Difference between 4.2.2 and 4.4.4 for a VFX heavy movie

[PDR]

This article from Alister Chapman seems to indicate that 10bit isn't as important as a high data rate to keep down noise, and that 10bit eats into your data rate.

http://www.xdcam-user.com/forum3/viewtopic.php?f=27&t=38

The first thing to consider is that a 10 bit codec requires a 30% higher bitrate to achieve the same compression ratio as the equivalent 8 bit codec. So recording 10 bit needs bigger files for the same quality.

He refers to "codecs", but what he's saying about 8bit vs. 10bit only applies to uncompressed acquisition formats. This wouldn't apply to something like AVC-I for example.

Properly implemented 10-bit compression (e.g. some AVC impementations) is actually more efficient than 8-bit encoding from the equivalent 8bit codec. It doesn't take 30% more bandwidth, you actually can use on average 5-15% LESS bandwith for equivalent quality, or better quality at equivalent bandwidths. ie. Higher PSNR at equivalent bitrates.

Have a look at page 3 of this Broadcast Engineering article for more info
http://broadcastengineering.com/hdtv/avch-encoding/index2.html

Some more info from Ateme (<2MB)
http://www.mediafire.com/?wy88vih0a0uc7f9



Now he is absolutely correct about low noise being important, but "noise" from camera and/or sensor is different than "noise" generated by compression.

Actually better compression or uncompressed preserves more noise from the signal sent by the camera that the onboard compression wouldn't have preserved. People always complain that uncompressed or low compressed HDMI/HDSDI captures are more noisy than the native onboard codec, but higher detail goes hand in hand with higher noise. You can selectively denoise in post, but you cannot recover details that were not recorded in the first place

 

[Duke M.]

I understand what you're sayinig, and that's a very interesting article, but the article you are citing is for Peak SNR. As the article itself says PSNR has very little correlation to what the human eye perceives.

"A known problem with PSNR is a lack of correlation with the human visual impression."

Another issue is that on the very first chart on page three it shows no difference between 4.2.0 and 4.2.2 at bit rates below 70Mbps in h.264, and very little difference up to 100Mbps (which is where most people run a Nanoflash). Of course AVCHD only reaches 24Mbps implying no difference and AVC-I takes more bandwidth than an A-B-I version.



The charts that really show a difference at low bandwidths are the 12bit files. There the difference is noticeable at low rates.

"Now he is absolutely correct about low noise being important, but "noise" from camera and/or sensor is different than "noise" generated by compression. "

Agreed. Looking beyond that puzzling BBC report on resolution and noise for the F3 (said report seems to have several factual errors and who knows what settings) the Sony stats for the F3 claim extremely low noise. The video clips I've seen do seem to have really low noise. (My camera hasn't arrived yet to test.)

Assuming that the F3 is extremely low noise with certain settings the question that I see is specific to the F3. Can I get cleaner video from the F3 recording to a Nano at 220-280Mbps with I-frame at 8 bit, than I'd get from the Ki Pro Mini at 100Mbps? (Keeping in mind that the Ki Pro Mini is 10 bit and the Nano is 8 bit, and I don't see banding as an issue if the profile settings are close to what you want in the first place.)

In other words will the higher data rate offset any 10 bit noise advantages? Or the converse. If a higher data rate preserves noise, wouldn't a 10 bit data file preserve more noise too?

 

[PDR]

I understand what you're sayinig, and that's a very interesting article, but the article you are citing is for Peak SNR. As the article itself says PSNR has very little correlation to what the human eye perceives.

"A known problem with PSNR is a lack of correlation with the human visual impression."

Yes, no one except signal engineers use PSNR. It's a very problematic metric, however it's still the one of the few objective measures we have and still widely used. You can demonstrate similar findings with SSIM, which is another objective measure. But again, it can be problematic.

The point is, "30% more bitrate" for 10-bit vs 8-bit only applies to uncompressed.

Another issue is that on the very first chart on page three it shows no difference between 4.2.0 and 4.2.2 at bit rates below 70Mbps in h.264, and very little difference up to 100Mbps (which is where most people run a Nanoflash). Of course AVCHD only reaches 24Mbps implying no difference and AVC-I takes more bandwidth than an A-B-I version.

The HVS can clearly see the differences between 422 and 420 at almost any bitrate, except very low bitrates, on individual frames, but less so under normal viewing conditions. PSNR is not a good measure for this.

Agreed. Looking beyond that puzzling BBC report on resolution and noise for the F3 (said report seems to have several factual errors and who knows what settings) the Sony stats for the F3 claim extremely low noise. The video clips I've seen do seem to have really low noise. (My camera hasn't arrived yet to test.)

Assuming that the F3 is extremely low noise with certain settings the question that I see is specific to the F3. Can I get cleaner video from the F3 recording to a Nano at 220-280Mbps with I-frame at 8 bit, than I'd get from the Ki Pro Mini at 100Mbps? (Keeping in mind that the Ki Pro Mini is 10 bit and the Nano is 8 bit, and I don't see banding as an issue if the profile settings are close to what you want in the first place.)

