You can scale a 16 bit number into a 14 bit register..... and you end up with 14 bits of precision.... you try scaling it back up and you get two bits of random noise added on to the bottom of the signal.Orangutan said:
Upvote
0
EOS-1D X Mark II Claims of 15 Stops of DR [CR3]
- Thread starter Canon Rumors Guy
- Start date
Don Haines said:
Dead on. Once you discard information you cannot get it back. You can reorganize information though, and a compression curve applied to the raw oixel data before ADC could allow 15 stops of information to be compressed more usefully into 14 bits with only a small loss.
Upvote
0
I'd love to hear the fuller explanation of that. From what little I understand, the sensor has a capacity for a certain number of electrons. Ideally, you want to just count those electrons. For example, the 1DX has 88,600 (per Clarkvision). That's would require 17 bits to represent directly, so some magic is already occurring to scale that to 14 bits. Part of that, I presume, comes from "removing the noise." With my almost-nonexistent understanding of signal processing, I'm at a loss to understand how it's inherently harder to convert 17 bits of electrons to 14 bits of data than it is to convert 17 bits of electrons to 15 bits of data. (I'm assuming the 1DX2 will not have significantly higher FWC due to slightly smaller pixel size). I.e., it's not binary data until after the signal processing is done, so the analog signal can be sliced into slabs of any desired size.Don Haines said:
Again, I confess ignorance and curiosity.
Upvote
0
FWC can't be greater with smaller pixels on a denser sensor (22MP vs 18MP ... i Don't believe) But then again, as has been discussed at length on this forum over the years, the problem Canon sensors truly suffer from is the read noise, which robs a great deal of the native DR they actually capture prior to being processed OFF chip on a separate ADC. Now we have a FF sensor (NOT a super 35 like the C300II) with ON chip ADC. Let's take the 5D3 sensor which is capable of more DR than the rest of the system allows and make in ON chip ADC. I think it's entirely plausible that we see a lot more of its native capabilities (13 stops + perhaps) than we have before because we aren't losing signal to noise introduced along a signal path that no longer exists since the ADC process is now on board. That alone right there could make up a pretty significant portion of a cleaner signal, and that's NOT taking into account the newer die process Canon is apparently using since after the 5D3 to create the pixels for the 7D2 and 5DSR, neither of which benefitted from on chip ADC. The 1DX2 will be the true marker for Canon going forward. I think I got most everything right, but I'll bow to Jon's wisdom on the math here.
Orangutan said:
Upvote
0
jrista said:If true, this is very welcome. About time. That said, I'll believe it when I see it.
I am skeptical, as they claimed this before with the C300 II, and in testing it wasn't close. I am also curious whether it is an actual linear DR increase at the hardware level, or due to some kind if processing curve.
There was a follow up article including a response from Canon that indicated there were different camera settings needed to extract the maximum dynamic range from the sensor as well as showing how Canon is defining dynamic range a bit differently from Cinema5D. By Canon's definition, which is similar to how DxO defines it, and the "optimal" camera settings they were indeed hitting 15 stops.
Upvote
0
Video is a funny animal because most capture isnt done in RAW mode, whereas nearly all pro work in stills is. Like Sony S-Log and Canon C-Log, you need to use certain image profiles to get the flattest image and widest DR possible from the camera. Where as shooting in RAW, you get everything no matter what. All provided of course you expose correctly (no lens cap shots). I'd love to read the article you're speaking of. Have a link?
raptor3x said:jrista said:If true, this is very welcome. About time. That said, I'll believe it when I see it.
I am skeptical, as they claimed this before with the C300 II, and in testing it wasn't close. I am also curious whether it is an actual linear DR increase at the hardware level, or due to some kind if processing curve.
There was a follow up article including a response from Canon that indicated there were different camera settings needed to extract the maximum dynamic range from the sensor as well as showing how Canon is defining dynamic range a bit differently from Cinema5D. By Canon's definition, which is similar to how DxO defines it, and the "optimal" camera settings they were indeed hitting 15 stops.
Upvote
0
So they ARE using a compression curve to achieve it. Well, that is less than ideal. Sounds much the same as Sony using craw to preserve dynamic range with their lossy raw compression. Bummer. Definitely not good for Astro...we need linear signals. Might be fine for landscapes though.
