//===============================
//=====================================================================
void newiteration_residOpticalFlow_mt(int iter, const ap::real_1d_array& x, double f,const ap::real_1d_array& g, void *params)
{
globs_LBFGS_ *glob_param=(globs_LBFGS_*)params;
double normG=0.0;
for(int ii=g.getlowbound();ii<=g.gethighbound();ii++) normG+=g(ii)*g(ii);
normG=sqrt(normG);
cout<<"Iter="<<iter<<";fData="<<glob_param->fData<<";fSmooth="<<glob_param->fSmooth<<";fData+lambda*fSmooth="<<f<<";RMS(g)="<<normG/((double)(g.gethighbound()-g.getlowbound()+1))<<endl;
}
/*
*
* The cost function considered here is the following (with u(x,y,z),v(x,y,z),w(x,y,z) the flow at each pixel or super-pixel)
*
* E(u,v,w)=\sum_{s\inS}\sum_{n\in s} Huber(r(u_s,v_s,w_s),deltaHdataTerm) + \lambda \sum_{s \in S}\sum_{a\in Neigh(s) s.t. a<s} \left[ Huber(u_s-u_a,deltaHsmoothTerm) + Huber(v_s-v_a,deltaHsmoothTerm) + Huber(w_s-w_a,deltaHsmoothTerm) \right]
*
* S is the set of superpixels (stores in the partition). Thus, we impose that the flow is the same for voxels belonging to teh same supervoxel
*
* r(u,v,w)=I(x+u,y+v,z+w,t+1)-I(x,y,z,t)=imTarget(p+uvw)-imSource(p)
*
* I(x+u,y+v,z+w,t+1) and its partial derivatives need to be calculated using interpolation. As usual it is a trade-off between accuracy and speed.
*/
void funcgrad_residOpticalFlow_mt (ap::real_1d_array x, double& f, ap::real_1d_array& g, void *params)
{
#define INTERP_TRILINEAR //decides the kind of interpolation we want
globs_LBFGS_ *glob_param=(globs_LBFGS_*)params;
mylib::Partition* imTargetPartition=glob_param->imTargetPartition;
mylib::Array* imSource=glob_param->imSource;
mylib::Array* imTarget=glob_param->imTarget;
mylib::Array* imTarget_dx=glob_param->imTarget_dx;
mylib::Array* imTarget_dy=glob_param->imTarget_dy;
mylib::Array* imTarget_dz=glob_param->imTarget_dz;
vector<vector<pair<int,double> > >* partitionNeigh=glob_param->partitionNeigh;
//mylib::Region** regionPvec=glob_param->regionPvec;
double lambda=glob_param->lambda;
double deltaHsmoothTerm=glob_param->deltaHsmoothTerm;
//double deltaHdataTerm=glob_param->deltaHdataTerm;
//float* scale=glob_param->scale;
if(imSource->type!=mylib::UINT16_TYPE || imTarget->type!=mylib::UINT16_TYPE)
{
cout<<"ERROR: funcgrad_residOpticalFlow: code expects UINT16 images"<<endl;
exit(2);
}
//mylib::uint16* imSourcePtr=(mylib::uint16*)(imSource->data);
//initialize residual and gradient
int numPartitions=mylib::Get_Partition_Vertex_Count(imTargetPartition);
f=0;
double fSmooth=0;
memset(g.getcontent(),0,sizeof(double)*3*numPartitions);
int sizeX=x.gethighbound()-x.getlowbound()+1;
if(3*numPartitions!=sizeX)
{
cout<<"ERROR: funcgrad_residOpticalFlow: size of unknowns does not much number of image regions"<<endl;
exit(2);
}
int numPpos=0,numNeighPos=0;
//int *listedges=NULL;
//int nedges=0;;
double fAux,gAux,auxU,auxV,auxW;
//int ndims=imTarget->ndims;
//--------------------------
/*
//needed if we calculate data term directly in here instead of calling the function calculateDataTermOneRegion
mylib::Size_Type k;
mylib::Indx_Type p;
double auxI,der;
mylib::Coordinate *c=mylib::Make_Array(mylib::PLAIN_KIND,mylib::DIMN_TYPE,1,&ndims);
//memset(gAux,0,sizeof(double)*(ndims));
mylib::Region* regionP=NULL;
double* xx=new double[imTarget->ndims];
*/
//-------------------------
#ifdef INTERP_TRILINEAR
const mylib::Array** imTargetDer=(const mylib::Array**)malloc(sizeof(mylib::Array*)*imTarget->ndims);
imTargetDer[0]=imTarget_dx;imTargetDer[1]=imTarget_dy;imTargetDer[2]=imTarget_dz;
#endif
mylib::Use_Array_Basis(imTarget);//set basis to get all the coordinate indexes
//-------------------------------------------
/*
const int numThreads=12;
double* fVec=new double[numThreads];
boost::thread_group threads;
int numPini=0,numPend=0;
int step=(int)floor(double(numPartitions)/double(numThreads));
for (int i = 0; i < numThreads-1; ++i)
{
numPend+=step;
threads.create_thread(fDataTermThread(regionPvec,imTarget,imTargetDer,imSourcePtr,x.getcontent(),scale,deltaHdataTerm,fVec[i],g.getcontent(),numPini,numPend));
numPini+=step;
}
threads.create_thread(fDataTermThread(regionPvec,imTarget,imTargetDer,imSourcePtr,x.getcontent(),scale,deltaHdataTerm,fVec[numThreads-1],g.getcontent(),numPini,numPartitions));//residual
threads.join_all();
for (int i = 0; i < numThreads; ++i) f+=fVec[i];
