GH4 Can someone explain the 4:2:2 output from GH4?

[harddrive]

What you described is some form of Luma subsampling, and I don't know if that is something people do, because you want each pixel to have its own independent luma information.

Yes to the last bit, and in the simplest case (for luminance) just take the luminance pixels in 2x2 blocks and average the four. That will give you 1920x1080 "new" values for luma information, each independent. (Practically, I'm sure there's a better way to downscale.)

This is more of a theoretical thing you could do, but no one actually does.

If they don't, more fool them!

OK, more seriously, everyone has to work within the software they have and often exactly how these behave can be hidden. The real point is - it is theoretically quite possible to get "true" 1080 4:4:4 from 4k 4:2:0. Same obviously for 1080 4:2:2. Whether or not any given software is capable of doing it is another thing entirely. But there's no problem with the theory.

 

[joe12south]

Maths is hard! :cheesy:

Seriously, though, even if the concepts are hard for you to follow, it's trivial to do real world tests (with the sophisticated subsampling techniques commonly found in modern software) and observe the clear results. Put simply, you can reconstruct an amazing amount of "real" (useful) information if you manipulate suboptimal data within a known context.

 

[Barry_Green]

I'll try once more.

4k 4:2:0 means by definition:
Luminance - 3840x2160 samples Chrominance - 1920x1080 samples

1080 4:4:4 means by definition:
Luminance - 1920x1080 samples Chrominance - 1920x1080 samples

OK? So if you downsample the 4k luminance from 3840x2160 to 1920x1080 * and don't alter the chrominance * you end up with 1920x1080 samples for each.

In other words, 1080 4:4:4, it's as simple as that.

This post is exactly correct, with one minor minor clarification -- when talking about 4K, we're not talking about 4096, we're talking about Quad HD. Other than that, yes, this is how it works, and can be easily verified by simply taking a 4K still frame into PhotoShop and resizing it to 1920x1080. I have done so with a frame from the HC-X1000 (the 1/2.3" 4K Panasonic camera) and verified that it does indeed deliver 10-bit 4:4:4 after the downconversion.

 

[Barry_Green]

We also have to remember that chroma subsampling is a method of lossy compression.

Yes, it is -- its original intent was to cut down the bandwidth of a video transmission signal by discarding visually unnecessary information.

However, we should further clarify that it is only lossy if actual valid information is being discarded. In the case of single-sensor Bayer pattern sensors at their nominal resolution, no information is being discarded. 4K 4:2:0 from a 3840x2160 sensor maintains 100% of the luma and 100% of the chroma resolution that the sensor is capable of producing, so while it's a technically lossy compression system, it is actually (for a 4:2:0 sensor) completely lossless.

If we were talking about a three-chip system, where the native signal coming off the chips is enough to populate a 4:4:4 frame, then 4:2:0 would be highly lossy, discarding 75% of the sampled color information. But when talking about a Bayer sensor, the sensor itself has no more color information than that which can be contained in a 4:2:0 sampling system, so a 4:2:0 encoding of a Bayer-pattern sensor is sufficient to serve as a lossless encoding system.

Let's say you have a Foveon camera, or 3CMOS/3CCD camera that shoots 4k UHD. So no debayering involved. Each pixel has its own RGB values. The camera grabs a frame, and encodes it as YCC 4k 8bit 4:4:4, so in the resultant frame, every single pixel has a unique 8bit value for each colour channel, a maximum of 256 possible shades/tints for each channel.

In this example, 4:4:4 encoding is necessary to preserve all the information the sensor is delivering, yes.

If you then re-encode it to 4k 8bit 4:2:2, you have now lost half the colour information. You can't get it back. That colour information is gone.

Yes, this is a brutal culling, literally discarding half the horizontal color resolution. When this resulting image is viewed on a monitor or computer, some method of interpolation or pixel-doubling will be necessary to fill in the data that was discarded in the conversion to 4:2:2.

By resampling at 4:4:4, all you are doing now is looking at each pixel individually, instead of blocks of pixels. So you won't lose anymore information, but you won't gain any new information either.

