Data items in the EM_3D_REFINEMENT category record details
    about the class/particle refinement. 
  In random conical tilt, images were assigned angular positions through 
  rotational alignment and tilt-angles. From each different class, a 
  three-dimensional preliminary model is constructed. To improve the output, 
  those preliminary models from each class that have a high degree of 
  similarity are merged. In theory, these models corresponded to groups 
  of the same molecule just viewed from different orientations. Once all 
  the good random conical tilt models (and their corresponding particle 
  data sets) have been merged, iterative angular refinement is used to 
  improve the model's resolution. Equidistant projections are first 
  generated from the merged model. The entire particle data set (whether 
  the old random conical tilt experimental particles, or new untilted 
  experimental particles, or both) is then cross correlated to each 
  reference projection. A correlation coefficient is generated between 
  each experimental particle and reference projection. For each individual 
  experimental particle, it is matched to the reference projection that 
  gave the highest correlation coefficient. Therefore, it is assumed that 
  this particle matches the Euler angles of the reference projection.