C300: Actual Camera Resolution Before Aliasing

J

jcs

Veteran
[EDIT] After further discussion, .66 is sufficient (vs. .58), which matches what e.g. ARRI did with the Alexa for HD: 2880/.66 = ~1900. [/EDIT]

Recently there has been discussion regarding 4K vs. 1080p and actual camera resolution. After spending $16,000+ for a C300 II, I was a bit disappointed to see aliasing in 4K. After a bit of debate, research, and computer graphics tests to simulate sensor sampling & Nyquist, I reached the same conclusion as Jao did here, ".58 & .8" multipliers for Bayer sensors: http://lagemaat.blogspot.com/2007/09/actual-resolution-of-bayer-sensors-you.html

The summary in the real world (which matches Nyquist sampling theory), is Bayer sensors provide .58x the photosite resolution before aliasing begins, where extinction of detail occurs at .8x:

Bayer+resolution.jpg


Here is my 2x sampling simulation (2 pixel black & white pairs), showing extreme aliasing when rotated:

Lines44_4degSmall.png.3fa44548424ef92132e574a8cf82 55cb.png


Nyquist shows us we need >2x sampling to eliminate aliasing, thus here's 3 pixels instead of 2 to represent a line:

Lines44_4degAA_.png.612d29c6869e32392a780d2a0ef9de 56.png


Jao measured many real-world test charts for Bayer sensors, and he found .58x before aliasing, pretty much the same as predicted by the 3-pixel line test (where 4 pixels would be even better). 1/1.5 = .67, 1/2 = .5, (.67 + .5)/2 = .58! http://lagemaat.blogspot.com/2007/09/actual-resolution-of-bayer-sensors-you.html. So .58x is pretty much an excellent predictor of alias-free resolution possible with a Bayer sensor. More images and info here:
http://www.eoshd.com/comments/topic/25630-why-shooting-4k/?do=findComment&comment=205652

This explains why the C300 II, C200, and to a lesser extent, the C700, all alias in 4K: in order to get alias-free 4K, we need 3840/.58 = 6620 Bayer photosites.

One reason the Alexa (ALEV III sensor) looks so good is 2880*.58 = 1670, which is 83% of 1920, vs. 3840*.58 = 2227, which is 58% of 4K (e.g C300 II sensor). The key here is aliasing: whenever aliasing occurs, even if not substantially visible (such as Moire), unconsciously we can see it, which hurts perceived image quality. That's why Alexa sensors and film tend to look better to most people, and one of the secrets to the film look (along with sufficient organic noise/grain to provide dithering and texture).

Here the Sony F65 with its "8K" (kinda) sensor shows alias free 4K performance:

Where the C300 II (raw):


C200 (raw):


C700 (raw):


Varicam Pure (raw):


Varicam LT (raw):


All alias, as predicted by the math as they don't provide sufficient Bayer photosite resolution for alias-free 4K sampling.

When shopping for a new 4K camera, use .58x the Bayer photosite resolution to compute alias-free resolution performance. Additionally, the OLPF must be tuned to filter out frequencies higher than .58x, else aliasing will occur. The Alexa 65 with 6.6K meets the "True 4K" requirement as does the Sony F65 (and I'm sure their new CineAlta release due shortly). I haven't seen test charts for Red: they could also provide True 4K performance at 6.6+K, depending on OLPF used.
 
Last edited:
Yes, one needs a 4:2:2 for a claimed resolution - 4.2 MPX for 1080p, 17.6 for 4K. Which is why F65 and Alexa65 (not related) have a 17.6 MPX sensor.

But, on the low end - pending processing requirements and capabilities - more pixels deliver more detail up to the point where there are diminishing returns due to an eye and brain perception.
 
jcs said:
in order to get alias-free 4K, we need 3840/.58 = 6620 Bayer photosites
Click to expand...

If you're counting horizontal resolution you need to multiply by about 0.76, not 0.58.
0.76 x 0.76 = 0.58

Jao van de Lagemaat is counting total pixels (MP), and we usually count just one line (K).
A 5052 x 2842 sensor would be 14.4 megapixels.
Bayer reduces each side by 0.76, so it would be 3840 x 2160 of actual resolution, which is 8.3 megapixels.
8.3 / 14.4 = 0.58

So each side is reduced by about 3/4, and the total area is therefore is reduced by about one half (because 3/4 x 3/4 = 1/2)

This matches up with Graeme Nattress's measurements for the Red One camera: http://www.reduser.net/forum/showthread.php?9891-Red-One-Measured-Resolution
He's saying about 3/4, and Jao says 1/2, and they're both right. It's the same number of pixels, just different ways of counting them.


Jao's article is 10 years old. Even before then, I was hearing how you need get about 75% of a side's resolution out of Bayer.
What I see much less of is a similar scientific explanation for what resolution you should be going for.

