I think he meant cones. The rods can be ignored for purposes of photography.
-
Welcome to the new forums! Please read this first. For known issues we are working to resolve, click here.
The Top Layer
- Thread starter D Cox
- Start date
I don't think it matters much. The point is that the top layer on its own cannot distinguish one colour from another. To do this, you have to compare one layer with another, or one layer with the sum of the two others. No one layer gives colours.
But the top layer differs from the other two in being more evenly sensitive through the spectrum, and much less noisy.
Once you get colour differences in various parts of the image, and they are in order along the spectrum, you can transform these differences into RGB for display on a monitor. It doesn't matter that the spectrum sensitivity curves don't match those of human cones (although those of a Foveon are more like those of cones than are those of a Bayer mosaic) because the transforms deal with that.
I wonder if there is less noise reduction going on by default with the Quattro cameras vs. the Merrill cameras, and that's why we see some noise in some of the photos. As they get better at designing the processes for producing images from the raw data, Sigma just might get the Quattro down to a level where it is not only producing an image with more detail and better color than the Merrill, but less noise too. That will be interesting to see. Already I think I've witnessed a major improvement in the images with regard to micro-contrast. I don't remember seeing noise in a lot of the early photos though . . . just some of them. I look forward to seeing the results from the Quattro DSLR.
Roland
I hope you don't mind me joining in here, but how the monochromatic image is painted in on the Quattro is what interests me.
The brain pays more attention to edges and patterns than it does to the detail of smooth transitions. Could it be that the Quattro algorithm looks at colour and luminosity gradients across pixels that neighbour edges and tries to reassign the spectrum of the 5Mpix data on either side of the 20Mpix resolved edges?
My thought is that might explain some of the differences between the M and Q, or maybe not...
Andrew
I hope you don't mind me joining in here, but how the monochromatic image is painted in on the Quattro is what interests me.
The brain pays more attention to edges and patterns than it does to the detail of smooth transitions. Could it be that the Quattro algorithm looks at colour and luminosity gradients across pixels that neighbour edges and tries to reassign the spectrum of the 5Mpix data on either side of the 20Mpix resolved edges?
My thought is that might explain some of the differences between the M and Q, or maybe not...
Andrew
Roland Karlsson
Forum Pro
I have seen three theories how Quattro might work.
- Expand the lower layers to 20 Mpixel and use ordinary old fashioned Foveon algorithms. Personally I did not think this would work. You would get color artefacts. I have tested, and I got color artefacts.
- Bin the upper layer to 5 Mpixel. Use an ordinary Foveon old fashioned algorithm, then use the upper layer to add extra details. This was my first thought, and I saw that Arvo thought so to.
- The third one - the one I described here. The reason I think this is possible is the Merrill sensor. It is very noisy in the two lower layers. Still the resulting color image that SPP makes is surprisingly clean. So, I think the upper layer has to play a huge role here - somehow.
But, as a starting point I assume a solid RGB image with three good sharp channels is probably best. It is much more robust against editing. Maybe you want to increase saturation - then ouch if color sharpness is low.
Roland
Apologies for my poor use of language.
My proposal would give good colour sharpness in the sense of having sharp transitions from one area of colour to another. It would give good colour accuracy through the larger effective pixel size used by the colour algorithm.
It would only fall down in three areas, maybe:
1) For small complex shapes, where the foreground happens to give a similar top layer response to the background, even though the colours are very different. The algorithm would struggle to find the point of colour transition.
2) For wispy objects with visible structure but no defined edges or gradient clues to the colour transitions.
3) For areas of colour with subtle variations that interact with the 5Mpix scale to give artefacts at small scale steps in the top layer response at the 20Mpux scale - a sort of fine grain effect like noise but actually an artefact of the distribution of colour information by the algorithm.
Probably barking up the wrong tree...
Andrew
Apologies for my poor use of language.
My proposal would give good colour sharpness in the sense of having sharp transitions from one area of colour to another. It would give good colour accuracy through the larger effective pixel size used by the colour algorithm.
It would only fall down in three areas, maybe:
1) For small complex shapes, where the foreground happens to give a similar top layer response to the background, even though the colours are very different. The algorithm would struggle to find the point of colour transition.
2) For wispy objects with visible structure but no defined edges or gradient clues to the colour transitions.
3) For areas of colour with subtle variations that interact with the 5Mpix scale to give artefacts at small scale steps in the top layer response at the 20Mpux scale - a sort of fine grain effect like noise but actually an artefact of the distribution of colour information by the algorithm.
Probably barking up the wrong tree...
