GH4 Can someone explain the 4:2:2 output from GH4?
- Thread starter lifepo4
- Start date
joe12south
Veteran
Stuart U,
All else being equal, 1080P Prores 4:2:2 can store all of the information. No need to use 4:4:4 unless doing something else, something special, with the footage. (Certain types of VFX, for example.)
All else being equal, 1080P Prores 4:2:2 can store all of the information. No need to use 4:4:4 unless doing something else, something special, with the footage. (Certain types of VFX, for example.)
If you've got a 3840x2160 Bayer sensor camera and are shooting 4k, 4:2:0 sampling is optimal - it's as simple as that.
Thomas Smet
Veteran
To simplify all of this:
Both a bayer sensor and 4:2:0 have a loss of color data. Bayer is a bit more of a refined process at filling in those lost blanks than 4:2:0 but in the end you are dealing with pixel prediction in both cases so a loss is a loss and a pixel created with math is still a pixel created with math.
There are only two physical ways known to man and the law of physics to get a perfect 4:4:4 image. use a 3 chip sensor design or use a sensor 4x as large and down convert to 1/4 the resolution eg 3840x2160 to 1920x1080. In both cases each R,G and B channel will have 1920x1080 unique pixel values. Every other process is an interpolation or mathematical guess. Some a bit more refined and accurate than others but still pixel mush.
Both a bayer sensor and 4:2:0 have a loss of color data. Bayer is a bit more of a refined process at filling in those lost blanks than 4:2:0 but in the end you are dealing with pixel prediction in both cases so a loss is a loss and a pixel created with math is still a pixel created with math.
There are only two physical ways known to man and the law of physics to get a perfect 4:4:4 image. use a 3 chip sensor design or use a sensor 4x as large and down convert to 1/4 the resolution eg 3840x2160 to 1920x1080. In both cases each R,G and B channel will have 1920x1080 unique pixel values. Every other process is an interpolation or mathematical guess. Some a bit more refined and accurate than others but still pixel mush.
Thanks Thomas, all true.
Only other thing I'd add is that it makes design sense to match components in a system as reasonably as possible. If the sensor is capable of far better than the recording system, much of it's abilities will be wasted - so why not use a simpler and most likely cheaper sensor?
If the sensor is the limiting factor, why use a recording system that's is capable of far better than the sensor can deliver? All it will do is increase datarate and/or increase compression, and exacerbate storage and card speed issues for no real gain.
Which is why 4k 4:2:0 for a 4k bayer sensor makes sense.
Only other thing I'd add is that it makes design sense to match components in a system as reasonably as possible. If the sensor is capable of far better than the recording system, much of it's abilities will be wasted - so why not use a simpler and most likely cheaper sensor?
If the sensor is the limiting factor, why use a recording system that's is capable of far better than the sensor can deliver? All it will do is increase datarate and/or increase compression, and exacerbate storage and card speed issues for no real gain.
Which is why 4k 4:2:0 for a 4k bayer sensor makes sense.
Sangye
Member
I understand how you can get 4:4:4 from downscaling 4k to 2k. But how do you get 10-bit, as some people are claiming?
The way I see it, you're averaging four 8-bit images, and calling the result 10-bit. But it's really not the same. By that reasoning, downscaling 8-bit 8K footage to 2K would still be 10bit, only more truly 10-bit, since your sample size is sixteen rather than four. However, what if you first downscaled your 8-bit 8K footage to 4K, call it 10-bit, then downscale that to 2k. Do you now have 12-bit? I would say no, and there is clearly an error in the logic here.
However, averaging four 8-bit values together in 10-bit space, while not the same as true 10-bit, is still somehow better than 8-bit... right?
The way I see it, you're averaging four 8-bit images, and calling the result 10-bit. But it's really not the same. By that reasoning, downscaling 8-bit 8K footage to 2K would still be 10bit, only more truly 10-bit, since your sample size is sixteen rather than four. However, what if you first downscaled your 8-bit 8K footage to 4K, call it 10-bit, then downscale that to 2k. Do you now have 12-bit? I would say no, and there is clearly an error in the logic here.
However, averaging four 8-bit values together in 10-bit space, while not the same as true 10-bit, is still somehow better than 8-bit... right?
I find the colour space conversion relatively straightforward to visualise - I'm less sure about the 8-10 bit.
But to have a go at explaining the rationale.....
