Autofocuser method: Bahtinov vs gradient vs SEP. Linear, polynomial etc. Which is more reliable in your experience?

[pipnina]

I've been wrestling with getting the most accurate autofocus results recently. I thought I had my filter offsets nailed but while my OIII subs were coming through sharp last night, my Ha had donuts! Back to the drawing board.

I found running single-star (for speed) or full-frame both yielded rather inconsistent results. It would show me a V curve that looked to me like it landed at say, position 410 or so, but because sample 23 had a lower HFR at 386, it chooses 386 as the focus position. Then I run it on my Ha filter instead of Lum, and Ha comes out focused at 490, but my previous guess at filter offset was Ha should only be somewhere between 77 and 84 steps out, this is over 100!

 

I'm wondering if some people get better filter offset results with one-off bahtinov mask focusing. Then when imaging for real, using another method that doesn't involve touching the scope or going outside to add/remove masks to focus Lum and let the offsets sort out the filters actually in use?

[symmetal]

I use SGP and it overlays a best fit curve over the autofocus V curve, so tends to ignore the odd sample outside the norm. You need fairly steep sides to the V curve too for it to work well, set by your autofocus step size. If your best HFR is around 1.0 then the first and last samples should be around 5.0 or so.

For determining filter offsets it's best to run the autofocus several times and take the average reading. Also move the scope between repeated autocus runs to avoid a certain star concentration possibly throwing the result off slightly.

Alan

[Avocette]

Which autofocus software are you running? With your RPi4 I guessed you might be using Astroberry or StellarMate and hence Ekos. The latest version of KStars (3.6.0) (in StellarMate but not yet updated in Astroberry) is said to be yielding better results with a new version single pass linear process. The single pass establishes the V curve, finds the HFR minimum, the focuser is then displaced back to the other side of the curve, and starting from the original position moves again in the measurement direction to the step number where the minimum was found. Hence errors due to gearbox backlash should be eliminated.

I write this waiting excitedly for Astroberry to be updated, having just replaced two SW DC motor focusers with ZWO EAFs!

[pipnina]

2 minutes ago, Avocette said:

Which autofocus software are you running? With your RPi4 I guessed you might be using Astroberry or StellarMate and hence Ekos. The latest version of KStars (3.6.0) (in StellarMate but not yet updated in Astroberry) is said to be yielding better results with a new version single pass linear process. The single pass establishes the V curve, finds the HFR minimum, the focuser is then displaced back to the other side of the curve, and starting from the original position moves again in the measurement direction to the step number where the minimum was found. Hence errors due to gearbox backlash should be eliminated.

I write this waiting excitedly for Astroberry to be updated, having just replaced two SW DC motor focusers with ZWO EAFs!

I'm only on 3.5.9 at the moment on my desktop (currently my laptop runs the headless indi server, since my raspberry pi still has some glitches).

I'll see if I can get the newest version to control it with...

[pipnina]

Something I just noticed last night when trying out that new Kstars 1-pass linear fit.

If I run a pass of auto focus on luminance, and it comes out at 394 steps position, I can then take an image and it looks fine, quite nicely focused all things considered.

I can then do the same for Ha, it might come out at 478, and when I take an image it's similarly good.

So I set my filter offset to 84 on Ha (0 on lum).

Then I switch back to lum, the capture module in Ekos moves focus back by 84 and takes a photo. It looks fine, so in theory the focus offset is correct.

However! If I then ask for a Ha image again. It moves the focus forward by 84 again. Takes a picture, and it's got donuts! Even the diffraction spikes have split into two!

But moving back to lum remains focused, so it surely can't be backlash or the motion back would also have slack?

I adjusted the offset up to 116 manually as a stop-gap as that seemed to be much better focused than 84, despite the AF routine producing the 84 result and being accurate when it's made.

Another argument against backlash is that when doing single-pass linear fit, it moves right to left, finds the ideal spot, then winds left to right to the ideal spot, and that produces good results. With backlash the ideal spot would be found but not reached as the return movement would fall short?

 

Any ideas of what's going on here?