Yes-ish. You will gain no new samples, you will still be dealing with half horizontal resolution. However, a sophisticated conversion could be done which up-rezzes the 960x1080 color signal up to 1920x1080. It doesn't have to be as simple as just doubling the horizontal pixels, so you could gain some new information, but that new information would be *synthesized*, not *sampled*. It would not be sampled resolution, it would be synthesized up-rezzed resolution. It sounds a lot worse than it is though, because all Bayer images are basically synthesized instead of sampled, and yet they end up looking pretty good...

Even if you downsample to 1080p, only 1 of 2 things will happen, either pixels will be discarded completely, or a new pixel will be created which is a form of mathematical average of the other pixels surrounding it based on the algorithm used. Regardless of the algorithm used to downsample, information is lost.

Getting pedantic here but taking a stab at clarifying. If you're talking about your original hypothetical example of the natively-sampled 4:4:4, then yes of course you're correct, downsampling to 1/4 the size cannot be done in any manner other than losing information.

But, back to the original topic of this discussion thread, if you're downsampling directly from 4:2:0 4K to 1080p, it is true that you will lose luminance resolution, but you won't lose any color resolution. A 4:4:4 1080p encoding can hold all the color resolution from the original 4k bayer sensor scan.

And it could be argued that you may actually be gaining bit depth, if downsampling into a 10-bit frame; a cluster of four 8-bit pixels summed together can produce one 10-bit pixel (brightness only, no change to luminance bit depth).

As a 4:4:4 image, each pixel may have its own independent information, but that information is less than the original 4:4:4, and most likely less than the 1st re-encoding to 4:2:2.

If you're talking about an original 4k 4:4:4 downsampled to 1080p 4:4:4, then yes, of course it is true that each pixel holds less information than the four independent pixels that made it in the first place.

If we instead apply this idea to a 4:2:0 Bayer pattern downsampling to 1080p 4:4:4, then the statement is only partly true -- the luminance will be less; even at 10-bit depth there will simply be a significant loss in perceived luminance resolution. The chrominance, however, will be exactly the same as the original Bayer scan, a true 8-bit 4:4:4.

(Note we're ignoring the differences between whatever bit depth the sensor actually delivers, and the 8-bit encoding that we're presuming is taking place for all the digital encodes).

 

[h-munster]

Technically you can just sum the values, and throw it into a larger bit space. But it would look wrong.

Huh? The technique has been used for years, both at capture and in post. What looks wrong?

You would be losing colour information.

Certainly pixel location information would be lost, but not color depth. The loss of that location information is always the case when sacrificing resolution for increased bit depth.

Which is why that simplistic method isn't used.

It's used quite a bit.

Think of it like this, you have 2 pure (saturated) red channels, so each one has a value of 255. You sum them together for a value of 510 and put them into a 10bit space, which can support 0 to 1024. Now your pure red only looks ~50% as "bright".

That's probably not how such a conversion would be done, but, even so, there is still no loss of color depth in your example (except for the possibility that you might be putting more shades beneath the noise floor).

In summing two 8-bit pixels into a single 10-bit pixel, it is best to map the shades to span the entire 10-bit scale. In such a case the contrast would be correct, with no loss of color depth.

You could do it with 4 pixels instead, so you would have 4 8bit numbers, but you would still be losing information because you are essentially replacing 4 independent pixels with 1 pixel that is trying to approximate the information from those 4. It can't do that as accurately as having those 4 pixels, thus information is lost.

Yes. Again, information regarding the original pixel location is lost, but not color depth information. Such is the case whenever resolution is sacrificed for increased bit depth.

 

[atarijedi]

Yes, it is -- its original intent was to cut down the bandwidth of a video transmission signal by discarding visually unnecessary information.

However, we should further clarify that it is only lossy if actual valid information is being discarded. In the case of single-sensor Bayer pattern sensors at their nominal resolution, no information is being discarded. 4K 4:2:0 from a 3840x2160 sensor maintains 100% of the luma and 100% of the chroma resolution that the sensor is capable of producing, so while it's a technically lossy compression system, it is actually (for a 4:2:0 sensor) completely lossless.

If you have a sensor which outputs 4k/UHD, you then take that raw data, debayer it, and then chroma subsample it using a 4:2:0 ratio, you will be losing chroma information.

