Discussion in 'Olympus Cameras' started by kawhona, Oct 6, 2016.
Is there a technical reason that oly isn't producing 14 bit raw files ?
But I think you'd have to talk to an Olympus engineer to get an answer more detailed than that. AFAIK, no micro 4/3 camera does 14 bit files. The benefits of 14 over 12 bit have been dubious, I think.
But, it's 2 more!
Only thing I can think of is space. Three 12-bit entities fit into five bytes (36 bits just a bit too much for 4 bytes), but go to 14 bits and you need a sixth byte to store it all. 20% memory size increase. It may not be their reason, but its all that I can think of. Space.
It's also the entire processing pipeline, not just the storage requirements. All the circuits need to be wider, or 20% faster if they're serial.
In 99.99% of the pictures you won't see any difference anyway, and the remaining 0.01% is mainly self-delution.
Yes, it's more difficult to make good low noise 14bit analog-to-digital converter (ADC) and more data needs higher processing power.
Would there be difference in image quality between 12bit and 14bit files in case of M43 cameras?
Yes, if readout noise would be the same, it would be visible in RAWs at low ISO settings in shadows.
Difference would be similar like in this example. Both images had the same analog signal but quantization of this signal was different, in one case better than in the second one. I don't tell which one is which so everyone can decide how big difference it is.
Remember that this is the difference for the 4 year old Sony's 16MPix sensor which is in most of Olympuses.
Difference in case of new E-M1II would be much more visible if it has really 1 stop advantage over current cameras.
DxO measured the E-M1 dynamic range as 12.7 Ev, or 76.5 dB. 12-bit A/D quantization noise is 83.0 dB below full scale, or -83.0 dB. That implies that the E-M1 noise prior to the A/D is -77.6 dB. 14-bit A/D quantization noise is -95.1 dB. When fed the E-M1 noise, the output would be -77.5 dB. That is 1.0 dB better, or 0.17 Ev. How visible would that be? Is that about what you're showing above?
DxO measured the E-M1 dynamic range as 12.2EV. 12.7EV is for print, not dynamic range for real pixels. If you want measure true dynamic range of sensor, you need to use parameters of sensor in electrons. E-M1 pixels have full well of ~15-16 thousands electrons. Ideally sampled read out noise is ~1.3 electrons (btw, one of the best results for digital cameras).
With ideal ADC with the same noise, E-M1 would have already 13.6EV at base ISO.
With 14bit ADC it would be slightly less but it would be still over 13EV at base ISO.
My example in previous post is real example for my E-PM2.
One image is limited by 12bit ADC (is taken at base ISO) and the second one isn't (is taken at higher ISO settings).
There are of course technical reasons why this isn't so straightforward in practice (I saw examples where ADC adds some noise at low ISO but at high ISO this effect was negligible) but I think that we can expect 14bit ADC in M43 cameras in future.
Thanks for pointing that out! 12.2 EV is 73.5 dB. Analog noise prior to a 12-bit A/D would be -74.0 dB. A 14-bit A/D would yield -74.0 dB output, a 0.5 dB improvement, or 0.08 stops. Is that visible?
If the E-M1 II is one stop better than the E-M1 I, its noise from a 12-bit A/D would be -79.5 dB. Input noise would be -82.0 dB. Output from a 14-bit A/D would be -81.8 dB. That's 2.3 dB better, or 0.38 stops. I think I could see that pretty easily.
It's 12.2EV because it's already limited by its 12bit ADC! So your calculations work with wrong inputs.
Take another look. I work backwards to get A/D input noise.
No, you work with numbers already limited by ADC. If you want to obtain results for base ISO without ADC limitations you need to know readout noise and pixel capacity (full well) before analog-to-digital conversion. Look here please. Especially at figures 4., 5b. and 8a.
Here is theoretical dynamic range for E-M1 without additional noise added by ADC for DxOMark data.
(the result, ~13EV, is worst than my number 13.6EV based on info from the photonstophotos.net, which is maybe too optimistic):
It only matters if you can do the math.
All I'm doing is summing the power of the noise prior to the A/D and the power of the A/D quantization noise to get the output noise power. To get input noise power, I subtract the quantization noise power from the output noise power. That should work unless the two noises are not that statistically independent. I think they should be, except perhaps when their levels are similar. I don't see a problem unless it is related to statistical independence.
Pure math and physical sensor and electron sampling often don't match up exactly. As you say, you're dealing with statistics and probabilities. Oly's 12.2 stops of EV is what we get. We can theorize all we like about what that'll look like in the EM1mk2, but until we get hardware to test it's just that: theoretical. I don't think there'd be a significant boost with a 14bit pipeline and the added circuitry required might even degrade the signal.
also, OMG U GUYZ R SMRT! go take some pictures.
Rob, I'm trying to determine how much a 12-bit A/D might be limiting the performance of the E-M1 I and how much a 14-bit A/D might benefit the E-M1 II. I'm just curious.
I've asked Machi to take our discussion off-line since it seems to disturb some people.
Rob, Math works, including with statistics and probability. Which is why people are able to design the sensors which you are using to take pictures. Feel free to ignore the thread if it bothers you. There may be others who are interested.
I, for one, would like to see this continued online. The OP was a technical question and I think that your discussion has been pretty much on topic.
Separate names with a comma.