Thread Rating:
  • 0 Vote(s) - 0 Average
  • 1
  • 2
  • 3
  • 4
  • 5
Sigma lost it completely.
#41
You seem to have rewritten your old post totally, Russel.



And a lot of misconceptions too, in that new post.



You seem to not be aware that, besides chromatic information, each pixel also contains luminance information. And it is only this luminance information which delivers detail.

So, of course you are right in saying that only 25% of pixels actually record red light, And that only 25% of the pixels record blue light. And that 50% of the pixels record green light.



But ALL pixels record luminance, and after good de-mosaicking and interpolating, you get much more detail than you are suggesting.

The difference between Bayer CFA and Foveon/Sigma is very low, a bit more than 6% of resolution.

Of course, with only pure red (so no gradations of lighter red) with black patterns, or pure blue with black patterns, resolution can be less, But that is not really a significant aspect, just a theoretical one.

Do take note that the moment the red, green or blue subject has gradations lighter than pure red/green/blue, the other colored sites will start to supply luminance information, even if the subject still looks red.



Your assumption that colour temperature of light has any influence in what we are discussing is strange. 6500K light is NOT red, NOT blue and NOT green. So, all sensels will all deliver luminance information, same as with "white" light.



About high ISO noise, Foveon has its very own noise problem: Its sensor does not record all of the light. Where with bayer CFA, at all pixels all of the light of either the red, the green or the blue part of the spectrum get recorded,

Foveon sensors actually do not record red, green and blue. It is more like blue, yellow and red.

Of the blue, all blue gets recorded. However.... also part of the yellow and red gets recorded.

This means two things:

- The sensel is filled up not just with blue, but also probably with some yellow and some red, depending on the actual subject. It reaches filled status sooner than if it were only to record blue light.

- It also means that one has to do interpolation to determine how much of the recorded energy actually stems from the blue light spectrum.



Below the "blue" layer lays the yellow layer. This layer not only records yellow light, but also some of the red spectrum.

This means 4 things:

- The yellow light reaching the yellow layer is about 1/2 of the yellow light reaching the sensor, the other half has been captured in the blue layer.

- One has to use interpolation to determine how much of the recorded energy stems from actual yellow, and what is from the red spectrum.

- One has to interpolate to figure out what actually has been "green", as the recorded spectrum of light is yellow.

- What has been recorded in the yellow layer is still being polluted with red, meaning the sensel can fill up sooner than if it only would be recording yellow.



Below the yellow layer lays the red layer.

- The red layer only receives about 1/3rd of the red light that reached the sensor. The rest of the red is polluting the blue and yellow layer.



Where bayer sensors actually record 100% of the green part of the spectrum at the green sensels, with Foveon only about 50% of yellow gets recorded in the "green" layer. This is then also still polluted with red.

To get green values, one does not only have to interpolate the red away. but also interpolate against the blue channel, to determine green.



With red, only a small amount of the red light gets captured. Because this means a very noisy red channel, the red channel actually gets blurred as noise reduction measurement.



All in all, Foveon sensors have their very own noise problems, stemming from the loss of light through the different layers of silicon. And Foveon sensors have their own areas of "colour blindness".



As I have noted before, the difference between Bayer CFA sensors and Sigma/Foveon sensors in resolution recording is small, about 6-7% for normal scenes, if we only look at the CFA.

However, this is quite hard to "measure", as Foveon/Sigma do not supply AA-filters with the sensors.

Without AA-filters, the results always will look more detailed (even though, most of the detail is false detail from pixels just having edges.. and so, making edges. But AA-filters have little to do with CFA's.... they are about trying to not record the false detail.



So.. if one really wants to visually compare the sensors, one either has to test a SIgma/Foveon next to a Bayer-CFA sensor without AA-filter, or one needs to put an AA-filter suiting the sensor's pixel pitch in front of the Foveon sensor, so its false detail does not pollute the results.



What I was putting forward is that the perceived difference in detail has much more to do with the missing AA-filter, than with the way the colours are recorded. And that is actually just a fact.

That the foveon sensor can still have an advantage in unusual situations (like a black pattern on a pure primary colour) is obvious, and that it has its own disadvantages in certain situations is also known.



