bclaff_too wrote:
I hope you won't find my comment out of line because you might think it's simply semantics.
As the ISO setting is increased read noise does not become a "increasing" factor, if anything the effect is slightly reduced.
The key factor is that less signal is collected, Full Well Capacity (FWC) is not fully utilized.
I completely agree with you Bill - it's probably just semantics
Here's a PTC-type graph I did last year in my Dpreview S&T Thread you kindly helped me out in:
This is for ISO 100 so the read noise values will be different for the High ISO we're speaking to but the general trend of read noise vs shot noise as a function of FWC utilized (saturation level in this graph) is demonstrated.
snapsy wrote:
Note that as ISO increases the impact of read noise on total noise increases
I might be over my head technically as I'm far from an electrical engineer. But I know more about Poisson processes than I would care to elaborate. Nevertheless my understanding is that it would be the exact *opposite*. Lower importance of read noise at high ISO is also consistent with Canon sensors typically getting their butt whooped at ISO 100 and remaining competitive from ~ISO 400 onward.
At low ISO settings (<400) -> Shot noise is low (plenty of photons to go round) -> Higher impact of read noise
At high ISO settings (>1600) -> Shot noise is high (not many photons) -> Lower impact of read noise
In the middle (400-1600) things are more balanced.
By Snapsy's logic, Canon sensors would be competitive near base ISO (read noise does not matter and shot noise is constant for both) and compared to Sonikon would get relatively worse throughout the ISO range. Historically, the opposite has been true.
Please correct me if I am misunderstanding something. This has been a very informative discussion.
mb126 wrote:
I might be over my head technically as I'm far from an electrical engineer. But I know more about Poisson processes than I would care to elaborate. Nevertheless my understanding is that it would be the exact *opposite*. Lower importance of read noise at high ISO is also consistent with Canon sensors typically getting their butt whooped at ISO 100 and remaining competitive from ~ISO 400 onward.
At low ISO settings (<400) -> Shot noise is low (plenty of photons to go round) -> Higher impact of read noise
At high ISO settings (>1600) -> Shot noise is high (not many photons) -> Lower impact of read noise
In the middle (400-1600) things are more balanced.
By Snapsy's logic, Canon sensors would be competitive near base ISO (read noise does not matter and shot noise is constant for both) and compared to Sonikon would get relatively worse throughout the ISO range. Historically, the opposite has been true.
Please correct me if I am misunderstanding something. This has been a very informative discussion. ...Show more →
It's a matter of what tones of the image we're speaking to and whether the difference would be visible in the normal course of working of the image. At base ISO Canon is further behind than the competition in read noise than at High ISO but that difference is in the deep shadows, which usually isn't perceivable unless the shadows are raised in post. In contrast, at High ISO the shadows are shifted right into the lower midtones, so even though Canon's disadvantage is less at High ISO it's more visible in the normal course of working the image (ie, doesn't require pushing the shadows to see).
armd wrote:
Yes, they should represent little challenge as compared to a small passerine, though if you ask those who have been in the field with Canon for years, it is a fact of life. The AF is THE reason Art gave up Canon in favor of Nikon and Sony. There is nothing "wrong" with my gear (am a CPS member), have many years experience with Canon equipment, have calibrated all of the gear, etc. FWIW I was tracking using center AF point with surround on a Egret flying across our field of view with a sun angle to our back, at 9am against open water/trees. I've experienced this with GBH's, Sandhill Cranes, Eagles and other large birds for no real explicable reason other than focus simply doesn't lock on and track properly. It doesn't happen all of the time, but frequently enough to be frustrating. Five minutes later, I captured a beautiful series of a barn swallow under similar lighting - smaller subject, more erratic flight, etc. This kind of failure never happened with my 1d series but was more common with the 7d's, 5d's, etc. in spite of proper technique, settings etc. I've wondered for years whether it had something to do with separate processors for AF, voltage differences driving the lens, or some other factor?
My hope was the 90d would have an improved AF system which would surpass the 7dmkii/5dmkiv. I need more than 20 mp full frame otherwise, I'd be shooting with a 1dx2. With respect to the spot AF, I use it a fair amount for perched birds for improved accuracy.
