Increasing color depth and sampling for color correction

[poisondeathray]

I've read comments that you should transcode 8-bit 4:2:0 camcorder footage (e.g. HDV, AVCHD) to prores (mac) or cineform which is 10-bit 4:2:2 for keying and color correction because it "stands up better" and "can be pushed farther" etc...

My question is how is this possible?

Is it "magically" increasing the color information in terms of color depth (8-bit vs. 10-bit) and chrominance subsampling (4:2:2) from an 8 bit per channel, 4:2:0 source ? Are they just interpolating values or padding values ? What about errors in "guessing"?

When you import your native 8-bit 4:2:0 footage into your NLE or software, etc.., isn't it usually decompressed to RGB 4:4:4 internally anyways?

Even if you are working in 10-bit color depth project, wouldn't you need a 10bpc capable monitor like HP Dreamcolor to "reap" the benefits? or do the benefits hold, but you just can't "see" them?

It's true, when I am working with a 16bpc or 32bpc project (instead of 8bpc) in after effects for example, the color gradients become smoother and there is less banding - this is noticable, but what happens when you encode to the final 4:2:0 final format e.g. blu-ray or dvd etc...isn't that extra color infomation downsampled again ? Banding reappears. So what's the advantage of the digital intermediate then?

I don't see the benefit of that workflow? or am I missing something?

Thanks for any insight

[jagabo]

I would generally agree that using more spacial resolution (4:2:2 vs 4:2:0) and greater intensity resolution (10-bit vs 8-bit) can give you better results. Even starting with 8-bit 4:2:0 and upsampling to 10 bit 4:2:2 can improve results if you're filtering. And even if your final result is going to be 8 bit 4:2:0 again. The more bits you have while working the more accurate the final result will be. It also gives you the ability to randomize the last bit or two to reduce posterization problems.

Even on an 8 bit display you can see a minor difference in smoothness (less banding) when working with 10 bit data. I'll see if I can come up with some examples filtering 6 bit data natively, and 6 bit data upsampled to 8 bits, filtered, then downsampled back to 6 bits. I'm pretty sure the differences will be visible.

[vhelp]

In my experience, the extra bit-depth is beneficial in terms of color detail..more accurate colors bring out the image. I'm speaking from my own expience with working with laserdisc captures. I have many capture cards and ran them through their paces over the years with my laserdisc player, and I can tell you there is nothing better than a 10bit (internal processing) capture card then most 8bit ones. My previous 8bit laserdisc captures can't compare to the 10bit ones.

here are some examples of 10bit captures of laserdisc to help express, but I don't have the bandwidth to post new (8bit) captures, sorry.

--> https://forum.videohelp.com/topic362850.html#1929699

If you deal with a source in 10bit, assuming it was "captured" the color depth or detail will be greater, (better) but that will largely depend on the output source being captured, ie laserdisc. The capture card had nothing but plusses, a greate (composite) 2d-comb filter, 10bit, and the laserdisc output source originated from composite encoded data, everything was basically transparent in this case. But I suppose it would be similar to a degree in 8bit->10bit image conversions, algorithm-icly speaking of course, and the better the "decoder" the less problems in the gradient error you mentioned, ie start trek voyagers opening theme is a perfect example representing the error you mentioned. In fact, its still prevailent in the dvd series, last episode I just saw this error on was in s5e6, "Timeless" where the voyager ship passes through the suns flare. hmm. that would be a great challenge to fix, graphically, in adobe photoshop, as an excercise I think I'll etch that into my some day todo list

Still, I think the level of this error could be reduce, but the idea just escaped me at the moment..oh well..it was just in my head, the part of the cause and fix. Most of this error is user-error imho. That voyager example could have been fixed no problem. I've seen many "gradient" examples (mpeg-2 sources) where transients were smooth.

-vhelp 5229

[vhelp]

ooops, wrong thread..

-vhelp 5230

[poisondeathray]

Thanks for the replies guys,

I can definitely see the difference in project as it's being edited, but once converted to YV12 4:2:0 8-bit again, I don't see how the difference translates into final edit (i.e you see banding again)

I'm having difficulty with the concept of having something better than what you started with...This concept is so entrenched with video, and we preach that here everyday with re-encoding lossy formats... at best you should have the same quality, not better. How is this different? Not to mention cineform and prores are in fact lossy formats...

vhelp - I don't know anything about laserdisc, but capturing at a higher bit depth (and subsequently bit rate) makes that less lossy right? - that much I can understand. e.g. if I captured a video and used RGB 4:4:4 there is no color sub sampling vs. a YV12 4:2:0 capture format. So this would better represent the original capture information better, but how can it be "better" than the original?

