AVC Codec Curveball

[Barry_Green]

1. doesn't hvx200a shoot in DVCPRO HD that is able to do 100mbps.

Yes it does. And you're standing on the edge of a slippery slope...

Even as we established if the 21mbps PH mode can beat 48mbps equivalent of h264, surely the odds are less that it can stand up to 100mbps DVCPRO HD which is mroe than 4 times the bit rate.

<slip> and there you go, sliding down that slope, plunging down into the morass of confusion, deep in the pit where people think that "a megabit is a megabit".

So here's my question for you: I'll trade you a fifty for one of your twenty dollar bills. What do you say -- deal?

(wouldn't you want to ask "fifty of what"? Fifty pesos? Or fifty euros? makes a big difference, doesn't it?)

So that leads me to ask: "How is it possible that AF100 can then have higher image/visual fidelity than HVX200a?"

Because a peso isn't a euro.

And a megabit isn't a megabit.

Is the sensor the other variable that allows it to surpass HVX200a footage?

There are many things that contribute. The larger CMOS chip on the AF100 shows a lot less noise than the smaller 1/3" CCD on the HVX200A. The megapixels of the AF100's sensor make for sharper imagery than the spatial-offset HVX sensors. The codec is pretty much an equalizing factor; I would say that the two codecs are probably fairly similar in overall performance (although DVCPRO-HD is a dream to edit with).

So, again, separate the variables out. And the first thing to learn is that a megabit of this has no relationship to a megabit of that.

So here's the question that should boggle your mind next -- would 21mbps of AVCHD look better than 100mbps of DVCPRO-HD... FROM AN HPX3700?!?! That's the question. And I don't know the answer. But we do know that HVX200 to HMC150, they look about the same.

2. Which then further leads me to ask, how much of a difference does the sensor make?

The ONLY way to know, would be to have identical systems, and swap out only the sensor. And that's not possible.

And is anyone else troubled by the fact that Jan herself does not even know what sensor the AF100 is using?

Who says she doesn't know? Just because she doesn't tell us the details, doesn't mean she doesn't know the details.

Is this intentional corporate opaqueness in order to conceal the fact that AF100 in fact merely has a DSLR sensor with a slightly better codec so that people don't just call AF100 a "DSLR in a camcorder body!" ?

First of all -- who cares? Look at the images it delivers, and then decide whether it does the job you want done.

Secondly, it is most definitely not the same as in the GH1/GH2. Unquestionably it is vastly improved over the GH1, and the GH2 has all the same aliasing issues, that the AF100 overcomes. So if it's a GH2 sensor with a proper OLPF, then that's what it is. If it's a modified GH2 sensor with a lower pixel count and proper OLPF, then that's what it is. If it's a brand-new-from-the-ground-up design, then that's what it is. Does it matter to you? And if you knew the answer, what would you be able to do with that information?

Look, here's the deal -- I know folks are caught up in talking about this kind of stuff because there's no real footage to discuss. And all this whose-sensor-is-bigger-than-whose stuff will disappear once the footage hits the screens and the jaws hit the floor (or, don't, as the case may be; beauty is of course in the eye of the beholder). All I can say is -- wait for the footage. That's when the discussions of merit and substance can take place.

 

[Barry_Green]

Wow. Strong words. Strong, strong words. True even for the feature?

What I'm saying is that I don't think it's a necessity. Might want it for some certain shots, just like you might rent a fisheye lens for a certain shot, but I don't believe it's mandatory, and I'm now making the prediction that 95% of AF100 buyers will never bother to buy an external recorder.

I intend to test this more when I have a final-codec version, but right now, based on what I'm seeing, I just don't see why people would feel they "need" to use an additional recorder. On a $300k+ budget, I'd probably rent a Nano and use it, not just for its higher quality, but also for redundancy and the margin of safety that redundant recording would provide. But for most work, no, I just don't think folks are going to find it necessary.

 

[Barry_Green]

But I think your acceleration analogy is very applicable, because with video we're not just speaking of static image quality (the equivalent to horsepower). A racing vehicle needs a good drivetrain and suspension to effectively apply its engine torque to reaching top speed as quickly as possible. Similarly, a real-time AVC codec needs high CPU bandwidth and generous memory resources to handle rapid motion scenes in the limited time interval allotted to encoding each GOP.

