Robert Harris on The Bits - 8/3/04 column - OFFICIAL THREAD (1 Viewer)

[TedD]

quote:(It would probably not be correctable if it were due solely to 4:2:0 chroma subsampling.)



But since it is correctable in the decode chain, which is largely complentary to the encode chain, it should be correctable there as well.





Yes, but it is very slight. Unfortunately, I cant correct in finer than a single pixel increment.

Another common cause of color fringing is the mis-registration of the color separations that are required if the original negative color has degraded, or in the case of IB Technicolor. However these issues usually show up as a single color fringe. They can be highly variable in location and duration because of the differential shrinkage of the 3 B&W source elements.

Also, I didn't admit to this earlier, but I have had my share of MVS and VTAM as well as a large helping of AS400 in my past
Ted

 

[ChristopherDAC]

Wow, TedD, that's an impressive home theatre you have there! Any way I could get an invite to come up to Southlake some time and play some of my anime LDs through your S9?

Truth is, I'm not really a digital video geek -- I'm a video geek, or maybe just a geek, period. Here's a little explanation of the theory behind the Discrete Cosine Transform, which the Motion Picture Experts Group video compression system implements.



Fourier's theorem postulates that, basically, any arbitrary function of a parameter which can be drawn on graph paper [it has to have a non-infinite number of discontinuities, be bounded along its length by sufficiently large multiple of the exponential function, &c.] can be resolved into the sum of an arbitrarily great number of sine and cosine functions of the original parameter of varying frequencies and amplitudes. A video signal in a television system is such a function, being a function of time continuous everywhere and bounded by -0.7/+0.3 V typically, and can be so resolved. Owing to a quirk at the intersection of mathematics and physics which is video image scanning, the number of periodic functions this signal can be resolved into is not infinite: in fact it is strictly limited. Firstly, there is a definite maximum frequency, equivalent [by Nyquist's theorem] to the reciprocal of: one half the number of [visible and hidden] picture elements per line multiplied by the line time. For broadcast NTSC this is 4 MHz as per FCC ruling; for 480*704 DVD in the region of 5.5 MHz. Secondly, the spectral distibution is noncontinuous: the only frequencies which actually appear are [for interlaced video, from a live source; progressive scan and film source chanfge the setup slightly] even multiples of half the line rate plus or minus small integer [including zero] multiples of the frame rate. Incidentally, this distribution led colour-television designers to sneak in a second signal, modulated on a subcarrier which was an odd multiple of half the line rate, to transmit the colour information; unfortunately the hyperfast spectrum analyzer which would be required to perfectly extract this second signal has never been implemented.

Now, the upshot of this is that one can completely specify a video picture not only with the time-varying signal, but also with a set of ordered pairs in the frequency-magnitude plane. Further consideration establishes that every different image will have a different set of frequency/magnitude pairs, and that closely related images will have similar sets. The mathematical function describing how frequency is correlated to magnitude in a given image is called its spectral power distribution. It is not hard to see, if you have ever seen a demonstration of how the Fourier transform works [e.g. the adding up of sine waves to make a square wave] that there will be a few terms, mostly low-frequency, which will determine the overall "shape" of the image, and that these will be the terms with the largest amplitudes; the terms of lower amplitude will specify only the "texture" -- that is, whether edges are sharp or rounded-off, whether large flat spaces are blank or mottled, and so on. Sooner or later, if you had to write down all of the frequency terms in an image by hand, you would hit upon the scheme of writing down only those which were above a certain threshold of magnitude, on the principle that the ones you left out didn't really matter enough to be missed when the image was reconstructed. This is, in effect, what lossy MPEG compression does; the use of digital sampling, 2-dimensional rather than 1-dimensional spatial frequency matrices, blocks, macroblocks, interframe compression, and the like changes the details but not the principle.

