Sigma SD1 review

[Dave Millier]

One thing I noticed is that when attempting to equate pixel counts between the two technologies (in the P645 section), Michael appears to write that you have to account for both the demosaicing losses and the x2 factor often attributed to Foveon. Is this not double counting? If you call the Foveon sensor 30MP you have already adjusted for demosaicing losses. I can't see any way in which the Foveon sensor is equivalent to 40MP as the article implies!

[Graeme Nattress]

On resolution: So let's look at the Foveon SD1 RAW at 4704 x 3136, and for now, linear luma resolution to keep things simple. For a Bayer CFA to achieve a measured horizontal resolution of 4704, and assuming it uses a properly configured OLPF, it would need to have a sensor with 6030 photosites across (assume a 78% photosite to measured resolution factor which is not unreasonable).

Moving that back to MP, that equates to about a 24MP Bayer CFA sensor. Although such a sensor may be susceptible to mild chroma moire, it will be much less prone to luma aliasing than the un-filtered Foveon.

Of course, if we properly filter the Foveon so that it has similarly low levels of luma aliasing as the above 24MP Bayer CFA, it probably measures around 4000 x 2666 in terms of measured resolution (very OLPF dependent and hard to guess...) To get a similar measured resolution with similarly low luma aliasing we'd need 5128x3418 on Bayer CFA, around 18MP.

So what it comes down to is weighing up the factors....

1) does Chroma moire bother you in your shots, and if it appears, is it hard to remove?
2) does luma moire bother you in your shots. Luma moire is particularly difficult to deal with (unlike chroma moire which is relatively easy in most cases)
3) is colour accuracy important
4) is noise important, or in other words, how important is dynamic range to you, or looking another way, high ISO performance, as all these factors are inter-linked.

If lack of Chroma moire is the most important factor to you, then a Foveon sensor makes some kind of sense. If luma moire is important to you, go Bayer CFA with an OLPF. If such moire doesn't bother you, try a Bayer CFA without OLPF and see how you get on. If colour or noise are important, go Bayer CFA.

Graeme

[bjanes]

I think that the biggest problem with this "review" was that Mr Reichmann is not competent to write reviews invoving the digital imaging technology. That is not his strength. There were numerous errors and inaccuracies in the article - the one you quoted here is a great one (I had to read it twice to make sure he really said what he said and then I had a good long old fashioned  laugh rolling down the floor  ) - yet it was posted.

It is difficult to assess the reliability of statements made by a newbie with only 5 previous posts and with no CV (curriculum vitae) listing his qualifications or prior publications. I would suggest that Michael is eminently qualified to make a subjective review of the camera. He is a very experienced photographer and has worked with all types of cameras and has extensive contacts in the photo industry whose advice he can seek. His observations are likely valid, but his technical explanations for what he sees may or may not be accurate. I will leave that to the experts with advanced degrees and experience in digital imaging.

DXO is going to evaluate this camera some time in July and we will have some technical data with which to work and they may add their own comments. For example, their comparison of the Nikon D5000 and the Canon EOS 500D demonstrates how relatively high "color blindness" requiring large matrix coefficients can degrade image quality. Does this apply to the Sigma?

Regards,

Bill

[Bart_van_der_Wolf]

It will be interesting to see some test charts from the SD1, but it may be the case that the SD1 works better on "real subjects" than test charts.

Hi Erik,

A good, and meaningful, test(card) shot is as telling as "real subjects". But then there would no longer be discussions about mythical properties attributed to an overpriced camera/sensor ...

Cheers,
Bart

[Graeme Nattress]

Test cards are "real world" subjects, just rather boring subjects. The whole idea of test charts is to provide a repeatable scene to a camera so that we can ensure what we see is a result of the camera, not an unintended change in scene.

Aspects like dynamic range are notoriously hard to judge from a typical scene and oh-so-much more accurate with a good chart, for instance. Resolution is also a factor that is not easy to measure on a typical scene, but a well designed test chart can really help on. If you're concerned about aliasing and moire (which I think you should be) you need to know how susceptible your camera system is to such artifacts and how they effect the image and what can be done about it - and you need to know this before you go out and shoot your scene, not after when you're looking at the image in Photoshop and see the issues. For colorimetry, you need to know what colours are an issue, and how different light sources interact with your camera system. The spectral responses of different light sources will behave differently with camera systems, and good charts with measured light sources can really help you know in advance what will work best, and where problems may lie.

