10 Bit Cameras

[John Richard]

What are the opinions/facts on the Sony FS-100? does it have a true usable 10bit output that would make sense with the Gemini?

(Unfortunately the FS-100 does not have HD-SDI and relies on HDMI as the output - grrr)

[Piotr Wozniacki]

Unfortunately (from the viewpoint of Gemini potential buyers), the FS100 only outputs 8 bits via HDMI. So, investing in the Gemini for the FS100 would be an overkill, IMHO. It will make a perfect companion for the F3, though (I wish I could afford them both...).

Fortunately for the current nanoFlash users (myself included), after implementing TC through HDMI, our little marvels will serve this camera perfectly :)

Piotr

[Rafael Amador]

David,
I mostly agree with you.
Simplifying, you lose "1bit". Is like you end up working in 9b.
512 levels instead of 1024.
BUT that's still much better than 256.
Better picture sampling, and as you points too, easier to clean the noise.

What I do not agree with you when you say:
"If you have a set of scales accurate to + or - 2 grams, what would be the point of having a set of weights that would allow you to measure to 0.1 of a gram?"
Hey, those cameras processes at 12/14 bits before going 10b.
Our original scale is not set in grams or even decigrams, is milligrams.


Dan wrote:
"I do agree completely with Alister and others that some camera's that technically have a 10-Bit HD-SDI output do not have a low enough noise floor to make the 10-Bit truly useful".
Dear, Dan:
You don't get the full advantage of the 10b, but you go way much better than at 8b.

Originally Posted by Steve Kalle

Hi Dan,

I tried to find the post about 1bit per 6db of S/N but I haven't had any luck. I honestly thought that it was you who posted this. If it was you, can you provide an explanation again about 1-bit and 6db?

Thanks

Steve: that's not wrong. I more bit means duplicating the number of sampling levels, so duplicating the sensibility. That means 6dbs of dinamic range.

About the "EX-1 proved 48db sensitivity", up to the EX-1 brochure "the CAMCORDER" (In SxS recording) has a S/N of 54db.
The SDI signal doesn't fallow the same process.
We still talking about the EX-noise and the 10b issue without having yet shot any test at 10b Unc from that SDI.
We are mixing the concept of system noise with camera noise (that can be generated for a wrong setting), and compression noise and .
rafael

[David Heath]

Originally Posted by Rafael Amador

What I do not agree with you when you say:
"If you have a set of scales accurate to + or - 2 grams, what would be the point of having a set of weights that would allow you to measure to 0.1 of a gram?"
Hey, those cameras processes at 12/14 bits before going 10b.
Our original scale is not set in grams or even decigrams, is milligrams.

Ah, but the original scale is for the raw output from the chip, and before ANY of the necessary camera processing - knee compression, gamma compensation, gain, anything. That will inevitably scale many of the quantisation steps up a great deal, so if (say) a 12 bit system, a single bit difference pre-processing may equate to many bits post-processing. (And in other parts of the waveform many bits will be compressed into one.)

That's the whole reason for the necessity of the front end needing to be at a far better bit depth than the output.

To carry on the weight analogy, it's a bit like a "system" whereby you measure in hundreths of a gram, then give an "output" of 100x that weight. So if a "customer" comes to you with .04g, you give out 4grams. You may need milligram accuracy for the input - you only need gram accuracy for the output.

Maybe a better analogy would be if you are making something from drawings, and the drawings are to 1:10 scale. If you need an accuracy of at least a centimetre in the finished product, then obviously the drawings should be accurate to the nearest millimetre.

The 1 bit=6dB is true in principle, but in real life there are all sorts of complications. The idea of a camera having a defined s/n ratio is great - but in practice the figure will vary depending on whether it's measured in the highlights, lowlights, or mid-greys. I suspect manufacturers will want to publish the figures that look the best, and gloss over on what part of the curve the measurements are made. That's why I'd rather stick to statements like "if the noise is greater than the differences in level represented by the least significant bit, resolving to the 10 bit level becomes a bit pointless" and not get deeply into figures.

The science is that the human eye can only resolve something like 100 different shades of grey at a single time - and they can be *just about* represented within a 7 bit binary number. I have a book (not online - sorry) which shows a 6 bit greyscale with obvious banding, then the same greyscale with noise added, and the banding is totally masked. Practically, 8 bits becomes desirable - gives a little leeway - for straightforward viewing. The depth resolution is better than the human eye. Processing is a different matter, and 10 bit processing may be a very good thing even if the original acquisition has been at 8 bit - 10 bit acquisition only really becoming worthwhile if the camera noise level is low, very low.

[Rafael Amador]

David,
I agree with all your points. Is pure maths.
What I do not agree is with your starting point:

"Ah, but the original scale is for the raw output from the chip, and before ANY of the necessary camera processing - knee compression, gamma compensation, gain, anything".

When a camera claims to process in 12 or 14 bits, that counts for all the process from digitizing the RGB on the CCD/CMOS all the way (RGB matrix, knee, detail,..) to a 12/14b Y'C'bC'r/4.4.4.
Then the signal just need to be down-sampled (in bit depth and Chroma) to 10b YUV 422 for SDI out for to whatever codec the camera may uses.

See what the AF-100 brochure claims:
"The AG-AF100 series is equipped with a high-performance 18-bit digital signal processor (DSP) for image processing. Optimized for HD video recording, the DSP handles various image rendering processes, such as dynamic range stretch (DRS), gamma, 12-axis independent color correction* and detail enhancement, as well as conversion to HD/SD video formats - all with exceptional precision and high image quality.