Now the bitrate basically determines the number of frequency-magnitude pairs per image which are recorded. From the above discussion it should be clear that one basically sets the threshold of "least magnitude which will be recorded" based upon this number, and sets all the other pairs to zero, omitting them. The spectral power disribution now comes into play: if there are a few high-amplitude frequency components and many small to very-small amplitude components, one can get in the big ones and then as many of the small ones as will fit, and have a good-looking image. If, on the other hand, there are a large number of components right about the same amplitude, an amplitude great enough to make a visible contribution to the picture [film grain can come into play here], some of them are going to get cut and others are not, and the end result will be a very noisy picture. Edges [made sharp and regular by those low-amplitude components] will have rings around them from the first few low-amplitude components which got left in, planes will exhibit moire pattern or mottling, and so on. This is why two images with different "feels" will compress differently.



The end result is that picture quality is a synergistic function of bitrate and spectral power distribution in the original image. To give some idea of the practical results of these considerations I have assembled some image files: Windows .bmp bitmaps, equivalent to uncompressed video, and Joint Photographic Experts Group .jpeg compressed image files, which implement DCT.



Two bitmaps, 512*384 elements, 24 bits per sample, 589 878 bytes:

one with a high degree of order

the other rather random



Now, at JPEG 100% quality setting:

20.4 KB

17.4 KB

Notice that the edges are fuzzy if the image is blown up.



Now, at JPEG 10% quality setting:

4.60 KB

4.09 KB

Note that the orderly image is still legible, while the random one is smeared beyond recognition. I had hoped to use images of the same bitwise size, but as it happens I was not able. I consider the larger size of the "orderly" files a result of the rounded edges of the original font vs. the square dots in the "random" image.





Did that help anybody? Or was it so much gibberish?

 

[TedD]

ChristopherDAC,



You have a PM.



Ted

 

[Michel_Hafner]

Concerning the jpeg: it depends also how smart your jpeg

encoder is and what options you have. Take the original

and jpeg it in Photoshop at highest quality progressive

with several scans. It's still only ~25000 bytes but the

fuzzy edges are very much gone. That's not amazing since

the picture is basically red and black pixels which can

be losslessly and efficiently compressed with run length

enoding only (only two colors in the picture). For this

picture any result with fuzzy edges is simply very suboptimal

compression. Natural pictures are not like this so the

cosine transform approach makes a lot more sense there.

Concerning the mpeg: Most pictures that are jpeged are

difference pictures, not the originals, which makes

the result more complex and depend on the quality of

motion estimation from picture to picture. The more redundancy there is and you can find the better the picture

at the same bit rate. That is the reason why a supercomputer

could invest a lot of number crunching in analysing a

sequence and squeeze maximum quality out of it with a given

bit budget.

 

[Eric Stewart]

Ted,

On the topic of chroma misalignment vis-à-vis luma, a.k.a. chroma shift or chroma delay, leading to “color haloes around faces”:



There is indeed a possibility of variable color fringing from IB (imbibition) Technicolor matrices, or similar separation masters in other color systems, when the separate elements used in the film restoration and/or transfer have differentially shrunk over time. Robert Harris has of course been eloquent on this subject in his columns.



Leaving that possibility aside momentarily, you’ve shown that the “color haloes” problem can in fact, as you say, be corrected in the decode chain, at least in single pixel increments. That suggests it can be corrected in the encode chain.



I don’t quite know if it could be so nicely corrected if it were due solely to 4:2:0 chroma subsampling. Maybe yes, maybe no.



My information (see Poynton, p. 90ff.) is that MPEG-2 4:2:0 chroma subsampling is done in a way that nominally “cosites” each “fat subsample” that results from the subsampling. “Cositing” means that each chroma subsample is nominally centered at a horizontal position shared with associated specific luma samples. Its center is thus aligned in the horizontal direction with those luma samples’ centers.



The opposite of cositing is “interstitial siting,” in which subsamples’ centers align with points in between those of associated luma samples.



