Well, I wonder actually. It's the only reason why I still switch to SGP for an initial focus and then back to Linux, that focus routine is very robust in my experience. It may require more power, but if it provides better results I'm totally fine with that. False positives can happen, but can be rejected as outliers if you get data from many more stars. In Ekos my problem is similar to the ones above, it works fine in certain star-fields, but if I get unlucky Ekos selects double stars or a way too bright one or one at the edge, that happens just too often in my case.

Ok so I added "Full Field" option. It works fine under simulation, but I doubt it will be useful as full-frames are _slow_ and star-detect algorithm is quite susceptible to noise, but hey, it's there now. I also started researching wavelet-based approaches to measuring focus, but it's still in the early stages. If anyone has any good pointers on how to proceed in that regard, let me know.

If it is error prone, how difficult would it be to have the auto select algorithm for a single star to choose one from the center 50% of the frame?

It was selecting between 5% and 95%, now changed it from 15% to 85% of the field. Might make it configurable later.

I guess you found lots of research about it already, but I stumbled on this paper which I thought was nice: ojs.cvut.cz/ojs/index.php/ap/article/viewFile/1462/1294

That was an interesting read, but without some code, this could take months to develop with my limited time.

Yeah code would be awesome, but it's very hard to actual find, I'll try though. In the mean time, maybe another interesting paper providing formulas (which should be possible to code) is nice?

Just as a reference, there's a nice talk about full frame guiding on TAIC: (sorry that should indeed be a link to their website instead).
I think the focusing issue is touched at some point.

Ah, look at that, the authors of that last paper wrote an actual Python open-source program about it: github.com/johnmgregoire/PyWaveletPeakID

It's all on 1D data I think they mention, would that be problematic given the 2D nature of an image? I read somewhere that a scan in the x and one in the y direction would then work..

Stop me if I'm spamming, but here's an interesting one on 2D wavelet analysis including Matlab code (which I doubt is useful): link.springer.com/content/pdf/10.1007/978-981-10-2537-2_4.pdf

I don't clearly see what would be the benefit of this fine-grained star detection. I went through the document and observed that the quality metric is mostly the number of stars detected. While there would statistically be a better coverage over the full frame, isn't detecting one single star over 9-square quadrants enough to drastically improve the focus quality? And as mentioned, selecting stars so that they may not be mistaken with others. We perhaps only need a "usual" source detector for this, and a clever algorithm.

-Eric

Yes I agree, but to be able to detect more real stars in an image, requires a good algorithm also for that 1 star. It requires to be able to define the edges of the signal, the peak of it etc. If you get that better, it will automatically result in more stars to be detected. Having a good understanding of that peak shape is what makes a good FWHM measurement possible and hence a good focusing procedure.