Aliasing
- Thread starter Barry_Green
- Start date
Barry_Green
Moderator
It is possible. CommanderSpike's already done a bit of a homebrew version.
Problem is, if you eliminate all the aliasing, you're going to be left with only the true resolved detail that these cameras can provide, and that's not much more more than standard-def. It is the presence of aliasing that makes these cameras look sharp at all.
These cameras are trying to use a still-camera sensor to do high-speed motion photography. It's beyond what they were ever designed to do. So the pixel-binning (or line skipping, whichever/whomever you believe) is necessary to get the frame rate up to video speeds. There are compromises involved. If you believe the "six-pixel bin" approach as stated by Canon's Tim Smith, each binned pixel is made up of reading a combination of six sensor pixels. But because it's a Bayer pattern sensor, thos six pixels cannot be side by side -- it's not like a 3x2 block! It's six pixels extracted out of a matrix of at least 24 (in red and blue), which means there are gaps between the pixels. Which is why spurious aliased detail "gets through". So if you put on an OLPF that's strong enough to blur those little details out, something that blurs the entire six-pixel matrix into one big pixel, you'll be left with imagery that's way way way lower than HD. But it'd be accurate.
Which is, of course, why these cameras only resolve a medium-def image -- because, in terms of true resolving power, they can only resolve detail as small as that six-pixel bin. But if we examine that six-pixel bin as being out of a 24-pixel matrix, then that really means the OLPF would drop the resolution from the stated 18-megapixels, to 1/24 of that (which would mean a 750,000-pixel frame).
And, guess what -- let's do the math, an SD frame has about 350,000 pixels in it, and a 720p frame has about 920,000 pixels in it. 750,000 fits right between there, doesn't it? Which is a reaffirmation of what I've been saying -- the true resolved detail on these systems is somewhere between standard-def and 720p. At best.
Any additional "sharpness" being detected is aliasing that's leaking through and contaminating the image. (and yes, I'm well aware that people like that contamination and think it adds to the image).
Problem is, if you eliminate all the aliasing, you're going to be left with only the true resolved detail that these cameras can provide, and that's not much more more than standard-def. It is the presence of aliasing that makes these cameras look sharp at all.
These cameras are trying to use a still-camera sensor to do high-speed motion photography. It's beyond what they were ever designed to do. So the pixel-binning (or line skipping, whichever/whomever you believe) is necessary to get the frame rate up to video speeds. There are compromises involved. If you believe the "six-pixel bin" approach as stated by Canon's Tim Smith, each binned pixel is made up of reading a combination of six sensor pixels. But because it's a Bayer pattern sensor, thos six pixels cannot be side by side -- it's not like a 3x2 block! It's six pixels extracted out of a matrix of at least 24 (in red and blue), which means there are gaps between the pixels. Which is why spurious aliased detail "gets through". So if you put on an OLPF that's strong enough to blur those little details out, something that blurs the entire six-pixel matrix into one big pixel, you'll be left with imagery that's way way way lower than HD. But it'd be accurate.
Which is, of course, why these cameras only resolve a medium-def image -- because, in terms of true resolving power, they can only resolve detail as small as that six-pixel bin. But if we examine that six-pixel bin as being out of a 24-pixel matrix, then that really means the OLPF would drop the resolution from the stated 18-megapixels, to 1/24 of that (which would mean a 750,000-pixel frame).
And, guess what -- let's do the math, an SD frame has about 350,000 pixels in it, and a 720p frame has about 920,000 pixels in it. 750,000 fits right between there, doesn't it? Which is a reaffirmation of what I've been saying -- the true resolved detail on these systems is somewhere between standard-def and 720p. At best.
Any additional "sharpness" being detected is aliasing that's leaking through and contaminating the image. (and yes, I'm well aware that people like that contamination and think it adds to the image).
I think the bendy propeller looks AMAZING. i know its inaccurate.... but God its beautiful!
Postmaster
Dark Side of the Camera
Actually I don`t get it.
Why do they use those Bayer sensors? (yeah they are cheaper) :violin:
Why not 3 traditional chips in full size/full HD matrix and do away with all that debayer, rolling shutter, Jello, moire Voodoo Shmoodoo?
In my book they should have phenomenal low light sensitivity and great DOF.
To complicated? To expensive?
Frank
Why do they use those Bayer sensors? (yeah they are cheaper) :violin:
Why not 3 traditional chips in full size/full HD matrix and do away with all that debayer, rolling shutter, Jello, moire Voodoo Shmoodoo?
In my book they should have phenomenal low light sensitivity and great DOF.
To complicated? To expensive?
Frank
Al MacLeod
Well-known member
Manufacturers have long known that to soak the market for all it's worth, you have to appeal
to every niche you can. If you accommodate everybody with one camera, you have wasted a lot of market potential.