Here’s how the MPEG-2 4:2:0 chroma subsampling works. Start with a 2x2 array or “quad” of Y’ or luma samples. This quad is initially associated with a 2x2 quad of Cb chroma samples and also with a 2x2 quad of Cr chroma samples. The following process applies separately and equally to both Cb and Cr, so I will simply refer to “chroma.”



The four chroma samples in their 2x2 quad have to be reduced to one “fat subsample.” This is done by digital filtering.



Specifically, the simplest imaginable digital filter that could be used for MPEG-2 4:2:0 chroma subsampling (see Poynton, p. 93) considers the four chroma samples that match up with the four luma samples in a 2x2 quad. It also considers the two chroma samples immediately to the left of that 2x2 chroma quad, for a total of six chroma samples in a 3x2 array, which is the input to the filter.



The filter gives the two middle chroma samples in each row of this array a weight of 1/4, and the other four chroma samples a weight of 1/8, in coming up with a weighted average which becomes the output value of the “fat subsample.”



By its design, this simple averaging filter ignores chroma samples immediately to the right of the 3x2 array. It thereby sites each subsample one-half pixel to the left of where it “should” be vis-à-vis the associated 2x2 array of luma samples.



The subsample is thus cosited (aligned) with the two leftmost samples in the 2x2 array of luma samples. That is, it is not sited “interstitially” between the leftmost and rightmost samples, as it would have been for the sort of 4:2:0 subsampling done for JPEG/JFIF or MPEG-1.



The “fat subsample” is, however, sited interstitially with respect to the top two and bottom two luma samples in the 2x2 luma quad. That is, there is supposedly no chroma displacement up or down with respect to the luma quad.



Poynton quickly adds that this simplest imaginable “averaging” filter produces pesky aliases when applied to high-quality motion video. It needs to be, and is, supplanted by a more sophisticated filter such that each chroma subsample will “take contributions from” several surrounding chroma samples, not just the two neighbors to its immediate left.



Even so, I assume that each subsample remains influenced more by its “left-neighborhood” than by its “right-neighborhood.” Such filtering would tend not only to blur a sharp vertical edge transition in the chroma component, but also to “delay” the “peak” of that transition with respect to the associated luma transition’s “peak” (assuming a left-to-right horizontal scan). That is, the “peak” of the chroma transition would lie slightly to the right of the “peak” of the luma transition on the display screen.



In the space between the “luma-edge peak” and the delayed “chroma-edge peak,” the value of the chroma — its hue and saturation — would still be somewhat determined by hues and saturations that derive from pixels that lie well to the left of the “chroma-edge peak.”



That alone might be enough to explain color-delay “haloes” along, say, a neck line or cheek line, since these lines are in effect sharp vertical edge transitions.



Notice that I said earlier that each “fat subsample” is nominally centered one-half pixel to the left of where it “should” be. This was interpreted as a shorthand way of saying that the digital filter that derived the subsample weighted neighboring samples to the left of “this” subsample left more heavily than those to its right. Result: an effective “chroma delay,” as well as some amount of “chroma blur.”



So, now, that’s what the encoder does. What does the MPEG-2 decoder do with these chroma subsamples? I assume that in the simplest case it simply ignores the nominal leftward cositing of each subsample and assumes the subsample to be sited interstitially with respect to the associated quad of luma samples — i.e., it aligns the subsample with the middle of that quad.



That in effect slides the subsample rightward roughly one-half pixel – making the onset of each already-blurred-and-delayed “chroma-edge peak” even more delayed, with respect to the “luma-edge peak.”



Such behavior on the part of the decoder could further explain “color haloes” along, say, a neck line or cheek line.



However, it is hard to see how this type of explanation alone could explain chroma misalignments in a downward, vertical direction (see my example from A Room with a View). Since MPEG-2 4:2:0 chroma subsampling nominally sites the subsamples at the same vertical position as the associated 2x2 luma quads, how could the same sort of explanation account for downward chroma displacements?



