Continuing the discussion from Requirements for overhauled calibration task?:
I have run into an issue implementing the image characterization task. Iterating detect/measure/fit PSF measurement is not converging for my test data (which is an exposure from obs_test). In more detail, CharacterizeImageTask does the following; note the extra steps before the detect/measure iterations:
- add a simple PSF model to the exposure (if it doesn’t have one, and it usually won’t), like the old task. Measurement requires this.
- run repair, leaving cosmic rays in, as the old task does.
- subtract an initial estimate of background, like the old task, since detection gives insane results otherwise
- iterate the following N times (N is a config parameter):
- detect sources
- measure sources
- fit a PSF
I observe that the sigma of the fit PSF does not converge, though it does not vary much. For example starting with a naive estimate of 1.5 pixels, I see .55, 1.58, 1.50, 1.56, 1.55, 1.58, all based on a single source, alas, then 1.55 based on 2 sources and the next iteration fails with no eigen images. I will append a log as a reply, since attaching text files is not supported.
I realize this could be mis-configuration, but it worries me that it does not converge. Any ideas for improving convergence?
Also, I think @jbosch was hoping not to need an initial estimate for PSF FWHM, but I don’t see how to avoid it with the algorithm.