Why doesn't MPEG2 use 4:1:1 for interlaced material?

[GlennChan]

Ok, my second post here. :D

Backgrounder:
http://www.hometheaterhifi.com/volume_8_2/dvd-benchmark-special-report-chroma-bug-4-2001.html

The wikiepdia contains semi-accurate information about the 4:1:1 scheme, and pictures of 4:2:0 flaws (with both 4:2:0 interlacead and progressive sampling). [*I noticed some errors in my own work on the wikipedia, my mpeg2 decoder doesn't obey co-siting.]
http://en.wikipedia.org/wiki/Chroma_subsampling

I'm curious... why didn't the MPEG2 standard just go with 4:1:1 for interlaced material in the first place? 4:1:1 doesn't suffer from the interlaced chroma problem / combing artifacts at all.

Does it have something to do with PAL color encoding?

[Alan Roberts]

All digital coding for 625-line systems is 4:2:0. All digital coding for 525-line systems is 4:1:1. This is all to do with the ratio of horizontal-to-vertical resolution of the systems, and little to do with analogue coding of PAL or NTSC. However, the same statement couild be said about PAL and NTSC, where analogue PAL is mostly broadcast on 625 systems, where the line-averaging in the decoding reduces vertical bandwidth by an acceptable amount which would be poor in 525-line systems.

[infocus]

All digital coding for 625-line systems is 4:2:0. All digital coding for 525-line systems is 4:1:1.
Surely DVCPro is 4:1:1 in PAL, and NTSC DVDs are 4:2:0, Alan?

(And welcome, Glenn. A very interesting link you provided on your second post!)

[Alan Roberts]

Quite right, I'd forgotten that one, and it's why DVCPro at 25Mb/s isn't approved of in broadcast, while DVCPro at 50 is (because it's 4:2:2).

[steve]

ISTR that 4:2:0 matched PAL (and SECAM) better because the chroma signal is effectively averaged by the decoding process anyway. D-PAL (delay line) and S-PAL (human eye averaging) designs create a lower vertical resolution by this averaging of the chroma image. Thus it was deemed to be more acceptable to have a similar balanced X Y reduction in resolution for DVD, particularly where the DVD programme might be modulated back onto a PAL carrier for viewing as composite (yuk!) or S-Video.

That's how I remember something but I could be wrong.

Steve

[Alan Roberts]

Nothing wrong there, you're absolutely right.

4:2:0 for 625 systems delivers 720x576 luma and 360x288 chroma.
4:2:0 for 525 systems delivers 720x480 luma and 360x240 chroma.

4:1:1 for 625 systems delivers 720x576 luma and 180x576 chroma.
4:1:1 for 525 systems delivers 720x480 luma and 180x480 chroma.

Clearly, 4:2:0 is better for 625 than is 4:1:1, and would be better for 525 as well in analogue systemjs where the horizontal "pixel count" is closer to 480 than 720.

[infocus]

ISTR that 4:2:0 matched PAL (and SECAM) better because the chroma signal is effectively averaged by the decoding process anyway.
Yes, and the original question should really have had the qualifier "for NTSC". The problem described in the original link I don't believe exists in PAL because of exactly that averaging process.

Obviously 4:4:4 would not suffer any such problems, but when bandwidth is scare chroma subsampling is an obvious way to lower rates. The first step then becomes 4:2:2 - but what if that saving isn't enough? Then it has to 4:1:1 or 4:2:0. My understanding about the first DV formats (NTSC and PAL DVCPro) was that the equipment started off as digital islands in an analogue sea - they would normally either be interfaced to analogue equipment, or other digital recorders via analogue connections.

In such cases 4:1:1 makes a lot of sense, 4:2:0 can easily give cumulative chroma shifts over successive generations in the case above. But the bad side to 4:1:1 is the lack of horizontal chroma resolution. When the image stays in the digital domain, or is not intended for production or multiple generations, (and DVDs are explicitly NOT intended to be further copied!) 4:2:0 makes far more sense - the horizontal gain is worth more than the vertical loss if you like. Of course, the worst of all worlds is 4:1:1 production and 4:2:0 distribution - the end result is effectively equivalent to 4:1:0 - and that's what happens in 60Hz countries if you produce in Dv and output to DVD. Which is one very good reason why 4:2:2 systems are sen as far more important in NTSC countries than PAL ones.

Finally, the linked article makes it very clear that the problem arises as a combination of 4:2:0 and an interlaced system. With a progressive system, it goes away - 4:2:0 simply means a reduction in chroma resolution of 50% compared to luminance both horizontally and vertically. With a progressive system, 4:4:4 is obviouly optimum for production (especially if a lot of processing is to be done), whilst 4:2:0 may be a sensible choice for distribution and economical production. 4:2:2 makes much less sense out of the interlaced world - why have vertical chroma resolution twice that of horizontal?

[GlennChan]

I was specifically interested in interlaced material, where 4:2:0 runs into complications.

If you sample progressively:
You take the whole frame and divide it into 2x2 blocks. Each chroma sample will be taken from two fields, and will describe the chroma for two fields. This is problematic when there is motion, since:
The first row of luma is for time period A.
The second row of luma is for time period B.
The chroma sample for the 2x2 block is for time periods A and B.

What happens is that the chroma will lead and lag the motion. This can appear as an artifact where you have a person gesturing when they are talking, and in sporting events. It appears as grey area along the edge where the movement is.

If you sample interlaced, then you break the frame into two fields. The chroma samples are either for time period A or B. However, this introduces other problems such as comb artifacts and the effective resolution is roughly halved. You can get rid of the comb artifacts, but it takes a good motion detector (you do it in conjunction with de-interlace).

This is why PAL DVCPRO broke with the DV format and went with 4:1:1 instead of 4:2:0. *PAL DV uses a different 4:2:0 sampling scheme than in MPEG2. In the vertical dimension, I believe PAL DV 4:2:0 uses co-siting whereas in MPEG2 it uses interstitial siting.

4:2:0 for 625 systems delivers 720x576 luma and effectively 360x144 chroma. (Sort of. Comb artifacts require processing to get rid of.)
4:2:0 for 525 systems delivers 720x480 luma and effectively 360x120 chroma. (Sort of. Comb artifacts require processing to get rid of.)

4:1:1 for 625 systems delivers 720x576 luma and 180x576 chroma.
4:1:1 for 525 systems delivers 720x480 luma and 180x480 chroma.

You could also consider screen aspect ratio, and pixel aspect ratio.

4:2:2 makes much less sense out of the interlaced world - why have vertical chroma resolution twice that of horizontal?
This is to avoid the chroma 'breaking up' on movement.

[infocus]

Glenn - I didn't word my last post very well. To avoid any confusion, I was agreeing that 4:2:2 had much merit for an interlace system, but much less so for a true progressive one. For the latter (progressive), 4:4:4 is obviously great if the bandwidth is available, but if not, then halving the chroma resolution horizontally AND vertically seems to make sense - 4:2:0.

[Alan Roberts]

In the early days of digits, when digital kit was analogue in/out (i.e. early 1970s, for transmission links to transmitters), 4:1:1 made a lot of sense because the chroma bandwidth of analogue composite broadcast didn't need more. But it started to get used in production more and more, and then the limited bandwidth of 4:1:1 chroma made colour keying very poor, so 4:2:0 was adopted as a better compromise.

Clearly, 4:2:2 is better, and 4:4:4 better still, but bandwidth (or more strictly, data rate) is always a limiting factor, and we can't afford 4:4:4 even for top-end broadcast acquisition.