Andrew
It seems pretty straight forward to me. They make the image the same way they would with a Merrill, but with the exact same green and red values in four adjacent locations, based on the information from the green and red layers of the sensor. Each and every pixel will be slightly different, based on the differences from one blue "pixel" to the next having an affect on those "pixels" made of green and red. This would actually require less processing, because the computer or camera should only have to process 1/4 as much the information from each of the green and red layers. In fact, we're talking about roughly 1/3 as much information compared to each of those two layers of a Merrill sensor, but with just 1/3 more information in the top layer. That means that overall, there would be less processing required . . . or maybe about the same amount of processing, compared to what is necessary for a Merrill camera.
As far as knowing the algorithm - why do we need to know that? Can't we just use our own algorithm? Or couldn't we use a modification of the old one that just doubles the "pixels" of the green and red layers in the x and y direction?
Maybe I'm just too simple-minded to understand. I've been told that I "just don't understand" before. (Somehow it turns out that most of the time I actually DO understand, and people are either over-complicating things in their own minds, or they don't realize the complexities involved with something I understand. I really don't believe all this stuff is rocket science though. Of course, I could be wrong. It's happened before.)
Roland, maybe you can help me understand the complexities that make it a really difficult thing to make an image from a raw file. Here is how I imagine the data in a raw file to look, after it has been deciphered:
(This is of course a small sample of a much larger quantity of data, and Layer 1 would have twice as many rows and columns as layers 2 and 3 . . . in a Quattro raw file. And of course, there would be more information included - meta-data and the embedded jpeg, and maybe some other stuff.)
Layer 1
10110010010110 10110110010110 10110001010110 10110110010010 10110011010100
10110010010100 10110010010110 10110010010110 10110010010110 10110010010101
10110010010110 10010110010110 10101010010110 00110010010110 11010110010110
10110010010110 10111010010110 10100110110100 10110010010110 10110010010110
10110010010110 10110010010110 10110110010101 00110010010110 11010010010110
Layer 2
10110010010110 10110110010110 10110001010110 10110110010010 10110011010100
10110010010100 10110010010110 10110010010110 10110010010110 10110010010101
10110010010110 10010110010110 10101010010110 00110010010110 11010110010110
10110010010110 10111010010110 10100110110100 10110010010110 10110010010110
10110010010110 10110010010110 10110110010101 00110010010110 11010010010110
Layer 3
10110010010110 10110110010110 10110001010110 10110110010010 10110011010100
10110010010100 10110010010110 10110010010110 10110010010110 10110010010101
10110010010110 10010110010110 10101010010110 00110010010110 11010110010110
10110010010110 10111010010110 10100110110100 10110010010110 10110010010110
10110010010110 10110010010110 10110110010101 00110010010110 11010010010110
Is this about right?
(This is of course a small sample of a much larger quantity of data, and Layer 1 would have twice as many rows and columns as layers 2 and 3 . . . in a Quattro raw file. And of course, there would be more information included - meta-data and the embedded jpeg, and maybe some other stuff.)
Layer 1
10110010010110 10110110010110 10110001010110 10110110010010 10110011010100
10110010010100 10110010010110 10110010010110 10110010010110 10110010010101
10110010010110 10010110010110 10101010010110 00110010010110 11010110010110
10110010010110 10111010010110 10100110110100 10110010010110 10110010010110
10110010010110 10110010010110 10110110010101 00110010010110 11010010010110
Layer 2
10110010010110 10110110010110 10110001010110 10110110010010 10110011010100
10110010010100 10110010010110 10110010010110 10110010010110 10110010010101
10110010010110 10010110010110 10101010010110 00110010010110 11010110010110
10110010010110 10111010010110 10100110110100 10110010010110 10110010010110
10110010010110 10110010010110 10110110010101 00110010010110 11010010010110
Layer 3
10110010010110 10110110010110 10110001010110 10110110010010 10110011010100
10110010010100 10110010010110 10110010010110 10110010010110 10110010010101
10110010010110 10010110010110 10101010010110 00110010010110 11010110010110
10110010010110 10111010010110 10100110110100 10110010010110 10110010010110
10110010010110 10110010010110 10110110010101 00110010010110 11010010010110
Is this about right?
richard stone
Senior Member
That is an interesting observation, Yes, blotches are mostly in the mid or mid-low tones.
Personally I think the blotches are created by the difference between the middle and the lower layer. And I guess that the dark tones are mostly taken from the top layer. That makes sense if the blotches are not found in the dark tones.
That is probably the scary truth.We're only saved by SPP - the best de-blotcher on the planet.
victorgv
Senior Member
- Messages
- 2,295
- Solutions
- 1
- Reaction score
- 1,569
iso 1600
That is an interesting observation, Yes, blotches are mostly in the mid or mid-low tones.
