AVCHD video at higher bitrates

[Dave Blackhurst]

Guy -
This was the reason compression was originally NECESSARY - it wasn't THAT long ago that I was speaking with someone about editing/producing AUDIO with a computer and was told there simply wasn't enough horsepower or capacity at that time to do it at a cost effective price point...

Editing DV was I'm sure a challenge at first, and I certainly remember the groans from my comupter the first time I tried to edit HDV... and the resultant upgrade. AVCHD taxed the machine I had the first time I tried to edit THAT, so of course an upgrade became necessary...

BUT, back to my point, if one has virtually infinite storage/bandwidth/processing/graphics display power, in theory no compression would be needed, as you could simply display a representation of ALL the 1's and 0's.

BIG IF! And not something that one can expect unless your budget is under the "black projects" military funding... As a practical matter, content and programs have an amazing ability to expand to fill the latest hard drives and use the full capacity of the latest hardware, but MUST be able to run on reasonably up to date systems. I had to upgrade the in-laws, as their "adequate" machine (for e-mail and docs and pictures to some extent) totally choked on video... adequate simply was not quite up to date for modern video display/playback.

Back to Arkadys theory - compression is sometimes called "loss-less", but typically is "lossy", just simply because in order to compress you have to have an algorithm that is in all likelyhood imperfect, and some method of reducing the amount of 1's and 0's while retaining a sufficient amount of the original information to offer an "adequate" playback of the original source.

What constitutes "adequate" varies greatly, my camera has multiple bitrate settings and even DV options, I set the highest on the theory that the more data points, the more accurate and less potentially "noisy" (more signal, less noise) my file will be when I go to work with it and play it back.

SO, presuming that a de-compressed 24Mbps date stream has captured a sufficiently detailed "snapshot", but can now reference adjoining "snapshots" in order to add additional information while re-compressing, it's reasonable to suspect that you could produce a "better" series of "snapshots", with fewer compromises in the data integrity.

It would also result in a significantly larger file size, and potential incompatibilities with playback devices (I can burn a BR file to a regular DVD and I've had good luck playing it back ONCE I lowered the bitrate to around 17Mbps - the 8Mbps looked too degraded to use, IMO).

The "goal" is a balance at each stage of maximum "signal" (data points) with the least amount of loss or noise introduced into the workflow. The burning of an SD DVD by editing/rendering from the original HD files is a good example - if you convert the HD files to SD, THEN edit/render, you started with a significantly lower number of data points.

Generally I like to keep the bitrates/signal levels as "hot" as possible without overloading the inputs/sensor/bandwidth/hardware. What we can do NOW vs. what we could do 5 years ago (let alone 10 or 15, or heaven forbid the dark ages 25+ years ago!) is significant due to improvements in computers generally.

I like to tell my kids about how my first computer had a 30 MEGABYTE hard drive and like 256K of RAM (and don't forget those big computing devices they kept in the basement when I was in college - they always had to feed these strange beasts all these "cards", which all the computer techs used to carry around in boxes... large heavy boxes...).

Remember the now totally useless 1.2M floppy??? Fortunately I just found some cheap card readers on eBay that fit that slot, W7 immediately wants to grab onto the fast flash cards and use them to increase performance!

My Sony CX550 has 64 GIGABYTES, My main box has got 3+ terabytes of storage, and even on the "obsolete" hand me down machines, 2GB of RAM...

Compression is a "necessary evil", but that doesn't mean one has to just accept that it's fully optimized, thus Arkady's line of experimentation is interesting and valid.

[David Heath]

Originally Posted by Arkady Bolotin

The problem here is practicality. The maximum bitrate of each video standard (24 MBps for AVCHD, 25 MBps for HDV, 40 MBps for Blu-ray Disc) is not determined pure theoretically but in accordance with the corresponding media. Thus, 24 MBps bitrate is the upper limit for flash memory media, 25 Mbit/s is the maximum bitrate for magnetic tape, and 40 Mbit/s is the maximum for optic disk media.

This means that even if we manage to squeeze some additional detail by the increase of bitrate, the resultant video could be played back only on above-the-mainstream computers powered with 3.0 GHz or better processors and equipped with RAID-type hard drives.

