DR at the single pixel level.

However if you scale down the image you gain 1 bit of SNR and DR when you reduce width and height by a factor of 2. This even works when you use 14 bit for the RAW image and 16 bit for the digital processing afterwards. This is simple statistics. If you take a picture at 22 MP at 13 stops of DR and scale it down to 1.4 MP you get 15 stops of DR. Therefore HD-ready video with 15 stops of DR is totally believable with the current sensor technology.

One way to get more than 14 stops of DR out of 14 bits is to use a nonlinear A/D-curve and use digital processing to linearize it afterwards in 16 bits. This would reduce the SNR a bit but it's not really a problem because due to poisson noise you don't have 16 bits of SNR anyways. The only difficulty is to make a good low noise nonlinear amplifier.

So, yes, there are some tricks to get more than 14 stops of DR out of a 14 bits A/D-converter.

 

[PureClassA]

Oh! Ok. You originally meant a 22MP 14bit vs. 19MP 16 bit. I'm with ya.

it's only 2/14ths larger.

So a 22MPixel 16 bit file will look much like a 19MPixel 14 bit file, assuming the same number of bits are thrashing at the bottom.

Forgive my ignorance. Ya lost me. You're saying a 16bit RAW file from 22MP sensor will have a similar file size to a 14bit RAW file from a 19MP sensor?? I'm missing something. How could that be smaller/same size?

16bits per pixel * 19,000,000pixel = 304,000,000bits
14bits per pixel * 22,000,000pixel = 308,000,000bits

 

[heptagon]

I'll believe it when I see it!

This. So far everything is smoke and mirrors.

It starts to mean something to me if the sensor technology 1) really improves the images I make, 2) makes it into the 6DII and 3) comes at an affordable price point.

 

[jrista]

Well, if we assume 22mp, then at 14-bit:

Size = 22000000 * 14 / 8 = 38.5mb

And at 16-bit:

Size = 22000000 * 16 / 8 = 44.0mb

The uncompressed raw size of every light sensitive pixel on the sensor would require 5.5 additional megabytes per image. Not really all that much. This ignores the fact that the masked border pixels are also included, along with some metadata, so the difference might be 6mb in the end. Anyway, it shouldn't be a big deal going from 14 bits to 16 bits for a sensor of lower resolution like this.

Now, if we were talking 50mp, that might be a bit of a heftier growth in file size:

Size = 51600000 * 14 / 8 = 90.3mb
Size = 51600000 * 16 / 8 = 103.2mb

Gain of over 13mb per image there. Of course, with a 5Ds you would likely be taking far fewer shots at a lower frame rate, in the case of landscapes possibly as little as one shot per minute. With a 1D X II you could be taking 15 frames per second in regular bursts. If you regularly bring home a couple thousand shots with a 1D X, then you would need an additional 12 gigs to handle it at 16-bit.

At 16 bit, how much larger of a file are we talking here though? It's pretty substantial. I think Canon would give serious consideration to that. As previously mentioned, the full "15 stops of DR" Canon is boasting is likely something firmware driven that can/will only be recoverable within proprietary software like DPP. Once the RAW file gets imported and interpreted, it can be exported in a loss-less TIFF for LR or PS manipulation if desired. I'm not saying they won't go 16bit RAW files, just saying I'd be surprised if they did. The DR on current Canon bodies is more than satisfactory for most pros, however I understand there are SOME shots where more would help and that there are those who specialize in certain styles where more is commonly demanded. Canon isn't a panacea for everyone, but do serve the majority of pro togs very well. That being said, if you have a shot where you really need to squeeze every drop of highlight and shadow manipulation out of the widest image range possible on the new Canon models, you'll need to use DPP for those particular shots first. (I would suspect)

who's to say it's 14 bit RAW data?

canon was the first to go 14 bit, they could conceivably be the first to go to 16 bit (in the non medium format space that is).

Historically the first camera they upped the bit depth on was the 1 series (the 1D Mark III was the first with 14 bit RAW data).

there's alot of talk / whining / hand wringing and none of you explored the possibility that canon just simply went to 15/16 bit RAW files

*IF* canon implemented the ADC patents which are dual slope ADC - they are not ADC in the traditional sense, but a time to digital value. quantifying that could easily occur with a variety of bit depths.

