Sigma SD1 review

[joofa]

With a Foveon by shooting for black and white, you throw away it's one advantage, it's co-sited colour, but you get it's noise and ISO performance disadvantages.

Bayer filter CFA could loose upto perhaps 2/3 of light compared to a monochrome sensor. I would assume that Foveon treated as monochrome would also have this light gathering advantage, which should also help with noise in certain situations.

[Graeme Nattress]

Well, we know chroma noise is an issue - but if you just look at luma noise (the Foveon papers read of treating their "green" as luma) do you see an advantage over a Bayer CFA?

Graeme

[Graeme Nattress]

I'm not worried so much about shot noise as read noise.

[joofa]

Sorry, i accidently deleted my message to which you replied. I am copying it here again:

 

Well, we know chroma noise is an issue - but if you just look at luma noise (the Foveon papers read of treating their "green" as luma) do you see an advantage over a Bayer CFA?

Graeme

I think I can see less light collection in green pixels on Bayer CFA treated as luminence vs. monochrome sensor as shown in the image below taken with two exact same sensors except one being Bayer and other mono.
 


Since the green pixels on Bayer are darker compared to mono pixels the shot noise component should be more in Bayer. If Foveon pixels can be treated as mono pixels in the same way, I would think they should have similarly better shot noise component.

Joofa

I'm not worried so much about shot noise as read noise.

Depends upon the situation. Can't be summarily dismissed, IMHO. Furthermore, if Bayer really looses upto 2/3 of light compared to mono, and we can assume Foveon may be operated as mono easily, then under some low light situations the better light gathering of mono should help in reducing the perceptibilty of read noise.

In addition, a Foveon treated as mono might require less lighting to get equivalent brightness as the above image shows. Again lighting situation or budget can't be summarily dismissed, IMO.

Joofa

[Ray]

Ray, so it sounds like you like the look of no OLPF. If colour accuracy is not an issue for you, and you generally shoot at low ISO where noise is less of an issue, then great. Other thing you can do is shoot with a higher resolution Bayer CFA camera and downsample it -  that can induce the same kind of lack-of-OLPF look you like (depending on choice of downsampling filter), could also lead to still lower noise.

Graeme

Graeme,
I certainly do like the look of any crisp and detailed image with lots of micro-contrast, provided color redition is also pleasing. Don't we all?

I'm very much of the persuasion that a camera is a tool for a specific purpose. I prefer to use a cropped-format DSLR when I want the longest reach with my longest telephoto lens, which is the Canon 100-400/5.6 IS zoom, but I also prefer full-frame when I want the widest FoV.

Sometimes I will choose a camera, such as the Nikon D7000, because of its low noise, low weight, high DR, and HD video capability.

If I already owned a Sigma 800/5.6 lens with Sigma mount, I might be excited at the prospect of fitting to it an 'effective' 30mp cropped format body.

However, I don't own such a lens and don't intend buying one, partly because it doesn't have IS. I wouldn't consider buying the Canon 800/4.5 either, because it's not only heavy but also very expensive; more expensive than the Sigma. In fact, the Sigma 800 with SD1 should be less expensive than the Canon 800/5.6 IS with 7D.

I have no problem if you want to describe noise reduction as a type of interpolation. I was speaking from the layperson's understanding of interpolation as the creation of entirely new values (as opposed to adjusted values), as in upsampling an image which results in the creation of totally new pixels which didn't exist before the upsampling.

In concept, I see a difference between the creation of completely new values for the missing two primaries in the Bayer Array, estimated from the values of surrounding pixels, and the noise reduction applied to existing values of all 3 primaries in the Foveon pixel.

But let's not get hung up on verbal definitions. If the two processes come under the same technical umbrella of interpolation, then so be it.

What intersts me more are the advantages of the SD1 as a tool which may be able to produce results, in certain circumstances, that no other camera can.

What do you predict the comparison would be if I were to remove the AA filter from my 15mp Canon 50D and, using the equivalent quality and FL of lens, compare it with the SD1 at base ISO, ensuring equal FoVs of course?

