<classes> | File with class averages |
<3d model> | Preliminary 3d model to clean up |
[sym=<sym>] | Symmetry specification |
[proc=<nproc>] | Number of processors to use |
[3dit=<n iter>] | Number of 'normal' cleaning iterations |
[3dit2=<n iter>] | Number of 'zeroing' cleaning iterations |
[keepthr=<thr>] | Threshold for exclusion of class-averages (see make3diter) |
[keepsig=<sigmult>] | Alternate specification of threshold (see make3diter) |
[speed=<1-5>] | Speed in make3diter, convergence rate, dfl=1 |
[fixamp] | Uses the Fourier amplitudes from the original image |
clean3d classes.5.img threed.5a.img sym=d7 3dit=2 3dit2=2
This program performs iterative real-space refinement to make a 3D model more closely match a set of class-averages. There are 2 versions of the routine. The '3dit=' routine simply tries for the 'median' match. The '3dit2=' routine drives the densities towards zero as much as is permitted by the class averages.
The algorithm used by this program is quite difficult to describe. The basic idea is to calculate the differences between each class average and a corresponding projection of the model, then compute a 3D 'error model' describing the error in each pixel compared to unfiltered back-projection. However, this isn't strictly accurate, since not just the average error value is calculated for each voxel. In reality, we calculate a statistical set of predicted errors for each voxel. '3dit' uses the median value of the predicted error. '3dit2' uses a value between the median value, and the max/min value which would move the density towards zero. For a real explanation, you'll have to dissect the code. We'd like to retire this algorithm and replace it with something more mathematically robust, but in testing so far, it has produced the best corrections.
The '3dit2' algorithm can get a bit carried away if you run it for too many iterations. From experience we suggest '3dit2=2 3dit=1' will generally provide good results. Using the 'automask2' option of proc3d is also a good idea. during iterative refinement.