Upvote
0
frankchn said:
This is very interesting/scandalous. Makes for good yellow journalism/blog.
It's all video measurements so there's a ton of post processing on top of everything else, whereas DXOmark (faulty in its own respects) at least measures raw readout data... less useful, but... more meaningful for stills.
It seems like Canon has sensor technology that's on par or close enough to on par with Sony's that for stills you're getting the same DR. For video it should serve you well, too.
The Alexa is special. Its dual gain path is magic. Watch the Revenant. Nothing compares for pure high quality capture, but they're using a 65mm (medium format) sensor to get there and that sensor is an "unofficially" dramatically improved version of a dual gain path-way-out-of-your-league 90w-to-use-it monster of a sensor so....
What we can take from this is: Canon decided to catch up with Sony.
The D5 and 1DX2 might have a lot in common. But with Canon and Nikon trading advantages...
...and ideally both pushing for huge innovation in their competitors, by pushing their specific buttons.....
Game on.
Upvote
0
Mt Spokane Photography said:
I do find this statement a bit odd, though I agree with some. When my camera clips highlights, and I try to expose for the highlights, making me raise the shadows, with noise and banding, the 11 stops of DR my sensor gives me is not enough. That does not make a 15 stop DR - sensor not necassary. Perhaps 13 stop would do in my case, which is still 2 stop more than I get from my current setup. Hence, a 15 stop sensor would be great, and more than enough.
Upvote
0
In general it is even possible to squeze 20db of scene dynamic range into 14 bit ADC output.jrista said:
For best results it should be of course analog signal compression before ADC input and not digital compression like done by Sony.
Kind of S-Log analog compressor curcuit which could be switched on/off and also have adjustable gain curve via camera settings.
Problem though that using 14 bit ADC would not give any real gain for tonal range which is important. So higher than 14 stop per pixel DR using 14 bit ADC still would be some compromise/tradeoff which is not OK for some of the applications where tonal range fidelity is important.
Upvote
0
If Canon manages to match Sony/Nikon in the low ISO dynamic range arena, that will be a massive achievement already.
Now, can they get their DPAF to work in AF Servo mode? That will totally awesome.
What about getting their standard AF sensor to match Nikon's 3D tracking capabilities?
Now, can they get their DPAF to work in AF Servo mode? That will totally awesome.
What about getting their standard AF sensor to match Nikon's 3D tracking capabilities?
Upvote
0
Pitbullo said:
Mt Spokane statement is perfectly correct.
DR is the ratio between most strongest and most weakest signals.
Strongest signal is limited by fotocell saturation point, weakest signal by the circuit noise floor.
So analog DR (before ADC) could be improved by raising saturation point with given circuit noise floor level or by reducing noise level or doing both at the same time.
Then 14bit ADC would be limiting factor, so to squeze more analog signal DR to ADC limits it would be requied to use pre-ADC analog compression circuit (with S-log gain curve).
On the other hand statement that 15 stops DR is more than enough is more than odd /strange.
More DR means better overall sensor quality especially more DR at high ISO.
So having more DR is the same as having more money - the more the better )
Upvote
0
Policar said:
Cinema5D doesn't measure sensor DR, it's measurement is based on 'how people usually shoot'. So, he uses different exposures and ISOs for different cameras. If you did that for a still camera you would be called a moron.
Upvote
0
Woody said:
For 1DX II I am more interested in high ISO DR improvements and do not care much about low ISO DR. If it be better then good, if the same as before I am OK with that.
For general walkaround camera I already have better performance within ISO range from 100 to 12800 using my Sony a7s and a7rII than using my 1DX .
But high ISO performance for 1DX II is extremely important for this camera intended applications.
My wish that it would be at least 1 stop better than a7s and a7rII at high ISO.
Best would be to have 1 stop better performance at ISO 25600 than Sony a7s.