But, back to the original topic of this discussion thread, if you're downsampling directly from 4:2:0 4K to 1080p, it is true that you will lose luminance resolution, but you won't lose any color resolution. A 4:4:4 1080p encoding can hold all the color resolution from the original 4k bayer sensor scan.

Here's the thing, you either have to chroma subsample before or after resizing the frame. If you chroma subsample a 4k 4:2:0 to 4k 4:4:4, you are not gaining anything, and you are not losing anything. If you then resize that 4k 4:4:4 to 1080p 4:4:4, you will lose both luma and chroma information as the pixels in the 1080p frame are interpolated from the 4k frame. If you weren't losing any information, then technically you could go back to 4k 4:2:0 and it would look exactly the same, but we all know it won't.

And it could be argued that you may actually be gaining bit depth, if downsampling into a 10-bit frame; a cluster of four 8-bit pixels summed together can produce one 10-bit pixel (brightness only, no change to luminance bit depth).

You can gain bit-depth by resizing a 4k frame to 1080p. But you are still losing information as you interpolate 4 pixels into 1, as that 1 pixel can't accurately represent what those 4 pixels represented.

Huh? The technique has been used for years, both at capture and in post. What looks wrong?

Simple of summing pixels has never really been used in image scaling. Interpolation algorithms are what is used.

Certainly pixel location information would be lost, but not color depth. The loss of that location information is always the case when sacrificing resolution for increased bit depth.

I didn't say colour depth would be lost, I said colour information. If you have a 2x2 block of pixels, red and blue, and interpolate it into 1 purple pixel. You have just lost colour information.

It's used quite a bit.

Except it isn't. The algorithms used for image scaling interpolation are quite complex, and aren't just simple summing.

That's probably not how such a conversion would be done, but, even so, there is still no loss of color depth in your example (except for the possibility that you might be putting more shades beneath the noise floor).

In summing two 8-bit pixels into a single 10-bit pixel, it is best to map the shades to span the entire 10-bit scale. In such a case the contrast would be correct, with no loss of color depth.

Yes. Again, information regarding the original pixel location is lost, but not color depth information. Such is the case whenever resolution is sacrificed for increased bit depth.

Like I said before, I never mentioned losing colour depth. I stated losing colour information. They are not the same thing.

Colour depth (bit depth) is simply the range of shades/tints a pixel can have in each channel. Colour information is what colour that pixel is, that is the information that is lost when your chroma subsample at a lower ratio, or scale(resize) a frame down.

 

[Joshua Cadmium]

Yes, it is -- its original intent was to cut down the bandwidth of a video transmission signal by discarding visually unnecessary information.

However, we should further clarify that it is only lossy if actual valid information is being discarded. In the case of single-sensor Bayer pattern sensors at their nominal resolution, no information is being discarded. 4K 4:2:0 from a 3840x2160 sensor maintains 100% of the luma and 100% of the chroma resolution that the sensor is capable of producing, so while it's a technically lossy compression system, it is actually (for a 4:2:0 sensor) completely lossless.

If we were talking about a three-chip system, where the native signal coming off the chips is enough to populate a 4:4:4 frame, then 4:2:0 would be highly lossy, discarding 75% of the sampled color information. But when talking about a Bayer sensor, the sensor itself has no more color information than that which can be contained in a 4:2:0 sampling system, so a 4:2:0 encoding of a Bayer-pattern sensor is sufficient to serve as a lossless encoding system.

I'm sorry, but I agree with AtariJedi - this is wrong. There's no way that 4:2:0 would ever be lossless.

4:2:0 is averaging all color information into 2x2 solid blocks of color. Debayering, though, will give you per pixel color information.

Once you move away from the raw sensor data, the image starts to degrade. Raw on the Blackmagic cameras, for instance, is way better than even 4:2:2 ProRes HQ.

 

[Joshua Cadmium]

Here's a good overview of the math behind Bayer vs Full Color: http://www.dpreview.com/articles/7886047725/sigmainterview .

Check out this image in particular, though:

This shows the math behind Bayer being only 71% (1 / square root of 2) as good as a full color system. However, as an VP with Sigma attested in the article, "clever demosaicing reduces the gap". In other words, this math is a more simplistic debayer. You're going to get better than 71%, but not 100%.

