Final output of the Foveon X3 ?

[IamJacksBrain]

Boris,

I thank you for attempting to apply some practical information to an academic discussion. I do think there may have been a couple of problems with your methodology though. (Keep in mind this is new to all of us and my methodology may be flawed too.)

I think a fairer assessment could be made if you upsampled the foveon image to match the same resolution as your other image, and to do away with the sharpening in Photoshop. (Of course if it were really fair then these printing tests would occur with raw images instead of jpegs, but whatever.)

The reason I think the foveon image should be upsampled is because a print with fewer pixels isn't going to look as good regardless of quality of the pixel information. (A picture printed with only 20 pixels should look twice as good as a picture with only 10 pixels, even if those 10 are more accurate.) Also, since bayer cameras apply their interpolation before the image leaves the camera to acheive their resolution it seems to me that applying interpolation to a foveon image to acheive the same resolution would be a good test.

Just my thoughts.

Mario

I printed two of the sd9 samples from Phil's site

All on a Epson 1280 and high gloss paper. I applied a auto level
and sharp 200%-.03-0
to all in Photoshop.

 

[David Barkin]

Hi Ron

You are resonding to technical posts with technical posts. Fine. This post is not meant as a critique of that.

But what I'd like to know is if we're comparing apples and oranges or apples and apples.

Is there a realistic comparison possible? Is the Foveon chip equal to a bayer chip twice it's size, 50 percent?

Is the fact that there are different implementations of the bayer chip a problem and to what degree?

If you can make such a comparison let it be on the basis of the best existing bayer pattern.

I don't mean to put you on the spot. Exisiting terminology and the lay understanding of this is that when a camera is rated at 6MP it's going to be twice as powerful as a 3MP. Along comes Foveon and the exisiting ratings don't seem to mean much.

So my question really boils down to assigning ratings to chips that will make some sense to the person going out to buy a camera.

Dave

No.

The advantages are in better reproduced colour and lack of moiré
patterns in certain situations.

There is no real gain in sharpness or picture resolution.

The goal is to have improved sharpness and picture resolution.
Moire is a symptom of the undelrying problem Foveon is trying to
fix. It's not the problem.

--
Ron Parr
FAQ: http://www.cs.duke.edu/~parr/photography/faq.html
Gallery: http://www.pbase.com/parr/

 

[David Barkin]

Hi Mario

Good post! Yes interpolate the Foveon up and see what comes out. I print at 16 x 20 and always interpolate up anyway!

Dave

I thank you for attempting to apply some practical information to
an academic discussion. I do think there may have been a couple of
problems with your methodology though. (Keep in mind this is new
to all of us and my methodology may be flawed too.)

I think a fairer assessment could be made if you upsampled the
foveon image to match the same resolution as your other image, and
to do away with the sharpening in Photoshop. (Of course if it were
really fair then these printing tests would occur with raw images
instead of jpegs, but whatever.)

The reason I think the foveon image should be upsampled is because
a print with fewer pixels isn't going to look as good regardless of
quality of the pixel information. (A picture printed with only 20
pixels should look twice as good as a picture with only 10 pixels,
even if those 10 are more accurate.) Also, since bayer cameras
apply their interpolation before the image leaves the camera to
acheive their resolution it seems to me that applying interpolation
to a foveon image to acheive the same resolution would be a good
test.

Just my thoughts.

Mario

I printed two of the sd9 samples from Phil's site

All on a Epson 1280 and high gloss paper. I applied a auto level
and sharp 200%-.03-0
to all in Photoshop.

 

[Ron Parr]

So my question really boils down to assigning ratings to chips that
will make some sense to the person going out to buy a camera.

I'm afraid that there's no magic formula. It's like trying to compare stereo equipment by using the specifcations. The specifications give you some rough guidelines, but in the end there are many factors that can't be captured by a few numbers. You need to use your ears and go with the one that sounds best.

For the Foveon vs. Bayer thing. I'm not interested in trying to predict some magic multiplier that will convert from one to the other. In most of my posts, I'm just trying to address some misunderstandings about what the technology does and what the motivations behind it are.