Art moving from Canon to Nikon is one that is not fully understood by most including you and I. One thing for sure it was highly political and obviously caused a lot of bad feelings on Art's part. I don't think his opinion is completely baseless but it is obviously fuelled by business concerns and a lot of emotion. He states his reasons but some insiders have stated others. Do we know and should we care? No. We just need to get what works for us. I believe that a less nebulous and more level-headed explanation comes from Ari Hazeghi. He also moved from Canon to Nikon for AF but he doesn't really have any skin in the game in terms of following or sponsorship. He stated quite clearly that Canon's AF is quite strong and Canon may well be ahead of Nikon in terms of lenses and extenders but he found that Nikon had the advantage of locking onto birds during complex action BUT even as someone who pursues avian photography at that level those situations only represent about 5% of his frames. As someone who shoots Canon and Nikon (Nikon occasionally now) I believe this is more realistic than some of the evangelical, hyperbolic system switchers we hear making noise on social media. Certainly - 100% a 1DxII will drive your 500LII faster than your 5D4 will. It would acquisition and lock more confidently because the it has a better AF system and the battery has more voltage to drive it faster (using a grip on the 5D4 will not change the voltage output). The problem is made worse with the use of extenders. I don't know if that plays into your equation but owning several bodies I know it first hand. So I do not believe that you will find your solution in the 90D. I doubt it will have an AF system more robust than the 5D4 and we already know that it uses a LP-E6N. There isn't a lot of mystery as to what kind of performance you'll get trying to drive a 500LII for BIF with a 90D. I've used the D500 with the 500PF. It is an excellent combo and performs better than Canon's APS-C cameras with big whites but, marginally. Maybe you should get a hold of a 1DxII and see what happens. You could borrow your friend's gear too. You have stated that you have observed his side-by-side results but have you tried the gear yourself? Maybe a switch to Nikon is your solution? The only thing that doesn't sit right is your inability to capture egrets with your setup. I used the 5D4 with the 500LII and the 400DOII with 1.x and 2x extenders pretty much all winter and I didn't have a problem with things a lot smaller and faster than egrets albeit in steady, level flight with the 2x on. I'm CPS too and Canon has given my gear a clean bill of health in the past when they were clearly wrong. Maybe something to consider?
Here is Ari's blog entry about his switch. You can also read some of his reviews. He talks about why the 1DxII drives heavy lens elements faster. Best of luck with your issue.
32.5 mp, doesn't that make Canon the leader? Have any of the other companies come out with 32mp yet? Plus it bodes well for an 80+ mp ff sensor, havingthe same pixel density
I hope Canon can match the DR. Mostly to get people who want to run down Canon over this issue to shut up already, because I'm sick of hearing about it.
I doubt they come out with a camera that had worse high ISO than its predecessor. Which is probably hard, which is probably why 1) no one else has done it & 2) it took so long.
mb126 wrote:
I might be over my head technically as I'm far from an electrical engineer. But I know more about Poisson processes than I would care to elaborate. Nevertheless my understanding is that it would be the exact *opposite*. Lower importance of read noise at high ISO is also consistent with Canon sensors typically getting their butt whooped at ISO 100 and remaining competitive from ~ISO 400 onward.
At low ISO settings (<400) -> Shot noise is low (plenty of photons to go round) -> Higher impact of read noise
At high ISO settings (>1600) -> Shot noise is high (not many photons) -> Lower impact of read noise
In the middle (400-1600) things are more balanced.
By Snapsy's logic, Canon sensors would be competitive near base ISO (read noise does not matter and shot noise is constant for both) and compared to Sonikon would get relatively worse throughout the ISO range. Historically, the opposite has been true.
Please correct me if I am misunderstanding something. This has been a very informative discussion.
At low ISO the signal is large, there are a lot of photoelectrons, the shot noise is the sqrt of the number of pe-, so the shot noise is high, the Signal to noise ratio is the signal/sqrt((input referenced read noise)^2 + number of pe-), so the SNR is high. Read noise << shot noise. Impact of read noise is small.
As the ISO goes up, the signal goes down, the shot noise goes down, the read noise does not change much, the signal drops faster than the shot noise, the SNR goes down. The shot noise is gets closer to the read noise as the signal is reduced. Read noise ~ shot noise. Impact of read noise is larger.
snapsy wrote:
I completely agree with you Bill - it's probably just semantics
Here's a PTC-type graph I did last year in my Dpreview S&T Thread you kindly helped me out in:
This is for ISO 100 so the read noise values will be different for the High ISO we're speaking to but the general trend of read noise vs shot noise as a function of FWC utilized (saturation level in this graph) is demonstrated.