But it occurs to me that it’s hard to find an example of an apparent downward shift that could not also be interpreted as a rightward shift. That is, it’s hard to find a ruler-straight, absolutely horizontal chroma edge.



For example, the A Room with a View screen shot has a color halo at the top of Daniel Day-Lewis’ forehead, at his hairline. But the hairline is curved and a bit tilted, so is this a downward chroma shift or a rightward shift? I note that the off-color halo does not extend all the way over to the right end of his hairline, where the lighting creates a dark shadow between hair and skin and then a relatively bright glare coming off his temple area. If downward displacement were occurring, there ought to be a halo between shadow-line and glare. There also ought to be a halo below his left sideburn in the less-zoomed-in image, and there doesn’t seem to be. See the bottom examples at:



[url=http://home.comcast.net/~epstewart/Test_Web_Page_1.htm]http://home.comcast.net/~epstewart/Test_Web_Page_1.htm [/url]



If you have this DVD, Ted, can you check to see whether it can be corrected with just a leftward movement of the chroma plane?



And, thanks again for all your interest and input.

 

[Eric Stewart]

Christopher,



Your explanation of the DCT (Discrete Cosine Transform) used in MPEG-2 compression was helpful to me. Thank you very much.

quote:If ... there are a large number of components right about the same amplitude, an amplitude great enough to make a visible contribution to the picture [film grain can come into play here], some of them are going to get cut and others are not, and the end result will be a very noisy picture. Edges [made sharp and regular by those low-amplitude components] will have rings around them from the first few low-amplitude components which got left in, planes will exhibit moire pattern or mottling, and so on. This is why two images with different "feels" will compress differently.

I'm especially interested in the edges that develop rings around them "from the first few low-amplitude components which got left in." I'd now like to ask you to apply that same concept to the use of the DCT in compressing specifically the chroma (i.e. Cb and Cr) components of the video.



Most of us probably visualize what you said as applying basically to the luma or black-and-white component (Y') of Y'CbCr video. The "ringing" probably looks just like "edge enhancement" in the Y' signal.



But wouldn't the same exact kind of thing tend to compromise Cb and Cr, too?



Or, if not, why not?



Thanks ... and keep up the good explanations.

 

[Eric Stewart]

And to all concerned:



I hope all my nattering about "chroma haloes around faces" hasn't obscured one of the major thrusts of this thread, which is to try to figure out why recent Miramax DVDs like Cold Mountain have the sorts of problems Mr. Harris spoke of in his recent column.



It seems the problems with that movie in particular lumped into two categories:



(1) possible poor compression or over-compression



(2) edge enhancement



Based on what Christopher said in post # 102 about MPEG compression using the Discrete Cosine Transform, it looks to me as if the "edge enhancement" issue could have been wholly or partly explainable as "edges made sharp and regular by those low-amplitude components [that also] have rings around them from the first few low-amplitude components which got left in."



This would also account for other sorts of messiness or blurriness in the encoded output, Christopher says, depending on the idiosyncratic "feel" of the image input to the encoder. So maybe the two problem categories are really one!



Mr. Maloney, are you still there?

 

[TedD]

Eric:



Don't apologize for the chroma alignment side trip.

It's all part of the bigger question: Why do some DVD's look far poorer than they have to? The same thought applies to mis-framed and excessively zoomed DVD's. Why???? Enquiring minds want to know, and in this case some of those minds may be in a position to do something about it.

Sorry, but I don't have "Room With A View". That's why I didn't post anything on it.



Also, on the chroma shift issue, we need to keep in mind that most DVD's don't have this issue, at least not to point of being as visible as these two examples. Also, remember that "The Producers" was off 2 pixels.



One other gotcha If you are using an HTPC with an ATI Radeon card and using the overlay surface:



The Radeon card's chroma plane can decouple from the luma plane and shift as many as 3 pixels if any edge of the overlay surface is positioned beyond the edge of the desktop surface. This might happen if you are zooming the image with Zoom Player or TheaterTek to eliminate black bars on your screen. Approximately every 4 increments the chroma and luma alignment will be OK.



