Dietmar and Jochen,
Yeah, it's early development, though working pretty well even so I must say. And frankly there's going to be a nearly infinite range of dataset types that will be thrown at this, so...a tough thing to do.
If I recall from the prior 1.8 starfishy discussion - the stellar profile and quality coloring is pulled up from the linear stage, but the white outlining can be affected by other actions such as your autodev?
My 2600 and D5300 are APS-C both more or less in the ballpark (
) of 6000x4000. Yes the blue boxing is autoscaling, but the question is what is SVD using as its model for that autoscaling. If it's some kind of average size of all the white outlines, it could end up too small if there's a field full of tiny noisy spots that got white outlined, leaving the box undersized for the fewer, larger, but actual quality stars outlined and that will be actually used for sampling.
I had gone back to some data from earlier in the year on M13 for some testing (was actually testing something else but stumbled into this). Data is so-so. I hadn't yet refined my Newt for collimation and everything else, and didn't yet have my EFW and RGB filters. So I combined an hour of L from the 2600 and an hour of OSC from the D5300. It still came up with a fine result in 1.8 I think, and might be in the gallery here back in May or so.
Running these tests, including just on the D5300 data alone before extraction (not linked), depending on lack of bin and stretch/noise I can come up with a SVD sampling mask that pretty much has no stars to choose at all - just tiny outlines of mostly invisible noise or perhaps faint stars?
Here though, composited as the registered LRGB, and not binned, I get a ton of unusable outlines but also a decent sampling of good candidates. But note the blue box size (I've had even smaller). Note the zoom % of course, so I'm way way deep. It's not really
too bad, but does strike me as undersized.
-
- SVD9 sample screen.jpg (94 KiB) Viewed 16711 times
But if I then undo that SVD, bin to 35%, and run SVD again - I can get blue boxing much more appropriate to the white candidate outlines. Of course the SNR is improved too which may assist the white outlining.
I haven't yet tried full size/no-bin on some of my recent data (almost all SHO) but will this weekend if I can, to see what more I can find.
This link has these old M13 files though - L and the extracted RGB, all registered; and the log I used for this testing if anyone wants to see the settings in crop, wipe, etc. and SVD masking.
https://drive.google.com/drive/folders/ ... sp=sharing