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

  
 
curious80
Offline
Details
p.3 #1High DR for 5d4
snapsy wrote:
snapsy wrote
curious80 wrote
snapsy wrote
curious80 wrote
snapsy wroteGamma conversion from raw to a color space can be non-linear and thus is not inhibited by bit depth. But independent of bit depth the A7s sensor is only capable of 13EV even though the implied linear limit would be 14EV for 14-bits (irrespective of Sony's tonal compression).
But raw image files don't store gamma converted data, they store the linear data, and with that you are limited to 14EV of maximum DR with 14-bits. And that too only if your A/D is up to the task. A 14-bit A/D doesn't produce actual 14-bit worth of data.

Note that cameras like D810 which have over 14 stops of DR on DxoMark have that DR when down-sampled to 8MP, which effectively gives you more bits per downsampled pixel. At pixel level D810 has 13.67EV of DR as per DxoMark (which admittedly is still very high for a 14-bit sample).

A7s is at a disadvantage here because it is natively only 12MP and each of its fat pixels gets only 14-bits to represent it. SLOG2 provides a non-linear mapping of the type that you are talking about allowing storing more stops of DR to be stores in 14-bits.
But the gamma conversion is non-linear during demosaicing, the same technique performed for slog2 that you ascribed to its ability to exceed the native bit range for DR.
You are missing the point that once the linear data has been sampled by a 14-bit A/D and stored in the RAW file as 14-bits then any additional data that was captured by the sensor in the 15th or 16-bit has already been thrown away. There is nothing you can do to reclaim that. The only way you can save any extra information captured by the sensor is if the non-linear mapping is done in the sensor itself before it is sampled by the A/D. Which I believe is what modes like SLOG2 do.
That's true, Sony may be doing non-linear encoding of the sensor values at the ADC for slog2. However that doesn't really increase captured DR when the underlying pixels and their charge capture+transfer mechanism is still principally linear. It simply applies a tone curve earlier in the process (against the raw data) so that the encoded ADC values can later be gamma-interpreted to a > 2^14 of tonal space. I'm not seeing how that's fundamentally different than simply re-interpreting non-slog2 raw data in the same fashion after the fact. Perhaps I'm overlooking something though.

Btw I'm less inclined to believe the A7s's true sensor-level DR is 15.3EV that can only be unleased via slog2 when you consider that slog2 is only available on the A7s at ISO 3200, which has 1/5 the FWC (and thus DR) than ISO 100.
Thinking it through more, an ADC-based slog2 would be fundamentally different vs simply reinterpreting non-slog2 raw data if the native DR capability of the sensor is greater than the underlying ADC bit depth + ADC noise. But I don't see the A7s's FWC and charge-transfer noise supporting 15.3EV and its on-chip parallel ADC adding an entire 1EV of noise. Also, considering that the A7s has only 13EV of DR at ISO 100 (post-ADC) I don't see how it can how slog2 can achieve 15.3EV of sensor-level DR when it operates at ISO 3200 for slog2.


Well I don't really know too much about the performance of today's ADCs. When I dealt with ADCs, 13 effective bits out of a 14-bit ADC was really as good as you could expect and the 13.67 EV DR that D810 pulls out of a 14-bit ADC would have been the more surprising part.

I am not saying that A7s sensor does have 15.3 EV of DR. Maybe you are right and it doesn't. All I am saying is that IF it does indeed have 15.3EV of DR then it could not be captured in a 14-bit RAW file and will require something like SLOG2 to be able to capture that.

Also I suspect you can't really interpret the ISO's the same way as a linear RAW file when you are dealing with modes like SLOG2. I am 99% sure that in the ISO3200 SLOG2 mode A7s uses its FWC and not 1/5 FWC.
Apr 10, 2015 at 06:21 PM
snapsy
Offline
Details
p.3 #2High DR for 5d4
curious80 wrote:
snapsy wrote
snapsy wrote
curious80 wrote
snapsy wrote
curious80 wroteBut raw image files don't store gamma converted data, they store the linear data, and with that you are limited to 14EV of maximum DR with 14-bits. And that too only if your A/D is up to the task. A 14-bit A/D doesn't produce actual 14-bit worth of data.

Note that cameras like D810 which have over 14 stops of DR on DxoMark have that DR when down-sampled to 8MP, which effectively gives you more bits per downsampled pixel. At pixel level D810 has 13.67EV of DR as per DxoMark (which admittedly is still very high for a 14-bit sample).

A7s is at a disadvantage here because it is natively only 12MP and each of its fat pixels gets only 14-bits to represent it. SLOG2 provides a non-linear mapping of the type that you are talking about allowing storing more stops of DR to be stores in 14-bits.
But the gamma conversion is non-linear during demosaicing, the same technique performed for slog2 that you ascribed to its ability to exceed the native bit range for DR.
You are missing the point that once the linear data has been sampled by a 14-bit A/D and stored in the RAW file as 14-bits then any additional data that was captured by the sensor in the 15th or 16-bit has already been thrown away. There is nothing you can do to reclaim that. The only way you can save any extra information captured by the sensor is if the non-linear mapping is done in the sensor itself before it is sampled by the A/D. Which I believe is what modes like SLOG2 do.
That's true, Sony may be doing non-linear encoding of the sensor values at the ADC for slog2. However that doesn't really increase captured DR when the underlying pixels and their charge capture+transfer mechanism is still principally linear. It simply applies a tone curve earlier in the process (against the raw data) so that the encoded ADC values can later be gamma-interpreted to a > 2^14 of tonal space. I'm not seeing how that's fundamentally different than simply re-interpreting non-slog2 raw data in the same fashion after the fact. Perhaps I'm overlooking something though.

Btw I'm less inclined to believe the A7s's true sensor-level DR is 15.3EV that can only be unleased via slog2 when you consider that slog2 is only available on the A7s at ISO 3200, which has 1/5 the FWC (and thus DR) than ISO 100.
Thinking it through more, an ADC-based slog2 would be fundamentally different vs simply reinterpreting non-slog2 raw data if the native DR capability of the sensor is greater than the underlying ADC bit depth + ADC noise. But I don't see the A7s's FWC and charge-transfer noise supporting 15.3EV and its on-chip parallel ADC adding an entire 1EV of noise. Also, considering that the A7s has only 13EV of DR at ISO 100 (post-ADC) I don't see how it can how slog2 can achieve 15.3EV of sensor-level DR when it operates at ISO 3200 for slog2.
Well I don't really know too much about the performance of today's ADCs. When I dealt with ADCs, 13 effective bits out of a 14-bit ADC was really as good as you could expect and the 13.67 EV DR that D810 pulls out of a 14-bit ADC would have been the more surprising part.

