fredmiranda.com Login
This topic is in the 2015 archive. You can read it and like its posts here. It is read only.

Archive 2015High DR for 5d4

  
 
trale
Offline
Details
p.4 #1High DR for 5d4
Theoretically speaking, can this approach be taken up a notch to reach 20 stops (the approx effective DR of human eye) in a single exposure, at the cost of even more resolution? We might end up with a "low-rez" camera of "only" 12.5MB, but if it means that finally a single shot can accurately portray what we remember seeing with our eyes, that seems like a tremendous breakthrough.

As it stands, even with 15 stops of DR, high contrasts scenes with light sources in the composition will still require multiple exposures to fully capture the scene.
Apr 13, 2015 at 03:39 AM
alundeb
Offline
Details
p.4 #2High DR for 5d4
trale wrote:
Theoretically speaking, can this approach be taken up a notch to reach 20 stops (the approx effective DR of human eye) in a single exposure, at the cost of even more resolution?


The limit for this approach with the current levels of saturation at low ISO and read noise at high ISO is around 15-17 stops.
Apr 13, 2015 at 04:04 AM
alundeb
Offline
Details
p.4 #3High DR for 5d4
dhphoto wrote:
wouldn't this immediately render the 5Ds/r rather obsolete?

I can't see why Canon would do this to themselves.


Dual ISO readout at full resolution requires double the processor / data pipeline speed. The 5D III is the only "high-end" current Canon camera with a single Digic processor. If they make the 5D IV roughly the same pixel count and Dual Digic, they have the processing power to do dual ISO readout without reducing the framerate. The 5Ds will already have Dual digic, and would not have the speed to do this without halving the frame rate.

It seems that with Canon we will now have to choose between one of these:

Normal MP, high FPS, normal DR (1DX)
Normal MP, normal FPS, extended DR (5DIV)
High MP, normal FPS, normal DR (5Ds)

I would have preferred High MP, low FPS and extended DR, but doubt it will come anytime soon.
Apr 13, 2015 at 05:25 AM
PhilDrinkwater
Offline
Details
p.4 #4High DR for 5d4
alundeb wrote:
I would have preferred High MP, low FPS and extended DR, but doubt it will come anytime soon.


If that was the reason for the 5ds not being able to have high DR, I imagine you're not the only one to have that view...
Apr 13, 2015 at 06:20 AM
dhphoto
Offline
Details
p.4 #5High DR for 5d4
alundeb wrote:
I would have preferred High MP, low FPS and extended DR, but doubt it will come anytime soon.


I expect a majority here would, which is why I said Canon was shooting themselves in the foot. They make a big fuss about the oversized 7D2 sensor in the 5Ds/r, I think they are unlikely to put significantly superior image quality in a cheaper model.
Apr 13, 2015 at 06:58 AM
alundeb
Offline
Details
p.4 #6High DR for 5d4
PhilDrinkwater wrote:
alundeb wroteI would have preferred High MP, low FPS and extended DR, but doubt it will come anytime soon.
If that was the reason for the 5ds not being able to have high DR, I imagine you're not the only one to have that view...


Dual ISO readout also requires more complex pixel layout, which is more difficult with higher pixel density. So it is probably not so easy as to just lower the frame rate on the 5Ds.

Still, DR is only one parameter.

What other sensor than Canon 5Ds does:
50 MP or more
High selctivity CFA
Extremely low thermal noise
Lossless encoding
No nasty sensor glass reflections
Electronic first curtain shutter
Compatibility with everything from extremely wide to long and T/S glass

Apr 13, 2015 at 07:02 AM
snapsy
Offline
Details
p.4 #7High DR for 5d4
alundeb wrote:
The maximum dynamic range of linear encoded data does not equal the number of bits, because the qantization noise is much smaller than 1 LSB. The quantization noise of an ideal converter is maximum +/- 1/2 LSB. If the data is uniform, the quantization noise as an RMS value is LSB / sqrt(12).

This means that 14 bits of linear data can carry theorethically up to 15.8 stops of dynamic range.