My understanding is that RGB24 is 8bits per channel right? 8red , 8green, 8 blue ? (and the RGB32 saves the other 8 for alpha).

If i converted my 8-bit 4:2:0 format to 8-bit 4:4:4 uncompressed RGB24, how does that stack up to a 10-bit 4:2:2 format like cineform? Would that even make a difference, since internally, most NLE's convert to RGB anyways (I think?)...

Can you have a 10-bit uncompressed format? 10red, 10green, 10blue ?

Still a bit confused...

[edDV]

My question is how is this possible?

The main reason for upscaling 8 bit 4:2:0 to 10 bit 4:2:2 with ProRes or Cineform is to mix 8 bit material into a 10 bit project. Apple also has the Apple Intermendiate Codec (AIC) which is intended for 8 bit 4:2:0 in and out.

Digital intermediate codecs in general recompress GOP based formats into individual frames for easier editing and less generation loss (after conversion).

Is it "magically" increasing the color information in terms of color depth (8-bit vs. 10-bit) and chrominance subsampling (4:2:2) from an 8 bit per channel, 4:2:0 source ? Are they just interpolating values or padding values ? What about errors in "guessing"?

They interpolate new values and/or add noise to randomize the added bit depth. The first conversion is lossy but after conversion, subsequent processes have less generation loss. Cineform uses wavelet compression which is better and more computationally efficient for repetated layering or resizing frames.
http://techblog.cineform.com/?p=1280

[poisondeathray]

Thanks for reply edDV,

I understand why it's used as a digital intermediate for quicker editing and it's robustness over generations, but I'm more interested in the mechanics or theory or purported benefit for keying and color correction. I'm not convinced it's beneficial for those reasons.

And once your 10-bit, 16-bit or 32-bit project is converted to the final 8-bit 4:2:0 distribution format wouldn't those purported benefits be lost anyways?

They interpolate new values and/or add noise to randomize the added bit depth.

So how would this be beneficial in color correction? Wouldn't interpolating or adding noise would add data that wasn't in the original? Wouldn't there be errors in interpolation or added noise compared to the original?

If you were to compare the 8-bit 4:2:0 original to a cineform or prores intermediate, the latter 2 are lossy. Sure they have higher 4:2:2 chrominance subsampling, but it's based on the 4:2:0 original. The 10-bit is based on the 8-bit original. I can't see how this makes it "better" than the original?

I would generally agree that using more spacial resolution (4:2:2 vs 4:2:0) and greater intensity resolution (10-bit vs 8-bit) can give you better results. Even starting with 8-bit 4:2:0 and upsampling to 10 bit 4:2:2 can improve results if you're filtering. And even if your final result is going to be 8 bit 4:2:0 again. The more bits you have while working the more accurate the final result will be. It also gives you the ability to randomize the last bit or two to reduce posterization problems.

Sorry, I'm still having difficulty with this concept. The light just isn't turning on for me. I know if the original was 10bit 4:2:2, then everything is honky dory - that makes sense more color information. But starting from 8-bit 4:2:0 source?

The more bits you have while working the more accurate the final result will be
even when interpolated? isn't there more room for errors and yet again when you downsample? Isn't fewer lossy conversions better? How can you have more (accurate) information than what you started with? (I don't mean noise or padding) ?

[edDV]

My understanding is that RGB24 is 8bits per channel right? 8red , 8green, 8 blue ? (and the RGB32 saves the other 8 for alpha).

If i converted my 8-bit 4:2:0 format to 8-bit 4:4:4 uncompressed RGB24, how does that stack up to a 10-bit 4:2:2 format like cineform? Would that even make a difference, since internally, most NLE's convert to RGB anyways (I think?)...

Can you have a 10-bit uncompressed format? 10red, 10green, 10blue ?

Still a bit confused...

Full decompression would be less lossy than using a digital intermediate but that would require a multi-HDD RAID for playback or preview. 10 bit would require 20% more drives vs. 8bit. Simple conversion of 8 to 10 bits would not improve quality unless it was being filtered with other 10 bit original material. Broadcasters originate 10 bit to allow for serial processes with less rounding error per generation. 8 bit NLE's attempt all layering and filtering in one render pass. Internally 8bit RGB is expanded to many more bits through digital multiplication. For the final encode the expanded bits and 4:4:4 color sampling get reduced to the final format (e.g. 8bit 4:2:0 MPeg).

Digital intermediates use intraframe compression to keep bit rate within the range of a single HDD even when scanning or scrubbing.

[Cornucopia]

Yes, as edDV said, its real benefit is when/while doing Non-keyframe editing, COMPOSITING/MIXING, and Filtering & FX. Especially if you work with complex combinations of layers (with alpha masks) and color mixing calculations.

