Less Pixels More Bit Depth...

[Graeme Nattress]

REDCODE RAW also helps in keeping the files manageable, but yes, if you're going uncompressed, at least pack the bits! (Like a 10bit dpx does) or use some simple lossless compression.

Linear means 1 bit per stop. Yes, that's correct.

Graeme

 

[Michael Lindsay]

Hi Graeme

In theory is it possible to differently distribute encoding and still keep the other advantages of Raw?

thanks for the education.

Michael L

 

[Graeme Nattress]

You need a non-linear amp ahead of the a-to-d to do that, which can cause more problems that it's worth. It's also worth remembering that things like colour transform matrices need to work on linear light data, and thus you must have a reliable and accurate way to get back to linear if you make it non-linear.

Graeme

 

[Michael Lindsay]

Graeme

Thanks ..

Michael

 

[Daniel Browning]

I would prefer *more* pixels and *less* bit depth

I would prefer *more* pixels and *less* bit depth

I hope I don't annoy too many folks with the size of this post. If you were having trouble sleeping, maybe I can help with that, at least.

But what I and my trusty DI colorist want more than more pixels is--as the title of this post suggests--MORE BIT DEPTH.

I don't think you want more bit depth by itself: that just wastes recording space on random noise. To make use of 16 bits would require at least 15 stops of dynamic range, so that's what you're asking for: 4 more stops of dynamic range (AKA an 800% improvement, quite a tall request).

The 12-bits in the RED ONE are plenty sufficient to encode everything the camera can deliver: 11.3 stops of dynamic range. More bits would just bloat file sizes with no benefit.

In fact, most people don't use all 11.3 stops, because the last few stops are pretty noisy. If you only use 9.5 stops, you could do just fine with a 10-bit R3D and save a lot of room in the file size. (It would be nice if RED allowed photographers to choose lower precision recording to save space, but many customers would probably aim directly for their foot and shoot, so it probably wont happen).

If all RED ONE cameras got a free 16-bit ADC and 16-bit REDCODE upgrade tomorrow, it would not affect image quality at all, the only change would be bigger files sizes. If there was a 10-bit ADC that had less noise than the existing 12-bit ADC, the 10-bit would provide a better image. It would only be 9.6 stops, but they would be less noisy. But I suspect RED built the lowest-noise ADC they could.

Of course, for the highest quality, post-processing should occur with greater precision than the original file. 16-bit int or even 32-bit float will give slightly better results than grading in 12-bit lin or 10-bit log, particularly for bigger changes.

So what you probably intended to ask for is more dynamic range. If you're using all 11.3 stops now, and RED increases dynamic range per pixel by two stops in the Mysterium-X, to 13.3 stops, then a 14-bit recording would be sufficient to capture all the information put out by the sensor, and you can use all 13.3 stops.

But if you only use 9.3 stops right now with RED ONE because the last two stops are too noisy for your taste, then you only need 10 bits right now: the last two are being clipped to black. If RED increases dynamic range per pixel by two stops, and you want to start using 11.3 stops (out of the 13.3 total), then 12-bit will still be enough to capture all the information that you will use.

So there are three things:

• Number of stops the camera is capable of.
• Number of stops that most people will use.
• Bit precision needed to encode what the camera is capable of.
• Bit precision used by the ADC
• Bit precision recorded to the file.
• Bit precision used in post production

For RED ONE we have:

• Number of stops the camera is capable of: 11.3
• Number of stops that most people will use for 4K DI: 9.6
• Bit precision needed to encode what the camera is capable of: 12
• Bit precision needed to encode what most people will use for 4K DI: 10
• Bit precision used by the ADC: 12
• Bit precision recorded to the file: 12
• Bit precision used in post production: 10-32

It's important to know that the precision of the ADC is separate from what the camera is capable of using, and also what gets recorded. The Pentax K10D had a 22-bit ADC, but was only capable of 14-bits and only recorded 14-bits. (Actually, there were some reports of slight posterization, but the fixed pattern noise was so much stronger than the random read noise that most could not use all 14-bits anyway).