How do you arrive at 15.8 stops of DR from the RMS measurement of a mean 0 signal in a 14-bit ADC?
Apr 13, 2015 at 09:29 AM
Gunzorro
Offline
Details
p.4 #8High DR for 5d4
This is very exciting stuff, and sure to be implemented one way or another in the near future. I can see why Canon's management might well be confident that Canon has the best sensors and not need to buy another maker's product to put into Canon cameras. There is a long way to go wring the most out of Canon's current manufactured technology, especially on the processing side of things, where I see Canon continuing to have a strong lead over Nikon and Sony.

It should be an exciting couple of years coming up, and Canon owners will surely be glad they kept all their EF glass!
Apr 13, 2015 at 09:53 AM
alundeb
Offline
Details
p.4 #9High DR for 5d4
snapsy wrote:
alundeb wroteThe maximum dynamic range of linear encoded data does not equal the number of bits, because the qantization noise is much smaller than 1 LSB. The quantization noise of an ideal converter is maximum +/- 1/2 LSB. If the data is uniform, the quantization noise as an RMS value is LSB / sqrt(12).

This means that 14 bits of linear data can carry theorethically up to 15.8 stops of dynamic range.
How do you arrive at 15.8 stops of DR from the RMS measurement of a mean 0 signal in a 14-bit ADC?


15.8 stops of DR in a 14 bit ADC does not require a mean 0 signal. It requires a smallest mean signal of 1/sqrt(12). With a mean signal of 1/sqrt(12) or approximately 0.29 and some analog noise, the analog value of signal plus noise will sometimes be above 0.5 and then quantized as 1.

With some analog noise (if the noise is not present with the signal, we can add noise as dithering), the actual DR will be slightly lower than 15.8 stops, I haven't calculated that exactly.

Without analog noise or dithering, the only noise source is the quantization noise, and even though we cannot see any difference between a signal of 0 and 0.5, the noise is theorethically 1/sqrt(12) if the signal actually varies between 0 and 0.5.


Edited on Apr 13, 2015 at 10:07 AM · View previous versions
Apr 13, 2015 at 09:55 AM
snapsy
Offline
Details
p.4 #10High DR for 5d4
alundeb wrote:
snapsy wrote
alundeb wroteThe maximum dynamic range of linear encoded data does not equal the number of bits, because the qantization noise is much smaller than 1 LSB. The quantization noise of an ideal converter is maximum +/- 1/2 LSB. If the data is uniform, the quantization noise as an RMS value is LSB / sqrt(12).

This means that 14 bits of linear data can carry theorethically up to 15.8 stops of dynamic range.
How do you arrive at 15.8 stops of DR from the RMS measurement of a mean 0 signal in a 14-bit ADC?
15.8 stops of DR in a 14 bit ADC does not require a mean 0 signal. It requires a smallest mean signal of 1/sqrt(12). With a mean signal of 1/sqrt(12) or approximately 0.29 and some analog noise, the analog value of signal plus noise will sometimes be above 0.5 and then quantized as 1.

With some analog noise (if the noise is not present with the signal, we can add noise as dithering), the actual DR will be slightly lower than 15.8 stops, I haven't calculated that exactly.

How do you arrive at a mean signal of 1/sqrt(12)? To my understanding the 1/sqrt(12) is a measurement of the quantization noise relative to the signal, not a measurement of the signal itself. The signal itself for this discussion would have a mean of zero and an RMS of the signal's distribution at zero. Independent of this, the posterization of a mean 0 signal would make it unusable in the practical sense for digital imaging anyway so a theoretical DR is > 2^bit_depth isn't applicable (I've never read of an imaging ADC dithering signals at mean 0).
Apr 13, 2015 at 10:07 AM
alundeb
Offline
Details
p.4 #11High DR for 5d4
snapsy wrote:
alundeb wrote
snapsy wrote
alundeb wroteThe maximum dynamic range of linear encoded data does not equal the number of bits, because the qantization noise is much smaller than 1 LSB. The quantization noise of an ideal converter is maximum +/- 1/2 LSB. If the data is uniform, the quantization noise as an RMS value is LSB / sqrt(12).

