p.2 #3 · visualized read noise for TONS of different cameras
In the lens cap test, if the sensor simply set to black anything below a threshold, it would appear to have very little shadow noise - but that would prove nothing. How do I know that is **not** what is happening here with some of the sensors?
What you really want to do is to quantify the noise for black and different values of grey. Which is what DXOMark does and techradar.com claimed to do with the DXO Labs set-up
Bill Claff does his tests with black and white; but the white appears to be derived from the computer screen, which as a source of illumination, does not inspire my confidence.
Or one can look at actual photographs, which frankly are far more convincing in showing what a camera is capable of.
p.2 #4 · visualized read noise for TONS of different cameras
Arun Gupta wrote:
The 5D3 does have ISO 50, why is it missing from the chart?
it's basically fake on all of the cameras, I don't know why he put it there for the 5D2 to be honest, or why it looks different. It's the fakest of fake ISOs. The camera sets ISO100 when it shoots at ISO50 and just pretend it doesn't basically. It measures to same read noise as ISO100 on 5D2.
p.2 #5 · visualized read noise for TONS of different cameras
Arun Gupta wrote:
In the lens cap test, if the sensor simply set to black anything below a threshold, it would appear to have very little shadow noise - but that would prove nothing. How do I know that is **not** what is happening here?
that would define your black point though, so if they clipped it all of a sudden three stops higher the thing is then the black level would be 3 stops higher and it'd be automatically noted as lost from dynamic range, those three stops
nikon does often clip the below black level minus signal (right at the natural black point of the sensor, otherwise they couldn't get such high numbers for total stops of DR), but all of the above black level noise is there, this can be bad for some astronomy applications since you can't stack away back the noise in the natural way, but I don't think it matters much otherwise
p.2 #6 · visualized read noise for TONS of different cameras
Arun Gupta wrote:
In the lens cap test, if the sensor simply set to black anything below a threshold, it would appear to have very little shadow noise - but that would prove nothing. How do I know that is **not** what is happening here with some of the sensors?
What you really want to do is to quantify the noise for black and different values of grey. Which is what DXOMark does and techradar.com claimed to do with the DXO Labs set-up
Bill Claff does his tests with black and white; but the white appears to be derived from the computer screen, which as a source of illumination, does not inspire my confidence.
Or one can look at actual photographs, which frankly are far more convincing in showing what a camera is capable of.
yes you want to look at say SNR at middle gray too, but that is a totally different test than this one and has nothing to do with deep shadow performance and dynamic range, it is important too and it tells you how noisy all of the parts that are not very dark will be, which is important too (and for scenes with low dynamic rnage, it's basically the onyl part that matters)
p.2 #7 · visualized read noise for TONS of different cameras
Skibum5: thanks for the link. Visual tests like these are VERY useful. Hope that Canon will wake up and do something about their low ISO performance.
The recent release of the 60Da proves Canon treats astronomy applications very seriously and they will NEVER clip their black levels like Nikon.
Tester in A2B appears to have a very bad copy of the 7D with horrible banding. The copy I used to own definitely did not appear like that.
I am surprised the 40D has very good low ISO performance in this test, something not reflected in other measurements (e.g. DXOMark).
Another thing: it appears the 5D3 DOES improve (~ 0.5 stop) on low ISO dynamic range over the 5D2, but not significant like the Nikon.
Edit: I do not know why the Nikon cameras have totally black background in low ISO whereas the Canon images appear dark blue. Perhaps black level clipping at low ISO helps to reduce noise more than we think?
p.2 #9 · visualized read noise for TONS of different cameras
Nor does this test really indicate anything about dynamic range. There could be 8 stops available above what is shown here before the sensor saturates, or there could be 12. Or there could be five. Nothing in the data presented here tells us what the brightest value is that the sensor can faithfully record.
p.2 #10 · visualized read noise for TONS of different cameras
Arun Gupta wrote:
Nor does this test really indicate anything about dynamic range. There could be 8 stops available above what is shown here before the sensor saturates, or there could be 12. Or there could be five. Nothing in the data presented here tells us what the brightest value is that the sensor can faithfully record.
Agreed. These results only demonstrate noise performance near black levels. What surprises me is the nikon black patches are pitch black while the canon ones are dark blue. Seems that Nikon clipping is at work here.
p.2 #11 · visualized read noise for TONS of different cameras
thw2 wrote:
Skibum5: thanks for the link. Visual tests like these are VERY useful. Hope that Canon will wake up and do something about their low ISO performance.