[Sareesh Sudhakaran]

Sudhakaran, the problem with Foveon is is not that it samples RGB at the same spatial point, but that to enable it's sampling of RGB at the same point it uses silicon thickness as a colour filter and that leads to poor and noisy colour. With a Bayer CFA you get good colorimetry and lower noise, but you sacrifice the layered arrangement. With Bayer CFA you can always produce a sharper image and the appearance of spatially co-located colour through downsampling. With Foveon there's very little that can be done about the colour.
Graeme

Just curious, but is it possible for a separate circuit be built to read the array like it was a Bayer array and then use that mathematics to further process the RAW signal (or even add to it for more latitude in post processing) from the Foveon photosites?

[Graeme Nattress]

Sareesh, once you've read the RAW you can interpret it how you want. That won't help as it's how the Foveon does colour filtering that is the issue with it.

Ray, "I certainly do like the look of any crisp and detailed image with lots of micro-contrast, provided color redition is also pleasing. Don't we all?" - not, we don't all! Crisp and detailed - yes, but not so crisp or so detailed that sharp edges look jaggedy or fine detail descends to noise. I like a very naturalistic image - one that is how I see the world where you're not aware you're viewing pixels. With Foveon images I'm always keenly aware that the image is made up of pixels. I'd take soft (yes, soft, not out of focus, but soft) smooth and creamy over edgy any day.

Yup - same umbrella. The problem here is not you or what you think of as interpretation or not, but the Foveon mantra that is often promulgated by it's proponents that "we don't interpolate". Similarly they'll show the depth diagram of the pixel with bright red, green and blue layers, whereas in the case of a Bayer CFA the dye filters really are bright red, green and blue, but in the Foveon the equivalent colours of the "colour filter" are muddy and muted.

Ray - it's a 5 micron pixel, right? So a 3.9 micron pixel on a Bayer CFA would give the same telephoto reach measured luma resolution? Canon 600D is getting closer at 4.3 micron. DSLRs don't usually drop much below 5 micron for image quality (noise) reasons. Sounds like you'd have to do experiments with how sharp your glass is too, and how steady the camera and subject are as at that distance any movement will eat any rez advantage very quickly.

BTW, I think the "effective 30mp" is a rather exaggerated number. I think Foveon sans-OLPF to a Bayer-CFA with OLPF is about 24mp, but lower luma aliasing on the Bayer due to it using an OLPF. Remove OLPF from both and you'll get equivalent measured resolution around 20mp (harder to guess, especially as I like OLPFs and think they're necessary, not optional) on the Bayer to the 15 on the Foveon. The OLPF takes most of the resolution hit, not the mosaic.

Graeme

[Ray]

Ray - it's a 5 micron pixel, right? So a 3.9 micron pixel on a Bayer CFA would give the same telephoto reach measured luma resolution? Canon 600D is getting closer at 4.3 micron. DSLRs don't usually drop much below 5 micron for image quality (noise) reasons. Sounds like you'd have to do experiments with how sharp your glass is too, and how steady the camera and subject are as at that distance any movement will eat any rez advantage very quickly.

BTW, I think the "effective 30mp" is a rather exaggerated number. I think Foveon sans-OLPF to a Bayer-CFA with OLPF is about 24mp, but lower luma aliasing on the Bayer due to it using an OLPF. Remove OLPF from both and you'll get equivalent measured resolution around 20mp (harder to guess, especially as I like OLPFs and think they're necessary, not optional) on the Bayer to the 15 on the Foveon. The OLPF takes most of the resolution hit, not the mosaic.

Graeme

Graeme,
I'm well aware of the benefits of good glass and a steady camera. The Foveon sensor has been around for a long time and a number of comparisons have appeared on the internet which suggest the resolution and detail is close to that from Bayer sensor of double the pixel count, perhaps slightly less.

I've frequently seen comparisons betwee MFDBs and DSLRs of similar pixel count. The MFDB has a slight edge, but nothing as great as double the pixel count.

I've also seen comparisons on MaxMax of 5D images before and after the AA filter has been removed. Again, the differences are subtle, and certainly not equivalent to a camera with double the pixel count, with AA filter.

Please explain.

[Graeme Nattress]

"number of comparisons have appeared on the internet which suggest the resolution and detail is close to that from Bayer sensor of double the pixel count, perhaps slightly less" - the ones I've seen compare a Foveon based camera to a Bayer CFA with OLPF. They're also comparing aliases from the Foveon as if they were real detail which they're not. Unless you use a proper test chart that shows up aliasing, you're not getting a reasonable answer here.