No, Arkady, that's not true. Firstly you're confusing MegaBITS (written Mbs) with MegaBYTES (written MBs) and Mbs and MBs mean very different things. (In this context, 1MBs = 8Mbs).

AVC-HD has a maximum data rate of 24Mbs (NOT 24MBs) - but that's not the upper limit for flash memory media by far - even SDHC will go several times higher, Compact Flash a lot faster still, and P2 and SxS into the hundreds of Mbs.

As far as the basic argument goes, then video compression works by discarding the least significant data, hopefully insignificant enough that it's loss won't be noticed - at least for casual viewing. But once discarded, that's it - it's gone. To use your analogy, then if we start with 25.63452985032, the act of (lossy) compression may truncate it to 25.6345299 - and that's it. Feed that number into the best computer in the world and it can never know that the next few digits should be ....5032. As far as this goes, that's been lost for ever.

Originally Posted by Dave Blackhurst

....presuming that a de-compressed 24Mbps date stream has captured a sufficiently detailed "snapshot", but can now reference adjoining "snapshots" in order to add additional information while re-compressing, it's reasonable to suspect that you could produce a "better" series of "snapshots", with fewer compromises in the data integrity

What you're suggesting is effectively interpolation. It's true that it may make something "look" better superficially, but it won't - can't - add back in data that's been lost.

As example, let's think of a still photo, 1,000x1,000 pixels in dimension, viewed as an image of 10"x10", and full of good detail. Now we downscale to a image of just 100x100 pixels but still want to view a 10"x10" image. There are two basic ways we could do it.

The first would be to simply make each pixel fit an area of 0.1"x0.1". It would work - but have a very blocky look to it.

Better would be to upscale the image back to 1,000x1,000 - effectively "guessing" nine values in between the ones we know. That would smooth out the blockiness, make it more viewable, but wouldn't get back the detail that had been lost. It would look nicer to the eye - but be much softer than the original.

Mathematically, think of a string of numbers, which may represent brightness levels on a line across an image - let's say 1, 1, 1, 1.4, 1.6, 1.9, 2, 2, 2 and so on. Now we "compress" it by removing 1/2 of the values to get 1, 1, 1.6, 2 etc. Can anyone without access to the original ever work out what the inbetween numbers ever were?

First attempt may just be to repeat each number twice - so 1, 1, 1, 1, 1.6, 1.6, 2, 2, etc. Well, not good. Interpolation can be better (averaging before and after values) so 1, 1, 1, 1.3, 1.6, 1.8, 2, 2 etc - a closer representation - but still not correct. You can never replace real data once it's been discarded.

Now real video compression is far more complicated, obviously, and DCT etc techniques are far cleverer than simply dropping samples, let alone getting into frame by frame techniques. But the basic principle remains - once you discard data, there's no getting it back. Sorry.

As far as AVC-HD goes, then Guy is absolutely right. You could encode at a higher datarate - but then it wouldn't be AVC-HD! The implication may be that an NLE (say) wouldn't be able to cope with it.

Is the spec likely to be extended? I think it's quite likely to end up with a mode to cover 1080p/50, but I doubt anything else. The real question is what would be the point? If you compare it to (say) XDCAM 35Mbs, AVC-HD may *THEORETICALLY* be able to match it for quality - but at the expense of a lot of complexity, processing power etc. The gain then becomes (for a given quality) a smaller file size, the disadvantage far greater complexity.

Now that 35Mbs can be easily recorded to cheap media like SDHC, the incentive to compress the video ever harder becomes far less. The original rationale behind developing AVC-HD was to enable decent quality on cheap memory - not have to use P2 or SxS. Once memory advances to the point where a good quality codec can be recorded to cheap memory (think of the Canon cameras and 50Mbs 422 XDCAM to Compact Flash) much of the rationale behind AVC-HD dies. The advantages aren't worth the disadvantages.

[Robert Young]

Originally Posted by Arkady Bolotin

Why it has been completely missed by all of the engineers, software developers, & video professionals who work with these issues on a daily basis, all with the goal of trying to squeeze every fractional improvement possible out of the image data.

No, it hasn’t. Improving video quality by means of increasing bitrates is well-known and well-studied theoretical approach.