There's no question in my mind that AVC has proven itself to be a highly efficient video codec in non-real-time applications such as Blu-ray disc encoding. But this bitrate-saving encoding efficiency requires considerably more horsepower than more straightforward high-bitrate techniques. How effectively and reliably its advanced coding techniques can be applied within the real-time hardware constraints of the AF100 is something I look forward to evaluating, hopefully in the very near future.

Exactly. And this is the kind of hardware support that they can afford to put in a $5,000 camea body, that they can't put in a GH1 or a DSLR or a $800 Canon consumer camcorder.

 

[Homunculus]

And if you knew the answer, what would you be able to do with that information?

Rule the world.

Kidding aside. Yes Jan 100% unequivocally stated a couple weeks ago she has NO clue what the sensor is inside. I can find you the exact post if you don't believe me. Jan has no idea what sensor the AF100 is using, she said the engineers never told her.

Now the question is should that matter? The answer to me is you're right it doesn't because if the aF100 IS a "dslr in a camera body" why should that matter as long as the other factors of it overcome the problems that DSLR's had. After all when one says "dslr in a camera body" it seems to have a negative connotation like "it's a pig with lipstick on" but in reality there's nothing wrong with DSLR's other than the few shortcomings they have so if it IS a DSLR brain but with the DSLR short comings banished then there IS nothing wrong with that. Shoot isn't a Red One basically an advanced DSLR? Ok maybe I'm going too far with that but, point is if the AF100 DOES in fact have the GH2 sensor but with a better codec, or with modifications that make it better than a GH2 then that's cool. My only concern is that if it HAS the gh2 sensor but with the same codec and basically the same image quality then to me I need to know this information in order to decide what I will be buying. Sure for many of you the extra 4000$ is worth it for the other amenities like XLR inputs, form factor, etc, etc, but for me it's not because I don't make 100,000$ a year like some of you do and in fact I am broke so it's very important for me to know which choice to make. Imagine if you walked into a car dealership and said

YOU: "Hello sir I'm interested in purchasing this Pontiac Firebird."
Salesman: "Ah, excellent choice sir, it's a great car."
YOU: "Yes thank you. Now may I first ask, which engine does this car have, is it the 5.0L 300 horsepower or the 5.7L 375 horsepower engine?"
Salesman: "I'm sorry sir. I'm afraid I can't divulge that information to you. In fact the engineers that built this car put a surprise engine in there for you. But hey don't worry, the car is GREAT! I assure you."
YOU: "Umm...yes. But you see I'm putting down a major life investment with this amount of money on this vehicle. You're telling me I can't even see what engine that is in there?"
Salesman: "I'm afraid not sir. You'll just have to take my word for it. This car is GREAT!"
YOU: "Uh yeah. I think I'll go ahead and take a rain check on that, thanks for your time."

Imagine if that were to occur?

 

[Barry_Green]

That's not how it occurs though. You're making it sound like a Lexus is the exact same car as the half-price Scion, and the only difference is the badge that they slapped on it. But that's not the case.

I don't know if the answer is yes or no. For someone to whom that matters, well, I guess they'd buy the Chevy. And the GH2 is on the market for those folks.

For those to whom the Cadillac matters a lot more than whether or not the engine block is the same, they'll go for the Cadillac version.

I wouldn't classify the AF100 as a hobbyist camera. Hobbyists should be looking at the GH2. If you make your living off of your tools, then the AF100 is a no-brainer. If you're a hobbyist with not that much money to spend, then the GH2 is a no-brainer. If you're in-between, then ... yeah, that's where you have to spend your time doing your research to make that decision.

I'm sure when the final unit is released, specifications will be published, or at least many of the specs. But we're not at that point yet.

 

[Homunculus]

exactly I'm one of the poor sobs stuck in the middle. Definitely not a 'hobbyist' but not making a living from it either. It's my passion but for now I'm still trying to make festival films on my own dime and am not a professional freelance DP

 

[Barry_Green]

Well, yeah, then you're on the fence. You'll have to decide if the extra features, convenience, and alias-free images of the AF100, are worth the cost to you. Or if you'd be better off saving $4,000 and going with a GH2 and being able to shoot stills also.

 

[Homunculus]

Well, yeah, then you're on the fence. You'll have to decide if the extra features, convenience, and alias-free images of the AF100, are worth the cost to you. Or if you'd be better off saving $4,000 and going with a GH2 and being able to shoot stills also.

yes only problem is so far the gh2 footage has looked atrocious to me and half of the af100 has looked amazing. other half is split between looking bad to mediocre. so it SEEMS af100 visual quality (only thing that matters to me, not the extras) will be worth it to make that purchase.