I have nothing against a foveon sensor, in fact, I would not mind at all if my camera had one (that would behave in higher ISO shooting too). As long as it would have a suitable AA-filter, as I am not impressed by false detail, aliasing, moire and that "oversharpened" look.
#42
BC: You seem to have rewritten your old post totally, Russel. And a lot of misconceptions too, in that new post.



Sure Brightcolours, what do I know? After all, my only experience is writing Bayer filter array demosaicer software from scratch.



BC: You seem to not be aware that, besides chromatic information, each pixel also contains luminance information.



You figure I somehow missed that as I was writing the software? Yeah, each pixel also contains luminance information. The relative contributions of red, green and blue to luminance is about red 30%, green 59%, and blue 11%. Perhaps that's why I mentioned in my post that noise goes up a lot when you don't have much info from the relatively sweet green pixels. The contribution to luminance of, for example, blue information, isn't much.



BC: And it is only this luminance information which delivers detail.



Not sure what you're getting at, if there's alternating patches of hue that all have the same luminance, there'd be *no* detail?



BC:...But ALL pixels record luminance



Not a heck of a lot of perceived luminance information in the blue pixels, you seem to be glossing over "details" such as the fact that though, say, blue pixels contribute *something* to luminance, it's not a lot.



BC: ...Do take note that the moment the red, green or blue subject has gradations lighter than pure red/green/blue, the other colored sites will start to supply luminance information, even if the subject still looks red.



Have already covered this topic in my post, pointing out that as subjects *tend* towards missing more and more colors, performance *tends* downward. I.e. sure, Bayer sensors are better with some subjects. Heck they're great with gray subjects.



BC: Your assumption that colour temperature of light has any influence in what we are discussing is strange. 6500K light is NOT red, NOT blue and NOT green. So, all sensels will all deliver luminance information, same as with "white" light.



Your assumption that colour temperature of light has no influence on what we are discussing is strange to someone who's written demosaicing software. If the nonetheless bright ambient light is missing, especially, lots of blue, you have to amplify the noise-prone blue pixels lots more to get your final color-balanced image.



By the way, don't have to be a software author to observe this. Haven't you noticed the greater noise in incandescent-lit photos than greenish-fluorescent-lit or daylight-lit photos at same exposure value? I have.



BC: About high ISO noise...Where bayer sensors actually record 100% of the green part of the spectrum at the green sensels, with Foveon only about 50% of yellow gets recorded in the "green" layer...All in all, Foveon sensors have their very own noise problems, stemming from the loss of light through the different layers of silicon.



Nobody has said that any given Foveon sensor must have lower noise than any given Bayer sensor. What we can say is that while both Foveon and Bayer sensors probably change noise level with subject color, Foveon sensors don't ALSO change resolution with subject color. Which Bayer sensors do. And the Foveon edge is way more than 6% *in the* highly colored areas.



BC: As I have noted before, the difference between Bayer CFA sensors and Sigma/Foveon sensors in resolution recording is small, about 6-7% for normal scenes, if we only look at the CFA.



Foveon sensors are way better than 6% better in maintaining resolution in colored parts of scenes. I suppose you can add up how many parts of normal scenes are highly colored, and get the resolution edge down to 6% for the whole image, but you are glossing over the cases where the highly colored parts of the image might be of above average visual interest.



BC: However, this is quite hard to "measure", as Foveon/Sigma do not supply AA-filters with the sensors.



Since they have *less* need of one for most pictures, this is not cheating, it's a Foveon advantage.



BC: Without AA-filters, the results always will look more detailed



We agree.



BC: So.. if one really wants to visually compare the sensors, one...



Really what one should do is refrain from criticizing the SD1 sensor until one of us has had a chance to compare images of colored objects side-by-side with a 15 megapixel Bayer sensor of the same sensor size, and see which one if any looks better. If the Bayer camera has an AA filter it should be left in place as any normal user would do. My own impression of sample photos thus far is that the SD1 sensor output quality is superb, you seem to be talking an awful lot of theory and not talking much about sample photos.



BC: What I was putting forward is that the perceived difference in detail has much more to do with the missing AA-filter, than with the way the colours are recorded. And that is actually just a fact.



Saying it's a fact that a Bayer camera can get practically the same detail appearance/resolution from its blue pixels than from its possibly-missing-in-action green ones, doesn't make it so. Will point out for example the relative raggedness of the reddish sample photo shown a few posts back.