I am not saying that A7s sensor does have 15.3 EV of DR. Maybe you are right and it doesn't. All I am saying is that IF it does indeed have 15.3EV of DR then it could not be captured in a 14-bit RAW file and
...Show more →

Based on the midtone noise performance of the A7s's slog2 video vs A7s's non-slog2 ISO 100 video it's definitely operating at the 1/5 FWC. The footage has the noise-equivalent of a 12MP ISO 3200 still downsampled to the equivalent 1080 or 4K video resolution.
Apr 10, 2015 at 06:33 PM
curious80
Offline
Details
p.3 #3High DR for 5d4
snapsy wrote:
curious80 wrote
snapsy wrote
snapsy wrote
curious80 wrote
snapsy wroteBut the gamma conversion is non-linear during demosaicing, the same technique performed for slog2 that you ascribed to its ability to exceed the native bit range for DR.
You are missing the point that once the linear data has been sampled by a 14-bit A/D and stored in the RAW file as 14-bits then any additional data that was captured by the sensor in the 15th or 16-bit has already been thrown away. There is nothing you can do to reclaim that. The only way you can save any extra information captured by the sensor is if the non-linear mapping is done in the sensor itself before it is sampled by the A/D. Which I believe is what modes like SLOG2 do.
That's true, Sony may be doing non-linear encoding of the sensor values at the ADC for slog2. However that doesn't really increase captured DR when the underlying pixels and their charge capture+transfer mechanism is still principally linear. It simply applies a tone curve earlier in the process (against the raw data) so that the encoded ADC values can later be gamma-interpreted to a > 2^14 of tonal space. I'm not seeing how that's fundamentally different than simply re-interpreting non-slog2 raw data in the same fashion after the fact. Perhaps I'm overlooking something though.

Btw I'm less inclined to believe the A7s's true sensor-level DR is 15.3EV that can only be unleased via slog2 when you consider that slog2 is only available on the A7s at ISO 3200, which has 1/5 the FWC (and thus DR) than ISO 100.
Thinking it through more, an ADC-based slog2 would be fundamentally different vs simply reinterpreting non-slog2 raw data if the native DR capability of the sensor is greater than the underlying ADC bit depth + ADC noise. But I don't see the A7s's FWC and charge-transfer noise supporting 15.3EV and its on-chip parallel ADC adding an entire 1EV of noise. Also, considering that the A7s has only 13EV of DR at ISO 100 (post-ADC) I don't see how it can how slog2 can achieve 15.3EV of sensor-level DR when it operates at ISO 3200 for slog2.
Well I don't really know too much about the performance of today's ADCs. When I dealt with ADCs, 13 effective bits out of a 14-bit ADC was really as good as you could expect and the 13.67 EV DR that D810 pulls out of a 14-bit ADC would have been the more surprising part.

I am not saying that A7s sensor does have 15.3 EV of DR. Maybe you are right and it doesn't. All I am saying is that IF it does indeed have 15.3EV of DR then it could not be captured in a 14-bit RAW file and will require something like SLOG2 to be able to capture that.

Also I suspect you can't really interpret the ISO's the same way as a linear RAW file when you are dealing with modes like SLOG2. I am 99% sure that in the ISO3200 SLOG2 mode A7s would uses its FWC and not 1/5 FWC.
Based on the midtone noise performance of the A7s's slog2 video vs A7s's non-slog2 ISO 100 video it's definitely operating at the 1/5 FWC. The footage has the noise-equivalent of a 12MP ISO 3200 still downsampled to the equivalent 1080 or 4K video resolution.


You really have to understand that it is a non-linear gain curve unlike the normal ISO ratings where each intensity gets amplified by the same gain. In normal linear ISO3200 mode the all the values get multiplied basically by 32. So while you bring up the low intensities to high level, all the high intensities above the 1/32 of FWC get saturated. Which is why you only use 1/32 of the FWC. In the SLOG2 mode however the gain curve is such that the lower intensities get amplified a lot more than the higher intensities (which is what a log function does). So the gain curve is such that at the canonical 18% gray it will indeed correspond to a gain level of 32 consistent with ISO 3200. However the gain will be lower and lower as you go up the intensity level until at the top of the gain curve the gain would be 1 such that you will hit the saturation point at the same level as if you were shooting ISO 100. And in the process it will retain all the dynamic range.

Does that make sense?

Apr 10, 2015 at 07:04 PM
curious80
Offline
Details
p.3 #4High DR for 5d4
Basically calling this mode as ISO3200 is just an arbitrary choice, since the "ISO gain" in this case is different for each part of the intensity curve, . If you go by the definition that DxoMark uses for ISO, then this would be ISO 100 because DxoMark uses a definition based on the saturation point which is independent of the tone curve. The definition used by camera manufacturers however is based around 18% gray which would make this ISO3200.
Apr 10, 2015 at 07:16 PM
snapsy
Offline
Details
p.3 #5High DR for 5d4
curious80 wrote:
snapsy wrote
curious80 wrote
snapsy wrote
snapsy wrote
curious80 wroteYou are missing the point that once the linear data has been sampled by a 14-bit A/D and stored in the RAW file as 14-bits then any additional data that was captured by the sensor in the 15th or 16-bit has already been thrown away. There is nothing you can do to reclaim that. The only way you can save any extra information captured by the sensor is if the non-linear mapping is done in the sensor itself before it is sampled by the A/D. Which I believe is what modes like SLOG2 do.
That's true, Sony may be doing non-linear encoding of the sensor values at the ADC for slog2. However that doesn't really increase captured DR when the underlying pixels and their charge capture+transfer mechanism is still principally linear. It simply applies a tone curve earlier in the process (against the raw data) so that the encoded ADC values can later be gamma-interpreted to a > 2^14 of tonal space. I'm not seeing how that's fundamentally different than simply re-interpreting non-slog2 raw data in the same fashion after the fact. Perhaps I'm overlooking something though.

