AVC Codec Curveball

[Homunculus]

i'm a composer and come from audio as well. I can tell you right now with confidence that MOST audio PROFESSIONALS cannot tell teh difference between 16bit and 24bit. this is FACT. I have seen it tested SEVERAL TIMES by professional studio engineers, composers, producers, etc.

It seemed like 1 or 2 out of 10 poeple CLAIMED to be able to hear differnce of 16bit and 24bit recording. I personally can't hear the damn difference and I have 2000$ studio monitors

[Rick Burnett]

i'm a composer and come from audio as well. I can tell you right now with confidence that MOST audio PROFESSIONALS cannot tell teh difference between 16bit and 24bit. this is FACT. I have seen it tested SEVERAL TIMES by professional studio engineers, composers, producers, etc.

It seemed like 1 or 2 out of 10 poeple CLAIMED to be able to hear differnce of 16bit and 24bit recording. I personally can't hear the damn difference and I have 2000$ studio monitors

I too come from the audio world, and I use ADAM A7/SUB8 monitors and record on Apogee Rosetta 200, Digidesign Rack 3 and 2 Motu 896HD interfaces. The difference comes into play on what you edit on. Since very few things work in 16-bit INT, no one ever sees a problem. Every major DAW takes that 16-bit or 24-bit and places it in a 32-bit floating point representation. Once you've done that, you have so much incredible headroom, I've never run into issues.

I too have also experimented with ProTools|HD and 48-bit INT audio and I don't buy it as being superior. I've done EXTENSIVE testing, with both ProTools Native and HD and with the limited amount of processing I do, saying MAYBE 4 levels deep of EQ, Verb, limiting, etc, I don't see the degradation. This includes using low pass filtering which is some of the harshest on quantization errors and still, nothing audible.

What IS true though is that if you look at hardware designed for 16-bit (which will be older) and look at newer stuff that is 24-bit (or higher), what is important is the noise floor. Most newer equipment is going to be MUCH quieter and therefore, if you record at 16-bit or 24-bit, the noise is still going to be around the same, at least from what I have recorded on. But like I said, no one edits in 16-bit, or at least, no one I know.

With video the same is true. Throwing more bits at a sensor that isn't sensitive enough (too much noise) isn't going to help. However, I can say that the difference between 4:2:0 and 4:2:2 IS noticeable when you do visual effects work. I can speak mostly from using the 7D, but the color contamination you get at contrast boundaries between different colors makes rotoscoping with automatic tools really difficult. In addition, the same is true of fine detail and green screening. You don't want the green screen contaminating your subject, but it does and it is a pain to work with. When I look at the pixel zooms of many of the AF100 examples, I can see clearly that when I am doing VFX shots, I will want to use an external recorder to help me in post. For other shots, I don't think it is necessary.

[hjulenissen]

24 bit recording in nice to have, not only for processing, but also for headroom. 16bit audio sounds great, we don't really listen low enough to tell the difference between 16bit and 24bit distribution IMHO.

Agree. In itself, >~14 bits seemingly adds no audible benefit (when properly dithered), but recording practice makes it beneficial to record with the highest possible resolution. It does not add significant cost either, so why not.


A difficult speaker will be limited by a cheap amp. It's not a theoretical difference.

Depending on where you put the threshold for "cheap". $100 amplifiers are limited in the basic specs that manufacurers can afford. $500 or $1000 amplifiers does not differentiate significantly in measured specs, but rather by "magic non-measurable properties". So yes, really cheap amplifiers do often distort sound audibly (depending on the load), but the price point where we go from engineering to snake-oil is a lot lower than the audiophiles (and some recording technicians) tends to think.


Monster Cable? They have many patents on product display shells:) They did some good designs that everybody copies, and that's something by itself. Some of their stuff is actually useful, RCA with a split cylinder, things like that.

I dont contest that. Solid, well-engineered cables is nice to have, and should not cost more than a few dollars from a reputable "pro" source.


