AVS Forum banner

1281 - 1300 of 1453 Posts

·
Registered
Joined
·
860 Posts
Following on from Miki and my posts last night, I have decided to follow up with a few details as to why we have come to the conclusion that for now, in our opinion, the Three Matrix Correction technique would be our choice for probe matching at this point in time going forward.

After Miki made his post last week, we began to discuss a few things because theoretically TMC should indeed be more accurate than FCMM. So, we decided to look into this further and I took a series of measurements across the two picture modes (FCMM vs TMC) with the Spectro (Jeti 1501) over a standard 7^3 Anisometric cube patch set, and to both our surprise TMC actually seemed a little more accurate. There honestly wasn’t much in it, but it showed me the results given in the initial charts were actually reversed to our initial conclusions based on the corrected Klein measurement results alone.

So, we returned and took another look at FCMM vs TMC, and I have to say I am suitably impressed and I know Miki is too. To begin with I wanted to compare for myself how LightSpace FCMM compares with creating a matrix in Chromasurf and storing it within the Klein memory slots. LightSpace uses 3 patches of each RGBW at a patch value of 240 for each, and averages to give a final result to create the matrix. Chromasurf uses a single patch from RGBW with a value of 255, and somehow (I’m not sure how) balances luminance within the correction. But I also created and stored a Klein internal matrix based on 240 value patches too, wondering if I would see any tangible difference. Validating these three corrections against the Jeti over 125 patches in a 5^3 Anisometric built in patch set shows that on average, the method used in LightSpace is more accurate.



This gave us a good foundation to build upon with the implementation of TMC as we already know the method used for correction in LightSpace (with its averaging of three patches of each colour) is already very good. So, I measured a 7^3 Anisometric patch set with the Jeti, followed by the Klein with the FCMM probe match active. I then provided Miki with the BCS file for the Klein measurements and the BPD’s used in the correction matrix. He then extracted that data, converted it to TMC, reapplied the newly corrected data for me to compare against the Jeti.



Over 343 points, it corrected many points that were outside of NIST tolerance, and overall reduced Avg dE2K across the measurement range of colours. Greyscale Avg dE remains remarkably similar which is to be expected. What we can gather from this is that greyscale should remain consistent with colour accuracy improved. Looking at the points volumetrically in cube form, you can actually see where the errors are reduced.


With FCMM the biggest error is in Cyan with a dE of 0.4461. The error range has had to be scaled to show what is already such a low error more clearly, so although these bubbles look big you can see why it had to be scaled to this size in the following image.


This is TMC scaled to the exact same range as above. I would ask you to spot the difference, but I don’t think it is necessary. That Cyan error has been reduced to dE 0.2176, and all other (already small) errors that are in the above FCMM image have virtually disappeared.


We are not talking huge numbers here as the K10-A already works very well with FCMM, but it is a significant improvement nonetheless and I can foresee users with lower end probes such as the i1d3 seeing even more of an advantage to using this. As I still have an i1d3 to hand, that will be the next thing I look at after a day or two break from it all.

Although the implementation we have used is of a manual export -> convert -> import nature via an excel conversion sheet created by Miki using the iterative step method of TMC, it is actually a fairly quick, easy and consistently reliable method. It is converted from an initial FCMM match by using the BPD files used in the correction, and obviously the better you can get this (e.g. within NIST) then the better the resulting conversion will be. However, we have already seen others who previously could never get a valid NIST acceptable correction using FCMM easily get valid corrections using TMC from just the same 4 patches read with each probe. In other words, it just works.

We are still looking at MVPM (Multi-point Volumetric Probe Matching) but considering the ease in which the Three Matrix Correction can be applied, it is currently our preferred method for probe matching. For details on how TMC can be implemented into your LightSpace/ColourSpace profiles, then you would need to contact @Anger.miki directly where he can provide you with further details. He can be reached via PM here or he has also provided an email address in a previous post.

We would both like to express our full gratitude once again to @bobof for his time and expertise with all the help he has given this past week, and for the beautiful cube representation images I have used above.
 

·
Registered
Joined
·
4,893 Posts
Indeed, Miki and I have been testing extensively for about a week now. In fact very soon after he made his post. There was a reason we began the testing that will be explained in due course, but we have so much data to look through still and sort out properly, most of which is just a multitude of various different test that produce identical results. Once it has been thoroughly sorted through, organised and compiled into something easier to present, we will happily post our findings.

What we can say with complete confidence though is that the new Bodnar Method for probe matching (known as TMC, or Three Matrix Correction) does provide some tangible benefit over the traditional FCMM way of probe matching on consumer WOLED displays, and considering how easy it is to implement (even though we have to do it manually at this point), it would absolutely be my own personal choice to use from here going forward.

Special thanks also go to @bobof for his help, advice, and analysis skills over the past week.