This usually manifests itself by a womans lips with lipstick moving off of its proper location on the face. Solic color filled letters in a title are another obvious area to check.



On a brighter topic:



If anyone is interested in a reference or close to reference DVD from WB, check out "New York Minute".



Not exactly serious cinema, but then the Olsen twins are pretty easy on the eyes

Now, we have have a comparable example of reference to near-reference from Miramax: "Ella Enchanted".



Not exactly serious cinema either, but it shows that Miramax CAN, in fact, produce a high quality DVD.



Ted

 

[David Grove]

Gibbs Phenomenom







There have been some very nice explanations of Fourier Transform and sampling principles. Since we're going down that path a little, here is a little more "grist" for the mill.



This post is just to contribute a little more on Gibbs.



ChristopherDAC didn't mention it explicitly, but his discussion implied it. Gibbs is part and parcel of Fourier Theory.



As ChristoperDAC and others have so well explained, smooth waveforms can be represented as a sum of sine waves. The problem comes when you want to represent a waveform that isn't smooth. "Not smooth" means a waveform that has sharp corners. For example, a square wave.



Even then, the mathematics says you can get as close as you want ON AVERAGE, which is actually pretty good. Unfortunately, "on average" isn't quite the same as getting as close as we want ABSOLUTELY, which is, well, perfect.



For smooth waveforms the difference between "on average" and "absolutely" doesn't really matter that much. But, for waveforms with sudden transitions (discontinuities), it matters, as the link below illustrates.



Here is a site that illustrates the Fourier Theory Gibbs Phenonenon.



[url=http://www.sosmath.com/fourier/fourier3/gibbs.html]http://www.sosmath.com/fourier/fourier3/gibbs.html[/url]



IT IS NOT NECESSARY TO UNDERSTAND OR EVEN LOOK AT THE EQUATIONS TO APPRECIATE THE ESSENCE OF WHAT IS HAPPENING.



I posted this site because of the animated display of the waveform, which speaks qualitative volumes.

In the illustration, the original waveform is in grey. See how one can start with a single sine wave, and then start adding additional sine waves on top of the first to approximate the original waveform.



Observe how, as additional sine waves are "added in", the result gets closer and closer to the original square wave. This is what the Fourier Series is all about : One can find a proper and sufficient set of sine waves that sum to almost any given waveform.



The practical value is that, to reconstruct the original waveform, it is necessary only to tell someone what sine waves are required. As the frequency of the sine waves to be included increases, their amplitude usually decreases. So, eventually we just say we're good enough, and ignore the rest. So, it is not necessary to tell someone each and every value of each and every point of the original waveform, we just need identify the significant sine wave components.



In other words, it is quicker and easier to say, "Give me 3 cups of a 60 Hz. sine wave, 1 teaspoon of a 120 Hz. one, etc." than to laboriously, point-by-point supply individual values of every point of the entire original waveform. (This brief discussion omits phase for pedagogical resons).



An analogy might be that it is easier and cheaper to send a recipe than to actually send a cake.



So, we are now able to communicate to someone an arbitrary waveform, but can do it much faster by communicating the sine wave components than actually sending the individual points of the actual waveform. Voila! We have achieved data compression.



These very same principles apply in higher dimensional waveforms, such as images, too.

After all, that's the whole point-- to pack a whole movie onto the limited space of a DVD. (There is insufficient space on DVD to hold every pixel of every frame.)



Now, back to Gibbs.



A professor I once had used to wear a baseball cap with "TANSTAAFL" emblazened on it. It stood for "There Ain't No Such Thing As A Free Lunch." In this case, a mathematical consequence of getting close enough "on average" is the overshoot and "ringing" at the sharp transitions (discontinuities). This overshoot and ringing is called the Gibbs Phenomenon (named after the person who studied and described it mathematically). We'd prefer to get close enough "absolutely", which would avoid the whole Gibbs thing, but as long as 2+2=4, we're stuck with Gibbs when using Fourier techniques.



