hvx200 pixel count revealed

[Thomas Smet]

Originally Posted by Jan Crittenden Livingston

Thomas from reading your response I would say that you missed some very important information in the article. The signal that comes off of the chip is an analog signal and as such is then captured as a 1080P signal with the 14Bit A/D. It is then with the 19 bit DSP that the signal is manipulated into what ever signal it is that you need. It is a true $;@:2 in HD as it is in SD, DVBCPRO50.

As I suspected all along with the release of numbers, people do not understand. They see a number and that is all they get. Look at the second table and look at the numbers that are effective after pixel shift. You need to appreciate the engineering here as it has its impact, moreso than you seem to give it credit.

Please do not take this as being aimed solely at you as I have bee reading this all over, it is more aimed at all that do not fully understand what that article says. You have to read and understand every part of it as all parts are important. The second stage trsansformation is just as important as the initial chip count and the Spatial Offset. The third part is that DVCPRO HD is a 4:2:2 format, DVCPRO50 is a 4:2:2 format. This camera can intercust with Varicam and the SDX900. You can green screen withit and it does not have the 4:2:0 outline that is on every HDV green screen I have seen.

Hope this helps,

Jan

Jan this is exactly why I said it isn't 4:2:0 but somewhere in between 4:2:0 and 4:2:2. Pixel shift works by shifting the green channel to create more luma detail. This does not help the chroma detail in any way except for a slight boost to green tints. There is no way you can suck 960x1080 chroma pixels from 960x540.

Is it much better than 4:2:0? Well of course it is. Any fool can see that we are getting better than 4:2:0.

Even Juan from Reel-Stream has said that using the pixel shift from the DVX100 chips to get HD does not give chroma equal to luma.

In order to have a 4:2:2 and 1080 you would have to have 640x1080 chroma channels. Isn't it true then that through pixel shift the 960x540 luma gets sampled to 1920x1080 and then compressed to 1280x1080? With pixel shift how can all three color channels be doubled the same value when it is a well known fact that pixel shift mainly helps luma?

For keying and 4:2:2 there should be the exact same amount of vertical chroma detail as there is luma detail. If the vertical values differ in any way then it isn't 4:2:2 because 4:2:2 means only a reduction in horizontal chroma detail. Pixel for pixel the vertical value of the luma and chroma are not the same. If no pixel shift was used at all and the 960x540 samples were just upscaled well then yes the luma channels and chroma channels would be exact. Pixel shift however has different weight values on different colors. Maybe if I was shooting a commercial of the Green Giant in front of a green screen (oh god what a nightmare) then yes your chroma channels would also be high in value.

[Steve Mullen]

Originally Posted by David Heath

Jan, Thomas and I are obviously not alone in this, so perhaps it would be worthwhile to have an explanation as to how pixel shift technology can improve colour resolution as well as luminance beyond the native chip resolution? To finally lay this matter to rest?

The PS is only relevant for luma and in THEORY generates 1920 by 1080 pixels. Thus, there are 2X more luma than chroma PER ROW and 2X PER COLUMN exactly as required for 4:2:2 sampling.

Every CCD element in the R, G, B CCDs yields one REAL chroma sample. So there should be 960 horizontally and 540 vertically for RBG. In theory there is a luma to chroma ratio of 2X -- exactly what is required for 4:2:2 sampling. And, there are equal numbers of Cr and Cb (960x540) which is the 2:2.

So -- in theory -- the DSP does yield 4:2:2 video.

There are three areas to consider:

First, for DVCPRO 720 there must be 480 and 720 Cr and Cb samples. Clearly there are NOT enough vertical chroma samples. <<< CORRECTION MADE

Second, for DVCPRO 1080 there must be 640 and 1080 Cr and Cb samples. Clearly there are enough chroma samples. <<< CORRECTION MADE

Third, there are not enough luma samples because PS doesn't generate 2X -- it generates about 1.25X more luma samples. <<< CORRECTION MADE

So -- in reality -- the DSP does not have 4:2:2 video. Moreover, there is not enough VERTICAL chroma samples to meet the needs of the DVCPRO codec. <<< CORRECTION MADE

We only need to look at Adam's tests to see that HVX200 luma resolution is much lower than any other 720p camera -- including the old single CCD JVC. And, we can see from the tests that the HVX200 has 1080 resolution that is both neither greater than its "720" resolution and far less than any other 1080 camcorder.