I do not believe PRNU should be lumped together with read noise and shot noise. Photo Response Non Uniformity (PRNU) (a.k.a. Fixed Pattern Noise) is not noise in the same sense as read noise and shot noise. PRNU is fixed, measurable, and correctable.
dclark wrote:
I do not believe PRNU should be lumped together with read noise and shot noise. Photo Response Non Uniformity (PRNU) (a.k.a. Fixed Pattern Noise) is not noise in the same sense as read noise and shot noise. PRNU is fixed, measurable, and correctable.
Dave
Photo-Response Non-Uniformity (PRNU) is only correctable if it is Fixed Pattern Noise (FPN).
PRNU that is temporal rather than fixed is not correctable.
In any case, PRNU is so low for most modern sensors that it isn't a practical concern.
At the signal levels where PRNU makes a measurable contribution SNR is so high that it doesn't matter.
So is this data analysis predictive of how the 90D is going to compare to the 80D for apparent noise and DR (shadow pushing) at the un-normalized pixel level? We live in the age of 5K (15 MP) displays, an 8 MP average isn't too informative comparing a 24 to 32 MP sensor. Thanks for any clarification that can be provided.
Jeff Nolten wrote:
So is this data analysis predictive of how the 90D is going to compare to the 80D for apparent noise and DR (shadow pushing) at the un-normalized pixel level? We live in the age of 5K (15 MP) displays, an 8 MP average isn't too informative comparing a 24 to 32 MP sensor. Thanks for any clarification that can be provided.
cgarcia's measurements indicate the 90D will have modestly better per-pixel DR, ie better SE at its native 32.5MP resolution vs the 80D's native 24.2MP resolution.
mb126 wrote:
I might be over my head technically as I'm far from an electrical engineer. But I know more about Poisson processes than I would care to elaborate. Nevertheless my understanding is that it would be the exact *opposite*. Lower importance of read noise at high ISO is also consistent with Canon sensors typically getting their butt whooped at ISO 100 and remaining competitive from ~ISO 400 onward.
At low ISO settings (<400) -> Shot noise is low (plenty of photons to go round) -> Higher impact of read noise
At high ISO settings (>1600) -> Shot noise is high (not many photons) -> Lower impact of read noise
In the middle (400-1600) things are more balanced.
By Snapsy's logic, Canon sensors would be competitive near base ISO (read noise does not matter and shot noise is constant for both) and compared to Sonikon would get relatively worse throughout the ISO range. Historically, the opposite has been true.
Please correct me if I am misunderstanding something. This has been a very informative discussion.
In old sensors lacking on-sensor ADC, the signal had to travel a long distance to the external ADC. During the travel, some small amount of electrons were lost. In the hightlights that was not a big problem (e.g. 30 among 25000) but in the shadows they were a significative amount of the total signal transported (e.g. the same average of 30, now among 60). Old Nikon and Sony sensors were no different in this regard.
Raising the ISO the signal is amplified before being delivered to the ADC, so the same 60 electrons are now converted to e.g. 480 electrons at ISO 800. Since the amount lost during the travel remains about the same (30) but compared with a signal of 480 the actual loss is greatly reduced (these 30 represent only about 4 from the original signal).
Raising the ISO decreases the actual read noise in electrons (despite measured in data numbers in the RAW files it apparently increases). So at high ISO, despite the read noise becomes more relevant, it goes down to competitive levels. That explains why Canon always had good high ISO metrics despite its "outdated" ADC tech. And the new in-sensor ADC has dramatically reduced the read noise, despite Canon is not yet as refined as Sony. We mostly notice the improvement in the shadows of low ISO images, and high ISO amplification continues doing his old job to a lesser extent.
Now lets try an exercise: imagine a new sensor with twice the pixels. That causes the transported signal per pixel to halve, but the total aggregated signal at the image level usually remains the same if the manufacturer has done right its job. Since noises fortunately aggregate following the root mean square rule, it is not required to halve the read noise per pixel to keep the same performance: it suffices reducing it by 30% to 21 electrons in our example. I'm not an expert and I have no clue how they do this, but lately they achieve it. In the past I think that not always was the case.
cgarcia wrote:
In old sensors lacking on-sensor ADC, the signal had to travel a long distance to the external ADC. During the travel, some small amount of electrons were lost. ...