Take another look at



[url=http://www.sosmath.com/fourier/fourier3/gibbs.html]http://www.sosmath.com/fourier/fourier3/gibbs.html[/url]



Note that the overshoot and ringing get more and more "squished", and the summed waveform becomes more and more square looking, as more and more sine waves are included. The waveform looks more and more like a square wave. On average. This is good. On average. But, also note that the maximum overshoot, even though increasingly squished together, remains at the same peak value. Therein is the rub of Gibbs. That is, the height of the overshoot NEVER gets any less (and NEVER goes away), no matter how many sine waves you include in the sum.



One of the fascinating things about Fourier analysis is that the mathematical principles apply in both the untransformed domain (the original waveform, or, in our area of interest, an image), and also in the transformed domain (frequency plane). So, Gibbs also applies to discontinuities in the frequency domain.



One might ask, "So what?" Well, in our "recipe" we may have retained all the sine waves below a certain frequency, and ignored those above a certain frequency. Many folks use the phrase "brick wall filter" or "perfect filter" to describe that-- a constant value, followed by zero. But, whatever one calls it, it sure sounds like a "square wave", doesn't it? (Although this time in the frequency domain.) Well, it is. And now we know what happens with square waves-- Gibbs.



In other words, a "brick wall" filter in the frequency domain introduces ringing back in the original image when the sine waves (DCT coefficients) are used to perform the "reconstitution" of the image.



There are ways to reduce the effect-- for example changing the shape of the "wall" from sharp edges to some other shape with rounded edges (instead of a brick wall, use a haystack), but that has it's own set of pros and cons. Alternatively, one can filter in the original untransformed domain, but that, too, has its own set of pros and cons (for example, in many cases, such filtering is computationaly impractical).



Bottom line: TANSTAAFL



Studios have to choose their priorities and compromises.



Some choices may be more effective, but more expensive, than others.

DG

Edit: See later post. It isn't all Gibbs' fault.

 

[Mike Maloney]

Hi all. -- I'm back.



I have been swamped at work and it has not been easy to catch up on this thread. I actually chose a good time to be away from the forum, since the discussion of theory explains what I see in day-to-day experience.



The main thing I hope readers will take away from the discussion is that it is not easy to make pleasing bitstreams. MPEG compression is a compromise. Why use a brick-wall filter at all? Because not using one will cause other artifacts to appear. The highest fidelity (for the video) bitstreams would be made by limiting the audio to one stereo stream, encoding the movie at the highest bitrate possible and letting the movie run over however many discs that would be required. Is that what you really want?



One thing that has not been brought up is that the MPEG2 specification is now ten years old. What were you watching television on ten years ago? Most likely it was a CRT display device of some kind, and it certainly did not display non-interlaced video. As display devices get better (and bigger), and as you (the consumer) become more sophisticated, expect to see more MPEG artifacts. The studios are probably pretty happy to see this discussion happening so early in the DVD life-cycle; you all are ready to buy HD-DVDs. (I just hope they dump MPEG and switch to wavelet compression, but that's another discussion.)



My job is to make the best looking bitstreams I can.

To that end, this discussion has been helpful since I understand what is driving y'all crazy. Unfortunately, after all my testing, I didn't find a silver bullet. There is no combination of control settings that will make a perfect bitstream. But I do have an even better eye for what bothers all of you who have posted here and I will make better bitstreams because of it.



Cheers,

Mike



P.S. - The reason for the note of finality in the last paragraph is that as I said in the first paragraph; I'm swamped and I'll have little time to participate for the forseeable future.



M

 

[ChristopherDAC]

Mr. Maloney: Thank you. That is what we will all profit from -- DVDs being produced by people who know something about what should come out of the process, about what is going on in the process, and what to do about it, who work on a case-by-case basis to get what looks best.



All:



Please note: my use of JPEG images was meant to be illustrative only.



