Initial_classification_FD gives incorrect angles

[janden]

Specifically, it doesn't seem to give the same angles as the old Initial_classification when dealing with flipped/reflected images. To see this, consider the test script

n = 1024;
sigma = 0.5;

n_nbor = 3;

f = load('cleanrib.mat');
vol = f.volref(1:4:end,1:4:end,1:4:end);
vol = vol/max(vol(:));

qs = qrand(n);

im_clean = cryo_project(vol, qs);

im = im_clean+sigma*randn(size(im_clean));

[sPCA_data, ~, ~, recon_spca] = data_sPCA(im, sigma^2);

[class, class_refl, class_rot] = Initial_classification_FD(sPCA_data, n_nbor, false);

fname = tempname;
WriteMRC(im, 1, fname);
im_reader = imagestackReader(fname);

tmpdir = tempname;
mkdir(tmpdir);

[~, ~, im_ave_file, ~] = ...
    align_main(im_reader, class_rot, class, class_refl, sPCA_data, n_nbor, 0, 1:n, recon_spca, tmpdir);

delete(fname);
rmdir(tmpdir);

im_ave = ReadMRC(im_ave_file);

delete(im_ave_file);

disp(norm(im(:)-im_clean(:))^2/norm(im_clean(:))^2);
disp(norm(im_ave(:)-im_clean(:))^2/norm(im_clean(:))^2);

Now adding the line

class_rot(class_refl==2) = mod(class_rot(class_refl==2)+180, 360);

just after the call to Initial_classification_FD brings down the MSE from about 0.65 to 0.50. The difference is even greater in less noisy situations, when the MSE for the class averages will actually be larger than those of the original images when not correcting for the angles.

Presumably this has something to do with how the sPCA coefficients and/or bispectrum is computed in the new version. That being said, adding the above line to the Initial_classification_FD should solve the problem.

[janden]

Fixed (hopefully) in 39950dd.