Main Page | Modules | Namespace List | Class Hierarchy | Alphabetical List | Class List | Directories | File List | Namespace Members | Class Members | File Members

EMAN::Refine3DAlignerQuaternion Class Reference
[a function or class that is CUDA enabled]

Refine alignment. More...

#include <aligner.h>

Inheritance diagram for EMAN::Refine3DAlignerQuaternion:

Inheritance graph
[legend]
Collaboration diagram for EMAN::Refine3DAlignerQuaternion:

Collaboration graph
[legend]
List of all members.

Public Member Functions

virtual EMDataalign (EMData *this_img, EMData *to_img, const string &cmp_name="sqeuclidean", const Dict &cmp_params=Dict()) const
 To align 'this_img' with another image passed in through its parameters.
virtual EMDataalign (EMData *this_img, EMData *to_img) const
virtual string get_name () const
 Get the Aligner's name.
virtual string get_desc () const
virtual TypeDict get_param_types () const

Static Public Member Functions

AlignerNEW ()

Static Public Attributes

const string NAME = "refine_3d"

Detailed Description

Refine alignment.

Refines a preliminary 3D alignment using a simplex algorithm. Subpixel precision. Target function for the simplex algorithm is a rotation along an arbitrary axis defined by a quaternion, whose rotation magnitude is defined by the vector length (hence the simplex varies the vecotr component of the quaternion). In addition the simplex varies translation. Using quaternions avoids gimbal lock. The simplex algorithm moves the function downhill in a ameboa like fasion, hence it may get stuck in a local minima if the two 3D models are already roughly aligned.

Parameters:
xform.align3d The Transform storing the starting guess. If unspecified the identity matrix is used
stepx The initial simplex step size in x
stepy The initial simplex step size in y
stepz The initial simplex step size in z
stepn0 The initial simplex step size in the first quaternion vecotr component
stepn1 The initial simplex step size in the second quaternion vecotr component
stepn2 The initial simplex step size in the third quaternion vecotr component
spin_coeff The multiplier appied to the spin (if it is too small or too large the simplex will not converge)
precision The precision which, if achieved, can stop the iterative refinement before reaching the maximum iterations
maxiter The maximum number of iterations that can be performed by the Simplex minimizer
maxshift Maximum translation in pixels in any direction.
Author:
John Flanagan (with code recyled from David Woolford)
Date:
Feb 3rd 2011

Definition at line 982 of file aligner.h.


Member Function Documentation

virtual EMData* EMAN::Refine3DAlignerQuaternion::align EMData this_img,
EMData to_img
const [inline, virtual]
 

Implements EMAN::Aligner.

Definition at line 988 of file aligner.h.

References align().

00989                         {
00990                                 return align(this_img, to_img, "sqeuclidean", Dict());
00991                         }

EMData * Refine3DAlignerQuaternion::align EMData this_img,
EMData to_img,
const string &  cmp_name = "sqeuclidean",
const Dict cmp_params = Dict()
const [virtual]
 

To align 'this_img' with another image passed in through its parameters.

The alignment uses a user-given comparison method to compare the two images. If none is given, a default one is used.

Parameters:
this_img The image to be compared.
to_img 'this_img" is aligned with 'to_img'.
cmp_name The comparison method to compare the two images.
cmp_params The parameter dictionary for comparison method.
Returns:
The aligned image.

Implements EMAN::Aligner.

Definition at line 1575 of file aligner.cpp.

References EMAN::EMData::cmp(), EMAN::EMData::get_ndim(), EMAN::EMData::get_xsize(), EMAN::Dict::has_key(), ImageDimensionException, NullPointerException, EMAN::Cmp::params, EMAN::EMData::process(), refalin3d_perturbquat(), EMAN::EMData::set_attr(), EMAN::Dict::set_default(), status, t, x, and y.