Thankfully these are worst-case numbers, but the design although very clever, can only have real-world performance that is about 20% - 30% better than these numbers. Best-case numbers are only about 35% - 40% better.

Moreover, as far as I can tell, scaling these data to 1280x640 for 1080i does not increase resolution -- although the 1080 codec can carry additional information so it should look clearer.

However, I suspect the best-case numbers come from more edge enhancement dialed in. So you really need to see if you would accept that much enhancement. For going to film -- you really can't have much, if any edge enhancement.

Lastly, there are those who say the image has noise. I can't say! But, calculations can ADD noise and scaling can AMPLIFY noise so I would not be surprised that noise -- especially for 1080 video -- might be higher than some folks would like.

[William Gardner]

Steve and others,

Question: A raw 4:2:2 format would be able to accurately represent something like a red line followed by a green line followed by a red line and so on, yes? Since the subsampling of the chroma is only in the horizontal direction?

But with the pixel shifting in 1080 mode, there is no way for the HVX200 CCDs to capture red coming and going with each vertical line, nor for it to capture the green coming and going with each vertical line. It would seem that for this case, the CCDs would capture a constant color red field and a constant color green field, resulting in a single constant color image, no? This would seem to be very close to what a native 4:2:0 format would do to the same image, no?

So, am I missing something? Or isn't this fairly close to 4:2:0 for vertical resolution? And for green screen work where the background is truly only a single color, isn't this likely to be closer to the results that could be obtained with 4:2:0?

Please clarify and enlighten, if possible,

Thanks,
Bill

[William Gardner]

Steve wrote:
"The PS is only relevant for luma and in THEORY generates 1920 by 1080 pixels. Thus, there are 2X more luma than chroma PER ROW and 2X PER COLUMN exactly as required for 4:2:2 sampling.

Every CCD element in the R, G, B CCDs yields one REAL chroma sample. So there should be 960 horizontally and 540 vertically for RBG. In theory there is a luma to chroma ratio of 2X -- exactly what is required for 4:2:2 sampling. And, there are equal numbers of Cr and Cb (960x540) which is the 2:2.

So -- in theory -- the DSP does yield 4:2:2 video."

Confusion: I thought that 4:2:0 has 2x per row and 2x per column, and that 4:2:2 had 2x per column but 1x per row (i.e., full vertical resolution).

And if the theory gives you luma at 1920x1080 but chroma at 960x540, then isn't there 4x more luma, exactly as required by 4:2:0, but half as much as required by 4:2:2?

Bill

[Stephen L. Noe]

Originally Posted by William Gardner

And if the theory gives you luma at 1920x1080 but chroma at 960x540, then isn't there 4x more luma, exactly as required by 4:2:0, but half as much as required by 4:2:2?

Bill

I think along the same lines Bill. The extra row is completely interpolated and scaled to 1920x1080. You can't get something from nothing and maybe nearest neighbor or some other interpolation is used.

[John Cordell]

Great thread -- this is exactly a subject I had wanted someone who knows something to discuss.

Steve Mullen: Agreeing with others here, but it does seem your claim of full chroma for 1080 4:2:2 doesn't jibe with my understanding of the chroma notation (which, by the way, is totally goofball to begin with). I thought 4:2:2 meant "skipped" chroma for each pixel as you move across a row horizontally, but if you traverse a column up or down you have new chroma info at each successive spot. Did you get that wrong perhaps?

One thing I've been thinking about is the "bounding area" for each of the final generated pixels. (By that I mean you trace a line that includes all the area under any R, G or B component that's used in the calculation of any given pixel in the final 1920x1080 grid.) The bounding shapes that result from non-pixel shifted 4:2:0 are quite different than the shapes that HV pixel shifted create. The way they overlap is quite different, and the pattern on the HVX does seem at first glance to be "better" in some quantifiable way, but also seems like it would have some source color patterns that it does well at and other not so well at.

