From DXOMark: More pixels offsets noise3

Started May 30, 2009 | Discussions
Steen Bay Veteran Member • Posts: 7,418
Re: Bob..

bobn2 wrote:

Steen Bay wrote:

bobn2 wrote:

Steen Bay wrote:

Thanks. Reading Emil's reply again, I can see that he actually did
explain it nicely (and you too, of course), but still I find it a bit
counter-intuitive, that a gaussian blur affects the visible noise in
the downsampled image, but not the visible detail. Anyway, I'll take
your word for it, that it's the best method when downsampling, also
to prevent aliasing. (And it would of course be pretty easy to check,
downsampling the 50D resolution testchart shot, with and without the
gaussian blur)

Here you are 50D downsampled to 6.7MP, this time cubic, top no blur,
bottom blur.

Bob

Interesting, thanks. Isn't the top one, with no gaussian blur, the
sharpest, with most detail and noise? Does it have any
aliasing/artifacts? ISO 3200?

Could be I've overdone the blur a tad - if I got it right it
shouldn't affect sharpness. Could well be aliasing in the top one,
you'd need to hunt around the fabrics.
These are all ISO3200, same CR2's as PIX used.

I actually wondered a bit about that, because 40D only has ISO 3200 when "H", "ISO expansion" is activated. Is ISO 3200 a 'real' ISO on 40D, or just a 'pushed' ISO 1600? Doesn't that make a difference?

John1940
John1940 Senior Member • Posts: 2,820
Re: Bob..

Steen,

I think that one of the issues with regard to comparing the 50D and 40D is that the pixel spacing (horizontally or vertically) only drops by about 18% in going from the 40D to the 50D.

Here is a test that might prove something regarding the effectiveness of doing Gaussian blur before down-sampling with a series of successive doubling pixel-to-pixel spacing changes. It’s one that I have done before many times but I did not do the blurring first. However, I still have the evidence and may do some studies if I ever get the time.

I have scanned many 35-mm Kodachrome slides using a Canon 4000 dpi film scanner. When I first got the machine almost six years ago I decided to do quite a few tests with film and digital cameras I had at the time. My newest camera was a Canon 300D (Rebel) DSLR. I also had (and still have) a 35-mm Canon film body (which takes EF lenses) and had a Minolta 35-mm system with six lenses.

In one series of tests (using several slides but no DSLRs) I scanned each slide at 4000 dpi, 2000 dpi, 1000 dpi and 500 dpi successively without touching or moving the slide at all. In round numbers I got a 20 Mp, 5 Mp, 2.5 Mp, and 1.25 Mp image in Tiff-8 or Tiff-16 format (or RGB format). The scanner can do 8 bits or 12 bits per color at every pixel location. Each pixel therefore uses either 3 bytes or 6 bytes of memory. I then did various (but not scientifically rigorous) comparisons by down-sampling but did not do Gaussian blur with the higher resolutions first.

The potential value of this test method (which I’m sure has been used by many people in many organizations going back decades) is that all other variables are removed since the source of the image is the same and the scanner is the same. Also, the steps in resolution are large enough to show the primary advantage of the higher pixel count (detail). I never found any increase in noise in my unscientific experiments. But I never looked at whether Gaussian blur before down-sampling would make what I thought was a good result even better in favor of the higher-resolution scans.

I think that the scanner fires the same number of photons (as such things go) for each pixel location. If I’m right, that takes another variable off the table, or at least makes it knowable.

At this point I’m in the 50D-is-better camp for detail and for being no worse for noise if down-sampled to 10 Mp with Gaussian blur done first. But, unlike with the scanner, it’s not easy to do accurate side-by-side tests with the 40D and 50D. There are too many variables that cannot be controlled easily.

-- hide signature --

John1940

 John1940's gear list:John1940's gear list
Sigma DP3 Merrill
John Sheehy Forum Pro • Posts: 21,710
Re: Bob..

Steen Bay wrote:

Thanks. Reading Emil's reply again, I can see that he actually did
explain it nicely (and you too, of course), but still I find it a bit
counter-intuitive, that a gaussian blur affects the visible noise in
the downsampled image, but not the visible detail.