This portion isn't right. Electrons aren't lost but noise is introduced.
I'm not sure if it's 1/f noise, perhaps some EE will chime in.
But the longer the trace the more noise.
Jim Kasson frequently computes PDR but less often recently as he always gets the same results
PDR uses a different lower bound but it's not simply shifted.
8MP normalized Engineering Dynamic Range (EDR) (DxOMark Landscape Dynamic Range) is based entirely on Read Noise in DN and resolution.
PDR is based on the Photon Transfer Curve (PTC) and a particular resolution depended Signal to Noise Ratio (SNR).
PTC is a function of Read Noise and gain (also PRNU but that's no relevant to PDR at such low Signal values) .
Gain relates directly to Full Well Capacity (FWC), so PDR is influenced by FWC.
Finally, PDR is measured from an actual test image so other sources of noise, such as Fixed Pattern Noise (FPN), could influence the result.
Tech-savvy people interested in this topic might want to examine the PTC for the 80D.
On this PTC the black dot indicates where PDR is located. ...Show more →
I understand the PDR measure as a try to measure the "acceptable DR" from a image quality perspective. But I think that for practical purposes it is a flawed metric. And not just because it is difficult to understand or to be calculated.
Since all sensors (of the same size) have about the same light gathering efficiency, same SNR at high signal levels, and identical SNR performance due to photon noise... adding the SNR with a photonic component to the DR equation, even the smallest bit of it, IMO simply makes more difficult to compare how sensors actually perform in the metric which actually differentiates them (read noise and other sources hidden on it).
And yes, someone could argue that it has no sense to look at the very deep shadows. But one moment... really?
The PDR graph for the 80D you linked places the lower acceptable level at about 3.5 EV above the noise floor. That is, all between 0-3.5 EV is considered noise and "not usable DR". But as soon as I push the exposure 2 stops in some deep shadows, half of that area, previously in the black, will become visible. So I'm very interested on how much data is still under the PDR threshold (which unfortunately differs across cameras). I don't care if there is noise there: it there is still more signal than noise. The PDR hiddens me the shadows footroom and turns it variable across cameras (as you said, it is not a mere shift).
At ISO 25600 typically there are only 6 EV of "EDR" but I don't imagine nobody using that ISO throwing away the lower 3.5 EV because they are too noisy.
I see that the PDR for the Sony A7R3 is reportedly of 11.6 EV, 3 EV bellow the EDR. So the PDR measure thinks that the 3 lower EV above the noise floor are not usable... well: If I compare a ISO 100 shot pushed 3 stops with another at ISO 800, both are not that much different. Yes, the ISO 100 shot is noisier, but the shadows just surfaced are still acceptable on that sensor, and most of them come from the area discarded by PDR.
bclaff_too wrote:
This portion isn't right. Electrons aren't lost but noise is introduced.
I'm not sure if it's 1/f noise, perhaps some EE will chime in.
But the longer the trace the more noise.
Regards,
You are right... some appear and some other randomly dissapear!
cgarcia wrote:
I understand the PDR measure as a try to measure the "acceptable DR" from a image quality perspective. But I think that for practical purposes it is a flawed metric. And not just because it is difficult to understand or to be calculated.
Since all sensors (of the same size) have about the same light gathering efficiency, same SNR at high signal levels, and identical SNR performance due to photon noise... adding the SNR with a photonic component to the DR equation, even the smallest bit of it, IMO simply makes more difficult to compare how sensors actually perform in the metric which actually differentiates them (read noise and other sources hidden on it).
And yes, someone could argue that it has no sense to look at the very deep shadows. But one moment... really?
The PDR graph for the 80D you linked places the lower acceptable level at about 3.5 EV above the noise floor. That is, all between 0-3.5 EV is considered noise and "not usable DR". But as soon as I push the exposure 2 stops in some deep shadows, half of that area, previously in the black, will become visible. So I'm very interested on how much data is still under the PDR threshold (which unfortunately differs across cameras). I don't care if there is noise there: it there is still more signal than noise. The PDR hiddens me the shadows footroom and turns it variable across cameras (as you said, it is not a mere shift).