Surprised you didn't ask why I put 10bit YCbCr tonality only 2 bits over 8bit YCbCr when the math would put that 6bits over:) It's mathematically imprecise, but even 8bit luma after gamma is very high and an extra bit in each chroma channel is generous considering an RGB input. That brings the number to 11 millions, a very high figure for a camcorder. For the computer graphics formats in 8bit and 16bit RGB we have more but no camcorder gets billions, even recording linear 12bit, because of the processing required to output useful, white balanced, color space mapped, gamma corrected, output. Video capable of full 8bit RGB would look as clean as a graphics rendering.

The top scoring sensor for tonality on DXOMark:

http://dxomark.com/index.php/en/Camera-Sensor/All-tested-sensors/Phase-One/P65-Plus

At an effective ISO44, tonal range is 10 bits and color sensitivity 26bits. 16bit analog to digital conversion, large pixel CCD, slow readout, and still no billions of colors.
[/quote]
Yes, I am aware of those tests. I am guessing that the highest end of still-image cameras perform better than any moving-image camera in this respect, so that the ability for video cameras to deliver "true" 10bpp is even less.

According to Poynton, humans can perceive a 100:1 contrast grayscale gradation of 14 bits linear or 9 bit with gamma at a "given point of adaptation". 8 bits with gamma is enough for transparent viewing at 50:1 contrast.

I believe that luma and luminance is what is most important, quantization errors in chroma channels is usually (?) not that big of a problem.

I also believe that most (re)quantizations should ideally be carried out using dithering (or sensor noise, although that is probably not ideal).

-h

[hjulenissen]

I too come from the audio world, and I use ADAM A7/SUB8 monitors and record on Apogee Rosetta 200, Digidesign Rack 3 and 2 Motu 896HD interfaces. The difference comes into play on what you edit on. Since very few things work in 16-bit INT, no one ever sees a problem. Every major DAW takes that 16-bit or 24-bit and places it in a 32-bit floating point representation. Once you've done that, you have so much incredible headroom, I've never run into issues.

I think that it is useful to split the discussion into "sensor resolution" and "processing resolution". My claim is that ~14 true bits end-to-end resolution seems to be enough for human listeners. If you are going to connect the ADC directly to the DAC, then that means that both can be 14 true bits.

If you are going to use any processing in-between (and in practice, you probably are), then intermediate resolution should be high enough not to lead to significant losses. If you are going to do heavy gain processing, more than 14 bits ADC could be warranted for (~20bits is available for cheap).

It is very hard to say anything about the consequence of internal dsp resolution from the outside. Clever algorithm design can often compensate for what seems like low-resolution architecture.


With video the same is true. Throwing more bits at a sensor that isn't sensitive enough (too much noise) isn't going to help. However, I can say that the difference between 4:2:0 and 4:2:2 IS noticeable when you do visual effects work. I can speak mostly from using the 7D, but the color contamination you get at contrast boundaries between different colors makes rotoscoping with automatic tools really difficult. In addition, the same is true of fine detail and green screening. You don't want the green screen contaminating your subject, but it does and it is a pain to work with. When I look at the pixel zooms of many of the AF100 examples, I can see clearly that when I am doing VFX shots, I will want to use an external recorder to help me in post. For other shots, I don't think it is necessary.
Yes, I see the green-screen issue. Are there no clever algorithms that try to do this directly on chroma-subsampled material? Do you find any differences for non-oversampled Bayer sensors vs 3CCD solutions in this respect?

-h

[BobDiaz]

Here's your answer:

To hell with calculations and screwing around with math. Just look at the images. Put them to the test.
...

Barry,

This whole thread is like a flash-back to the early days of the HMC-200 and the endless debate over the image quality. My favorite was those who would try to use a calculator to judge the quality of an image they never saw. :2vrolijk_08:



Bob Diaz

[Rick Burnett]

I think that it is useful to split the discussion into "sensor resolution" and "processing resolution". My claim is that ~14 true bits end-to-end resolution seems to be enough for human listeners. If you are going to connect the ADC directly to the DAC, then that means that both can be 14 true bits.