BC: [would like Foveon]...As long as it would have a suitable AA-filter, as I am not impressed by false detail, aliasing, moire and that "oversharpened" look.



If on the average, Foveon images look better to 51% or more of viewers than Bayer sensors of the same resolution and post-processing, they're just better. Whether you like the reasons for people's preferences or not. There is no law making camera manufacturers put on AA filters. If AA filters are needed for the Foveon, or are not needed for Bayer cameras, all will come out in image samples. IF AA-filterless Foveon cameras only "seem better" to most people for whatever reasons you may not like, then they're better.



BC it's just pretty outrageous for you to write theoretical SD1 criticisms before you or anyone on the Forum's had a chance to thoroughly check out side-by-side images at same ISO and sensor size. Save all your stuff until after you've looked at comparisons and seen with your own eyes that the Foveon's only "6% better".
#43
I am a software developer too, do I care much about your "i have written blahblah"? No, it does not impress much. That you copied an example demosaicking algorithm for your very own hobby converter, great. get over it, though.

As you might be aware, some RAW convertors and not as good as others in recovering detail. Fill in your own conclusion.



I already am clear on that you love the false detail aliased results. That you love the oversharpened look, the stair stepping of every edge.



That you think that is "better", is your problem/preference. But it is not more actual detail.

Like I said, until you have a Bayer CFA sensor without AA-filter, or a Sigma sensor with AA-filter, you have to be careful with comparisons, as you are unable to filter the false detail visually.



And there are ways to compare those: The Leica M9 has no AA-filter. There are a few files to be found online for Cameras, like a 5D, that have their AA-filter removed. There are some MF backs without AA-filter.



It is stupid of you to say that I am "criticizing" the SD1, that that is "outrageous". I am not attacking any camera, least of all the SD1. I am just pointing out that that what people like YOU seem to love (the false detail, the exaggerated defined edges) mostly is from the missing AA-filter, and has less to do with the lack of CFA. No idea how that can be seen as ""criticizing" the SD1 "outrageously". That is not about SD1, that is about Sigma not putting in AA-filters on their sensors (when they actually should, as any signal processing person will tell you).



That you like that is fine. That I don't, should be fine too. Some people love to turn up the bass and the volume, to be impressed when listening to music. I am not one of those either.



fact: Sigma/Foveon results APPEAR more detailed, but actaully just show false detail, like an oversharpened photo from a bayer CFA photo will show too.

Fact: Sigma/Foveon sensors would benefit from AA-filters, to prevent the creating of edges where in the projected image there are not such defined edges on pixel borders (aliasing), and to prevent moire in areas of high frequency detail.



I am fine with people being impressed with AA-filterless results. They should be fine with someone pointing out that the extra detail they are perceiving mostly is false detail due to the square, hard edged nature of pixels.
#44
[quote name='Brightcolours' timestamp='1306251143' post='8790'] The difference between Bayer CFA and Foveon/Sigma is very low, a bit more than 6% of resolution.[/quote]

Bold claim, do you have any calculations to back you up?



[quote name='Brightcolours' timestamp='1306251143' post='8790'] Of course, with only pure red (so no gradations of lighter red) with black patterns, or pure blue with black patterns, resolution can be less, But that is not really a significant aspect, just a theoretical one.[/quote]

You haven't tried photographing any clothing, have you? Because if you had, you wouldn't dismiss it so easily as "theoretical". The thing is that any garment with the colour on the darker side of red or blue suffers from tremendous drop in detail, and they're not so uncommon. For example, shooting something navy already feels as if you just upped your iso by few steps.



The example below *is* in focus, however detail is just poor. Of course there's no such an issue if you shoot gray scarf instead.



[Image: sample4.jpg]



Another thing I find really annoying is the ability of camera to pick up colour moiré everywhere. Yes, low-pass filter suppose to fix that. No, it doesn't work in lots of situations. Try shooting some fine fabric under a bright light and you'll see what I mean. And the problem is that post-processing routines you would use to correct that quite often do lead to a loss of detail.



Keeping all of that in mind I was more than willing to give the new sensor a try. Obviously with the current pricing it's not going to happen and it's a pity.
#45
[quote name='Lomskij' timestamp='1306311475' post='8798']

Bold claim, do you have any calculations to back you up?