This means that 14 bits of linear data can carry theorethically up to 15.8 stops of dynamic range.
How do you arrive at 15.8 stops of DR from the RMS measurement of a mean 0 signal in a 14-bit ADC?
15.8 stops of DR in a 14 bit ADC does not require a mean 0 signal. It requires a smallest mean signal of 1/sqrt(12). With a mean signal of 1/sqrt(12) or approximately 0.29 and some analog noise, the analog value of signal plus noise will sometimes be above 0.5 and then quantized as 1.

With some analog noise (if the noise is not present with the signal, we can add noise as dithering), the actual DR will be slightly lower than 15.8 stops, I haven't calculated that exactly.
How do you arrive at a mean signal of 1/sqrt(12)? To my understanding the 1/sqrt(12) is a measurement of the quantization noise relative to the signal, not a measurement of the signal itself. The signal itself for this discussion would have a mean of zero and an RMS of the signal's distribution at zero. Independent of this, the posterization of a mean 0 signal would make it unusable in the practical sense for digital imaging anyway so a theoretical DR is > 2^bit_depth isn't applicable (I've never read of an imaging ADC dithering signals at mean 0).


Why does the signal have to be 0?

One common definition of DR uses SNR = 1.

Apr 13, 2015 at 10:10 AM
wayne seltzer
Offline
Details
p.4 #12High DR for 5d4
Isn't the Sony Exmor sensor have better DR due to lower read noise at low iso because their patented architecture which has the ADC digitizing closer to the source?
They need to improve there rather than dual iso approach.
At some point Canon will realize that funding 2nd rate sensor design is not worth it and not adding value to thr product over using the industry best sensor tech.
Apr 13, 2015 at 10:11 AM
snapsy
Offline
Details
p.4 #13High DR for 5d4
alundeb wrote:
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wroteThe maximum dynamic range of linear encoded data does not equal the number of bits, because the qantization noise is much smaller than 1 LSB. The quantization noise of an ideal converter is maximum +/- 1/2 LSB. If the data is uniform, the quantization noise as an RMS value is LSB / sqrt(12).

This means that 14 bits of linear data can carry theorethically up to 15.8 stops of dynamic range.
How do you arrive at 15.8 stops of DR from the RMS measurement of a mean 0 signal in a 14-bit ADC?
15.8 stops of DR in a 14 bit ADC does not require a mean 0 signal. It requires a smallest mean signal of 1/sqrt(12). With a mean signal of 1/sqrt(12) or approximately 0.29 and some analog noise, the analog value of signal plus noise will sometimes be above 0.5 and then quantized as 1.

With some analog noise (if the noise is not present with the signal, we can add noise as dithering), the actual DR will be slightly lower than 15.8 stops, I haven't calculated that exactly.
How do you arrive at a mean signal of 1/sqrt(12)? To my understanding the 1/sqrt(12) is a measurement of the quantization noise relative to the signal, not a measurement of the signal itself. The signal itself for this discussion would have a mean of zero and an RMS of the signal's distribution at zero. Independent of this, the posterization of a mean 0 signal would make it unusable in the practical sense for digital imaging anyway so a theoretical DR is > 2^bit_depth isn't applicable (I've never read of an imaging ADC dithering signals at mean 0).
Why does the signal have to be 0?

One common definition of DR uses SNR = 1.


The quantization of the LSB on imaging sensor ADCs is at a mean signal level of zero, with a +/- distribution around zero. I'm still not seeing how you arrive at 15.8EV of DR from the 14-bit quantization. Can you supply the fully formula you're using so that it's clearer? Also, what is the use of > 2^bit_depth DR if the result is fully posterized, irrespective of what the precise DR measurement is between 14.0EV and 15.8EV?
Apr 13, 2015 at 10:39 AM
curious80
Offline
Details
p.4 #14High DR for 5d4

The most probable explanation why they can get as high DR as 15 stops with slog2 at ISO 1600 is:

The sensor runs at base ISO, the shadows and midtones are lifted, and the extra stops of DR are encoded into the highlights. Effectively, the midtones get pushed 4 stops, and the highlights get a smooth rolloff. The image then apperas as ISO 1600, but without the higlight clipping that would occur if the sensor ran at ISO 1600.


