David Rasberry
Well-known member
- Joined
- Aug 20, 2008
- Messages
- 4,266
- Reaction score
- 0
- Points
- 0
Whatever you choose to encode it to when you debayer the raw files.
Welcome to our community
Be a part of something great, join today!
confused.. Is RED one's color sampling 4:2:2?
- Thread starter Beom Koh
- Start date
Whatever you choose to encode it to when you debayer the raw files.
Charles Angus
Well-known member
- Joined
- Jun 13, 2007
- Messages
- 1,055
- Reaction score
- 0
- Points
- 0
- Location
- Toronto
- Website
- www.charlesangustaylor.com
Just since it seems relevant to this thread, I would point out what 444 and 422 mean, as I understand it.
422 refers to a luma/chroma, and 444 is generally used to refer to RGB images (I'm not aware of a 444 YUV system, but there could be one). 422 refers to the process of taking a full-resolution RGB image, splitting it into one luma and two chroma channels, and then scaling the chroma channels down to 50% in the horizontal direction to save space.
Hopefully that makes it a little clearer why the terminology doesn't really apply to a RAW camera.
Now, people might say that a raw file is like a 422 file, because of this. The luma channel is mostly created from the green channel, and the two chroma channels are largely constructed from the red and blue channels, respectively. A Bayer-pattern sensor has twice as many green photosites as either red or blue photosites. It then interpolates the values it didn't record. Some people look at the lower amount of red and blue photosites as being analogous to 422 subsampling, but it's really a whole different kettle of fish.
422 refers to a luma/chroma, and 444 is generally used to refer to RGB images (I'm not aware of a 444 YUV system, but there could be one). 422 refers to the process of taking a full-resolution RGB image, splitting it into one luma and two chroma channels, and then scaling the chroma channels down to 50% in the horizontal direction to save space.
Hopefully that makes it a little clearer why the terminology doesn't really apply to a RAW camera.
Now, people might say that a raw file is like a 422 file, because of this. The luma channel is mostly created from the green channel, and the two chroma channels are largely constructed from the red and blue channels, respectively. A Bayer-pattern sensor has twice as many green photosites as either red or blue photosites. It then interpolates the values it didn't record. Some people look at the lower amount of red and blue photosites as being analogous to 422 subsampling, but it's really a whole different kettle of fish.
paulherrin
Well-known member
it is definitely a much different thing, not just a semantics issue. the process and principles behind cameras that produce 4:2:2 video is absolutely different than how a RED gets an image. also, the final resulting image is much higher quality in terms of accuracy, resolution aside. and it's definitely easier to work with for vfx purposes. death to chroma sub-sampling.
so dan, just to clarify, you're recommending sharpening the luma apart from the chroma to be good practice for general purposes?
so dan, just to clarify, you're recommending sharpening the luma apart from the chroma to be good practice for general purposes?
Eki Halkka
Well-known member
- Joined
- Sep 24, 2008
- Messages
- 515
- Reaction score
- 0
- Points
- 0
- Age
- 58
- Location
- Helsinki, Finland
- Website
- www.kolumbus.fi
If that comparison was to be made, at 100% size (4K) the subsampling would be nominated something like 2:2:0 - there's a green sample for every 2nd and blue / red samples for every 4th resulting image pixel. At 2K, the sampling would be more or less 8:4:4 - there's two green and one red / blue photosites for each resulting pixel.
But as said, this is comparing apples to oranges.
When one uses the r3d files directly in After Effects for keying, they should simply be considered as true RGB, as there's no need to filter the chroma channels before keying like with 4:2:2 or 4:2.0.
Eki Halkka
Well-known member
- Joined
- Sep 24, 2008
- Messages
- 515
- Reaction score
- 0
- Points
- 0
- Age
- 58
- Location
- Helsinki, Finland
- Website
- www.kolumbus.fi
Well, not really "absolutely" different - Red just saves the image earlier in the chain, and thus saves in a different format - and leaves most of the usually in-camera processing to post. If one saves out a, say, a 4:2:2 quicktime or avi file from a red original, the resulting file is equivalent to the camera originals of most other cameras.
It works nicely usually. Especially with chroma-subsampled files.
Dan Hudgins
Well-known member
- Joined
- Feb 16, 2007
- Messages
- 1,369
- Reaction score
- 0
- Points
- 0
- Location
- San Francisco, CA USA
- Website
- www.dancad3d.com
OLPF
OLPF
The OLPF filter is not 100% effective, so some chroma moire passed through as does some aliasing.
I asked if the blur radius is set to the green or red/blue distance and seem to have gotten the answer that RED uses a filter mid-way between the two, meaning that you get more red/blue alaising and moire than green aliasing and moire, by more I mean in relation not that it is a large value before compression and anti-OLPF sharpen.
Anyway, because of the lower pixel fill factor for the red/blue the OLPF is always a bit less effective on the red/blue than the green anyway.
Since the luma is made of red/blue plus green, in a ratio that uses more green most of the time, the luma will have less aliasing than the chroma. So you can sharpen the luma more than the chroma in any given image, probably, before residual aliasing may become noticable (remember that the OLPF is not 100% effective, so some small amount of aliasing is in the image data, the more you sharpen the more you counter the OLPF to some degree, at least when working with True RAW data, in the wavelet compressed data you may also have a kind of fixed/moving pattern noise mixed into the low level aliasing and image data down at the level of the random noise).
Because the chroma may hold more aliasing and moire (as well as noise from the chroma matrix for white balance), if you don't sharpen it, then you would not bring it's artifact contrast higher making it maybe noticable.
If there is a large difference between the luma and chroma sharpen you may see some color "bleed" at sharp edges of bright colored subjects, but in a 4K or 5K image at normal viewing distance of 3x the image diagonal, your eyes might not be able to notice that much.