Charts are certainly important in camera development, but I think a good photographer needs to understand their camera via charts in such a way that they can maximize performance in the field, and minimize the chance of an anomalous or sub-standard result occurring.

Measurements don't tell you all you need to know about image quality, but what they do tell you is very useful and usually correlates very well with subjective opinion.

Graeme

[ErikKaffehr]

Hi bart,

I expect that the resolution test in DPReview will show contrast inversion, but I'm not sure that it would be visible in common subjects.


Otherwise I agree with your view.

Best regards
Erik

Hi Erik,

A good, and meaningful, test(card) shot is as telling as "real subjects". But then there would no longer be discussions about mythical properties attributed to an overpriced camera/sensor ...

Cheers,
Bart

[Graeme Nattress]

Luma aliasing shows up on a variety of common subjects - clothing detail and distant brickwork being typical examples. If you're just shooting trees and landscapes it's not so much of an issue, although I and others find it adds an un-natural "crunchiness" to the edges of most in-focus objects and fine detail - your opinion might be that you like that.

Graeme

[joofa]

Resolution is also a factor that is not easy to measure on a typical scene, but a well designed test chart can really help on.

Test charts have some desirable properties such as controlled settings and repeatability, as you mention. However, it is a fair question to guage the "resolution" of real images or that of typical scenes, which opens up a whole new world of statistics compared to traditionally used methods on test charts. Please see the following link and the messages in that thread to guage the utility of such methods that work on real scenes:  

http://forums.dpreview.com/forums/read.asp?forum=1018&message=38170122

Joofa

[feppe]

It is difficult to assess the reliability of statements made by a newbie with only 5 previous posts and with no CV (curriculum vitae) listing his qualifications or prior publications.

Well, at least we have a picture of Aku Ankka here, and he seems rather authoritative to me.

"Aku Ankka" is the Finnish translation of "Donald Duck." No, I'm not making this up.

[michael]

One thing I noticed is that when attempting to equate pixel counts between the two technologies (in the P645 section), Michael appears to write that you have to account for both the demosaicing losses and the x2 factor often attributed to Foveon. Is this not double counting? If you call the Foveon sensor 30MP you have already adjusted for demosaicing losses. I can't see any way in which the Foveon sensor is equivalent to 40MP as the article implies!

Huh? I never wrote that the SD1 was equivalent to 40MP. Please reread what I did write.

It's not double counting to reduce one and increase the other, for the reasons given, if the point is to try and find an equivalence.

Michael

[deejjjaaaa]

I'd be curious to see how the Bayer cameras would do against the SD1 with a floating point raw converter, like RPP.  The amount of detail one gets out of RPP compared to ACR is pretty stunning.

does "stunning" have any numbers or it is another "6 stops of DR" ?

[kwalsh]

Resolution is also a factor that is not easy to measure on a typical scene, but a well designed test chart can really help on.

I am an engineer who loves measuring things, but I'll play some devil's advocate here.

I think your conflating "test chart" and "consistency" inadvertently.  Yes, consistent scenes are vital for comparisons.  That doesn't mean that a test chart that produces numerical metrics is necessarily very useful.  Rarely are the arbitrary numerical metrics derived of much use.  DxOMark is a classic example.  The full SNR curves they produce are actually pretty useful for comparing sensors.  Their scores, and even their DR and SNR curves much less so, often to the point of being misleading.  A consistently shot studio scene with real objects in a high contrast setting is far more useful to a photographer when it comes to comparing cameras.

I cherry picked the quote above as a classic example.  Most resolution tests are completely senseless as they test a portion of the MTF curve that is least relevant to actual photography and ignore the lower frequency portions of the MTF curve that play a much larger role in image formation.  Hence the loads of lens "tests" that rate superzooms with very high resolution scores when in fact in practice a real world image at a real world print size from the superzoom will appear far less detailed than a prime with similar "resolution" but a much more robust low frequency MTF.