Let's assume on our 8 bit system black is represented by binary 0, white by 255, and on the 10 bit system by black 0, white 1023 (10 "1"'s)
Let's further assume a mid-grey that is 120 (say) on the 8 bit system, and 480 on the 10 bit system.
A simple algorithm for doing the downscale would simply take groups of 4 pixels, and add the values to get a single 10 bit value. So 120+120+120+120=480.
But it's possible that slight increase in level may make one of the original group of four be 121, whilst the rest remain 120 - the summation would then give 481 for that slight increase in input. A single bit in the 10 bit system, just what you'd expect, and the hope would be another slight increase will mean two of the pixels will be 121, two will still be 120...... and so on.
That's the way I've heard the theory explained, and in principle it should work. But..... it's making a number of assumptions which may not be valid, and I stress "may not". In particular, then in the example above, it's CONCEIVABLE that it won't act as hoped for. It's conceivable that from the original four 120 value state, then as the light is slowly increased then what was described above doesn't happen - but rather that nothing happens for a while, then all the values go to 121 at pretty much the same time. Hence the 10 bit values jump from 480 to 484 to 488...... You've got 8 bit values in a 10 bit wrapper.
It also takes no account of noise, and that could mask any theoretical benefits of 10 bit resolution anyway. There's be no point in looking at a set of kitchen scales, with the finest division being 1/10 ounce, and writing down you had weighed 8.163 ounces, would there? So what's the point of worrying about bit depth in a noisy video system? High end equipment is different (better s/n) - which is why 10 bit has real significance.
So that's the two sides of the argument. Practically, all I would say is that a downscale from 8 bit 4k to 10 bit 1080 isn't going to do any harm, albeit increasing the file size. Indeed, a few years back, when the EBU did some exhaustive codec tests, they gave XDCAM 422 full broadcast approval, even as an 8 bit codec. But it was mentioned that for *production* and post work, then even with 8 bit acquisition, there may be benefits in using a 10 bit system for post work.
If that raises as many doubts as answers, I'm sorry! All I will say for certain is that for a camera like the GH4 there is little point in worrying about whether the recording system is 10 bit rather than 8 bit in the first place, or 4:2:2 as opposed to 4:2:0. Move up to the F55 or the forthcoming Varicam, and 10 bit takes on a whole new significance.
Last edited:
h-munster
Active member
To clarify the scale that's involved, let's assume that each pixel in our 8-bit group of four is at maximum value -- 255. So, 255 + 255 + 255 + 255 = 1020.harddrive said:
Now we have a total of 10-bits worth of values in each pixel of our converted image.
That's exactly what happens. If one scans a row of pixels from left to right in a digital image, the colors and values usually start to vary at some point. Scanning the next row of pixels below, the points where such value transitions occur might differ from the row above by one or two pixels. Such slight differences in the location of these transitions is where/why the resolution factor of color depth is so important.harddrive said:
That only happens in the smooth areas of the image, As mentioned earlier in this thread, if banding exists in the 8-bit original image, it will appear in the 10-bit version, unless something artificial is introduced (dithering, noise, etc).harddrive said:
However, in the areas of the image that have more complexity, there is more variation within the 8-bit group of four pixels.
Actually, one of the primary motivations in sacrificing resolution for increased bit depth is to reduce noise.harddrive said:
Of course, the other main reason is to not lose color depth when down-scaling resolution.
Barry_Green
Moderator
Ah, well, that's why you're having a hard time understanding it -- in the 10-bit conversion you don't average at all. You add, but don't average.
If handled properly, the downscaled 4K 8-bit can certainly become 1080p 10-bit. As others have explained, it's really rather simple. Just remember that, unlike with the color resolution, with the luminance downscale there is no averaging being done. We're summing the pixels, not averaging them.