I read once that color tv availability was delayed for a time because the manufacturers wanted to fully saturate the market for black and white sets. Waste not want not!
to every niche you can. If you accommodate everybody with one camera, you have wasted a lot of market potential.
I read once that color tv availability was delayed for a time because the manufacturers wanted to fully saturate the market for black and white sets. Waste not want not!
adkimery
Veteran
Well, like Barry said these are still cameras first and foremost w/bolted on video functionality because many still photojournalists are now being asked to shoot video for the web. The fact that video people are lusting for them and trying to use them for more than what the cameras are designed for is the fault of the video people, not the camera makers.
W/regards to 3xCCD vs 1xCMOS, AFAIK, 1xCMOS doesn't require as much space, doesn't require as much power, doesn't produce as much heat, and is overall a less expensive technology to go with. I think you also run into lens compatibility problems between camera systems that use 3xCCDs and camera systems that use 1xCMOS.
-A
Barry_Green
Moderator
Because sensors are monochrome. The only way to get color is to use three chips and a prism, or stick a bayer cfa in front of the chip.
Using three chips will get you away from debayering, but you'll still be subject to rolling shutter, jello, and moire.
Rolling shutter is integral to today's CMOS video chips, and it happens whether you have one chip or three. And jello is one of the effects of rolling shutter.
Moire is one of the most objectionable ways that aliasing manifests itself, but it's not tied inherently to Bayer filters, it's endemic to an undersampling video system.
The reason they use one chip in these SLRs is because of the simulation of the way a film camera works. They can use the same lenses, with the same flange focal distance, with one chip. That's not possible with three chips because you have to build an optical prism assembly to split the beam of light across three chips, and that takes up space, which blows the relationship of the lens-to-chip distance, meaning those lenses won't focus properly. It's more important to them to maintain the lens relationship, than it is to avoid Bayer de-mosaic'ing.
Barry great article, thanks. Too bad my 5d is such a compromise! Couldn't one bypass the compression with some software config and get the entire uncompressed sensor readout via hdmi? I was reading about some thought on that from cinema5d. Anyway I enjoyed learning and reading whY you said. I guess now I can see and agree with why the BBC won't use the 5d
Lenilenapi
Carbonite Member
Barry, what camera did you use for the DSLR res chart images? Some of these cameras have more aliasing than others don't they?
Also a friend has been pointing out artifacting in out of focus areas of Canon 5D footage, that looks kind of like soft rectangular patches . Do you know what that is caused by?
Lenny Levy
Also a friend has been pointing out artifacting in out of focus areas of Canon 5D footage, that looks kind of like soft rectangular patches . Do you know what that is caused by?
Lenny Levy
Last edited:
Barry_Green
Moderator
I haven't tried, but I would imagine the SD downconversion should be pretty good. The EX1 makes cleaner HD than DSLRs do, so downconverted EX1 footage would look at least as good, and probably better, than downconverted DSLR footage.
Barry_Green
Moderator
Those were from my 7D.
From what I've seen, they all seem pretty comparable in terms of total amount; maybe the GH1 is a little less than the Canons but it's not a big difference. What is a huge difference is the way the Canons turn to colored mush, the GH1 doesn't do that. The Pentax K-7 does something about halfway between, it's not as colorfully mushy as the Canons but it's not as monochrome as the GH1.
Not off the top of my head but I would expect some type of fixed-pattern noise? Does it happen in underexposed out-of-focus patches?
Barry_Green
Moderator
I don't have an LP, only circular...
Just out of curiousity, what would happen if a Sigma Foveon X3 sensor were used? It definitely has a sensor that works in a completely different way and doesnt use a bayer pattern.The results for stills from the Sigma cameras are often quite beautiful as well. Could additionally get rid of a lot of artifacts:
http://en.wikipedia.org/wiki/Foveon_X3_sensor
http://en.wikipedia.org/wiki/Foveon_X3_sensor
Barry_Green
Moderator
The Foveon is effectively a three-chip system, it's just that all three chips are sandwiched into one physical place instead of being in three separate spots and having light directed at them with a prism. With all three chips cosited like that, it acts to the lens as a one-chip system.
Foveon inspires loyalty and hatred like no other technology I've seen. Graeme Nattress doesn't seem too big of a fan, with his main gripe seeming to be that silicon is a lousy optical filter, making the longer wavelengths noisier to gain up.
But inherently, using a Foveon could turn this whole situation upside down, depending on how they read the chip out, and what native res they put on the chip. If they read it as three separate chips, reading each color separately, then you could do proper pixel binning without skipping gaps, and therefore probably get much cleaner performance. But if they read the chip as one chip, maybe not so possible.
Foveon inspires loyalty and hatred like no other technology I've seen. Graeme Nattress doesn't seem too big of a fan, with his main gripe seeming to be that silicon is a lousy optical filter, making the longer wavelengths noisier to gain up.