Btw I'm less inclined to believe the A7s's true sensor-level DR is 15.3EV that can only be unleased via slog2 when you consider that slog2 is only available on the A7s at ISO 3200, which has 1/5 the FWC (and thus DR) than ISO 100.
Thinking it through more, an ADC-based slog2 would be fundamentally different vs simply reinterpreting non-slog2 raw data if the native DR capability of the sensor is greater than the underlying ADC bit depth + ADC noise. But I don't see the A7s's FWC and charge-transfer noise supporting 15.3EV and its on-chip parallel ADC adding an entire 1EV of noise. Also, considering that the A7s has only 13EV of DR at ISO 100 (post-ADC) I don't see how it can how slog2 can achieve 15.3EV of sensor-level DR when it operates at ISO 3200 for slog2.
Well I don't really know too much about the performance of today's ADCs. When I dealt with ADCs, 13 effective bits out of a 14-bit ADC was really as good as you could expect and the 13.67 EV DR that D810 pulls out of a 14-bit ADC would have been the more surprising part.

I am not saying that A7s sensor does have 15.3 EV of DR. Maybe you are right and it doesn't. All I am saying is that IF it does indeed have 15.3EV of DR then it could not be captured in a 14-bit RAW file and will require something like SLOG2 to be able to capture that.

Also I suspect you can't really interpret the ISO's the same way as a linear RAW file when you are dealing with modes like SLOG2. I am 99% sure that in the ISO3200 SLOG2 mode A7s would uses its FWC and not 1/5 FWC.
Based on the midtone noise performance of the A7s's slog2 video vs A7s's non-slog2 ISO 100 video it's definitely operating at the 1/5 FWC. The footage has the noise-equivalent of a 12MP ISO 3200 still downsampled to the equivalent 1080 or 4K video resolution.
You really have to understand that it is a non-linear gain curve unlike the normal ISO ratings where each intensity gets amplified by the same gain. In normal linear ISO3200 mode the all the values get multiplied basically by 32. So while you bring up the low intensities to high level, all the high intensities above the 1/32 of FWC get saturated. Which is why you only use 1/32 of the FWC. In the SLOG2 mode however the gain curve is such that the lower intensities get amplified a lot more than the higher intensities (which is what a log
...Show more →

That would imply that the A7s's ADC is capable of per-pixel variable gain at the 30hz full-sensor sampling rate the A7s supports for video. I'm not saying that's impossble but I've never heard of a sensor that can do that.
Apr 10, 2015 at 07:18 PM
curious80
Offline
Details
p.3 #6High DR for 5d4
snapsy wrote:
curious80 wrote
snapsy wrote
curious80 wrote
snapsy wrote
snapsy wroteThat's true, Sony may be doing non-linear encoding of the sensor values at the ADC for slog2. However that doesn't really increase captured DR when the underlying pixels and their charge capture+transfer mechanism is still principally linear. It simply applies a tone curve earlier in the process (against the raw data) so that the encoded ADC values can later be gamma-interpreted to a > 2^14 of tonal space. I'm not seeing how that's fundamentally different than simply re-interpreting non-slog2 raw data in the same fashion after the fact. Perhaps I'm overlooking something though.

Btw I'm less inclined to believe the A7s's true sensor-level DR is 15.3EV that can only be unleased via slog2 when you consider that slog2 is only available on the A7s at ISO 3200, which has 1/5 the FWC (and thus DR) than ISO 100.
Thinking it through more, an ADC-based slog2 would be fundamentally different vs simply reinterpreting non-slog2 raw data if the native DR capability of the sensor is greater than the underlying ADC bit depth + ADC noise. But I don't see the A7s's FWC and charge-transfer noise supporting 15.3EV and its on-chip parallel ADC adding an entire 1EV of noise. Also, considering that the A7s has only 13EV of DR at ISO 100 (post-ADC) I don't see how it can how slog2 can achieve 15.3EV of sensor-level DR when it operates at ISO 3200 for slog2.
Well I don't really know too much about the performance of today's ADCs. When I dealt with ADCs, 13 effective bits out of a 14-bit ADC was really as good as you could expect and the 13.67 EV DR that D810 pulls out of a 14-bit ADC would have been the more surprising part.

I am not saying that A7s sensor does have 15.3 EV of DR. Maybe you are right and it doesn't. All I am saying is that IF it does indeed have 15.3EV of DR then it could not be captured in a 14-bit RAW file and will require something like SLOG2 to be able to capture that.

Also I suspect you can't really interpret the ISO's the same way as a linear RAW file when you are dealing with modes like SLOG2. I am 99% sure that in the ISO3200 SLOG2 mode A7s would uses its FWC and not 1/5 FWC.
Based on the midtone noise performance of the A7s's slog2 video vs A7s's non-slog2 ISO 100 video it's definitely operating at the 1/5 FWC. The footage has the noise-equivalent of a 12MP ISO 3200 still downsampled to the equivalent 1080 or 4K video resolution.
You really have to understand that it is a non-linear gain curve unlike the normal ISO ratings where each intensity gets amplified by the same gain. In normal linear ISO3200 mode the all the values get multiplied basically by 32. So while you bring up the low intensities to high level, all the high intensities above the 1/32 of FWC get saturated. Which is why you only use 1/32 of the FWC. In the SLOG2 mode however the gain curve is such that the lower intensities get amplified a lot more than the higher intensities (which is what a log function does). So the gain curve is such that at the canonical 18% gray it will indeed correspond to a gain level of 32 consistent with ISO 3200. However the gain will be lower and lower as you go up the intensity level until at the top of the gain curve the gain would be 1 such that you will hit the saturation point at the same level as if you were shooting ISO 100. And in the process it will retain all the dynamic range.

Does that make sense?
That would imply that the A7s's ADC is capable of per-pixel variable gain at the 30hz full-sensor sampling rate the A7s supports for video. I'm not saying that's impossble but I've never heard of a sensor that can do that.


I don't know the details of how it gets implemented in hardware. But this is what the SLOG2 mode does. Look there is no way to create extra dynamic range in the software by any tonal manipulation, even more so for A7s where the 4K stream is a direct 1-to-1 pixel readout without any down-sampling.

Apr 10, 2015 at 07:27 PM
kezeka
Offline
Details
p.3 #7High DR for 5d4
NCAndy wrote:
arbitrage wroteAnd for those not knowing what the C300II is doing it is taking two readouts at two different ISOs and combining them for a claimed 15 stops of DR. Canon even mentions the DR in the C300II press release where as they have never mentioned number of stops before.
Hopefully those two readouts are instantaneous at the time of exposure so they blend perfectly into one raw file with greater DR. As I understand it, the ML solution is to take two exposures at different ISOs to blend in software.