Test charts, with the right context, can be useful to an engineer as validation.  For a photographer they tend to be confusing at best and misleading at worst.  Shoot consistent real world scenes, not test charts, if your audience is photographers.

Again, playing devil's advocate here, agree with most of what you say.

Ken

[Graeme Nattress]

Ken, I don't really disagree with you too much. If you've got the location to set up a permanent test scene, I think that's a superb "test chart" that can offer valuable data, and present it in a form that's very understandable and comparable across cameras. Most of us don't have that luxury, but a measured light source and some charts is pretty much as good as we can get.

As for resolution, there are indeed people who will ascribe a single figure to resolution and indeed, that's practically meaningless. It's still worse when that figure calculates out to a greater resolution than there are photosites on the sensor! Giving a full MTF plot is indeed useful, and yes, looking at what the low frequencies are doing is very important, especially with lenses.

For resolution though, it's hard to get real-world scenes that allow for ease of measurement and clearly show aliasing. I think when we look at a camera in a critical way, it's very important to actually go out and shoot with the camera system, taking the image all the way through post production. However, good charts are invaluable tools, when well used, to quantify what you're seeing. Indeed, good charts can still be confusing, but then it's our job to educate.

Graeme

[Quentin]

Just a personal view, but there is nothing more uninteresting (to me, as a photograher) than another debate about Bayer / Foveon resolution equivalence etc.  The only thing that matters to me is what the files look like on screen or in print.  I like the look of Foveon files a lot (maybe I'm that crunchy brickwork guy!) but not enough to sink silly money in to a camera probably worth $2k or so at most.  I'd love to see what a 40mp or more Foveon sensor could do in a decent MF body, but I guess I will be even older and wrinklier than I am now before I find out (if I ever do).

[Graeme Nattress]

Quentin, I appreciate such a debate can be rather boring :-) I wish it wasn't the topic of conversation as I think with Foveon that there are more important issues like the colorimetry to discuss. However I do think it was good of Michael to get some Bayer CFA cameras that don't use an OLPF because (in terms of luma resolution) it's the OLPF (or lack of it) that people are seeing on Foveon cameras, not an inherent Foveon v CFA advantage in resolution. Yes, there is a luma resolution advantage to not having to demosaic, but it's not as great a detriment to the resolution as the OLPF is. I think we all wish we didn't need an OLPF - they're expensive to make, and the reduce good resolution as well as bad, but as long as we're not using really crap lenses or are running right into diffraction limits at usable apertures, aliasing will be a problem and I think the majority of camera manufacturers agree that it's better to have a lower amount of aliasing artifacts than to have absolutely maximal resolution. Especially for me, the dangers of aliasing are a lot worse with motion than for stills. With a still you can "fix" a single frame, but for motion, it gets expensive, and motion shows up aliasing much more clearly than stills as the aliases move in the opposite direction to the motion of the objects that are aliasing. On the chroma side, as pointed out above, you're trading inherent lack of chroma moire for necessarily strong chroma NR and poor colour accuracy. I think photographers need to know what the tradeoff is there, and what it means to them and their photographs.

The real aspect not really addressed in the review, and one that is, or at least should be, of great interest to photographers, is dynamic range, and that's an area where the Foveon could have issues. I'd be very keen to see how it performs there.

Michael did a great job discussing the workflow and ergonomic issues, and those are the aspects we can't measure on charts - they need a competent human.

Graeme

[michael]

Graeme,

I avoided the black hole of dynamic range, simply because without a Stauffer Wedge or similar it's almost impossible to measure, and visual judgement of this parameter is notoriously unreliable.

Michael

[Graeme Nattress]

A black hole it is indeed!

Graeme

[Ray]

I thought it had been fairly well established over the years that each Foveon pixel, consisting of 3 stacked sensels, is approximately equivalent to two Bayer type sensels, or slightly less, in terms of apparent resolved detail and sharpness.

However, until more extensive comparisons are made with this new model, we won't know for certain. It could be the case, as the pixel count of the Bayer type sensors has increased (and will contunue to increase) and their AA filters have become increasingly weaker, the resolution advantage of that lack of an AA filter in front of the Foveon sensor is accordingly reduced slightly.