If you are going to use any processing in-between (and in practice, you probably are), then intermediate resolution should be high enough not to lead to significant losses. If you are going to do heavy gain processing, more than 14 bits ADC could be warranted for (~20bits is available for cheap).

It is very hard to say anything about the consequence of internal dsp resolution from the outside. Clever algorithm design can often compensate for what seems like low-resolution architecture.

Yes, I see the green-screen issue. Are there no clever algorithms that try to do this directly on chroma-subsampled material? Do you find any differences for non-oversampled Bayer sensors vs 3CCD solutions in this respect?

-h

I do agree with the 14->14 as well, but here in this forum, I'd assume most people are doing some sort of processing to the audio, true? But, even worse than that is most people like to comment on 16-bit versus 24-bit without any tangible experience (not speaking of people in this thread, but in general). I'll give a perfect example, I bought both the Sony PCM-D50 audio recorder and the Zoom H4. WITHOUT QUESTION the Sony PCM-D50 was superior in terms of converter noise. I bought these to use on set to pick up environmental noises, ie, birds, ac units, room tone, traffic, etc. The H4 was VERY noisy for this sort of application (although works well for voice). Both were the same bit depth, yet, the results, easily measurable, were worlds apart. (And yes, I used both the internal mics on the H4 and a $2k external mic, it was NOT a mic issue).

Of course, let's be realistic, even when using a dedicated boom mic or lav, the noise you pick up is still greater than that of the noise of the pre-amp. Your effective dynamic range between the environment and your subject is typically the limiting factor, not the A/D (although I would argue that most cameras have really crappy built in mics, and that IS a limiting factor if you are trying to use them).

Pertaining to the greenscreening, I have not found any clever mechanisms yet. The macroblocking with H.264 based compression seems to cause a lot of the problems. When I did some work with 3CCD it was easier, but some of those cameras were pretty noisy. I'd say the majority of my work has been with CMOS though.

Since I rotoscope on a frame by frame basis, the best comparison I can give is that I had a picture I took of the same scene that i had video for. I downsampled the image to 1080p and then did the rotoscoping on it and it was real easy, the tool didn't get confused at the boundaries. Same for green screen. Ideally, I'd like to take a larger bayer pattern image scaled down to 1080p and use raw, but I really believe using one of the external solutions will suffice.

[Lpowell]

My claim is that ~14 [audio] true bits end-to-end resolution seems to be enough for human listeners. If you are going to connect the ADC directly to the DAC, then that means that both can be 14 true bits.
16-bit would be adequate for audio capture if digital recorders had built-in headroom limiters with analog-like saturation response. In practice, to handle the live dynamics of highly percussive musical instruments, it's best to have 24 bits of headroom available to avoid clipping the extreme peaks. You almost never hear these extreme levels of dynamic range in popular music releases due to the overwhelming amounts of compression used in mastering. There are some examples of uncompressed audio in early digital jazz recordings from the 80's, and their recording levels sound very quiet to modern sensibilities.

[hjulenissen]

16-bit would be adequate for audio capture if digital recorders had built-in headroom limiters with analog-like saturation response. In practice, to handle the live dynamics of highly percussive musical instruments, it's best to have 24 bits of headroom available to avoid clipping the extreme peaks. You almost never hear these extreme levels of dynamic range in popular music releases due to the overwhelming amounts of compression used in mastering. There are some examples of uncompressed audio in early digital jazz recordings from the 80's, and their recording levels sound very quiet to modern sensibilities.
I use ReplayGain to attenuate all those "modern" recording whose dynamics are crushed (and/or clipping applied) in a cheap attempt to impress on the radio. A high crest-factor is the norm in acoustic music. Just like many visual scenes are highly dynamic and applying HDR+tone-mapping commonly give them an artificial look (even though more detail is squeezed into the limited-dynamic display).

But this is recording engineer technique, and probably way off topic. My only claim was that 14 bits unprocessed seems enough. If you want to do dynamics processing (hopefully from esthetic reasons, not to "beat Britney by 2dB"), then more recording resolution could be warranted.

