S1H Raw Format?

[Zeng]

58 minutes ago, HockeyFan12 said:

How do you get more than 12 stops of dynamic range from a 12 bit ADC readout? Isn't this impossible by definition? Wouldn't there by a hard ceiling of 12 stops from a 12 bit ADC? See the depreview link above.

My issues are that I don't know why Cinema5D uses SNR=2 as the noise floor or why they place the first wedge of the Xyla chart below clipping. SNR=2 seems arbitrary, and their best explanation of why they choose this is it's what gets them a 14 stop measurement from the Alexa, but Arri intentionally underrates its sensors and if you talk with their reps they'll tell you the Amira, Mini, etc. are 15+ stops. And the official number from Arri has always been 14+ anyway. That's why I see the comparisons as useful, but to me the actual measurement I'd derive from the S1H's result are the 13.8 they find at SNR=1 plus a bit more from them placing the first wedge below clipping. So about 14 stops. Likewise all their numbers seem low to me.

Numbers like "usable dynamic range" and Cinema5D's decision to sort of arbitrarily decide a cut off point for noise have been discussed in the past and never that productively. Canon took issue with their results on the C300 Mk II and released their own tests that correlate more closely with 14-15 stops if read traditionally rather than by Cinema5D's arbitrary (however useful for comparison) metrics. If you're interested in useable dynamic range in the field I find Geoff Boyle's over/under tests on CML more helpful anyway as you can see an entire scene and where different channels clip and there are some cameras that clip certain colors sooner or less attractively. Of course the Alexa does best. 

If the S1H results are heavily influenced by noise reduction I don't know. I didn't know the Ursa Mini had no NR at that, that's impressive if true. I haven't had issues with noise reduction on either camera, but I have read about the S1H having ghosting issues for other people so clearly others have. I'm not really concerned with that, though, as I'm not planning to buy either camera. I'm just trying to understand how you can get 14 stops of dynamic range from a 12 bit ADC.

In my experience and tests Arri are the only ones keeping their DR rating more or less real. All others are counting nonexistent stops digging them out in the shadows. There will always be controversy as to what counts as a useful stop in the shadows vs the noise floor, especially given all the NR processing almost everyone is using. So one can find an additional stop in the shadows at the cost of chroma, etc. And it quickly becomes religious with the brand fans. Best advice - test your camera of choice, find its  DR limits as acceptable to you personally, and forget the numbers.

[thebrothersthre3]

1 hour ago, HockeyFan12 said:

How do you get more than 12 stops of dynamic range from a 12 bit ADC readout? Isn't this impossible by definition? Wouldn't there by a hard ceiling of 12 stops from a 12 bit ADC? See the depreview link above.

My issues are that I don't know why Cinema5D uses SNR=2 as the noise floor or why they place the first wedge of the Xyla chart below clipping. SNR=2 seems arbitrary, and their best explanation of why they choose this is it's what gets them a 14 stop measurement from the Alexa, but Arri intentionally underrates its sensors and if you talk with their reps they'll tell you the Amira, Mini, etc. are 15+ stops. And the official number from Arri has always been 14+ anyway. That's why I see the comparisons as useful, but to me the actual measurement I'd derive from the S1H's result are the 13.8 they find at SNR=1 plus a bit more from them placing the first wedge below clipping. So about 14 stops. Likewise all their numbers seem low to me.

Numbers like "usable dynamic range" and Cinema5D's decision to sort of arbitrarily decide a cut off point for noise have been discussed in the past and never that productively. Canon took issue with their results on the C300 Mk II and released their own tests that correlate more closely with 14-15 stops if read traditionally rather than by Cinema5D's arbitrary (however useful for comparison) metrics. If you're interested in useable dynamic range in the field I find Geoff Boyle's over/under tests on CML more helpful anyway as you can see an entire scene and where different channels clip and there are some cameras that clip certain colors sooner or less attractively. Of course the Alexa does best. 