You haven't tried photographing any clothing, have you? Because if you had, you wouldn't dismiss it so easily as "theoretical". The thing is that any garment with the colour on the darker side of red or blue suffers from tremendous drop in detail, and they're not so uncommon. For example, shooting something navy already feels as if you just upped your iso by few steps.



The example below *is* in focus, however detail is just poor. Of course there's no such an issue if you shoot gray scarf instead.



[Image: sample4.jpg]



Another thing I find really annoying is the ability of camera to pick up colour moiré everywhere. Yes, low-pass filter suppose to fix that. No, it doesn't work in lots of situations. Try shooting some fine fabric under a bright light and you'll see what I mean. And the problem is that post-processing routines you would use to correct that quite often do lead to a loss of detail.



Keeping all of that in mind I was more than willing to give the new sensor a try. Obviously with the current pricing it's not going to happen and it's a pity.

[/quote]

That is a nice example, that red knitwear. However, to fully show a difference, you do need to shoot a similar gray example, and convert both to grey scale in order to bypass our own eye's trouble with seeing red in detail.

Without any other sample, it is difficult to quantify anyway.



Also, the difference in detail depends a LOT on the RAW converter you use. (That is a plus for Foveon/Sigma: the RAW converter does not make such a big difference in detail. But a difference in colour, though).



My 6-7% in about normal scenes, not red knitwear. Of course there are situations thinkable which show a bigger difference (like your red knitwear). Just like there are situations thinkable where Sigma/Foveon struggles (it is "colour blind" for certain specific hues, for instance).



I have not calculated the 6-7% myself, I have read articles studying the difference, or rather, the influence Bayer CFA has on luminance information.

I can not just produce articles I read a while ago, but a quick google found this interesting page:

http://www.xs4all.nl/~bvdwolf/main/foto/...er_cfa.htm

A host of interesting articles can be found on the subject of demosaicking, luminance and chrominance information:

http://citeseer.ist.psu.edu/viewdoc/summ....1.14.4073

http://www.site.uottawa.ca/~edubois/CFA/



The moire issue.... my camera never has shown any moire, either luminance only or with chrominance problems (at least, I have not come across it myself). That is very camera specific: if a camera shows colour moire, its AA-filter is not optimal.

Moire areas already have lost the detail, as they have overwritten it with their own pretend-detail. So getting rid of the moire always will mean the loss of the "fake" detail, whether there are colour artifacts in the detail or if it is only luminance moire.



I too would like to try a foveon style sensor, but with AA-filter and, like you, for a more "normal" price. And a Canon lens mount <img src='http://forum.photozone.de/public/style_emoticons/<#EMO_DIR#>/wink.gif' class='bbc_emoticon' alt='Wink' />.



There are also other CFA's which do better jobs than the Bayer CFA. It will take a lot for a manufacturer to switch to alternatives, though, concerning processing software for instance.
#46
Final price is 6.9K. Still outrageous.



http://www.bhphotovideo.com/c/product/73...amera.html
#47
[quote name='Yakim' timestamp='1306317465' post='8802']

Final price is 6.9K. Still outrageous.



http://www.bhphotovideo.com/c/product/73...amera.html

[/quote]

So... we should view it as a competitor to a Canon EOS 1Ds Mk IV and Nikon D4x?
#48
Hasselblad's four-shot mode to create "full colour" image seems like a viable alternative to Foveon. I'm just curious if something like that could be implemented into an SLR which already sports sensor-shift based image stabilization (patent issues aside <img src='http://forum.photozone.de/public/style_emoticons/<#EMO_DIR#>/dry.gif' class='bbc_emoticon' alt='<_<' /> ). If so, some Sony A850 would beat crap out of SD1 both resolution and price wise...
#49
[quote name='Brightcolours' timestamp='1306318102' post='8803']

So... we should view it as a competitor to a Canon EOS 1Ds Mk IV and Nikon D4x?[/quote]



No. We should view it as a competitor to the Pentax 645D. After all, Sigma said so themselves.... <img src='http://forum.photozone.de/public/style_emoticons/<#EMO_DIR#>/smile.gif' class='bbc_emoticon' alt='Smile' />
  


Forum Jump:


Users browsing this thread:
1 Guest(s)