A Gaussian blur loses contrast at high image frequencies in both noise and subject detail. It prevents or reduces the aliasing of both in the downsizing. Aliased detail is detail that looks sharp, but actually has less detail than the source. Aliased noise is stronger than it would be if properly sampled.

-- hide signature --

John

John Sheehy Forum Pro • Posts: 21,710
Re: Bob..

Steen Bay wrote:

I actually wondered a bit about that, because 40D only has ISO 3200
when "H", "ISO expansion" is activated. Is ISO 3200 a 'real' ISO on
40D, or just a 'pushed' ISO 1600? Doesn't that make a difference?

The actual exposure is what makes the real difference.

The 40D shot is 1/400 and the 50D shot is 1/500, so, unless something in the setup gave different illumination, the 50D may be exposed 1/3 stop less.

-- hide signature --

John

photo nuts Senior Member • Posts: 1,364
Regarding Pixmantra's stand on DXOMark..

I have to agree some of the quantitative results do not agree with real world shooting experience.

For example, I have owned and used the Canon 400D and 450D quite extensively. In their analysis, the 400D is equal, if not slightly better than the 450D at high ISO. I call that BS. The 400D produces white speckles in the image at high ISO. This type of random noise is extremely hard to remove while the 450D is free from such ugly white spots. Even the-digital-picture has the same findings when they convert their RAW images. But somehow this was not captured in DXOMark's quantitative tests.

I do not understand this.

But it does show some kind of serious flaw in DXOMark's tests.

bobn2
bobn2 Forum Pro • Posts: 61,383
Re: Bob..

Steen Bay wrote:

bobn2 wrote:

Steen Bay wrote:

bobn2 wrote:

Steen Bay wrote:

Thanks. Reading Emil's reply again, I can see that he actually did
explain it nicely (and you too, of course), but still I find it a bit
counter-intuitive, that a gaussian blur affects the visible noise in
the downsampled image, but not the visible detail. Anyway, I'll take
your word for it, that it's the best method when downsampling, also
to prevent aliasing. (And it would of course be pretty easy to check,
downsampling the 50D resolution testchart shot, with and without the
gaussian blur)

Here you are 50D downsampled to 6.7MP, this time cubic, top no blur,
bottom blur.

Bob

Interesting, thanks. Isn't the top one, with no gaussian blur, the
sharpest, with most detail and noise? Does it have any
aliasing/artifacts? ISO 3200?

Could be I've overdone the blur a tad - if I got it right it
shouldn't affect sharpness. Could well be aliasing in the top one,
you'd need to hunt around the fabrics.
These are all ISO3200, same CR2's as PIX used.

I actually wondered a bit about that, because 40D only has ISO 3200
when "H", "ISO expansion" is activated. Is ISO 3200 a 'real' ISO on
40D, or just a 'pushed' ISO 1600? Doesn't that make a difference?

In practice, no. Remember, at 3200 ISO the D is maybe 9 stops or so, that means he 14 bit ADC is capturing 5 bits more information than it needs to. In that circmstance, a noiseless digital scaling might actually be better than a noisy analog amplification (although with a properly designed feedback amplifier, increasing the gain won't disproportionately increase the noise) - much too much is made of 'fake' high ISO's. In some MF backs, all ISO's bar the base are 'fake'.

-- hide signature --

Bob

Steen Bay Veteran Member • Posts: 7,418
Interesting..

bobn2 wrote:

Here's one I prepared earlier.
40D and 50D, processed in dcraw at defaults, no other processing.
Rerezzed to 30MP,22MP,15MP,10MP,6.7MP. Upsampling straight Lanczos.
Downsampling appropriate blur + Lanczos. No other processing applied.
Sorry for just 200x200 crops, makes it faster to view in DPR

It's actually pretty interesting (and a bit confusing :)) to see images compared at so many different output (or simulated print) sizes. To see how much the impression of noise/detail in the image depends on the 'magnification'.

Steen Bay Veteran Member • Posts: 7,418
Exposure?