I believe that ML stripes the sensor into half low ISO half high ISO (100/3200 or 1600 in most cases) and reads them SIMULTANEOUSLY. Hence the loss of half of the resolution in your highlights and half of the resolution in your low light areas but in doing so you gain a few stops of DR. Useful when you don't give a fruit loop about resolution and need DR. This is why you can take 1/4000s photos of race cars and not have weird artifacts from two separate exposures being blended.

If Canon is implementing this at the hardware level, one would hope they have boosted the sensor resolution somewhat significantly (30mp maybe?) to compensate for the loss of resolution at the far reaches of the DR bell curve in exposures.
Apr 10, 2015 at 08:48 PM
johnctharp
Offline
Details
p.3 #8High DR for 5d4
curious80 wrote:
Does that make sense?



That's actually very well put, and makes more sense than it should
Apr 10, 2015 at 08:57 PM
StillFingerz
Offline
Details
p.3 #9High DR for 5d4
These last few replies, well most in this DR discussion, are simply wonderful, informative, and as civilized as could be, bravo gentlemen, do keep this going. I understand little that DXO provides, but do enjoy this level of banter and knowledge

Cheers,
Jerry
Apr 10, 2015 at 09:14 PM
snapsy
Offline
Details
p.3 #10High DR for 5d4
curious80 wrote:
snapsy wrote
curious80 wrote
snapsy wrote
curious80 wrote
snapsy wroteThinking it through more, an ADC-based slog2 would be fundamentally different vs simply reinterpreting non-slog2 raw data if the native DR capability of the sensor is greater than the underlying ADC bit depth + ADC noise. But I don't see the A7s's FWC and charge-transfer noise supporting 15.3EV and its on-chip parallel ADC adding an entire 1EV of noise. Also, considering that the A7s has only 13EV of DR at ISO 100 (post-ADC) I don't see how it can how slog2 can achieve 15.3EV of sensor-level DR when it operates at ISO 3200 for slog2.
Well I don't really know too much about the performance of today's ADCs. When I dealt with ADCs, 13 effective bits out of a 14-bit ADC was really as good as you could expect and the 13.67 EV DR that D810 pulls out of a 14-bit ADC would have been the more surprising part.

I am not saying that A7s sensor does have 15.3 EV of DR. Maybe you are right and it doesn't. All I am saying is that IF it does indeed have 15.3EV of DR then it could not be captured in a 14-bit RAW file and will require something like SLOG2 to be able to capture that.

Also I suspect you can't really interpret the ISO's the same way as a linear RAW file when you are dealing with modes like SLOG2. I am 99% sure that in the ISO3200 SLOG2 mode A7s would uses its FWC and not 1/5 FWC.
Based on the midtone noise performance of the A7s's slog2 video vs A7s's non-slog2 ISO 100 video it's definitely operating at the 1/5 FWC. The footage has the noise-equivalent of a 12MP ISO 3200 still downsampled to the equivalent 1080 or 4K video resolution.
You really have to understand that it is a non-linear gain curve unlike the normal ISO ratings where each intensity gets amplified by the same gain. In normal linear ISO3200 mode the all the values get multiplied basically by 32. So while you bring up the low intensities to high level, all the high intensities above the 1/32 of FWC get saturated. Which is why you only use 1/32 of the FWC. In the SLOG2 mode however the gain curve is such that the lower intensities get amplified a lot more than the higher intensities (which is what a log function does). So the gain curve is such that at the canonical 18% gray it will indeed correspond to a gain level of 32 consistent with ISO 3200. However the gain will be lower and lower as you go up the intensity level until at the top of the gain curve the gain would be 1 such that you will hit the saturation point at the same level as if you were shooting ISO 100. And in the process it will retain all the dynamic range.

Does that make sense?
That would imply that the A7s's ADC is capable of per-pixel variable gain at the 30hz full-sensor sampling rate the A7s supports for video. I'm not saying that's impossble but I've never heard of a sensor that can do that.
I don't know the details of how it gets implemented in hardware. But this is what the SLOG2 mode does. Look there is no way to create extra dynamic range in the software by any tonal manipulation, even more so for A7s where the 4K stream is a direct 1-to-1 pixel readout without any down-sampling.


I agree, which gets us back to the original discussion on whether the A7s has 15.3EV of DR that's only reachable via slog2 or that slog2 even yields more capture DR. If slog2 turns out to be a tone curve applied to the raw data that doesn't actually change the operation of the ADC in terms of variable gain then slog2 is simply a mechanism to compress a steeper tone curve into the existing 2^14 range of encoded data (or actually 2^8 since the encoded video output of the A7s is 8 bits rather than 14, even for its uncompressed HDMI output), resulting in a greater output/rendered DR but no effect on actual sensor-level captured DR.
Apr 11, 2015 at 08:52 AM
curious80
Offline
Details
p.3 #11High DR for 5d4
snapsy wrote:
curious80 wrote
snapsy wrote
curious80 wrote
snapsy wrote
curious80 wroteWell I don't really know too much about the performance of today's ADCs. When I dealt with ADCs, 13 effective bits out of a 14-bit ADC was really as good as you could expect and the 13.67 EV DR that D810 pulls out of a 14-bit ADC would have been the more surprising part.

I am not saying that A7s sensor does have 15.3 EV of DR. Maybe you are right and it doesn't. All I am saying is that IF it does indeed have 15.3EV of DR then it could not be captured in a 14-bit RAW file and will require something like SLOG2 to be able to capture that.

Also I suspect you can't really interpret the ISO's the same way as a linear RAW file when you are dealing with modes like SLOG2. I am 99% sure that in the ISO3200 SLOG2 mode A7s would uses its FWC and not 1/5 FWC.
Based on the midtone noise performance of the A7s's slog2 video vs A7s's non-slog2 ISO 100 video it's definitely operating at the 1/5 FWC. The footage has the noise-equivalent of a 12MP ISO 3200 still downsampled to the equivalent 1080 or 4K video resolution.
You really have to understand that it is a non-linear gain curve unlike the normal ISO ratings where each intensity gets amplified by the same gain. In normal linear ISO3200 mode the all the values get multiplied basically by 32. So while you bring up the low intensities to high level, all the high intensities above the 1/32 of FWC get saturated. Which is why you only use 1/32 of the FWC. In the SLOG2 mode however the gain curve is such that the lower intensities get amplified a lot more than the higher intensities (which is what a log function does). So the gain curve is such that at the canonical 18% gray it will indeed correspond to a gain level of 32 consistent with ISO 3200. However the gain will be lower and lower as you go up the intensity level until at the top of the gain curve the gain would be 1 such that you will hit the saturation point at the same level as if you were shooting ISO 100. And in the process it will retain all the dynamic range.

