Visualizing the 3D color cube and the A7S green/yellow skintone issue

[jcs]

It appears the issue with Slog2 and S-gamut with the A7S is related to the nonlinear S-gamut Color Mode. My current work-around is to use Cinema Color Mode. If there is a significant benefit to using S-gamut, it appears some kind of 3D LUT is needed to fix it.

 

I few years ago I wrote custom software to do 3D color correction (for stereoscopic 3D images). The 3D color cube was directly displayed and could be edited, and it was also possible to render color swatch slices to check for issues with clipping (cube elements outside the unit cube space) or other mapping issues (such as color merging leading to banding/quantization).

 

Color cube display (unedited / unit-normalized cube shown (actual distorted cube for current color correction not shown)):

 

Color swatches (cube slices) showing range/clipping issues:

 

It may be possible to fix the S-gamut color mode issue by creating a 3D LUT which fixes the green/yellow splotches in skintones and perhaps also the magenta splotches in shadows. Since green and magenta are color complements, it kind of has the feel from physics describing a form of instability. That is, the system is stable in a narrow range, and when outside that range it becomes unstable, going extremely green or magenta, depending on which side of the parabolic curve center the color(s) are on.

 

[edit: hosted pics on my server]

[Inazuma]

Can you use tinypic.com or a similar service to host the pics?

[jgharding]

I can't see those images, perhaps use a free online pic host:

 

postimage.org

 

tinypic.com

 

There are Sony official LUTs to go from S-gamut to 709 too:

 

LUTS:

 

RosolveCubeFiles.zip â€2699 KB

 

DOCUMENTATION:

 

SonyLookProfilesSummaryV1_1.pdf â€450 KB

 

EDIT: found the images, I don't understand what they represent though :S

[sunyata]

Not sure about the "nonlinear S-gamut Color Mode", but I was doing some research and found a Sony whitepaper that had some useful formulas for s-log conversion, although it's a little dated. In the appendix there was this conversion matrix for S-Gamut to a "conventional" color space (HDW-F900) that is described as:

 

4.1 S-Gamut (Color Space Conversion)

4.1.1 F900 Color Space

 

"This color space should be used when users would like to match colors of images (like F35) to that of conventional Sony’s camcorders representing the HDW-F900. S-GAMUT to that for conventional cameras is shown below:"

 

 

(it's kinda camera specific)

 

But I tested these values anyway in a Nuke ColorMatrix node on A7s footage, after a manual sLog2 to linear conversion, and it actually looked pretty good (color-wise), maybe still a tad desaturated. Still looking for XYZ. 

 

[Updated]

 

 

Had to use Alexa LogC to Video gamma, extended to extended, noscale, then apply an "S-Gamut to sRGB" 3x3 transform.

 

So basically, got in range of correct with gamma and then the S-Gamut to sRGB, although now slightly over-saturated with blues compared to the F35 to F900 transform. It still needs work, but it's not too over stylized. Just a 1D gamma from Arri and then a 3x3 color transform, you can add whatever gamma and grade you need after.

 

Still trying to make a better S-log2 to linear 1D transform.

[TerrenceWilkins]

I'm stacking Impulz luts and getting this grade out of camera http://imgur.com/a/7AwhJ (stills from unchanged SLOG 2 footage)

[jcs]

I posted the 3D cubes to show what 3D LUTs look like when rendered. Understanding what happens to color spaces when the 3D RGB space is warped is helpful for understanding color transform issues. While the cubes are rendered with piece-wise linear line segments, the interpolation can be smooth/curved using tricubic interpolation (I wrote that software before general-purpose GPUs existed- tricubic interpolation is a bit slow in CPUs).

 

The first image shows the 3D LUT / cube, and the second image shows color swatches which represent a distorted 3D LUT: the green areas on the edges show where clipping has occurred. I found it helpful to visualize the 3D LUT and color swatches that cut through the 3D cube to better understand what the transform was doing as well as prevent strange issues when generalizing a transform to work on all possible input images.

 

JG- thanks for posting the LUTs. Terrence those look pretty good!

 

sunyata- Samuel H. (creator of the flaat profiles etc.) mentioned that S-gamut was non-linear. I haven't researched it; the transformation you posted is affine linear- though if applied after a non-linear slog2 to linear 1-D LUT the final linear transform will have a global change to color (the same way adjusting WB in post would work).

 

I'll give your and JG's 3D LUTs a try with Slog2+S-gamut which shows yellow splotches (before any transformations) and see if they go away. Thanks for posting the LUTs!

 

[edit]

Slog3 info: http://www.nzcine.com/wp-content/uploads/2014/07/TechnicalSummary_for_S-Gamut3Cine_S-Gamut3_S-Log3_V1_00.pdf (also shows full pseudocode for Slog3 to linear). Slog3+S-gamut3 is independent of WB transform while Slog2+S-gamut is not. Slog3 is more accurate. Why did Sony put Slog2 in the A7S instead of Slog3? ;)

[Luke Mason]

I think you guys are over complicating things, S-GAMUT on A7S uses x.v.Color to caputre extended gamut (1.8 times of sRGB), no post-processing softwares support this colour space, it will be scaled back to standard sRGB upon ingestion, which may cause issues.