The second equation also shows why, on a Bayer sensor, red and blue (basically chroma) will be as good as green (basically luma) at only a square root of 2 or 71% resize. [R[SUB]avg (r,b) [/SUB]* √ 2 = R[SUB]avg (g) [/SUB]] . Remember, though, that this is simplistic debayering.

Now, that being said, would I like a Sigma Foveon 4K video solution? Yes - absolutely yes (http://www.dpreview.com/reviews/sigmadp1/page20.asp) but it's probably not going to happen anytime soon.

The point, though, is that 4:2:0 is not going to accurately capture everything. Of course Bayer is not as good as full RGB color (although I think it is better than 3CCD - no optical artifacts from the prism and dichroic filters, no artifacts from the sensors being offset, etc.), but it's better than what others are presenting here.


Check out these papers for more info:

http://web.archive.org/web/20130226103859/http://www.unc.edu/~rjean/demosaicing/demosaicing.pdf

http://research.microsoft.com/pubs/102068/demosaicing_icassp04.pdf

http://www.csee.wvu.edu/~xinl/papers/demosaicing_survey.pdf

http://hal.archives-ouvertes.fr/docs/00/68/32/33/PDF/AEIP_SOUMIS.pdf

Also, here is a nice comparison of a bunch of commercial debayering: http://www.linuxphoto.org/html/test_demosaicing.html

 

[h-munster]

Simple of summing pixels has never really been used in image scaling. Interpolation algorithms are what is used.

In capture, summing sometimes occurs before any "algorithms." Otherwise, summing can be a part of the "algorithms," both at capture and in post.

I didn't say colour depth would be lost, I said colour information. If you have a 2x2 block of pixels, red and blue, and interpolate it into 1 purple pixel.

A purple pixel? That would seem to be adding unnecessary complexity. The idea is to sum pixels of the same color channel into a single pixel of that same color channel, but with a higher bit-depth.

You have just lost colour information.

You don't really lose any color information -- just resolution/pixel location information.

Except it isn't. The algorithms used for image scaling interpolation are quite complex, and aren't just simple summing.

Again, summing during capture is often done at a lower level, prior to any "algorithms" (if they come into play at all). Otherwise, summing is inherent in the "algorithms."

Like I said before, I never mentioned losing colour depth. I stated losing colour information. They are not the same thing.

Don't want to argue semantics, but color depth is really the entire possible range/scope of color information of an image/system.

Colour depth (bit depth) is simply the range of shades/tints a pixel can have in each channel.

Bit depth and color depth are not the same property. That is a common misconception.

As you state, bit depth is the number of shades possible per channel in a digital image/system.

Bit depth is actually a factor of digital color depth, as is resolution. If color depth were only bit depth, then 1-bit images would not be able to exhibit rich colors and shading, and analog imaging would not be possible (no bit depth in analog images). The basic relationship between color depth, bit depth and resolution in digital imaging is expressed with this formula:
COLOR DEPTH = (BIT DEPTH x RESOLUTION)[SUP]3[/SUP]

Colour information is what colour that pixel is, that is the information that is lost when your chroma subsample at a lower ratio, or scale(resize) a frame down.

By "pixel," I think you mean "pixel group," as in an "RGB pixel group."

You don't have to lose any color information when you lower the resolution -- merely sum the color channels of the pixels in the original image, and then use a greater bit depth to contain those new, larger scaled values.

 

[harddrive]

Check out this image in particular, though:

This shows the math behind Bayer being only 71% (1 / square root of 2) as good as a full color system. However, as an VP with Sigma attested in the article, "clever demosaicing reduces the gap". In other words, this math is a more simplistic debayer. You're going to get better than 71%, but not 100%.

To be precise it shows the GREEN channel (with two samples instead of four) being 71% as good as a full system

The second equation also shows why, on a Bayer sensor, red and blue (basically chroma) will be as good as green (basically luma) at only a square root of 2 or 71% resize. [R[SUB]avg (r,b) [/SUB]* √ 2 = R[SUB]avg (g) [/SUB]] . Remember, though, that this is simplistic debayering.

That is absolutely, positively, NOT what the second equation says. You're misreading the diagram.

If there were four red photosites/2x2 block, it's saying (left hand side) the average resolution for all axes would be sq rt 2 x the no in the brackets (so about 1.414x"M"). (I can't type the fractions, so lets call it "M" it's the same in all four cases.