If levels are set to avoid clipping, and no compressor/limiting is done, do you think that the limitations of (well done) 14 bits can be audible in realistic source material?

-h

[aquavideo]

Homunculus, I understand your interest in finding out if it is the same sensor but you have to realize there are many reasons they may not want that info out early, be it competition or the ability to make alterations up until the last minute. And you may never find out for sure unless someone does an iFIXIT type of tear down - which is a bit expensive when you are talking a $5K camera. Seems like Jan does a pretty good job of keeping up on all of those details, but you have to remember that Panasonic USA is essentially a marketing/distribution arm so not necessarily constant communication with the engineers on the internal details.

I've met a few of those engineers at NAB and such from both Sony and Panasonic and they tend to be VERY careful about what they say.

[Otis Grapsas]

The actual sensor used is not very important. It doesn't matter if it's identical with something else. There is nothing inherently bad about these sensors. The dSLR problem is not the sensor, it's the still camera design and the compromises required to keeep the still photography functionality.

[Lpowell]

If levels are set to avoid clipping, and no compressor/limiting is done, do you think that the limitations of (well done) 14 bits can be audible in realistic source material?
Theoretically, yes, but in a live recording situation you can't always predict the maximum transient level you'll need to handle. In addition, digital level meters on portable recorders aren't precise enough to capture the fastest peaks, they're only intended as a rough guide. The safe bet is to record in 24-bits with average levels around -18db. That provides at least 16 bits of resolution for even the quietest passages.

[Homunculus]

The actual sensor used is not very important. It doesn't matter if it's identical with something else. There is nothing inherently bad about these sensors. The dSLR problem is not the sensor, it's the still camera design and the compromises required to keeep the still photography functionality.

but then the point and question becomes, how do you differentiate between a 1000 and 5000 price point if both of your cameras, one consumer dslr, one prosumer cinecam have identical sensors and just about the same codec etc?

Meaning if you have a gh2 acvhd for 1000 then a af100 that has the same gh2 sensor with also acvhd, how do you reconcile it being 4000$ more expensive? what else is there inside that should in theory be making the af100 look much better than gh2 if it's NOT the sensor?
is it the optical filters or what

[hjulenissen]

The actual sensor used is not very important. It doesn't matter if it's identical with something else. There is nothing inherently bad about these sensors. The dSLR problem is not the sensor, it's the still camera design and the compromises required to keeep the still photography functionality.
I thought it was the 18megapixel-optimized OLPF, low bandwidth sensor readout resulting in line-dropping (aliasing and reduced sensitivity) as well as low framerates.

If you could have the basic 5Dmk2 sensor properties only properly prefilterd/downsampled to 1080p24/1080p60, stored as high-bitrate h264, would this not make it quite good for what people try to use it for? I see that ergonimics, manual controls, viewfinder, audio i/o etc may still be lacking, but those are only indirectly affecting video-quality.

-h

[Otis Grapsas]

I thought it was the 18megapixel-optimized OLPF, low bandwidth sensor readout resulting in line-dropping (aliasing and reduced sensitivity) as well as low framerates.

If you could have the basic 5Dmk2 sensor properties only properly prefilterd/downsampled to 1080p24/1080p60, stored as high-bitrate h264, would this not make it quite good for what people try to use it for? I see that ergonimics, manual controls, viewfinder, audio i/o etc may still be lacking, but those are only indirectly affecting video-quality.

-h

I believe a simple change to the OLPF would make it perform pretty well.

Fixing a dSLR for alias is not really an engineering challenge. The OLPF requires a new design, but I know two factories that can build the part and handle the conversion. The rest of the limitations would still break the deal for many people though and if you actually ask people if they want a permanent change that is not compatible with still mode, many will say no.

[Otis Grapsas]

but then the point and question becomes, how do you differentiate between a 1000 and 5000 price point if both of your cameras, one consumer dslr, one prosumer cinecam have identical sensors and just about the same codec etc?

