Common Lines: fuzzy mask & Gaussian filter

[j-c-c]

Many of the common-lines methods in the MATLAB code apply a fuzzy_mask to the projections images and a Gaussian_filter to the polar Fourier images prior to estimating orientations. This is not implemented in here in the Python code.

[janden]

Both of these should help (fuzzy mask reduces influence of noise in the corners which we know is not informative and Gaussian filtering makes shift estimation slightly more likely to succeed). Would be worth to implement this at some point (can be done as extra preprocessing steps).

[j-c-c]

Turns out we are already using a Gaussian filter for the common-lines methods. It was "hiding" here in the _generate_shift_phase_and_filter() method, line 447.

[garrettwrong]

Yeah, I thought we talked about it, but I didn't want to jump the gun. Cool, thanks. If it's just the fuzzy mask, I think that is implemented as a method in ASPIRE already. Optimistically could just optionally enable. 📿

[j-c-c]

Tested out the CLSyncVoting with and without a "fuzzy_masked" source for a couple different levels of SNR. Saw pretty good improvement when using the fuzzy mask. I will open a PR and add a fuzzy mask option to the common line methods. Might be good to use the mask by default.

Here are the results comparing estimated rotations to ground truth ("mad" is mean_angular_distance in degrees):

For SNR = 0.5:
unmasked mse = 3.18, masked mse = 2.55
unmasked mad = 75, masked mad = 65

For SNR = 0.75:
unmasked mse = 2.78, masked mse = 0.44
unmasked mad = 67, masked mad = 20

Here is the script I used:
fuzzy_script.zip

[janden]

That's pretty encouraging, especially for high SNR. Might make sense to have it as a default. From what I recall, it's pretty standard in the Matlab ASPIRE examples, so it makes sense.

[j-c-c]

This was implemented in #1013.

Closing.