The recent release of the 60Da proves Canon treats astronomy applications very seriously and they will NEVER clip their black levels like Nikon.
Tester in A2B appears to have a very bad copy of the 7D with horrible banding. The copy I used to own definitely did not appear like that.
I am surprised the 40D has very good low ISO performance in this test, something not reflected in other measurements (e.g. DXOMark).
Well, as stated they didn't normalize things. So you are comparing 10MP vs say 21MP for the 5D2 as if they were the same scale, which is not fair. Also, as pointed out, this doesn't show max saturation, which comes in a bit low for 40D at ISO100 so that counters the lower read noise as 100% view. However, my tests and DxO do show the 40D and 5D2 having about the same 100% view read noise which could fit what is shown here if you account for the lower saturation point.
However, visually, I, as well as others, have noted that the ISO100 deep shadows do have a nicer look, 100% view on 40D than the 5D2.
Anyway due to this and that you can't read DR directly off this, but it does give a sense of what things look like right around the black level, all cams at 100% view and not normalized. And it's clear, even as is, the Canon DSLR lag a bit at low ISO and that the 5D3 does something better for near black level noise than the 5D2 does from ISO1600 and up.
It's also best to do this using true RAW software, that looks at it before any processing at all, but I woudn't be surprised if he is just using ACR with all sliders at zero (there is no guarantee that sliders at zero do the same thing for each camera though). That said, in the broader sense, things look reasonable, near black 40D looking nicer than near black 50D/5D2 (although again this isn't quite fair since the 40D has much lower resolution and somewhat lower saturation point for ISO100) and recent Nikons have great near black performance, 5D3 improves near black performance vs 5D2 at higher ISOs.
p.2 #12 · visualized read noise for TONS of different cameras
Arun Gupta wrote:
Nor does this test really indicate anything about dynamic range. There could be 8 stops available above what is shown here before the sensor saturates, or there could be 12. Or there could be five. Nothing in the data presented here tells us what the brightest value is that the sensor can faithfully record.
true to an extent, for DSLRs the saturation point are never nearly so different as to be multiple stops though in any recent cams, probably always within 1 stop and likely within 1/2 stop I think
it does let you see what it does near blackpoint though at 100% view and you get a general sense of some things
p.2 #14 · visualized read noise for TONS of different cameras
skibum5 wrote:
Wow, someone put together quite the comprehensive visual comparison, sooo many different bodies compared: http://a2bart.com/tech/allcamdknz.htm
One thing to keep in mind is that he didn't account for differences in gain delivered by stated ISO number (for instance the 5D2 and 50D, in particular, give less gain for a given ISO than many of the others, so in reality, would look worse than shown here).
Another thing to keep in mind is that he didn't account for MP differences, these are all 100% crop comparisons, so the lower MP cameras here should actually look a bit worse in comparison than they do compared to high MP ones like 5D2/5D3/D800/Medium Format, in particular.
Also unless he used a true RAW examination program the results might be muddled since who knows what ACR does for each camera behind the scenes, it may do different baked in NR for each and set base levels differently, etc.
So it's not really quite fair test, but quite interesting all the same.shots probably are taken ISO3200 and under, but for scenes with tons of near black areas at super high ISO the 5D3 definitely should be better than the D800....Show more →
Thanks for the link. Interesting, but disappointing too.
I'm just speculating, but I don't think MP matters. What's showing is relative luminosity (ideally all squares should be black). Regardless of the sensor, the possible errant and unwanted luminosity noise added to each image, at each respective ISO, is relevant and shown (color noise doesn't matter here). Nor does the size of the blobs matter much; it's the overall luminosity that's important (not counting Canon's banding). Hint - squint and the blobs don't matter.
As long as he uses the same mfg's RAW engine we're getting as close to an apples-to-apples comparison as we can. It may be flawed because some algorithms may be newer than others, but it still reflects what users will see.