Exactly my point.
Recompressing source footage to higher bit rates has been studied extensively, is the basis of many commercial products, and, although the higher bit rate codec provides many benefits, it does not produce images that are better quality than the original footage. That seems to be the the consensus of the people who work in this field, and those who design and sell this type of product.

[Robert Young]

Originally Posted by Bruce Dempsey

Something similar occurs in the HDV world. How is it that when a scene is recorded as hdv and downconverted to sd then played back on an upconverting dvd player it looks better than a straight sd recording and play back on the same upconverting dvd player?

This is a completely different situation alltogether.
You are recording at a high resolution (HDV- lots of information captured), converting to a lower rez, lower data rate (DVD), versus recording at a low resolution (DV), converting to a different low rez, lower data rate (DVD).
Of course the high resolution sourced footage will make a better looking DVD.

[Arkady Bolotin]

Guy:

Regarding your first argument about comparison between maximum bitrate for Blu-ray disc (40 MBps) and your “crappiest USB key” which can handle 7 MB per second data transfer rate, I must say you are mixing up two things.

In digital multimedia bitrate represents the encoding level (the rate of compression): the lower bitrate, the higher compression level (and vice versa). So, 40 MBps for Blu-ray means that after data compression, each second of BD video playback uses only 5 Mbytes of data.

Meanwhile, the data transfer rate (DTR) is the amount of data that is moved from one place to another in a given time. So, the speed with which your USB key can transfer the information is 7 MB per second.

Regarding your remark about 4:2:2 chroma sub-sampling: its implementation might come true sooner than you expected. Even the first standardization of AVCHD (completed in May 2003) had the extension enabled Y’CbCr 4:2:2 and Y’CbCr 4:4:4. Right now, these options are included only in the highest profiles of the AVCHD format (Hi422P and Hi444P).

David:

When you said that 24 Mbit/s “is not the upper limit for flash memory media by far - even SDHC will go several times higher, Compact Flash a lot faster still, and P2 and SxS into the hundreds of Mbs", you unfortunately did the same mistake as Guy did: you mixed up the compression level of AVCHD standard and the DTR of memory media.

Your second argument, which states that “video compression works by discarding the least significant data”, holds true only in the case of lossy compression of data.

Meanwhile, lossless data compression allows original data to be reconstructed from the compressed data exactly. And, as we know, the H.264/AVC compression (used by AVCHD) is very close to lossless one.

Robert:

You speak semantic again. What does it mean “although the higher bit rate codec provides many benefits, it does not produce images that are better quality than the original footage”?

AVCHD video at higher bitrates might not be practical (or commercially applicable) but still provides more detailed images. Isn’t it “better quality”?

Dave:

Very well written, Dave. Nice essay, I was really enjoyed reading it. Kudos!

[Robert Young]

Originally Posted by Arkady Bolotin

Robert:

You speak semantic again. What does it mean “although the higher bit rate codec provides many benefits, it does not produce images that are better quality than the original footage”?

AVCHD video at higher bitrates might not be practical (or commercially applicable) but still provides more detailed images. Isn’t it “better quality”?

I'm not sure we are talking about the same thing:
If you mean recording at higher bit rates, of course the quality is better.
I've been under the assumption that we were discussing recording at lower bitrates and then transcoding/recompressing that original footage to a higher bit rate.
The point that I and David are making is that recompressing low bit rate footage to a higher bit rate format does not improve the image quality. It does not retrieve, or reconstruct any data that was lost in the original acquisition compression.

[Guy McLoughlin]

Originally Posted by David Heath

If you compare it to (say) XDCAM 35Mbs, AVC-HD may *THEORETICALLY* be able to match it for quality

It's been established by many people evaluating the new Panasonic AF-100 camera ( and by Barry Green using a Panasonic AVCHD recorder connected to a Sony EX-1 camera ) that the Panasonic implementation of it's 24 Mbs AVCHD CODEC is superior to Sony's implementation of 35 Mbs XDCAM CODEC. You seem to be very stubborn about not acknowledging this when it's been shown over and over again. I have yet to see ONE example where the Sony XDCAM 35 Mbs produces a better image than Panasonic AVCHD 24 Mbs.

Originally Posted by David Heath

- but at the expense of a lot of complexity, processing power etc. The gain then becomes (for a given quality) a smaller file size, the disadvantage far greater complexity.