Does that make sense?
That would imply that the A7s's ADC is capable of per-pixel variable gain at the 30hz full-sensor sampling rate the A7s supports for video. I'm not saying that's impossble but I've never heard of a sensor that can do that.
I don't know the details of how it gets implemented in hardware. But this is what the SLOG2 mode does. Look there is no way to create extra dynamic range in the software by any tonal manipulation, even more so for A7s where the 4K stream is a direct 1-to-1 pixel readout without any down-sampling.
I agree, which gets us back to the original discussion on whether the A7s has 15.3EV of DR that's only reachable via slog2 or that slog2 even yields more capture DR. If slog2 turns out to be a tone curve applied to the raw data that doesn't actually change the operation of the ADC in terms of variable gain then slog2 is simply a mechanism to compress a steeper tone curve into the existing 2^14 range of encoded data (or actually 2^8 since the encoded video output of the A7s is 8 bits rather than 14, even for its uncompressed
...Show more →

It is pointless to apply slog2 after you have already converted to 14-bits or 8-bits. If you did, it will just have a hole in the data with nothing at the values corresponding to the extra stops of DR. It will only harm you because you will use even fewer bits for the data that you actually have and will assign some of the bits to the data which does not exist.

Of course just because slog2 captures 15.3EV of data off the sensor, doesn't automatically mean that the sensor actually has 15.3EV DR. It could still be that some of this data is nothing but noise. At the same time it would be inconceivable that Sony will provide this mode unless they believe that the sensor can provide close to this level of DR. Otherwise as I said earlier, slog2 will do more harm than good. These modes are a tradeoff where you sacrifice precision of higher intensity data to make room for more low-intensity data. They could be off by a small amount - maybe half a stop or something like that. But there is no way they will waste a precious 1-2 bits of the 8-bit output just to fill it with 2.3EV of noise and leave only 6-7 bits to encode the real 13EV DR.

In fact if we go by the idea that this is just a normal ISO3200 mode followed by software manipulation of tone curve, then that would mean that the final output actual has only about 10EV of real data and about 5EV of junk. Lets just say that the chances of that being the case are very very very slim
Apr 11, 2015 at 12:13 PM
snapsy
Offline
Details
p.3 #12High DR for 5d4
curious80 wrote:
snapsy wrote
curious80 wrote
snapsy wrote
curious80 wrote
snapsy wroteBased on the midtone noise performance of the A7s's slog2 video vs A7s's non-slog2 ISO 100 video it's definitely operating at the 1/5 FWC. The footage has the noise-equivalent of a 12MP ISO 3200 still downsampled to the equivalent 1080 or 4K video resolution.
You really have to understand that it is a non-linear gain curve unlike the normal ISO ratings where each intensity gets amplified by the same gain. In normal linear ISO3200 mode the all the values get multiplied basically by 32. So while you bring up the low intensities to high level, all the high intensities above the 1/32 of FWC get saturated. Which is why you only use 1/32 of the FWC. In the SLOG2 mode however the gain curve is such that the lower intensities get amplified a lot more than the higher intensities (which is what a log function does). So the gain curve is such that at the canonical 18% gray it will indeed correspond to a gain level of 32 consistent with ISO 3200. However the gain will be lower and lower as you go up the intensity level until at the top of the gain curve the gain would be 1 such that you will hit the saturation point at the same level as if you were shooting ISO 100. And in the process it will retain all the dynamic range.

Does that make sense?
That would imply that the A7s's ADC is capable of per-pixel variable gain at the 30hz full-sensor sampling rate the A7s supports for video. I'm not saying that's impossble but I've never heard of a sensor that can do that.
I don't know the details of how it gets implemented in hardware. But this is what the SLOG2 mode does. Look there is no way to create extra dynamic range in the software by any tonal manipulation, even more so for A7s where the 4K stream is a direct 1-to-1 pixel readout without any down-sampling.
I agree, which gets us back to the original discussion on whether the A7s has 15.3EV of DR that's only reachable via slog2 or that slog2 even yields more capture DR. If slog2 turns out to be a tone curve applied to the raw data that doesn't actually change the operation of the ADC in terms of variable gain then slog2 is simply a mechanism to compress a steeper tone curve into the existing 2^14 range of encoded data (or actually 2^8 since the encoded video output of the A7s is 8 bits rather than 14, even for its uncompressed HDMI output), resulting in a greater output/rendered DR but no effect on actual sensor-level captured DR.
It is pointless to apply slog2 after you have already converted to 14-bits or 8-bits. If you did, it will just have a hole in the data with nothing at the values corresponding to the extra stops of DR. It will only harm you because you will use even fewer bits for the data that you actually have and will assign some of the bits to the data which does not exist.

Of course just because slog2 captures 15.3EV of data off the sensor, doesn't automatically mean that the sensor actually has 15.3EV DR. It could still be that some of
...Show more →

It's not pointless because without an slog2 curve applied you'll get only 6-7 of output DR, the same as you'd get with a standard tone curve on a JPEG. And just like a JPEG the output of the A7s's video mode is gamma-converted processed data, not raw, both in non-slog2 and slog2. By compressing the tone curve and then allowing that compression to be unwound in post in a non-linear way they're simply fitting more DR into the data, at the expense of tonality of course.
Apr 11, 2015 at 12:32 PM
kezeka
Offline
Details
p.3 #13High DR for 5d4
The folks over at canonrumors are going on about a canon patent that can simultaneously read all pixels at high and low ISO, meaning a better version of magic lanterns dual ISO that requires no post processing interpolation between the high and low ISO stripes.

http://image-sensors-world.blogspot.com.es/2013/08/canon-files-for-dual-range-column.html

Regardless, whatever Canon is doing with the C300 II sensor, it is good to see them at least trying something new rather than continuing to slightly modify their old sensor stack.
Apr 11, 2015 at 12:42 PM
curious80
Offline
Details
p.3 #14High DR for 5d4
snapsy wrote:
curious80 wrote
snapsy wrote
curious80 wrote
snapsy wrote
curious80 wroteYou really have to understand that it is a non-linear gain curve unlike the normal ISO ratings where each intensity gets amplified by the same gain. In normal linear ISO3200 mode the all the values get multiplied basically by 32. So while you bring up the low intensities to high level, all the high intensities above the 1/32 of FWC get saturated. Which is why you only use 1/32 of the FWC. In the SLOG2 mode however the gain curve is such that the lower intensities get amplified a lot more than the higher intensities (which is what a log function does). So the gain curve is such that at the canonical 18% gray it will indeed correspond to a gain level of 32 consistent with ISO 3200. However the gain will be lower and lower as you go up the intensity level until at the top of the gain curve the gain would be 1 such that you will hit the saturation point at the same level as if you were shooting ISO 100. And in the process it will retain all the dynamic range.