At ISO 25600 typically there are only 6 EV of "EDR" but I don't imagine nobody using that ISO throwing away the lower 3.5 EV because they are too noisy.
I see that the PDR for the Sony A7R3 is reportedly of 11.6 EV, 3 EV bellow the EDR. So the PDR measure thinks that the 3 lower EV above the noise floor are not usable... well: If I compare a ISO 100 shot pushed 3 stops with another at ISO 800, both are not that much different. Yes, the ISO 100 shot is noisier, but the shadows just surfaced are still acceptable on that sensor, and most of them come from the area discarded by PDR....Show more →
I think perhaps we should discuss this "offline" rather than dragging this thread off topic.
But clearly I disagee. PDR is a much better metric than 8MP EDR or any other metric I'm aware of.
cgarcia wrote:
Interesting! (I dind't know somebody took the effort to do a video).
To properly calculate the dynamic range, the image must have at least some pixels overexposed. The youtube video uses a image from the 5D (I hope) which was 12 bit, so in that case it is still correctly exposed for the task (upper values close to 4095, but maybe not just at that level, because old sensors didn't get to the very top numbers at ISO 100).
There is a usage example here (which somebody copied from my original post I can't find):
In canon it is very easy to calculate the dynamic range, because the sensor has an area with masked pixels (not receiving light). We can measure the noise on this area on any photo and compare it with the upper (brightest) level in the rest of the image, taking into account that the "0" is usually shifted. The maths used by the application are simple and well explained on the Internet:
range of values 0-16383 (14 bit)
black point (the "0", or average of all dark pixels): 512
white point (brightest pixels): 16383
read noise (standard deviation of the dark pixels): 2.80
90D DR at 32MP: log2((16383 - 512) / 2.80) = 12.47 EV
normalization 32MP -> 8MP: log2(sqrt(32/8)) = 1.0 EV
90D DR at 8MP: 12.47 + 1.0 = 13.47
This is the "engineering" DR. Or just DR, because PDR is a less known method published by a highly respectable but single site on the Internet IIRC (PDR uses another definition, essentially shifting up the lower bound to some more "photographically aceptable" value taking into account the sensor SNR at low signal levels).
In sensors which have no masked pixels (as the Sony), we need to shoot a 1/8000 frame with the lens cap on to measure the read noise. Another sensors not even have the 0 shifted (don't add e.g. 512) so it is not possible to calculate the standard deviation because the negative signal is lost (the only way it would be measuring patches with diminishing exposure and build a model to predict the lower limit).
It is recommendable to use the green channel because Canon has joined to the manufacturers scaling the other colors (R and B) to apply some kind of white balance "normalization" to the RAW data.
The dcraw tool used to get the raw data don't works with CR3 files, so you need to convert them with the Adobe DNG Converter and use the DNG file, which fortunately includes also the masked pixels (I tested it with CR2 files and I'm sure that DNG doesn't alters the raw data). There is a RAW library out there and some day I'll try it, because dcraw seems dead.
EDIT. Clarkvision has only a minor flaw in my opinion: the author likes to compare the maximum signal level at ISO 100 (full well) with the read noise at higher ISO, which decreases (measured in electrons) as ISO increases. So it gets astounding DR measures for Canon sensors in his tables, as good or better than other manufacturers. IMHO we should compare both values coming from the same ISO level (as everybody else does)....Show more →
Somebody is me ^^ and yes, it is a raw file from a 5D.