Essentially, as long as the ADC has enough precision for the camera, it doesn't matter how much higher it is (it could be 1000 bits and it wouldn't matter), you still only want to record enough precision that the entire signal is preserved, and then some (one half-stop of totally random noise).

Perhaps Mysterium-X will look like this (guessing):

• Number of stops the camera is capable of: 13.3
• Number of stops that most people will use for 4K DI: 11.6
• Bit precision needed to encode what the camera is capable of: 14
• Bit precision needed to encode what most people will use for 4K DI: 12
• Bit precision used by the ADC: 16
• Bit precision recorded to the file: 14
• Bit precision used in post production: 16

And perhaps Monstro will go like this (guessing):

• Number of stops the camera is capable of: 15.3
• Number of stops that most people will use for 4K DI: 13.6
• Bit precision needed to encode what the camera is capable of: 16
• Bit precision needed to encode what most people will use for 4K DI: 14
• Bit precision used by the ADC: 16
• Bit precision recorded to the file: 16
• Bit precision used in post production: float

Here's an analogy: say it's your job to count how many rain drops fall into a bucket per second. Your given a paper with only two columns, so you have two digits of precision. It starts raining and you very carefully write down numbers:

1.2
1.5
1.3
1.3
1.4

And so on. Soon you get so good at it that you think two columns is not enough precision, and you could do three columns:

1.35
1.46
1.20

But then someone points a big sprinkler right at the bucket and it sends hundreds of drops that are mixing with the true rain drops, and you can't tell them apart. So your precision goes down.

Rain drops are like photons, the true "signal" that the sensor tries to record. Sprinkler drops are like electrons, the read noise that we don't want to record. The number of columns is like the bit depth, or precision, of the ADC. It only makes sense to record numbers that have meaning; if they are lost in the noise, then using more precision will not give us any information: you might as well make up random numbers.

Here's an example of bit depth in current products: the D300. In 12-bit mode, the dynamic range is such that the last bit is not even used. 11 bits would have been fine. The Sony sensor Nikon used for the D300 doesn't have a 14-bit mode (the ADC is on-chip), but Nikon needed one in order to score Marketing points, so the engineers did a quadruple readout (reading the sensor four times for one shot), then did a little processing on all four readouts to remove some of the noise. This improves things significantly so that the camera has less noise and the 12th bit is now being utilized. The last two bits in 14-bit are still just random noise and can be discarded. So Nikon is essentially offering a very slow (1.5 FPS or something) method of reducing read noise.

Nikon has done the same thing with the D3X. Using the 12-bit Sony A900 sensor and reading it four times (at a very slow 1.8 FPS) to improve the noise a little bit so they can call it 14-bit.

[10,000 character limit reached. Continued in next post...]

 

[Daniel Browning]

...continued

...continued

[16-bit] is already supported in the still world and it's about time for those of us in Film Post to use it

The 16-bit recording of MFDB is just a marketing ploy. This demonstration shows that the last four bits do nothing but record random noise: http://forums.dpreview.com/forums/read.asp?forum=1019&message=29766597

Also, Emil Martinec demonstrated that even 14-bit precision in most cameras is wasteful:

http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/

In particular, look at the demonstration on page three:

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

Here is the 12-bit version (look at the huge gaps in the histogram):

Here is the 14-bit version (histogram is filled out):

I can't tell any difference between the two.

So I would say that what you and your colorist (and all of us) really want is higher SNR, more dynamic range, better response to underexposure, less noise, better compression, less quantization, etc. When that is achieved, I expect an increase of bit depth to go along with it.

I'm pretty sure the Monstro sensors are 16 bit.

I saw that, too. It wasn't specified if just the ADC was 16-bit or the REDCODE along with it. RED needs to put 16-bit on the specs for marketing reasons, whether they are used to bloat the file or not is something else entirely. Everyone else records more bits than they actually need because it helps their snake oil marketing. I expect RED will only have 16-bit ADC, but wont actually record any more bits than are actually needed. It's like a reverse bait-and-switch: sell the snake oil (22-bit ADC!), but secretly replace it with something *useful* (e.g. 14-bits REDCODE). At least, that's what I would do.