 

[jrista]

For those who are interested, I use this SNR formula (well, a more complicated version, but this is the stuff that matters for daytime terrestrial photography) frequently in my astro work:

SNR = ImageSignal/SQRT(ImageSignal + DarkCurrent + ReadNoise^2)

While in a pure signal, noise is simply the SQRT(ImageSignal), in a digital camera we also have dark current and read noise. Dark current is negligible for most daytime astrophotography, although with older cameras (like the 5D II) even an exposure of a few seconds (for say long exposure water photography) would suffer from increased color noise and hot pixels. These days, DSLRs and mirrorless have very low dark current, so that term can usually be ignored.

That leaves the image signal and the read noise to be added in quadrature. When you have a strong signal, over midtone gray, read noise doesn't matter much. As you drop farther and farther below midtone gray, read noise can matter more and more, depending on how high it is. If you have a faint image signal of 100e-, and 30e- read noise:

SNR = 100/SQRT(100 + 30^2) = 100/SQRT(100 + 900) = 100/(SQRT(1000) = 100/31.62 = 3.15:1 ~= 10dB

If you reduce read noise to 3e-, on the other hand:

SNR = 100/SQRT(100 + 3^2) = 100/SQRT(100 + 9) = 100/SQRT(109) = 100/10.44 = 9.58:1 ~= 19.63dB

In terms of stops, the camera with 3e- read noise has a stop and a half advantage over the camera with 30e- read noise. It was demonstrated some time ago that the Canon sensors themselves seem to be capable of up to around 15.6 stops of DR, given the intrinsic noise that comes from electronics on the sensor itself. It's their downstream ADC units that seem to add the most noise. If Canon could resolve that downstream noise problem, then I don't doubt that they could deliver true, raw 15 stops of dynamic range.

The question is...have they fixed the ADC units and reduced the noise from them?

I don't think it is just as simple as moving the ADC units onto the sensor die. That allows increased parallelism, and lower ADC operating frequency. But that was not all that Sony Exmor or the CP-ADC Toshiba sensors did. They also improved other things, such as moving the clock generator to a remote area of the sensor to prevent it from introducing noise into the ADC units, by adding logic to tune each ADC unit to it's column and eliminate banding, by improving the signaling that drives all the circuitry, etc. Canon would have to do more than just move the ADC units onto the sensor die...but given recent patents, I think they have the technology to do what's necessary.

I just hope they actually DO...I would rather see a real hardware level gain in dynamic range, due to a reduction in read noise, than the use of some kind of artificial gain due to compressing the high dynamic range analog signal into a lower dynamic range 14-bit output via some kind of tone or compression curve.

 

[rrcphoto]

I don't think it is just as simple as moving the ADC units onto the sensor die. That allows increased parallelism, and lower ADC operating frequency. But that was not all that Sony Exmor or the CP-ADC Toshiba sensors did. They also improved other things, such as moving the clock generator to a remote area of the sensor to prevent it from introducing noise into the ADC units, by adding logic to tune each ADC unit to it's column and eliminate banding, by improving the signaling that drives all the circuitry, etc. Canon would have to do more than just move the ADC units onto the sensor die...but given recent patents, I think they have the technology to do what's necessary.

this gets lost in the noise. I don't think canon in one fell swoop is going to catch up to 2-3 generations of exmor sensors. I suspect this will take at least one more generation.

While they have the patents, it's alot to do in one generation, and those patents are all pretty new.

 

[jrista]

I don't think it is just as simple as moving the ADC units onto the sensor die. That allows increased parallelism, and lower ADC operating frequency. But that was not all that Sony Exmor or the CP-ADC Toshiba sensors did. They also improved other things, such as moving the clock generator to a remote area of the sensor to prevent it from introducing noise into the ADC units, by adding logic to tune each ADC unit to it's column and eliminate banding, by improving the signaling that drives all the circuitry, etc. Canon would have to do more than just move the ADC units onto the sensor die...but given recent patents, I think they have the technology to do what's necessary.

this gets lost in the noise. I don't think canon in one fell swoop is going to catch up to 2-3 generations of exmor sensors. I suspect this will take at least one more generation.

