GH3 m43rumours Panasonic GH4: can it do 4:4:4 at 1080p?

This possibility of getting better color space and bit-depth from downscaling a 4K image to a 2K image is very interesting to me. I can't wait to see it demonstrated once the GH4K has been released.

I found a website that performed a simple demonstration of this process by creating an 8-bit greyscale diagonal graduation, downscaling it, and outputting at 16-bits. I actually found it hard to discern the difference between the two images from the website, so I downloaded them, opened them in separate viewers, and placed them directly next to each other. Then I could see a slight difference. So I'm convinced that this process does lead to higher quality images, but I'm not convinced that it'll make a discernible difference to the untrained eye. Here's the website.

http://www.slashcam.de/artikel/Grundlagen/4K---Von-RAW-bis-4-2-0.html

It's in German. I used Google Translate to read it, but you don't need to in order to understand the photos and their meaning.
 
Jim Simon said:
But how do you know that? How do you know that the software isn't throwing out that extra info and also downconverting it to the same 4:2:0?
Click to expand...

Because that isn't how image processing works. Typically when you down convert something it doesn't just throw out every other pixel. It does a mathematical formula on every pixel to calculate the new value. If a chroma block is 2x2 pixels in size it makes sense that if you reduce the resolution by 1/4 that 2x2 pixel block will turn into a 1x1 pixel block. The only way you would end up with 4:2:0 again is if you rendered the down conversion to a 4:2:0 format. Even if it did just throw out every other pixel you are throwing out both the luma and chroma samples every other pixel so they end up having an equal number.

Don't believe me do the test yourself with 1080p 4:2:0 video. Scale it down to 960x540 and render it to a 4:4:4 format like ProRes4444 or the animation Quicktime codec. You will see the advantage right away.
 
From EOSHD.com...
"The way this works is quite technical but it essentially amounts to using the extra pixels present in the 4K files to rebuild the lost colour information. Neighbouring pixels are summed to create a super pixel with a greater bit depth and better sampling. It requires a lot of processing power and a finishing codec good enough to store the extra colour data (10bit 4:4:4, with CineForm and ProRes being suitable choices) which is why the camera doesn’t do it internally. It does require a transcoding step in post but the increase in quality over the 8bit internal 1080p codec will be marked. The message is clear. Even if you’re working in 1080p, shoot 4K on the GH4 if you want the best quality."


That's pure fakery. What matters to me is the ORIGINAL RAW Clut. It's not 10Bit 4:4:4. So they are trying to convince you that by widening the color space via dithering ( AKA Interpolation) that 'magically' a wider gamut has been recovered. That's incorrect because the the original space off the sensor was 8Bit 4:2:0
The 4K>2K hat trick is just that. A TRICK A trick because pixel density has NOTHING at all to do with capture color gamut. Sorry to be so harsh. You will discover this when someone runs Empirical Chroma Key tests.
Panasonic
Please stop this nonsense!
 
Zephyrnoid said:
That's incorrect because the the original space off the sensor was 8Bit 4:2:0
Click to expand...

I thought it sounded weird. The image was originally encoded at 8 bit. I don't know of any way to make that genuinely 10 bit except to reencode the original image off the sensor. Anything else is just padding.
 
I always understood it to be 4:4:4 on down conversion. Just like when you take 1080p Full HD down to SD. It becomes the best looking Standard Definition one could get over any native shooting SD camera. Bottom line is you will see the difference. People who are interested in finishing in 2K or Full HD will certainly benefit from this. All that other technical mumbo jumbo discussions don't matter to me. Call it 4:4:4...call it fakery...U WILL C A DIFF. :)
 
Zephyrnoid said:
That's pure fakery. What matters to me is the ORIGINAL RAW Clut. It's not 10Bit 4:4:4.
Click to expand...

I don't have a dog in this race, but I think EOSHD's assumption is correct. It's math. If you downsample 4 pixels that are 8-bit, you get one 10-bit pixel (256*4=1024). Same with turning 4:2:0 into 4:4:4. The trick, as I understand it, is having to do an intermediate conversion to 16-bit RGB in something like After Effects in order to have the headroom to do the downsample without losing information. Sounds plausible to me.