Does that make sense?
That would imply that the A7s's ADC is capable of per-pixel variable gain at the 30hz full-sensor sampling rate the A7s supports for video. I'm not saying that's impossble but I've never heard of a sensor that can do that.
I don't know the details of how it gets implemented in hardware. But this is what the SLOG2 mode does. Look there is no way to create extra dynamic range in the software by any tonal manipulation, even more so for A7s where the 4K stream is a direct 1-to-1 pixel readout without any down-sampling.
I agree, which gets us back to the original discussion on whether the A7s has 15.3EV of DR that's only reachable via slog2 or that slog2 even yields more capture DR. If slog2 turns out to be a tone curve applied to the raw data that doesn't actually change the operation of the ADC in terms of variable gain then slog2 is simply a mechanism to compress a steeper tone curve into the existing 2^14 range of encoded data (or actually 2^8 since the encoded video output of the A7s is 8 bits rather than 14, even for its uncompressed HDMI output), resulting in a greater output/rendered DR but no effect on actual sensor-level captured DR.
It is pointless to apply slog2 after you have already converted to 14-bits or 8-bits. If you did, it will just have a hole in the data with nothing at the values corresponding to the extra stops of DR. It will only harm you because you will use even fewer bits for the data that you actually have and will assign some of the bits to the data which does not exist.

Of course just because slog2 captures 15.3EV of data off the sensor, doesn't automatically mean that the sensor actually has 15.3EV DR. It could still be that some of this data is nothing but noise. At the same time it would be inconceivable that Sony will provide this mode unless they believe that the sensor can provide close to this level of DR. Otherwise as I said earlier, slog2 will do more harm than good. These modes are a tradeoff where you sacrifice precision of higher intensity data to make room for more low-intensity data. They could be off by a small amount - maybe half a stop or something like that. But there is no way they will waste a precious 1-2 bits of the 8-bit output just to fill it with 2.3EV of noise and leave only 6-7 bits to encode the real 13EV DR.

In fact if we go by the idea that this is just a normal ISO3200 mode followed by software manipulation of tone curve, then that would mean that the final output actual has only about 10EV of real data and about 5EV of junk. Lets just say that the chances of that being the case are very very very slim
It's not pointless because without an slog2 curve applied you'll get only 6-7 of output DR, the same as you'd get with a standard tone curve on a JPEG. And just like a JPEG the output of the A7s's video mode is gamma-converted processed data, not raw, both in non-slog2 and slog2. By compressing the tone curve and then allowing that compression to be unwound in post in a non-linear way they're simply fitting more DR into the data, at the expense of tonality of course.


Of course you will always apply a gamma curve, but it will not be slog2. slog2 is made to fit 15.3EV of data into those 8-bits. If you start with data which has 13EV of DR then you will apply a different curve which can fit up-to 13EV of data into 8-bits. If you start with data that has 10-EV of DR (as ISO3200 has on A7s as per DxoMark) then again you would apply a different curve which can fit those 10-EV of data into 8-bits. If you apply slog2 on data which only has 10EV of data, then you will end up with a final 8-bit output in which only the top intensity values corresponding to the top 10EV of slog2 curve will have useful data. The whole range of lower intensities corresponding to the lower 5EV will be unused or garbage. In my books that is certainly pointless, unless you could think of a reason to do that
Apr 11, 2015 at 12:46 PM
curious80
Offline
Details
p.3 #15High DR for 5d4
Note that the usable dynamic range of the A7s in SLOG2 was tested as 14.1 stops here:

http://nofilmschool.com/2014/07/sony-a7s-dynamic-range-arri-alexa-amira

And this is not the absolute SNR=1 DR which would be higher. Even if this was the absolute DR, it would not be possible to get this through any software manipulation. Your sensor has to be capturing at the very least 14.1 stops of DR to get this result.
Apr 11, 2015 at 01:10 PM
snapsy
Offline
Details
p.3 #16High DR for 5d4
curious80 wrote:
snapsy wrote
curious80 wrote
snapsy wrote
curious80 wrote
snapsy wroteThat would imply that the A7s's ADC is capable of per-pixel variable gain at the 30hz full-sensor sampling rate the A7s supports for video. I'm not saying that's impossble but I've never heard of a sensor that can do that.
I don't know the details of how it gets implemented in hardware. But this is what the SLOG2 mode does. Look there is no way to create extra dynamic range in the software by any tonal manipulation, even more so for A7s where the 4K stream is a direct 1-to-1 pixel readout without any down-sampling.
I agree, which gets us back to the original discussion on whether the A7s has 15.3EV of DR that's only reachable via slog2 or that slog2 even yields more capture DR. If slog2 turns out to be a tone curve applied to the raw data that doesn't actually change the operation of the ADC in terms of variable gain then slog2 is simply a mechanism to compress a steeper tone curve into the existing 2^14 range of encoded data (or actually 2^8 since the encoded video output of the A7s is 8 bits rather than 14, even for its uncompressed HDMI output), resulting in a greater output/rendered DR but no effect on actual sensor-level captured DR.
It is pointless to apply slog2 after you have already converted to 14-bits or 8-bits. If you did, it will just have a hole in the data with nothing at the values corresponding to the extra stops of DR. It will only harm you because you will use even fewer bits for the data that you actually have and will assign some of the bits to the data which does not exist.

Of course just because slog2 captures 15.3EV of data off the sensor, doesn't automatically mean that the sensor actually has 15.3EV DR. It could still be that some of this data is nothing but noise. At the same time it would be inconceivable that Sony will provide this mode unless they believe that the sensor can provide close to this level of DR. Otherwise as I said earlier, slog2 will do more harm than good. These modes are a tradeoff where you sacrifice precision of higher intensity data to make room for more low-intensity data. They could be off by a small amount - maybe half a stop or something like that. But there is no way they will waste a precious 1-2 bits of the 8-bit output just to fill it with 2.3EV of noise and leave only 6-7 bits to encode the real 13EV DR.