01577 {
01578         
01579         if (!to || !this_img) throw NullPointerException("Input image is null"); // not sure if this is necessary, it was there before I started
01580 
01581         if (to->get_ndim() != 3 || this_img->get_ndim() != 3) throw ImageDimensionException("The Refine3D aligner only works for 3D images");
01582 
01583 #ifdef EMAN2_USING_CUDA 
01584         if(EMData::usecuda == 1) {
01585                 if(!this_img->getcudarwdata()) this_img->copy_to_cuda();
01586                 if(!to->getcudarwdata()) to->copy_to_cuda();
01587         }
01588 #endif
01589 
01590         float sdi = 0.0;
01591         float sdj = 0.0;
01592         float sdk = 0.0;
01593         float sdx = 0.0;
01594         float sdy = 0.0;
01595         float sdz = 0.0;
01596         bool mirror = false;
01597         
01598         Transform* t;
01599         if (params.has_key("xform.align3d") ) {
01600                 // Unlike the 2d refine aligner, this class doesn't require the starting transform's
01601                 // parameters to form the starting guess. Instead the Transform itself
01602                 // is perturbed carefully (using quaternion rotation) to overcome problems that arise
01603                 // when you use orthogonally-based Euler angles
01604                 t = params["xform.align3d"];
01605         }else {
01606                 t = new Transform(); // is the identity
01607         }
01608         
01609         float spincoeff =  params.set_default("spin_coeff",10.0f); // spin coefficient, controls speed of convergence (sort of)
01610         
01611         int np = 6; // the number of dimensions
01612         Dict gsl_params;
01613         gsl_params["this"] = this_img;
01614         gsl_params["with"] = to;
01615         gsl_params["snr"]  = params["snr"];
01616         gsl_params["mirror"] = mirror;
01617         gsl_params["transform"] = t;    
01618         gsl_params["spincoeff"] = spincoeff;
01619         Dict altered_cmp_params(cmp_params);
01620         
01621         const gsl_multimin_fminimizer_type *T = gsl_multimin_fminimizer_nmsimplex;
01622         gsl_vector *ss = gsl_vector_alloc(np);
01623         
01624         float stepi = params.set_default("stepi",1.0f); // doesn't really matter b/c the vecor part will be normalized anyway
01625         float stepj = params.set_default("stepj",1.0f); // doesn't really matter b/c the vecor part will be normalized anyway
01626         float stepk = params.set_default("stepk",1.0f); // doesn't really matter b/c the vecor part will be normalized anyway
01627         float stepx = params.set_default("stepx",1.0f);
01628         float stepy = params.set_default("stepy",1.0f);
01629         float stepz = params.set_default("stepz",1.0f);
01630         
01631         //gsl_vector_set(ss, 0, stepw);
01632         gsl_vector_set(ss, 0, stepi);
01633         gsl_vector_set(ss, 1, stepj);
01634         gsl_vector_set(ss, 2, stepk);
01635         gsl_vector_set(ss, 3, stepx);
01636         gsl_vector_set(ss, 4, stepy);
01637         gsl_vector_set(ss, 5, stepz);
01638         
01639         gsl_vector *x = gsl_vector_alloc(np);
01640         gsl_vector_set(x, 0, sdi);
01641         gsl_vector_set(x, 1, sdj);
01642         gsl_vector_set(x, 2, sdk);
01643         gsl_vector_set(x, 3, sdx);
01644         gsl_vector_set(x, 4, sdy);
01645         gsl_vector_set(x, 5, sdz);
01646         
01647         gsl_multimin_function minex_func;
01648         Cmp *c = Factory < Cmp >::get(cmp_name, altered_cmp_params);
01649                 
01650         gsl_params["cmp"] = (void *) c;
01651         minex_func.f = &refalifn3dquat;
01652 
01653         minex_func.n = np;
01654         minex_func.params = (void *) &gsl_params;
01655         
01656         gsl_multimin_fminimizer *s = gsl_multimin_fminimizer_alloc(T, np);
01657         gsl_multimin_fminimizer_set(s, &minex_func, x, ss);
01658         
01659         int rval = GSL_CONTINUE;
01660         int status = GSL_SUCCESS;
01661         int iter = 1;
01662         
01663         float precision = params.set_default("precision",0.01f);
01664         int maxiter = params.set_default("maxiter",100);
01665         while (rval == GSL_CONTINUE && iter < maxiter) {
01666                 iter++;
01667                 status = gsl_multimin_fminimizer_iterate(s);
01668                 if (status) {
01669                         break;
01670                 }
01671                 rval = gsl_multimin_test_size(gsl_multimin_fminimizer_size(s), precision);
01672         }
01673 
01674         int maxshift = params.set_default("maxshift",-1);
01675 
01676         if (maxshift <= 0) {
01677                 maxshift = this_img->get_xsize() / 4;
01678         }
01679         float fmaxshift = static_cast<float>(maxshift);
01680         
01681         EMData *result;
01682         if ( fmaxshift >= (float)gsl_vector_get(s->x, 0) && fmaxshift >= (float)gsl_vector_get(s->x, 1)  && fmaxshift >= (float)gsl_vector_get(s->x, 2))
01683         {
01684                 float n0 = (float)gsl_vector_get(s->x, 0);
01685                 float n1 = (float)gsl_vector_get(s->x, 1);
01686                 float n2 = (float)gsl_vector_get(s->x, 2);
01687                 float x = (float)gsl_vector_get(s->x, 3);
01688                 float y = (float)gsl_vector_get(s->x, 4);
01689                 float z = (float)gsl_vector_get(s->x, 5);
01690                 
01691                 Transform tsoln = refalin3d_perturbquat(t,spincoeff,n0,n1,n2,x,y,z);
01692                         
01693                 result = this_img->process("xform",Dict("transform",&tsoln));
01694                 result->set_attr("xform.align3d",&tsoln);
01695                 result->set_attr("score", result->cmp(cmp_name,to,cmp_params));
01696                 
01697          //coda goes here
01698         } else { // The refine aligner failed - this shift went beyond the max shift
01699                 result = this_img->process("xform",Dict("transform",t));
01700                 result->set_attr("xform.align3d",t);
01701         }
01702         
01703         //EMData *result = this_img->process("xform",Dict("transform",t));
01704         delete t;
01705         t = 0;
01706         gsl_vector_free(x);
01707         gsl_vector_free(ss);
01708         gsl_multimin_fminimizer_free(s);
01709 
01710         if ( c != 0 ) delete c;
01711         return result;
01712 }

