Canon XL2 CCDs are 4:3.

[A. J. deLange]

RE: "This proves that regardless of the DV compression, it is possible to see more information if the source material is more detailed."

This statement, at first blush, appears to violate the sampling theorem. A sensor with 720 pixels in its horizontal dimension cannot represent more than 360 cycles per picture width. A sensor with more than that (e.g. the 960 in the XL2) can capture higher spatial frequencies but they cannot be correctly displayed if the image is downsampled to 720. "Proper" downsampling would require that all frequencies above 360 CPW be attenuated below the visibility threshold. If this is not done then those frequencies will appear as aliases. "Proper" downsampling requires a "brick wall" filter i.e. one whose response goes from unity gain at 359.9999 CPW to 0 at 360.0001. Such a beast does not exist. Practical filters must start to roll off gain well below 360 in order to have sufficient attenuation above 360. The wider this so called transition band the easier the filter is to implement i.e. the fewer "taps" it requires or, put another way, the shorter its impulse response needs to be. Impulse response duration is important when working with images because of its effects on the edges of the picture.

So why does 35 mm downsampled to DVD look better than the stuff coming out of our XL2's? I believe that the answer is that in the telecine process the frames are substantially oversampled which allows the interpolation filter to have a narrower transition band. Thus we get closer to 360 CPW, the fundamental limit of the medium than we do with the filter implemented in the camera.

What about 4:3 where there is no downsampling and we can't blame the low pass filter? I'm afraid the answer there is that the camera, though its chips have 720 pixels, is not capable of anything near 360 CPW resolution. This could be caused by MTF limitations in the lens or other optical components (which doesn't bode well for an HD version of the XL series) or by the infamous Kell effect or combinations of the above. I find the manufacturers very disingenuous when it comes to discussions of the resolutions of these cameras (and not just Canon). There's lots more to it than pixel counts! Why won't they show us resolution targets or MTF curves?

To summarize: The absolute limit of the DV system is 360 cycles per picture width. The fact that telecine video looks sharper on DV than XL2 video means that DV is not the limiting factor in the XL2's resolution. So in essence I agree with the quoted statement up to somewhat less than the ultimate limit which DV does impose: 360 CPW.

[Andre De Clercq]

Synchronious sampling a CCD structure (e.g 720 hor samples for 720 hor pix) is not a nyquist problem. And indeed higher pix counts need to be downsampled with all the low pass filtering problems and limits involved. The basic problem is the optical sampling by the physical CCD structure: the scene is being sampled by 720 (hor) photosensors( simular story is true in vert direction). If this scene contains spatial (horizontal) components which can interfere with the pixalated CCD structure we get optical aliasing which cannot be reduced by whatever filtering on the electrical signal. So optical lowpass filtering (preferable steep) is the only way to go. Consumer cams quite often only have the MTF and/or diffraction limits as optical low pass filtering. Not at all steep and thus introducing alising vs res limits as a trade off. If we got a cam with an infinite number of pix there would not be optical aliasing, and just an ideal (digital) lowpass filter would result in 'perfect' DV samples. The approach on the optical oversampling (a multitude of 720 hor pix) and steep filtering is what happens in 35 mm telecine processing., and this makes the DV pix better.

[Richard Hunter]

<<<-- Originally posted by Rob Lohman : Besides the excellent point made by Marty you also have an
actual increase in resolution: the vertical resolution. Let me
clarify by a short example:

Normal NTSC DV is 720 x 480 pixels.

A fake 16:9 camera that does an anamorphic stretch does the
following:

1. crop the signal to 720 x 364

2. stretch that back out to 720 x 480 (upsampling)

(or it can do the stretch first and then a crop, but that's the exact
same thing)

So you are now working with a 720 x 364 pixel image instead of
720 x 480! Which looses you 116 lines of resolution (which would
be the same if you letterboxed it for example)!

A true 16:9 camera like the Canon samples the image in 16:9
mode at 960 x 480: -->>>

Hi Rob. I agree that native 16:9 is higher resolution than stretched 16:9, but what I was questioning was whether it is higher than native 4:3, if both are sampled at 720x480. I still don't really see that it can be so. Maybe your point is that most cameras will not offer both native 16:9 AND native 4:3, and must compromise on one of them? I would agree with that.

Regarding Marty's post, I think the example of comparing the quality of a 35mm image with DV footage is not so relevant. If I compare XL2 and VX2000 4:3 DV footage, I can also see differences in clarity that are obviously not related to resolution.

Even if we talk about sampling 16:9 at 960x480, it is proportionally the same number of horizontal pixels as sampling 4:3 at 720x480. Remember that the XL2 shows a wider view in 16:9 mode, and 1/3 more width than 4:3 requires 1/3 more pixels to maintain the same resolution. So how can we expect any more clarity or higher resolution from a 16:9 image sampled at 960x480 pixels compared with a 4:3 image sampled at 720x480. Even allowing for perfect downsampling (which is another can of worms), I thought the results should be exactly the same (except one is wider).