Which is what I had explained earlier. And I think from snapsy's last response to me, he also seems to have agreed to this. The only point of contention now is whether this gain curve gets applied to the data before quantization. Or is it the case that 14-bit quantized data is read from the sensor and then the gain is applied to these 14-bit quantized values to get the final 8-bit slog2 values. My contention has been that if slog2 curve really stores 15.3EV of DR in 8-bits then you can't create that by post-processing 14-bit linear RAW data. (Though after this discussion I searched around and found no clear mention anywhere that slog2 stores 15.3EV of DR. Some places include claims from Sony that the A7s sensor delivers 15.3EV of DR, but other places like the press release on dpreview don't seem to mention the 15.3EV number).

alundeb wrote:
The maximum dynamic range of linear encoded data does not equal the number of bits, because the qantization noise is much smaller than 1 LSB. The quantization noise of an ideal converter is maximum +/- 1/2 LSB. If the data is uniform, the quantization noise as an RMS value is LSB / sqrt(12).

This means that 14 bits of linear data can carry theorethically up to 15.8 stops of dynamic range.



Hmmm but unless the ADC is assigning the values in somewhat non-linear way, you just don't have enough bits to encode that many levels. If your whitest tone is recorded as 16383 (2^14 -1), then the tone 1 stop below it would be recorded as 8191, and so on and if you keep on going like that in a linear manner than you would hit 0 before you get a chance to encode the tones corresponding to the 15th and 16th stop.
Apr 13, 2015 at 11:13 AM
alundeb
Offline
Details
p.4 #15High DR for 5d4
snapsy wrote:
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wroteThe maximum dynamic range of linear encoded data does not equal the number of bits, because the qantization noise is much smaller than 1 LSB. The quantization noise of an ideal converter is maximum +/- 1/2 LSB. If the data is uniform, the quantization noise as an RMS value is LSB / sqrt(12).

This means that 14 bits of linear data can carry theorethically up to 15.8 stops of dynamic range.
How do you arrive at 15.8 stops of DR from the RMS measurement of a mean 0 signal in a 14-bit ADC?
15.8 stops of DR in a 14 bit ADC does not require a mean 0 signal. It requires a smallest mean signal of 1/sqrt(12). With a mean signal of 1/sqrt(12) or approximately 0.29 and some analog noise, the analog value of signal plus noise will sometimes be above 0.5 and then quantized as 1.

With some analog noise (if the noise is not present with the signal, we can add noise as dithering), the actual DR will be slightly lower than 15.8 stops, I haven't calculated that exactly.
How do you arrive at a mean signal of 1/sqrt(12)? To my understanding the 1/sqrt(12) is a measurement of the quantization noise relative to the signal, not a measurement of the signal itself. The signal itself for this discussion would have a mean of zero and an RMS of the signal's distribution at zero. Independent of this, the posterization of a mean 0 signal would make it unusable in the practical sense for digital imaging anyway so a theoretical DR is > 2^bit_depth isn't applicable (I've never read of an imaging ADC dithering signals at mean 0).
Why does the signal have to be 0?

One common definition of DR uses SNR = 1.
The quantization of the LSB on imaging sensor ADCs is at a mean signal level of zero, with a +/- distribution around zero. I'm still not seeing how you arrive at 15.8EV of DR from the 14-bit quantization. Can you supply the fully formula you're using so that it's clearer? Also, what is the use of > 2^bit_depth DR if the result is fully posterized, irrespective of what the precise DR measurement is between 14.0EV and 15.8EV?


The definition of DR I use is maximum signal divided by (signal where SNR = 1). That is the same definition used by ISO and also in the DxOmark measurements.

Assuming that quantization noise is the only noise, the noise level is 1/sqrt(12).

The formula I use to get 15.8 stops is: log(16383/(1/sqrt(12)))/log2
Apr 13, 2015 at 11:16 AM
alundeb
Offline
Details
p.4 #16High DR for 5d4
curious80 wrote:
Hmmm but unless the ADC is assigning the values in somewhat non-linear way, you just don't have enough bits to encode that many levels.


