Sensor ?

[Erik Magnuson]

I would like to see proof of this. It's certainly not my experience,
and I do have two CFA cameras, albeit not SLRs.

Simple: look at DPR's resolution charts. The 450D gets avg of 2250 line pairs per pixel height for a height of 2848 pixels. That's 79% linear or 62% area of pixel resolution. The nikon D3 comes out at 66% area. I thought there was a camera that came close to 70% area with in-camera processing but I could be wrong.

For raw processing, photozone.de measures values over 2150 for the 350D with 2304 vertical pixels. That's 93% linear and 87% area.

It may be possible to simulate a greater level of detail by, for
example, using clever interpolation and sharpening techniques, but I
cannot see that a CFA sensor can ever correctly detect a piece of
detail which, without the AA filter, would illuminate a single pixel.

It depends on the contrast and the hue changes involved -- it not that hard for B&W.

--
Erik

 

[DaSonyGuy]

If the sd 14 is considerd a 4.7 mp sensor (capturing 3 colors at each
pixel ) then wouldn"t a 12 mp beyer sensor really be a 4.0 mp X 3
colors = 12 ???..

Not quite...Bayer sensors have twice as many green photosites as they do red or blue photosites so a 12mp Bayer sensor is really a 6mp green + 3mp red + 3mp blue = 12mp total sensor.

This means that although the SD14's 14mp Foveon X3 sensor has 1.3mp less green photosites than a 12mp Bayer sensor at the same time it has 1.7mp more red photosites and 1.7mp more blue photosites.

It is the extra red and blue photosites that give the Foveon sensor a much greater colour gamut and therefore why it is able to more faithfully record the true colours of the subject being photographed, in most situations.

As for resolution, several side by side comparison tests definitely seem to confirm that in practice the 14mp Foveon X3 sensor resolves about the same as a good quality 12mp Bayer sensor.

--
DSG
--



--
http://sigmasd10.fotopic.net/

 

[Roland Karlsson]

It may be possible to simulate a greater level of detail by, for
example, using clever interpolation and sharpening techniques, but I
cannot see that a CFA sensor can ever correctly detect a piece of
detail which, without the AA filter, would illuminate a single pixel.

It all depends on the model used when doing the reconstruction. If you assume slow color changes and you assume that edges follows some reasonable spline or something - then you can do quite good reconstruction. One typical example is images of a city. The windows and all horizontal and vertical lines are super sharp. All the trees - with their fuzzy leaves - have a much lower resolution. This looks strange IMHO. And - of course - grid structures - eg blinds - creates big problems.

So - CFA reconstruction may sometimes result in very high resolution - when the model matches reality.

--
Roland

 

[Roland Karlsson]

It is the extra red and blue photosites that give the Foveon sensor a
much greater colour gamut and therefore why it is able to more
faithfully record the true colours of the subject being photographed,
in most situations.

WOA! Not so fast DSG.

This is definitely not the case.

Any difference in color gamut is only due to the spectral response of the X3 color channels.

--
Roland

 

[gbainbridge]

Well I don't know how those figures were obtained, but let's take an actual photographic test - the DP1 versus D60 test from Dpreview:

http://bainb.dyndns.org/Sigma%20Forum/DP ! D60_Comp.jpg

In the DP1 section, blowing it up we can see that, ignoring the aliasing above and below it, the last point where the DP! sensor correctly displays 9 individual lines is at 1800 lines per image height:

http://bainb.dyndns.org/Sigma%20Forum/DP ! Res.jpg

as the image height is 1760 pixels, this is not surprising, meaning that it has individual pixel resolution.

Now blow up the D60 image.



Despite the part of the chart that the reviewer has chosen, the reality is that you can only count 9 individual lines at the very top of the image, and even there they are quite smeared. By the time you have got down to 2000 lines per image height, although at first sight it looks O.K. in the normal size image, in the blow-up you can only count 8 lines. By coincidence the very top of the image also corresponds to 1800 lines per image height, which for the D60 with an image height of 2592 pixels works out at 69% of the vertical resolution or 48% of area.

I'm sorry, but despite any claims otherwise, any apparent luminance detail greater than the number of green pixels on a CFA sensor is a result of clever software, and is not actually detected by the sensor.

--
Thanks,
Gary.

 

[gbainbridge]

It all depends on the model used when doing the reconstruction. If
you assume slow color changes and you assume that edges follows some
reasonable spline or something - then you can do quite good
reconstruction.

Hence my point about clever interpolation and sharpening algorithms.

IF the subject is black and white and IF the software is clever enough to understand that, then you can ignore the colour filters and produce greater detail, but the software can never know whether that small single pixel sized black spot is really black or simply the wrong colour for that particular pixel.

--
Thanks,
Gary.

