autofocus failure in bad seeing or bad initial guess

[nlourie]

the 5-point + parabolic fit approach to autofocus does not do a good job when the seeing is bad (~3-5 arcsec), or when the estimated focus point is not within ~50-100 micron or so of the focus at any given point. In practice the parabola is just too far off on one side. Expanding the parabola to a wider throw doesn't necessarily help, since far away from the focus the parabolic approximation just breaks down and kills the fit.

In bad seeing, the focus change gets so soft that the parabola also doesn't do a good job. In some cases the data points are so noisy that the fit totally breaks down and yields an upside down parabola with some random center within the throw.