If the S1H results are heavily influenced by noise reduction I don't know. I didn't know the Ursa Mini had no NR at that, that's impressive if true. I haven't had issues with noise reduction on either camera, but I have read about the S1H having ghosting issues for other people so clearly others have. I'm not really concerned with that, though, as I'm not planning to buy either camera. I'm just trying to understand how you can get 14 stops of dynamic range from a 12 bit ADC.

There is simply no NR when shooting in RAW CDNG. With highlight recovery and NR in davinci added they were able to get a 13 stop read. That said their Alexa measurement is based on prores not RAW, so the Alexa probably has even higher dynamic range in RAW. 

[HockeyFan12]

38 minutes ago, Zeng said:

In my experience and tests Arri are the only ones keeping their DR rating more or less real. All others are counting nonexistent stops digging them out in the shadows. There will always be controversy as to what counts as a useful stop in the shadows vs the noise floor, especially given all the NR processing almost everyone is using. So one can find an additional stop in the shadows at the cost of chroma, etc. And it quickly becomes religious with the brand fans. Best advice - test your camera of choice, find its  DR limits as acceptable to you personally, and forget the numbers.

On one hand I think the S1H probably actually measures as 14 stops (the Cinema5D chart even indicates as much–13.8 stops at SNR=1 plus a bit from the first wedge having a little room to go before clipping) and the C300 Mk II probably measures at 14-15 in the lab. So I can't fault any manufacturer for making claims that are basically accurate. Some of Red's claims seem a little out there, though.

On the other hand, yeah, the C300 Mk II has an ugly noise texture and color channels clip earlier than claimed and the sensor has streaks when hit by bright lights. So in practice it's not that great. While the Alexa actually is super noisy in the shadows but the noise texture looks organic so in practice it's even better than it measures.

I'm not even trying to argue about one camera over the next, I just don't get how the 12 bit ADC in the S1H is producing 14 stops of measurable dynamic range. And yes, that's what the Cinema5D Xyla chart shows if read traditionally, even though I agree the noise floor isn't useful in practice, the Xyla chart shows 14 stops.

I do find these measurements helpful:

https://cinematography.net/CineRant/2018/07/30/personal-comments-on-the-2018-cml-camera-evaluations/

The Alexa actually doesn't have that much more DR total with the noisy shadows but it has gobs more highlight detail.

I really don't want to turn this into a camera debate. I just don't think the S1H has a 12 bit ADC in video mode. Either that or I'm misunderstanding something and would like to understand better.

[thebrothersthre3]

1 minute ago, HockeyFan12 said:

 

The Alexa actually doesn't have that much more DR total with the noisy shadows but it has gobs more highlight detail.

I really don't want to turn this into a camera debate. I just don't think the S1H has a 12 bit ADC in video mode.

I am pretty sure the 12 bit ADC is just the sensor limitation based off specs on paper. I don't really think it can be disputed unless you think Sony is lying for some reason about their own sensor's capabilities. 

[HockeyFan12]

Just now, thebrothersthre3 said:

I am pretty sure the 12 bit ADC is just the sensor limitation based off specs on paper. I don't really think it can be disputed. 

Thanks, do you have a source for this? What paper are you referencing? 

I don't have a horse in this race, I've used the S1H and was impressed by it so the results are there and I don't really care about the specs that much. It wouldn't matter to me either way. I just don't understand how this could be the case because it directly contradicts the dpreview article I posted, which is how I understand these things work.

[thebrothersthre3]

53 minutes ago, HockeyFan12 said:

Thanks, do you have a source for this? What paper are you referencing? 

I don't have a horse in this race, I've used the S1H and was impressed by it so the results are there and I don't really care about the specs that much. It wouldn't matter to me either way. I just don't understand how this could be the case because it directly contradicts the dpreview article I posted, which is how I understand these things work.