So -- if you had a sheet of black with one tiny gray rectangular speck on it, that might be sampled in a 1080p system as one pixel and it might (in an 8-bit system) have a value of 128. Pure black would be zero, and pure white would be 255 in this system. Let's say the gray speck is three pixels wide, so it gets sampled like this:
_____________
128 | 128 | 128
_____________
In an 8-bit system at 4K resolution, there would be four pixels sampled in the same space where only one 1080p pixel was sampled, and each 1080p pixel would result in 4 samples in a 2x2 pattern. Those four pixels might all be 128, but they might not. Let's say there's an extremely mild brightness difference from the left side being a tiny bit darker, and the right side being ever so slightly lighter:
_______________________
127 128 | 128 128 | 128 129
128 128 | 128 128 | 128 128
_______________________
Now, if this was an 8-bit system, you'd average those all together, and each pixel would come out to be 128. But, when you downconvert it and sum each block together, you get individual 1080p single pixels with slightly different values. The middle pixel would have a value of 512 (128+128+128+128), which, visually, would be a middle gray in a 10-bit system (0 = pure black, 1023 = pure white). But the left pixel would be the tiniest bit darker (511) and the right pixel would be the tiniest bit brighter (513). So each of our 4K pixels is using 8-bit values, but our summed 1080p 4-into-1 pixel is now representing a 10-bit value (511, 512, and 513).
So you have three distinct (but very similar) shades of gray. The 8-bit system would throw away the difference, but an 8-bit->10-bit conversion would preserve the difference.
It is at this point that we run into some potential hiccups. First of which is, as harddrive said, if there's visible noise in the system, it can easily cause fluctuations that can mask out or overpower the fine differences between the shades. So if your camera is exceptionally clean, then 10-bit is of value, but if your camera has some noise in the signal then 10-bit may actually net out to be nothing more than wasted space.
Secondly, can you see the difference? Maybe, maybe not -- it depends on your monitor and graphics card and display device. Some LCD panels are only 6-bits, some are 8-bits only. And some graphics cards may only process or display 8 bits. You might have 10-bit data, but it may look visually identical to 8-bit data simply because your computer system is incapable of displaying the finer shades. It may be counterproductive for people to turn to YouTube or Vimeo footage to try to see the differences, when those systems are almost certainly truncating their video footage into 8-bit...
Last edited:
thorn
Active member
But as has been said so many times now on this thread - only if the source justifies it.
Progressive or interlace? That makes a big difference. But this thread is about 4k 4:2:0, and how it can be scaled down to 1080, 4:2:0 1080 is a different matter. Though it depends where it comes from - if a 1920x1080 Bayer sensor, 4:2:0 or 4:2:2 won't make much difference for all the reasons over the last few pages. A camera with 3 1920x1080 chips is a different matter.
Barry_Green
Moderator
Quoted for truth. 4:2:0 in interlaced is a nightmare. 4:2:0 in progressive is much more reasonable. And 4:2:2 (from a source that justifies it) can be much better.
A GH4 (or HC-X1000 or other 8-bit 4:2:0 4K camera) downconverting to 1080p would be very well served by being recorded as 4:2:2. But a single 1080p sensor, recording 4:2:2, is pretty much pointless (unless you plan on downconverting to SD, I guess...)
Again, QFT. Three chips can deliver a full 4:4:4 signal, and recording in 4:2:2 will show definite and substantial benefits over recording that signal in 4:2:0.
Joshua Cadmium
Well-known member
I'm still sticking to what I wrote, which is that 4:2:2 is going to look substantially better. Again, GH4 4:2:0 as recorded comes with a lossy color space conversion plus compression, which means you don't have the original color fidelity anymore.
If you want to describe all Bayer sensors as inherently 4:2:0, it makes sense on its face, except for the fact that interpolation is better with more advanced debayers, so that the net result is better than 4:2:0.
For instance, Red (who I know is not unbiased) explains it like this: "those with a video encoding background may want to try and apply the 4:2:2, 4:1:1, etc. categorizations to a Bayer sensor, but this terminology is intended for compression methodologies and final images, not the sensors themselves. A 4K Bayer sensor is capable of producing full 4K 4:4:4 RGB files, for example; 4:2:2 is what could be applied to this file afterwards." - http://www.red.com/learn/red-101/bayer-sensor-strategy
Arri agrees with this as well, which is why the Alexa (which has a Bayer sensor) can output 4:4:4 RGB and 4:2:2 YCbCr.
You don't want to loose any color information, even if it does have interpolated data, since the chroma subsampling is a step you can't come back from.
Once subsampling occurs (because it can only be done in YCbCr) the original color fidelity is compromised.
It's easy to comprehend - since you're not working with the raw data, any step in the compression chain is going to negatively impact the image.
Sangye
Member
Thank you @Barry_Green! That was an amazingly clear explanation.