With MF you're changing a whole bunch of variables - you're using a different lens and sensor size which makes things tricky.

On the 5D don't they just remove half the OLPF - the other half being stuck to the sensor, or is that a different camera I'm thinking of?

I find that with my Bayer CFA (+OLPF) camera systems I can measure limiting resolution of about 78% of the horizontal pixel count. That's like saying you get 60% of the MP count as real measured resolution with low aliasing too. If we assume that half of that loss from 100% is OLPF and half the mosaic pattern (not an unreasonable assumption) that gives a just under 80% of the MP count as real measured resolution figure. Or in other words, to get a measured 15mp resolution that would imply needing to start with 19mp.

The problem is when you eyeball this stuff rather than measure it, you (especially looking at a scene rather than a chart) will see aliasing as real resolution. So when you compare a camera with low aliasing with one with high aliasing, you really do need to measure rather than look. Of course, if you like the look of the aliasing then eyeball away as measurements mean little in that case, but personally, I prefer high measured resolution without excessive aliasing rather than potentially higher resolution that is corrupted by aliases.

As I'd mentioned earlier, photography is a wonderful pursuit as it combines the technical and the creative. In this case, measuring resolution and aliasing will tell you what's going on and  will inform you of why you're perceiving the image as it is, but it doesn't dictate your personal imaging aesthetic that obviously differs from mine with regards to aliasing.

Graeme

[Ray]

The problem is when you eyeball this stuff rather than measure it, you (especially looking at a scene rather than a chart) will see aliasing as real resolution. So when you compare a camera with low aliasing with one with high aliasing, you really do need to measure rather than look. Of course, if you like the look of the aliasing then eyeball away as measurements mean little in that case, but personally, I prefer high measured resolution without excessive aliasing rather than potentially higher resolution that is corrupted by aliases.

As I'd mentioned earlier, photography is a wonderful pursuit as it combines the technical and the creative. In this case, measuring resolution and aliasing will tell you what's going on and  will inform you of why you're perceiving the image as it is, but it doesn't dictate your personal imaging aesthetic that obviously differs from mine with regards to aliasing.

Graeme

Graeme,
It sounds to me as though you are ultra-critically tuned to any hint of aliasing when you view an image.

Do you see any aliasing in the image at Dpreview of the old guy snoozing, posted by cinefeel, linked below?

http://forums.dpreview.com/forums/readflat.asp?forum=1027&thread=38932863

[Graeme Nattress]

Sure, the hair and eyebrows look a bit "crunchy" - but that shot is not anywhere near as bad as some Sigma images. I find if I add a minute blur to that shot, I find it looks more "natural" and less plasticky to me. I really don't like what the camera does with fields of grasses in landscape shots where instead of naturally blurring out into the distance, you get a descent into noise.

But now we're getting into the aesthetics of imaging. My tech talk here is not about what we like and don't like, but why the image looks the way it does - which is predominantly the lack of optical low-pass filtering. Where it would be interesting to go from here is to add an OLPF to the Sigma and compare with / without, and also to deconvolve the OLPF and compare that to without. Then you'd get a much better idea of exactly what it is you're seeing.

Graeme

[Dave Millier]

Hi Graeme

I certainly appreciate your explanations and your insights and I also do not buy into some of the rather extreme claims for the Foveon approach. But I also believe that images are best judged by your eyes, not theory when considering aesthetic matters (engineering issues are of course best dealt with more formally). 

As far as the resolution claims for Foveon are concerned, you may be technically correct but I can't help feel that your own preferences tend to lead you to over-emphasise the negative aspects of Foveon.  I use a SD9, DP1 and SD14, not as my main cameras, but mainly for experimentation with alternative processes. And in my opinion, (even ignoring the hype), the images look pretty good, especially in good light. The colour is attractive and can be made acceptably natural and realistic with DNG profiles. The resolution in prints of the 4.6MP sensor seems to me to be the equal of 12MP class Bayer cams (I have a 5D Mk1, 450D for comparison).  As you say, without some form of bandwidth limitation, aliasing must be present, but mostly it doesn't seem to be visible. You see it most often in architectural type subjects and sometimes in grass stems for example that are orientated on a diagonal but Photozoom upsampling methods usually elliminate it.