While they have the patents, it's alot to do in one generation, and those patents are all pretty new.

I agree that it probably won't happen in one fell swoop, however I disagree that it COULDN'T happen in one fell swoop. It happened like that before.

The patents were recently granted, but if you look at the dates on most of them, they are not new by any means, two years old or older (I think one was from 2011 even.)

Canon has a bad habit of sitting on lucrative technology. I've never understood it, but they have had patents for some pretty kick-ass technology since 2008...they just haven't employed it. At least, not in their DSLRs...some of the technology did find it's way into their smaller form factor sensors for P&S cameras, and those were the most advanced sensors Canon ever manufactured, using a 180nm process, copper interconnects, etc. There is some kind of lethargy in Canon's larger sensor division that just keeps them from moving forward at none other than a snails pace.

Canon fans often ridicule Nikon, which has their own business problems. However that has never kept them from pushing the technology envelope on every front at all times. The D500 is, in my humble opinion, a phenomenal camera. It's a refinement of technology on every front, not just the ergonomics or just the AF or just the sensor...it packs a ton of high end functionality at every level. I hear all the arguments about Canon needing to continue delivering the excellent customer service and reliability they always have...but if a company that is struggling even more than Canon and has a much smaller budget for R&D can push the technology envelope ever farther, why can't Canon? Canon has BILLIONS to play with...

Anyway. Someday I'm sure they will catch up. I would just prefer they do it in one fell swoop, because I believe they can.

 

[Proscribo]

this gets lost in the noise. I don't think canon in one fell swoop is going to catch up to 2-3 generations of exmor sensors. I suspect this will take at least one more generation.

While they have the patents, it's alot to do in one generation, and those patents are all pretty new.

I don't see why Canon couldn't do it if they wanted, as Sony isn't the only one who has good sensors. Actually, according to DXO, Toshiba and Samsung have BETTER sensors than Sony, altho they are APS-C, also Toshiba's sensors are now pretty much Sony's.

 

[AndreeOnline]

All this assumes integer calculations - which isn't a bad assumption at all. But it might be that Canon chooses to store the luminance levels as a floating-point number. It's more computational work, but it could definitely be done to reduce the data to a (say) 14 or even 12 bit number.

Well, obviously they don't "store the luminance levels as a number" at all. It's a transfer function that de-linearizes the raw data.

And I'm not sure if I'm reading you correct here, but I get the feeling you think that you could gain more DR by using a "finer scale". I see variants of this reasoning from time to time. That is why I tried to address it earlier and make it clear that you can't just "define more stops at the cost of fewer luminance levels per stop".

Anyway. There's little point discussing it here. The thread is already all over the place with some information mixed with lots of speculation. =)

To get back on an easier to follow line of thought, think of it this way:

EOS-1DX Mark II Claims 15 Stops of DR

• …
• …
• …
• …
• …and 15 stops of DR when using Canon Log 2 shooting videos in 4k mode.

 

[jrista]

All this assumes integer calculations - which isn't a bad assumption at all. But it might be that Canon chooses to store the luminance levels as a floating-point number. It's more computational work, but it could definitely be done to reduce the data to a (say) 14 or even 12 bit number.

Well, obviously they don't "store the luminance levels as a number" at all. It's a transfer function that de-linearizes the raw data.

And I'm not sure if I'm reading you correct here, but I get the feeling you think that you could gain more DR by using a "finer scale". I see variants of this reasoning from time to time. That is why I tried to address it earlier and make it clear that you can't just "define more stops at the cost of fewer luminance levels per stop".

Anyway. There's little point discussing it here. The thread is already all over the place with some information mixed with lots of speculation. =)

To get back on an easier to follow line of thought, think of it this way:

Aye, increasing bit depth when the camera is only capable of 12 stops of dynamic range only increases the number of levels of noise you can differentiate, but it doesn't really improve the dynamic range of the camera.

