The new auto focus routine has generally been very successful for me. It definitely is a major improvement over the old routine in cases where there is a central obstruction and far out of focus, such that stars are donuts.
Many star fields, however, have enough stars that many multiple stars groups are detected as a single star, which has a very negative impact on the accuracy of the HFR calculation. This has been widely commented on in this forum. Here is an example:
There is a very high percentage of multiple star detections.
I understand that making the star detection routine reject these involves a lot of cpu overhead and makes the routine slow on slower pcs.
I think there is an easy solution to this problem that will reject all (or almost) all of these with no significant overhead:
Select stars using current method.
Eliminate the 50% (or whatever) of the stars with the highest HFR values. This will automatically reject all the multiple star selections and some of the single star selections, in almost all cases.
Voila! No more multiple star selections.
As long as there are at least a few stars selected, this should work fine. From my past detailed analysis of auto focus I have discovered that even using just one star gives excellent results, very comparable to using 300. So throwing half of them away, the half that contains the multiples, will improve the accuracy, in many cases dramatically.
I guess this may or may not affect AF. I can see how the mean HFR might be off, but we don’t really care about the absolute accuracy of the HFR, just relative accuracy. As focus changes, the size of the clustered stars will change with it. This would produce normal focus curves. I think a problem may occur when a cluster that was once selected as three stars (or whatever) is now selected as three individual stars. I have never personally had an issue with it.
Can you show that the accuracy of the curve is affected by this (with an AFPack)?
If required, I think it would be better to reject stars where the flux to radius ratio is disproportionately low… rejecting the highest HFR values might be bad.
I don’t think this is really the case. Assume a single star doubles in size. The diameter of a pair of the same size stars will not nearly double when each of the stars individually doubles in size. I think this would produce only a slighter worse HFR, whereas the single star HFR would be much worse. This has the affect of minimizing the true range of HFR values.
+1 here, multiple stars may have a very high apparent HFR not always related to the used star size since the resulting HFR depends also on the stars distance rather on the diameter. Near the focus when some stars are like pinpoint some remains detected with an huge HFR. It is difficult to believe in a very good result with wrong data. Ignoring bigger stars may be a single solution to test, no very difficult to code I think, SGP can ignore the smaller signal why not the bigger?, big HFR often (always?) correspond to saturated stars or multiple stars, the number of remaining ‘correctly detected’ stars are, in the majority of case more than sufficient to achieve a good focus (theoretically one star is enough) and are in a very close HFR range, so easy to filter.
Overall the current focus system works well for me with sct, but I agree that outlier rejection would be beneficial. The problem isn’t just that some stars skew the average by being artificially large - the big problem is when those stars are sometimes counted in the average and sometimes not. That will cause kinks in the curve that throw off the final focus value. That matters even more when there are only a few stars available - which can happen at long focal length.
So I agree that a simple rejection algorithm could help a lot when you have perhaps 4 or more stars.
Hello,
thank you Frank for your comment, I agree that the number of detected stars during the AF can heavily vary, and thus false big can significantly affect some of the AF frames more than others. But having plenty of stars (in the milky way for exemple) can simply product plenty of false big stars to be detected… so it is difficult to predict the real effect on V curve and the focus quality. Rejection filter for big stars could close the debate, but apparently this point does not want to be discussed.
Pascal