[Thomas Smet]

Actually for 4:2:2 and 720p DVCPROHD needs 480x720 chroma samples and for 1080p you need 640x1080. The 480x360 and 640x540 you describe is for 4:2:0 and is exactly what you get with HDV.

[Steve Mullen]

You are all correct! I was wrong about 4:2:2. I've made corrections to my post.

All of this means the DSP can NOT yield 4:2:2 and there is inadequate vertical resolution for both 720p and 1080i DVCPRO50 and DVCPRO HD codecs. (Unless, I'm wrong again.)

Unless Jan can explain how Panasonic can in good faith claim PS yields 1920 horizontal pixels (to get the "4" of 4:2:2) and point-out where the missing vertical chroma samples come from -- some real issues have been opened that I never even considered.

[Dan Euritt]

"Is the 960x540 pixel matrix, progressively scanned at 60 frames per second, good enough for HD after pixel shift and after Panasonic’s new DSP running their newly coined “Advanced Progressive Technology” APT processing? The short answer is YES INDEED, when compared with the HDV 1/3” CCD camcorders."

http://www.coax.tv/

[Brian Luce]

Originally Posted by Steve Mullen

You are all correct! I was wrong about 4:2:2. I've made corrections to my post.

All of this means the DSP can NOT yield 4:2:2 and there is inadequate vertical resolution for both 720p and 1080i DVCPRO50 and DVCPRO HD codecs. (Unless, I'm wrong again.)

Unless Jan can explain how Panasonic can in good faith claim PS yields 1920 horizontal pixels (to get the "4" of 4:2:2) and point-out where the missing vertical chroma samples come from -- some real issues have been opened that I never even considered.

so now we're saying not even dvcpro50 is 4:2:2? wow, the news gets worse all the time. i thought dvcpro50 was already confirmed at 4:2:2.

[Steve Mullen]

Originally Posted by Dan Euritt

http://www.coax.tv/

Very good anylsis, but it has one flaw: The writer, ex of Panasonic, says, "While Panasonic asserts that pixel shift (spatial CCD offset of ½ pixel) can extend resolution by a factor of 1.5, BBC has stated that a more likely real-world factor is 1.33. Not much of a difference, but let us use a middle road factor of 1.4."

Unfortunately, the BEST-case is about 1.4. The WORST-case is about 1.08. The TYPICAL-case is about 1.25. (The BBC is very accurate -- so why did he feel the need to round-up?) Therefore, you need to lower all his calculated values.

But, perhaps more serious -- he like me -- never calculates chroma resolution. Which seems to be where we are now.

His comments about JPEG2000 are very interesting. Sony clearly is not going that way. JVC is clearly staying with ProHD -- which certainly is a pro format -- revealing a bit of bias on his part.

So is he suggesting Panasonic will be switching to JPEG200? He said, "We predict that many professional camcorders will use JPEG2000" at 50Mbps to 75Mbps for recording to a removable hard disk and solid state (P2) memory.

NAB 2007 could offer some serious surprises if he is correct!

[Bob Grant]

The other issue I see with using a smaller number of larger sized elements in the CCDs is the impact on noise. As stated the larger elements should yield less noise at a given light level however the size of the noise blocks will equate to the size of the CCD elements, assuming the noise is coming from the CCDs themselves. I suspect that's the case and might be further compounded by the chips running hotter due to the progressive scanning.
The lack of line averaging due to the progressive scanning is also costing around 6dB in S/N which is probably cancelled out by the larger CCD elements.
Still we're left with large blocks of noise, unless I'm thinking this through wrongly we're effectively getting almost SD sized noise in a HD image.
That's almost certainly further compounded by the cinelike gamma used in this camera, pushing the gain up at the bottom half of the curve is going to bring the dark current noise up with it.

[Stephen L. Noe]

Originally Posted by Bob Grant

The lack of line averaging due to the progressive scanning is also costing around 6dB in S/N which is probably cancelled out by the larger CCD elements.
Still we're left with large blocks of noise, unless I'm thinking this through wrongly we're effectively getting almost SD sized noise in a HD image.
That's almost certainly further compounded by the cinelike gamma used in this camera, pushing the gain up at the bottom half of the curve is going to bring the dark current noise up with it.