But inherently, using a Foveon could turn this whole situation upside down, depending on how they read the chip out, and what native res they put on the chip. If they read it as three separate chips, reading each color separately, then you could do proper pixel binning without skipping gaps, and therefore probably get much cleaner performance. But if they read the chip as one chip, maybe not so possible.
strangways
Well-known member
Foveon
Foveon
As the owner of a Foveon-sensored camera which I both love and hate, I can say for certain that it does have its weaknesses, such as plenty of crosstalk between colors, but it does have a few advantages that could help avoid aliasing.
First, it does not really require an OLPF, and the first Sigma SLRs to use the sensor omitted them.
Second, the sensor itself can do pixel binning and thereby increase the readout speed. Foveon calls this "Variable Pixel Size"
Current Bayer HDSLRs need to do pixel binning on another chip, and have to do a full readout of the sensor first. Because they cannot do that fast enough, they resort to line skipping instead.
The problem is that VPS does not provide very high resolution. The Fx17-78-F13D sensor used in the current Sigma cameras can only do a full un-binned sensor readout of 2652x1768 pixels at 5 fps, and to get it up to 30 fps, you have to bin it to 640x480.
It can also do 700×525 at 20fps, which leads me to believe 25 (and 24 fps) would fall somewhere around 610x457, which is less than standard def.
That would, however, be a true resolution, in that it would not have aliasing, and no blur from an OLPF, so it would be actually resolving the same as a 3-CCD or 3-MOS video camera with the same pixel resolution (as long as that video camera was not using pixel shifting, and was recording 4:4:4 color.)
Foveon
As the owner of a Foveon-sensored camera which I both love and hate, I can say for certain that it does have its weaknesses, such as plenty of crosstalk between colors, but it does have a few advantages that could help avoid aliasing.
First, it does not really require an OLPF, and the first Sigma SLRs to use the sensor omitted them.
Second, the sensor itself can do pixel binning and thereby increase the readout speed. Foveon calls this "Variable Pixel Size"
Current Bayer HDSLRs need to do pixel binning on another chip, and have to do a full readout of the sensor first. Because they cannot do that fast enough, they resort to line skipping instead.
The problem is that VPS does not provide very high resolution. The Fx17-78-F13D sensor used in the current Sigma cameras can only do a full un-binned sensor readout of 2652x1768 pixels at 5 fps, and to get it up to 30 fps, you have to bin it to 640x480.
It can also do 700×525 at 20fps, which leads me to believe 25 (and 24 fps) would fall somewhere around 610x457, which is less than standard def.
That would, however, be a true resolution, in that it would not have aliasing, and no blur from an OLPF, so it would be actually resolving the same as a 3-CCD or 3-MOS video camera with the same pixel resolution (as long as that video camera was not using pixel shifting, and was recording 4:4:4 color.)
Last edited:
sumitagarwal
Member
Barry, thanks for trying to tackle this in a logical and mathematic fashion, it is a huge asset to the community. I've been following your advice since I got the DVX100P years ago (recently sold to my company as I bought the 7D!.
Your numbers put the 7D at about 1155 x 650 when in 1080P mode.
I was going to try to do an equivalent set of calculations for other Canon cameras, but then realized I couldn't fully wrap my head around how you arrived at the 24 photosites/pixel, so I couldn't appropriately determine what that number would be for the other cameras.
Would you mind walking us through a numerical assumption on how the 1DM4 would perform and what the 5DM2 and the T1i should be like?
My guess is that the lower the native resolution of the sensor, the better the resulting binned image because fewer pixels need to be binned and thus you are not spacing out each data point as far. I would expect the 7D to perform better than the 5DM2, but then would expect the 1DM4 and T1i to perform better than the 7D.
To save you a couple minutes (should you agree to help) I've calculated the 16:9 image areas for each:
7D: 5184 x 2916 = 15,116,544
5DM2: 5616 x 3159 = 17,740,944
1DM4: 4896 x 2754 = 13,483,584
T1i: 4752 x 2673 = 12,702,096
How about a hypothetical camera with interchangeable OLPFs, so you could choose the one appropriate for the job? Or even, and this is way out, interchangeable sensors? Keep all the computational gubbins in the camera body, and swap sensors as needed? So you could have a FF 24Mp sensor for stills, a DX 16 (Say) Mp one for sports / wildlife and a DX / APS-C / s35 3+ Mp one optimised for cine?
Everyone wins, the marketing chaps can sell you alternative sensors, or upgraded bodies and you (The shooter) only need to buy the sensor you need.
Bet Canikon don't see it that way though.
Dave
Everyone wins, the marketing chaps can sell you alternative sensors, or upgraded bodies and you (The shooter) only need to buy the sensor you need.
Bet Canikon don't see it that way though.
Dave