Richard

[Jean-Philippe Archibald]

An interesting thread to follow. If one of you know the answer, I would be interested to know how the SDX900 acheive to provide the differents aspects ratios. Three 16X9 2/3" with a smaller patch for 4:3, or the same as Xl2 (but bigger CCDs of course)?

[Rob Lohman]

Richard: the XL2 offers both native 16:9 *AND* native 4:3

The 16:9 is NOT SAMPLED at 720 x 480, it is STORED at 720 x 480.

That is a WHOLE different thing. The CCD's are sampled at 960 x 480!

That is downsampled to 720 x 480. Again, if you want to strictly
talk resolution the INCREASE is NOT in the HORIZONTAL BUT in
the VERTICAL. You get 480 lines of (original) pixels instead of 364!

So the sampling IS higher and the end result IS better.

You are mixing horizontal with vertical. And yes, you are right
that the field widens so more information is in the horizontal
(which should perhaps yield a higher compression level).

Your comparison to the VX2000 actually has everything to do with
resolution since the sensors on the XL2 use more pixels (which is
whole other thing to talk about...) and the lens is probably sharper
as well. But that's for another thread <g>

[A. J. deLange]

Lets assume for a moment that the MTF of the lens is what limits the resolution and further assume that one views at the optimum distance where 1 horizonatal scan line subtends 1/60 degree. If one puts the camera on a tripod, frames a shot and records a few seconds in both 4:3 and 16:9 the image impinging on the CCD block will be exactly the same but the 16:9 horizontal data has to be compressed by 720/960 to fit into the 720 pixels DV allows so that a picture sample which covers 1 unit of width on your fovea in 4:3 covers 1.33 units in 16:9. The 16:9 picture is, thus, 33% blurrier in the horizontal direction but equally sharp in the vertical.

[Richard Hunter]

Hi Rob. As I mentioned, I am not comparing stretched and native 16:9, I am trying to compare native 16:9 and native 4:3. The vertical resolution is therefore fixed at 480 (for NTSC) so is not a variable in this case.

Even if there is no downsampling loss going from 960 to 720 samples (which I think is not feasible in real life) my point is that 960 samples for a 16:9 horizontal line is the same resolution as 720 samples for a 4:3 horizontal line, so I don't see how native 16:9 can be described as higher resolution than native 4:3.

Regarding the VX2000 and XL2 images, I am comparing 4:3 video from both cameras, not native 16:9 with stretched 16:9. I believe the resolution of both in 4:3 mode is 720x576 (I am using PAL) and the sensors are using exactly the same number of pixels. As far as I can see, the differences are not related to resolution at all. When you say that the sensors on the XL2 use more pixels, does this apply in 4:3 mode?

Richard

[Chris Hurd]

<< I don't see how native 16:9 can be described as higher resolution than native 4:3. >>

Well, strictly speaking, no it is not. But the main point here is that there's no loss of resolution in 16:9 mode, as there always had been in the past prior to the availability of native 16:9 camcorders such as the XL2.

<< When you say that the sensors on the XL2 use more pixels, does this apply in 4:3 mode? >>

Yes, see the comparison chart at www.dvinfo.net/canonxl2/articles/article06.php

[A. J. deLange]

It may be getting down to semantics but if resolution means the ability to resolve then there is a loss of horizontal resolution in the XL2 in 16:9 mode relative to 4:3 because the same number of pixels are being used to cover 33% more world in the horizontal direction. If you hang a resolution chart on the wall, frame and focus and capture some video in 4:3 then change to 16:9 without changing anything else then upsample the captured images so they have the same number of pixels vertically (resulting in more samples in the horizontal direction for 16:9) and compare you will get something like this:

http://www.pbase.com/image/37816574

These images are crops from the centers of the captured images and have been upsampled to be 400 x 400 pixels each. The 16:9 sample is on the left; the 4:3 on the right. That the horizontal resolution is better with 4:3 is quite clear from examination of the Group 1 vertical bars. What you get in return for this lost resolution is more picture. Now it is true that we are appreciably better off than in the XL1 where 16:9 was obtained by tossing out lines at the top and bottom of the picture.

[Barry Green]

I'm not sure I follow what you're demonstrating here. Are you saying that you digitally resized (up or down) the 16:9 image? To what purpose -- just to maintain the same visual aspect ratio? Because the resizing process may account for some of the difference here...

A fundamental difference between 16:9 and 4:3 is that they use the same # of pixels, but the shape of the pixels is what changes. To be a fair test of what is actually being resolved by the camera, I'd think you shouldn't do any resizing of the image, just let one be fat and the other be skinny and let's see what the actual pixels look like.