One more question. I mulling over a few different workflows for getting the best quality 2K/1080p output from my 4K/UHD source footage:
1. I can downscale all of my footage and convert it to ProRes at the same time, using 5DtoRGB Batch. This is what I used to use with my GH2, and it still works well with my GH4. I use ITU-R BT.601, Full Range of Luminance, and sometimes apply Technicolor CineStyle as a default LUT if I shot flat but don't want to do heavier color-grading later.
2. I could do the same as above, minus the LUT, using Adobe Media Encoder.
3. I could pass my 4K / UHD footage through Adobe Media Encoder, converting it into a 4K / UHD ProRes 422 preset, thereby putting the 8-bit values into a 10-bit container, and ostensibly allowing for 10-bit calculations to happen automatically when I downscale it later?
4. I could import the original 4K/UHD footage directly into Premiere on a 4K or UHD sequence, and when I'm ready to export, choose a 2K or 1080p output profile (usually h.264). In this case, I would assume that toggling on "Render At Maximum Depth" would be advisable?
5. I could import the original 4K/UHD footage directly into Premiere, on a 10-bit 2K or 1080p sequence, and scale the footage down by 50%. I would assume that here I should still enable Render At Maximum Depth at export?
Basically, the tools most readily at my disposal are 5DtoRGB Batch, Adobe Media Encoder, and Adobe Premiere. As a Creative Cloud subscriber I do also have access to After Effects, but I don't want to add that to my workflow if I don't need to. So, what's the best way to get 10-bit 4:4:4 from my GH4's footage, using these tools?
edit -- over on EOSHD, a user shared a command-line tool for converting the GH4's 4K footage into 4:4:4 10-bit 2K. But isn't it possible and much easier using one of the options above?
One more question. I mulling over a few different workflows for getting the best quality 2K/1080p output from my 4K/UHD source footage:
1. I can downscale all of my footage and convert it to ProRes at the same time, using 5DtoRGB Batch. This is what I used to use with my GH2, and it still works well with my GH4. I use ITU-R BT.601, Full Range of Luminance, and sometimes apply Technicolor CineStyle as a default LUT if I shot flat but don't want to do heavier color-grading later.
2. I could do the same as above, minus the LUT, using Adobe Media Encoder.
3. I could pass my 4K / UHD footage through Adobe Media Encoder, converting it into a 4K / UHD ProRes 422 preset, thereby putting the 8-bit values into a 10-bit container, and ostensibly allowing for 10-bit calculations to happen automatically when I downscale it later?
4. I could import the original 4K/UHD footage directly into Premiere on a 4K or UHD sequence, and when I'm ready to export, choose a 2K or 1080p output profile (usually h.264). In this case, I would assume that toggling on "Render At Maximum Depth" would be advisable?
5. I could import the original 4K/UHD footage directly into Premiere, on a 10-bit 2K or 1080p sequence, and scale the footage down by 50%. I would assume that here I should still enable Render At Maximum Depth at export?
Basically, the tools most readily at my disposal are 5DtoRGB Batch, Adobe Media Encoder, and Adobe Premiere. As a Creative Cloud subscriber I do also have access to After Effects, but I don't want to add that to my workflow if I don't need to. So, what's the best way to get 10-bit 4:4:4 from my GH4's footage, using these tools?
edit -- over on EOSHD, a user shared a command-line tool for converting the GH4's 4K footage into 4:4:4 10-bit 2K. But isn't it possible and much easier using one of the options above?
Last edited:
Thomas Smet
Veteran
Yes, no special tool needed. Honestly just use 4k material in a 1080p timeline and you will gain most of the benefits on the fly. Assuming your machine can handle that load that is. Or better yet just edit as 4k and scale when you are finished. Keying 4k 4:2:0 and down converting after the fact would also yield 4:4:4 looking HD. You have to really think of the 4k to HD thing as not really gaining chroma but throwing out luma data to match the limited chroma. Down converting after your edit can yield the same overall end results. Give or take a bit of precision.
Thomas Smet
Veteran
Significant is a bit of a stretch here. Even jpeg on still images throws out color data because we as humans don't see color very well. The larger your image becomes to less likely it is to see chroma sub sampling. We are talking a very subtle difference here and no where near a value of significance. This whole 4:2:0 thing has been largely over exaggerated and beat into our heads from the 4:1:1 DV days and the fact that the first 4:2:0 cameras used interlaced 4:2:0 sampling patterns. Even 4:2:2 die hards when faced with proper 4:2:0 HD material realized that it wasn't as bad as they were left to believe.
