What's the Kell factor in 720p?

michael

According to a brief Google search, some folks put it as low as 0.67 and others say it's as high as 0.9.

Do we have a lines of res on that? Isn't 1080i like in the 800+ ball park. What would 720p be?


According to a brief Google search, some folks put it as low as 0.67 and others say it's as high as 0.9.

See the ASTC DTV comparison chart at http://www.evansassoc.com/lib/Atsc-dtv.html -- they're assuming a Kell factor of 0.9 for all progressive modes including 720p.

0.67 comes from http://www.dtvforum.info/lofiversion/index.php/t531.html

For anybody who may be wondering what the heck we're talking about, see "The Kell Factor Explained" at http://members.aol.com/ajaynejr/kell.htm

Good Links! What I wanted exactly!

Thanks DVINFO Master............


[/url]

I've got to re-read Poynton, but I think things are a bit more complex than that:

Interlaced video is filtered to stop line twitter, progressive, as true progressive for a true progressive display is not.

Both interlaced and progressive video are effected by the Kell factor, which relates to the perceived resolution of a display.

The "interlace factor" and the "Kell factor" are two different things, but it sounds like a lot of people lump them together. As far as I understand, Kell did his experiments with progressive video.

Anyway, let me re-read Poynton and sumarise back here....

Graeme

Don't forget the Nyquist limit. Aaaaggghhh! The Nyquist limit!

Oh there you go, you had to bring up seaquest..Oh oh I mean Nyquist.

Yeah Nyquist. Good band, right. Ohhh, Sh&* , I can't know everything.

What the hell are Nyquist's for gods sakes?

Michael, - not knowing


Don't forget the Nyquist factor. Aaaaggghhh! The Nyquist factor!

See "Sampling and Nyquist's Theorem for Audio and Video"

located at http://members.aol.com/ajaynejr/nyquist.htm

and then see also "Nyquist and TV Lines, what is the resolution of a digitized TV image"

located at http://www.spectra-one.com/digitalvideo.html#digtv

I've got to re-read Poynton, but I think things are a bit more complex than that:

Interlaced video is filtered to stop line twitter, progressive, as true progressive for a true progressive display is not.
e

I too have wondered if the .7 Kell factor is multiplied by a the interlace filter factor.

So I created a model of interleave handing line pairs usind row-pair summation. Then used the created frames as the basis for Kell.

I've concluded Kell of .7 is applied equally to interlace and progressive V rez. The reasons P looks so much higher in V res I don't think is based upon two filter functions killing I V rez. Otherwise, NTSC V rez wouldn't be about 340-lines.

And, yes Kell is .7 done on P video.

Steve! Somehow I just *knew* this thread would attract you. And there really hasn't been any substance to it yet, until now! Heh.

My own understanding -- correct me if I'm wrong -- is that we have the Kell factor to thank for 720p looking as good as, or in some cases better than 1080i, but I often wonder if I'm not confusing Kell vs. the interlace factor.

I'm getting a headache- too much for a Sunday morning.. and to think, all these years, I've just used my eyeballs and trusted the messages I recieved from my brain!

no, seriously, thanks for the info, interesting!

Now that we have the factor, what is the scientific rez of 720p, approximate area?

I'm getting a headache- too much for a Sunday morning.. and to think, all these years, I've just used my eyeballs and trusted the messages I recieved from my brain!

no, seriously, thanks for the info, interesting!

It depends on what you're imaging. If you lined up perfectly, pixel-for-pixel, with a 1280x720 grid (on a piece of paper, for example), then 720p would yield 1280 x 720.

Interlaced video never will. The interlaced factor is always going to cost you about 30% of your resolution -- the interlaced factor meaning that fields are blended together to reduce flicker. So the very best you'll ever get out of a 1920x1080i system would be about 1920x800x30.

Not so with 720p. With 720p, every pixel stands alone, and is not de-interlaced or filtered. So you could, under ideal circumstances, get 1280x720x60.

When the Kell factor comes into play is when you haven't lined up exactly perfectly with the source material. When the grid on the paper lines up on the CCD perfectly, you will have perfect line-for-pixel reproduction on a progressive camera (although never on an interlaced camera). But when the grid lines and the CCD pixels don't line up perfectly, when lines straddle pixel boundaries, that's when the Kell factor comes into play. The Kell factor is an attempt to take into account the resolution that's lost when the pixels and lines aren't perfectly matched, so some gray area comes into play.

Thanks Barry!


Michael


It depends on what you're imagin
.