 

[TimTheFoolMan]

OU: "Hello sir I'm interested in purchasing this Pontiac Firebird."
Salesman: "Ah, excellent choice sir, it's a great car."
YOU: "Yes thank you. Now may I first ask, which engine does this car have, is it the 5.0L 300 horsepower or the 5.7L 375 horsepower engine?"
Salesman: "I'm sorry sir. I'm afraid I can't divulge that information to you. In fact the engineers that built this car put a surprise engine in there for you. But hey don't worry, the car is GREAT! I assure you."
YOU: "Umm...yes. But you see I'm putting down a major life investment with this amount of money on this vehicle. You're telling me I can't even see what engine that is in there?"
Salesman: "I'm afraid not sir. You'll just have to take my word for it. This car is GREAT!"
YOU: "Uh yeah. I think I'll go ahead and take a rain check on that, thanks for your time."

Imagine if that were to occur?

This is remarkably plausible, when the reality is that without opening the hood, the driver wouldn't have a clue which engine was which. They want to know, just so they can impress their friends with the knowledge. The difference is, the salesman would just make something up instead of saying that they don't know.

I'm currently being asked to shoot an indie film for some guys who are bound and determined to go out and buy a 5Dii, even though my GH13 is available. Do they care what the images look like from the GH13? Nope. They've heard that someone else has a 5Dii, and none of their buddies have a GH13, so they want what their friends have, with no regard to what it means.

It's a telling point that they haven't, themselves, shot anything. Ever.

P.S. The scenario you're describing happened in 1978 when the T/A 6.6 (W72 400ci "serious" engine) was release alongside of the more sedate 6.6L (L78 403 ci engine) Trans Ams. There was about 20hp (per Wikipedia) difference between the engines from the factory, and "car guys" would look down their noses at anything that wasn't the T/A 6.6.

 

[Otis Grapsas]

Like I said, using PSNR only to compare different codecs will only take you so far - for really comparing you need to measure subjective response in a viewing test.

-h

Without ABX you could use the forum responses for that. It does not really apply to post processing though.

 

[Otis Grapsas]

Who are "we"? I think there are other, better, more established ways to express codec capabilities.

There are intra-prediction-only methods ("Intra"), and intra+inter methods ("Inter"). MJPEG, JPEG2k/DCI and DV-codec (and I assume "AVCIntra") belongs to the first cathegory. MPEG2/4, h.264, AVCHD, (and what you call "GOP") all usually belong to the latter (although there is nothing stopping you from using only the intra-parts of any codec).

We = those who work with codecs. We are discussing very standard implementations here. We don't have to dig very deep, we use what the camera manufacturers offer. It's a practical approach vs a theoretical one. The GOP aspect is indeed the most important difference and what most people use as a reference for the codec type.

 

[Otis Grapsas]

As you point out, a direct comparison of the AVC codecs alone would require a custom hardware setup. What I'm interested to examine is how stock AF100 footage will compare to optimized GH13 footage. It's a practical concern - I want to know how to make best use of each camera's capabilities.

Aside from image sensor differences, I can control virtually all other recording variables in both GH13 and AF100. Calibrating the two ISO scales is what I think may require some experimentation.

I think you can use the AF100 for comparing codecs. It can connect to anything. It's 8bit so there is a limitation to that, but it does offer the common hardware basis and we can actually use the real time implmentations that are actually used for video instead of offline encoding in editors and transcoding tools.

 

[Otis Grapsas]

So far, 21mbps AVCCAM is doing that -- and it's doing it better than 48mbps of Canon h.264, or 35mbps of MPEG-2.

They are all far worse than what you would expect given the bitrate and the encoding scheme. I imagine the hardware and the low power designs are very limiting.

 

[Otis Grapsas]

another question. Has anyone considered the thought that 10bit and 4:2:2 are actually just as crappy as 8bit 4:2:0 and are indistinguishable from it?

All the tests I've seen so far point to them being almost indistinguishable. Just curious cus there's been so much hullaballoo over 10bit 422, but maybe we should already start looking ahead to 12bit 444 or whatever is next because 10bit 422 doesn't seem to visually be demonstrably and appreciably different than 8bit 420. Yes it's always better to have more bytes to play with but I'm just saying, is this a valid argument? I mean I know 10 bit has 4 times the depth of 8bit in the color palette from 256 to 1000 some odd I believe but ...why is this barely ever noticeable at all?