Totally agree with you here. There is no free lunch, so AVCHD decompression requires a lot of processing power. With the release of the AF-100 camera less than one month away, I expect Apple is going to have to finally make Final Cut Pro AVCHD friendly. ( all of the worst examples of AVCHD editing that I've seen have all been done by FCP users )

Originally Posted by David Heath

Once memory advances to the point where a good quality codec can be recorded to cheap memory (think of the Canon cameras and 50Mbs 422 XDCAM to Compact Flash) much of the rationale behind AVC-HD dies. The advantages aren't worth the disadvantages.

I disagree here. I expect AVCHD to continue to be extended over the next couple of years, and eventually see an AVCHD 4:2:2 35/50 Mbs standard added to the mix. ( Panasonic will eventually build a "big brother" to the AF-100 )

[Guy McLoughlin]

Originally Posted by Arkady Bolotin

In digital multimedia bitrate represents the encoding level (the rate of compression): the lower bitrate, the higher compression level (and vice versa). So, 40 MBps for Blu-ray means that after data compression, each second of BD video playback uses only 5 Mbytes of data.

The compression level directly relates to the amount of space on disk/disc required to store a still or video image. With playback the media containing the video file has to maintain a sustained minimum data transfer rate equal to the compression rate of the image to guarantee proper playback. Thus a video file compressed at 40 Mbs requires playback media that can deliver a sustained 40 Mbs data transfer rate to deliver a proper image. ( assuming the rest of the electronic processing path can handle this amount of data )

Many of the new Blu-ray players have a USB port so that you can indeed copy your BD video file to a USB key to play your video at full resolution on your TV set. The Panasonic Blu-ray players even understand the AVCHD file format, so you can pop an SDHC card directly from your camera and in to a SDHC slot on their player to play the video footage you just shot seconds earlier. ( no additional editing/processing required )

[Arkady Bolotin]

Robert:

I very much regret, but I suspect either you haven’t read my respond to David’s comments or – if you have – you did this carelessly.

Nevertheless, let me repeat my arguments again.

Your statement that “recompressing low bit rate footage to a higher bit rate format does not improve the image quality” and another one that “it does not retrieve, or reconstruct any data that was lost in the original acquisition compression” would be correct under the following two conditions:

1. The H.264/AVC compression (used by the AVCHD format) was of lossy type, which discards (loses) some of the data, in order to achieve compression.

2. The snapshot fragments in the beginning of this thread did not exhibit any difference; or those fragments did not exist at all (as well as my experiments).

However, both these conditions are false. First, by its results the H.264/AVC compression can be classed as lossless one; second, the fragments do exist and they do show evidence of improvements in the detail department.


Guy:

Yes, it’s correct that there is a correlation between the maximum multimedia bitrate and the data transfer rate (DTR) for the corresponding media (if this is what you are trying to say).

For example, the Blu-ray disc max bitrate of 40 Megabit/s means that the corresponding media – which is an optical disk – must provide such speed of data transferring that is not less than 5 Megabytes per second.

Analogously, the AVCHD max bitrate of 24 Megabit/s stays that the media – flash memory – must supply data at the speed not less than 3 Megabytes per second.

[David Heath]

Originally Posted by Arkady Bolotin

David:

When you said that 24 Mbit/s “is not the upper limit for flash memory media by far - even SDHC will go several times higher, Compact Flash a lot faster still, and P2 and SxS into the hundreds of Mbs", you unfortunately did the same mistake as Guy did: you mixed up the compression level of AVCHD standard and the DTR of memory media.

No, no mistake by either of us. My answer you quote above was a direct response to your earlier statement that ” Thus, 24 MBps bitrate is the upper limit for flash memory media, 25 Mbit/s is ………….” You are clearly referring to flash memory – not AVC-HD – and that is what Guy and I responded to.

Also note that AVC-HD is not (as you seem to imply) uniquely tied to flash memory anyway. The spec (AVCHD - Wikipedia, the free encyclopedia ) says:

Developed jointly by Sony and Panasonic, the format was announced in 2006 ........AVCHD is a file-based format and ........., video can be recorded onto DVD discs, hard disk drives, non-removable solid-state memory and removable flash memory..............