Personally I think the blotches are created by the difference between the middle and the lower layer. And I guess that the dark tones are mostly taken from the top layer. That makes sense if the blotches are not found in the dark tones.
That is probably the scary truth.We're only saved by SPP - the best de-blotcher on the planet
Roland Karlsson
Forum Pro
That is also an interesting observation.
So - blotches normally are found on certain areas.
Mid tones - and normally grayish.
Hmmmm ... I have seen it on a reddish brown surface also.
So ... I am not sure.
But - it seems like it is very hue/darkness dependent. Looking at an image, the blotches are not found everywhare, rather in certain areas.
Last edited:
B
Buhl213
Guest
In other words, where the differences between the 3 layers are minimal, and minor differences from e.g. microlenses/incident angles, play a larger part?
There are blotches all over the CC, most evident and strongest is the patch on the left hand panel rightmost green patch. Other blotches are seen on patch rims on almost every patch on both panels. But that is obviously OK, if you're after B&W prints. Having said that I'll say, my, new to me, very much used DP2M will be here on Sunday, and I can do my own blotch tests, or just take some photos around.
Once, I lit a letter-size white card (in portrait orientation) from about a foot above with small light bulb - so as to use the 'inverse square law' to get a gradient of illuminance down the paper. With the right exposure, the onset and fade-out of the infamous blotching is quite evident.
Kendall Helmstetter Gelner
Forum Pro
This is a different way of describing a theory I've been trying to get across to people for some time now, perhaps your way they can understand better why it might work...
I don't know that it really works that way, just that it's a technique that could easily account for most fine detail being resolved in the Quattro, most especially things like fine detail against a blue sky.
The way I had described the "reassignment" was that every adjoining pixel had the potential to give you information for reducing the averaging caused by the lower 5mp layer combining four pixels worth of data. So for instance if one neighboring 4x4 quad was all blue, and a single top sensor value on the next quad over had the same value as the top layer values on the blue quad - you could subtract the lower layer values from the data under the "blue" quad from the quad with the single pixel, and only be averaging three potential pixels worth of data, not four. That would increase color accuracy no matter how you then calculated the remaining three pixels.
richard stone
Senior Member
I thought about this for a while and yes, of course, it is ISO 1600. And yet still, there is the blotching, right in a textured area in the lower to mid-tone. And we could very likely expect the same blotching in an ISO 400 image of the scene.
I do not see obvious blotching in the image on the scale presented in the solid colors. Could blotching somehow, in some circumstance, be induced, perhaps by insufficient lighting on the solid colors in the image? Possibly, but I do not see it as an obvious issue in the image as presented.
Blotching is not some random occurrence. We seem to be looking at noise and frequency and the intersection and interaction of the noise from the various sensor levels and (if what I am seeing is part of it) from image detail resolution (a luminance function).
Richard
That is an interesting observation, Yes, blotches are mostly in the mid or mid-low tones.
Personally I think the blotches are created by the difference between the middle and the lower layer. And I guess that the dark tones are mostly taken from the top layer. That makes sense if the blotches are not found in the dark tones.
That is probably the scary truth.We're only saved by SPP - the best de-blotcher on the planet
Kendall Helmstetter Gelner
Forum Pro
I like the idea of thinking of it as a "voting" process, since I imagined it looking at all neighboring pixels to determine how it thought it should break apart the lower layers...
This has been described as Bayer-like in concept, but it's not because it would only use the data present in the quad - it would just use other quads to determine the distribution of the lower level data among the four output pixels. A Bayer system would use the actual data from spatially removed sensors.
About how to look for that processing difference between an M and a Q - I'm not really sure how you could tell if this process were in use. The only reason i think it might be is because you can see single pixel level detail from a Q image, and overall a Q image pretty much always looks a little sharper than a Merrill image so you know they have to have some way to distinguish more than 15MP worth of data, so some system that makes use of the data the top layer records must be in use - and the system we are thinking of is the only way I can think about the processing that would seem to have that effect.
I wish I had more time period...
--
---> Kendall
http://www.flickr.com/photos/kigiphoto/
http://www.pbase.com/kgelner
http://www.pbase.com/sigmadslr/user_home
Last edited:
Roland Karlsson
Forum Pro
Time? Yeah ... that is one of the more sparse resources it seems.
Nice that some do try to think
Here is another try at thinking.
Lets say that you first duplicate the top layer to all three layers. Then you will get a monochrome image of high quality, that in the case of Merrill will be a saturated red after conversion.
Then you use some secret and clever algorithm to modify the middle and low layers, without adding noise, so that they look similar to the actual middle and low layers. And, and here comes the big problem, without adding halos and without adding desaturation of small details.
What do you say?
Similar threads
Keyboard shortcuts
- f
- Forum