One cannot measure noise or DR for one pixel only. The signal level we use is the mean value, and the noise level is the standard deviation. As long as there is noise, the mean value or the noise level is not limited by the quantization step.
Apr 13, 2015 at 11:21 AM
snapsy
Offline
Details
p.4 #17High DR for 5d4
alundeb wrote:
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wroteHow do you arrive at 15.8 stops of DR from the RMS measurement of a mean 0 signal in a 14-bit ADC?
15.8 stops of DR in a 14 bit ADC does not require a mean 0 signal. It requires a smallest mean signal of 1/sqrt(12). With a mean signal of 1/sqrt(12) or approximately 0.29 and some analog noise, the analog value of signal plus noise will sometimes be above 0.5 and then quantized as 1.

With some analog noise (if the noise is not present with the signal, we can add noise as dithering), the actual DR will be slightly lower than 15.8 stops, I haven't calculated that exactly.
How do you arrive at a mean signal of 1/sqrt(12)? To my understanding the 1/sqrt(12) is a measurement of the quantization noise relative to the signal, not a measurement of the signal itself. The signal itself for this discussion would have a mean of zero and an RMS of the signal's distribution at zero. Independent of this, the posterization of a mean 0 signal would make it unusable in the practical sense for digital imaging anyway so a theoretical DR is > 2^bit_depth isn't applicable (I've never read of an imaging ADC dithering signals at mean 0).
Why does the signal have to be 0?

One common definition of DR uses SNR = 1.
The quantization of the LSB on imaging sensor ADCs is at a mean signal level of zero, with a +/- distribution around zero. I'm still not seeing how you arrive at 15.8EV of DR from the 14-bit quantization. Can you supply the fully formula you're using so that it's clearer? Also, what is the use of > 2^bit_depth DR if the result is fully posterized, irrespective of what the precise DR measurement is between 14.0EV and 15.8EV?
The definition of DR I use is maximum signal divided by (signal where SNR = 1). That is the same definition used by ISO and also in the DxOmark measurements.

Assuming that quantization noise is the only noise, the noise level is 1/sqrt(12).

The formula I use to get 15.8 stops is: log(16383/(1/sqrt(12)))/log2

Why are you dividing the bit-depth scale by 1/sqrt(12)? You seem to be mixing units from the quantization error proof with the bit-depth scale. The proper calculation (source, page 6) yields a 100% SNR (and thus DR) result of 6.02(nbits) + 1.76db, which equates to 14.33EV of DR for a 2^14 ADC.
Apr 13, 2015 at 12:53 PM
alundeb
Offline
Details
p.4 #18High DR for 5d4
snapsy wrote:
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wrote15.8 stops of DR in a 14 bit ADC does not require a mean 0 signal. It requires a smallest mean signal of 1/sqrt(12). With a mean signal of 1/sqrt(12) or approximately 0.29 and some analog noise, the analog value of signal plus noise will sometimes be above 0.5 and then quantized as 1.

With some analog noise (if the noise is not present with the signal, we can add noise as dithering), the actual DR will be slightly lower than 15.8 stops, I haven't calculated that exactly.
How do you arrive at a mean signal of 1/sqrt(12)? To my understanding the 1/sqrt(12) is a measurement of the quantization noise relative to the signal, not a measurement of the signal itself. The signal itself for this discussion would have a mean of zero and an RMS of the signal's distribution at zero. Independent of this, the posterization of a mean 0 signal would make it unusable in the practical sense for digital imaging anyway so a theoretical DR is > 2^bit_depth isn't applicable (I've never read of an imaging ADC dithering signals at mean 0).
Why does the signal have to be 0?

One common definition of DR uses SNR = 1.
The quantization of the LSB on imaging sensor ADCs is at a mean signal level of zero, with a +/- distribution around zero. I'm still not seeing how you arrive at 15.8EV of DR from the 14-bit quantization. Can you supply the fully formula you're using so that it's clearer? Also, what is the use of > 2^bit_depth DR if the result is fully posterized, irrespective of what the precise DR measurement is between 14.0EV and 15.8EV?
The definition of DR I use is maximum signal divided by (signal where SNR = 1). That is the same definition used by ISO and also in the DxOmark measurements.