Meaning if you have a gh2 acvhd for 1000 then a af100 that has the same gh2 sensor with also acvhd, how do you reconcile it being 4000$ more expensive? what else is there inside that should in theory be making the af100 look much better than gh2 if it's NOT the sensor?
is it the optical filters or what

The feature set alone would be worth the difference. I'm not sure AF100 looks better than a GH2, and I don't see how why it should look different on the sensor level. It has less aliasing, a better codec and better processing which explains any quality differences.

[hjulenissen]

I believe a simple change to the OLPF would make it perform pretty well.

Fixing a dSLR for alias is not really an engineering challenge. The OLPF requires a new design, but I know two factories that can build the part and handle the conversion. The rest of the limitations would still break the deal for many people though and if you actually ask people if they want a permanent change that is not compatible with still mode, many will say no.
Video-optimized OLPF would fix the aliasing issues. But the camera will still throw away 3/4 of received photons or something due to line-skipping. It will also not do 60 frames/fields per second unless one use the 720p mode with even more line-skipping/aliasing issues.

But if a good 1080/720 OLPF was available, and 1080p50/60 at "180 degrees shutter" was possible (additional cooling? Firmware hacks?)

What do you think would be missing in terms of pro video IQ? (Nevermind the fact that most buyers are non-pro and dont want to ruin the still-mode). Is there something inherently bad about its demosaic, nr, color processing, 8bpp gamma, lossy encoding etc that seriously limits its use?

Like the audio-discussion has shown, I think it makes sense to discuss quantization, lossy encoding & friends in terms of a)What would have been needed in an ideal world (given perfect in-camera exposure, color-matrix etc), and b)how large additional margins makes sense for typical post adjustements.

[Otis Grapsas]

Video-optimized OLPF would fix the aliasing issues. But the camera will still throw away 3/4 of received photons or something due to line-skipping. It will also not do 60 frames/fields per second unless one use the 720p mode with even more line-skipping/aliasing issues.

But if a good 1080/720 OLPF was available, and 1080p50/60 at "180 degrees shutter" was possible (additional cooling? Firmware hacks?)

What do you think would be missing in terms of pro video IQ? (Nevermind the fact that most buyers are non-pro and dont want to ruin the still-mode). Is there something inherently bad about its demosaic, nr, color processing, 8bpp gamma, lossy encoding etc that seriously limits its use?

Like the audio-discussion has shown, I think it makes sense to discuss quantization, lossy encoding & friends in terms of a)What would have been needed in an ideal world (given perfect in-camera exposure, color-matrix etc), and b)how large additional margins makes sense for typical post adjustements.

A third party modification would only change the OLPF. The rest of the implementation is a given. If they could do better readout, they would be doing it.

Quality standards vary from person to person depending on the use. I wouldn't use a dSLR because I find the codecs inferior to those of a good camcorder. If they could fix aliasing and the codec quality issues, it would be very close to a camcorder in image quality, like the AF100.

If you can nail the look in-camera and skip all post processing, you can get very good results out of common lossy codecs, but motion quality will still be a compromise vs a high end camera. You know, many people read our posts and many absorb them as objective information, which is a mistake. When deciding on a camera model, it makes sense to rent for a day. There are many people who love the Sony EX1 and others who crossed EX1 out on their list when they noticed the motion quality. Renting for a day is very cost effective and saves a lot of time compared to listening to opinions on the internet. You never know where one is coming from, right?

RANT mode!

There are lots of people, reviewers, bloggers etc, that practically work for the manufacturers. You can tell by the actual words they use when reviewing a competitive product. I'm SO tired of these camcorder gurus who haven't written a gamma curve in their life but manage to shape the opinions of thousands, and it's always about pushing the lowest quality camcorders, the ones that make the most money in each segment. They sell the cameras before they are even manufactured! Practically anyone can see right through them after a couple of years of following their activity and most people suspect the type of relationship they have with specific manufacturers, but newbies are easy to influence. It's natural to trust those who appear to hold some weight in the discussion and environment. It is a measure of authority and we know how people respond to authority.