One of the frustrating things that surrounds the Foveon debate is for people to tend to take a position for or against the tech (Sigma hype partly to blame). I wish people would be more neutral and try to accept that there is something about Foveon images that is attractive and whilst it might never be mainstream, it is a valuable alternative and concentrate on getting the best out of it rather than simply dismissing it (or hyping it)...

Sareesh, once you've read the RAW you can interpret it how you want. That won't help as it's how the Foveon does colour filtering that is the issue with it.

Ray, "I certainly do like the look of any crisp and detailed image with lots of micro-contrast, provided color redition is also pleasing. Don't we all?" - not, we don't all! Crisp and detailed - yes, but not so crisp or so detailed that sharp edges look jaggedy or fine detail descends to noise. I like a very naturalistic image - one that is how I see the world where you're not aware you're viewing pixels. With Foveon images I'm always keenly aware that the image is made up of pixels. I'd take soft (yes, soft, not out of focus, but soft) smooth and creamy over edgy any day.

Yup - same umbrella. The problem here is not you or what you think of as interpretation or not, but the Foveon mantra that is often promulgated by it's proponents that "we don't interpolate". Similarly they'll show the depth diagram of the pixel with bright red, green and blue layers, whereas in the case of a Bayer CFA the dye filters really are bright red, green and blue, but in the Foveon the equivalent colours of the "colour filter" are muddy and muted.

Ray - it's a 5 micron pixel, right? So a 3.9 micron pixel on a Bayer CFA would give the same telephoto reach measured luma resolution? Canon 600D is getting closer at 4.3 micron. DSLRs don't usually drop much below 5 micron for image quality (noise) reasons. Sounds like you'd have to do experiments with how sharp your glass is too, and how steady the camera and subject are as at that distance any movement will eat any rez advantage very quickly.

BTW, I think the "effective 30mp" is a rather exaggerated number. I think Foveon sans-OLPF to a Bayer-CFA with OLPF is about 24mp, but lower luma aliasing on the Bayer due to it using an OLPF. Remove OLPF from both and you'll get equivalent measured resolution around 20mp (harder to guess, especially as I like OLPFs and think they're necessary, not optional) on the Bayer to the 15 on the Foveon. The OLPF takes most of the resolution hit, not the mosaic.

Graeme

[Graeme Nattress]

Dave, hard to be neutral when it comes to aesthetics! On the numbers side, I'm trying my best to give the numbers that I've measured and they're repeatable and consistent with other measurements of similar cameras.  I wish people would shoot better charts on the Foveon and then I'd have better numbers to go on there - so instead I just give it a 100% luma resolution factor (usually reckon Bayer CFA with OLPF around 78%).

The "for or against" is indeed a polarized reaction, and unfortunate. Without it we'd be able to have much more interesting discussions on the colorimetry and resolution and dynamic range aspects.

Aliasing is most easily visible when you get moire - then it stands out like a sore thumb. When it's present in an image that doesn't have a repetitive line structure it manifests in more subtle forms - crunchy edges, stair-step edges, diminishing detail descending into noise etc., but these are harder effects to observe and are most easily confused with real resolution. I'm sensitive to such effects, and doubly so because I deal with motion imagery where the temporal aspect can make aliasing appear more obvious.

As you mention, stair-stepping can be dealt with by up-samplers that work on an edge-adaptive basis, but similarly the per-pixel softness of a OLPF using camera can be dealt with by de-convolution or downsampling.

What I'm keen for photographers to understand is what measurable aspects of photography correlate with what they see visually. I think that's useful and powerful for them to know so that they can understand why they like what they like and how they can use that to their visual advantage.

Graeme

[hjulenissen]

Graeme,
I certainly do like the look of any crisp and detailed image with lots of micro-contrast, provided color redition is also pleasing. Don't we all?
...

I dont know what micro-contrast means.

I like crisp images with _real_ detail.

For many organic non-man-made objects, I can live with lots of aliasing, and it can add to the perceived crispness.