Oversampling.
 
Zephyrnoid said:
From EOSHD.com...
"The way this works is quite technical but it essentially amounts to using the extra pixels present in the 4K files to rebuild the lost colour information. Neighbouring pixels are summed to create a super pixel with a greater bit depth and better sampling. It requires a lot of processing power and a finishing codec good enough to store the extra colour data (10bit 4:4:4, with CineForm and ProRes being suitable choices) which is why the camera doesn’t do it internally. It does require a transcoding step in post but the increase in quality over the 8bit internal 1080p codec will be marked. The message is clear. Even if you’re working in 1080p, shoot 4K on the GH4 if you want the best quality."


That's pure fakery. What matters to me is the ORIGINAL RAW Clut. It's not 10Bit 4:4:4. So they are trying to convince you that by widening the color space via dithering ( AKA Interpolation) that 'magically' a wider gamut has been recovered. That's incorrect because the the original space off the sensor was 8Bit 4:2:0
The 4K>2K hat trick is just that. A TRICK A trick because pixel density has NOTHING at all to do with capture color gamut. Sorry to be so harsh. You will discover this when someone runs Empirical Chroma Key tests.
Panasonic
Please stop this nonsense!
Click to expand...

Where does the quote you posted say anything about gamut? Seems like you are confusing gamut and bit depth or I am misunderstanding what you are saying.

One thing to consider is that you probably can't get both better bit depth AND better sampling from downsampling 4K 8bit 4:2:0 to 2K 10bit 4:4:4. For example, the red and blue channels are both 2048x1080 so it is not getting supersampled and may not get scaled at all. On the other hand, the luma channel is downsampling so conceivably the new pixels could have a new bit depth given the math of adding together 4 8bit pixels to get one new one. So to some degree there is both better bit depth and better color sampling, but it may not be quite the same as a 10bit 444 original since the R/B channels may just be 8bit data moved into a 10bit container. On the other hand, it is possible that if the scaling algorithm does use the luma channel to interpolate values for R and B, it may be possible to have a bit depth increase in the color channels. For example, the YUV 420 may in processing be first converted to 4K RGB before downscaling, in which case the actual higher luma resolution may have an effect on the interpolated R and B channels which may increase the precision of the color channels when scaled down. Seems like something that may need testing. In either case, though, it does seem concievable that what once was 4K 24bit (8bpc) 420 would effectively become at least 2K 26bit 444 if not 2K 30bit 444
 
If you consider andrew reid as an expert image sensor engineer and not a sensationalist BLOGGER then go ahead get excited :)
 
To be fair, Andrew didn't just make this up. He has asked quite a few camera specialists and they have confirmed this down-scaling to 4:4:4 to be true.

Me, I couldnt care less. The GH4 is going to be a great camera either way. Has anyone confirmed yet that the on screen info doesn't disappear after 10 seconds........
 
vtwiz said:
Me, I couldnt care less. The GH4 is going to be a great camera either way. Has anyone confirmed yet that the on screen info doesn't disappear after 10 seconds........
Click to expand...

Yes, been waiting for this info - Driftwood was supposed to get some hands-on time and ask the Panasonic guys about this...
 
dcloud said:
If you consider andrew reid as an expert image sensor engineer and not a sensationalist BLOGGER then go ahead get excited :)
Click to expand...

This has nothing to do with image sensor technology and is all about image processing 101. Any time you take a large image and scale it by 1/4 resolution you are creating new interpolated pixels. This process reduces the ratio of chroma pixels to 4:4:4. This is a very basic concept and one you can test now with 1080p material by down converting to 960x540. Don't believe people then try it yourself.
 
photo.jpg

Since some people are having trouble believing the downversion upsampling I did a small test. Posting this as a png image to keep it 4:4:4.

1. Rendered out some 3D hair on a sphere against solid green as a 1920x1080 4:4:4 still image in png format. Best way to start with a 4:4:4 sample.
2. rendered a copy of the image as a 4:2:0 video.
3. Imported both png version and 4:2:0 video version into After Effects. Put them side by side in a 1920x1080 comp.
4. Down converted to 960x540. Same ratio as down converting 4k to HD.
5. Rendered out a new 4:4:4 png image which I now posted.

