4k vs more DR/color resolution w/ 1080p....moot point?

[Razz16mm]

From what I seen at CES all the the TV companies are putting more DR and a wider color gamut in UHD TVs than HDTVs. Just because some people can't see the difference in a 2mp image from an 8mp image does not mean technology will come to a stop.

REC2020 compliant UHD does have a much wider DR and color gamut. But if you downscale and encode to standard HD, your are in the smaller REC709 color space, not UHD color space.
8-bit can't resolve the full gamut of UHD even if you are shooting 4k. You have to be in 10 bit or 12 bit deep color mode to make use of it.

 

[harddrive]

Let me go through the details because I think there's some confusion as to what I meant with the 10 bit thing...

In 4K 420 footage (let's make it UHD for simplicity), you have 3840x2160 8-bit luminance values, and 1920x1080 8-bit chroma value pairs. It's clear already that you're going to end up with 444 8-bit color because all the info is there to begin with.

Fully agreed - 4K 420 gives you 1920x1080 chroma samples - the same number as needed for 2K 444.

For luminance, you're mixing four 8-bit values to get the value of the new (bigger) pixel. When you mix four values ranging from 0 to 255, how many options do you have for the result? 1024 different options. That's one 10-bit luminance value, which you get by adding the four 8-bit values, instead of averaging them. You could also average them and stay in 8-bit world, but then you're throwing information away.

That is only true if the variations about the mean of the four initial values are random (such as for noise). In the scenario you describe, that won't be necessarily true due to rounding in forming the 8 bit values. All the values may get rounded the same way, and all have the same skew - hence my previous example:

Imagine as an example 4 pixels in a block, three of which have the 10 bit value 0000000011 (3) and one has 0000000100 (4). In an 8 bit system, rounding will give all four the same value - 00000001 (equivalent to 4in the 10 bit system).

Convert to 2K, and the averaged (10 bit) value in the first case will be 13/4=3.25 - which will round to 0000000011 (3).

Do the same averaging derivation for 10 bit for the original 8 bit case, and the result will remain 4 (0000000100)

That's an extreme example, obviously, but it's to try to show the principle that once you've lost integrity by rounding to 8 bit, you can't necessarily retrieve it by the averaging process.

You make a valid point here demostrating that the result can't be real 10 bit. But won't there be in most cases a result from averaging those four 8bit pixels that could be described with more precision by 10bit?

Answer to the second sentence is it depends on the camera noise level. If it's relatively noisy (compared to quantisation levels) the answer is yes - think of it as the case where variations about the mean are random. If the camera has a very low noise figure, and assuming a uniform scene, then all the values in each block are likely to take the same 8 bit value. Raise the level slightly and all the levels will move up by one bit at the same time. Average, and allow for 10 bit, and you don't really gain anything - you get the same increases, but in chunks of 100 in 10 bit value.

None of that is to say there isn't a benefit in moving 10 bit after acquisition - there is. But that's also true for when you're in 2K throughout, when there is no rescaling going on. (10 bit gives less repetitive rounding when the material is processed.) Again, a lot depends on noise levels. Acquisition at 10 bit only really has significance with very low noise cameras.

 

[Guest]

...Acquisition at 10 bit only really has significance with very low noise cameras.

What do you mean?

 

[harddrive]

8-bit can't resolve the full gamut of UHD even if you are shooting 4k. You have to be in 10 bit or 12 bit deep color mode to make use of it.

I disagree, that's not the meaning of gamut in this context.

Bit depth relates to the fineness of difference that can be represented between two nearly identical colours. Gamut relates to the extremes - the limit of what colours can be represented in a system.

Look at http://en.wikipedia.org/wiki/Gamut , and especially the diagram right of page.

True spectral colours are represented by the boundary of the curved space - what a system can define is contained within the triangle, and that's the gamut for the system. Take (say) a true spectral green of wavelength 500nm and NO system will be able to truly represent it - only approximate to it. (The nearest point on the triangle to that spot on the chart.) But in general, the wider the gamut, the closer the approximation can be.

 

[harddrive]

What do you mean?

Assuming a uniform scene, which we'll say can be represented by an 8 bit value of 00001000, then think what video noise actually means.

It may mean that for six consecutive values we get 00000110, 00000111, 00001100, 00000111, 00001001, 000001011 - instead of six values each of exactly 00001000 in an ideal world.

(In other words, we get 6, 7, 12, 7, 9, 11 instead of six values all of 8)

Now, let's say we can define values more precisely, (using decimals for ease) to say the nearest tenth. We may now find that in the absence of noise, the true value is not 8 but 8.1. But with the noise our samples may become 5.8, 7.1, 11.7, 7.3, 8.8, 11.2. But what's it gained? If we can't measure the value of 8.1 to nearer than about +or- 4, then what's the point of recording the measurements to the nearest .1?

So with a camera where the noise level is significantly greater than the level represented by a single bit in an 8 bit system, all that 10 bit recording will do is more accurately define the noise level. Worse, then all else equal , a 10 bit system will require 25% higher data rate than an equivalent 8 bit system. If bandwidth is limited, then the 10 bit system may require higher overall compression to reach the limit - and the negative effects of that may far outweigh any benefits that 10 bit may bring in other ways.

 

[Guest]

Assuming a uniform scene, which we'll say can be represented by an 8 bit value of 00001000, then think what video noise actually means.

It may mean that for six consecutive values we get 00000110, 00000111, 00001100, 00000111, 00001001, 000001011 - instead of six values each of exactly 00001000 in an ideal world.