In other words will the higher data rate offset any 10 bit noise advantages? Or the converse. If a higher data rate preserves noise, wouldn't a 10 bit data file preserve more noise too?

Does the Ki Pro mini only record 100Mb/s ?

Prores needs ~160-190Mb/s to get similar results to 280Mb/s I-frame MPEG2 in terms of noise (similar numbers for DNxHD). But the benefit of 10-bit in terms of smooth gradients and post production workflow might be important as well depending on the type of content you are shooting, and what you plan to do in post

 

[macgregor]

This seems to stir debate every once and a while, I've posted this a couple times, so I'm sorry if I sound like a broken record posting it again. Here is a quote from John Galt, lead designer of the genesis camera, one of the only bayer pattern cameras in the world outputting true 4:4:4.

The Genesis was not bayer pattern. ;D

 

[TimurCivan]

Genesis was a manly ccd....

 

[nyvz]

Genesis was a manly ccd....

It was an RGB stripe pattern, right?

The Genesis was not bayer pattern. ;D

I know you guys are just being facetious, but basically the particular color pattern or sensor technology don't really change the meaning of the original statement since the idea is essentially that Genesis is an oversampled camera with at least as enough photosites and color samples at the sensor level to take advantage of 4:4:4 or RGB recording

 

[tom.wong]

for all intensive purposes, f35 and genesis are almost identical cameras. sony even designed the sensor for panavision, so they are gonna have a lot of similarities. different ergonomics, little bit different processing, maybe not even identical sensor but all based off the same principles. and I believe panavision uses their own custom log curve (Panalog) vs sony's S-Log. Both are RGB stripe.

DP's I know who use the genesis find it ergonomically very close to use a film camera, both can get pretty hulky though. I also heard f35 got a recent firmware update that boosts its native sensitivity to 800 now, not sure if that carried over to the Genesis too, currently rated somewhere in the 400-640 range.

 

[nyvz]

Regardless of the math, here is something I can tell you from practical experience.

The more data your fx artist has to work with, the easier it will be and the better the final composite will look.

I do notice a difference between 4:2:2 and 4:4:4 10bit s-log files. Especially for keying greenscreens.

If you can't afford the storage to shoot 4:4:4 slog for the entire show, then at least try to splurge for the greenscreens.

As I type this I am looking at a 4:2:2 log greenscreen in Nuke that came from an Arri Alexa.
Last month I had some Alexa greenscreens of comparable quality that were 4:4:4: log.
Without a doubt the 444 files felt fatter and keyed better.

In this case bigger really is better.

I agree that there is likely to be some benefit from 4:4:4 but it is hard to quantify how much. Your anecdote about the working with alexa footage is interesting, but it is hard to take as evidence since there are so many possible confounders that would detract from making any kind of causal correlation since you were comparing footage from two different shooting situations a month apart and you may have expected the 444 to be better if you knew before looking at it that it was 444. Do you have more details about what you observed and about the differences between the footage that you can share? Were the scenes exposed similarly with the same ISO settings?

Also, this is the F3 forum, so it may be worth considering for those really weighing the benefits of 4:4:4 on an F3, that the F3 sensor is only 3.36MP (~2444x1374 or something, I dont have the numbers from sony in front of me just that calculated resolution). That means that the sensor has ~1222x687 red and blue color sampling resolution at the sensor level. 4:2:2 gives you 960x1080 recorded color resolution and 4:4:4 gives you 1920x1080 recorded color resolution, but if we figure in the best case scenario the debayer and downsample does not cause a hit to color resolution in processing, and that resolving power will be limited by the weakest part of the system, in an F3 one might say that 4:2:2 will give you in the range of 960x687 color resolution, and 4:4:4 will give you 1222x687, a difference of only about 27% rather than the 100% one would expect from 4:4:4. Hard to say how much this affects real world performance, but when considering the cost and workflow implications of 4:4:4 it may be a tough decision for some if it is for only a 27% increase in color resolution.

I know good debayering makes this a bit more complicated than counting red and blue filtered photosites, but I don't imagine it can do that much to create red or blue pixel data where there were no samples at the sensor, right?

 

[macgregor]

That´s a very good perspective nyvz.

 

[zeke]

Here is a link to a reprint of an article about John Knoll at ILM on the difference in 4:2:2 and 4:4:4 for heavy VFX use on the final 2 Star Wars movies (Episoe 2 - 3)..

http://pro.sony.com/bbsccms/ext/cinealta/docs/FV_StarWarsReprint_5-05.pdf

Puts it in laymans terms, and seems appropriate for the OP original question.

