Wally Pfister talks about RED

[Daniel Browning]

If I'm not mistaken, Nyquist is half the sampling frequency, so this would be about 46 cycles/mm for Red's sensor?

Twice that, actually. The sampling frequency is one line per 0.185mm, or 185 lines per mm (4096 / 22.12mm). Typically, cycles are measured in line pairs, and taking the normal assumption of 2 lines per linepair, at 4K 16:9 you get a Nyquist of 93 lp/mm.

So in theory would the red camera be resolving 67% mtf at 46 cycles/mm?

67% at 93 lp/mm, yes.

Is that with or without taking the low pass filter into account?

Without.

An unfiltered image sensor, even with 100% fill factor, will have lots of spurious resolution above Nyquist -- high frequency detail that *should* be invisible to the sensor, but isn't, and so contaminates the lower frequencies (i.e. aliasing).

When you add in the OLPF, the contrast is greatly reduced at those higher frequencies (though not eliminated altogether). The side effect is that contrast is also reduced at low frequencies -- ones that did not need any filtering, but they have to get it anyway because of the nature of physics. Fortunately, in many cases this loss of contrast (e.g. 50% MTF at 30% lower spatial frequencies) is easy to restore with sharpening, as long as noise is not a problem.

How then does [Bayer] interpolation factor into this? Wouldn't the Nyquist frequency for red or blue, in absence of green data from which to interpolate, then fall as low as 13 cycles/mm?

Yes - the red and blue are sampled at half the rate of green (46 vs 93 lp/mm). The neat benefit of Bayer interpolation is that it gives you full luminance resolution, almost as if by magic. The "downside" is that you do have less color resolution than a 3-chip of the same pixel count - but nobody wants color resolution anyway. 99% of 3-chip cameras throw away half the color resolution from the start by using 4:2:2.

That difference in color resolution only matters if you are shooting something with very high frequency color details. In practice, most people find that even the existing color resolution is far more than necessary. You can blur red and blue a whole lot more (e.g. Gaussian blur in in LAB mode in Photoshop) and it still wont affect the image, because you still have full luminance resolution.

It would be so much simpler if digital camera manufacturers published mtf charts of their systems at default settings.

Agreed. RED is doing better than most - they published MTF charts, dynamic range tests, spectral QE graphs, etc. - but as a whole, the industry publishes very little technical information.

I think they would do it if we still demanded it, but we don't. Now we just want fast, easy, and simple; learning technical stuff is not on anyones priority list. Why publish MTF charts, spectral QE graphs, dynamic range data, etc. when 99.999% of your customers are not interested in learning more than "where's the record button?" I think film users had a lower percentage of non-technical users (maybe 98%?), so there was enough interest to establish a precedent of publishing technical information.

I suppose, in theory, one could really boost the mtf of any system simply by sharpening the hell out of it. What's the mtf of hdr imagery? A lot of that stuff has rather insane amounts of wide-radius sharpening applied to it.

I prefer to call it "local contrast enhancement" instead of HDR -- though it's often not an "enhancement". The sharpening MTF goes way over 100% because it's no longer just transmitting the full amount of contrast from the scene to the display, but going beyond that and inventing contrast that could never have existed in the scene itself.

It may not be aggressive to your taste, but you must admit red's low pass filtering is pretty aggressive by industry standards.

It depends on the "industry". I was comparing it to film and high-end video cameras, where aliasing is generally very low.

Look at the 7d, for instance, which looks way sharper than it has any right to until artifacts quickly appear.

I would hardly consider 7D the standard. Canon just allowed photographers to store the liveview stream -- they didn't actually design it for video. The aliasing of the 7D was not a design choice, it was a nasty compromise dictated by the slow sensor. If you want to see what it looks like when Canon actually engineers something correctly, look at the new Canon XF300/305 -- it has very little aliasing. Furthermore, DSLR video is still a very tiny fraction of all video cameras.

In any case, 7D video footage, like that from most DLSRs, has such bad aliasing problems that it borders on the ridiculous. The artifacts are far worse than almost any other video camera in this century. Most people either like the artifacts, or they use such thin DOF that there is very little detail in the frame to begin with -- so very little aliasing. Much of the time they miss focus completely, so it might as well be SD -- but at least it doesn't have any aliasing.

And most medium format digital backs for still photography (and the leica m9) have no low pass filter at all.

There are several reasons for that. First, for the same reason that MFDB have poor low light performance, low dynamic range for their size, poor color accuracy (metameric failure), poor sensitivity, slow frame rate, and a host of other deficiencies compared to 35mm DSLRs: low sales volume. They can't afford to setup a lab to grow and polish Litium Niobate crystal of the size, grade, and purity needed at the sales volumes they can support.

Furthermore, the cost scales exponentially with area because even the tiniest defect will show up on the image, thanks to being so close to the sensor. Even though MFDB is only three times more area than 35mm, the cost can be an order of magnitude (or more) higher. More importantly, Canon/Nikon ship millions of units a year, compared to less than 6,000/year for all MFDB combined, so economies of scale is a huge factor.