 @androidlad seems to be the guy to go to on sensor information. I'm not sure where he sources this stuff. Maybe he'll chime in

[androidlad]

3 hours ago, HockeyFan12 said:

How do you get more than 12 stops of dynamic range from a 12 bit ADC readout? Isn't this impossible by definition? Wouldn't there by a hard ceiling of 12 stops from a 12 bit ADC? See the depreview link above.

My issues are that I don't know why Cinema5D uses SNR=2 as the noise floor or why they place the first wedge of the Xyla chart below clipping. SNR=2 seems arbitrary, and their best explanation of why they choose this is it's what gets them a 14 stop measurement from the Alexa, but Arri intentionally underrates its sensors and if you talk with their reps they'll tell you the Amira, Mini, etc. are 15+ stops. And the official number from Arri has always been 14+ anyway. That's why I see the comparisons as useful, but to me the actual measurement I'd derive from the S1H's result are the 13.8 they find at SNR=1 plus a bit more from them placing the first wedge below clipping. So about 14 stops. Likewise all their numbers seem low to me.

Numbers like "usable dynamic range" and Cinema5D's decision to sort of arbitrarily decide a cut off point for noise have been discussed in the past and never that productively. Canon took issue with their results on the C300 Mk II and released their own tests that correlate more closely with 14-15 stops if read traditionally rather than by Cinema5D's arbitrary (however useful for comparison) metrics. If you're interested in useable dynamic range in the field I find Geoff Boyle's over/under tests on CML more helpful anyway as you can see an entire scene and where different channels clip and there are some cameras that clip certain colors sooner or less attractively. Of course the Alexa does best. 

If the S1H results are heavily influenced by noise reduction I don't know. I didn't know the Ursa Mini had no NR at that, that's impressive if true. I haven't had issues with noise reduction on either camera, but I have read about the S1H having ghosting issues for other people so clearly others have. I'm not really concerned with that, though, as I'm not planning to buy either camera. I'm just trying to understand how you can get 14 stops of dynamic range from a 12 bit ADC.

The ADC bit-depth numbers you see (12, 14 or even 16) are quantisation accuracy, they themselves do not act as ceilings or floors.

It's the SNR measurement threshold that sets the floor, the last few stops are always buried in noise, and noise reduction is an extremely effective way to increase SNR, especially in video mode due to the temporal nature. That's how they got more than 12 stops of DR from 12bit ADC.

Current ALEXA sensors use 14bit ADC, but because the ADCs are off-chip, they themselves introduce a bit more noise than modern on-chip ones, so the camera's 16bit ISP does a bit of noise reduction (and other ARRI magic). For this reason, 14 stops is a good benchmark, so I wouldn't look at SNR=1 results, but they publish all the numbers and I respect each individual's own interpretation.

 

 

 

[TheBoogieKnight]

2 minutes ago, androidlad said:

The ADC bit-depth numbers you see (12, 14 or even 16) are quantisation accuracy, they themselves do not act as ceilings or floors.

It's the SNR measurement threshold that sets the floor, the last few stops are always buried in noise, and noise reduction is an extremely effective way to increase SNR, especially in video mode due to the temporal nature. That's how they got more than 12 stops of DR from 12bit ADC.

 

 

 

 

II don't think this is true. You can't get more dynamic range from a 12 bit readout than 12 bits although you can get more *perceived* dynamic range if you start off with more bits and add dither. This is what happens in the audio industry (I worked/lectured in Music Tech for many years). But it actually reduces the true dynamic range, it's just that you can hear (or see in a camera) below the noise floor due to the dithering removing  quantisation distortion.

[HockeyFan12]

47 minutes ago, TheBoogieKnight said:

II don't think this is true. You can't get more dynamic range from a 12 bit readout than 12 bits although you can get more *perceived* dynamic range if you start off with more bits and add dither. This is what happens in the audio industry (I worked/lectured in Music Tech for many years). But it actually reduces the true dynamic range, it's just that you can hear (or see in a camera) below the noise floor due to the dithering removing  quantisation distortion.

