Measurement algorithms often want an estimate of the noise that doesn't include contributions from object shot noise (or at least not the shot noise of detected objects). DM-32355 will provide code to do this, but we probably actually want to run that code up front and save the results somewhere (e.g. in Exposure, or in cell-based coadd data structures). In some contexts (e.g. single-cell measurements) we may not always have enough background/object contrast to run the algorithm robustly, and in others its just wasteful to do it repeatedly.
It's worth thinking about:
- Whether to save a constant per visit+detector image, a full variance image for each detector+image, or something binned or smoothly varying.
- Whether to coadd a constant per cell or a full variance image per cell.
- Whether to actually save and/or propagate the fit parameters (see DM-32355) in bins or as per-image constants instead of the background estimates.
If we end up needing to propagate pixel covariance information through coaddition, we may want to consider rolling that up into the same object(s) as this information - effectively a class that attempts to represent "noise information not captured in the usual variance plane".