Assuming that quantization noise is the only noise, the noise level is 1/sqrt(12).

The formula I use to get 15.8 stops is: log(16383/(1/sqrt(12)))/log2
Why are you dividing the bit-depth scale by 1/sqrt(12)? You seem to be mixing units from the quantization error proof with the bit-depth scale. The proper calculation (source, page 6) yields a 100% SNR (and thus DR) result of 6.02(nbits) + 1.76db, which equates to 14.33EV of DR for a 2^14 ADC.


The maximum signal in your source is a sine wave with peak-peak value equalling the ADC full scale, and the amplitude used in the max SNR calculation is the rms value of the sine. For an image sensor, the maximum signal is a unipolar DC value and not a sine wave. This gives a signal gain of 2*sqrt(2) or 9 dB over the formula in your source.
Apr 13, 2015 at 01:21 PM
snapsy
Offline
Details
p.4 #19High DR for 5d4
alundeb wrote:
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wroteHow do you arrive at a mean signal of 1/sqrt(12)? To my understanding the 1/sqrt(12) is a measurement of the quantization noise relative to the signal, not a measurement of the signal itself. The signal itself for this discussion would have a mean of zero and an RMS of the signal's distribution at zero. Independent of this, the posterization of a mean 0 signal would make it unusable in the practical sense for digital imaging anyway so a theoretical DR is > 2^bit_depth isn't applicable (I've never read of an imaging ADC dithering signals at mean 0).
Why does the signal have to be 0?

One common definition of DR uses SNR = 1.
The quantization of the LSB on imaging sensor ADCs is at a mean signal level of zero, with a +/- distribution around zero. I'm still not seeing how you arrive at 15.8EV of DR from the 14-bit quantization. Can you supply the fully formula you're using so that it's clearer? Also, what is the use of > 2^bit_depth DR if the result is fully posterized, irrespective of what the precise DR measurement is between 14.0EV and 15.8EV?
The definition of DR I use is maximum signal divided by (signal where SNR = 1). That is the same definition used by ISO and also in the DxOmark measurements.

Assuming that quantization noise is the only noise, the noise level is 1/sqrt(12).

The formula I use to get 15.8 stops is: log(16383/(1/sqrt(12)))/log2
Why are you dividing the bit-depth scale by 1/sqrt(12)? You seem to be mixing units from the quantization error proof with the bit-depth scale. The proper calculation (source, page 6) yields a 100% SNR (and thus DR) result of 6.02(nbits) + 1.76db, which equates to 14.33EV of DR for a 2^14 ADC.
The maximum signal in your source is a sine wave with peak-peak value equalling the ADC full scale, and the amplitude used in the max SNR calculation is the rms value of the sine. For an image sensor, the maximum signal is a unipolar DC value and not a sine wave. This gives a signal gain of 2*sqrt(2) or 9 dB over the formula in your source.


Quantizied DR is the signal peak to stddev irrespective of what the unbounded scale of the input peak into the ADC is. Why would the relative peak have any effect on the ADC's DR calculation? Both get hashed down to quanitization units, converting an infinite set of input measurements (photons, electrons, etc..) into a discrete set of digital output values. The only effect the relative peak-to-stddev differences have is on the resolution of each discrete step.
Apr 13, 2015 at 01:46 PM
snapsy
Offline
Details
p.4 #20High DR for 5d4
alundeb wrote:
curious80 wroteHmmm but unless the ADC is assigning the values in somewhat non-linear way, you just don't have enough bits to encode that many levels.
One cannot measure noise or DR for one pixel only. The signal level we use is the mean value, and the noise level is the standard deviation. As long as there is noise, the mean value or the noise level is not limited by the quantization step.


But DR is limited by one or the other (quantiziation bit-depth or noise) so the effect is the same. In other words, increasing the bit-depth but keeping noise the same just means wasted ADC resolution (Ie, oversampling noise). Conversely, keeping bit-depth the same but decreasing noise means undersampling.
Apr 13, 2015 at 01:47 PM
FM ForumsCanon ForumJoin Upload & Sell
  
 

You are not logged inLog inRegister