Upvote
0
Here are your 15 stops broken down to discreet levels. Each stop must per definition be a doubling of the previous value in a linear representation:
0. Stop: 0-1
1. Stop: 1-2
2. Stop: 2-4
3. Stop: 4-8
4. Stop: 8-16
5. Stop: 16-32
6. Stop: 32-64
7. Stop: 64-128
8. Stop: 128-256
9. Stop: 256-512
10. Stop: 512-1024
11. Stop: 1024-2048
12. Stop: 2048-4096
13. Stop: 4096-8192
14. Stop: 8192-16384
15. Stop: 16384-32768
Since it would be very impractical to assign 1 value to the first stop and 16384 values to the last stop, we use a logarithmic function to distribute the values in a non-linear fashion. The log curve will assign roughly the same amount of discreet values per stop.
Now, you might think that you can take these values and, with the help of a log curve, break them out to even more "stops" by using a flatter profile. But the only thing that means is that you start to define fractions of a real, linear light stop. So that isn't possible.
In practical terms, on a signal level, Dynamic Range (DR) in dB is calculated:
DR (db) = 20 log (Peak signal at Full Well Saturation/r.m.s. Noise)
For an image sensor, in essence, each stop will correspond to a 6dB change.
15 stops would in theory need a 90dB sensor. The C300 mkII achieves 67dB (which Canon is pretty proud of). 67dB is equivalent to 11.17 stops. So, there is a discrepancy here.
My own little theory is that cameras recording log images do some significant pulling of the signal (under exposing) and then use the ISO to boost the levels back up (and recover the shadows). Kind of like a dual ISO thing.
Meaning if you shoot at ISO800 you get 3 stops highlight protection (based off of ISO100). 9 real stops + 3 ISO stops = 12 stops. A Sony that requires ISO3200 for log would give you a whopping 5 stops extra. 9 real stops + 5 ISO stops = 14 stops.
This technique would be especially advantageous if the sensor has low readout noise—something the Sony sensors have been good at. And now the new generation of Canon sensors too, I think.
Another example of this is Canon's HTP that gives you an additional stop of highlight recovery, but the lowest available ISO is 200. Or, as in my 1Dc: ISO400 (required for log) would give me 10+2=12 stops.
0. Stop: 0-1
1. Stop: 1-2
2. Stop: 2-4
3. Stop: 4-8
4. Stop: 8-16
5. Stop: 16-32
6. Stop: 32-64
7. Stop: 64-128
8. Stop: 128-256
9. Stop: 256-512
10. Stop: 512-1024
11. Stop: 1024-2048
12. Stop: 2048-4096
13. Stop: 4096-8192
14. Stop: 8192-16384
15. Stop: 16384-32768
Since it would be very impractical to assign 1 value to the first stop and 16384 values to the last stop, we use a logarithmic function to distribute the values in a non-linear fashion. The log curve will assign roughly the same amount of discreet values per stop.
Now, you might think that you can take these values and, with the help of a log curve, break them out to even more "stops" by using a flatter profile. But the only thing that means is that you start to define fractions of a real, linear light stop. So that isn't possible.
In practical terms, on a signal level, Dynamic Range (DR) in dB is calculated:
DR (db) = 20 log (Peak signal at Full Well Saturation/r.m.s. Noise)
For an image sensor, in essence, each stop will correspond to a 6dB change.
15 stops would in theory need a 90dB sensor. The C300 mkII achieves 67dB (which Canon is pretty proud of). 67dB is equivalent to 11.17 stops. So, there is a discrepancy here.
My own little theory is that cameras recording log images do some significant pulling of the signal (under exposing) and then use the ISO to boost the levels back up (and recover the shadows). Kind of like a dual ISO thing.
Meaning if you shoot at ISO800 you get 3 stops highlight protection (based off of ISO100). 9 real stops + 3 ISO stops = 12 stops. A Sony that requires ISO3200 for log would give you a whopping 5 stops extra. 9 real stops + 5 ISO stops = 14 stops.
This technique would be especially advantageous if the sensor has low readout noise—something the Sony sensors have been good at. And now the new generation of Canon sensors too, I think.
Another example of this is Canon's HTP that gives you an additional stop of highlight recovery, but the lowest available ISO is 200. Or, as in my 1Dc: ISO400 (required for log) would give me 10+2=12 stops.
Upvote
0
dilbert said:
I know, I'm only addressing the movie mode here. Should have made that clear.
Upvote
0
Similar threads
