Comments

Bill Ravens wrote on 5/16/2007, 2:06 PM
yes..DV specification is 4:1:1
GlennChan wrote on 5/16/2007, 7:02 PM
1080 HDV is:
1440x1080 luma
720x540 chroma
If the footage is interlaced, chroma performance is effectively around 720x270.

NTSC DV is:
720x480 luma
180x480 chroma

PAL DVCPRO:
720x576 luma
180x576 chroma

PAL DV + DVCAM:
720x576 luma
360x288 chroma
If the footage is interlaced, chroma performance is effectively around 360x144.

They all generally look good. Extremely saturated colors will stress the color subsampling though... the worst probably being green text on a magenta background (blue/yellow and red/cyan look bad too). But it's unlikely that you'd run into that situation.
farss wrote on 5/17/2007, 2:37 AM
A better approach seems to be to capture the HDV and convert to SD Sony YUV which gives you 4:2:2.

Bob.
DJPadre wrote on 5/17/2007, 4:59 AM
now this is one thing that has always confused me Bob..

how does one gain a further 2 points in the colour spacing by a simple straightforward downconversion?

My biggest flag wave with Digibeta was 422.. DVCpro50 the same.. the colour space for both was lovely and rich and i could do a hellofalot of work with it, in virtually any NLE i threw it on..
Also the results spoke for themselves.
I mean if i can save myself afew hundred a pop without the need to hire a HVX everytime i needed 422 i'll take it....

Can someone explain, in lamens terms, how 420 hdv is converted to 422 by downconverting? what am i missing here.. ??
I mean if i can do this i will and it will save me at least 400bux a job on hire fees by doing it myself..

Ive read about this somewhere before but for teh life of me, im still clueless..

If i can indeed convert HDV to 422 SD, then i am definately going back to HDV and rethnking my ideals.. I mean, i always thoght beign 420 hdv would be identical in pal DV 420... to MY naked eye, i see no difference when using a Z1... mind u from memory im prolly a lil off centre as i sold my Z's about 6 months ago..

any info greatly appreciated
farss wrote on 5/17/2007, 5:21 AM
My understand is that you've got enough chroma data in HDV to get more chroma sampling during the downcovert.
Remember that those 4:2:0 or whatever numbers are relative to the number of pixels in the frame. So to do down conversion the 1440x1080 gets converted to 4:4:4 at 1920x1080, nothing real gained so far. But then that gets sampled at 4:2:2 during the downconvert to 720x480, hope you can see how that helps.

In other words for every SD pixel there's 4 pixels in the HDV frame. If you were to look at only the "V" channel you'd see it was very soft compared to the "Y" channel in the HDV however as you're downscaling anyway....

Bob.
DJPadre wrote on 5/17/2007, 5:50 AM
ok..
so what about Pal... ? that woudl be something like 4:2:1.5 (1.5 or something) considering were runnign a slightly higher resolution... yes...

But consider Progressive scan.. if the HDV material is Progressive and your scaling down to progressive, then this is all good, but what about interlaced HDV scaled down to progressive SD
Isnt the temoral field resolution only 540...? i mean in theory its lower than progressive Pal SD by pure numbers, so the scale threshold of the pixels themselves is actually scaling UP to 576p not down..... i mean this is purely 50i HDV to 25p, not 1080p to 576p
Or am i just getting all confused here.. LOL
farss wrote on 5/17/2007, 7:24 AM
With PAL yes, but it's still going to be better than 4:2:0!
However I think the mpeg-2 chroma sampling in HDV is different also to DV with interlaced and progressive, I do stress the "think" part of that too.
Progressive to progressive from say the V1 for sure even with PAL you're going to be doing very nicely.

Interlaced to progressive, hm, might again need to look into how HDV samples chroma in interlaced.

Much is made of the advantages of P over I and technically it's all true. But in practice I'm not 100% certain. In theory a camera running in P can have more V res as it doesn't do line pair averaging as there's no issue with line twitter due to interlacing. Great theory. Except almost no display device can handle that much vertical res as they're almost all interlaced. This really came as a bit of a shock to me, our LCD HDTVs are trying to emulate an interlaced display using "Bob", they're supposed to workout that the PsF fed to them is really P but fail. End result is high vertical res causes all manner of uglies. So, if those magic cameras shooting P are delivering all that vertical res how come they're not having problems too.
My understanding of all this is that SD PAL is limited to around 480 lines res, don't matter if it's P or I, anymore and the displays go nuts. We tried this with 1080p and a test grid of alternate black and white lines. Nothing much outside of a 24" PC display would handle it, haven't tried it with SD but I'd imagine the same thing would happen.

So yes, shooting HDV in 25p on a V1P and downscaling to 4:2:2 should yield some damn fine results, 1080i I really don't know what the exact number would be.
This stuff gets so confusing all I think it's worth looking at the theory is as a path to doing real world tests. In the end that's what really counts. I see the other problem as all these numbers and comparisons are always "all else being equal" but it very rarely is.

Bob.
GlennChan wrote on 5/17/2007, 10:55 AM
In PAL DV, the 4:2:0 is co-sited both vertically and horizontally. You use linear interpolation between the points (i.e. for the missing values, you take the average of the neighboring chroma samples).

This is opposed to interstitial chroma, where you take the chroma value you have to copy it over. Interstitial chroma gives you boxy-looking results. Vegas' NTSC DV and SonyYUV codecs do this (it's more obvious in the DV codecs). Haven't tried the PAL/4:2:0 codecs.
I'm not sure if DV is supposed to be interstitial or co-sited (have read/heard different things, and never read the DV spec).

Perhaps a clearer example:
Suppose you had chroma samples c1 c2 c3 c4 c5 c6
When you subsample, you filter and throw half of them away.

In co-sited (and using linear filtering; other filtering is possible),
c3 = 0.25 * c2 + 0.5 * c3 + 0.25 * c4
c5 = 0.25 * c4 + 0.5 * c5 + 0.25 * c6

On decode:
new c3 = c3
new c4 = 0.5 * (c3 + c5)

In interstitial:
c3 = 0.5 * (c3 + c4)
new c3 = c3
new c4 = c3

If you draw it out, the co-sited chroma is located exactly over a pixel. Interstitial chroma is located between pixels.

1b- In 4:2:2, the chroma is co-sited horizontally. Vertically is not applicable, since vertically you have full chroma resolution.

1c- In MPEG2-based formats, the chroma is co-sited horizontally. This makes sense since it can be source from 4:2:2 sources, so it makes sense to keep the chroma co-sited (and co-sited looks better anyways).

Vertically however, the chroma is interstitial. I'm not 100% sure as to the technical reason. But essentially, this is a different scheme than that used in PAL DV (except for PAL DVCPRO, which is 4:1:1).

2- In practice, there is a huge mix-up between co-sited and interstitial chroma. An ideal system would stick to either format/scheme, but in practice everything is all over the place. Most NLE codecs decode the chroma as if it were interstitial, even if it was encoded co-sited.

By going interstitial (even if the chroma were encoded for co-sited chroma), you get a minor advantage in that you don't get generation loss, though anything going to its first generation (i.e. titles) will look very bad. You also don't get a problem with co-sited where you can see a difference between 1st and 2nd generation stuff (i.e. so you only need to recompress cross dissolves, and leave everything else alone).