John Sheehy wrote:

Steen Bay wrote:

I actually wondered a bit about that, because 40D only has ISO 3200
when "H", "ISO expansion" is activated. Is ISO 3200 a 'real' ISO on
40D, or just a 'pushed' ISO 1600? Doesn't that make a difference?

The actual exposure is what makes the real difference.

The 40D shot is 1/400 and the 50D shot is 1/500, so, unless something
in the setup gave different illumination, the 50D may be exposed 1/3
stop less.

John

Yes, that would be a disadvantage for the 50D, but the brightness level seems pretty much the same. If anything the 50D image actually seems a tab brighter. But suppose it's possible that 50D uses a different tonecurve.

bobn2
bobn2 Forum Pro • Posts: 61,383
Re: Interesting..

Steen Bay wrote:

It's actually pretty interesting (and a bit confusing :)) to see
images compared at so many different output (or simulated print)
sizes. To see how much the impression of noise/detail in the image
depends on the 'magnification'.

It is a bit confusing, isn't it. Hopefully, it makes the point about the difficulty of straightforward extrapolation from one comparison. Given that many rely on 'experts' to do that comparison for them, it would be nice if the 'experts' actually understood the issues.
--
Bob

bobn2
bobn2 Forum Pro • Posts: 61,383
Re: Exposure?

Steen Bay wrote:

John Sheehy wrote:

Steen Bay wrote:

I actually wondered a bit about that, because 40D only has ISO 3200
when "H", "ISO expansion" is activated. Is ISO 3200 a 'real' ISO on
40D, or just a 'pushed' ISO 1600? Doesn't that make a difference?

The actual exposure is what makes the real difference.

The 40D shot is 1/400 and the 50D shot is 1/500, so, unless something
in the setup gave different illumination, the 50D may be exposed 1/3
stop less.

John

Yes, that would be a disadvantage for the 50D, but the brightness
level seems pretty much the same. If anything the 50D image actually
seems a tab brighter. But suppose it's possible that 50D uses a
different tonecurve.

Not in this case, the images are from raw and the tone curve applied in dcraw. It's the default straight line, hence PIX's comments about the less than optimum conversion.

-- hide signature --

Bob

Steen Bay Veteran Member • Posts: 7,418
Resampling..

bobn2 wrote:

Steen Bay wrote:

bobn2 wrote:

Steen Bay wrote:

bobn2 wrote:

Steen Bay wrote:

Thanks. Reading Emil's reply again, I can see that he actually did
explain it nicely (and you too, of course), but still I find it a bit
counter-intuitive, that a gaussian blur affects the visible noise in
the downsampled image, but not the visible detail. Anyway, I'll take
your word for it, that it's the best method when downsampling, also
to prevent aliasing. (And it would of course be pretty easy to check,
downsampling the 50D resolution testchart shot, with and without the
gaussian blur)

Here you are 50D downsampled to 6.7MP, this time cubic, top no blur,
bottom blur.

Bob

Interesting, thanks. Isn't the top one, with no gaussian blur, the
sharpest, with most detail and noise? Does it have any
aliasing/artifacts? ISO 3200?

Could be I've overdone the blur a tad - if I got it right it
shouldn't affect sharpness. Could well be aliasing in the top one,
you'd need to hunt around the fabrics.
These are all ISO3200, same CR2's as PIX used.

I actually wondered a bit about that, because 40D only has ISO 3200
when "H", "ISO expansion" is activated. Is ISO 3200 a 'real' ISO on
40D, or just a 'pushed' ISO 1600? Doesn't that make a difference?

In practice, no. Remember, at 3200 ISO the D is maybe 9 stops or so,
that means he 14 bit ADC is capturing 5 bits more information than it
needs to. In that circmstance, a noiseless digital scaling might
actually be better than a noisy analog amplification (although with a
properly designed feedback amplifier, increasing the gain won't
disproportionately increase the noise) - much too much is made of
'fake' high ISO's. In some MF backs, all ISO's bar the base are
'fake'.

Bob

OK, that's not an issue, but about the downsampling, don't you prefer the image without gaussian blur yourself? To me the 'blurred' one looks like an image with to much NR.

Steen Bay Veteran Member • Posts: 7,418
Yes, a bit tricky..