In the end, you should read reviews from people you trust, like Phil, and use your eyes to decide which produces the best images.

--
Ron Parr
FAQ: http://www.cs.duke.edu/~parr/photography/faq.html
Gallery: http://www.pbase.com/parr/

 

[Mark Palumbo]

I agree about doing the upsampling of the Fovean image and also to NOT blindly apply USM to all images. That actually seems to be unfair to me not all cameras or even image to image require the same amount of USM if any is needed at all. To just apply it for the sake of application would skew the results, unless that's what one intends.

I thank you for attempting to apply some practical information to
an academic discussion. I do think there may have been a couple of
problems with your methodology though. (Keep in mind this is new
to all of us and my methodology may be flawed too.)

I think a fairer assessment could be made if you upsampled the
foveon image to match the same resolution as your other image, and
to do away with the sharpening in Photoshop. (Of course if it were
really fair then these printing tests would occur with raw images
instead of jpegs, but whatever.)

The reason I think the foveon image should be upsampled is because
a print with fewer pixels isn't going to look as good regardless of
quality of the pixel information. (A picture printed with only 20
pixels should look twice as good as a picture with only 10 pixels,
even if those 10 are more accurate.) Also, since bayer cameras
apply their interpolation before the image leaves the camera to
acheive their resolution it seems to me that applying interpolation
to a foveon image to acheive the same resolution would be a good
test.

Just my thoughts.

Mario

I printed two of the sd9 samples from Phil's site

All on a Epson 1280 and high gloss paper. I applied a auto level
and sharp 200%-.03-0
to all in Photoshop.

 

[Boris]

Mario

Fair enough for me, this how I would use the camera....take a picture...load it in photo shop, auto level, sharpen and resize and print. I don't interpolate my pictures out of the camera, never have, tried all those programs don't work for me. I could have upsampled (more?) in camera with the s2 and that woudn't have been fair. It is a good 3mp chip..... probably will beat out the 4mp chips and come close to the current 6mp DSLR, It is not going to chase the D100, D60 and S2 into the dustbin.....
Only my opinion of course
Boris

I thank you for attempting to apply some practical information to
an academic discussion. I do think there may have been a couple of
problems with your methodology though. (Keep in mind this is new
to all of us and my methodology may be flawed too.)

I think a fairer assessment could be made if you upsampled the
foveon image to match the same resolution as your other image, and
to do away with the sharpening in Photoshop. (Of course if it were
really fair then these printing tests would occur with raw images
instead of jpegs, but whatever.)

The reason I think the foveon image should be upsampled is because
a print with fewer pixels isn't going to look as good regardless of
quality of the pixel information. (A picture printed with only 20
pixels should look twice as good as a picture with only 10 pixels,
even if those 10 are more accurate.) Also, since bayer cameras
apply their interpolation before the image leaves the camera to
acheive their resolution it seems to me that applying interpolation
to a foveon image to acheive the same resolution would be a good
test.

Just my thoughts.

Mario

I printed two of the sd9 samples from Phil's site

All on a Epson 1280 and high gloss paper. I applied a auto level
and sharp 200%-.03-0
to all in Photoshop.

 

[Stuffy]

Although Norm Koren's simulation shows that X3 should have very slightly better resolution than the D60 or D100:

http://www.normankoren.com/Tutorials/MTF7.html#Res_summary

This is discounting differences relating to moire and noise.

BTW, that X3 sensor is 3.43 megapixels (real pixels, of course), not 3.0.

--Stuffy

General consensus among people profecient and experienced in
digital photography is, that foveon will be comparable with sensors
1.4 to 1.6 times larger than 3 mp. that means that it will produce
the same level of detail as 4.2 - 4.8 megapixel bayer sensor. the
colour fidelity of foveon will probably be higher, however it will
not be able to capture more detail than physically possible. so far
nobody was able to prove this statement wrong. naturally it all
depends on the amount of detail in the scene, however how often
will you take picture of a plain sheet of plastic which has little
detail in it ?

 

[caterpillar]

So my question really boils down to assigning ratings to chips that
will make some sense to the person going out to buy a camera.