NormQ wrote:
Art moving from Canon to Nikon is one that is not fully understood by most including you and I. One thing for sure it was highly political and obviously caused a lot of bad feelings on Art's part. I don't think his opinion is completely baseless but it is obviously fuelled by business concerns and a lot of emotion. He states his reasons but some insiders have stated others. Do we know and should we care? No. We just need to get what works for us. I believe that a less nebulous and more level-headed explanation comes from Ari Hazeghi. He also moved from Canon to Nikon for AF but he doesn't really have any skin in the game in terms of following or sponsorship. He stated quite clearly that Canon's AF is quite strong and Canon may well be ahead of Nikon in terms of lenses and extenders but he found that Nikon had the advantage of locking onto birds during complex action BUT even as someone who pursues avian photography at that level those situations only represent about 5% of his frames. As someone who shoots Canon and Nikon (Nikon occasionally now) I believe this is more realistic than some of the evangelical, hyperbolic system switchers we hear making noise on social media. Certainly - 100% a 1DxII will drive your 500LII faster than your 5D4 will. It would acquisition and lock more confidently because the it has a better AF system and the battery has more voltage to drive it faster (using a grip on the 5D4 will not change the voltage output). The problem is made worse with the use of extenders. I don't know if that plays into your equation but owning several bodies I know it first hand. So I do not believe that you will find your solution in the 90D. I doubt it will have an AF system more robust than the 5D4 and we already know that it uses a LP-E6N. There isn't a lot of mystery as to what kind of performance you'll get trying to drive a 500LII for BIF with a 90D. I've used the D500 with the 500PF. It is an excellent combo and performs better than Canon's APS-C cameras with big whites but, marginally. Maybe you should get a hold of a 1DxII and see what happens. You could borrow your friend's gear too. You have stated that you have observed his side-by-side results but have you tried the gear yourself? Maybe a switch to Nikon is your solution? The only thing that doesn't sit right is your inability to capture egrets with your setup. I used the 5D4 with the 500LII and the 400DOII with 1.x and 2x extenders pretty much all winter and I didn't have a problem with things a lot smaller and faster than egrets albeit in steady, level flight with the 2x on. I'm CPS too and Canon has given my gear a clean bill of health in the past when they were clearly wrong. Maybe something to consider?
Here is Ari's blog entry about his switch. You can also read some of his reviews. He talks about why the 1DxII drives heavy lens elements faster. Best of luck with your issue.
Art has described himself as sucks at BIF - he is basically a bird portrait man with an emphasis on composition and colour whereas Ari is a dedicated BIF whizz. I can't compete in any way with Arbitrage for his extreme BIF photography but I have never had any problems getting a high proportion of tack-sharp images of large birds in flight or even fast flying puffins etc using a 5DIV, 5DSR, 7DII etc with 100-400mm II, 400mm DO II etc even with extenders when I get them in the central 9 points.
bclaff_too wrote:
I think perhaps we should discuss this "offline" rather than dragging this thread off topic.
But clearly I disagee. PDR is a much better metric than 8MP EDR or any other metric I'm aware of.
I wouldn't mind if you have the discussion here. It is relevant as to "How much do we know about the 90D dynamic range yet". On the other hand, it would likely be an endless discussion.
My opinion is that PDR is a very sensor good metric, perhaps the best all-round metric. But as with all single number metrics, it does not tell the whole story. It will depend on your application which metric is most useful. For me, full well capacity (how is the maximum image quality) and fixed pattern noise (ten times more annoying than random noise) is almost all I care about. From a single PDR value at base ISO, it is impossible to tell either of those. (You can tell what FWC has to be as a minimum, but not how two cameras actually compare). To be fair, it is not possible with EDR either
To elaborate more on the full well capacity, cameras with good full well capacity and high read noise, such as the 5DS R, score relatively low on PDR but in reality have fantastic image quality under ideal conditions. If people read PDR numbers only, they may get the impression that many APS-C sensors are better, and even some 1" compact cameras would be close. That would be misleading in most situations. Even compared to other FF cameras with much higher dynamic range, I often find the 5DS R files to print better at 24x16" (That's mostly because of the resolution).
bclaff_too wrote:
I think perhaps we should discuss this "offline" rather than dragging this thread off topic.
But clearly I disagee. PDR is a much better metric than 8MP EDR or any other metric I'm aware of.
I have very much enjoyed your participation. Please continue.
I enjoy these too, but unfortunately a vast majority of people won't use the DR and high ISO capabilities despite the 90D seemingly being better. So many seem to give up with the 7D2 and 80D past 3200, many don't push shadows more than maybe 2 stops. Those folks either don't process their raw/jpg very well, have no patience to do, or just use the wrong tools, so these capabilities on the APS-C bodies are lost on them.
I see posts all the time like "I cannot push my 7D2 past 1600", etc. and I think they are really missing some capabilities/opportunities.
I do know that if Canon can get an APS-C to have something in striking distance of the 5D4, then a very high resolution FF mirrorless will be in high demand.