2.75 million colors for 8bit YCbCr
11 million for 10bit YCbCr
16.7 million for 8bit RGB
1 billion for 10bit RGB

but not for AF100, it's limited to 2.75million color 8bit YCbCr even if you record 10bit. The sensors we use today are between 0.1 million to 16 million depending on the ISO setting. But that's only tonality. The external recorders will also improve motion and reduce artifacting. There are many applications that need as much help as they can get in these departments.

In order to see how the codecs compare, you need some serious motion in frame or camera motion and you need to put them through post processing.

 

[Otis Grapsas]

The sensor makes a lot of difference. I have encoded uncompressed material from a range of sensors using identical post processing, so the only differences are inside the sensor block, on the sensor/ad converters. Since the sensors were different resolution, I adjusted the bitrates to equal bits/pixel on all sensors. That's the true datarate.

In tonality: With good codecs you start to see the limitations of the lower quality sensors. They have higher noise and inferior tonality to begin with. After a certain bitrate point, you improve the codec and the tonality stays the same. That point is very low in low end sensors.

In motion quality and artifacts: The better codecs always help.

 

[Homunculus]

2.75 million colors for 8bit YCbCr
11 million for 10bit YCbCr
16.7 million for 8bit RGB
1 billion for 10bit RGB

but not for AF100, it's limited to 2.75million color 8bit YCbCr even if you record 10bit. The sensors we use today are between 0.1 million to 16 million depending on the ISO setting. But that's only tonality. The external recorders will also improve motion and reduce artifacting. There are many applications that need as much help as they can get in these departments.

In order to see how the codecs compare, you need some serious motion in frame or camera motion and you need to put them through post processing.

those sound like big numbers but once again, here let me give you this example.

A human cell let's say has 1,000,000,000 atoms a rinovirus mitochondria let's pretend has 10,000,000,000,000 atoms. If I showed you both would you be able to tell me which has more atoms or HOW many atoms? Ofcourse not because there's a diminishing returns that occurs at certain quantum points of human perception.

So my point is, whether 8 bit has 1 million and 10 bit has 1,000,000,000,000,000,000,000,000,000,000,000 can you as a human with your limited eye sight even tell the difference? It seems in most cases the answer is NO as has been proven in several tests already.

So who cares if it's 30 septillian more colors the point is can the human eye even distinguish that many? Probably not

 

[Otis Grapsas]

16 million colors is 256 shades per color. If you have a sky or a light creating a gradient on a monochromatic wall, you can notice 2.75 vs 11 million after color correction and the codecs add to the problem. A light gradient on the wall could be from 140/255 to 180/255 changing over a 1000 pixel distance. That creates distict steps on an 8bit YCbCr codec. If the gradient is in even lower levels, the codec will compromise it further and it will have blocks after a certain point. It's not something you want to boost in post. These things are important if you see them in post in the material you are actually shooting and find them objectionable. You have to put the budget into the equation also.

 

[Homunculus]

16 million colors is 256 shades per color. If you have a sky or a light creating a gradient on a monochromatic wall, you can notice 2.75 vs 11 million after color correction and the codecs add to the problem. A light gradient on the wall could be from 140/255 to 180/255 changing over a 1000 pixel distance. That creates distict steps on an 8bit YCbCr codec. If the gradient is in even lower levels, the codec will compromise it further and it will have blocks after a certain point. It's not something you want to boost in post. These things are important if you see them in post in the material you are actually shooting and find them objectionable. You have to put the budget into the equation also.

that's funny because someone already did a scientific test with his sony EX and compared 8 bit to 10 bit where he deliberatley ccreated a gray gradient and BOOSTED it in post and compared and the 8bit and 10 bit gradients were indistinguishable. did you see that link? can't remember where it's posted. but it pretty much debunks what you said!

 

[Homunculus]

my point being that practice is often different than theory. in theory there's huge mathematical disparities. in practice it appears the average human being can't even see the difference in those disparities with his meager human eye sight

but you're welcome to prove me wrong with proof. show me some 8bit vs. 10bit gradients!

 

[Steve Kahn]

that's funny because someone already did a scientific test with his sony EX and compared 8 bit to 10 bit where he deliberatley ccreated a gray gradient and BOOSTED it in post and compared and the 8bit and 10 bit gradients were indistinguishable. did you see that link? can't remember where it's posted. but it pretty much debunks what you said!

Homunculus can you look for that link?