There are also people who are so involved in specific implementations that they can't be objective any more. You will not hear good words about dSLRs from Jim Jannard. You won't hear good words about compression from me because I obviously love uncompressed imaging, I'm spending all my time on it, it's even a camcorder that might be a commercial product soon, in other words competitive to hm everything else? So take my words with a grain of salt, the involvement in my project is so strong that it must be pushing through. I do like uncompressed a lot, but I discuss with people that are only interested in shooting filmmaking. They want to shoot demanding material and make the most out of post. They don't want to keep people waiting on set in order to finetune gamma/matrix/etc and they don't want wasted shots. They will pay for the flexibility to handle these things in post and the creative possibilities that provides, because it's a good investment. They want to do film and digital cinema delivery so they appreciate the uncompressed high precision recording and get intraframe formats when they have an option. They come from digibeta, 16mm and 35mm, HDCAM, and DVCPROHD Panasonics. But that's not the typical customer for a camcorder. Cinematography is not the typical use for an affordable camcorder. It's quite possible that my perspective does not even apply to most of the people reading these threads, it refers to a different usage scenario. But I believe this is a filmmaking forum, and this is why I post here. Some people might be interested in the technical implications of camera formats in post.

To sum it up, I'm not the one to ask about what will work for a specific user and I can't quantify it in bits, PSNR or anything else. The discussion is technical and theoretical and it could also be misleading. You never know until you try the actual camcorder/implementation, take it through your type of post and put it against your personal criteria. It costs what? 5% of the camera cost? Very good investment vs blindly getting a $10,000 package that will not do what you expect it to do.

[Otis Grapsas]

A sample from Alexa and Ki-Pro (1080p version available):

http://www.youtube.com/watch?v=xixmVTz5nNc

In this case the limit is the format because Alexa has very good design and a very large pixel. It would look even better with uncompressed and higher precision recording. This must be the best lit Alexa sample so far. The quality of overexposure handling and color are evident even in the 1080p youtube stream. This is no cinegamma curve for sure:)

Use the same recorder with a consumer/prosumer quality camcorder and you will get results similar to the in-camera recording because the camera is the limit. Same clipping, same shadow noise look, same color response. The limit is the sensor quality and the design quality. I understand how sensor quality limits these cameras, but I don't understand why they don't invest a little more in improving the design of the processing, gamma, matrix etc.

[roei z]

Heh... done in Israel.. :)

this is words apart from avchd 24mbs

[Otis Grapsas]

Heh... done in Israel.. :)

this is words apart from avchd 24mbs

Yes, but it is actually inferior! AVCHD is Uncompressed->24mbit. This is Uncompressed->KiPro codec->4mbit for youtube. 1/6th the AVCHD rate and an additional compressed intermediate that also damaged quality.

What we are seeing in this sample (besides good lighting) is good pixel quality and processing quality. These two survive any codec and no codec or recording method will add these to the output of an affordable camcorder.

[hjulenissen]

RANT mode!

There are lots of people, reviewers, bloggers etc, that practically work for the manufacturers. You can tell by the actual words they use when reviewing a competitive product. I'm SO tired of these camcorder gurus who haven't written a gamma curve in their life but manage to shape the opinions of thousands, and it's always about pushing the lowest quality camcorders, the ones that make the most money in each segment.
I believe that if the discussion contains measurable, physically sound stuff, along with "see for yourself examples", those problems are reduced.

Not saying that everything relevant about a camera can be (easily) measured, nor that every measurement is relevant. But doing that stuff in addition to subjective "how did it feel to your hands", "was the gui intuitive", "what did the noise look like to you" seems to aid the debate. Anyone can say that a video looks flawless or like crap. And they do on the net.

One could easily make a parody out of the opposite of what you are describing. Those that talk for hours about differences that are undetectable to the naked eye unless heavy edited first as if it is making the image unwatchable. Those that hypothesise about how good stuff they could make "if they only had 4:4:4 wide-gamut 10-bit progressive intra-frame lossless raw" (to use every current buzzword in a non-coherent sentence), instead of going out and using what they have. Those that seem to believe that broadcast or theatres have no limitations.

-h