Try to tell the difference between the two. Remember one started out as 1920x1080 4:4:4 and the other started out as 1920x1080 4:2:0. After Effects doesn't even have the worlds best down conversion method and the results are still very impressive. A more sophisticated compositing application will give even better results. The way Shake used to be able to do.

note: This does not test 10bit since posting a 10bit image isn't really an option. For that you will have to do the test yourself.
 
Last edited:
4:4:4 , yes mathematically it's a fact - because Y'CbCr 4:2:0 3840x2160 means you're starting with 1920x1080 CbCr channels to begin with. You just resize the Y' plane to 1920x1080 and you have full color (each pixel in Y' has a corresponding value in Cb and Cr, ie. there is no chroma subsamping)

But ONLY the Y' signal will be potentially 10bit when downsampling, and it can be expressed as 0-1023 . The CbCr channels are only 1920x1080 in 8bit 3840x2160 4:2:0 footage, values expressed from 0-255. This means it's impossible to get true 10 bits from the color difference channels because they weren't 3840x2160 to begin with

After Effects works in RGB. Because the 8bit CbCr channels are expressed from 0-255, and the equations for converting between YCbCr and RGB are multiplicative, the resultant 1080p in RGB will be quantized to 8bit.
 
PDR nailed it.

When you have a source that has a two pixels of green, one pixel of red, and one pixel of blue, it should be fairly simple to see how those could all be combined into one pixel of R/G/B. Therefore the simple concept of 3840x2160 4:2:0 being downconvertable into 1920x1080 4:4:4 should be pretty obvious. As I said before, we don't yet know how much detail is contained in that 3840x2160, but however much is there, the ratio of color to luminance will increase in the downconversion.

As for 10-bit, again, remember that in terms of brightness we have four individual pixels that are being converted down to one. If you just do a simple summation, you can store all the brightness values of the four source 8-bit pixels into the 10-bit destination pixel. In 8 bits you can store a range of 0-255, in 10 bits you can store 0-1023. So even if you had four source pixels at absolute maximum brightness (255), you can store all that in one single 10-bit pixel (4 x 255 = 1020). Does that actually give you any "more" info? Yes, but perhaps not how you might think. It's possible people might be thinking that having more bits will store more dynamic range (i.e., a wider range from black to white). And that's not what will happen. Pure white (at 255, in the 8-bit system) is still going to be pure white once it's downconverted. What the 10-bit system will do is give you finer steps between the minimum and the maximum. So there will be no difference between absolute pure white in an 8-bit system (255), and pure white in a 10-bit system (1023), as far as the overall level of brightness. But in an 8-bit system, the tiniest bit darker would result in a reading of 254, which would equate to 1019 in the 10-bit system. Yet the 10-bit system is capable of storing 1022, 1021, and 1020 as well.

With color it's more complex. You're sampling one pixel out of four for color information. It should generally do well, but what if you had a weirdo case of in each group of 4 pixels, your picture happened to have image data that equated to three red and one blue. What would your color sampling turn out to be? It depends on where those pixels line up on the sensor; if your blue pixel sensor happens to line up over the blue pixel, then yeah, you'll get blue, but if that blue pixel sensor happens to line up over one of the reds, you won't get any blue info at all for that 2x2 pixel block. This is one of those "what if" cases that won't ever amount to a hill of beans, but if we were making a case for absolute color fidelity, you'd have higher color fidelity from a three-chip 1080p sensor block than you would get from a 4K 4:2:0 chip that downconverts to 1080p.

So yes, you should get 10-bit shading in the downconversion, but -- again, as PDR said, only in brightness. You won't get 10-bit color, at all. In 4K 4:2:0 vs. 1080p 4:4:4, there's 4x as much brightness information in the 4K, but identically the same color information. So the brightness information can be averaged together and represented in a 10-bit word, but the color information won't change at all, it'll still be 8-bit.

So, effectively, with ideal downconversion, you should be looking at 1080p with 4:4:4 8-bit color sampling, and 10-bit brightness.
 
Barry_Green said:
PDR nailed it.