We will share more sometime in the near future.

Stay safe everyone!
I am assuming that all the same guidelines for meter distancing, spot size, positioning, etc., are the same for TCM as for FCMM? I hadn't seen those factors mentioned, but maybe they don't need to be.
 

·
Registered
Joined
·
17,805 Posts
I am assuming that all the same guidelines for meter distancing, spot size, positioning, etc., are the same for TCM as for FCMM? I hadn't seen those factors mentioned, but maybe they don't need to be.
All prior setup recommendations still apply to TCM.
 

·
Registered
Joined
·
4,289 Posts
The reason it’s being called the Bodner method is because there is a tradition color science, when you come up with or discover something, it gets named in your honor.

Examples:

Barten Ramp

Hunt Effect

Judd-Voss CMF

McAdam Ellipses

Pointer’s Gamut

Yoshi Method = FCMM

Bodner Method = TMC


Ben Bodner, who is a color scientist from RIT that is working for LG electronics, deserves a tip of the hat for his research that produced this novel approach to profiling colorimeters.
 

·
Registered
Joined
·
860 Posts
I am assuming that all the same guidelines for meter distancing, spot size, positioning, etc., are the same for TCM as for FCMM? I hadn't seen those factors mentioned, but maybe they don't need to be.
The Bodnar Method (for those that want that name mentioned) or TMC is calculated from an initial 4 patch read with each probe of just RGBW just as it is in the Yoshi Method for FCMM. So yes, meter setup will be the same. But then in my opinion correct setup of all equipment, including meters is critical to getting the best results possible at all times. For example when taking non contact measurements it is always best practice to ensure FOV is correctly matched, meters properly aligned, level and perpendicular to the display.

Whilst it is appreciated that it is not always that easy to do such things in a home environment, you can be assured that such steps were taken during these tests.
 

·
aka jfinnie
Joined
·
4,215 Posts
Awesome work @liberator72 and @Anger.miki, and kudos to Ben Bodner for developing the method.

Thanks to @liberator72 for providing the datasets to play with.

As you mentioned the cubes presented have an exaggerated scaling applied to the patch dE... The cubes are generated using a python script and the plot.ly data visualisation libraries. The scaling is set such that the diameter of each patch (coloured per patch RGB values) is (10*dE2k)^3. This has the effect of the spheres rapidly getting smaller below 0.1dE, and rapidly getting larger above 0.1dE. This is done to accentuate features in the dataset, which otherwise are very close and hard to discern using linear methods.

It will be interesting to subject other probes to such analysis - my understanding is the particular highlighted region in the FCMM cube will be where the particular probe isn't a good match for the standard observer, and hence is causing a failure of FCMM with WRGB displays as the FCMM assumes an RGB additive display.

I'm interested to know in practice what issues there are along the boundary regions of the TMC. TMC approximates the subpixel mixing of the display, but the actual mixing in practice for a given RGB triplet will depend on display settings unless you can get into some fully native panel gamut mode all the way from RGB input values to the panel WRGB. I think you'd need some densely populated odd-shaped profiles concentrating on the border regions to find out what sort of scale of issues you may find there due to the mismatch. It doesn't appear there were any points in particularly significant error on the 7^3 profile, and you'd hope if you ended up with a point or two in a larger profile feeding into a LUT engine that the engine would be smart enough to discard.

I've previously provided the code for the patch sorter I developed to reduce display drift (which incidentally works by making the same assumptions about patch power per TMC - ie only max 3 subpixels active out of 4, with W being used to substitute RGB where possible). The code for generating these 3D views is even rougher than that was (!), has some plot.ly integration issue I've not fully got to the bottom of and really could do with some kind of GUI, so for now I won't be uploading that. The method is simple enough so with any luck someone will integrate it into some proper software before I get to go near it again... hehe. In the meantime if anyone has interesting cases they'd like plotting the format you'd need to provide me with is a CSV as follows: patch number,R,G,B,dE.

Looking at the points volumetrically in cube form, you can actually see where the errors are reduced.


With FCMM the biggest error is in Cyan with a dE of 0.4461. The error range has had to be scaled to show what is already such a low error more clearly, so although these bubbles look big you can see why it had to be scaled to this size in the following image.


This is TMC scaled to the exact same range as above. I would ask you to spot the difference, but I don’t think it is necessary. That Cyan error has been reduced to dE 0.2176, and all other (already small) errors that are in the above FCMM image have virtually disappeared.

We are not talking huge numbers here as the K10-A already works very well with FCMM, but it is a significant improvement nonetheless and I can foresee users with lower end probes such as the i1d3 seeing even more of an advantage to using this. As I still have an i1d3 to hand, that will be the next thing I look at after a day or two break from it all.
 