delete[] fVec;
*/
//---------------------------------------------------------------------
/*
//single thread execution
for(int numP=0;numP<numPartitions;numP++)
{
numPpos=numP*ndims;//for parallelization
calculateDataTermOneRegion(regionPvec[numP],imTarget,imTargetDer,imSourcePtr,x.getcontent()+numPpos,scale,deltaHdataTerm,fAux,g.getcontent()+numPpos);
f+=fAux;
//for(int aa=0;aa<ndims;aa++) g(numPpos+1+aa)+=df[aa];//by passing g directly we avoid double copy
}
*/
//not needed anymore since I precomputed before minimization
//mylib::Free_Region(regionP);
//-------------------------------------------------------------------------
/*
//calculate smooth term (value and gradient) using only 2n-connectivity regions (very limited)
listedges=mylib::Get_Partition_Neighbors (imTargetPartition, numP, &nedges);
mylib::P_Edge* pEdge=NULL;
for(int ii=0;ii<nedges;ii++)
{
//pEdge->region1 < pEdge->region2 ALWAYS
pEdge=mylib::Get_Partition_Edge (imTargetPartition, listedges[ii]);
if(numP==(pEdge->region1))//to avoid double counting we only include edges with numP<neigh
{
numNeighPos=3*(pEdge->region2);
auxU=x(numPpos+1)-x(numNeighPos+1);
auxV=x(numPpos+2)-x(numNeighPos+2);
auxW=x(numPpos+3)-x(numNeighPos+3);
HuberCostAndDer(auxU,deltaHsmoothTerm,fAux,gAux);
fSmooth+=fAux;
g(numPpos+1)+=lambda*gAux;// *1.0
g(numNeighPos+1)-=lambda*gAux;
HuberCostAndDer(auxV,deltaHsmoothTerm,fAux,gAux);
fSmooth+=fAux;
g(numPpos+2)+=lambda*gAux;// *1.0
g(numNeighPos+2)-=lambda*gAux;
HuberCostAndDer(auxW,deltaHsmoothTerm,fAux,gAux);
fSmooth+=fAux;
g(numPpos+3)+=lambda*gAux;// *1.0
g(numNeighPos+3)-=lambda*gAux;
}else{
break;//pEdge->region1 < pEdge->region2 ALWAYS
}
}
*/
//--------------------------------------------------------------------------------------------
//calculate smooth term (value and gradient) using regions closer than maxDistance (precalculated ahead of time)
numPpos=0;
for(int numP=0;numP<numPartitions;numP++,numPpos+=3)
{
vector<pair<int,double> >* edges=&((*partitionNeigh)[numP]);
double lambdaAux;
for(vector<pair<int,double> >::const_iterator iter=edges->begin();iter!=edges->end();++iter)//we do not need to avoid double counting because it was already done during neighbor list creation
{
numNeighPos=3*(iter->first);
lambdaAux=lambda*(iter->second);
auxU=x(numPpos+1)-x(numNeighPos+1);
auxV=x(numPpos+2)-x(numNeighPos+2);
auxW=x(numPpos+3)-x(numNeighPos+3);
HuberCostAndDer(auxU,deltaHsmoothTerm,fAux,gAux);
fSmooth+=fAux*(iter->second);
g(numPpos+1)+=lambdaAux*gAux;//*1.0
g(numNeighPos+1)-=lambdaAux*gAux;
HuberCostAndDer(auxV,deltaHsmoothTerm,fAux,gAux);
fSmooth+=fAux*(iter->second);
g(numPpos+2)+=lambdaAux*gAux;//*1.0
g(numNeighPos+2)-=lambdaAux*gAux;
HuberCostAndDer(auxW,deltaHsmoothTerm,fAux,gAux);
fSmooth+=fAux*(iter->second);
g(numPpos+3)+=lambdaAux*gAux;//*1.0
g(numNeighPos+3)-=lambdaAux*gAux;
}
}
//--------------------------------debug------------------------
/*
cout<<"DEBUGGING!!! funcgrad_residOpticalFlow"<<endl;
//cout.precision(20);
//cout.setf(ios::fixed,ios::floatfield);
cout<<"fData="<<f<<";fSmooth="<<fSmooth<<";fData+lambda*fSmooth="<<f+lambda*fSmooth<<endl;
cout<<"x=[";
for(int ii=0;ii<3*numPartitions;ii+=3)
{
cout<<x(ii+1)<<" "<<x(ii+2)<<" "<<x(ii+3)<<";"<<endl;
}
cout<<"];"<<endl;
cout<<"g=[";
for(int ii=0;ii<3*numPartitions;ii+=3)
{
cout<<g(ii+1)<<" "<<g(ii+2)<<" "<<g(ii+3)<<";"<<endl;
}
cout<<"];"<<endl;
//exit(2);
*/
//-------------------------------------------------------------
//--------------------------------------debug: write out gradient based on superpixels-----------------------------------------
/*
for(int numP=0;numP<numPartitions;numP++,numPpos+=3)
{
//regionP=mylib::Record_P_Vertex(imTargetPartition,numP,1,0);//regions can have holes
regionP=regionPvec[numP];
for (k = 0; k < regionP->rastlen; k+=2)
{
for (p = regionP->raster[k]; p <= regionP->raster[k+1]; p++)//perform computation on voxel p
{
cout<<p<<" "<<numP+1<<" "<<g(3*numP+1)<<" "<<g(3*numP+2)<<" "<<g(3*numP+3)<<";"<<endl;
}
}
//mylib::Free_Region(regionP);
}
exit(2);
*/
//------------------------------------------------------------------------------------
#ifdef INTERP_TRILINEAR
//if(xx!=NULL) delete[] xx;
if(imTargetDer!=NULL) free(imTargetDer);
#endif
//save to display iteration info
glob_param->fData=f;
glob_param->fSmooth=fSmooth;
f+=lambda*fSmooth;
}