John1940 wrote:

Steen,

I think that one of the issues with regard to comparing the 50D and
40D is that the pixel spacing (horizontally or vertically) only drops
by about 18% in going from the 40D to the 50D.
Here is a test that might prove something regarding the effectiveness
of doing Gaussian blur before down-sampling with a series of
successive doubling pixel-to-pixel spacing changes. It’s one that I
have done before many times but I did not do the blurring first.
However, I still have the evidence and may do some studies if I ever
get the time.
I have scanned many 35-mm Kodachrome slides using a Canon 4000 dpi
film scanner. When I first got the machine almost six years ago I
decided to do quite a few tests with film and digital cameras I had
at the time. My newest camera was a Canon 300D (Rebel) DSLR. I also
had (and still have) a 35-mm Canon film body (which takes EF lenses)
and had a Minolta 35-mm system with six lenses.
In one series of tests (using several slides but no DSLRs) I scanned
each slide at 4000 dpi, 2000 dpi, 1000 dpi and 500 dpi successively
without touching or moving the slide at all. In round numbers I got a
20 Mp, 5 Mp, 2.5 Mp, and 1.25 Mp image in Tiff-8 or Tiff-16 format
(or RGB format). The scanner can do 8 bits or 12 bits per color at
every pixel location. Each pixel therefore uses either 3 bytes or 6
bytes of memory. I then did various (but not scientifically rigorous)
comparisons by down-sampling but did not do Gaussian blur with the
higher resolutions first.
The potential value of this test method (which I’m sure has been used
by many people in many organizations going back decades) is that all
other variables are removed since the source of the image is the same
and the scanner is the same. Also, the steps in resolution are large
enough to show the primary advantage of the higher pixel count
(detail). I never found any increase in noise in my unscientific
experiments. But I never looked at whether Gaussian blur before
down-sampling would make what I thought was a good result even better
in favor of the higher-resolution scans.
I think that the scanner fires the same number of photons (as such
things go) for each pixel location. If I’m right, that takes another
variable off the table, or at least makes it knowable.
At this point I’m in the 50D-is-better camp for detail and for being
no worse for noise if down-sampled to 10 Mp with Gaussian blur done
first. But, unlike with the scanner, it’s not easy to do accurate
side-by-side tests with the 40D and 50D. There are too many variables
that cannot be controlled easily.

John1940

I still use my 'old' 40D. Can't afford, or justify, to update every year, but I'm sure that the 50D is the better camera, and I'm looking forward to a 18-20mp 60D, which I'll probably get. And yes, comparing cameras and images can be a bit tricky, which all these long threads should prove

bobn2
bobn2 Forum Pro • Posts: 61,383
Wanna be confused a bit more...

Same thing with D3 and D3x, at native res of each, 200 and 6400ISO. (For interest only, the D3 has several optimisations for high ISO and the D3x for low ISO that have nothing to do with pixel density)

-- hide signature --

Bob

bobn2
bobn2 Forum Pro • Posts: 61,383
Re: Resampling..

Steen Bay wrote:

OK, that's not an issue, but about the downsampling, don't you prefer
the image without gaussian blur yourself? To me the 'blurred' one
looks like an image with to much NR.

I think that's to do with the 'quality' of the blurred and downsampled noise. Different cameras and converters produce different styles of noise, and different LPF functions also. In this case, I do like the quality of the raw noise better. Also, you don't get a clear idea of the visual quality of the final product from normal viewing distance from pixel peeping. If the pixels were print size rather than monitor size, you'd be hard pressed to see any difference.

-- hide signature --

Bob

OP berleconi Contributing Member • Posts: 720
Re: Interesting..

Some suggested the same part of the sensor area,
Displayed the same size on screen.
Also same target the same percent on the sensor.

The example was not – what was different?

Berl.
http://FreePhotoSoftware.shorturl.com

Steen Bay Veteran Member • Posts: 7,418
Another possible issue..

bobn2 wrote:

Steen Bay wrote:

John Sheehy wrote:

Steen Bay wrote:

I actually wondered a bit about that, because 40D only has ISO 3200
when "H", "ISO expansion" is activated. Is ISO 3200 a 'real' ISO on
40D, or just a 'pushed' ISO 1600? Doesn't that make a difference?