I'm afraid that there's no magic formula. It's like trying to
compare stereo equipment by using the specifcations. The
specifications give you some rough guidelines, but in the end there
are many factors that can't be captured by a few numbers. You need
to use your ears and go with the one that sounds best.

For the Foveon vs. Bayer thing. I'm not interested in trying to
predict some magic multiplier that will convert from one to the
other. In most of my posts, I'm just trying to address some
misunderstandings about what the technology does and what the
motivations behind it are.

In the end, you should read reviews from people you trust, like
Phil, and use your eyes to decide which produces the best images.

--
Ron Parr
FAQ: http://www.cs.duke.edu/~parr/photography/faq.html
Gallery: http://www.pbase.com/parr/

Ron,

I totally agree with this. It is similar to the CD vs vinyl record debate. Those with sharp ears and technical saavy will argue that CDs are not faithful audio reproductions, but history has already voted on this. In the end, how would the general consumer accept it? Besides, this is too early to argue certain technical merits (sad) until Phil comes out with the production review (which will be some time still).

And I agree with the others here too that the initial impressions are quite favorable. The leaves are indeed more alive. The texture seems more "film-like." The color, except for Phil's lady seems yellowish, but I presume this was under conditions that Phil did have no time to control or study.

I am more fearful, that if the foveon chip proves to be superior in a more controlled test, IF other manufacturers will jump the bandwagon. The diffusion of this technology must come from them. Sigma has locked on the technology with the lens mount. If they and maybe another camera maker decide to make a go at it (maybe Minolta?), they have to come out with a lens mount that will attract current nikon, canon users. They must also come out with a mid-level, if not entry level solutions if they are to make a difference.

Your points though are well made and much appreciated.

--

• Caterpillar

 

[Sean Keegan]

Chris, Joe said it has 10.62 million light sensors which is completely correct. They are just at 3.54 million individual pixel locations. Current sensors do not even come close to that many light sensors.

Regards,
Sean

Look ... a 3.54mps X3 has 3X3.54 (10.62) light sensors

Incorrect. It has 3.54m sensors, and this is vitally important.
There are not 3 sensors at each location magically measuring
"red", "green" and "blue" light. There is a sensor which is
charged according to the light that falls on it. However, this
sensor is not a 2-dimensional rectangle, it has depth and the
charge generated by light falling on it will be distributed
throughout this depth.

It so happens that the charge varies with depth according to the
wavelength of the light, and Foven takes advantage of this by
sampling the "bucket" at 3 different points.

However, it's nothing like the highly sanitised view you seem to be
taking where there are nice distinct layers, marked "red", "green"
and "blue", with nice little boundaries between them. In reality,
it's much, much more messy than that. The absorbsion is a
statistical process, and the "red" light will actually be
distributed throughout the "bucket", mingled with other parts of
the spectrum. It's just that there's likely to be more of it at a
given depth.

When you sample this "red" light, you're also getting plenty of
yellow, green and even blue light thrown in for good measure. The
clever bit is being able to tidy this mess up, by comparing the
"red" sample with the "green" and "blue" samples to try and work
out what really is red, and what is just garbage that leaked from
elsewhere in the bucket.

If you just took the direct outputs from each of the 3 samples in a
Foveon pixel, and made an image based on those with no processing,
you'd most likely end up with a sludgy grey/brown mess.

So whilst you may have to do lots of tricky maths with a mosaic
sensor to get an estimate of the colour at each location, you laso
have to do lots of tricky maths with the Foveon sensor, and what
you get is, again, merely an estimate of the colour at any given
position. The nature of the maths is different, but they're both
still statistical estimates. Which one is better will be revealed
by the final images.

 

[JoePhoto]

So whilst you may have to do lots of tricky maths with a mosaic
sensor to get an estimate of the colour at each location, you laso
have to do lots of tricky maths with the Foveon sensor, and what
you get is, again, merely an estimate of the colour at any given
position. The nature of the maths is different, but they're both
still statistical estimates. Which one is better will be revealed
by the final images.