For many man-made (or highly structured natural objects), aliasing looks really bad to me, and I know of no bullet-proof limited effort post-processing procedure to fix it.

-h

[Ray]

I dont know what micro-contrast means.

Have you tried searching the internet for a definition and explanation. 

Following is a good explanation relating to the concept of lens contrast, from a tutorial on this site at:

http://www.luminous-landscape.com/tutorials/understanding-series/lens-contrast.shtml

When we talk about lens contrast, we're not talking about that quality. What we're talking about is the ability of the lens to differentiate between smaller and smaller details of more and more nearly similar tonal value. This is also referred to as "microcontrast." The better contrast a lens has (and this has nothing to do with the light­dark range or distribution of tones in the final print or slide) means its ability to take two small areas of slightly different luminance and distinguish the boundary of one from the other.

The article also explains how subtle differences in color in adjacent pixels can add or create microcontrast, which may not be apparent if the image were B&W.

Now, I don't pretend to be as technically savvy as many others who contribute to this site, and what I'm about to write may simply be incorrect, or a load of bunkum.

However, when I see an image of a Bayer Array pattern where every pixel is a particular intensity of either red, or blue or green, and it is explained that the demosaicing algorithm has to totally reconstruct two additional values for every single pixel in the RAW image, ie,  create completely new values of red and blue for every green pixel, and completely new values of green and red for every blue pixel, and completely new values of green and blue for every red pixel, I'm quite amazed that the final result appears as good as it does, because two thirds of all values comprising the image are total inventions.

Of course, it's understood that such invented, new values are based upon the values of surrounding and adjacent pixels. They are not plucked out of thin air, I know. 

I can't help wondering what would happen if one were to photograph a test chart containing just a Bayer Array, ie. millions of red, green and blue squares, setting up the camera so that every red, green and blue square on the test chart were aligned with every red, green and blue pixel on the camera's sensor.

Whether one is 'for' or 'against' the Foveon concept, it seems clear to me there are two major factors contributing to the enhanced effect of detail (or crispness, or crunchiness, however you want to put it).

One factor is its lack of an OLPF, a charcteristic which the SD1 shares with most MFDB backs and a charcteristic which I imagine 'effectively' enhances the 'microcontrast' properties of whatever lens is attached to the camera.

The other factor is the difference in color interpolation methods. In the case of the Foveon, we are at least starting from real values which are spatially accurate even if some intensity is lost or degraded by the absorption qualities of the silicon.

I would guess, (and again I admit I'm not technically knowledeable on the processes), that here we have another source of enhanced microcontrast.

Adding these two factors, or sources of enhanced microcontrast, we get an effect which is significantly greater than we would expect from the omission of just the OLPF.

[joofa]

However, when I see an image of a Bayer Array pattern ... I'm quite amazed that the final result appears as good as it does, because two thirds of all values comprising the image are total inventions.

The image signal is changing slowly compared to the RGB grid in Bayer CFA; so a lot of it is not invention as the missing values of a certain color are either (more or less) the same as the next pixel of the same color or "shared" by pixels of other colors that are closer. When this condition does not hold as the signal changes more rapidly aliasing happens.

I can't help wondering what would happen if one were to photograph a test chart containing just a Bayer Array, ie. millions of red, green and blue squares, setting up the camera so that every red, green and blue square on the test chart were aligned with every red, green and blue pixel on the camera's sensor.

As noted above Bayer CFA are not designed for this situation as this is not a real world scene. A real world scene has some redundancy (equivalently stated above as varying slowly compared to the RGB grid in Bayer CFA) so that the Bayer scheme suffices.

In the case of the Foveon, we are at least starting from real values which are spatially accurate even if some intensity is lost or degraded by the absorption qualities of the silicon.

In theory Foveon color can be made pretty accurate with the use of certain specialized filtering techniques. If Sigma colors are not accurate enough, then, may be they are not using such techniques, but I'm intrigued and interested in knowing why is that. May be there is a reason of practical implementation versus a theoretical approach.

Joofa

[Ray]

The image signal is changing slowly compared to the RGB grid in Bayer CFA; so a lot of it is not invention as the missing values of a certain color are either (more or less) the same as the next pixel of the same color or "shared" by pixels of other colors that are closer. When this condition does not hold as the signal changes more rapidly aliasing happens.