If you start with 8bit4:2:0 and do only keyframe editing onto a 8bit4:2:0 final master, you won't notice really ANY difference vs. bumping up.
BUT
If you do the hard stuff mentioned above WHILE staying in 8bit4:2:0, you're basically going down 2, 3, 4 or more generations, quality-wise. And compounding the already inherent banding.
So bumping up TEMPORARILY bypasses the EXPECTED loss which is the byproduct of that kind of manipulation. Works the same way for audio too.

Scott

[poisondeathray]

Simple conversion of 8 to 10 bits would not improve quality unless it was being filtered with other 10 bit original material.

This is what I'm getting at. The claims are that converting 8-bit material to 10-bit 4:2:2 would be "better for color correction". I can't conceptually see how this is beneficial (disregarding the ease of editing etc...). All the other stuff you've mentioned I understand.

[edDV]

They interpolate new values and/or add noise to randomize the added bit depth.

So how would this be beneficial in color correction? Wouldn't interpolating or adding noise would add data that wasn't in the original? Wouldn't there be errors in interpolation or added noise compared to the original?

If you were to compare the 8-bit 4:2:0 original to a cineform or prores intermediate, the latter 2 are lossy. Sure they have higher 4:2:2 chrominance subsampling, but it's based on the 4:2:0 original. The 10-bit is based on the 8-bit original. I can't see how this makes it "better" than the original?

I'm not arguing upconversion is beneficial for color correction. The randomization from wavelet encoding minimizes banding from rounding errors during filtering or resizing. You are trading some distortion (noise) for less artifacting. During spacial translations, 4:4:4 computation has the benefit of keeping all component pixels in the same X, y, z space thus avoiding color smear.

[poisondeathray]

Thanks for replying cornucopia,

If you do the hard stuff mentioned above WHILE staying in 8bit4:2:0, you're basically going down 2, 3, 4 or more generations, quality-wise. And compounding the already inherent banding.
So bumping up TEMPORARILY bypasses the EXPECTED loss which is the byproduct of that kind of manipulation. Works the same way for audio too.

My understanding is internally, the 8-bit 4:2:0 source is decompressed to 4:4:4 RGB anyways (e.g. in premiere). Is this wrong?

I'm not following you, if everything is the same original generation, how do you incur more losses than upon final export? I'm not talking about multiple generations here. Or am I misunderstanding something?

I understand how using cineform would be better for multiple generations vs. a 8-bit 4:2:0 codec. That isn't what we are discussing. If I were to use multiple generations, I usually use lossless formats.

Please, I only am concerned about the purported benefits for color correction/keying of using a digital intermediate vs. the original on the current generation, not about the other things like ease of editing, or higher bit depth and less color sampling is better on subsequent genrations, which is pretty clear...

[edDV]

The key concept for a NLE-processor is all computation is done in the final encode pass. The 8 bit YCbCr processed frames are converted once to 4:4:4 RGB and then back to the export format. While in RGB space, bit depths expand with each multiply but all gets rounded back to 8bit for export. Block based MPeg will result in more block edge oriented rounding errors vs the more randomized Cineform wavelet pixels.

8 bit NLE's don't perform as well for serial processes where intermediate effects are stored as 8 bit and then recycled as source for another level of processing. The 8bit rounding errors build up with each pass. Broadcast and film production flows often are sequential so 10, 12 or 14 bit source and process flows are used.

[edDV]

Please, I only am concerned about the purported benefits for color correction/keying of using a digital intermediate vs. the original on the current generation, not about the other things like ease of editing, or higher bit depth and less color sampling is better on subsequent genrations, which is pretty clear...

For color correction in FCP, Premiere, Vegas, etc. I see no bit stretch advantage for 8 bit source. There may be some proprietary tech in products like the DaVinci that I'd have to research.

As for keying, there may be edge blend advantages from 10 bit upconversion. Keep in mind that consumer digital luminance is encoded 16-235 which only has 219 levels. MPeg block errors within 219 levels can be seen.

10 bit upconversion would interpolate 64-960 or 896 levels of gray. This presumably would be done in combination with a deblocking filter.

[Cornucopia]

Chroma upsampling increases the # of pixels in the chroma world, but usually does so by interpolation, or BLENDING. While that's usually BAD in terms of overall resolution/quality (mainly in the luminance channel), for THIS PURPOSE it has the benefit of SMOOTHING the transitions and making the keyhole less blocky/jaggy.
You could have just done BLUR on the chroma channel without the uprezzing/upsampling, but then instead of the chroma looking SMOOTHED, it would just have looked BLURRED and of LOWER resolution.

Does that answer your specific question about keying?