It may not be directly relevant, but of all the DSLR cameras on the market now that currently provide 14 bit RAW files, I believe there is no clear evidence that all those bits are actually justified.

Correct. The best of them (Nikon D3) only need 13 bits, and most only need 12 bits (or less). If they combined the low read noise of ISO 1600 with the high FWC of ISO 100 (as in patent 6734905), then the D3 would need 15 bits, but until then 13 is the max.

The only thing you'd get out of a 16 bit Red One is more accurate noise, as Graeme has said over and over and over again. It may help with banding though...

By banding, I assume you mean posterization. If it helped with that, then it would mean there was more signal to record. A too-small bit depth causes quantization error (AKA quantization noise), and then it causes posterization (banding). Most of the time people see posterization in post production because of their display or processing (e.g. 8-bit display or 10-bit processing can easily show posterization that isn't there in the raw file).

In the world of DSLRs going from 10 bits to 12 bits to 14 bits has resulted in better color reproduction across the board, broader ISO ratings, etc. (although admittedly, there are other technological things improving at the same time).

All the improvements in those areas have been due to other reasons. Color reproduction improved due to resolution, noise, color filter arrays that more closely match the human eye, and software. ISO ratings improved due to higher quantum efficiency (15-50%) and less read noise. The references above show that, for almost all DLSR cameras, 12-to-14 bits made no difference.

As long as read noise, FWC, and QE stays the same per area, every increase in resolution can be accompanied by a decrease in bit depth to capture the same dynamic range. If technology improves per area (as it has in the past), the same or more precision is required. The ideal camera, which wont be possible for decades, if ever, is a photon counter. It has so much resolution (we're talking gigapixels), that the chance of two photons striking the same pixel is vanishingly small. The bit depth required is just 1: on or off.

In post processing, you can convert such a 1-bit image up to any desired bit depth. Here's a demonstration by Jay Turberville:

http://forums.dpreview.com/forums/read.asp?forum=1018&message=30280847

(Click thumb for full image)



You can see that a one-bit precision captures all sorts of tonal gradations. There are also plenty of methods for handling color. "Some researchers are actually trying to build one right now: Donald Figer and his colleagues at Rochester Institute of Technology got $2.8M grant to develop a light detector that would count individual photons for astronomy applications":

http://image-sensors-world.blogspot.com/2008/10/rit-team-to-develop-photon-counting.html

There's also a paper by Eric Fossum, inventor of CMOS sensors:

"What To Do With Sub-Diffraction-Limit (SDL) Pixels? -- A Proposal for a Gigapixel Digital Film Sensor (DFS)"
http://ericfossum.com/Papers/2005 Program What to Do with Sub-Diffraction Limit (SDL) Pixels.pdf

In it he proposes grouping 1-bit pixels into clumps, like film grain, for post-processing.

So what I would actually like to see is less bit depth, and more resolution.

 

[Marcus la Grone]

cranking the number of clean bits

cranking the number of clean bits

The sensor array fundamentally accounts for a lot of the noise down in the floor. While well designed silicon can greatly help the noise floor, there is a simple way to help even "simple" arrays: good heat management. I come from a scientific imaging background and the standard trick we often use to get more 'good bits' is to cool the array. Going from ambient down to -40F (-40C... fancy that) adds about 3 to 4 useful bits. The problem is no one wants to carry the extra battery around for the TE coolers (or, as is my present case, do all their photography outside in Alaska). Well designed heat pipes, however, can greatly improve wast heat and therefore noise. If the guys at RED say they can get us 16 good clean bits, I'll be impressed, but happy. But then again, so far I'm already impressed with what they have done.

 

[Daniel Browning]

While well designed silicon can greatly help the noise floor, there is a simple way to help even "simple" arrays: good heat management.

That's certainly the case for any exposure longer than 1/10 second, but typical photography is not affected by thermal noise to any measurable degree. Dark current of about 10 e-/sec is typical for a modern pixel at room temperature, and temporal noise is only (approx.) square root of that. All the other noise sources (e.g. read noise) are so much higher that thermal noise can't be noticed at 1/10 second.