 

S-GAMUT, although shares the same name between many Sony cameras (F5, F55, F65, A7S, FS700 etc.), are very different from each other, for example, S-GAMUT on F5 uses standard BT.709 with reduced saturation, while on F55/F65, it's a very wide proprietary colour gamut that pretty much covers REC.2020.

[sunyata]

Yes, that's understood Luke... it's just an experiment, nothing empirically correct (not yet anyway). The issue was trying to fix the greenness in the skin tones, I personally don't use 3D LUT's for baking in color transforms, I primarily use them for pre-vis and only use 1D transforms for converting log to lin etc, because a lot of what I render has to leave the main plate untouched. In this case though it's an interesting challenge.

 

jcs--

 

The difference btwn sLog2 and sLog3.. more detail in the low end, less in the highlights, people claim more noise vs Slog2.. 1.5 more advertised stops.

--

It does look like with what Sony is trying to do with the A7s, i.e. cram a lot of dynamic range into 8bit 4:2:0, sLog3 might not be a good fit, harder to correct, you may have to stretch the top end and compress the toe even more. And it looks like it would be more prone to clipping... hate to sound like a broken record but these profiles need 10bit.

[jcs]

I think you guys are over complicating things, S-GAMUT on A7S uses x.v.Color to caputre extended gamut (1.8 times of sRGB), no post-processing softwares support this colour space, it will be scaled back to standard sRGB upon ingestion, which may cause issues.

 

S-GAMUT, although shares the same name between many Sony cameras (F5, F55, F65, A7S, FS700 etc.), are very different from each other, for example, S-GAMUT on F5 uses standard BT.709 with reduced saturation, while on F55/F65, it's a very wide proprietary colour gamut that pretty much covers REC.2020.

 

Hey Luke- where is this documented? I did a brief search and found no math for Slog2 and its S-gamut. If the A7S's Slog2/S-gamut is custom for the camera, has Sony created A7S LUTs/matrices for conversion to linear/709?

 

sunyata- thanks, that's the gamma curve from the PDF; could create something close with curves by hand, however it appears the issue is with S-gamut (since Cinema Color Mode appears to do better than S-gamut).

 

Ultimately, we'd like to understand how the transforms work and what we need to get the proper conversions for the A7S in Premiere Pro and FCPX etc. (for most folks; Resolve would also be good too if the Slog2/S-gamut is different for the A7S)).

[sunyata]

From the pdf.

 

Recording system (movies):

 

Color space = xvYCC standard (x.v.Color when connected via HDMI cable) compatible with TRILUMINOS Color

 

http://www.sony.net/Products/di/en-us/translation_img/products/vq5f/overview/spec.pdf

 

this is probably the reference. 

 

b.t.w. the sLog2 vs sLog3 image was just to show how much higher middle gray was on sLog3.. 

[jcs]

Again, the luma curve (gamma) doesn't appear to be the issue with color- it's the Color Mode: S-gamut, which appears to be coupled to WB (from other docs and forum comments). That means it appears S-gamut changes with WB (the Slog2 version, not the Slog3 version which does not change with WB). Kholi on dvxuser swapped out S-gamut and used the Pro Color Mode and folks are getting good results without the yellow/green issue (he made other changes as well). I also found Pro to work well however to our eyes we like Cinema Color Mode better.

 

While we can tweak things in camera and in post, it would be helpful if Sony could release full specs (whitepaper) for the A7S's Slog2+S-gamut along with any LUTs for use with common NLEs (PPro, FCPX, Avid, Resolve, etc.). The limited docs state that the A7S's Slog2+S-gamut (8-bit) is not the same as the pro version (10-bit).

[sunyata]

Again, the luma curve (gamma) doesn't appear to be the issue with color- it's the Color Mode: 

 

jcs... I was responding to your comment below.

 

 Slog3 is more accurate. Why did Sony put Slog2 in the A7S instead of Slog3? ;)

 

... not suggesting it was a fix for the color issues.

[blafarm]

... has Sony created A7S LUTs/matrices for conversion to linear/709?

 

This thread is a bit old, but I though I'd chime in with a piece of information that may already be widely known.

 

Sony is working very closely with Atomos on the Shogun, which will be capable of displaying an input signal with a 3D LUT on it's Spyder-calibrated 1080p Rec.709 screen.

 

Atomos has told me that Shogun will have a Sony-supplied SLog2 3D LUT either at introduction -- or shortly thereafter.

 

And, as users will be able to import their own LUTs into the Shogun, one might assume that the Sony LUT might be exportable.