(In other words, we get 6, 7, 12, 7, 9, 11 instead of six values all of 8)

Now, let's say we can define values more precisely, (using decimals for ease) to say the nearest tenth. We may now find that in the absence of noise, the true value is not 8 but 8.1. But with the noise our samples may become 5.8, 7.1, 11.7, 7.3, 8.8, 11.2. But what's it gained? If we can't measure the value of 8.1 to nearer than about +or- 4, then what's the point of recording the measurements to the nearest .1?

So with a camera where the noise level is significantly greater than the level represented by a single bit in an 8 bit system, all that 10 bit recording will do is more accurately define the noise level. Worse, then all else equal , a 10 bit system will require 25% higher data rate than an equivalent 8 bit system. If bandwidth is limited, then the 10 bit system may require higher overall compression to reach the limit - and the negative effects of that may far outweigh any benefits that 10 bit may bring in other ways.

Assuming a uniform scene,

(EDIT/updated comment: you did say this^. Which makes your point clear, but also seems too theoretical, as a proper scene would have adequately exposed areas, and most cameras I have access to are pretty clean. )


I see. Though, noise pollution is hardly uniform across a scene, unless doing a lens cap test. So, a blue sky behind a darker lit character would still benefit from 10bit, even if the shadowed character is a bit noisy.

Regardless, 8bit does not bother me in a working environment, for the most part, it is family videos and the occasional feature that I get into the high end quality.


What I find funny, is that the family and hobbyist are the majority of people that benefit from high end video, where as professionals rarely can get their work displayed in high fidelity. Yet, while it is little extra cost to add bit depth and resolution these days, it is still considered a "pro" feature. Few pros need it! It is largely hobbyist and a few family people wanting 4K 15stops DR 10/12bit. I am glad BMCC is out in the wild now. Though, their 4K is still out of reach of the average citizen. Soon, hopefully. $2500 with media seems to be the threshold.

 

[Samuel H]

ok, that's a fully uniform scene with noise. Now consider a linear gradient with 0 on the left and max on the right, without noise to start with.

It's a 3840-wide image to begin with, with one 8-bit luminance value per pixel. The first 33 pixels are zero, then the next 33 are one. On our final 2K image with 10-bit luminance values (each one is the sum of four original 8-bit values), we will have zeros for the first 16 pixels, then a 2, then 4 for the next 16 pixels. That "2" is a 10-bit value, something that you wouldn't have if you averaged instead of adding (i.e. if you stayed in 8-bit instead of going to 10-bit).

It is clear that with just one value in the newly created depth, banding is not going to look very different. But it's something. And it gets bigger the more you complicate the example.

Say the first row is the same as before, but the second one has twice the frequence: it goes from min to max, then to min again, in the same 3860 pixels.
First row: first 33 values are zero, nexe 33 are one.
Second row: first 15 values are zero, next 16 are one, next 15 are two, next 16 are three.

Adding up to get 10-bit 1920 image, we will have:
seven zeroes, one one, eitght twos, one four, seven sixes, nine eights
only those zeroes, fours and eights lie in the original 8-bit space, everything else is newly created channel depth
now half the values (16 out of 33) are in the newly created depth

Take a real image, and around 75% of the resulting values will end up being in that newly created depth. Unless the codec in the original image is screwing things up, which it probably will.



Now, about noise:
Of course a camera with less noise will be better placed to take advantage of this.
And noise makes all these calculations much more complex (but I don't think this will change the final conclusion).
In any case, consider the following: Stu Maschwith (my color guru) says he always starts his workflow with a mild denoise, then he adds grain at the end; he says that this is a way to get something similar to what you'd get if you shot 10-bit: it breaks the original banding and macroblocking and creates artificial intermediate values. This downsampling we're talking about now is a much cleaner version of that (because the original information is actually enough to get you 10-bit).


Edit:
if you now want to think about compression...
even with the simplest form of compression, that gradient will be first ran through several differences of differences filters
simplest form: substract left pixel, now the image becomes: 28 zeroes, 2 ones, 3 twos
if we discard the new channel depth and stay at 8-bit (average instead of adding values), we have to encode: 31 zeroes, 2 ones.
what I mean to say is: it's not clear that you need 25% more bandwidth to encode a 10-bit image, because you're very unlikely to find big values; your Huffman table will have much longer words, but you'll hardly ever use them, and how the distribution of smaller words changes is not clearcut at all; it can have a very minimal impact, or it may need twice the bandwidth

 

[Razz16mm]

I disagree, that's not the meaning of gamut in this context.

Bit depth relates to the fineness of difference that can be represented between two nearly identical colours. Gamut relates to the extremes - the limit of what colours can be represented in a system.

Look at http://en.wikipedia.org/wiki/Gamut , and especially the diagram right of page.

True spectral colours are represented by the boundary of the curved space - what a system can define is contained within the triangle, and that's the gamut for the system. Take (say) a true spectral green of wavelength 500nm and NO system will be able to truly represent it - only approximate to it. (The nearest point on the triangle to that spot on the chart.) But in general, the wider the gamut, the closer the approximation can be.

As with broad dynamic range though, 8-bit values won't produce very good color resolution in a wide RGB gamut. The two are closely related.

 

[True North]

I recently came across Shane Hurlbut's blog (HurlBlog). I know its quite well known so forgive me if I am stating the obvious.

He gives a fairly detailed account of how they chose the cameras for 'Need for Speed' and he goes in depth on how he tests for highlights and dynamic range etc....

Lots of video example comparisons between Alexa, Red, C500, BlackMagic etc... If you are interested in this topic it really is worth checking out his blog.

http://www.hurlbutvisuals.com/blog/topics/equipment/

 

[Guest]

I want one of those Vizio reference TV's.