Hi,
Currently in local refinement, there is an option for applying gaussian priors during local refinement, which as I understand it applies a penalty for rotation/offset changes that increases continuously with difference from the initial value.
This can work well, but I suspect we end up losing some particles that have genuinely large orientation errors that we would like to correct (without disrupting the ones that are already happy).
To that end, I wonder if it would be worthwhile to test adding Geman-McClure priors for local refinement, where the penalty still increases with distance, but then saturates, to allow for genuine outliers to not be so restrained. These are currently used with quite good effect for flexible fitting in Coot.
Cheers
Oli
1 Like
Hey @olibclarke – interesting! Do you have an idea of how large these “outlier” orientation errors might be? If they’re quite large, this might be a bit tricky since we typically only search relatively narrow rotation / shift extents in local refinement anyways (3 * sigma of the prior by default) and the “saturation” of more robust priors like GM might have little effect by that point.
1 Like
Hard to say for sure without testing with synthetic data where there is a ground truth available. But I guess the point is that it would allow you to use a tight prior with a larger search range than the current default (which would be pointless in the current implementation as larger deviations would be too heavily penalized)?
1 Like
Yep, assuming the priors are turned on. Ok, thanks for the suggestion, we’ll give this some thought!
1 Like