The math is not so tricky. It's described in this Kodak patent:

http://patft.uspto.gov/netacgi/nph-Parser?Sect1=PTO2&Sect2=HITOFF&p=1&u=/netahtml/search-bool.html&r=1&f=G&l=50&co1=AND&d=ft00&s1= '5668596'.WKU.&OS=PN/5668596&RS=PN/5668596

It basically involves one matrix multiplication per pixel. Since
you need to do at least one such multiplication anyway for white
balance, you can just premultiply your white balance matrix by the
correction matrix. In actual implementation, it boils down to
changing a few numbers in ROM. There is no extra computation done.

For a Bayer pattern sensor, the color filters and substrate also
behave stochastically. The difference is that there is somewhat
more overlap with the Foveon approach.

There is NO "overlap" with the Foveon X3 approach - as long as you are relating it to 3.54mps. With Bayer/interpolation ... there is overlap as soon as you go above 1.5mps.

--
Ron Parr
FAQ: http://www.cs.duke.edu/~parr/photography/faq.html
Gallery: http://www.pbase.com/parr/

--
Thanks for reading .... JoePhoto

( Do You Ever STOP to THINK --- and FORGET to START Again ??? )

 

[JoePhoto]

Mario
Fair enough for me, this how I would use the camera....take a
picture...load it in photo shop, auto level, sharpen and resize
and print. I don't interpolate my pictures out of the camera, never
have, tried all those programs don't work for me. I could have
upsampled (more?) in camera with the s2 and that woudn't have been
fair. It is a good 3mp chip..... probably will beat out the 4mp
chips and come close to the current 6mp DSLR, It is not going to
chase the D100, D60 and S2 into the dustbin.....
Only my opinion of course
Boris

Your "opinion" could be wrong, (in my opinion).

It is NOT a 3mps chip ... it is a 3 X 3 mps chip ... (actually a 3.54 X 3 = 10.62mps chip .... that compares to a Bayer equivalent of 4 X 3.54 = 14.16mps).

So it will be VERY INTERESTING to see how it compares to the new 14mps camera announcement.

I thank you for attempting to apply some practical information to
an academic discussion. I do think there may have been a couple of
problems with your methodology though. (Keep in mind this is new
to all of us and my methodology may be flawed too.)

I think a fairer assessment could be made if you upsampled the
foveon image to match the same resolution as your other image, and
to do away with the sharpening in Photoshop. (Of course if it were
really fair then these printing tests would occur with raw images
instead of jpegs, but whatever.)

The reason I think the foveon image should be upsampled is because
a print with fewer pixels isn't going to look as good regardless of
quality of the pixel information. (A picture printed with only 20
pixels should look twice as good as a picture with only 10 pixels,
even if those 10 are more accurate.) Also, since bayer cameras
apply their interpolation before the image leaves the camera to
acheive their resolution it seems to me that applying interpolation
to a foveon image to acheive the same resolution would be a good
test.

Just my thoughts.

Mario

I printed two of the sd9 samples from Phil's site

All on a Epson 1280 and high gloss paper. I applied a auto level
and sharp 200%-.03-0
to all in Photoshop.

--
Thanks for reading .... JoePhoto

( Do You Ever STOP to THINK --- and FORGET to START Again ??? )

 

[Ron Parr]

There is NO "overlap" with the Foveon X3 approach - as long as you
are relating it to 3.54mps. With Bayer/interpolation ... there is
overlap as soon as you go above 1.5mps.

I think didn't quite get the kind of overlap we're talking about.

There is plenty of overlap in the absorption curves for each photosite. I suggest you read the patent and look at Figure 8.

http://l2.espacenet.com/dips/bnsviewer?CY=ep&LG=en&DB=EPD&PN=US5965875&ID=US+++5965875A1+I+

--
Ron Parr
FAQ: http://www.cs.duke.edu/~parr/photography/faq.html
Gallery: http://www.pbase.com/parr/

 

[Erik Magnuson]

At the moment we can only
discuss "theoritical" possibilities. And again ... we are
SPECIFICALLY discussing the "resolution" possibilities of the X3
technology - and comparing it to Bayer/Mosiac technology.

The only problem is, that like I said, theoretical cameras can only take theoretical pictures.