The D4 kicks butt (sony came a long way from the A900). Canon 7d has bad shadow banding at low ISO (worse than the "identical" sensor 60d) but is almost identical to the new 5d3 in luminosity noise. Canon G1X is seriously flawed (pathetic?) since it's a "new" sensor, yet turns in worse luminance noise levels at ISO 100-200 compared to the older G11/G12, and Pentax did extremely well on some of their cameras.
p.2 #15 · visualized read noise for TONS of different cameras
jamesf99 wrote:
I'm just speculating, but I don't think MP matters. What's showing is relative luminosity (ideally all squares should be black). Regardless of the sensor, the possible errant and unwanted luminosity noise added to each image, at each respective ISO, is relevant and shown (color noise doesn't matter here). Nor does the size of the blobs matter much; it's the overall luminosity that's important (not counting Canon's banding). Hint - squint and the blobs don't matter.
It depends on the downsampling algorithm, and on when the downsampling is performed. Remember that in the RAW file, the black point is not numerical zero, but an offset, for example 1024. The read noise histogram is distributed symmetrically around that offset. The values to the left represent impossible negative luminance and have to be removed for a correct black point by either the raw converter, or later by adjusting levels. If the downsampling is performed before the offset subtraction, the histogram wlll become narrower around the offset. Then after black point subraction, the luminance of the "noise floor" will be reduced.
p.2 #16 · visualized read noise for TONS of different cameras
alundeb wrote:
It depends on the downsampling algorithm, and on when the downsampling is performed. Remember that in the RAW file, the black point is not numerical zero, but an offset, for example 1024. The read noise histogram is distributed symmetrically around that offset. The values to the left represent impossible negative luminance and have to be removed for a correct black point by either the raw converter, or later by adjusting levels. If the downsampling is performed before the offset subtraction, the histogram wlll become narrower around the offset. Then after black point subraction, the luminance of the "noise floor" will be reduced. ...Show more →
I'm not sure i follow you completely, or why the downsampling event is as important. . What I was speculating is that IF there's a similarity of conversion tool(s) operating uniformly (downsampling done consistently for all samples), then there is at least some basis for assuming this is what an end user would see.
In the real world the algorithms will almost certainly vary across brands, some establishing the black point at 1024 or elsewhere, but unless it's changed later in levels, I don't think the user will have any control. Is there something else I missed here?
p.2 #17 · visualized read noise for TONS of different cameras
jamesf99 wrote:
I'm not sure i follow you completely, or why the downsampling event is as important. . What I was speculating is that IF there's a similarity of conversion tool(s) operating uniformly (downsampling done consistently for all samples), then there is at least some basis for assuming this is what an end user would see.
In the real world the algorithms will almost certainly vary across brands, some establishing the black point at 1024 or elsewhere, but unless it's changed later in levels, I don't think the user will have any control. Is there something else I missed here?
In some old versions of ACR you could set the black point negative, but that seems not to possible anymore. It allowed "negative luminance" values to be represented in the image.
I am not sure how raw converters like CaptureOne with implicit downsampling treat this, but it is possible that you will get a better result converting directly to a smaller RGB file than if you downsample later.
p.2 #18 · visualized read noise for TONS of different cameras
alundeb wrote:
In some old versions of ACR you could set the black point negative, but that seems not to possible anymore. It allowed "negative luminance" values to be represented in the image.
I am not sure how raw converters like CaptureOne with implicit downsampling treat this, but it is possible that you will get a better result converting directly to a smaller RGB file than if you downsample later.
I've used most versions (CS-CS5) but didn't remember that (negative luminance) was possible. Was that in the v7.0 plugin (the only one I skipped)? I also can't say that I ever thought about it having this kind of effect, but it's good to know.
Sounds like this would be a worthy, and expensive, test for someone with lots of tools and time to run. Any takers out there?
p.2 #19 · visualized read noise for TONS of different cameras
I don't remember which version(s) of ACR it was, but I think skibum5 and thedigitalbean were aware of it at the time.
After I placed my order for a Nikon D800E, I have stopped worrying about read noise at low ISO, and won't volunteer for more testing. But the engineer in me still wants to at least try to share a tiny beam of light on the subject on occasion.
p.2 #20 · visualized read noise for TONS of different cameras
jamesf99 wrote:
As long as he uses the same mfg's RAW engine we're getting as close to an apples-to-apples comparison as we can. It may be flawed because some algorithms may be newer than others, but it still reflects what users will see.
Young couple, John and Jane, enters wedding photographer's studio. During the negotiations,
John says, "Before we sign the contract, we want to see lens-cap-on-the-camera shots, raised at least 4 stops. I read on fredmiranda.com, where pro- photographers hangout, that this is what we'll see, and I don't want Jane to be disappointed."