SD or DV to be more precise was horrible because 4:1:1 is impossible to get back any decent level of prediction. Combine that with the fact that you had such a tiny image to begin with and there just wasn't a lkot of color data there. HD was a huge game changer because now those color dots were 1/4 the size and smaller. It was now significantly harder to see chroma sub sampling in an image. That combined with decent progressive 4:2:0 made for a very clean image where chroma could be reconstructed very well for human display.
4k takes this to a whole new level where those pixels are microscopic to the human eye. SD if it had perfect 4:2:0 would have had 360x240 chroma samples. 4k now has 1920x1080 chroma samples. That is 24x the amount of color data in an image compared to when many of us attempted to pull keys on SD material.
It is all about ratios. SD it was pretty normal to say you really had to have 4:2:2 because well we really did. With HD those color dots became so much smaller that even 4:2:0 still put SD to shame. You are still talking 6x the amount of color data just by moving to 4:2:0 HD. 4k at 4:2:0 has color pixels as small as if you had HD at 4:4:4. So if a production felt that HD at 4:2:2 was amazing then 4k can be much worse and still have more color detail.
SD 4:4:4 < HD 4:2:2 < 4k 4:2:0
Even if you look at the sampling this way each generation is still significantly better then the previous.
Barry_Green
Moderator
There may be two discussions going on here. The argument that harddrive is making, if I'm not mistaken, is that 4:2:0 encodes all the available color RESOLUTION from a Bayer sensor. Discussions about color space conversions etc. may be relevant in discussing overall image fidelity, but in terms of actual resolved detail, 4:2:0 encoding is sufficient to handle anything a Bayer sensor can output (assuming, obviously, that the frame size is the same as the native frame size of the Bayer sensor.)
Well, yes, that could be argued, but -- you're discussing interpolation methods at that point. There is no factual increase in actual resolved detail. And the question is whether the in-camera conversion is done better than it could be handled in post, etc.
While technically true on the face of it, this is profoundly misleading. I mean, first of all, all RGB files are inherently stored as 4:4:4; it cannot be any other way. That's really the inherent meaning of RGB, is that there is pixel data for red, green, and blue stored for every pixel. But where this becomes extremely misleading is in the idea that just because you can pour a gallon of gas into a five-gallon container, that doesn't mean you have five gallons of gas. And scaling up Bayer's blue and red channels through interpolation and up-rezzing may result in creating a color value for each pixel, but it most definitely does not represent what true 4:4:4 sampling would have delivered.
Again, outputting 4:4:4 does not mean the same thing as sampling 4:4:4. I can import VHS tapes onto my computer and export them as 1080p 4:4:4, but I think you and everyone else would agree with me that while technically the frame I exported would be 1920x1080, and while technically there would be red, green, and blue color info in each pixel position, there's no way anyone would confuse that VHS-sourced material for true 1080p 4:4:4. It is, again, pouring a gallon of gas into a 5-gallon can. You can CALL it a 5-gallon gas can, but the fact of the matter is that there's only a gallon of gas in it.
And the same goes with any Bayer-sourced footage; the sensor is incapable of sampling more data than can be represented in a 4:2:0 encode. A Bayer sensor may be equivalent to a four-gallon gas tank, and there's only one gallon of red in it. Interpolation and up-rezzing are not resolution and they're not sampling, they're synthesizing artificial data using educated guesswork.
Not to say it can't work surprisingly well; obviously it can. But it isn't resolved detail, it's synthesized guesses at what the color values might have been had there actually been some sampling take place at that particular pixel. It is a necessary evil, and it is a fundamental requirement when using a Bayer sensor. The only alternatives, to get actual higher resolved color information, would be to use three chips and a prism (as broadcast cameras typically do), or three chips sandwiched together (Foveon) or massively oversample and downconvert (as the F35 and Genesis did with their RGB Stripe sensor system, where it took a 12-megapixel sensor to output a 2-megapixel 4:4:4 image).
Exactly - though to be a pedant "virtually all" may be more precise. And it's also with one highly important caveat. That we're talking about a 4:2:0 recording system where "Y" is the same as the Bayer sensor dimensions.
So, for a quad-HD Bayer sensor, with 3840x2160 photosites, we're talking about a recording system where the Y signal is recorded to 3840x2160 and hence the chroma signals are 1920x1080 each.
I have no doubt Barry is fully aware of that - but it's important to stress, as quotes can get taken out of context. (And that doesn't just apply to digital video.....