So 5K gets you true 4K. Great. It's not like 4K is the 88 mph of cameras, where once you reach it, your camera will light up and travel back in time.
So your camera is true 4K.
Well, my camera is true 3.2K.
And Bob's camera is true 5.3K.

What if all you need is true 2K? Then you're just wasting dynamic range, ISO, frame rates, etc.
What if you actually need true 7K? Then none of us have reached the goal.

How much resolution do you need --- or want, whatever.
What reasons did you use to reach that magic number?
 
Last edited:
IMHO, not much of this matters in most of our real worlds and only camera nerds care about this stuff. None of it makes you more money or helps you tell better stories. Cool and fun to geek out on and learn but in the real, non-test bench, non chart world, few can see it, notice it or care. I agree with the findings in regards to the Arri cameras in particular but I feel there is much more to their secret sauce than side resolution. Just my .02 worth as a mostly medium to lower end of the scale working DP.
 
Last edited:
combatentropy said:
If you're counting horizontal resolution you need to multiply by about 0.76, not 0.58.
0.76 x 0.76 = 0.58

Jao van de Lagemaat is counting total pixels (MP), and we usually count just one line (K).
A 5052 x 2842 sensor would be 14.4 megapixels.
Bayer reduces each side by 0.76, so it would be 3840 x 2160 of actual resolution, which is 8.3 megapixels.
8.3 / 14.4 = 0.58

So each side is reduced by about 3/4, and the total area is therefore is reduced by about one half (because 3/4 x 3/4 = 1/2)

This matches up with Graeme Nattress's measurements for the Red One camera: http://www.reduser.net/forum/showthread.php?9891-Red-One-Measured-Resolution
He's saying about 3/4, and Jao says 1/2, and they're both right. It's the same number of pixels, just different ways of counting them.


Jao's article is 10 years old. Even before then, I was hearing how you need get about 75% of a side's resolution out of Bayer.
What I see much less of is a similar scientific explanation for what resolution you should be going for.

So 5K gets you true 4K. Great. It's not like 4K is the 88 mph of cameras, where once you reach it, your camera will light up and travel back in time.
So your camera is true 4K.
Well, my camera is true 3.2K.
And Bob's camera is true 5.3K.

What if all you need is true 2K? Then you're just wasting dynamic range, ISO, frame rates, etc.
What if you actually need true 7K? Then none of us have reached the goal.

How much resolution do you need --- or want, whatever.
What reasons did you use to reach that magic number?
Click to expand...

Good point regarding Jao's .58 coming from total megapixels. I got .58 from doing computer graphics tests and realized that 2 pixels is not enough to represent lines without aliasing. See the rotated line examples in the first post. Studying aliasing in general with computer graphics, we need 3 pixels, which form a simple sinusoid, so that we can rotate the lines without aliasing. So if 2 pixels represents Nyquist at = 2x, 3 pixels is Nyquist at >2x. The line test used 64 pixels in the first aliased case with 2 pixels (alternating black and white lines, max possible resolution in the perfectly aligned sensor case). The 3 pixel case used 43 pixels, and 64/43 = 1.49 = 1.5x. I had previously guessed at 2x based on the F65 test chart results (~8K to get 4K). So 1/1.5 = .67 and 1/2 = .5. Those looked like they'd average to something near .58, (1/1.5 + 1/2)/2 = .58! However Jao as you pointed out was using total mpix for his test chart analysis, so .67 could be closer to accurate. .76 seems a bit high based on the computer graphics test, so it would be cool to see test charts which support .76. Do you know of any? (Understand .76*.76 = .58, however I haven't seen actual test charts which show this (no aliasing at .76)). Another way to provide a 'proof' would be doing the computer graphics test with .76 lines instead of .67: intuitively we'd get some aliasing.

The point is, 4K cameras (such as the C300 II), can't capture 4K without aliasing. Just raising awareness that some cameras marketed as 4K really aren't if we include little or no aliasing as a requirement for True 4K. See the above test charts for real world results: Only the ~8K F65 (~.5x) provides alias free 4K...
 
Last edited:
puredrifting said:
IMHO, not much of this matters in most of our real worlds and only camera nerds care about this stuff. None of it makes you more money or helps you tell better stories. Cool and fun to geek out on and learn but in the real, non-test bench, non chart world, few can see it, notice it or care. I agree with the findings in regards to the Arri cameras in particular but I feel there is much more to their secret sauce than side resolution. Just my .02 worth as a mostly medium to lower end of the scale of working DP.
Click to expand...

Aliasing can be pretty nasty, including creating ruined and unusable shots from fine fabrics (suits etc.) and brick walls. So this post is more about aliasing wrecking things (as well as unconscious reduction in filmic quality: less organic) when manufacturers skimp on Bayer photosites along with sufficient OLPFs to eliminate aliasing.
 