Yeah, this was my understanding, too, and what the dpreview article states, which is why I'm confused. I get that with Canon Log and Arri Log the data is converted from a 16 bit linear signal (or whatever it is) to a 12 bit logarithmic one (or 10 bit with 60p on the C200). So you can store more dynamic range in a raw file than its bit depth. This just usually doesn't happen with still cameras, usually it's linear data that hasn't been compressed as much.

But I thought that an ADC quantized (if that's the right term) a linear signal and so each bit represents another order of magnitude of signal strength represents another potential stop of dynamic range. But of course the sensor's noise floor and full well capacity are the other limiting factors so it's not like ADC equals dynamic range, either, just potential dynamic range.

I thought. I dunno, last time I tried to wrap my head around something like this I got it all wrong, so I'm probably just confused. I still don't get it, but it's probably just me.

Carry on...

[androidlad]

21 minutes ago, TheBoogieKnight said:

II don't think this is true. You can't get more dynamic range from a 12 bit readout than 12 bits although you can get more *perceived* dynamic range if you start off with more bits and add dither. This is what happens in the audio industry (I worked/lectured in Music Tech for many years). But it actually reduces the true dynamic range, it's just that you can hear (or see in a camera) below the noise floor due to the dithering removing  quantisation distortion.

Cameras don't just read the sensor then dump the data straight to the cards, do they?

Noise floor can be lowered by noise reduction.

[TheBoogieKnight]

5 minutes ago, androidlad said:

Cameras don't just read the sensor then dump the data straight to the cards, do they?

Noise floor can be lowered by noise reduction.

Yes but not below the bit-depth. And dynamic range can't be increase over the bit-depth. You can't get more a 12-bit signal from 12 bits unless (as I said earlier) you start with a higher bit-depth and add dither. This gives you more perceived dynamic range.

[TheBoogieKnight]

Would any of you lot buy one now the R5 is released then? I can get one for under 3.5k with the Ninja and grip. Seriously considering it but just scared of regretting it horrible if the R5 comes in at a similar price point and is as amazing as it's looking. I've got an S1 though so it would mean selling my L Mount kit etc. etc..

[sanveer]

.

[sanveer]

1 hour ago, TheBoogieKnight said:

This is what happens in the audio industry (I worked/lectured in Music Tech for many years).

And perhaps dentistry too ...

 

24 minutes ago, TheBoogieKnight said:

Would any of you lot buy one now the R5 is released then? I can get one for under 3.5k with the Ninja and grip. Seriously considering it but just scared of regretting it horrible if the R5 comes in at a similar price point and is as amazing as it's looking. I've got an S1 though so it would mean selling my L Mount kit etc. etc..

You sound awfully familiar. Are you from Egypt, by any chance?

🤣

[deezid]

58 minutes ago, TheBoogieKnight said:

Yes but not below the bit-depth. And dynamic range can't be increase over the bit-depth. You can't get more a 12-bit signal from 12 bits unless (as I said earlier) you start with a higher bit-depth and add dither. This gives you more perceived dynamic range.

You will still be able to see patches in darker areas, doesn't mean the data is usable without huge amounts of noise reduction which is why the S1H is able to achieve almost 13 stops of 'usable' dynamic range, too bad Imatest doesn't care about motion artifacts caused by temporal noise reduction. I wouldn't mind temporal noise reduction that doesn't affect motion performance - HLG e.g. is perfectly fine on this camera achieving 12.2 actually usable stops. 😉

BMD, RED and Canon on the other hand use spatial noise reduction to boost their Imatest performance which doesn't create any motion related artifacts but sacrifices texture, darker areas chroma information often seems to be wiped out, skin may look waxy even, the Pocket 4K being the worst case. Cameras like the Pocket 6K and Ursa G2 and newer Canon cameras and any RED camera ever seem to be pretty well balanced though.