That said, personally, I prefer the smoother, finer grain of noise from a 16-bit camera than I do from a 12 bit camera. When you are limited to 12 bits, or even 14 bits, the noise grains are a bit harsher. With the full range of 16 bits, noise does look cleaner and smoother, and the results just look better IMO. Of course, this is going off of what I see with astro CCD cameras which are usually 16-bit. Those cameras tend to be quite expensive because of the high grade engineering and high quality signal processing, which results in nearly perfect gaussian characteristics:

In my testing, Canon DSLRs don't have a perfect gaussian noise...there are other characteristics that show up quite readily in an FFT:

 

[PureClassA]

I don't think it is just as simple as moving the ADC units onto the sensor die. That allows increased parallelism, and lower ADC operating frequency. But that was not all that Sony Exmor or the CP-ADC Toshiba sensors did. They also improved other things, such as moving the clock generator to a remote area of the sensor to prevent it from introducing noise into the ADC units, by adding logic to tune each ADC unit to it's column and eliminate banding, by improving the signaling that drives all the circuitry, etc. Canon would have to do more than just move the ADC units onto the sensor die...but given recent patents, I think they have the technology to do what's necessary.

I just hope they actually DO...I would rather see a real hardware level gain in dynamic range, due to a reduction in read noise, than the use of some kind of artificial gain due to compressing the high dynamic range analog signal into a lower dynamic range 14-bit output via some kind of tone or compression curve.

Your knowledge is always a marvel. Thanks. And having looked at all the amazing patents Canon has filed and published in recent years, I would tend to agree that there is no reason (apart from perhaps cost) they can't implement a major overhaul in one big step. And it's not just when they are filed, Canon would have been maturing those technologies for some time prior to even filing them. That's not really so much a measure in the timeline of where they are in terms of market-ready mass production. "Hey we know EXACTLY how to do it, we just have to drop a ton of time and money to make it happen." It's like Maeda said himself, even he is frustrated with how slow Canon moves.

I would have to assume if they are now producing the ADCs ON sensor die, to me that seems the biggest manufacturing cost obstacle to overcome. Seems like the rest is rather mundane in comparison. That begs the question again, "Well then, why wouldn't they?" Of course this is all academic until we get a damn camera...

 

[yeahright]

All this assumes integer calculations - which isn't a bad assumption at all. But it might be that Canon chooses to store the luminance levels as a floating-point number. It's more computational work, but it could definitely be done to reduce the data to a (say) 14 or even 12 bit number.

Well, obviously they don't "store the luminance levels as a number" at all. It's a transfer function that de-linearizes the raw data.

And I'm not sure if I'm reading you correct here, but I get the feeling you think that you could gain more DR by using a "finer scale". I see variants of this reasoning from time to time. That is why I tried to address it earlier and make it clear that you can't just "define more stops at the cost of fewer luminance levels per stop".

Anyway. There's little point discussing it here. The thread is already all over the place with some information mixed with lots of speculation. =)

To get back on an easier to follow line of thought, think of it this way:

Aye, increasing bit depth when the camera is only capable of 12 stops of dynamic range only increases the number of levels of noise you can differentiate, but it doesn't really improve the dynamic range of the camera.

That said, personally, I prefer the smoother, finer grain of noise from a 16-bit camera than I do from a 12 bit camera. When you are limited to 12 bits, or even 14 bits, the noise grains are a bit harsher. With the full range of 16 bits, noise does look cleaner and smoother, and the results just look better IMO. Of course, this is going off of what I see with astro CCD cameras which are usually 16-bit. Those cameras tend to be quite expensive because of the high grade engineering and high quality signal processing, which results in nearly perfect gaussian characteristics:

In my testing, Canon DSLRs don't have a perfect gaussian noise...there are other characteristics that show up quite readily in an FFT:

the fact that the FFT-spectrum is not flat means that the noise is not WHITE, but not that it is not GAUSSIAN.

 

[rfdesigner]

(2D FFTs)

For those that don't understand 2D FFTs. The ideal response to an image of noise is a flat noise plane with a single white dot in the middle (representing the overall average rightness). Low frequencies are in the centre, high frequencies are towards the edge.

