Spatially non-uniform filtering is used to regularize the half-maps each iteration. The refinement iterates through the Expectation (alignment), Maximization (reconstruction), and Regularization steps until convergence. The actual maps that are output each iteration are those directly after the maximization – so the
map_sharp outputs for non-uniform refinement are raw reconstructions without local filtering done. The regularization step thus is intended to produce more accurate alignments. @apunjani may be able to comment more on the algorithmic differences; you can also check out the Nature Methods paper for a more detailed description.
As far as we are aware, there doesn’t seem to be a standard post-processing procedure, so local filtering may or may not be helpful for model building. Doing local resolution estimation --> local filtering --> sharpening might produce a map that looks a bit more interpretable than just the globally filtered map
PS: if you’re using the legacy version, there are a few differences between the new and legacy versions, and there’s a thread discussing the other map outputs (
map_filtered.mrc here: How is the Non-Uniform Refinement map_filtered map created). We generally suggest the “NEW” version be used as we’ve obtained superior map quality with it.