Therefore .... it remains to be seen if it is true .. but the
POTENTIAL is there for the X3 to be DOUBLE the resolution of
existing 6mps. (And without color morie/artifacting).

The main problem with your theory is that mosiac sensors appear to produce better resolution in the real world than you give them credit for. There is a lot of room for tweaking in demosiacing algorithms (which bring me back to why you have to compare implementations and not theories.)

The main objection I had in this thread was the description of the X3 output as "true color" as if it were capable of measuring the precise amount of red, green, blue light falling on it. It just ain't so.

--
Erik

 

[Ger Bee]

--- I've done a few prints and there seems to be plenty of resizeability going on. Keep in mind that at a 3MP 3X capture at 100% is the same as a CCD capture but delivering 1/3rd – sort of anyway (you know the RGBG neighbouring colour co-ordinates added together to make one colour shade but involving the use of 9 pixels shared 3x3x3 so one ends up with the image data from only 1 in 3 pixels roughly per CCD MP rating actually being recorded).

Ergo one could say that a 6MP CCD is delivering just 2MP of finished data which then goes on to be interpolated further and you can see yourself that the images are fine if not even good.

Now the point is a 3X capture is delivering 3MP of data which then goes on to be interpolated further. From my simple and none scientific quick tests, I’d say this hold true to a high degree.

Another way of looking at is – (loosely) is that a 3MP 3X = a 9MP CCD. However, again I’d say we would find it is more equal to a 6MP camera in practice.

If you get your mind around that loose concept then one can see how the Foveon colour stays truer as one “enlarges” them and should appear sharper as it relies on colour accuracy to deliver “apparent” image sharpness at the same time.

Yet another way to imagine the difference – (loosely) – imagine the CCD capture device, interpolating your image FIRST, in the camera, but with the Foveon 3X you interpolate AFTER on your PC.

But with all that said – one must, as stated, try a few prints fro oneself to see if one is satisfied.

 

[MarvinK]

I don't think the super CCD has much to offer for this discussion (or any discussion of resolution and interpolation). In fact, I think it's more of a marketer's creation than a genuine technical advance.

After all, any CCD array can become a super CCD by simply rotating its grid orientation by 45 degrees in the image plane. What does that really buy? Not a whole lot, IMO, since it takes the same number of measurements in the same area as the non-super-CCD of the same number of effectiove pixels.

MarvinK

Just thought I'd follow up on my follow-up:

A few things came to mind after I wrote this.
...
...
4. I wonder what advantages there would be, if any, to Foveon
adopting a
pixel layout like the one on the Super-CCD??

Of course, it's all guessing until we start seeing reviews from
production cameras, yeah?

-gl

 

[Erik Magnuson]

After all, any CCD array can become a super CCD by simply rotating
its grid orientation by 45 degrees in the image plane.

Well, they would have to license/evade Fuji's patents.....

What does that really buy?

It measureably DOES improve vertical and horizontal resolution. So for all of you who mainly take photos of resolution charts, it's a real technical advance ;-)

And theoretically it also allows

• Improved resolution and S/N ratio
• Higher sensitivity
• Horizontal/vertical skipped readout
• High-speed electronic shutter

Hey, if X3 advocates can argue from marketing slides, why not SuperCCD? What? You say the S2 implementation doesn't show all these advantages? Well, that's obviously not important because theoretically they all exist!

Actually, there is a serious point hidden among the sarcasm. In a way it's a shame that only Fuji is pursuing SuperCCD because that makes it impossible to determine if any limitations are due to the implementation or the concept.

--
Erik

 

[Ger Bee]

It's my understanding that in fact it takes twice as many sensors
to make one image pixel then with the Foveon. On a D100 for
example, you are in effect having a 3MP camera. It then
interpolates up to get the 6MP.

Try this - each pixel records a value of 1/3 as one needs three pixel readings to make a guess at the colour to be assigned to any group of pixels. The actual shade of each pixel is known, as this is the actual reading from the vibrations from each pixel that is assigned to picture capture.