If there are dots or stripes in the FFT space then that means there are repeating patterns in the image space, even if you can't seem them, which is often the case.

 

[FEBS]

So, yes, there are some tricks to get more than 14 stops of DR out of a 14 bits A/D-converter.

What you explain is indeed the base of oversampling. To get a picture 22Mb @ 15bit DR instead of 14 bit you need an effective sensor of 22Mb x 4 = 88Mb sampled by a 14bit Analog to Digital Convertor. Then by scaling down, and using a good software low-pass filter you can get 22Mb @ 15bit in theory as in practice we also don't get 14stops of DR from a 14bit convertor due to the noise.

But 88Mb sensor and a 14bit ADC does not seems the way to me how Canon will get the 15 stops of DR.

 

[jrista]

All this assumes integer calculations - which isn't a bad assumption at all. But it might be that Canon chooses to store the luminance levels as a floating-point number. It's more computational work, but it could definitely be done to reduce the data to a (say) 14 or even 12 bit number.

Well, obviously they don't "store the luminance levels as a number" at all. It's a transfer function that de-linearizes the raw data.

And I'm not sure if I'm reading you correct here, but I get the feeling you think that you could gain more DR by using a "finer scale". I see variants of this reasoning from time to time. That is why I tried to address it earlier and make it clear that you can't just "define more stops at the cost of fewer luminance levels per stop".

Anyway. There's little point discussing it here. The thread is already all over the place with some information mixed with lots of speculation. =)

To get back on an easier to follow line of thought, think of it this way:

Aye, increasing bit depth when the camera is only capable of 12 stops of dynamic range only increases the number of levels of noise you can differentiate, but it doesn't really improve the dynamic range of the camera.

That said, personally, I prefer the smoother, finer grain of noise from a 16-bit camera than I do from a 12 bit camera. When you are limited to 12 bits, or even 14 bits, the noise grains are a bit harsher. With the full range of 16 bits, noise does look cleaner and smoother, and the results just look better IMO. Of course, this is going off of what I see with astro CCD cameras which are usually 16-bit. Those cameras tend to be quite expensive because of the high grade engineering and high quality signal processing, which results in nearly perfect gaussian characteristics:

In my testing, Canon DSLRs don't have a perfect gaussian noise...there are other characteristics that show up quite readily in an FFT:

the fact that the FFT-spectrum is not flat means that the noise is not WHITE, but not that it is not GAUSSIAN.

Well, the 5D III is neither gaussian nor white, really. Here is the actual noise (not the FFT) and it's histogram:

Here is the actual noise and histogram for the QSI sample:

The standard deviation and characteristic of the noise from the QSI, which is also 16-bit, is significantly cleaner. A D800/D810 has noise much closer to the QSI than to the 5D III, for what it's worth. The standard deviation is larger, but not that much larger. Anyway, this kind of noise quality is why astrophotographers spend the big bucks on a high quality CCD camera.

 

[deleteme]

We’re told that Canon will claim 15+ stops of dynamic range for the new Canon EOS-1D X Mark II. This claim was also made for the Cinema EOS C300 Mark II, which may be true at the hardware level, but in practice it may not actually perform to that specification. Let’s hope it’s actually the case with the new 22mp sensor.</p>
<p>Specifications for this camera have been extremely slow to come in from known sources. There is definitely a much tighter ship being run at Canon, but we do expect more to leak out as we approach an announcement in the next 4-8 weeks.</p>
<span id="pty_trigger"></span>

It was inevitable that the DR bragging race would be an integral part of marketing.
Now we have to see if it really exists.

 

[PureClassA]

It was inevitable that the DR bragging race would be an integral part of marketing.
Now we have to see if it really exists.

That's why we're here... :

 

[PhotographyFirst]

Hasn't it already been shown that Canon sensors could produce 15 stops of DR for a long time now?

It's really just a matter of extracting that 15 stops of DR and getting it into a file for the memory card, which they have not done as of yet.

I am betting a large sum of internet points that if Canon becomes the new stills DR leader, the next point of contention for hating Canon will probably be the new automated AFMA system Nikon now has. Once that hits the streets, everything less will be incapable of taking even mediocre photos.