Originally Posted by Arkady Bolotin

However, both these conditions are false. First, by its results the H.264/AVC compression can be classed as lossless one; second, the fragments do exist and they do show evidence of improvements in the detail department.

However good AVC-HD may be, I really don’t think it can be called “lossless” by any imagination. How great the losses are will vary hugely with picture content. But if your point about it being “lossless” was correct, then surely your whole argument falls down anyway? If 24Mbs AVC-HD is lossless, how can it be improved on by reencoding to a higher bitrate? “Lossless” implies “perfect”, and how can you improve on that?

Examining the samples you show more closely, I’m not really sure either is much different to the other. And are you certain they both correspond to EXACTLY the same frame? They are different – but I’m not convinced one is actually BETTER than the other. Likewise, have you compared a succession of frames throughout that shot? Without doing that, there’s no way to eliminate differences caused by random variations on a pixel basis during recompression.

Originally Posted by Guy McLoughlin

I expect AVCHD to continue to be extended over the next couple of years, and eventually see an AVCHD 4:2:2 35/50 Mbs standard added to the mix. ( Panasonic will eventually build a "big brother" to the AF-100 )

At heart, H264 and MPEG2 are both built on the same foundation. H264 has an additional range of tricks it can call on to help, but the improvement they can give is not linear with bitrate. The tricks work far better at lower bitrates, give far less relative improvement at higher bitrates.

So a higher than 24Mbs rate of AVC-HD will progressively lose it’s theoretical advantage over comparable MPEG2, whilst being far more difficult to work with. The more the bitrate gets upped, the less and less point there is to AVC over MPEG2.
If a big brother to the AF100 emerges, I’d expect it to be AVC-Intra 100 to SDXC cards, which potential “big brother” users would find far more useful. (And SDXC – not SDHC – already has the speed required for 100Mbs.)

[Dave Blackhurst]

It struck me that the single linear/plane digit isn't at all a good analogy, except on a most simplistic level... and a more complex example is necessary to explain why, with "interpolation", it is possible to "recreate" otherwise "lost" data.

Even in a still photo, you have not a linear series of numbers (each pixel is not represented by A digit from 0-9, rather a SERIES of digits), but rather a matrix of "numbers", In other words there is not just "X", but "X" and "Y", and except on the very edges, where there wouldn't be adjacent pixels, each pixel has EIGHT adjacent pixels from which to potentially draw "information" Each pixel, is not entirely independent, but is related to the surrounding pixels - part of any algorithm is the discarding of "redundant" information while leaving a marker to "restore" that information upon decompression, by referencing the surrounding pixel data.

I believe each pixel, being comprised of multiple bits, in effect would also have a third dimensional component.

So in order to more accurately understand the potential, one needs to envision that EACH pixel has a given depth, and eight adjacent "information sets", of some "depth". It doesn't take very long to realize the data volume we are dealing with, and why compression becomes a potential necessity in order to deal with that volume.

We all realize that in the single linear digit example, once a number is GONE, it's GONE... but for the sake of argument when dealing with compression schemes, you have to look at it three dimensionally - IOW, the algorithm takes a look at numbers (data bits) which can be "reduced" because of redundancy, and recreated from the surrounding data points during decompression. The higher the level of compression (or the lower the level of horsepower/hamsters available to process the data), the more data points must be discarded, thus the greater the reduction in the quality of the eventual reconstruction.

If we ignore the "edges" (worst case being a corner, with 7 adjacent points, or an edge with 10), each digital pixel "data poiint" actually has 26 directly adjacent points in a 3 dimensional (X, Y, & Z) matrix... depending on the efficiency and accuracy of the compression scheme (as noted, they do vary), you can see where there is much more information to draw from - and that's before you begin to consider that the surrounding pixels provide a degree of information that can be utilized, to lesser and degrading degrees, the farther away they are from the "target pixel"...

Add to that, as soon as we speak of video, we now are not talking about a snapshot, but 30 per second, in motion, which adds another very complex set of dimensions, and I think I just gave myself a headache!

Keep in mind that things like face tracking, auto focus,exposure, etc, are crunching these data points in REAL TIME, or at least as fast and the little processing hamsters can run, in order to deliver the best possible results!



I'm hoping my offspring who seem to excel in mathematics are better prepared for the digital world - I understand films, emulsions, stuff I can handle and "see"... once you start converting everything into 1's and 0's it gets so much more complex to get a handle on!!