The issue of pricing is interesting. Whilst I sympathise with the frequently expressed outrage at the high price of the SD1 compared to other models of cameras that appear to offer more features, I see lots of products on the market that are outrageously priced in my view, and I therefore don't buy them.

It could be the case that Sigma are not particularly worried if you don't buy the SD1 because their manufacturing is currently geared for only a small output. Maybe the recent earthquake and tsunami in Japan has affected this pricing decision.

Let's suppose, for example, that Sigma is able to produce only 500 units a month, as a result of disruptions in Japan.

Ask yourselves which is better from Sigma's point of view, to sell 500 units a month at $2,000 each, resulting in a shortage of supply, long back-orders and angry customers, or to sell 300 units a month at $7,000 each, resulting in plenty of stock in all the retail outlets and warehouses, and a gradual accummulation of unsold stock which can be later sold at a substantial discount.

I notice from rereading Michael's article, that the SD1 has another unique feature, apart from its being the highest resolving cropped format camera on the market. It has a user-removable IR filter allowing the camera to be used for infrared photography.

[ErikKaffehr]

Hi,

I may agree on resolution figures being less than relevant, but I have not seen any example of a really well corrected lens that did not have high resolution.

I'd say that much of the contradiction regarding resolution and contrast is coming from the fourties, before antireflex coatings were perfected. A lens which corrected most aberrations well had many surfaces and therefore a lot of flare. We are now in much better control of flare but the perception of contradiction of "resolution" and contrast are still with us.

A well constructed lens today has both high MTF at all frequencies and high resolution.

Best regards
Erik

Ken, I don't really disagree with you too much. If you've got the location to set up a permanent test scene, I think that's a superb "test chart" that can offer valuable data, and present it in a form that's very understandable and comparable across cameras. Most of us don't have that luxury, but a measured light source and some charts is pretty much as good as we can get.

As for resolution, there are indeed people who will ascribe a single figure to resolution and indeed, that's practically meaningless. It's still worse when that figure calculates out to a greater resolution than there are photosites on the sensor! Giving a full MTF plot is indeed useful, and yes, looking at what the low frequencies are doing is very important, especially with lenses.

For resolution though, it's hard to get real-world scenes that allow for ease of measurement and clearly show aliasing. I think when we look at a camera in a critical way, it's very important to actually go out and shoot with the camera system, taking the image all the way through post production. However, good charts are invaluable tools, when well used, to quantify what you're seeing. Indeed, good charts can still be confusing, but then it's our job to educate.

Graeme

[Dave Millier]

Michael

I can only try and interpret what you wrote here:

"If we assume (and I did going into this test) that a Foveon sensor actually has comparatively about double its linear resolution (so about 30MP for the SD1) and a Bayer equipped camera such as the 645D has about 1/3rd less effective resolution than its actual pixel count (again about 30MP for the 645D), this could actually be a competitive IQ horse race."

The Sd1 has an actual 15MP sensor, the 645 a 40MP sensor. We know that Foveon and Bayer MP are not worth the same, and to come to an agree equivalence, we can make some assumptions (as you do). We can either:

- assume the Foveon is worth more than is nominal 15MP (say 2x or 30MP)

or

- assume Bayer is worth less than its nominal pixel count (say 2/3 or 30MP)

If we do the first, we are describing 30MP equivalence vs 40MP. If we do the second, we are describing 15MP vs 30MP equivalence.

The two methods don't arrive at 30MP equivalence for both and certainly shouldn't both be employed to come to a figure that suggests that the two cameras are competitive!

What you should do is say that demosacing loses about 30% of resolution and AA filtering (for those who have it) about another 20% so a camera like the A900 is only resolving something equivalent to about a 12mp Foveon.  The 645 without an AA filter probably resolves about the same as a 30MP Foveon would. The SD1 with 15MP Foveon pixels has only half the effective pixel count of the 645. Hardly a "competitive IQ horse race"....

Huh? I never wrote that the SD1 was equivalent to 40MP. Please reread what I did write.

It's not double counting to reduce one and increase the other, for the reasons given, if the point is to try and find an equivalence.

Michael