This statement is not at all clear to me. Do you mean the camera is moving?
If the camera is moving during the exposure, then I understand the values of red, green and blue as recorded will be different, compared to a perfectly stationary camera, but that doesn't change the situation with regard to the fact that 2/3rd of the total number of values comprising the image have to be invented.

I understand, or at least I'm receptive to the idea, that there are degrees of mathematical probability involved in the demosaicing process.

For example, it may be calculated by the demosaicing algorithm used, that there's a 65% probability that a particular value of blue, in relation to a particular green pixel, is accurate, as calculated from surrounding pixels.

My point would be, in defense of the Foveon process, that there's a 98% chance that the blue pixel is accurate, in the same circumstances, because the blue wavelength is received  at the top of the sensel, and maybe an 80% chance, or 70% chance that the red pixel is accurate after appropriate interpolation is made (if that's the right word) after taking silicon absorption charcateristics into consideration.
 

In theory Foveon color can be made pretty accurate with the use of certain specialized filtering techniques. If Sigma colors are not accurate enough, then, may be they are not using such techniques, but I'm intrigued and interested in knowing why is that. May be there is a reason of practical implementation versus a theoretical approach.

Agreed! Except, of course, it needs to be stated that the Foveon colors are 100% accurate spatially, if not in intensity. (But I sometimes wonder if I'm speaking gobblegook  ).

[hjulenissen]

This statement is not at all clear to me. Do you mean the camera is moving?

I guess he is saying that the light falling onto two neighbor sensels tends to be "similar", just like the weather today tends to be similar to the weather yesterday. How often do you photograph "bayer-like patterns of abrupt on-off transitions perfectly aligned to the pixel grid"?

There are many reasons why they could be similar: camera movement, scene movement, intrinsic properties of light-reflecting objects, diffraction, lense-flaws, etc. The point is that the Bayer sensor seems to still be a good compromise between the virtues that we humans can sense, the way in which light tends to hit a sensor in typical applications, and the process-technology that has been viable to use for the last decades.

-h

[hjulenissen]

Have you tried searching the internet for a definition and explanation.  

Yes, but it seems to be a constructed word with no technical or practical use. Your link made no more sense than other sources I have seen. We might as well succumb to senseless audiophile b.s. like "ethearility", "3-dimensionality", "space", etc. Mumbo-jumbo words invented due to a lack of understanding of how technical aspects affect the subjective impact (which is difficult).

I much prefer writers admiting that they have no idea _why_ a given lense/camera helps them make appealing images, but that it does.

-h

[Dave Millier]

Agreed.

For many years the concept of resolution seems to have been accepted (the ability to separate two closely spaced points) and so has acutance (the rapidity of the transition between two adjacent tones). Together these terms seem sufficient to describe the subjective impression of detail.

What exactly micro-contrast could mean, I have no real idea. To me it sounds like a synomyn for "fine detail" or "high resolution" or perhaps "acutance". In Foveon land they have also invented the term "fast edge roll-off" also unexplained but sounding remarkably like acutance. 

As you note the hifi mags polluted the audio community with a bunch terms that meant nothing but which through continuous repetition became familiar and sounded as though it was conveying something. I have no way of telling whether my subjective impression of "silky highs" and "taut bass" and "air" corresponds with anyone else's (including the author's) but we can sure convince ourselves we are talking about the same thing. Even though my sensitivity to the silkiness of the highs seems to be inordinately influenced by what I ate for lunch.  Micro contrast, 3D, verisimiltude seem to be doing a similar job. They tell us nothing meaningful but the allow people with similar brand loyalties to forge a connection that seems emotionally meaningful. It's particularly helpful at seeing off outsiders because if they don't understand the terms it just proves they lack the particular sensitivity that is required to join the elite club... 

 

Yes, but it seems to be a constructed word with no technical or practical use. Your link made no more sense than other sources I have seen. We might as well succumb to senseless audiophile b.s. like "ethearility", "3-dimensionality", "space", etc. Mumbo-jumbo words invented due to a lack of understanding of how technical aspects affect the subjective impact (which is difficult).

I much prefer writers admiting that they have no idea _why_ a given lense/camera helps them make appealing images, but that it does.

-h