I am finding this discussion fascinating, I think we are actually much on the same page, as there is an acknowledgement that all AVCHD is not necessarily "alike" - I've been shooting AVCHD cameras for a while, the first ones IMO were "better" or equal to HDV, but as things have advanced, they have allowed me to get much better footage, often in much tougher shooting conditions, than any HDV/tape camera I could afford - some of that is hardware, but I have to presume that algorithms are also being tweaked to improve performance.

My opinion is that it's taken a few generations to get to where, even with dedicated pixel peeping, it's hard to argue with the performance... but we're close... at the least we are approaching the point of diminishing returns under the existing standards.

Before one says that just because the current state of the art is "the best that can be achieved", one should consider that man isn't supposed to fly any more than the humble honeybee... it's only a matter of time, creative thinking, and dedicated effort, before the "impossible" becomes reachable... barriers are meant to be broken!

Remember too that scientists/engineers in the lab must eventually be told they need to release their "monsters" to pay the bills, otherwise they'd always be tweaking and improving and we'd never get to play with the new toys... and so next years new toys will almost always bring a years worth of creativity and tweaking to the table... such an exciting and fun time to be dealing with technoliogy!

[Robert Young]

Originally Posted by Arkady Bolotin

...would be correct under the following two conditions:

1. The H.264/AVC compression (used by the AVCHD format) was of lossy type, which discards (loses) some of the data, in order to achieve compression.

2. The snapshot fragments in the beginning of this thread did not exhibit any difference; or those fragments did not exist at all (as well as my experiments).

I disagree with your assumptions:
1) AVCHD is not considered a lossless codec- in fact, most consider it to be very lossy. It does not tolerate repeated recompression well at all. This is why it's often converted to DI for extensive, or multi system editing. To decompress AVCHD from the original file and recompress it again to AVCHD- no matter what the data rate- will actually cause you to lose quality
2) Honestly, I don't think I can actually see any significant quality difference in the two frame grabs from your original post.
My main point throughout this discussion is that the notion of recompressing original footage to a higher data rate is exactly what occurs when we use a DI codec for editing. It is established knowledge that the high data rate DI image is not of higher quality than the original lower data rate acquisition footage.
The companies that develop and sell DI products readily acknowledge this fact.
Clearly, if there were even a slight chance that the image quality was improved by this process, these companies would be marketing their products extensively on this basis.

[Arkady Bolotin]

David:

I must confess I like your arguments, they‘re logical, smart and mostly flawless. Therewithal, I found few of them being wrong.

First, regarding your definition of lossless compression as “perfect”, I have to say lossless does not mean perfect, it means “without loss” (simply because without proper definition word “perfect” has no meaning).

Second, to your question how I can improve on that, I answer this. By partly decompressing AVC-compressed video (rendering it at higher bitrates) we lessen a processor load and – as result –number of artifacts during replay.

Third, concerning you reservation whether we can classify AVC compression as lossless. Technically speaking we cannot, but since this compression design includes many lossless compression algorithms, in many cases we may expect the exact (or almost exact) reconstruction of data compressed with H.264/AVC. In which cases and to what extent it would be exact, this – as you properly put – would apparently depend on footage content.

Regarding your observation of the frame fragments, I have nothing to add. As you understand, it was virtually impossible to extract the same frame from two different clips, so your assumption that they are different just because of random pixel variation should stay.


Dave:

You did it again! Very, very nice! Even though I have studied it for many years, never before theory of data compression had seemed so appealing to me! For that I want to thank you!


Robert:

Sorry Robert, I have noticed your reply just in the moment I’ve finished my responds to the previous posters. But I believe most of what I have written to David would be relevant to your comments. If not, let me know, I’ll answer to them later. Again, my sincere apologies.

[David Heath]

Originally Posted by Arkady Bolotin

First, regarding your definition of lossless compression as “perfect”, I have to say lossless does not mean perfect, it means “without loss”.....

I can't help feeling that's just playing with words, sorry.

Originally Posted by Arkady Bolotin

By partly decompressing AVC-compressed video (rendering it at higher bitrates) we lessen a processor load and – as result –number of artifacts during replay.