Having now experienced TedD's home theatre, I have to say that he has equipment which allows him to see what is there; his setup is capable of mirror realism from good material. This is the sort of experience everyone should have at least once, hopefully without having to use some kind of mind-altering substance

. I will endorse his descriptions of the phenomena he is observing, with one reservation: under the circumstances the unusual picture characteristics in question make the picture look less bad to my eyes than simply alien. Some of the distortion I observed made the images look like nothing I am aware of on the Earth. If these DVDs were billed as imported fresh from the planet Thrakkarzorg, or as containing surrealist films, I would not be disturbed by the contents at all.



I want to reiterate my conviction that the picture characteristics of the source material, both the obvious and the subtle which may be described as "atmosphere" or missed totally [particularly on smaller displays], will play a large part in determining the resemblance between the original film and what comes out of the MPEG decoder onto your [or at least TedD's

] screen. Nevertheless there is a great deal of preprocessing and filtering which could be done [preferably, to my mind, after the video transfer, leaving same untouched] to make an image more amenable to the MPEG transmission channel. Conversely, there is apparently a considerable deal of processing, either deliberate or inadvertent [particularly in the operation of the encoder] which has absolutely the opposite effect, being in fact an engine for producing bad DVDs. The lack of standardisation or benchmarking of the functionality of the transfer-to-pressing chain makes it extremely difficult to track down what needs to be done in any specific case, and this is undoubtedly only complicated by the inadequacy of studio equipment. If we the public wish to enjoy good and reliable quality on DVD [as in other parts of our lives] we need to become involved in the process!



And yes, differential ringing in the two chroma components could certainly account for bizarre colour shifts. That may not be the actual explanation in this case, but it is a danger in any colour television system -- in fact, the NTSC system includes by design a region of frequencies in which detail is rendered along one chroma axis but not the other, corresponding with the orientation of human visual sensitivity, so all kinds of chroma differential atrefacts are made possible. This is less of a concern with two components of the same bandwidth but it is certainly still a live one.

 

[Michel_Hafner]

Well, Mr. Maloney, you still have not told us if the EE and

ringing was added by your MPEG encoding or if it was already

on the master you got. In any case it should not be there

and is avoidable to a large degree, and not only at 9 Mbit/s

but with average bit rates quite some lower. There are

countless DVDs that can testify to this.

 

[Eric Stewart]

David Grove,



"Muchas gracias" for your essay on the Fourier Transform and the Gibbs Phenomenon, a.k.a. "ringing." The "moving graphs" on the web page you referenced are a "must see." They show a "square wave" -- which is much like a pair of sharp edge transitions in an image, say a dark telepone pole against a bright sky -- being approximated by a sum of sine waves ...



... but because of inherent overshoot at the edges of the square wave, what results is a sort of "goalposts" effect. Also illustrated is something called a sigma-approximation which can minimize the Fourier overshoot -- see the second graph. The math is frankly beyond me, but wouldn't you say using this sigma-approximation would be a good idea for video encoding?



But, then again, I'm not clear on exactly what the relationship is between the Fourier Transform and the Discrete Cosine Transform used by MPEG? Can you enlighten us on that point? Do the phenomena of which you speak carry over to DCT?

 

[Robert Harris]

I'm told that the problems were not a part of the element turned over to the post facility. They were a part of the element produced and delivered by them after their work.



RAH

 

[Eric Stewart]

Mike Maloney,



I'd like to express my personal appreciation for your contributions to this thread. When you are less swamped or would just like to pop up for air, I hope you'll rejoin us.



And when you do, it would be nice if you could answer Michel Hafner's concern point blank: "if the EE [edge enhancement] and ringing was added by ... MPEG encoding or if it was already on the master ... " of Cold Mountain.



There's also a third possibility: both of the above. There very well could have been EE on the master, but then it may have been exacerbated by the ringing associated with David Grove's contribution to this discussion ...