 

[Roland Karlsson]

... but the software can never know whether that
small single pixel sized black spot is really black or simply the
wrong colour for that particular pixel.

That is only true for super sharp lenses and no AA filter. And to avoid situations like that is why you add the AA filter.

But - the same goes really for any sensor that does not have 100% fill factor. The information that falls in between the detectors is not detected. So - if you are unlucky - you might miss something - i.e if you are taking images of the starry sky.

--
Roland

 

[gbainbridge]

That is only true for super sharp lenses and no AA filter. And to
avoid situations like that is why you add the AA filter.

I accept that, but once again the AA filter is reducing the level of detail that the sensor can detect. That's why there are so many people out their removing or changing their AA filters on their CFA cameras*.

--
Thanks,
Gary.

Like an acquaintance of mine who replaced the AA filter on his D200 with one from a D70, and reported a major improvement in resolution.

 

[Erik Magnuson]

In the DP1 section, blowing it up we can see that, ignoring the
aliasing above and below it,

Well, yes this is part of the problem. We are talking about "resolutions" that are above Nyquist so even with an AA filter there can be positional errors.

Despite the part of the chart that the reviewer has chosen, the
reality is that you can only count 9 individual lines at the very top
of the image, and even there they are quite smeared.

No, they are there, just very low contrast. Increase the contrast and it becomes much clearer:



The upper line is the 2000 and the lower the 2200 level. See the 9 lines now?

I'm sorry, but despite any claims otherwise, any apparent luminance
detail greater than the number of green pixels on a CFA sensor is a
result of clever software, and is not actually detected by the sensor.

This is like claiming that because X3 doesn't measure R, G, B directly, the RGB output "is a result of clever software, and is not actually detected by the sensor." The output in both cases is a derived product after much processing.

--
Erik

 

[Roland Karlsson]

OK - thats the old "9 converging lines" example

What you see here depends on your own imagination mainly

Look at the images carefully - maybe at 200% depending on how good your sight is.

What do you see?

This is what I see:

The Foveon image contains a striped pattern of 1 pixel wide parallel lines. A pattern that now and then is disturbed by some fuzziness. There are no (0) converging lines - not a trace of it. So - the Foveon image is totally bogus. It does not resemble the original image at all. It is all invented.

The Bayer image (or rather the image with an AA filter) contains very fuzzy converging lines. Ohhh ... they are aliased - lots of it. But - it is a distorted image of converging lines - all the way. So - it is a nice try at showing the actual image.

What did you see? Before and after reading what I saw?

What you really should do now is upscaling the images using some smoothing algorithm - e.g. bicubic. I think you will be surprised.

--
Roland

 

[JohnK]

As for resolution, several side by side comparison tests definitely
seem to confirm that in practice the 14mp Foveon X3 sensor resolves
about the same as a good quality 12mp Bayer sensor.

I agree, having informally compared DP1 to 5D, that's my conclusion, FWIW. BTW I'm a professional designer, painter, scenic artist and have a pretty good eye for detail
--
JohnK
Take a picture, it'll last longer.

 

[1chaz]

Another thing I have noticed is the large raw file size for a 4.7 mp 11 to 15 mb per shot when my 20d is 6 to 7 mp in raw maybe one is compressed and one is not ??. I do know that I like this camera and the results that you can get with a little forethought and planning , I will post some pics as soon as I have a chance FWIW am a Lithographer by trade
--

personnaly I think both sensors have thier positives and negitives what I want is the gigapixel sensor

 

[Erik Magnuson]

Another thing I have noticed is the large raw file size for a 4.7 mp
11 to 15 mb per shot when my 20d is 6 to 7 mp in raw maybe one is
compressed and one is not ??.

Both are compressed which is why the size varies. Both also contain an embedded jpeg -- which also varies in size due to compression. The SD14 embeds a highest quality jpeg which tends to inflate the raw file size somewhat.

--
Erik

 

[rocklobster]

The fact is that usually half of the pixels in a bayer CCD are Green with the remaining half divided equally between red and blue. Bayer advocates will try to convince you, because the luminance information is predominantly in the Green area of the spectrum, that the real resolution is closer to half of the total. But then you have anti-aliasing and other filtering that reduces the resolution.

Ultimately though, the proof is in the picture!

Cheers

 

[Roland Karlsson]

The fact is that usually half of the pixels in a bayer CCD are Green
with the remaining half divided equally between red and blue. Bayer
advocates will try to convince you, because the luminance information
is predominantly in the Green area of the spectrum, that the real
resolution is closer to half of the total. But then you have
anti-aliasing and other filtering that reduces the resolution.