In the actual (Bayer) case, (right hand side) with one red photosite/2x2 block it's giving the figure as 1/sqrt2 xM - or about 0.707 M. In other words, half on each axis what a full system would give you.

This is spelt out at the bottom of the diagram: The resolution of a system measuring red or blue at all positions will be 2 times greater than that of one measuring at every forth [sic] position. The only thing I'd add to that would be "on each axis", and really it shouldn't come as any surprise. If you've got 1920x1080 red or blue photosites in a 3840x2160 Bayer sensor, red or blue resolution can never exceed 1920x1080.

If we make the assumption that green resolution represents luminance, and red, blue chrominance, then that diagram shows us that for a given sensor resolution (call it B), Bayer filtration will give us .707B for luminance, .5B for chrominance on each axis. Half sensor resolution for chrominance on each axis is equivalent to 4:2:0.

Now, that being said, would I like a Sigma Foveon 4K video solution? Yes - absolutely yes (http://www.dpreview.com/reviews/sigmadp1/page20.asp) but it's probably not going to happen anytime soon.

OK, the maths is above, but many might say "what about the real world?", and be sceptical. Your link above answers that better than I ever could.

What Barry and I are saying is, in essence, that for a Bayer sensor the chrominance resolution will be worse than for luminance - are we agreed?

The way to check that is to use a coloured chart. If the performance on coloured lines is the same as on black-white, we're obviously wrong. If the performance on coloured lines is substantially worse than black-white, we're right. Does that make sense?

So - look at Joshuas link above, and I for one are in no doubt at all. Ignore the absolute camera-camera comparisons - what's significant is how EACH CAMERA varies between b/w and coloured lines. I can't do better than quote from the article: ".......it only becomes apparent on the color test chart how well the camera resolves color detail. The DP1 [Foveon, or "4:4:4" sensor] produces detailed and clean images for all color combinations while the Nikons' Bayer sensor gets visibly confused by some of the color combinations which results in artifacts and blurring."

The real killer is the red/blue chart (hardly surprisingly) but whilst the Foveon manages it pretty well (cf the b/w chart) the Bayer sensors don't do remotely as well.

That's not to dismiss Bayer per se of course - the article goes on to say "....like in the black and white scene the D60 and D40 can partially make up for the DP1's better per pixel resolution by simply outputting larger images". In other words, higher pixel counts, or oversampling. But it proves pretty conclusively that for any given Bayer sensor, it's chroma resolution is not as good as for luminance - see above for the maths.

More sophisticated deBayering can indeed help - but only up to a point. It can make better "guesses" as to what the red, green (say) values might be expected to be at a pixel where we only have a real value for blue - but they are still guesses, or interpolations if you want to be technical! It can't put back information that has been lost for ever by the lower number of red, blue photosites compared to green.

 

[Stuart U]

So what is recommended as far as downsizing 4K to 1080: ProRes 4:2:2 or 4:4:4?
Or…?

 

[h-munster]

So what is recommended as far as downsizing 4K to 1080: ProRes 4:2:2 or 4:4:4? Or…?

The more color information that you have in the original image, the better your results can be. So, unless there is some practical/limiting reason to do otherwise, it is best to start with a 4:4:4 image of the highest bit depth (or raw).

 

[Stuart U]

MR. Munster: I understand that but there are a few practical limitations with starting with 4:4:4 or raw. One being I'm using a GH4. So now that I've shot GH4 4K and wish to finish in 1080 ProRes, what would you scholars recommend?

 

[h-munster]

Start out with as much color depth as you can, which means shoot with the highest res/quality that is possible/practical on your GH4.

If you have a Ki Pro Quad or an Atomos Shogun, use the highest res at 10-bits. If not, record internally at the highest quality (cinema 4K at 8-bits and 4:2:0?).

 

[Stuart U]

1: I have already shot 4K. Best quality in camera.


2: I now wish to finish in 1080, let's say ProRes, and maintain best quality in post throughout. is 4:4:4 recommended? Or is that unnecessarily bloating file sizes? Or 4:2:2? Or…?


Anyone care to weigh in with their best practice, or provide recommendations…?