MFDB have some great advantages, but unfortunately not enough to stem of death of Medium Format manufacturers (from over a dozen down to just 4 now). They can't afford to do their own sensor design, so they have to buy off-the-shelf parts from Kodak and Dalsa, which have far lower performance (per area) than DSLRS and don't come with AA filters.

The higher the resolution, the more likely that motion blur, depth of field, lens aberrations, diffraction, and other factors will prevent fine detail (and aliasing). MFDB are often used with thin DOF, where the aliasing only occurs on the plane of critical focus, such as the eyes and not the hair. They're also often used at very deep DOF, where diffraction reduces the aliasing.

Furthermore, with higher resolution, the less likely that aliasing will contaminate low spatial frequencies (only the higher ones). This means that the aliasing may not be visible in normal sized prints (e.g. magazine spread) even if it is in large ones (40x30).

There is also momentum. Some people have been practicing unsafe imaging (with no protective anti-alias filter) for so long that they have contracted a nasty STD: Sharpness Transfixion Disease. It caused them to be transfixed with sharpness above all else, even if the image is riddled with artifacts. Of course, what they see as "sharpness" and "3D", uninfected people see as jaggies, stair-stepping, sparkling, "snap to grid", wavy lines, bands, patterns, fringing, popping, strobing, noise, false detail, and of course moiré. MFDB manufacturers can't rip the aliasing artifacts away from these STD-infected customers too suddenly, or they will experience extreme withdrawals.

Even my experience with old betacams is that many of them alias really quickly and really seriously. Whether or not it's smart to do as much, it seems camera companies have put a premium on apparent sharpness over a lack of aliasing for quite some time now.

I'm sure you're right about the betacams. In older or interlaced cameras there are other issues that look similar to aliasing, such as interlace twitter and chroma crawl. But I was thinking more about quality HD video cameras. I have hosted some of the data about this:

http://thebrownings.name/WHP034/

 

[Matt W.]

Wow, thanks for the detailed response. Although some of it is beyond me, the rest makes a lot of sense and is very interesting.

I'm getting way beyond my comfort zone here, but using the Nyquist–Shannon sampling theorem wouldn't 93 cycles/mm translate into 46 cycles/mm without aliasing? That's where I was getting confused. But I suppose even the relatively strong low pass filter on the red allows for some aliasing somewhere and so the red is likely resolving much better than 46 cycles/mm. And I don't know how mtf correlates with the Nyquist–Shannon theorem if it has anything to do with it at all. Probably unrelated.

Too bad about medium format backs. They always seemed so cool to me and the samples from the 80 megapixel leaf back are amazing, but I've heard they are difficult to work with--and that sounds like a lot of ugly image artifacts to deal with. Not that I can afford one, anyway. Still, if, without an anti-aliasing filter, they're approaching 67% luminance mtf at better than 100 cycles/mm that's still really impressive. That's as high as or higher than the best black and white films (tmax 100) and in color and without the grain.

 

[Daniel Browning]

You're welcome.

using the Nyquist–Shannon sampling theorem wouldn't 93 cycles/mm translate into 46 cycles/mm without aliasing?

Nyquist is normally given as half the sampling frequency. The RED ONE sampling frequency is 185 L/mm, so you divide that by two to get Nyquist: 93 lp/mm.

That said, dividing by two is not enough to prevent aliasing, because we're not dealing with bandlimited signals. Imagine a black line right next to a while line, each taking up the width of exactly one pixel. If they aligned perfectly with the pixel grid and had no OLPF, you would see perfect black next to perfect white. But if you moved the sensor slightly, so that half the black line was on one pixel and the other half was on another pixel, you would see a gray image instead of black and white. So two pixels is not enough to prevent aliasing -- in practice you need at least 2.5 (or better, 3) pixels to record a transition from black to white. That is what the OLPF will do, hopefully: prevent any harsh transitions (that may be a product of chance) and make the resolution more smooth and natural.

But I suppose even the relatively strong low pass filter on the red allows for some aliasing somewhere

Yep -- can't get rid of it completely. Sometimes you can fix it in post (such as when only the high frequencies are aliased, or when red/blue are aliased but green is not).

Too bad about medium format backs. They always seemed so cool to me

Me too. I think a 6x9 camera would be about the largest I would want to handhold, so it would be great to have one if someday there is a breakthrough in silicon wafer manufacturing costs (and all the other costs that make them $40,000).

 

[Matt W.]

One last thing (I am procrastinating working on grad school applications today so I am posting compulsively all over the internet):

If you start with 186 pixels/mm, that translates to 93 cycles/mm as the sampling frequency in cycles (two pixels per cycle), which should then translate to 46 cycles/mm sampled without aliasing according to Nyquist, I think? So anything that's resolving more detail than that risks aliasing. It seems to me you'd need to divide by two twice to get maximum resolution sampled perfectly in cycles/mm. I may be mistaken. But I figure it all works out somehow and clearly with more detail than this would indicate.

If you're looking for handheld 6x7, the Mamiya 7 looks amazing. Unfortunately it's film and still pretty expensive.