virtual string EMAN::Refine3DAlignerQuaternion::get_desc  )  const [inline, virtual]
 

Implements EMAN::Aligner.

Definition at line 998 of file aligner.h.

00999                         {
01000                                 return "Refines a preliminary 3D alignment using a simplex algorithm. Subpixel precision.";
01001                         }

virtual string EMAN::Refine3DAlignerQuaternion::get_name  )  const [inline, virtual]
 

Get the Aligner's name.

Each Aligner is identified by a unique name.

Returns:
The Aligner's name.

Implements EMAN::Aligner.

Definition at line 993 of file aligner.h.

00994                         {
00995                                 return NAME;
00996                         }

virtual TypeDict EMAN::Refine3DAlignerQuaternion::get_param_types  )  const [inline, virtual]
 

Implements EMAN::Aligner.

Definition at line 1008 of file aligner.h.

References EMAN::TypeDict::put().

01009                         {
01010                                 TypeDict d;
01011                                 d.put("xform.align3d", EMObject::TRANSFORM,"The Transform storing the starting guess. If unspecified the identity matrix is used");
01012                                 d.put("stepx", EMObject::FLOAT, "The initial simplex step size in x. Default is 1");
01013                                 d.put("stepy", EMObject::FLOAT, "The initial simplex step size in y. Default is 1");
01014                                 d.put("stepz", EMObject::FLOAT, "The initial simplex step size in z. Default is 1." );
01015                                 d.put("stepn0", EMObject::FLOAT, "The initial simplex step size in the first quaternion vecotr component. Default is 1." );
01016                                 d.put("stepn1", EMObject::FLOAT, "The initial simplex step size in the second quaternion vecotr component. Default is 1." );
01017                                 d.put("stepn2", EMObject::FLOAT, "The initial simplex step size in the third quaternion vecotr component. Default is 1." );
01018                                 d.put("spin_coeff", EMObject::FLOAT,"The multiplier appied to the spin (if it is too small or too large the simplex will not converge).  Default is 10.");
01019                                 d.put("precision", EMObject::FLOAT, "The precision which, if achieved, can stop the iterative refinement before reaching the maximum iterations. Default is 0.01." );
01020                                 d.put("maxiter", EMObject::INT, "The maximum number of iterations that can be performed by the Simplex minimizer. Default is 100.");
01021                                 d.put("maxshift", EMObject::INT,"Maximum translation in pixels in any direction. If the solution yields a shift beyond this value in any direction, then the refinement is judged a failure and the original alignment is used as the solution.");
01022                                 return d;
01023                         }

Aligner* EMAN::Refine3DAlignerQuaternion::NEW  )  [inline, static]
 

Definition at line 1003 of file aligner.h.

01004                         {
01005                                 return new Refine3DAlignerQuaternion();
01006                         }


Member Data Documentation

const string Refine3DAlignerQuaternion::NAME = "refine_3d" [static]
 

Definition at line 75 of file aligner.cpp.


The documentation for this class was generated from the following files:
Generated on Mon May 2 13:29:20 2011 for EMAN2 by  doxygen 1.3.9.1