In fact if we go by the idea that this is just a normal ISO3200 mode followed by software manipulation of tone curve, then that would mean that the final output actual has only about 10EV of real data and about 5EV of junk. Lets just say that the chances of that being the case are very very very slim
It's not pointless because without an slog2 curve applied you'll get only 6-7 of output DR, the same as you'd get with a standard tone curve on a JPEG. And just like a JPEG the output of the A7s's video mode is gamma-converted processed data, not raw, both in non-slog2 and slog2. By compressing the tone curve and then allowing that compression to be unwound in post in a non-linear way they're simply fitting more DR into the data, at the expense of tonality of course.
Of course you will always apply a gamma curve, but it will not be slog2. slog2 is made to fit 15.3EV of data into those 8-bits. If you start with data which has 13EV of DR then you will apply a different curve which can fit up-to 13EV of data into 8-bits. If you start with data that has 10-EV of DR (as ISO3200 has on A7s as per DxoMark) then again you would apply a different curve which can fit those 10-EV of data into 8-bits. If you apply slog2 on data which only has 10EV of data, then
...Show more →

Whether the A7s has 15.3EV or 13EV the point remains that slog2 allows a wider-than-bit-depth DR to fit into the data, a point I think we're in agreement on. Where I think we still disagree is that slog2 was necessary to exceed the A7s's 14-bit ADC depth because I haven't seen any evidence that the A7s can actually do 15.3EV. For me the discussion turned when I realized that with video were not talking about raw 14-bit data but instead processed 8-bit output, which fully explains why a non-linear encoding scheme like slog2 is utilized; it's to fit > 2^8 DR into an 8-bit stream rather than necessarily to fit a > 2^14 DR. It also supports the notion that this DR and tonal compression is occurring long after the ADC and instead downstream in Sony's imaging ASIC - meaning the ADC is still pumping out 14-bit raw linear data irrespective of whether slog2 is being used or not.
Apr 11, 2015 at 01:22 PM
curious80
Offline
Details
p.3 #17High DR for 5d4
snapsy wrote:
curious80 wrote
snapsy wrote
curious80 wrote
snapsy wrote
curious80 wroteI don't know the details of how it gets implemented in hardware. But this is what the SLOG2 mode does. Look there is no way to create extra dynamic range in the software by any tonal manipulation, even more so for A7s where the 4K stream is a direct 1-to-1 pixel readout without any down-sampling.
I agree, which gets us back to the original discussion on whether the A7s has 15.3EV of DR that's only reachable via slog2 or that slog2 even yields more capture DR. If slog2 turns out to be a tone curve applied to the raw data that doesn't actually change the operation of the ADC in terms of variable gain then slog2 is simply a mechanism to compress a steeper tone curve into the existing 2^14 range of encoded data (or actually 2^8 since the encoded video output of the A7s is 8 bits rather than 14, even for its uncompressed HDMI output), resulting in a greater output/rendered DR but no effect on actual sensor-level captured DR.
It is pointless to apply slog2 after you have already converted to 14-bits or 8-bits. If you did, it will just have a hole in the data with nothing at the values corresponding to the extra stops of DR. It will only harm you because you will use even fewer bits for the data that you actually have and will assign some of the bits to the data which does not exist.

Of course just because slog2 captures 15.3EV of data off the sensor, doesn't automatically mean that the sensor actually has 15.3EV DR. It could still be that some of this data is nothing but noise. At the same time it would be inconceivable that Sony will provide this mode unless they believe that the sensor can provide close to this level of DR. Otherwise as I said earlier, slog2 will do more harm than good. These modes are a tradeoff where you sacrifice precision of higher intensity data to make room for more low-intensity data. They could be off by a small amount - maybe half a stop or something like that. But there is no way they will waste a precious 1-2 bits of the 8-bit output just to fill it with 2.3EV of noise and leave only 6-7 bits to encode the real 13EV DR.

In fact if we go by the idea that this is just a normal ISO3200 mode followed by software manipulation of tone curve, then that would mean that the final output actual has only about 10EV of real data and about 5EV of junk. Lets just say that the chances of that being the case are very very very slim
It's not pointless because without an slog2 curve applied you'll get only 6-7 of output DR, the same as you'd get with a standard tone curve on a JPEG. And just like a JPEG the output of the A7s's video mode is gamma-converted processed data, not raw, both in non-slog2 and slog2. By compressing the tone curve and then allowing that compression to be unwound in post in a non-linear way they're simply fitting more DR into the data, at the expense of tonality of course.
Of course you will always apply a gamma curve, but it will not be slog2. slog2 is made to fit 15.3EV of data into those 8-bits. If you start with data which has 13EV of DR then you will apply a different curve which can fit up-to 13EV of data into 8-bits. If you start with data that has 10-EV of DR (as ISO3200 has on A7s as per DxoMark) then again you would apply a different curve which can fit those 10-EV of data into 8-bits. If you apply slog2 on data which only has 10EV of data, then you will end up with a final 8-bit output in which only the top intensity values corresponding to the top 10EV of slog2 curve will have useful data. The whole range of lower intensities corresponding to the lower 5EV will be unused or garbage. In my books that is certainly pointless, unless you could think of a reason to do that
Whether the A7s has 15.3EV or 13EV the point remains that slog2 allows a wider-than-bit-depth DR to fit into the data, a point I think we're in agreement on. Where I think we still disagree is that slog2 was necessary to exceed the A7s's 14-bit ADC depth because I haven't seen any evidence that the A7s can actually do 15.3EV. For me the discussion turned when I realized that with video were not talking about raw 14-bit data but instead processed 8-bit output, which fully explains why a non-linear encoding scheme like slog2 is utilized; it's to fit > 2^8
...Show more →

Snapsy, I think you are getting close to understanding the whole thing but are still missing out on a couple of points.

First, yes log-gamma curves have been in use on video cameras for a while. They are nothing new to A7s. As you said the aim of those curves is to fit a wider than 8EV DR into 8-bit video output. So then the next question becomes that how much DR you need to pack into those 8-bits, and that determines the profile of the gamma-curve that you apply. There are other curves such as s-log which also compress higher DR into 8-bits. However slog-2 has been used in A7s as it can compress an even larger DR range into 8-bits than is possible with slog.

The discussion started from your suggestion that sony uses slog2 to somehow come up with 15.3EV DR in software even though the sensor doesn't have 15.3EV DR. What I have explained to you is that it is not possible. If you start from 14-bit linear RAW data and then apply a gamma curve designed to pack 15.3EV of data into 8-bits, then you will have no data to place in the intensity ranges corresponding to the lowest 1.3EV. You will simply use up the range of intensities corresponding to the upper 14EV of the slog2 curve and some of the lowest intensities will be unused. And if you use ISO3200 where the actual DR captured is 10EV as per DxoMark, then the final 8-bit slog2 output will have about 5EV of noise+unused data. No one in their right mind will use a 15EV curve to encode 10EV of useful data, because that will mean you will sacrifice a lot in the tonality just to store 5EV worth of noise.