I don't like the "cut out" chroma look that some edge finding chroma filters can make, since the chroma edge is sharper than it would be formed by the lens, so for my taste its better to have a 1/2 pixel blur on the chroma edges rather than have the chroma edges with the "cut out and pasted on" color look which screams "digital" where as the softer look is more "film like" where the dyes bleed to a radius of 1 to 2 pixels somewhat due to the way color film processing works and the way light bounces around in movie film (which is why they use JET backing for anti-halation).
http://photo.net/film-and-processing-forum/009eqD
For doing a good job sharpening RGB images made from a Bayer sensor camera, its better I think to split the processing into luma and chroma and optimize the algorithms for each to make the sum of both take into account that the OLPF is not a 100% solution to the issue of chroma moire, aliasing, and noise build up from the need for anti-OLPF sharpening.
If you are working with 5K images from a Bayer camera, and always going down to 1920x1080 for your DCP (1920x1080 seems to be what people are using 1998x1080 seems to be not what is being used or what is projected in the viewable screen area, anyone have experence with making DCP want to talk about that issue?) then you need some blur in both the luma and chroma before resize or you will show aliasing in the end result because you need image spots to be at least four to nine pixels area to not alias, as anyone who does computer graphics knows, you cannot show moving images that have 100% edge contrast to the single pixel without aliasing because a 1/2 pixel move causes the point to be one pixel wide then two pixels wide then one pixel wide. If you sit far back your eyes may not see the moving edge or spot changing from one to two pixels to one pixel as it moves, but closer you will see a pulsation if the contrast is high. So how sharp you want your 5K images depends on the end use size and how it will be viewed. The de-Bayer gives best results when it is optimized for the end use size, and its better to make those adjustments to the luma and chroma bandpass in the de-Bayer, rather than after making an RGB image and re-processing that RGB image (as is being done with output from REDCINE-X) since you may add noise combining the luma with the chroma in making the RGB image since the luma will (or at least could if it is white balanced and anti-OLPFed) have higher noise to mix in.
Another issue is CA from the lens, if you sharpen the chroma it tends to make the CA more noticable.
OLPF
The OLPF filter is not 100% effective, so some chroma moire passed through as does some aliasing.
I asked if the blur radius is set to the green or red/blue distance and seem to have gotten the answer that RED uses a filter mid-way between the two, meaning that you get more red/blue alaising and moire than green aliasing and moire, by more I mean in relation not that it is a large value before compression and anti-OLPF sharpen.
Anyway, because of the lower pixel fill factor for the red/blue the OLPF is always a bit less effective on the red/blue than the green anyway.
Since the luma is made of red/blue plus green, in a ratio that uses more green most of the time, the luma will have less aliasing than the chroma. So you can sharpen the luma more than the chroma in any given image, probably, before residual aliasing may become noticable (remember that the OLPF is not 100% effective, so some small amount of aliasing is in the image data, the more you sharpen the more you counter the OLPF to some degree, at least when working with True RAW data, in the wavelet compressed data you may also have a kind of fixed/moving pattern noise mixed into the low level aliasing and image data down at the level of the random noise).
Because the chroma may hold more aliasing and moire (as well as noise from the chroma matrix for white balance), if you don't sharpen it, then you would not bring it's artifact contrast higher making it maybe noticable.
If there is a large difference between the luma and chroma sharpen you may see some color "bleed" at sharp edges of bright colored subjects, but in a 4K or 5K image at normal viewing distance of 3x the image diagonal, your eyes might not be able to notice that much.
I don't like the "cut out" chroma look that some edge finding chroma filters can make, since the chroma edge is sharper than it would be formed by the lens, so for my taste its better to have a 1/2 pixel blur on the chroma edges rather than have the chroma edges with the "cut out and pasted on" color look which screams "digital" where as the softer look is more "film like" where the dyes bleed to a radius of 1 to 2 pixels somewhat due to the way color film processing works and the way light bounces around in movie film (which is why they use JET backing for anti-halation).
http://photo.net/film-and-processing-forum/009eqD
For doing a good job sharpening RGB images made from a Bayer sensor camera, its better I think to split the processing into luma and chroma and optimize the algorithms for each to make the sum of both take into account that the OLPF is not a 100% solution to the issue of chroma moire, aliasing, and noise build up from the need for anti-OLPF sharpening.
If you are working with 5K images from a Bayer camera, and always going down to 1920x1080 for your DCP (1920x1080 seems to be what people are using 1998x1080 seems to be not what is being used or what is projected in the viewable screen area, anyone have experence with making DCP want to talk about that issue?) then you need some blur in both the luma and chroma before resize or you will show aliasing in the end result because you need image spots to be at least four to nine pixels area to not alias, as anyone who does computer graphics knows, you cannot show moving images that have 100% edge contrast to the single pixel without aliasing because a 1/2 pixel move causes the point to be one pixel wide then two pixels wide then one pixel wide. If you sit far back your eyes may not see the moving edge or spot changing from one to two pixels to one pixel as it moves, but closer you will see a pulsation if the contrast is high. So how sharp you want your 5K images depends on the end use size and how it will be viewed. The de-Bayer gives best results when it is optimized for the end use size, and its better to make those adjustments to the luma and chroma bandpass in the de-Bayer, rather than after making an RGB image and re-processing that RGB image (as is being done with output from REDCINE-X) since you may add noise combining the luma with the chroma in making the RGB image since the luma will (or at least could if it is white balanced and anti-OLPFed) have higher noise to mix in.
Another issue is CA from the lens, if you sharpen the chroma it tends to make the CA more noticable.
paulherrin
Well-known member
thanks dan, very informative.