Fix Version/s: None
As Bob Armstrong noted in
Bob recalls seeing some effects due to this change, and the commit message of the changeset (made on the LSST stack but not on the HSC stack) states:
for LSST sdss shape chooses a slightly different footprint. See computeAdaptiveMomentsBBox in SdssShape.cc.
Rewrite set_amom_bbox using afw::geom operators
This changes results slightly in lsst_dm_stack_demo (as the the region of
pixels used in the fit can differ by 1 pixel due to different rounding).
This ticket is to more deeply investigate the effects of this change and, in particular, see if it contributes to differences seen between the stacks: e.g. the increasing with magnitude offset of the SDSS trace radii:
Of note, the difference does not show up in a comparison of the psf model trace radii:
- relates to
DM-9535 Assess whether differences in Brighter-Fatter implementations are contributing to the trace radii differences: LSST vs. HSC
DM-6463 Please provide how-to-reproduce instructions for LSST/HSC comparison epics
DM-6815 Update LSST full-stack processing configuration to match best practice from HSC
DM-9028 Process the HSC RC dataset with the latest hscPipe 4.0.5
I have run the HSC stack through single frame processing for the HSC-I subset of the COSMOS RC dataset with and without this adaptation (having gone down a bit of a rabbit hole not recognizing at first that the HSC stack defaults to LOCAL image coordinates!). Of note, I also set processCcd.calibrate.measurePsf.reserveFraction=0.0 for these runs to eliminate differences in PSF model star selection contributing to differences in the processing.
The following plots show some measurement comparisons of matched catalogs between these two runs:
Comparison of Gaussian Fluxes:
Comparison of SDSS trace radii:
Comparison of model PSF trace radii (at the position of *calib_psfUsed stars)
While some scatter is introduced, it is a very small amount, and we see from the good agreement between the SDSS trace radii, this particular difference between the stacks is not contributing to the issue highlighted above, let alone any major differences in their processing results. As such, I consider this issue sufficiently explored and, while it would preclude any bitwise identical expectations, the differences introduced are at a very small level and not a worry.
A next likely candidate for the trace radii differences will be explored in
Bob Armstrong, would you mind having a look and see if you agree with my analysis/conclusions here?
I agree, that this makes only minor differences and shouldn't impact the problems you are seeing. Note, that even setting the reserveFraction to zero will not necessarily give you the same PSF stars because the initial detection algorithms are slightly different.
I don't think so as I am comparing two HSC 4.0.5 stack runs here: one with the nominal HSC setting of the bbox used in adaptive moments, one using a branch that adopts the LSST-style bbox setting. Otherwise, all should be exactly the same.
I think the easiest way to approach this is to apply the changes made on LSST to the HSC stack and see compare the output for the two HSC runs (with and without adaptation) as well as between the LSST and HSC stack runs.