)And that's the case with the Arri quote. The Alexa oversamples to get 1080 - by which it's meant that the basic sensor dimensions are greater than 1920x1080, and the derivation of the 1080 signal involves not just debayering, but deBayering and downconversion. It's typical in such case to oversample by at least sq rt 2 in each dimension - which means a rough doubling of the total photosite count. Do that and the output luminance resolution should then closely match the recording 1920x1080 for Y - but the chrominance resolution will be about 70-75% of that. So not quite worth 4:4:4 - but much better than 4:2:0 is capable of.
Once you start oversampling and downconverting, then yes, better than 4:2:0 chroma subsampling becomes increasingly worthwhile.
And again, exactly so, and well put. The size of the container doesn't mean it's filled. Obvious with everyday objects, maybe less so with resolution and recording formats and subsampling - but the analogy is spot on.
Barry has really answered it all, but I'll just pick up on one of Joshua's points, "Once subsampling occurs (because it can only be done in YCbCr) the original color fidelity is compromised." Which is absolutely true. But the subsampling in what we're talking about has been done by the sensor long before we think about chroma subsampling in the 4:x:x sense. Using the 3840x2160 as a base reference, the Bayer sensor has "subsampled" the red and blue information down by a factor of two by the fact that each of those colours only has 1920x1080 photosites. In Barrys analogy, the sensor gives you a gallon of red gas, deBayering will put it in a four gallon container, then chroma subsampling to 4:2:0 will pour it back into a one gallon for transport. (Where the display device should interpolate it back again.) Leaving it in the four gallon can for transport won't mean you'll bring any more gas back - just give you a more awkward thing to carry!
If anybody is still unhappy with this, just two points. For such a 3840x1920 sensor, red and blue photosites only number 1920x1080. It's therefore totally impossible to get red or blue resolution any higher than 1920x1080. That's the theory in a nutshell. All you can do is interpolate - pour a given quantity into a larger container. It won't give any more petrol.
Second point is to just see it in practice, and I can't do better than again refer to the link Joshua posted in post 48 with the coloured resolution charts - http://www.dpreview.com/reviews/sigmadp1/page20.asp . If a Bayer sensor did give "4:4:4" performance - taken to mean equal R,G,B (and hence luminance) resolution figures - the colours of the lines on the chart wouldn't make any difference to the results, would they? (As indeed, they don't for the Foveon chip.) But they obviously do. For any of the Bayer chips, the resolution results for coloured lines are far worse than for the equivalent black/white chart for that camera, and worst of all is when there is only red and blue. I take that as perfect experimental proof of the theory that Barry and I have been saying.
Last edited:
Yes, it's everything to do with ratios, so increase the basic resolution and you increase the chroma resolution in proportion. With 4k, and with 4:2:0, the chroma resolution by itself becomes the same value as luminance for HD!
I've brought up the question before of why - for 4:2:2 and 4:1:1 - is the subsampling always horizontal, never vertical? The answer is because all the TV standards at the time were interlace. It becomes relatively easy to interpolate chroma samples horizontally along the line - far more difficult vertically because of the interlace. Which is why all the early digital standards sampled every line. If you're only sampling one chroma for every four of luminance, it seems more logical to go for halving on each axis - so how did 4:1:1 come about? Answer above - interlace. In the early days, it was common for digital recorders to input/output analog composite signals, and whilst 4:2:0 is fine at a single generation, it can give via interlace vertical colour shifts on multiple analog-digital-analog transcodes. Hence 4:1:1.
But I suspect you are in a 60Hz territory? Because in PAL land we hardly saw 4:1:1, it was only used for DVCPro. Both DV and DVCam were 4:2:0 for PAL. And by this time, the need for frequent A-D-A converting was far less, it was more normal to stay in the digital domain.
Since all main viewing formats are 4:2:0 (DVDs, digital broadcast), the problem in NTSC land wasn't so much 4:2:0 being bad in itself - but recording 4:1:1 and outputting 4:2:0 meant a double loss, only getting one in eight of the original chroma samples! Recording 4:2:2 did at least mean that after the 4:2:0 transcode for the DVD, you did at least get true 4:2:0.
And a large reason only 4:2:2 HD systems are considered true broadcast spec is again down to interlace - 1080i/25 or 1080i/30 are the most common systems for broadcast. Once you get to 4k and progressive, there's much less necessity for 4:2:2.