HDV problems gone?

[A. J. deLange]

Originally Posted by Ken Hodson

When you up-rez luma it is is unknown territory, being it has to completely guess what information would be there. When up-sampling chroma the outline of where information goes is already mapped out by the luma info.

The statement is true with respect to luma because luma uses the full bandwidth of the sensor in a 3 chip camera. In a Bayer camera (single chip) it doesn't and luma and chroma are both interpolated (approximated) though chroma more so. Because the chroma bandwidth of the channel (our eyes) is demonstrably less than the luma bandwidth we subsample to say 4:2:2. We have irrevocably thrown away information when we do this but it's information that the eye can't use (under normal viewing conditions) so we can get away with it. If we try to throw away more information (by going to 4:1:1 or 4:2:0) we rely on the same principal - that the eye can't use the information we tossed. But in this case, as many can see the difference, the theory falls down and it becomes an engineering trade between quality and bandwidth. There is no way to use luma to accurately reconstruct chroma. If there were the sampling schemes would employ them. There are ways to use luma to approximate what missing chroma might be and these are used all the time. It is the latter that schemes like sharpening, uprezzing and even the processing of data from Bayer masked (single chips) employ

It isn't imagining how things will look. Based on the luma pattern and surrounding chroma, good software can figure it out quite well.

There are no algorithms which can restore the information which has been thrown away. At the risk of being repetitive there are many algorithms which can guess what >might< have been in the missing slot (called generically "interpolators" though a variety of techniques other than simple linear interpolation have been used) and reconstruct a pleaing picture - perhaps even more pleasing than the original to the point where...

I have been more then impressed.

...one is quite impressed. But if one looks closely at such images (i.e. subtracts them from the originals) one finds distortion has increased. But if it looks better, who cares up to the point where the distortions become plainly visible.

Combining the luma info with the half sparce 2:0 of the HDV frame gives a very logical path of where chroma info should go, given good software.

If by this you mean combining Y with R-Y and B-Y to get R and B (and eventually G) that's true but there is no information in Y about where B and R are likley to go except that the values were band limited to the Nuyquist frequency of the original sampling. But as we have no energy above the Nyquist frequency at the current sampling rate (and we don't) we can only make something up above that frequency if we want to increase detail. In nonlinear processing this can be done but it is very risky so the more usual technique is to stuff 0's in to the data steam and low pass filter it thus producing intermediate samples which are below the >lower< Nuyquist frequency. This is actually using the chroma signal itself to tell you "where the info should go". After upsampling other techniques like peaking can be used to, for example increase >apparent< chroma sharpness.

I frames are 5 times per second for 720pHDV and twice per second for 1080iHDV.

True. My mistake.

The theory that much is lost from one HDV frame to the next is much misunderstood. Each frame isn't a degraded form of the frame before it under most circumstances. The differences between each frame is what is recorded, not the whole frame in a degrading fashion, which saves a lot of redundant data. It is not a case of every frame degrading horribly between each I frame as you suggest .

In MPEG 2 each macroblock is examined to see how far it moved from one frame to the next (if at all). In later codecs rotations and distortions are also estimated. The motion is recorded (motion vector) and the frame reconstructed. The difference between the reconstructed frame and the actual frame is then quantized and recorded along with the motion vectors. The quantization level is set according to the fixed available bandwidth. When the camera is focused on a still scene the system works very well. The data load is 2 I frames per second. Theoretically all motion vectors are 0 and all differential data is 0. When things move the story is different. No estimation scheme gives error free estimates so motion vectors are not perfect and thus the differential data load goes up. When it reaches a certain point the system has to quantize more coarsly and data is lost. As with lost temporal samples data lost to quantizing can never be restored though again a good reconstruction algorithm can produce pleasing pictures. Reconstructed (i.e. non I) frames may contain a lot of distortion which is sometimes quite visible depending on the codec. The codec used by my local PBS station, for example, is terrible. If anything moves it's a mess. OTOH the codec on my XL-H1 is excellent. Haven't had it fall down on me to date.

My point about the use of the intermediate codecs is that editing HDV natively is difficult because if you want to cut on a B frame or P frame you have to reconstruct that frame which often involves frames which come after it. This is awkward in the timeline and more so when rendering but I guess it can be done. Seems easier to convert IBBBP.... to IIIII.... for editing and that is, AFAIK, the approach most people use. The larger point I was trying to make is that sampling theorem is sampling theorem. It doesn't matter whether the signal is one dimensional (audio) two dimensional (a picture) or three dimensional (a movie with x, y and time being the three dimensions) Mr. Shannon's wisdom still applies.