Barry, your description of the interlace factor is great, but your description of Kell sounds more like the description of a Nyquist factor. I've still not read Poynton over again to get this right in my head though... Kell is often made up to be some kind of über-factor that covers interlace/Nyquist and whatever else reduces perceived resolution, but Kell itself is applicable to both interlace and progressive, as would a Nyquist factor.

In any sampled system, there are two filters - the anti-alias filter on the input and the reconstruction filter on the output. The beam of a CRT does a sort of gaussian reconstruction, which would have very much a different visual appearance of, say, an LCD screen, so does this effect any of the above factors, in that Kell was working originally with progressive video on a CRT??

Graeme

Well, that is indeed a question -- and when do we attribute resolution loss to the interlace factor, what circumstances have us attribute it to Nyquist, and at what point do they interrelate to Kell?

Seems to me the Kell factor should be equally applicable to horizontal pixels as well. The interlace factor only ever affects vertical resolution, but half-sample offset (whether Kell or Nyquist) should affect horizontal as much as vertical...

Perhaps as was said earlier, the Kell factor is a catch-all that covers all of the above?

1080 x .7 is just over 700-lines. And 480 x.7 is about 340 which definitely is NTSC V rez. So the Kell seems accurate for both NTSC and 1080i.

720 x .7 is just over 504-lines.

Your explanation of Kell is corrrect for progressive yet remains equally correct for interlace even though row-pair summation SEEMS to make interlace seems different.

Essentially, Kell in progessive covers the probability of non-alignment (thereby causing mixing of lines) while in interlace it covers the FORCED mixing of two lines by row-pair summation.

Amazing they both come out to the same number!

Barry, your description of the interlace factor is great, but your description of Kell sounds more like the description of a Nyquist factor.

Totally different. Nyquist forces the addition of optical and electrical low-pass filters (before the A/D) at under half the frequency response (H V rez) of the lens and CCD to avoid A/D sampling down at 2X higherto not yield aliasing.

Both factors plus MTF limit sharpness making all claims of CCD and format rez damn near meaningless!

"Both factors plus MTF limit sharpness making all claims of CCD and format rez damn near meaningless!" - Nyquist certainly doesn't make these resolutions meaningless!!!

The Nyquist limit of, say 720 pixels horizontal, would be a "resolution" of 720, but that's only if you have perfect brick-wall filters and the like. A very reasonably amount for a deduction to account for the slope of the filters is about 30%. That is because resolution in pixel terms or TVlines terms is based upon lines, not line pairs. Nyquist deals with cycles / frequency / line-pairs. You need twice as many sample points (ie pixels) as their is cycles, ie line pairs.

Graeme

Ok, it looks to me that we have three different factors here that lower resolution (or should I say, perceived resolution):

Nyquist (caused by pre-filtering to stop aliassing, ib both horizontal and vertical directions, usually done by an optical filter)

Interlace (a vertical filtering to stop line twitter)

and Kell.

According to Poynton, Kell tried to quantify the the resolution loss in the TV system when it was noted that the image had a lower resolution than had been predicted based upon Nyquist's work. This was back in the 30s, and was indeed with progressive images. The first figure Kell gave in 1934 was 0.64, and a later 1940 paper by Kell gives the factor as 0.8, but it should be noted that the Kell factor is a subjective measurement. It depends on the spot profile of the beam in the CRT, and hence, I believe that the Kell factor would be different depending on the technology the display uses - be it CRT, plasma or LCD, so there cannot be just one Kell factor.

The problem with the Kell factor, again, according to Poynton, is that it is ambiguously defined and therefore depending on the type and circumstances of the experiment you use to measure it, you'll get different results, and because it's subjective, different experimentors will get different results even when conducting the same experiment.

Because of this, Poynton refers to the above as the "Kell Effect", in that he was the one who discovered this resolution loss, but because it's hard to quantify and different circumstances will produce different results, he refuses to assign a number to the "Kell Factor".

As Kell's initial work only concerned progressive video, there was no accounting for interlace at all. Poynton comments that many people have lumped the interlace factor and Kell factor together into one über-factor, but I think this leads to further confusion.

So yes, there is a resolution loss above and beyond what would be predicted from sampling theory, and yes, this is above and beyond the filtering that is there to stop interlace twitter (if you're in an interlaced system) and yes, this is called the Kell Factor or Kell Effect, but it's very hard to put a number to it that is accurate as it will differ on so many factors including the viewer. It's roughly around 0.7, but again, I have no idea how that changes for modern viewing technologies like LCD or DLP or Plasma.