 

[nyvz]

Interesting article, though it is worth noting that it sounds a bit like they might be comparing 10b HDCAM SR 444 RGB to 8b HDCAM 311, which is a much bigger difference than just 422 vs 444. It doesnt seem to state that specifically but they mention that the old system uses 1440, which sounds like HDCAM, which makes sense since episode 2 came out in 2002 and HDCAM SR didnt come out until 2003. Also the article seems to say that a big part of the benefit of 444RGB VS 422YUV is skipping the YUV-RGB conversions, which makes some sense but it doesnt have much to do with resolving power so much as color reproduction.

Also, the comparison of capturing 444 vs 422 from an F950 may not be as relevant to the F3 question since the F900/F950 uses 3 2.2MP CCD imagers that with a gross pixel count of 6.6MP with equal numbers of photosites dedicated to each color vs the F3 with half as many total photosites and only 25% of them being dedicated to R and B. Basically the sensor design of an F900/950 is designed for 1080pRGB output, bayer sensors, on the other hand are really designed with chroma subsampling built in to some extent, and the F3 isnt oversampled enough to counteract that. The RGB vs YUV thing is interesting but I have no experience with video images captured in full RGB, but it makes a lot of sense that it might work better in general and with VFX renders since VFX and image sensors are all going to be natively RGB.

 

[Lenilenapi]

man this shxt is complicated!

 

[ddh]

John Knoll ... "Don’t Waste Bandwidth If you have 10-bit gamma with curves that protect the top end of the exposure range, you have all the benefits perceived in log plus the benefits from working in gamma space in shadow areas. If you record an image in log space instead, you’re wast- ing bandwidth."
I'm a bit lost here. Is Knoll recommending a different approach or a tweakiing of something?

 

[zeke]

nyvz, it's his observations about the benefits of 4:4:4 over 4:2:2 for VFX work that is the core of the article. In a nutshell, at 4:2:2 your color resolution is only half the pixel resolutiion, since only aprox. half the pixels are being sampled. Plus the potential loss of info when converting from YUV to RGB. Getting it into RGB in the first place will help you pull better blue/green screen keys. More detail available. Doesn't mean you can't use 4:2:2, or they wouldn't have been able to make Episode 2. But they were able to do much better work with 4:4:4. Not only color, but highligts too. I think RGB plus SLog should enable one to do even better green/blue screen than before.

 

[nyvz]

nyvz, it's his observations about the benefits of 4:4:4 over 4:2:2 for VFX work that is the core of the article. In a nutshell, at 4:2:2 your color resolution is only half the pixel resolutiion, since only aprox. half the pixels are being sampled. Plus the potential loss of info when converting from YUV to RGB. Getting it into RGB in the first place will help you pull better blue/green screen keys. More detail available. Doesn't mean you can't use 4:2:2, or they wouldn't have been able to make Episode 2. But they were able to do much better work with 4:4:4. Not only color, but highligts too. I think RGB plus SLog should enable one to do even better green/blue screen than before.

Right, but my point is that the information is inconsistent which makes it unclear that his observations are completely relevant to the 422 vs 444 comparison. It makes sense that they may have used HDCAM for episode 2, but that would mean when they moved to HDCAM SR 444RGB they were not just upgrading from 422 to 444, they would have been upgrading from 8 to 10bit, 1440x1080 luma sampling to 1920x1080 RGB, 480x1080 color sampling to effectively 1920x1080, 140Mbps to 440Mbps etc. It is confusing to me that they talk about the benefits of going 444 over 422 but only have as evidence a project where they went from 311 to 444.

We know the mathematical benefits of 444 over 422. What it sounds like we are exploring here are the practical benefits.

 

[zeke]

Right, but my point is that the information is inconsistent which makes it unclear that his observations are completely relevant to the 422 vs 444 comparison. It makes sense that they may have used HDCAM for episode 2, but that would mean when they moved to HDCAM SR 444RGB they were not just upgrading from 422 to 444, they would have been upgrading from 8 to 10bit, 1440x1080 luma sampling to 1920x1080 RGB, 480x1080 color sampling to effectively 1920x1080, 140Mbps to 440Mbps etc. It is confusing to me that they talk about the benefits of going 444 over 422 but only have as evidence a project where they went from 311 to 444.

We know the mathematical benefits of 444 over 422. What it sounds like we are exploring here are the practical benefits.

You are correct, practical benefits. One other thing, in the article they talked about video gamma (linear), yet with S-Log you can now use a more Filmic S curve gamma and not lose any information. Very cool.
From a strictly practical point of view, if you are doing heavy VFX type movie, then capturing in RGB saves you a lot of headaches even if all you're doing is set extension.
I think the only way to resolve this is doing a lot of test footage before making any commitments. With all the other variables (lenses, lighting...) it's the only approach that makes sense. I'm willing to bet, with proper over exposure of green screen vs the subject, you could get by with 4:2:2. But blue? Not so sure. Preparation, preparation, preparation.

 

[TimurCivan]

IT seems the long and short of it is this, when in 10 bit, the jump from 422 to 444 isnt very noticeable.