 

[h-munster]

2: I now wish to finish in 1080, let's say ProRes, and maintain best quality in post throughout. is 4:4:4 recommended? Or is that unnecessarily bloating file sizes? Or 4:2:2? Or…?

Anyone care to weigh in with their best practice, or provide recommendations…?

Take a look at this thread from another forum. It's been a while since I read through it, but, as I recall, some folks were reporting that their NLEs could do the transcode with no loss of color information.

Here is another thread from the same forum regarding a Mac command-line app that does the transcode.

 

[Joshua Cadmium]

That is absolutely, positively, NOT what the second equation says. You're misreading the diagram.

You're right, it's not what the second equation says, but that's not what I wrote. As I put it, "The second equation also shows why, on a Bayer sensor..." You're misreading me.

If you compare the equation for Bayer red & blue [ R[SUB]avg (r,b)[/SUB] ] to the equation for Bayer green [ R[SUB]avg (g) [/SUB]] - the difference for red and blue compared to green is only 1 / √2 . So, again, for a Bayer sensor, red and blue debayered will at least equal green debayered at a 71% resize. Of course it's not going to be as good as a full color sensor, but that was not what I was arguing.

I was trying to best represent your argument, which you have agreed I've done, in order to better refute it, all at the same time.

If we make the assumption that green resolution represents luminance, and red, blue chrominance, then that diagram shows us that for a given sensor resolution (call it B), Bayer filtration will give us .707B for luminance, .5B for chrominance on each axis. Half sensor resolution for chrominance on each axis is equivalent to 4:2:0.

Since we're getting more technical, I need to say that chrominance subsampling shouldn't really be used for raw data or even freshly debayed data, except to say that there has been no choma subsampling going on. While I get your argument that the numbers describe a ratio (and it does make sense on it's face) it is not entirely accurate.

Even in 4:4:4 sampling, you could be losing color information. That would happen if the data was transformed from pure RGB to the more common Y’CbCr, a holdover from black and white television signals, with inaccuracies in the process.

However, you have to transform color from RGB to Y’CbCr if you want to subsample chrominance (Cb and Cr). 4:2:2 and 4:2:0 do not exist in RGB land, only Y’CbCr. More importantly, because of this, Cb and Cr do not equal blue and red (http://wolfcrow.com/blog/understanding-luminance-and-chrominance/) . So, when you go to subsample Cb and Cr, it's not the red and blue directly from the sensor. You're slicing up data that's already had a lossy conversion. Not ideal.

Describing a Bayer pattern sensor as 4:2:0 also fails to take into consideration that debayering fills in the gaps for all three RGB colors. You will get full color information for each pixel (4:4:4, or non-subsampled RGB) Now, the colors will be interpolated at certain points, but there aren't going to be actual holes in the data.

Basically, even 4:2:2 can be far removed from the actual RGB data.

Now, again, I agree that not having to fill in the gaps on a Bayer sensor will be better. However, to your original argument, 4:2:0 will look substantially worse than the original Bayer RGB data.

To argue otherwise would ignore decades of research into debayering, and treats a 4:2:0 container as if the colors were still raw, when really they've already been transformed before being sliced up.

 

[Thomas Smet]

1: I have already shot 4K. Best quality in camera.


2: I now wish to finish in 1080, let's say ProRes, and maintain best quality in post throughout. is 4:4:4 recommended? Or is that unnecessarily bloating file sizes? Or 4:2:2? Or…?


Anyone care to weigh in with their best practice, or provide recommendations…?

Well now that kind of depends on you, what you are shooting, and what NLE you are using.

1. What looks good to you? Try both and see how they look. Use your own eyes and judge the quality. Is the 4:2:2 good enough? Chroma sub sampling is always treated like an evil but it doesn't have to be. The reason why video has chroma subsampling is because our eyes just don't handle color as well and there is a minimal visual loss. If 4:2:2 works for you then by all means it is a high quality standard. 4:4:4 is perfection of course but do you need it? That has been the point of others. What do you really need?

2. What kind o projects are you shooting? I hate to be one that applies quality rules to certain types of projects but you really do sometimes have to weigh this sort of thing. To be honest 4:4:4 gains you very little in most real world situations and going through all the effort can sometimes be a waste for little to no gain. There is also the point of starting to obsess over pixels instead of the content or the story. In the end the energy you put into the production of your project will matter 1,000x more then 4:2:0 vs 4:2:2 vs 4:4:4.