I hope that helps, and if anyone has any further insight, or even ideas on how we could conduct our own experiments to find our own personaly Kell factors, then that would be very interesting indeed.

Graeme

Ok, it looks to me that we have three different factors here that lower resolution (or should I say, perceived resolution):
e

What I think we all agree on is that the resolution we GET is much lower than what Sony or JVC advertise.

And this is true of both V and H rez.

Worse, it is possible create a higher rez camera -- as with the VX2100 -- by raising the anti-aliasing cut-off but introducing serious aliasing.

WE HAVE ALSO YET TO TALK ABOUT THE FACT THE SONY IS UNDER-SAMPLED. THE CCD REZ IS LESS THAN THE FORMAT. We typically sawthis in the early days of Hi8 and DV camcorders.

One Sony rep. claimed green shift increased H rez to 1920. I've estimated 1440 under ideal conditions. But the reality is that on a green & black rez chart there are only 960 H pixels. That's dramatically less than 1280 in 720p.

And we have Panasonic trying to sell a pro 1080i camera that scales-up 1280x720! At least Sony is dong 1080p and downscaling to 720p.

One Sony rep. claimed green shift increased H rez to 1920. I've estimated 1440 under ideal conditions.
Pixel-shifting 960 up to 1920 is impossible. Your estimate of 1440 under ideal conditions is correct. Graeme knows more about this than I do, but the BBC estimates that the most you'll get from pixel-shift is an increase in proportion of the square root of two (1.414 times as much horizontal resolution) and that a factor of around 30% increase is more realistic to expect.

It is a valid technique, it does work, and it works well, but it seems to primarily work when chroma sampling is less than 4:4:4. If you're shooting 4:2:2 or 4:2:0, pixel-shifting seems to be a no-lose proposition. You get more actual resolution, bigger pixels, less noise, better sensitivity, and better dynamic range. If you're aiming for 4:4:4, pixel shift would seem to interfere with the cleanest 4:4:4, but none of these cameras do 4:4:4 so it's really not an issue.

But the reality is that on a green & black rez chart there are only 960 H pixels. That's dramatically less than 1280 in 720p.
Yes, but the reality is that you're never going to be shooting a green & black rez chart. That's a test that's designed specifically to "break" the system. In any real-world situation, the pixel shifting is going to deliver better results overall than the non-shifted system will (excepting the case of 4:4:4).

In real-world circumstances, if even the tiniest hint of red or blue (or white or purple or cyan or orange or any color whatsoever, in any amount) is present, the pixel will be detected by two or three CCDs, and the pixel-shift resolution algorithm will be able to use both and construct a genuine, legitimate luma sample. Only when shooting a single pure color (red or green or blue) will you find resolution lacking, and that's just not going to happen in the real world.

And we have Panasonic trying to sell a pro 1080i camera that scales-up 1280x720!
Yes, and it looks fantastic. 1280 horizontal is all the system needs, because the codec only records 1280 horizontal pixels. Had they put 1920 pixels on the CCD they would have unnecessarily limited latitude and sensitivity, and increased noise, for no gain.

Not sure how 720 pixel-shifts up to 1080. That's a 50% increase, which is within the limit of horizontal pixel-shift, but I am still researching how vertical pixel-shift works.

At least Sony is dong 1080p and downscaling to 720p.
What Sony camera does 720p? I can't think of any Sony product that records 720p. And only the CineAlta F900 & F950 line support 1080p, the rest of their cameras are all 1080i-only, no?

There's a new Sony dsr450ws with a 24p progressive mode for 20k.

http://www.dynamix.ca/products/info.asp?id=373

Actually with some forms of pixel shifting technology resolution can be theoretically expanded infinitely with the resolvability of the glass being the limitation.

Actually the JVC HD10 camera with 1280 x 659 pixels has a resolution of 650 vertical tv lines and a resolution of 700 horizontal tv lines. I have always wondered how 1280 pixels could produce only 700 horizontal tv lines and it was explained that the resolution charts use a circular object with the height identical to the width so according to the 720p format the maximum horizontal resolution can never exceed the vertical resolution which is 720 lines. However if a resolution test chart was used that took into consideration widescreen aspect ratios JVC claims that the horizontal HDTV lines of resolution would be 1244.

So if the Sony camera is producing 960 TV lines of horizontal resolution that is 89 percent of the achievable limit of the fomat with the upper limit being 1080 horizontal tv lines of resolution.

"Actually with some forms of pixel shifting technology resolution can be theoretically expanded infinitely with the resolvability of the glass being the limitation."