For that matter just leave all these details aside. Can you explain to me that if it is just a standard ISO3200 capture followed by an slog2 curve applied in software, then how are the Cinema5D folks able to measure 14.1 stop of usable DR in slog2. You should not be able to measure more than 10.2 EV of DR in that case.
Apr 11, 2015 at 01:59 PM
snapsy
Offline
Details
p.3 #18High DR for 5d4
curious80 wrote:
snapsy wrote
curious80 wrote
snapsy wrote
curious80 wrote
snapsy wroteI agree, which gets us back to the original discussion on whether the A7s has 15.3EV of DR that's only reachable via slog2 or that slog2 even yields more capture DR. If slog2 turns out to be a tone curve applied to the raw data that doesn't actually change the operation of the ADC in terms of variable gain then slog2 is simply a mechanism to compress a steeper tone curve into the existing 2^14 range of encoded data (or actually 2^8 since the encoded video output of the A7s is 8 bits rather than 14, even for its uncompressed HDMI output), resulting in a greater output/rendered DR but no effect on actual sensor-level captured DR.
It is pointless to apply slog2 after you have already converted to 14-bits or 8-bits. If you did, it will just have a hole in the data with nothing at the values corresponding to the extra stops of DR. It will only harm you because you will use even fewer bits for the data that you actually have and will assign some of the bits to the data which does not exist.

Of course just because slog2 captures 15.3EV of data off the sensor, doesn't automatically mean that the sensor actually has 15.3EV DR. It could still be that some of this data is nothing but noise. At the same time it would be inconceivable that Sony will provide this mode unless they believe that the sensor can provide close to this level of DR. Otherwise as I said earlier, slog2 will do more harm than good. These modes are a tradeoff where you sacrifice precision of higher intensity data to make room for more low-intensity data. They could be off by a small amount - maybe half a stop or something like that. But there is no way they will waste a precious 1-2 bits of the 8-bit output just to fill it with 2.3EV of noise and leave only 6-7 bits to encode the real 13EV DR.

In fact if we go by the idea that this is just a normal ISO3200 mode followed by software manipulation of tone curve, then that would mean that the final output actual has only about 10EV of real data and about 5EV of junk. Lets just say that the chances of that being the case are very very very slim
It's not pointless because without an slog2 curve applied you'll get only 6-7 of output DR, the same as you'd get with a standard tone curve on a JPEG. And just like a JPEG the output of the A7s's video mode is gamma-converted processed data, not raw, both in non-slog2 and slog2. By compressing the tone curve and then allowing that compression to be unwound in post in a non-linear way they're simply fitting more DR into the data, at the expense of tonality of course.
Of course you will always apply a gamma curve, but it will not be slog2. slog2 is made to fit 15.3EV of data into those 8-bits. If you start with data which has 13EV of DR then you will apply a different curve which can fit up-to 13EV of data into 8-bits. If you start with data that has 10-EV of DR (as ISO3200 has on A7s as per DxoMark) then again you would apply a different curve which can fit those 10-EV of data into 8-bits. If you apply slog2 on data which only has 10EV of data, then you will end up with a final 8-bit output in which only the top intensity values corresponding to the top 10EV of slog2 curve will have useful data. The whole range of lower intensities corresponding to the lower 5EV will be unused or garbage. In my books that is certainly pointless, unless you could think of a reason to do that
Whether the A7s has 15.3EV or 13EV the point remains that slog2 allows a wider-than-bit-depth DR to fit into the data, a point I think we're in agreement on. Where I think we still disagree is that slog2 was necessary to exceed the A7s's 14-bit ADC depth because I haven't seen any evidence that the A7s can actually do 15.3EV. For me the discussion turned when I realized that with video were not talking about raw 14-bit data but instead processed 8-bit output, which fully explains why a non-linear encoding scheme like slog2 is utilized; it's to fit > 2^8 DR into an 8-bit stream rather than necessarily to fit a > 2^14 DR. It also supports the notion that this DR and tonal compression is occurring long after the ADC and instead downstream in Sony's imaging ASIC - meaning the ADC is still pumping out 14-bit raw linear data irrespective of whether slog2 is being used or not.
For that matter just leave all these details aside. Can you explain to me that if it is just a standard ISO3200 capture followed by an slog2 curve applied in software, then how are the Cinema5D folks able to measure 14.1 stop of usable DR in slog2. You should not be able to measure more than 10.2 EV of DR in that case.

I've been thinking about the slog2's ISO 3200. I still don't think typical ADCs can do what you describe, a per-pixel analog gain as a means to implement a non-linear tone curve; the curve parts seems simple enough but not the variable analog gain part. Perhaps what Sony is doing instead is a variable multiplicative gain of the data in the digital domain, either at the ADC or more likely inside their imaging ASIC; in other words, the ADC outputs ISO 100 data and the ASIC scales the per-pixel ADUs using the variable exposure technique you ascribe to the ADC. This seems more plausible on a technical basis since the ASIC is already designed to perform transformations on a large volume of data in real-time. The downside of this would be worse noise performance vs a nominal ISO 3200 from the ADC since the A7s is not an ISOless sensor in terms of read noise. I suppose this could be verified with a carefully controlled comparison of the A7s slog2 video vs a pair of downsampled stills images, one shot at ISO 100 and pushed in PP vs the same exposure shot at a nominal ISO 3200 and comparing the noise of all three.

An interesting side-note is that the A7II added support for slog2 vs the A7, and the A7II's slog2 ISO is 1600 vs the A7s's ISO 3200. This might hold a useful clue about which of our theories is more likely to be correct.
Apr 11, 2015 at 04:08 PM
mttran
Offline
Details
p.3 #19High DR for 5d4
Mescalamba wrote:
Well, its only way to do it without switching to better fab.. also nice thing is, it will work perfectly even after switching to new fab (actually, if Canon did this and they moved to lets say 90nm fab or even 65nm, then it would floored every competition).


+1
Apr 11, 2015 at 07:46 PM
alundeb
Offline
Details
p.3 #20High DR for 5d4
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.

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.
Apr 13, 2015 at 03:01 AM
FM ForumsCanon ForumJoin Upload & Sell
  
 

You are not logged inLog inRegister