When you have a source that has a two pixels of green, one pixel of red, and one pixel of blue, it should be fairly simple to see how those could all be combined into one pixel of R/G/B. Therefore the simple concept of 3840x2160 4:2:0 being downconvertable into 1920x1080 4:4:4 should be pretty obvious. As I said before, we don't yet know how much detail is contained in that 3840x2160, but however much is there, the ratio of color to luminance will increase in the downconversion.

As for 10-bit, again, remember that in terms of brightness we have four individual pixels that are being converted down to one. If you just do a simple summation, you can store all the brightness values of the four source 8-bit pixels into the 10-bit destination pixel. In 8 bits you can store a range of 0-255, in 10 bits you can store 0-1023. So even if you had four source pixels at absolute maximum brightness (255), you can store all that in one single 10-bit pixel (4 x 255 = 1020). Does that actually give you any "more" info? Yes, but perhaps not how you might think. It's possible people might be thinking that having more bits will store more dynamic range (i.e., a wider range from black to white). And that's not what will happen. Pure white (at 255, in the 8-bit system) is still going to be pure white once it's downconverted. What the 10-bit system will do is give you finer steps between the minimum and the maximum. So there will be no difference between absolute pure white in an 8-bit system (255), and pure white in a 10-bit system (1023), as far as the overall level of brightness. But in an 8-bit system, the tiniest bit darker would result in a reading of 254, which would equate to 1019 in the 10-bit system. Yet the 10-bit system is capable of storing 1022, 1021, and 1020 as well.

With color it's more complex. You're sampling one pixel out of four for color information. It should generally do well, but what if you had a weirdo case of in each group of 4 pixels, your picture happened to have image data that equated to three red and one blue. What would your color sampling turn out to be? It depends on where those pixels line up on the sensor; if your blue pixel sensor happens to line up over the blue pixel, then yeah, you'll get blue, but if that blue pixel sensor happens to line up over one of the reds, you won't get any blue info at all for that 2x2 pixel block. This is one of those "what if" cases that won't ever amount to a hill of beans, but if we were making a case for absolute color fidelity, you'd have higher color fidelity from a three-chip 1080p sensor block than you would get from a 4K 4:2:0 chip that downconverts to 1080p.

So yes, you should get 10-bit shading in the downconversion, but -- again, as PDR said, only in brightness. You won't get 10-bit color, at all. In 4K 4:2:0 vs. 1080p 4:4:4, there's 4x as much brightness information in the 4K, but identically the same color information. So the brightness information can be averaged together and represented in a 10-bit word, but the color information won't change at all, it'll still be 8-bit.

So, effectively, with ideal downconversion, you should be looking at 1080p with 4:4:4 8-bit color sampling, and 10-bit brightness.
Click to expand...

Unless you factor in chroma interpolation. We already know that a good chroma upsampler can turn a 4:2:0 video into a much closer to 4:4:4 video through interpolation. There is lots of software out there that does this very well. Not a perfect solution however but does help a lot. In this case you could get a form of interpolated 10 bit of color because you now have 2x2 pseudo samples of color. Of course they are not 100% accurate but there very well could be a different value for all four pixels. Since the GH4 is a bayer sensor with heavy color interpolation during the debayering process even the SDI 10bit HD output could be considered not as good as the 10bit from a 3 chip video camera. Good chroma interpolation could achieve a similar effect of creating new chroma pixels as the debayering process. So comapred to a 3 chip camera yes not as accurate but compared to what comes off a bayer sensor anyway the interpolated color may not be all that much worse.
 
vtwiz said:
To be fair, Andrew didn't just make this up. He has asked quite a few camera specialists and they have confirmed this down-scaling to 4:4:4 to be true.
Click to expand...
Well based on his articles he doesnt seem to be giving credits to those people he asked. Worded like they're of his own
 
dcloud said:
Well based on his articles he doesnt seem to be giving credits to those people he asked. Worded like they're of his own
Click to expand...

He mentioned right at the start of the article that he asked David Newman of Go Pro about the theory. He even posted a screen shot of the actual discussion with David. That is a lot more than a lot of other articles do to give credit to their source.
 
You must log in or register to reply here.
Back
Top