Registration of 180-degree fisheye images

Mmm, tricky one, I've not seen fisheye stacking yet here. What does APP say exactly when you try to stack those? Or what does it look like? I can imagine that the distortion that changes with different pointing might make it very difficult to align them.

It either can't register at all, no matching star pairs found, or gives a bad registration value, typically several tens of pixels RMS, with areas that register OK-ish and other areas that are clearly way off.

The image above is a lucky one where both colours happened to be close enough in pointing to give nice registration. Some others I tried look pretty bad if they register at all.

I tried various amounts of stars, pattern recognitions, distortion correction, and registration model. I haven't played around yet with distortion margin.

I was hoping that the Calibrated Projective registration model would be able to deal with the distortion but that didn't work (yet). It probably needs to know more about the distortion in the lens, whether it's F * tan(theta) or F * theta projection or something else. The lens I used is https://www.sigmaphoto.com/8mm-f35-ex-dg-circular-fisheye

If needed I can upload some of the stacks.

Let me ask Mabula first if it's even a possibility. I might then ask for the data, always interesting if we can implement something that might work with it.

I asked Mabula and this is indeed not really possible. Especially as the view changes, making the distortion very different.

OK, thanks for checking. I'll put my hopes on the "not really possible" as opposed to "not possible" and hope for a (more distant) future update. 😉

That's always a good thing, having a positive mindset. haha The problem with fisheye lenses is that distortion, it's very variable and is likely not possible because it would need to distort the image to be able to align them. It wouldn't look good. Keeping the same point direction might still kind of work like you already noticed.