Thats a too simplistic view. The green pixels are put in a checkerboard pattern. In some cases this helps the interpretation and in some cases it is the opposite. When it helps you can get 100% resolution - just as there was no missing samples. But - it might also result in strange meandering patterns and other irritating stuff. I personally think this checkerboard pattern is one of the major problems with Bayer CFA. It makes the image having very different resolution in different parts.

--
Roland

 

[gbainbridge]

Very low contrast, as looking at a histogram will show you. Well below what would normally be taken as the extinction point (-6db equivalent to 25% contrast)

This is like claiming that because X3 doesn't measure R, G, B
directly, the RGB output "is a result of clever software, and is not
actually detected by the sensor." The output in both cases is a
derived product after much processing.

No, because unlike CFA software the Foveon software is not adding spatial information that is not seen by the sensor.

The question is, how far do you take it? Clever software could count the number of lines and realise that they are diminishing to a point, then draw them all the way, apparently giving the camera infinite resolution.

As with Roland's earlier point, with a cityscape the software can recreate the straight lines of buildings and correctly infer detail which the sensor does not actually see, but as soon as you go to truly random detail, such as foliage or pebbles on a beach, the paradigm breaks down and the level of detail drops dramatically. e.g. in the thread starting at http://forums.dpreview.com/forums/read.asp?forum=1027&message=28857642

To end this thread (as I have to go to work!) here is a dramatic example of artificially created detail, from a picture taken of myself and colleagues when we were in Turin, using a certain CFA camera (which received a recommended rating from Dpreview):



look at it at 100% and there is no real detail at all:



--
Thanks,
Gary.

 

[gbainbridge]

I would not disagree with that. It's a result of aliasing, one of the disadvantages of not having an AA filter.

There is a valid argument that even the Foveon cameras should have AA filters. How efffective that would be I don't know.

However the reality is that I have never got the level of detail from any of my CFA cameras as I get from the Foveons- just bigger pictures.

Here is a good example (not embedded as it's full size):

http://bainb.dyndns.org/photo/convert.php?showorig=true&currname=Gary_B_SDIM1202_SD14.jpg&currdir=Photokina&showvideo=ture

--
Thanks,
Gary.

 

[Roland Karlsson]

I would not disagree with that. It's a result of aliasing, one of the
disadvantages of not having an AA filter.

There is a valid argument that even the Foveon cameras should have AA
filters. How efffective that would be I don't know.

However the reality is that I have never got the level of detail from
any of my CFA cameras as I get from the Foveons- just bigger pictures.

Here is a good example (not embedded as it's full size):

http://bainb.dyndns.org/photo/convert.php?showorig=true&currname=Gary_B_SDIM1202_SD14.jpg&currdir=Photokina&showvideo=ture

Very nice example. And yes - an AA filter would not improve the visual impact of this photo. But - what you really should do (for optimal image quality) is having by far more pixels and then an AA filter.

This is how you do for audio. An AA filter and then oversampling and then downsampling using low pass filtering - thats the method to get optimal quality.

If you dont oversample and have no AA filter - then you will get a distincter and more high frequency content sound with fewer samples per second. It will sound bright. But ... at the same time aliasing will introduce distortion of the sound.

We are not as sensitive to distortion of images than distortion of sound. Therefore - it is possible to get away with aliasing in images. Something the Sigma cameras show nicely.

But .... replacing sharp converging lines with sharp parallel lines does at last not work so well for me. Fuzzy converging lines is not so nice either - but it is nearer to reality.

--
Roland

 

[Roland Karlsson]

Hmmmm ... this was a faulty example IMHO. The strange look is due to excessive noise reduction. No low noise CFA images looks anything like this.

--
Roland

 

[Erik Magnuson]

Very low contrast, as looking at a histogram will show you. Well
below what would normally be taken as the extinction point (-6db
equivalent to 25% contrast)

Then the AA filter is working as designed. This detail is not high-confidence anyway.

No, because unlike CFA software the Foveon software is not adding
spatial information that is not seen by the sensor.

No, it's deriving color information not directly sensed: interpolation in a different dimension -- vertically instead of horizontally.

The question is, how far do you take it?

To the point where it works to provide an acceptable image most of the time and rarely provides an unacceptable image. By the time ink hits paper, all of the input samples have undergone extensive processing for either technology.

but as soon as you go to
truly random detail, such as foliage or pebbles on a beach, the
paradigm breaks down and the level of detail drops dramatically.

Then you switch to a different demosaicking algorthm or adjust the tuning of the current one to be less aggressive. Remember that much of the "detail" you will see with X3 at this same level of magnification will be no more accurate as it will be highly aliased. It may be more visually pleasing but that depends a lot on the details of the processing.

look at it at 100% and there is no real detail at all:

A 100% view is essentially a 32" wide print. Show me an X3 camera with an 18x zoom that can do better at ISO 800.

--
Erik