3. The NLE can make a pretty big difference. Sometimes just the act of using 4k material in a 1920x1080 (sequence/timeline/whatever your NLE calls it) is enough to gain the benefits of 4:2:0 down scaling to 4:4:4 on the fly without th need to render out 4:4:4 master files. Sure you may get less editing performance but the storage savings would be huge. Some NLE's down convert a bit better than others and some have different modes of down scaling. Take Adobe Premiere for example that uses different down scaling methods based on render settings and if you use the CPU or GPU for rendering. Even the lamest NLE with the poorest down conversion is still converting 4k samples down to HD samples and will see the benefits from the increased chroma ratio. The way those pixels are calculated may vary but you still end up with HD 4:4:4.

 

[harddrive]

You're right, it's not what the second equation says, but that's not what I wrote. As I put it, "The second equation also shows why, on a Bayer sensor..." You're misreading me.

OK - you put "red and blue (basically chroma) will be as good as green (basically luma) at only a square root of 2 or 71% resize. - which I took to mean red/blue will be as good as green - and they will all be at 71% of sensor resolution. I now realise that's not what you meant after:

If you compare the equation for Bayer red & blue [ R[SUB]avg (r,b)[/SUB] ] to the equation for Bayer green [ R[SUB]avg (g) [/SUB]] - the difference for red and blue compared to green is only 1 / √2 . So, again, for a Bayer sensor, red and blue debayered will at least equal green debayered at a 71% resize.

I think we're agreed that the chroma resolution will be about 71% of that for luminance - and the latter will be about 71% of the sensor dimensions? Hence chroma will be about 50% of sensor resolution?

What becomes important is what is referenced by "4" in the 4:x:x notation, which is probably best described as the luminance resolution of the recording system. And for the purposes of this thread we've been assuming 4k, at least the quad-HD variant of it, and also that the sensor is the same - 3840x2180.

If we were talking about an oversampled system (such as the F3 or the Alexa for 1080 HD) then I agree my argument is no longer valid - but that's because the luminance resolution of the recording system is then less than basic sensor resolution. It would indeed follow that the available chroma resolution in this case would be more than 4:2:0 sampling could capture. To put numbers to it, then assuming a Bayer sensor of 2715 horizontally, you should get lum res of about 1920, and chroma res of about 1360. Downscale, and record in a 1920x1080 4:2:0 system, and it's obvious that chroma resolution will be lost. That would only be capable of 960 horizontally - yet up to 1360 could be generated.

But that's not the case here, not when the recording resolution for luminance is the same as sensor resolution.

Describing a Bayer pattern sensor as 4:2:0 also fails to take into consideration that debayering fills in the gaps for all three RGB colors. You will get full color information for each pixel (4:4:4, or non-subsampled RGB) Now, the colors will be interpolated at certain points, but there aren't going to be actual holes in the data.

Well, it's not accurate to describe the Bayer sensor as 4:2:0 for a number of reasons, but going back to my first post in the thread I said "In other words, what's the point of anything higher than 4:2:0 sampling? Think of it as a container. If you've got 1 litre of liquid, what's the point of putting it in a 2 litre bottle? Similar here." And OK, it should have had the caveat about assuming recording resolution is the same as sensor resolution.

And whilst deBayering may "fill in the gaps", and give a value for G,B in a location where's there is only a R photosite (same for the other colours) it doesn't give "full color information for each pixel (4:4:4, or non-subsampled RGB)" - but I think you acknowledge that yourself when you say "Now, the colors will be interpolated at certain points......

And in practice, interpolating colour information at certain points, is exactly what happens when a 4:2:0 signal gets viewed, processed, whatever. Assuming the chroma resolution is only starting off at half sensor resolution (which I think we're agreed upon now?) the advantage of doing the interpolation then is that far less data needs to be stored, transferred, compressed etc

 

[Stuart U]

Thanks Thomas. I've been a (fine art) stills guy for a very long time, I have made a few shorts and music videos in the past and am now in the process of resurrecting that. I'm about to update my computer etc. and was really fishing for a good point of departure as far as testing workflow options are concerned. So many thanks for weighing in with your thoughts and sharing your experience.