["chroma" added in first sentence for clarity]

[David Kennett]

A. J., youve described things very well.

To clarify something, 4:2:2, 4:2:0, or 4:1:1 only reduce color RESOLUTION from the full 4:4:4. The only effect is resolution loss. It does not affect anything related to color depth or. color space. Assuming that we cannot see the color resolution loss, it certainly is reasonable that a chroma key could benefit from higher color resolution.

All the issues about MPEG encoding don't really have anything to do with color sampling. It is worth mentioning that for any given data rate, intERframe encoding will beat intRAframe encoding EVERY time, it just makes editing tougher. Converting to an intRAframe encoder from the MPEG intERframe encoding certainly is an advantage for complex editing. I think some CineForm marketing types have made too big an issue of color sampling.

Here are things I would do BEFORE improving color sampling.
1. Do away with interlaced scanning.
2. Use the best intERframe encoder (AVC or VC1), and use the highest data rate I could.
3. Make camera improvements (sensitivity, resolution, noise).
There are more, but I think you get the idea!

[Ken Hodson]

A.J I appreciate your thorough reply. I don't know if we disagree as much as I thought we did. We have gotten very technical, but at what end. We both know HDV doesn't produce junk on its non I frames, unless pushed in extremes, and even that varies between different cams. Editing native HDV isn't really advantageous. Native HDV at 4:2:0 isn't the best in many post situations. Especially ones that rely on chroma info. Products like AspectHD help very much in this regard, yet do not diminish the image one bit. I don't know what else to say. For me, intelligent upsampling to 4:2:2 has been a god-send in post(not to mention work flow simplification), and I would not be such a fan of HDV without it. Period. I think HDV capture is a great compression all around. But for me that is where it ends and I don't want to work with it from that point on.

David I agree with your points 1-3. Colour sampling isn't the most important part. In fact as I have noted I am very happy with software enhancements in post which do help overcome some of the few deficiencies in post when using a lower chroma sampled format. It is all still such a leap over SD DV that it is all good to me.

[Nick Hiltgen]

To get back to the original question, we used a variant of this box to drop the HDV footage into an avid before there was any support for 24F. It works rock solid (I'm told) with the z1u but we had a few issues with it. When you color correct footage imported through this box, you have the same leway you would if you were able to import the HDV directly into the computer. It's mostly for capturing time code and giving computers that need it deck control over an HDV device. To be completely honest I wasn't too impressed, but I think if they've finally got the bugs out then it might be something for those looking fr this very specific need.

[Ronald Lee]

If I may divert the discussion away from color space for a bit, suppose we shot in a progressive (say 24p HDV) format, rather than interlaced. That should remove the anti aliasing, sure, but if the footage was decompressed, would it show less loss than if we had used an interlaced original source?

Common sense says probably, but in practice?

Also, I don't suppose that artifacting in the blacks is an HDV problem is that 'gone' with this device is there?

The device still shows some promise with the better color space range though, I've taken 4:1:1 SD DV and put it into a 4:2:2 enviroment and there IS more leeway in what you can do with the color correction.

Ronald

[David Kennett]

Ronald,

My guess is that the additional leeway in color correction is your CC software's ability. Do you use the exact same software in both cases? Can you describe the limitations with 4:2:0?

Keep in mind, too, that the chroma information being reduced is only RESOLUTION. We have done nothing to reduce the number of available colors, or limit the richness or saturation of colors. The only possible observable effect would be the smearing of color about a clearly visible luminance edge.

4:2:2 makes 2 color samples of Pb and Pr for every 4 luminance samples (Y) ON EACH SCAN LINE, reducing the color HORIZONTAL resolution to half that of luminance. 4:2:0 ALSO reduces the vertical chroma resolution by half. So, in a sense, 4:2:0 just reduces the vertical resolution by the same amount as the horizontal resolution in 4:2:2. A difference that would very difficult to observe. Interlaced scanning does make the process more difficult, since adjacent lines accur in alternate fields, and there may be motion during that time.

Incidentally, the analog bandwidth reduction of chroma (in NTSC and PAL) was to about 1/3 that of the luminance (about 3:1:1).