A camera claiming "12/14b digital processing" and that 12/14b stoping at the raw output of the chip,that would be a scam.
I haven't read no where the supposed bit depth processing of the EX-1/3.
rafael

[David Heath]

OK, to try to make it clearer, yes, it's true that the entire early processing chain will be 12/14 bit - but that's not to say that the DATA within that will have that sort of resolution. And the processing will affect which of the range is relevant, which is not. So you may start with a nominal 12 bit number (say 000000101010) multiply it by (say) 16, and you get 001010100000. Now whatever number you start with, after a multiplication by 16 the last four significant bits will always be zero, and hence can be discarded. The analogy is that after "processing" all the information will be in the most significant 8 bits and it's pointless to keep the others.

The above is obviously highly simplistic - processing involves far more than multiplication by 16 :-) - but the principle does hold true. You DO need a high bit depth for initial processing, to ensure the calculations don't get rounded or overflow, but the likelihood is that at output the least significant bits won't be meaningful and it's sensible to just lose them. (Which is true even if 10 bit final output.)

The question is only where to draw the line. And one of the main factors in that will be the noise level of the camera after processing. Which brings us back to where this started - a noisier camera can have the line drawn at keeping a higher bit depth than one with an inherently lower s/n ratio. In practical terms, the most 8 significant bits are all that is relevant for the majority of cameras, there may be something to be gained for 10 bits for the most expensive.

[Rafael Amador]

Hi David,

Your maths are perfect, but for a completely different scenery.
You are talking about "multiplying x 16"; you are talking about about AMPLIFYING.

In the camera video signal processing there is not any AMPLIFYING involved.
Well, we have the "Gain" for desperate situations.
Out of the Gain, the multipliers you will find in a camera processor are values between "0" and "1".
The product of any multiplication can't ever be bigger than the multipliers, and can't ever be bigger than "1", so in fact what you are always doing is DIVIDING, reducing values.

- After the raw RGB is Gamma corrected (R'G'B'), Luma is calculates with this formula (709):
Y'= 0.2126 R' + 0.7152 G' + 0.0722 B').

- When you matrix R'G'B' to get C'b and C'r you go to similar calculations.

No one of those operations will generate "0's" at the right of your values.

YOU NEVER GET A BUNCH OF 000000 in the LESS IMPORTANT BITS.
All the range is full of detail.
When reducing bit-depth you won't ever be trashing a bunch of empty bits.

You also you forget the substance of what we are talking about.
You start from "OK, to try to make it clearer, yes, it's true that the entire early processing chain will be 12/14 bit - but that's not to say that the DATA within that will have that sort of resolution"
Well if you start with something as poor as"xxxxxxxxxx00", you are right, but that's not what you should be getting from a CMOS/CCD claiming 12/14 bits.
A camera just sample electric charge that may range all the way from "0 to 1,099 Volts".
Here things doesn't works in "steps" yet. Here we have infinite intermediate values to represent.
The more bit depth you use to sample that charge, the more definition will have. Noise included.
rafael

[David Heath]

Originally Posted by Rafael Amador

You are talking about "multiplying x 16"; you are talking about about AMPLIFYING.

In the camera video signal processing there is not any AMPLIFYING involved.

Yes there is, though it may not be obvious as such at first sight. The obvious examples are knee compression and gamma adjustment. Referenced to peak value, the former acts by division for the highest values and hence multiplication of the values corresponding to darker levels. Gamma correction is effectively multiplying the lowlight values by a far greater factor than highlights.

YOU NEVER GET A BUNCH OF 000000 in the LESS IMPORTANT BITS.
All the range is full of detail.
When reducing bit-depth you won't ever be trashing a bunch of empty bits.

No, practically I doubt you ever would get all 000's - I did say my analogy was "highly simplistic". In pactice, any manipulation is unlikely to be simple multiplication by a power of 2. But the principle remains true. You may need the high bit depth to be able to perform the calculations - after calculation the least significant bits will be random if not 0000. The fluctuations in them will ONLY represent noise and are not worth wasting bandwidth on preserving.

A camera just sample electric charge that may range all the way from "0 to 1,099 Volts".
Here things doesn't works in "steps" yet. Here we have infinite intermediate values to represent.
The more bit depth you use to sample that charge, the more definition will have. Noise included.
rafael

But just what is the point of more and more accurately defining the level any instantaneous value of noise is at? This is equivalent to trying to state a weight to a hundreth of a gram when your scales are only accurate to +/- a single gram.

Let me try one more example. Let's say a businessman wants to sell a product at the average price that 1,000 other stores charge, plus 1 penny. All the prices are in the range 15p to 31p - so can be defined with a two digit number. The first step is obviously to add all the 1,000 prices together, and let's say the result is 23,034 - obviously we need at least FIVE digit processing to start processing the data. Next, divide by 1,000, which may give 00,023.034 - which you may say means EIGHT digit processing, and finally add a penny to give 00,024.034 pence.

But do you think that is the price that will go on the ticket? Obviously not, and the practical answer will be "24 pence" - it makes sense to go back to a two digit number for output. The three most significant bits have served their purpose in processing, the last three likewise. An improved analogy would be to realise that his 1,000 competitors change their prices slightly day by day and the next day the average result may be "00.023.893 pence". The digits after the decimal point here represent noise, and it's pointless to worry about them.

So in the analogy a two digit number is quite adequate to define input and output, but you need five/eight digit numbers to process the calculation. And the principle for video is exactly the same. 12/14 digit processing does not necessarily mean 10 bit output will offer any advantage over 8 bit output - that will depend on the noise level.