The actual exposure is what makes the real difference.

The 40D shot is 1/400 and the 50D shot is 1/500, so, unless something
in the setup gave different illumination, the 50D may be exposed 1/3
stop less.

Yes, that would be a disadvantage for the 50D, but the brightness
level seems pretty much the same. If anything the 50D image actually
seems a tab brighter. But suppose it's possible that 50D uses a
different tonecurve.

Not in this case, the images are from raw and the tone curve applied
in dcraw. It's the default straight line, hence PIX's comments about
the less than optimum conversion.

When comparing the detail/sharpness from high resolution cameras like 40D and 50D, then it's pretty important that the focus plane is exactly the same, which (looking at the comparisons in the DPR reviews) seems to be a bit difficult to achieve.

ejmartin Veteran Member • Posts: 6,274
Re: Bob..

John Sheehy wrote:

Steen Bay wrote:

Thanks. Reading Emil's reply again, I can see that he actually did
explain it nicely (and you too, of course), but still I find it a bit
counter-intuitive, that a gaussian blur affects the visible noise in
the downsampled image, but not the visible detail.

A Gaussian blur loses contrast at high image frequencies in both
noise and subject detail. It prevents or reduces the aliasing of
both in the downsizing. Aliased detail is detail that looks sharp,
but actually has less detail than the source. Aliased noise is
stronger than it would be if properly sampled.

It seems though that some people (not you, as you have mentioned many times) actually like aliased detail and its illusion of sharpness, otherwise there wouldn't be a market for taking the AA filters off of sensors.

-- hide signature --
Steen Bay Veteran Member • Posts: 7,418
Re: Interesting..

berleconi wrote:

Some suggested the same part of the sensor area,
Displayed the same size on screen.
Also same target the same percent on the sensor.

The example was not – what was different?

Berl.

It's the same percentage of the sensor area in each of the comparisons, like for example 40D and 50D, both upsized to 30mp to simulate a big print. It's the 'simulated prints' that have different sizes, just like they should have.

bobn2
bobn2 Forum Pro • Posts: 61,383
Re: Interesting..

berleconi wrote:

Some suggested the same part of the sensor area,
Displayed the same size on screen.
Also same target the same percent on the sensor.

The example was not – what was different?

Imaging resource does not always set up their test scene the same, there are noticeable differences in the relative position of objects test to test, also the lighting.

-- hide signature --

Bob

bobn2
bobn2 Forum Pro • Posts: 61,383
Re: Another possible issue..

Steen Bay wrote:

bobn2 wrote:

Steen Bay wrote:

John Sheehy wrote:

Steen Bay wrote:

I actually wondered a bit about that, because 40D only has ISO 3200
when "H", "ISO expansion" is activated. Is ISO 3200 a 'real' ISO on
40D, or just a 'pushed' ISO 1600? Doesn't that make a difference?

The actual exposure is what makes the real difference.

The 40D shot is 1/400 and the 50D shot is 1/500, so, unless something
in the setup gave different illumination, the 50D may be exposed 1/3
stop less.

Yes, that would be a disadvantage for the 50D, but the brightness
level seems pretty much the same. If anything the 50D image actually
seems a tab brighter. But suppose it's possible that 50D uses a
different tonecurve.

Not in this case, the images are from raw and the tone curve applied
in dcraw. It's the default straight line, hence PIX's comments about
the less than optimum conversion.

When comparing the detail/sharpness from high resolution cameras like
40D and 50D, then it's pretty important that the focus plane is
exactly the same, which (looking at the comparisons in the DPR
reviews) seems to be a bit difficult to achieve.

The time taken to set up a perfectly comparable test could be long. Interesting to note, however that both the 50D/40D and D3/D3x examples clearly show the detail advantage of the downsampled high res camera over the native low res camera. I'm sure every such test would do so - you'd just have to ignore the noise differences of very different sensor technologies.
--
Bob

Keyboard shortcuts:
FForum MMy threads