I'd love to see a reference for this as it goes right against sampling theory. If you want to be able to expand resolution in this manner you'd need a lot more than just one shift. You'd need many ccds, each shifted from each other, which just isn't going to happen. As Barry pointed out, the only hard and fast mathematical theory of pixel shifting I could find showed that for a factor of n resolution increase you'd need n^2 shifts. Given that we have two shifts, that gives a theoretical max of 1.414 (root 2) and the BBC say in practical terms it's more like 1.3.

As for measured resolution on a bayer chip camera - there's no way they can get 1244 lines (not TV Lines) off a 1280 resolution sensor - it's just not going to happen. As for the FX1, I thought people had measured it at about 800 lines, which sounds reasonable for a pixel-shifted 960 sensor that's lens limited.

Graeme

Well of course you would need more than one shift. To resolve infinite detail you would need an infinite number of shifts. The efficiency of detail gained with each shift would be meaningless because of the sheer number of shifts involved would more than overcompensate for any efficiencey losses gained in the process. What is to be noted is that only one sensor is needed. All that is need is the ability to move this sensor to different locations in order to create more samples of information. For multiple shifts servo motors can accomplish the task. In the first run you offset the sensor half a pixel width horizontally then you take another sample after you move the sensor half a pixel vertically. Then you shift the sensor a quarter of a pixel width horizontally and vertically then shift it an eighth of a pixel width and then sixteenth and keep repeating the process until you are satisfied with the resolution gained. Using this technology one megapixel chips have been used to gain 20 megapixels of information but theoretically there is no limit to the detail that can be resolved other than the limits of the glass.

That's pretty much what the people were doing in the paper I read, but they were using the natural, random vibration of the CCD over time, to get high resolution stills out of it. Seemed to work very well from the pictures they showed.

However, certainly with regards to smaller chip video cameras, we'll reach lens limits long before we need to resort to such measures.

Graeme

As to the JVC HD10 video camera first of all it does not use a bayer CCD but rather a state of the art hybrid complimentary primary color filtration system that has a color resolving power close to the limits of what the MPEG2 compression codec allows. Other than improved low light performance the advantages of 3 CCDs are dubious because any additional information gained by the use of 3CCDs would be thrown away after compression. Many people accuse JVC of producing shoddy products but it was JVCs decision at the time that it would be better not add additional weight and charge the consumer a thousand dollars more for 3CCD technology that is really not needed especially since the camera did an outstanding job of color reproduction and resolved images with 3 times the detail. However had JVC known that the consumer would refuse to buy the camera regardless of its superior quality then 3 CCD technology would have been included even if it meant overcharging the consumer.

So the horizontal resolution of the single chip JVC HD10 camera stands at 1244 HDTV lines or 700 regular TV lines which is very close to the achievable limit of the format. This exceeds the spatial resolution of the Panasonic Varicam which uses a 960 x 720 chip. However the Panasonic has better temporal resolution.

The Sony HDV camera has a horizontal resolution of at least 800 regular TV lines which would equal 1440 HDTV lines of horizontal resolution.

Another example is that the JVC HD100 uses pixel shifting technology to generate the equivalent of 3 megapixel JPEG still images that can be shot continuously.

It doesn't matter what pattern it uses - it's still resolution reduced compared to a 3CCD solution. And all it takes is for you to look at the picture without any added electronic sharpness (oh dear, you can't) to see that it's soft and quite low rez.

I severely doubt that the benefits of 3ccds will be thrown away after compression, and indeed the HD100 looks just so much better than the HD1 or HD10.

Now don't get me wrong - I'm a strong supporter of single chip cameras, but they've got to have bigger chips, and they've got to have at least as many pixels on their one chip as the 3 chips would have in total. Do that, and you'd have a single chip camera that would be very superb indeed.

Graeme

1080 x .7 is just over 700-lines. And 480 x.7 is about 340 which definitely is NTSC V rez. So the Kell seems accurate for both NTSC and 1080i. 720 x .7 is just over 504-lines.

Amazing they both come out to the same number!

Steve, are you trying say that it not matter if footage is progressive or interlaced, Kell factor of .7 is added to pixel count to get actual resolution?

That would make PAL DVD 504x336 lines, HD100 896x504, and HDCAM 1008x756 lines. Sony actually advertises 1000 lines on HDCAM.

I thought that you consider vertical resolution only as "the" resolution of camera.

I also thought 720-60p resolution is same as 1080i-60i resolution. Interlaced resolution must go down with movement, doesn't? Further decrease in resolution due to HDV MPEG2 processing.