Precisely what I think. The noise in the CCD is multiplied and interpolated and becomes evident in the portion of the gamma curve where it has been boosted to meet the cinegama color matrix standard.

I'm not so sure about the chips running hot. I think it was Sony and JVC who's research revealed that the threshold was 960. This is why JVC was forced to use the two chip CCD block in order to get full rez 1280x720. I think Panasonic's chips are well under any heat threshold.

[David Heath]

Originally Posted by Steve Mullen

All of this means the DSP can NOT yield 4:2:2 and there is inadequate vertical resolution for both 720p and 1080i DVCPRO50 and DVCPRO HD codecs. (Unless, I'm wrong again.).

Having read over all the above, I feel a lot depends on definitions. The DSP could be described to yield 4:2:2 if equal weight was given to interpolated pixels as to 'real' ones. That's not the same as identical lum/chroma vertical resolution - and that is what would normally be inferred by the use of the term 4:2:2. To use it to refer to the signal coming from the DSP would seem misleading in this case.

Unless Jan can explain how Panasonic can in good faith claim PS yields 1920 horizontal pixels (to get the "4" of 4:2:2) and point-out where the missing vertical chroma samples come from -- some real issues have been opened that I never even considered.

Well - the term refers solely to a ratio, and the "4" can refer to whatever you like, the second two numbers then defining solely the ratio of chrominance to luminance pixels. Comparisons can only be then made as long as the value defining "4" is constant. Hence to directly compare 4:2:2 of DVCPRO HD (4 representing 960) and 4:2:0 of HDV1 (4 representing 960) would be meaningless.

Originally Posted by Bob Grant

Still we're left with large blocks of noise, unless I'm thinking this through wrongly we're effectively getting almost SD sized noise in a HD image.
That's almost certainly further compounded by the cinelike gamma used in this camera, pushing the gain up at the bottom half of the curve is going to bring the dark current noise up with it.

That raises some very interesting thoughts. I was intrigued a while ago by a remark in this thread http://www.dvinfo.net/conf/showthread.php?t=61086 . (See posts 14 onwards.) Jay Morris described that The noise is in the mid-dark areas. It's hard to explain. Looking at it closely, things that fall off in the dark blacks are clean and things that are properly exposed look good, but somewhere in the low middle is extra noise.

The blacks being clean, but dark greys more noisy sounds very peculiar to me, not like normal noise within a video system when the blacks would be affected as much as the 'low middle'. It's beginning to sound as though the signal processing has enabled higher resolutions (luminance, at any rate) to be obtained from lower resolution chips, but has brought other problems in it's wake.

[Steve Mullen]

Originally Posted by David Heath

Well -- the term refers solely to a ratio, and the "4" can refer to whatever you like, the second two numbers then defining solely the ratio of chrominance to luminance pixels. Comparisons can only be then made as long as the value defining "4" is constant.

The "4" and the two "2s" define the luma to chroma ratio -- 2:1. We know we have 960 RGB samples horizontally. So in order to have a ratio of 2:1, we must have 1920 luma samples horizontally. And, it can't have that.

However, I think the HVX200 really does BETTER than 4:2:2, because the ratio is LESS than 2:1!

If a camera had 4000 luma and 1000:1000 chroma. The ratio would be 4:1. If a camera had 3000 luma and 1000:1000 chroma. The ratio would be 3:1. If a camera had 2000 luma and 1000:1000 chroma. The ratio would be 2:1.
If a camera had 1000 luma and 1000:1000 chroma. The ratio would be 1:1.

The HVX200 has about 1240:960:960 which means the chroma and luma samples are nearly equal. That's good--not bad.

The problem is not the luma / chroma RATIO -- it is the inability to provide either 720 or 1080 vertical chroma samples as there are only 540 chroma samples.

Now about noise. Yes -- running progressive causes a sensitivity loss of 6dB -- 1 stop. But, it also cause a lower S/N ratio!

Do the "larger" CCD elements compensate?