·
Registered
Joined
·
970 Posts
The reason it’s being called the Bodner method is because there is a tradition color science, when you come up with or discover something, it gets named in your honor.

Examples:

Barten Ramp

Hunt Effect

Judd-Voss CMF

McAdam Ellipses

Pointer’s Gamut

Yoshi Method = FCMM

Bodner Method = TMC


Ben Bodner, who is a color scientist from RIT that is working for LG electronics, deserves a tip of the hat for his research that produced this novel approach to profiling colorimeters.
Murphy's law :D
 
  • Like
Reactions: WiFi-Spy

·
Registered
Joined
·
472 Posts
More about FCMM and TMC

I have been investigating the best profile matching for a 3DLUT generation. For this reason I have studied the various approaches based on my panel (Oled) reset to its native 3DLUT, Instruments a Klein and a Jeti and obviously LS (indeed the alpha version of CS, which demonstrated itself quite stable and rich of features).

What I know from CS tools, is that the panel has linear issues. This brings to another item to be taken in account during probe matching: the luma matching. I.e. if I don't perform readings in the same luma conditions, I am not matching apples with apples. Same applies to verifications. And with all OLED drifts controls, that's not an easy task. As approach I have used the Jeti to check the panel before starting measuring with the Klein. Off course keeping in account the difference of Y due to measured area reading (as said a nightmare).

Anyhow, I have verified the difference between my Jeti and Klein readings, using FCMM and TMC. TMC has been generated off line, based on a spreadsheet developed [email protected], I have contributed with the Matrix generation part. All has been (1) verified on a 6 cube, and taking in account the encouraging results of the TMC mode, I have verified with (2) primary saturation swaps (as addressing points on the boundary of the matrixes), and with (3) edge points.

PS: At the end I have decided for using TMC.

Here some more details about what I have found.
1. 6 cube. Here the points on a RGB cube

and the outcome.
I would say that there is clear advantage for TMC.

2. Same with primary saturation sweeps.

and outcome.

3. Things are a little more difficult with edge patches


They are 255 and the different in reading between Jeti and Klein is much more evident. E.g. an RGB pattern with 13 in any position requires 2 minutes Jeti reading vs. a couple of seconds with Klein (in low light averaging). So no surprise that on low luminance patches there is inaccuracy.

You can see here the outcome sorted by the original sequence and by the reference level. I have computed a different Norm Y based on the group of sequence.
Generally talking, It appears that FCMM are quite in line at the triangle vertex, and definitely superior with primary colour combination.
 

Attachments

  • Like
Reactions: mrtickleuk

·
Registered
Joined
·
472 Posts
Again on TMC vs. FCMM

I have tried to perform what I call an E2E verification.
Basically I have generated a 21 cube profile with no correction, using Klein Oled internal matrix, applied FCMM and TMC on the result and generated the relevant 3DLUT exactly with the same parameters. Keep in mind in the same session I have also measure (bpd) for correction on the unit 3DLUT. All thsi to have everything as much as possible alligned. Then I have uploaded the 3DLUT and measured with my Jeti using a 6 cube profile.
You can find here a picture with all the excel detail, and herein a summary (the excel file is also attached if you like to play with).


From a dE2000 perspective you can see that in average TMC is better. But if you compare value per value you can see that in 49 out of 215 patches TMC is worse that FCMM. I have filtered in the above picture when this difference is above 0.1), while the below one gives an idea about their distribution (TMC in green):




Here also with details about dY, dx, dy, dxy to references (stil filtered as above)



Playing with sorting and filtering I could not find any convincing explation. But maybe it's has no sense to look for that, as TMC as FCMM is based on 4 colours reading cannot be 100% accurate and in some case FCMM may be better as evident in the below picture
 

Attachments

  • Like
Reactions: mrtickleuk

·
Registered
Joined
·
970 Posts
Enrico a point is missing, 6^3=216. Anyway, nice to have another confirmation that TMC is better than FCCM on WOLED.
 

·
Registered
Joined
·
472 Posts
Enrico a point is missing, 6^3=216. Anyway, nice to have another confirmation that TMC is better than FCCM on WOLED.
Black is missing. Present in bcs but never measured. Waste of time having Jeti staying there for 130 seconds for something you know. Thanks to @Light Illusion CS beta that allows to do that effortless.



PS: I had little doubt on TMC (we started verifying that together when it appeared), but, let's me say, I was expecting dE improvment for all the patches. By testing the full chain I think I have spotted limits of 4 coulor matrix.
 

