nice! these numbers are super interesting, thanks for sharing! now it seems clear where your threshold of 320 comes from. interesting to see how it’s separated into the two modes. i wonder into which class lut profiles would fall, which map camera rgb to true xyz values. and yes, you would expect the RBF to have terrible scaling with the number of points.
1 Like
A small update:
The article we have written about the CLUT compression method used in G’MIC, has been accepted for publication at the CAIP’2019 conference (18th International Conference on Computer Analysis of Images and Patterns). That is great news!
We are currently writing a long version of the paper for a submission to a journal (which one is not decided yet).
We have also a French version of the paper, that has been accepted at the GRETSI’2019 conference (the main French conference about image and signal processing). which will be held in Lille / France, at the end of August.
Reviews for both English/French versions of the paper were really great, with no major concerns from the reviewers. That’s always nice to know we don’t have a lot of work to do before submitting the final version 
Now, we have to prepare the slides and/or posters for the presentations.
Still a lot of work to do, but having our work published somewhere as a scientific paper and at the same time implemented in an image processing software so that users around the globe can benefit from it is really satisfying for us (but it also doubles the amount of work, compared to a more classical research-only paper ).
10 Likes
The slides for the conferences are now ready!
Next two weeks, we have first a presentation at the GRETSI’2019 conference in Lille / France, then to the CAIP’2019 conference in Salerno. A busy start to the school year! 
3 Likes
A small update :
We have finally a long version of this CLUT compression work that has been accepted and published in SIIMS (SIAM Journal on Imaging Sciences), a good journal in our community.
We have also a reformatted version available on HAL (French equivalent of Arxiv), with lot more details and stuffs than in the CAIP’2019 conference paper.
You can get the .pdf here : https://hal-normandie-univ.archives-ouvertes.fr/hal-02940557/document
Probably not of broad interest, but maybe @hanatos could be interested ? (a bit of RBF inside ).
2 Likes
nice, thanks for the pointer. this has become quite an extensive writeup! i like the new visualisations of the RBF
i’m using what @gwgill called 2.5D luts in vkdt now for colour mapping. that is, a 2D thin plate spline RBF (i.e. log something radial function) on xy chromaticity values (or some similar space based on unbounded rec2020 rgb values), normalising away the input brightness (X+Y+Z, proportional to radiance or so). that’s because i’m mostly interested in replacing the dreadfully wrong input colour matrix transform by something better.
i’m hoping to do gamut mapping by just moving the control node points inwards, but didn’t push this to production quality yet.
1 Like
Did a talk today about CLUT compression/decompression in G’MIC. Just in case 
1 Like
really liked it, especially the practical part
Saw 5 or 10 minutes of the beginning (friends came around). Quite technical I thought…
Perhaps, in time, libraries of CLUTs will disappear. For I start with a good quality image that — for all its quality, does not convey the mood of my contemplated production. So after much work, some consultation, a bit of staring moodily out of windows, entertaining one or two thoughts that — in the next life — I shall take up something simple, possibly brain surgery, I arrive at an adjusted image that conveys the mood. Implicit between those two images is a CLUT. And that CLUT is not so much of a Holy Grail, just working data. The Holy Grail is, instead, that pair of images. And I might tweak one or another to generate a family of CLUTs, to be used and then cleared from storage when I need some space. But those two images — Ha! — They go in the vault.