Thus your resolution is not effected and one will still get 6MP from a 6MP chip (loosely) but one carries only 1/3rd (or ¼) of the total colour information – that’s actual information read from as many as nine pixels just to assign the central pixel with a guesstimated colour – it only reads actual vibrations and they are measured on a maximum of 256 individual shades – it knows the actual colour over every pixel and that never changes. The shade in each of the individual Red, Green & Blue filters can run from pure white to pure black and at least three readings are needed to offer any individual pixel its colour.

Thus one can say that one has 6MPs resolution (sharpness) but only 2MPMs of colour data collected and included within the final output (and if you like redistributed over the palette). Thus our CCDs can often give “flat” results but with adequate sharpness.

The Foveon depends on the actual radiation signature penetration depth and vibration to determine its colour on each and every pixel assigned to picture capture. It can do this very accurately so it can use its colours to define borders. The 3X can actually detects both the shade and actual colour for every pixel (the CCD can detect only the shade) so every tiny nuance can be recorded. This afford the image a sharpness derived from its accurate colour definitions – over any given surface distance the 3X will detect a colour change for every pixel whereas the CCD will assign colours based on thee pixels and will thus record shading less accurately and have a “soft” look to them.

For colour printing the effect is like needing 1 pint to fill a square.

From the CCD one is only given 1/3rd of a pint but one has to fill the square anyway.
From the 3x one is given most of the pint to fill the square.

Given that we are actually “happy” with the results from using 1/3rd of a pint to fill our squares from our CCD, if we then spread the 3X’s pint around more we get a result very similar to that of the CCD but because the colour also defines the image we also loose less sharpness – the CCD needs resolution to define sharpness. Think of that one like drawing our square and then washing our colour around INSIDE it – think of the 3x like painting the square with the same brush as we fill with.

 

[G. Lentz]

Actually, I was concerned that the discrete '1x1 matrix' arrangement of the Foveon may produce an unnaturally sharp, even grid-like image. Judging from the diagonal table edge boundary in the upper left quadrant of the new sample image



, this concern appears justified. A pixel array similar to the Super-CCD (which, I believe, consists of interleaved octoganol pixels, not just a rotated grid) may help to at least disquise this effect.

This discussion indeed deals with resolution. Perhaps we should be discussing final sharpness instead? After all, an image that is twice as sharp from the go, independent of DSP, could always be upsampled to a larger size, and as a result of the dithering involved may possibly even be more aesthetically pleasing in its more natural smoothness.

I'm not saying that the Foveon is twice as sharp as a Bayer, but there is a certain 'sharpness' already inherent in the design, by virtue of the fact that you don't have to perform extensive matrix computations to recover (read 'guesstimate') the original value of a given photopoint.

-gl

After all, any CCD array can become a super CCD by simply rotating
its grid orientation by 45 degrees in the image plane. What does
that really buy? Not a whole lot, IMO, since it takes the same
number of measurements in the same area as the non-super-CCD of the
same number of effectiove pixels.

MarvinK

Just thought I'd follow up on my follow-up:

A few things came to mind after I wrote this.
...
...
4. I wonder what advantages there would be, if any, to Foveon
adopting a
pixel layout like the one on the Super-CCD??

Of course, it's all guessing until we start seeing reviews from
production cameras, yeah?

-gl

 

[G. Lentz]

For some reason, I'm not seeing the url that I referenced in the previous post. I'll try again:

"



"

 

[JoePhoto]

There is NO "overlap" with the Foveon X3 approach - as long as you
are relating it to 3.54mps. With Bayer/interpolation ... there is
overlap as soon as you go above 1.5mps.

I think didn't quite get the kind of overlap we're talking about.

There is plenty of overlap in the absorption curves for each
photosite. I suggest you read the patent and look at Figure 8.

But we do not know yet how much that may affect image quality. They may have it "figured out".

http://l2.espacenet.com/dips/bnsviewer?CY=ep&LG=en&DB=EPD&PN=US5965875&ID=US+++5965875A1+I+

--
Ron Parr
FAQ: http://www.cs.duke.edu/~parr/photography/faq.html
Gallery: http://www.pbase.com/parr/

--
Thanks for reading .... JoePhoto

( Do You Ever STOP to THINK --- and FORGET to START Again ??? )