[TheBoogieKnight]

26 minutes ago, deezid said:

You will still be able to see patches in darker areas, doesn't mean the data is usable without huge amounts of noise reduction which is why the S1H is able to achieve almost 13 stops of 'usable' dynamic range, too bad Imatest doesn't care about motion artifacts caused by temporal noise reduction. I wouldn't mind temporal noise reduction that doesn't affect motion performance - HLG e.g. is perfectly fine on this camera achieving 12.2 actually usable stops. 😉

BMD, RED and Canon on the other hand use spatial noise reduction to boost their Imatest performance which doesn't create any motion related artifacts but sacrifices texture, darker areas chroma information often seems to be wiped out, skin may look waxy even, the Pocket 4K being the worst case. Cameras like the Pocket 6K and Ursa G2 and newer Canon cameras and any RED camera ever seem to be pretty well balanced though.

Why is HLG so much different to VLOG in terms of noise etc? They've definitely improved things on the S1 with firmware updats although I don't think the H got the same improvements yet (could be wrong)?

[TheBoogieKnight]

29 minutes ago, deezid said:

BMD, RED and Canon on the other hand use spatial noise reduction to boost their Imatest performance which doesn't create any motion related artifacts but sacrifices texture, darker areas chroma information often seems to be wiped out, skin may look waxy even, the Pocket 4K being the worst case. Cameras like the Pocket 6K and Ursa G2 and newer Canon cameras and any RED camera ever seem to be pretty well balanced though.

Have you used the Pocket stuff much?

[deezid]

 

12 minutes ago, TheBoogieKnight said:

Have you used the Pocket stuff much?

Yes, but sold it soon after the Braw update and went with the S1 which had nicer color rendition and better dynamic range as well as much more texture but at the time horrible blue clipping (Pocket 4K still has red clipping) and motion artifacts using V-Log. Switched to the S1H soon after and tried avoiding ghosting/smearing by overexposing heavily which often leads to thin looking footage, at least the S1H never had blue clipping.

Fun fact: The pre production model of the S1H at IBC in September 2019 with ProRes raw support had the blue clipping fixed while the other model didn't. There were two different firmware development branches.

 

14 minutes ago, TheBoogieKnight said:

Why is HLG so much different to VLOG in terms of noise etc? They've definitely improved things on the S1 with firmware updats although I don't think the H got the same improvements yet (could be wrong)?

HLG and any other profile but V-Log and Like709 for whatever reason don't have any motion artifacts but still use noise reduction even at the lowest setting which is fine to me.

The S1 had the noise reduction and blue clipping update for V-Log on November 19th already. It's both cleaner and free of artifacts on the S1.

[Lux Shots]

On 5/5/2020 at 10:09 PM, HockeyFan12 said:

Not being snarky, genuinely curious: how does the S1H achieve 14 stops of dynamic range with a 12 bit ADC? 

You must be mistaking the bit depth with dynamic range. The only thing they have in common is ration numbers. You can have 14 stops of dynamic range with a 10-bit ADC. Now you won't have fine gradation between the darkest and the brightest, but resolution and dynamic range are completely different.

[Lux Shots]

On 5/6/2020 at 5:55 PM, deezid said:

 

Yes, but sold it soon after the Braw update and went with the S1 which had nicer color rendition and better dynamic range as well as much more texture but at the time horrible blue clipping (Pocket 4K still has red clipping) and motion artifacts using V-Log. Switched to the S1H soon after and tried avoiding ghosting/smearing by overexposing heavily which often leads to thin looking footage, at least the S1H never had blue clipping.

Fun fact: The pre production model of the S1H at IBC in September 2019 with ProRes raw support had the blue clipping fixed while the other model didn't. There were two different firmware development branches.

 

HLG and any other profile but V-Log and Like709 for whatever reason don't have any motion artifacts but still use noise reduction even at the lowest setting which is fine to me.

The S1 had the noise reduction and blue clipping update for V-Log on November 19th already. It's both cleaner and free of artifacts on the S1.

Version 2.0 FW update V-Log setting now has noise reduction of -1, which should address any NR issues you are complaining about. May 25th will be a happy day for S1H users.