Sony says F65 has a Q67 sensor (or sensor pattern). So, it looks like 67% (17.6 megapixels in 4:2:2) of an RGB (26.4 MPX, 4:4:4) readout is it. And, as I had mentioned above, 17.6 MPX is what both F65 and Alexa65 have, albeit in different sizes.

It'd be interesting to see what the new Sony Cine Alta full frame camera will have. I speculate an RGB 26.4 MPX sensor ... which will probably be billed as "12K".
 
jcs said:
Aliasing can be pretty nasty, including creating ruined and unusable shots from fine fabrics (suits etc.) and brick walls. So this post is more about aliasing wrecking things (as well as unconscious reduction in filmic quality: less organic) when manufacturers skimp on Bayer photosites along with sufficient OLPFs to eliminate aliasing.
Click to expand...

If it's subtle, most clients never notice it, although if it's blatant, they are not happy campers. It usually comes down to proper wardrobe choice for me though, I rarely have issues with buildings and brick walls from a distance aliasing, it's almost always clothing.
 
My 4K monitor still shows aliasing on the Sony F65 vimeo video at 4K. Part of the problem of showing these tests is that the actual delivery, usually by streaming video, has its own limitations. Even when I downloaded the original F65 footage, I am still seeing aliasing and moire on a 4K full monitor playback.

In any case, this is why I think Canon will be delivering an 8K sensor sooner rather than later. The 4K from 8K seems like the simplest solution. I've always thought of the 4K from Canon's cinema line to be similar to the "FullHD" from the DSLR line in terms of relative resolution.

I do wonder if the internal 4K might have a bit of anti-aliasing built in and would have been interested to see a test.
 
Last edited:
On the one hand, I'm with puredrifting. Search for BMPCC on Youtube and just casually watch a few. Don't sherlock for moire. Just see if you can enjoy the video. Lots of great stuff. And yet it's as ghetto as it gets. The native sensor count is "only" 1920 x 1080, and there is no optical low-pass filter atop it. For those of you trying to please nervous producers, it's probably hard these days to get by with a camera like that (assuming it had ergonomics you wanted). But if you forget the nervous producer and your own OCD, and put yourself in the shoes of the average audience member watching it on their TV or laptop, you have to admit that we fret quite a bit over stuff that doesn't really matter.

On the other hand, if you put on your perfectionist hat and, if for no other reason than an intellectual exercise, ask what would it take to no longer worry about moire, then it seems you have to at least do what Canon did with the original C300: downsample 4 pixels to one. Aiming for 2K delivery? Shoot with a 4K Bayer sensor. Aiming for 4K? Shoot 8K. Etc.

And just to drive you crazy, there have been reports of moire even on the original C300. But really, it's so subtle, it's not worth fretting over, unless it's a very niche science project.

https://vimeo.com/57185034
http://www.xdcam-user.com/2012/02/c300-moire/

Bonus thorn: Even if your footage is devoid of moire, you can still get severe aliasing when it gets automatically scaled on someone's computer (http://www.dvxuser.com/V6/showthread.php?356295-GH5-Moire). It's rare, but since you're asking ;)
 
If aliasing and moire don't bother ya, sure, you've got a lot more options :)

I shoot mostly for our own projects, and when I saw moire on the C300 II, I was disappointed. Since it doesn't happen very often, it wasn't a big deal. Out of curiosity given my computer graphics and video game background, I looked more closely at camera sensors and aliasing to figure out what it takes to have alias- and moire- free video. Also, remember the 5D Mark II? It took amazing video however it suffered from moire. Canon released the 5D Mark III, and moire was pretty much gone. Then people complained it was too soft! I showed that with post-sharpening, the 5D3 was mostly alias-free as well as detailed enough for most uses (only in wide shots did the low resolution become significant). Thus we want detailed and alias free images.

Perhaps Vimeo showed aliasing with the F65 footage because it was a streaming issue? They use adaptive bitrate streaming, so until it's fully up to speed/cached, the lower resolution versions might be aliasing.

The .67 sensor resolution factor is a useful tool to determine if a particular camera can possibly be detailed and moire free at the same time.
 
FWIW, Canon has been showing its 8K camera for a couple of years now. It records externally to four Convergent Design Odyssey 7 units. Kind of reminds you of the first 4K consumer camcorder JVC Q10 that recorded to four different SD cards.

And, as I'd mentioned previously, it'd be interesting to see what Sony shows on the 6th. Given that theirs is a full frame, not s35, size sensor, one expects a higher pixel count. I figure it'll be usable 33 MPX at the very least.

8K monitors have already been out.

https://petapixel.com/2017/03/27/dells-retina-busting-worlds-first-8k-monitor-officially-sale-5000/
 
You must log in or register to reply here.
Back
Top