But that's not going to happen like that. If the processor can't cope it will stutter or drop frames - not just introduce a few more artifacts. And what basis do you have for saying that higher bitrate AVC will need less processor power than lower bitrate AVC?

Originally Posted by Arkady Bolotin

Technically speaking we cannot, but since this compression design includes many lossless compression algorithms, in many cases we may expect the exact (or almost exact) reconstruction of data compressed with H.264/AVC.

It may include some lossless algorithms, but it includes a lot of lossy ones as well! How much loss will depend on bitrate (higher the bitrate, greater the loss) but you be sure there will be some.

Originally Posted by Arkady Bolotin

Regarding your observation of the frame fragments, I have nothing to add. As you understand, it was virtually impossible to extract the same frame from two different clips, so your assumption that they are different just because of random pixel variation should stay.

Well, if the comparisons offered aren't of the same frame, pre and post recompression, I don't think any sensible conclusion at all can be drawn. How can you be sure that the differences between the two you've shown are due to your theory, and not simply randomness and noise!? Frankly, that's a far more likely explanation for what you're seeing.

If you can't make sure of a same frame comparison, (by looking at a frame, say, 15 frames after a vision cut?) the only other way to do a meaningful test would be to take a random sample of (say) 10 original frames, and a random sample of 10 recompressed frames. To make it scientific, you'd then need to present all 20 randomly to someone who didn't know which was which, and ask them to choose what they thought were the best 10.

If they picked out the 10 recompressed frames, I might start to get interested. I'm not holding my breath, though.

Dave - yes, my analogy was highly simplistic. But it does hold true IN PRINCIPLE for real world systems. What you are describing is lossless compression, and yes, it's obviously possible. But the level of compression that is being spoken of here is far, far higher than can be achieved losslessly, certainly for any real world images. The trick is to do it in such a way that the most insignificant details get discarded first. Ideally, only the ones that the eye is unlikely to notice anyway - though even that may reduce the scope for post manipulation.

Originally Posted by Dave Blackhurst

........there is an acknowledgement that all AVCHD is not necessarily "alike" - I've been shooting AVCHD cameras for a while, the first ones IMO were "better" or equal to HDV, but as things have advanced, they have allowed me to get much better footage, often in much tougher shooting conditions, than any HDV/tape camera I could afford - some of that is hardware, but I have to presume that algorithms are also being tweaked to improve performance.

Now that is without question. The AVC-HD spec allows for various "tricks" to be used in addition to basic MPEG2 to improve the quality/bitrate ratio - but how many is not specified. The spec only fully defines the DECODER, not the coder. An early AVC coder may not have been much better than basic MPEG2.

Hardly surprisingly, cheaper cameras aren't likely to be as good as more expensive ones, and the coders improve with general technical advances. I fully expect the AF100 to have a better coder than the HMC150, for example - technology moves on.

In the UK that was demonstrated recently with the HD broadcasts when the BBC got new hardware coders. (In the UK, HD is broadcast as H264.) They were able to substantially reduce the bitrate, with not much impact on overall quality.

[Dave Blackhurst]

In the end it's all a matter of degree...

You've got to remember that "reality" is not 1's and 0's, but rather a highly complex multidimensional environment... one that no two people can perceive in "exactly" the same way. ANY device or methodology to "capture" or record "reality" introduces another complex series of variables...

IOW, "perfect" doesn't exist. You're merely creating a representation, with some degree of degradation or enhancement at various stages.

I don't think the deficiencies of technology/flaws in reproduction/losses of data points count in the end - if the CONTENT has value or "moves" the viewer, it doesn't matter to the "average audience member" if it looks like poo, or has flaws...

While it's a noble task to improve upon technology and max out the capabilities of our "toys", one should remember that human perception fills in most of the "holes" anyway, sometimes in amusing ways (optical illusions illustrate how "easy" it is to fool the eye).

Most compression and interpolation schemes are just another way to achieve the most efficient balance between image quality and the capabilities of the hardware, while "fooling" the eye of the viewer in a pleasing way... and are inherently "flawed", regardless of whether mathematically they are "lossless" or "lossy".

Over time, with increases in hardware capabilities and improvements in software, we should get closer to "perfect", or at least the perception thereof. It still will come down to the skill and vision of the human talent that will determine the actual value of the end product.