Scott

[jagabo]

minimizes banding from rounding errors during filtering or resizing

This is the heart of the matter. Keep in mind that video calculations are usually done in an integer space. Having more granularity (10 bit's 0-1023 vs 8 bits's 0-255) gives more precision in the intermediate calculations.

Let me see if I can make this clear with a hypothetical example. This isn't necessarily a realistic situation but I need a situation that's simple yet has enough data to work with to show clear results. Let's say we have an 8 bit grayscale image and we want to perform an 8x enlargement. We're only going to look at three pixels in one dimension for simplicity. Here are the three pixels in our original image:

 5                       6                       7

Now we perform an 8x enlargement by simple integer interpolation:

 5  5  5  5  5  5  5  5  6  6  6  6  6  6  6  6  7

Since this is an integer space there are no values between 5 and 6 or between 6 and 7. The result is an image with obvious banding. A wide band of 5's, a wide band of 6's, etc.

Now let's convert our original image to a 10 bit colorspace by multiplying all pixels by 4:

20                      24                      28

Then we perform a 8x linear interpolation enlargement:

20 20 21 21 22 22 23 23 24 24 25 25 26 26 27 27 28

If we view that on a 10 bit display the results will be much smoother than the 8 bit enlargement. Of course, if we simply convert back to 8 bits by dividing all values by 4 we will get exactly the same results as with the 8 bit enlargement. But let's say we perform a simple dither of the 10 bit data. We will alternate between adding 1 and subtracting 1 at each pixel in the enlarged image:

21 19 22 20 23 21 24 22 25 23 26 24 27 25 28 26 29

Now when we convert this back to 8 bits by dividing by 4 we get:

 5  4  5  5  5  5  6  5  6  5  6  6  6  6  7  6  7

In the area between the 5 band and the 6 band there are pixels which are alternating between 5 and 6. And between the 6 band and the 7 band they are alternating between 6 and 7. When viewed on a 8 bit display, as long as you aren't too close, those areas where the pixels are alternating between 5 and 6 appear to have values between 5 and 6. They will look like they have a value of 5.5. Or 5.333 and 5.666, etc.

There is still some banding in the final image. This could have been reduced by adding and subtracting 2 or 3 instead of 1 at the 20 bit stage. And instead of simply alternating you could use a more sophisticated dithering pattern, or randomized numbers.

Now if you consider that calculations with real image data are 2 dimensional (width, height), or 3 dimensional (width, height, time), and that filtering may be more sophisticated than our simple linear interpolation, you can see that having more bits to work with can give you better results -- even if you convert back to 8 bits for final output.

[poisondeathray]

Chroma upsampling increases the # of pixels in the chroma world, but usually does so by interpolation, or BLENDING. While that's usually BAD in terms of overall resolution/quality (mainly in the luminance channel), for THIS PURPOSE it has the benefit of SMOOTHING the transitions and making the keyhole less blocky/jaggy.
You could have just done BLUR on the chroma channel without the uprezzing/upsampling, but then instead of the chroma looking SMOOTHED, it would just have looked BLURRED and of LOWER resolution.

Does that answer your specific question about keying?

Scott

Yes, that's the missing "mental link" I was looking for in terms of keying. Thanks Scott

But what about color correction and banding? Now in terms of banding, my understanding is this has almost entirely to do with bit depth, rather than the chrominance subsampling?

Higher bit depth definitely helps, at least in the editing stage. I see this all the time where you switch into a 16bpc or 32bpc in after effects, and PRESTO, all the banding and color transitions are smoothed over instantly. But this doesn't hold up very well when you export to 8-bit end format (i.e. banding reappears, unless you do some "tricks" like dithering and noise). Would a prores or cineform intermediate perform "better" in this case?

(Now I'm going to try to digest jagabo's message, might take me awhile )

Thanks alot for your explanations and help guys, much appreciated

[poisondeathray]

minimizes banding from rounding errors during filtering or resizing

This is the heart of the matter. Keep in mind that video calculations are usually done in an integer space. Having more granularity (10 bit's 0-1023 vs 8 bits's 0-255) gives more precision in the intermediate calculations.
....
Now if you consider that calculations with real image data are 2 dimensional (width, height), or 3 dimensional (width, height, time), and that filtering may be more sophisticated than our simple linear interpolation, you can see that having more bits to work with can give you better results -- even if you convert back to 8 bits for final output.

Thanks edDV and jagabo,

That example makes it perfectly clear to me now.

In after effects, you can use 16bit or 32bit per channel projects. As I mentioned earlier , banding is visiblity eliminated immediately when switching even on an 8-bit LCD display. (But of course the benefits are not translated to the 8-bit export entirely, but I can see from jagabo's #'s how it still would provide better results). So what benefit would a cineform or prores intermediate have over the native footage in this case? (lets disregard the other advantages like ease of editing etc....)