EDIT: that's not to say heat management is unimportant. Some say it's the biggest challenge in designing motion cameras (I wouldn't know). Can't have cameras catching fire or "letting the magic smoke out".

Personally, I'm very interested in astrophotography and timelapse, so I would love a Thermo-Electric cooler; I requested it in my DSMC wish list post.

 

[Marcus la Grone]

agree/disagree

agree/disagree

Going to have to sort of agree and sort of disagree with the short exposure statement. (Using DALSA arrays) At 30fps we've seen improvements from cooling; granted, they aren't as impressive improvements as they are for long exposures. On the other hand, while the noise may be "okay" at room temp, encased in a camera, most FPAs shoot up north of 45C. 10 e-/sec is actually kinda noisy in my book, if your well depth is 150k that is only 15k "happy range" aka: ~13-14bits.
wrt: title of the thread-- I'm greedy I want more pixels and more bits...

 

[Daniel Browning]

Going to have to sort of agree and sort of disagree with the short exposure statement. (Using DALSA arrays) At 30fps we've seen improvements from cooling; granted, they aren't as impressive improvements as they are for long exposures. On the other hand, while the noise may be "okay" at room temp, encased in a camera, most FPAs shoot up north of 45C.

Interesting. Thanks for the information.

 

[Ian Andolina]

Daniel: thank you for such detailed and fascinating posts!

 

[George Wilcox]

More bits on better sensors IS better....

More bits on better sensors IS better....

"Curiously, all the 14-bit cameras on the market (as of this writing) do not merit 14-bit recording. The noise is more than four levels in 14-bit units on all of these cameras (Nikon D3/D300, Canon 1D3/1Ds3 and 40D); the additional two bits are randomly fluctuating, since the levels are randomly fluctuating by +/- four levels or more. Twelve bits are perfectly adequate to record the image data without any loss of image quality, for any of these cameras (though the D3 comes quite close to warranting a 13th bit). "

http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p3.html#bitdepth

There is certainly much more to a clean image than the number of bits - the design of the sensor, micro-lenses, gap between sensors, materials, noise filtering, and so on. But if you look at comparisons of the images produced on different DSLR like I have rather obsessively, I can tell you that the higher bit depth sensors produce better images. Have a visit to http://www.imaging-resource.com to compare.

Comparing the new Sony A900 to the Canon EOS 5D MkII,the Sony uses 12bit A/D and the Canon uses 14bit. The Canon absolutely smokes the Sony like a cheap cigar! Especially at high ISO ratings. It also bests the EOS 1Ds MkIII which has the same bit depth and resolution. This shows that improvements in signal processing and sensor design play a key role, but pixel depth is important too. As an interesting point, a 12bit pixel can reproduce 4096 colors. A 14bit pixel can reproduce 16,384 colors and a 16bit pixel can produce 65,536 colors per pixel. So a 16bit pixel can reproduce 16 times the number of colors as a 12bit pixel can. That's a whole lot of richness and detail in highlights and shadows and alot more room to move, all yours for a modest 33% increase in bits. Unsubstantiated rumor has it that Canon's 1Ds Mk IV will sport 18bits/pixel with 262,144 colors per pixel, but I for one, look forward to 24 bit DSLR and cine cameras with 16.7 million colors per pixel. Now that would finally go toe to toe with film! Color depth is the weak link in digital imaging so far, but as you can see, Jim Jannard is dedicated to heading in the right direction. 16bit will be a huge step in the right direction. Bravo Jim; keep those delicious bits AND pixels coming!:w00t:

 

[Daniel Browning]

There is certainly much more to a clean image than the number of bits - the design of the sensor, micro-lenses, gap between sensors, materials, noise filtering, and so on. But if you look at comparisons of the images produced on different DSLR like I have rather obsessively, I can tell you that the higher bit depth sensors produce better images.

The 14-bit cameras produce better images exactly because of the other reasons you mentioned (and more), not the greater recorded bit depth.