·
Registered
Joined
·
1,081 Posts
I have tried to perform what I call an E2E verification.
Basically I have generated a 21 cube profile with no correction, using Klein Oled internal matrix, applied FCMM and TMC on the result and generated the relevant 3DLUT exactly with the same parameters. Keep in mind in the same session I have also measure (bpd) for correction on the unit 3DLUT. All thsi to have everything as much as possible alligned. Then I have uploaded the 3DLUT and measured with my Jeti using a 6 cube profile
I've done this a couple of weeks ago, but also tested volumetric probe match. cLUT validation was a 1.2K profile, not just 216 points. My results were slightly better than yours but very similar, which are very good results but naturally ur dE results are also depending how you set up the display in precal....

but, the real comparison is to compare this approach against the results of...

(1) using a traditional workflow for probe offset
(2) on same screen
(3) w/ exact same precal set up
(4) using same probes
(5) same sw versions
(6) do this within max. of 10 hrs screen drift, ie not more than 10 hrs usage on the screen
(7) run exact same main profile and validation profile

then pull diff of results - actual dE does not matter in this eval, only the dE diff matters...

- M
 

·
Registered
Joined
·
472 Posts
The 216 points are a quality assessment of the outcome. The main point is that in generating the LUT (21 cube) I have used both TMC and FCMM doing a single reading (something you cannot do with volumetric). In that session 4 colors reading were performed and verified with maximum care. I have choosen a 6 cube for verification to minimize the time, i.e. for being able to do all in a single session.

Yes, dE differences are the one of the most important and are the only dE related reported in my Summary sheet.

.
 

·
Registered
Joined
·
1,081 Posts
The 216 points are a quality assessment of the outcome.
I got that, ur validation sample size is too small. u waste time on a 21^3 yet validate (!) w/ 216 points. The validation is the most important part here.

The main point is that in generating the LUT (21 cube) I have used both TMC and FCMM doing a single reading (something you cannot do with volumetric).
yes u can.

btw, u've not done a single reading, u've done a single profile (consisting of thousands of single patch reads). :)
 

·
Registered
Joined
·
970 Posts
I think you can do that with MVPM if you have reproduced the bpd interpolation and the colour engine. Differently, you cannot apply it to a profile devoid of any correction.

Anyway @Iron Mike could you send me your best match (bdps and csv used for it) and verification, please? Since you and DeWayne don't use a Jeti, I would learn a lot looking into your files. Thanks in advance.
 

·
Registered
Joined
·
472 Posts
I got that, ur validation sample size is too small. u waste time on a 21^3 yet validate (!) w/ 216 points. The validation is the most important part here.

I was doing a real case. The final reading is there, with a reference spectro, for assessing its quality, not for measuring (I should have run the usual 10 cube I do for that). But what I see as tendence is likely true.

yes u can.

btw, u've not done a single reading, u've done a single profile (consisting of thousands of single patch reads). :)
For comparing volumetric you have to do two measurements, one with data not correct (that then you can correct to TMC and FCMM) and one with volumetric. The cannot be identical, despites all cautions you take. That was my point ....
 

·
Registered
Joined
·
78 Posts
Interesting the white point coordinates between liberator72 old whitepoint done with the i1d3 completely match between gen/raw CMF, and the WRGB EDR coordinates, while the newer probe doesn't, i wonder if this a indication that the older probe is more accurate, or that it at least measures the same as his.

BTW did you ever post coordinates for i1d3 with your newer whitepoint that you did with the jeti?
 

·
Registered
Joined
·
860 Posts
BTW did you ever post coordinates for i1d3 with your newer whitepoint that you did with the jeti?
I managed to gather readings from several i1d3's (both Retail and OEM) measuring the alternate whitepoint with the EDR active. Although they measure in the same "ballpark" they all differ, with only two coming fairly close to each other. So it seems random and likely varies by batch.

With the WOLED EDR in place they all measure anywhere in between x 0.310-0.314 - y 0.330-0.334. Where any particular meter falls within that is anyone's guess.
 

·
Registered
Joined
·
9,119 Posts
I managed to gather readings from several i1d3's (both Retail and OEM) measuring the alternate whitepoint with the EDR active. Although they measure in the same "ballpark" they all differ, with only two coming fairly close to each other. So it seems random and likely varies by batch.

With the WOLED EDR in place they all measure anywhere in between x 0.310-0.314 - y 0.330-0.334. Where any particular meter falls within that is anyone's guess.
Hmm, that's a pity. I have less need of it but still have an interest. Whereabouts were the two that were close?
 

·
Registered
Joined
·
860 Posts
Hmm, that's a pity. I have less need of it but still have an interest. Whereabouts were the two that were close?
In all honesty I really don't think it makes too much of a difference as the two that were close together were outliers. They all read differently but within that range. I could post each individual measurement, but it would be useless as there is no way to know if any other particular meter would match it. It would be just as useful to take a stab in the dark and pick a random value within the given range of x 0.310-0.314 and y 0.330-0.334 (with the WOLED EDR active).

What it does show though is that it is really beneficial to profile your meter to a Spectro so that you have some sort of reference.
 
1281 - 1300 of 1453 Posts
Top