... a "Gibbs phenomenon" associated with MPEG encoding, possibly triggered by an actual or effective brick-wall filter.



It's almost Zen: when life presents a seeming either/or, look for a possible both/and ...



Cheers,

 

[Eric Stewart]

Ted,



I thought you might like to know ...



... I've discovered that the North by Northwest "chroma shift" problem is not evident when I play the DVD via the DVI interface between my Samsung DVD player and my Samsung display.



Using this interface, the player apparently does something akin to your "X-1" chroma plane shift.



I think it's actually "a bit" more complicated '

' ... since that faint red edge your correction introduced at the left side of Cary Grant's face didn't show up.



Of course, I discovered this the day after I sent the A Room with a View disc back to Netflix ... '

'



Cheers,

 

[Eric Stewart]

RAH,

quote:I'm told that the problems were not a part of the element turned over to the post facility. They were a part of the element produced and delivered by them after their work.

Did you mean by "the element turned over to the post facility" the one prior to the so-called HD master of which Mr. Maloney spoke, and from which the HD master was derived? I assume in the case of Cold Mountain that this was not a film element but a video element, right? A digital interpositive, no?



Or did you mean by "the element turned over to the post facility" the HD master itself, from which the "compression master" was derived?



If we're very, very patient, we'll get to the bottom of this ...



Cheers,

 

[Eric Stewart]

Christopher,



Amen to your emphasis on the complexity of the "transfer-to-pressing chain."



And let's not forget the "encoded-bitstream-to-display chains" in our homes. I mentioned in another post that one such chain in my home corrects the "color haloes on faces" problem I've been beefing about ... while others do not.



I'm green with envy at your having been able to witness Ted's "encoded-bitstream-to-display chain" in person.



Cheers,

 

[Cassy_w]

The Question of Edge Enhancement...



Looking at some recent Disney/Miramax films on StarzHD, which are being shown as widescreen "Upconverts" (not actual 1080i) and comparing them to the DVD's, it's clear they have the exact same problems of Edge Enhancement, but without the loss in detail and compression issues of DVD.



However, I saw a few of these films in true High Definition on D-VHS (recorded in Canada) at a friend's place and they had no Edge Enhancement at all, but still appeared to be the same transfers.



What does this tell me? These films are transferred into High definition and look great. They are then downconverted for DVD and StarzHD and that is where the EE is added. From there the downcoverted master is compressed for DVD and released as the abominations we've seen in COLD MOUNTAIN, etc.



Please forgive my non-technical nature, guys. I am not an expert. Just a chick making oberservations with her husband. Please correct me if I am wrong.

 

[Robert Harris]

To the best of my knowledge, the CM element was taken directly from the digital files and a color-corrected tape without any of the problems mentioned delivered to the post facility.



Seemingly, the only way to get to the bottom of this would be to deliver the same element which was turned over to Mr. Maloney's facility to the top post houses in town, and see what each of them can come up with as a final result, pressed to a DVD ram or other carrying element.



Different post facilities use different software.



It is entirely possible that no matter how much effort Mr. Maloney and his co-workers put into a job, the result may continue to be problematic, not based upon their personal skills, but rather, upon the equipment available to them.



In all too many cases work goes to the lowest bidder turning out acceptable work. Is this the case here?



I don't know.



Should the post people at Disney and Miramax truly wish to step up to the challenge, this would be the obvious route to go.



As it is, taking extraordinary work by some of the world's top directors, DPs and special effects people, and turning out an unacceptable DVD is well...



unacceptable.



A contract for future work should go the post house turning out the best final product on a timely basis, and without going far astray from extant budgets.



So the question is now out there.



Is there anyone at Miramax or Disney with a desire for quality and the guts to move a test forward who has the corporate capability of doing so?



Alternatively, will corporate slam the door shut on such a test before it has a chance to be opened?



In the end the concept is very simple.



Film doesn't lie.



Neither does video.



When all of the excuses, discussions and hyperbole are set aside.



What does the DVD look like?



RAH