Comparing the new Sony A900 to the Canon EOS 5D MkII,the Sony uses 12bit A/D and the Canon uses 14bit. The Canon absolutely smokes the Sony like a cheap cigar!

First of all, the Canon JPEG engine is far better than Sony's, which accounts for the difference you see on that web site. RAW is a different story: at base ISO with a good raw conversion the Sony has better contrast, more resolution, and similar noise.

Second, correlation does not imply causation. I know a lot of people who measure fuel economy as "how much it costs to fill up". This works because as fuel economy goes up, manufacturers shrink the fuel tank and driving distance remains the same. If Marketing had their way, they would just make the gas tank smaller. This would fool said people into thinking fuel economy improved just because the "price at the pump" was lower, at least until they realized their measurement methodology was fundamentally flawed.

The bit depth is just one of the many differences between the two cameras, one cannot assume that it is the primary cause of the difference in image quality. In fact, for JPEG, the difference can not be related at all because the default black point and curves crush all detail in the last two bits anyway.

By the way, I just got my 5D2 a few days ago and I will be testing it as soon as I get the chance. One thing that I will test for is what bit depth is really needed for the camera. So far it looks very promising and has very low read noise.

Especially at high ISO ratings.

High ISO needs requires even *less* bit depth than base ISO. At ISO 3200, the 5D2 cannot make use of even 10 bits because so much is just read noise. The bit depth of the camera has to be designed for the amplification with the most dynamic range, which is base ISO in most cameras.

In any case, even if the Canon *was* better than the Sony, one could not assume that it was because it recorded more bits. It could be because the sensor has higher sensitivity and lower read noise. Or maybe the manufacturer switched to a sloppier color filter to improve sensitivity at the expense of spectral response (i.e. color accuracy), as Canon did with the 5D2.

Or maybe the new 14-bit ADC does have less read noise. Maybe it's just now two stops less. For example, if a camera had 11.0 stops of dynamic range and 12-bit ADC, then the 12 bits are sufficient to record the entire signal. If the same camera was upgraded with a 14-bit ADC which improved read noise by 0.5 stops, then it would have 11.5 stops of dynamic range. But that can still be recorded losslessly within 12 bits. Which is what manufacturers *should* do because it would be best for the photographer.

Instead, Marketing realizes that they can market additional bits, even if they are nothing but bloat, because people will assume that no company is evil enough to waste disk space just for marketing purposes. Unfortunately, Marketers are smart enough to do it in small increments, so that more people are fooled. If they jumped to 32 bit depth or 1000 bit depth immediately, no one would beleive it, but adding 2 fake bits here and 2 there is easier to sell.

One can truncate most 14-bit cameras to 12 bits in post production and never even notice the difference.

It also bests the EOS 1Ds MkIII which has the same bit depth and resolution. This shows that improvements in signal processing and sensor design play a key role, but pixel depth is important too.

Here's how you can measure the role of bit depth: truncate the bits in post production. If there's no difference after you lop them off, they were marketing bits. If it results in posterization, then they were real bits. All of the early 14-bit cameras had at least two marketing bits. Some of the more recent ones have only 1 marketing bit (e.g. D700 uses 13/14 bits).

As an interesting point, a 12bit pixel can reproduce 4096 colors. A 14bit pixel can reproduce 16,384 colors and a 16bit pixel can produce 65,536 colors per pixel. So a 16bit pixel can reproduce 16 times the number of colors as a 12bit pixel can. That's a whole lot of richness and detail in highlights and shadows and alot more room to move, all yours for a modest 33% increase in bits.

That's all true (setting aside Bayer interpolation), but it only applies if the sensor and electronics can keep up. Anyone can record 16 bits or 22 bits or a thousand bits, that's not hard. The hard part is building the sensor/electronics to actually put something useful in those bits.

For example, the Pentax K10D has a 22-bit ADC, but it is certainly not capable of 4.1 million values per pixel. It actually does use all 14 bits, but the last two stops have such strong pattern noise that no one can stand to use them in real life, so it's effectively 12 bits (despite the 22-bit ADC).

keep those delicious bits AND pixels coming!

Yes!