(btw, I do think your original premise is correct, and that there should be an overall slight loss in resolution in 16:9 as vs. 4:3, due to the 960->720 pixel conversion. I'm just not sure I understand quite what you've done here and whether that reflects it properly).

[Richard Hunter]

<<<-- Originally posted by Chris Hurd : << I don't see how native 16:9 can be described as higher resolution than native 4:3. >>

Well, strictly speaking, no it is not. But the main point here is that there's no loss of resolution in 16:9 mode, as there always had been in the past prior to the availability of native 16:9 camcorders such as the XL2.

<< When you say that the sensors on the XL2 use more pixels, does this apply in 4:3 mode? >>

Yes, see the comparison chart at www.dvinfo.net/canonxl2/articles/article06.php -->>>

Hi Chris. I think you are comparing native and stretched 16:9 again, which is not what I was questioning.

For the sensors pixel count in 4:3 mode, your comparison table shows just under 2% fewer pixels for the VX2100. All other things being equal (which I totally accept is not the case when comparing the XL2 and VX2000) I would not expect this difference in resolution to be very perceptible. If I am wrong, then the resampling from 960 to 720 horizontal pixels for 16:9 mode would give a huge reduction in perceived resolution, (notwithstanding the argument already put forward that we can still see more clarity from a high quality source image that has been downsampled to DV resolution). There just has to be a quality hit when going low rez, otherwise the solution to delivering high definition video at low bit rates would be very easy - just downsample it.

Richard

[Filip Kovcin]

sorry if this was already asked before - in this case, please send me just simple link to corresponding issue (subject).

who is producer/provider of canon's xl2 ccd?

i read (heard?) before it was panasonic (for XL1), but cannot confirm it. anyone knows?

when i tested both cameras dvx100A and brand new canon xl2 on ibc in amsterdam last september, the look and feel was almost identical - very similar color interpretation, especially when difficult colors are presented (i had special carton wiht bluish green which is my "ultimate" color tester), and the result was very similar if not identical.

so, my question is - is canon's ccd panasonic's one?

thanks,

filip

[Andre De Clercq]

What deLanghe compares by these transformed images is in fact the angular screen resolution as seen from a fixed distance at fixed picture hight. And indeed the number of cycles/degree is in favor of the (smaller) 4:3 picture if the basic number of pixels/lines is fixed in the images.

[A. J. deLange]

Barry,

Fair enough. The original frame grabs are at http://homepage.mac.com/ajdel/FileSharing7.html. A couple of things folks need to think about when viewing these. They are both 720 x 480 which is 3:2. To see them as they would be seen on a TV screen they need to be re-sampled to, respectively 16:9 and 4:3. That's what I did in Photoshop using bicubic interpolation with the final picture heights being the same in each case. Also, when you look at them on a computer monitor you have to be cognizant of the fact that resampling is taking place and that both the algorithm used by the viewing application and the interpolation (if any) done by your graphics card come into play. Photoshop is a good tool for viewing these frames because there is an "actual pixels" mode in which one pixel in the file goes to 1 pixel on your monitor and when scrutinized using the magnifier tools individual pixels are blown up to "tiles" so that the individual pixels in the original can be clearly discerned.

Happy Holidays to all!

[Barry Green]

Okay, I am officially in the A.J. deLange fan club now... I understand exactly what you're saying, I agree, and I can demonstrate it as well.

Here's an extraction from the XL2. This was an Accu-Chart, which is a high-definition test chart. We shot this exact same chart on the XL2 from a locked-down tripod shot, the only thing that changed at all was the switch on the XL2 from 4:3 to 16:9.

I took the horizontal resolution section and extracted a patch and compared them side-by-side, no resizing or anything.
http://www.icexpo.com/XL2-16x9-vs-4x3.JPG

The XL2 is quite clearly capable of resolving horizontal detail much more clearly in 4x3 mode than it is in 16x9 mode. In 4x3 mode the lines marked 500 are crisp, sharp, and cleanly resolved, and 600 looks pretty good as well. It even did a fair job at separating out the lines marked 700. In 16x9, 600 is not resolved well at all, and 700 turns to complete mush on the right side. I would tend to think the 960->720 downsampling has something to do with it; it would be interesting to see how the FX1 performs on the same chart, seeing as it has 16x9-shaped CCD's with wide skinny pixels (but also 960 of them).

I think I have that shot here somewhere, I'll have to look for it...
(edit: I don't have that shot. I do have an in-camera down-rez of shooting the Accu-Chart as HDV, and recording the firewire output as DV. In that mode, the FX1's horizontal resolution looks basically the same as the XL2's 16x9 mode).

(for reference, the Accu-Chart lists different horizontal resolution numbers depending on whether you're shooting 16x9 or 4x3, to take into account the effect of "TV Lines" of resolution. According to the accu-chart, the XL2 is showing about 450 lines in both 4x3 and 16x9 mode.)