canon d5 mk2 Chroma subsampling

[Ram Shani]

any one know's what is the camera chroma sudsampling?

[Daniel Browning]

It only reads every third line from the sensor, applies heavy noise reduction (even with NR: off) and 4:2:0 subsampling.

[Mark Hahn]

I'm not defending them, but when you only take the third line, you have to filter a lot or you'd get horrible moire. The fact that the filter also reduces noise is an unfortunate side effect. The plastic moire filter is designed for use with every line, not every third.

The "correct" solution would be to use all pixels and downsample in software or use raw, but I don't think that will happen until digic 999. That would require an immense amount of calculation.

[Daniel Browning]

Originally Posted by Mark Hahn

You have to filter a lot or you'd get horrible moire.

Given the fact that I *do* get horrible moire led me to believe that this was garden variety noise reduction, but if the moire would be even worse (hard to imagine) without this noise reduction, then it would be the correct trade off.

[Jon Fairhurst]

At least it's every third line, rather than every third line and every third pixel. We get moire, but only on horizontal lines. Vertical lines are clean.

I would love to see somebody develop an optimized filter for the camera. Sure, you'd lose some resolution, but it would give a pristine 720p result. For many, this would be an attractive tradeoff.

Then again, even in this digital age, many professional cable channels have NTSC in their chains. Compared to that, the 5D's moire is negligible.

[Mark Hahn]

Of course, if the camera cost $10,000, it would change the physical moire filter for different needs. You get what you pay for.

[Jon Fairhurst]

And it could have the right filter and cost much less, if it didn't have to take those pesky 21MP photos. ;)

Seriously, if Canon were to put this sensor into a stripped down video camera with manual controls, the right optical filter, and XLR audio, it would be a sweet piece of kit. Being Canon, I'd imagine that they'd be able to develop a high-end video lens for a step-up model - including motorized >10x 35mm zoom with settable rates, video-specific image stabilization, and video-specific auto-focus and momentary auto-focus features.

Imagine having that level lens for run-n-gun as well as the ability to use primes for narrative stuff.

Between the possibilities at Canon and Nikon and the developments at RED (not to mention plans from the majors that have yet to see the light of day), we have much to look forward to.

[Mark Hahn]

Originally Posted by Daniel Browning

Given the fact that I *do* get horrible moire led me to believe that this was garden variety noise reduction, but if the moire would be even worse (hard to imagine) without this noise reduction, then it would be the correct trade off.

Moire filtering and noise reduction are really two different animals . Noise reduction uses sophisticated algorithms to find the edges of objects and tries to leave them sharp. Then on the "flat" areas if assumes the perturbations are noise and filters them.

Moire has to be a low-pass filter that affects every single thing in the image, but when done right it only chops off frequencies above Nyquist, which are invisible anyway. It is impossible to build a perfect "brick-wall" filter that passes 100% of everything below Nyquist and 0% of everything above, but they try.

Now for the giant problem. When taking every third line, the Nyquist frequency is one-third what it is when taking every line. So the little plastic filter designed for every line is off by a factor of three in frequency. This actually doesn't reduce any more noise. Actually it passes way too much frequency which causes moire, not more reduction of noise.

The nyquist frequency is really weird. Assume you have a Nyquist limit at 100 Hz. If you try to pass 101Hz through it you get a wrap-around effect and the moire signal is 1 Hz. How weird is that? This causes the moire to appear often in big stripes.

So the obvious solution is to interchange the plastic moire filters like you do the viewer filters but they don't allow for that in the 5D2.

[Jon Fairhurst]

Great explanation, Mark!

There's one technical error, however, though the general idea is right:

Originally Posted by Mark Hahn

Assume you have a Nyquist limit at 100 Hz. If you try to pass 101Hz through it you get a wrap-around effect and the moire signal is 1 Hz.

Going beyond Nyquist is more like two mirrors (or a game of Pong), than a wrap around. A 101Hz signal in a 100Hz Nyquist system would give a result of 99Hz. Here are some examples that tell the story:

101Hz -> 99Hz
150Hz -> 50Hz
199 Hz -> 1Hz
201Hz -> 1Hz
250Hz -> 50Hz
299Hz -> 99Hz
301Hz -> 99Hz
350Hz -> 50Hz
399Hz -> 1Hz
etc....

Of course, the 5D MkII has an OLPF, so we can assume that the sensor never sees anything over 3x Nyquist in the vertical direction.

[Mark Hahn]

I stand corrected. I always get that confused.

[Jon Fairhurst]

No problem. The general idea was right. Frequencies that are too high become low frequencies. That's why a fine patterned shirt can result in big, low-frequency areas of light and dark.

[Mark Hahn]

That is what confuses me. In order to get a big moire band (near zero Hz), you would need a frequency that is almost double the Nyquist frequency. It would be easy to block that frequency, even with a cheap filter.

Of course the 5D2 lets three times the Nyquist frequency through, so every moire frequency is possible. The 5D2 video must have some kludge to make it look less bad.

[Jon Fairhurst]

This is where the right optical filter comes in. A filter that cuts off at 1.5x Nyquist or 2x Nyquist wouldn't eliminate moire, but it would help reduce it, and it should have no effect whatsoever on sharpness.

As you mentioned, however, there are no brickwall filters. You can't pass 0.99x Nyquist while removing 1.01x Nyquist with any filter not made of unobtainium.

For 1080p, something that passes 1.0x Nyquist, but removes, say, 1.5x Nyquist would be about as good as it gets, but you'd still have some moire - but less than we have now. For 720p, you might be able to get something that passes 0.67x Nyquist, and rejects 1.0x Nyquist and above. Given that, there would be no aliasing whatsoever.

Now, all we need to do is find those filters...

[Mark Hahn]

Originally Posted by Jon Fairhurst

This is where the right optical filter comes in. A filter that cuts off at 1.5x Nyquist or 2x Nyquist wouldn't eliminate moire, but it would help reduce it, and it should have no effect whatsoever on sharpness.

As you mentioned, however, there are no brickwall filters. You can't pass 0.99x Nyquist while removing 1.01x Nyquist with any filter not made of unobtainium.

For 1080p, something that passes 1.0x Nyquist, but removes, say, 1.5x Nyquist would be about as good as it gets, but you'd still have some moire - but less than we have now. For 720p, you might be able to get something that passes 0.67x Nyquist, and rejects 1.0x Nyquist and above. Given that, there would be no aliasing whatsoever.

Now, all we need to do is find those filters...

I only know electrical filters. I have no idea how plastic filters work. Could it be possible to add one to the the lenses or does it need to be immediately adjacent to the sensor? I suspect the latter.

[Jon Fairhurst]

I would think that the filter can be at the other end of the lens; soft and diffraction filters do work, after all. That said, the needed filter might vary depending on focal length and distance from the sensor.

One disappointment is that photo filters are desperately lacking in specs. About all they tell you is the name and diameter. I'm sure that plastic film suppliers and scientific optics companies have detailed specs, but either you need be in the market for factory quantities, or be willing to pay as much for the filter as for the camera.