Google Git
Sign in
android / platform / development / d732021a676b4563520bd80e9b090ecaec84ff19 / . / perftests / panorama / feature_stab / db_vlvm / db_rob_image_homography.cpp
blob: 82dec0cbe6149e28f939ca20997b9a4d0545f0de [file] [log] [blame]
/*
* Copyright (C) 2011 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/* $Id: db_rob_image_homography.cpp,v 1.2 2011/06/17 14:03:31 mbansal Exp $ */
#include "db_utilities.h"
#include "db_rob_image_homography.h"
#include "db_bundle.h"
/*****************************************************************
* Lean and mean begins here *
*****************************************************************/
#include "db_image_homography.h"
#ifdef _VERBOSE_
#include <iostream>
using namespace std;
#endif /*VERBOSE*/
inline double db_RobImageHomography_Cost(double H[9],int point_count,double *x_i,double *xp_i,double one_over_scale2)
{
int c;
double back,acc,*x_i_temp,*xp_i_temp;
for(back=0.0,c=0;c<point_count;)
{
/*Take log of product of ten reprojection
errors to reduce nr of expensive log operations*/
if(c+9<point_count)
{
x_i_temp=x_i+(c<<1);
xp_i_temp=xp_i+(c<<1);
acc=db_ExpCauchyInhomogenousHomographyError(xp_i_temp,H,x_i_temp,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+2,H,x_i_temp+2,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+4,H,x_i_temp+4,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+6,H,x_i_temp+6,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+8,H,x_i_temp+8,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+10,H,x_i_temp+10,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+12,H,x_i_temp+12,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+14,H,x_i_temp+14,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+16,H,x_i_temp+16,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+18,H,x_i_temp+18,one_over_scale2);
c+=10;
}
else
{
for(acc=1.0;c<point_count;c++)
{
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i+(c<<1),H,x_i+(c<<1),one_over_scale2);
}
}
back+=log(acc);
}
return(back);
}
inline double db_RobImageHomography_Statistics(double H[9],int point_count,double *x_i,double *xp_i,double one_over_scale2,db_Statistics *stat,double thresh=DB_OUTLIER_THRESHOLD)
{
int c,i;
double t2,frac;
t2=thresh*thresh;
for(i=0,c=0;c<point_count;c++)
{
i+=(db_SquaredInhomogenousHomographyError(xp_i+(c<<1),H,x_i+(c<<1))*one_over_scale2<=t2)?1:0;
}
frac=((double)i)/((double)(db_maxi(point_count,1)));
#ifdef _VERBOSE_
std::cout << "Inlier Percentage RobImageHomography: " << frac*100.0 << "% out of " << point_count << " constraints" << std::endl;
#endif /*_VERBOSE_*/
if(stat)
{
stat->nr_points=point_count;
stat->one_over_scale2=one_over_scale2;
stat->nr_inliers=i;
stat->inlier_fraction=frac;
stat->cost=db_RobImageHomography_Cost(H,point_count,x_i,xp_i,one_over_scale2);
stat->model_dimension=0;
/*stat->nr_parameters=;*/
stat->lambda1=log(4.0);
stat->lambda2=log(4.0*((double)db_maxi(1,stat->nr_points)));
stat->lambda3=10.0;
stat->gric=stat->cost+stat->lambda1*stat->model_dimension*((double)stat->nr_points)+stat->lambda2*((double)stat->nr_parameters);
stat->inlier_evidence=((double)stat->nr_inliers)-stat->lambda3*((double)stat->nr_parameters);
}
return(frac);
}
/*Compute min_Jtf and upper right of JtJ. Return cost.*/
inline double db_RobImageHomography_Jacobians(double JtJ[81],double min_Jtf[9],double H[9],int point_count,double *x_i,double *xp_i,double one_over_scale2)
{
double back,Jf_dx[18],f[2],temp,temp2;
int i;
db_Zero(JtJ,81);
db_Zero(min_Jtf,9);
for(back=0.0,i=0;i<point_count;i++)
{
/*Compute reprojection error vector and its Jacobian
for this point*/
db_DerivativeCauchyInhomHomographyReprojection(Jf_dx,f,xp_i+(i<<1),H,x_i+(i<<1),one_over_scale2);
/*Perform
min_Jtf-=Jf_dx*f[0] and
min_Jtf-=(Jf_dx+9)*f[1] to accumulate -Jt%f*/
db_RowOperation9(min_Jtf,Jf_dx,f[0]);
db_RowOperation9(min_Jtf,Jf_dx+9,f[1]);
/*Accumulate upper right of JtJ with outer product*/
temp=Jf_dx[0]; temp2=Jf_dx[9];
JtJ[0]+=temp*Jf_dx[0]+temp2*Jf_dx[9];
JtJ[1]+=temp*Jf_dx[1]+temp2*Jf_dx[10];
JtJ[2]+=temp*Jf_dx[2]+temp2*Jf_dx[11];
JtJ[3]+=temp*Jf_dx[3]+temp2*Jf_dx[12];
JtJ[4]+=temp*Jf_dx[4]+temp2*Jf_dx[13];
JtJ[5]+=temp*Jf_dx[5]+temp2*Jf_dx[14];
JtJ[6]+=temp*Jf_dx[6]+temp2*Jf_dx[15];
JtJ[7]+=temp*Jf_dx[7]+temp2*Jf_dx[16];
JtJ[8]+=temp*Jf_dx[8]+temp2*Jf_dx[17];
temp=Jf_dx[1]; temp2=Jf_dx[10];
JtJ[10]+=temp*Jf_dx[1]+temp2*Jf_dx[10];
JtJ[11]+=temp*Jf_dx[2]+temp2*Jf_dx[11];
JtJ[12]+=temp*Jf_dx[3]+temp2*Jf_dx[12];
JtJ[13]+=temp*Jf_dx[4]+temp2*Jf_dx[13];
JtJ[14]+=temp*Jf_dx[5]+temp2*Jf_dx[14];
JtJ[15]+=temp*Jf_dx[6]+temp2*Jf_dx[15];
JtJ[16]+=temp*Jf_dx[7]+temp2*Jf_dx[16];
JtJ[17]+=temp*Jf_dx[8]+temp2*Jf_dx[17];
temp=Jf_dx[2]; temp2=Jf_dx[11];
JtJ[20]+=temp*Jf_dx[2]+temp2*Jf_dx[11];
JtJ[21]+=temp*Jf_dx[3]+temp2*Jf_dx[12];
JtJ[22]+=temp*Jf_dx[4]+temp2*Jf_dx[13];
JtJ[23]+=temp*Jf_dx[5]+temp2*Jf_dx[14];
JtJ[24]+=temp*Jf_dx[6]+temp2*Jf_dx[15];
JtJ[25]+=temp*Jf_dx[7]+temp2*Jf_dx[16];
JtJ[26]+=temp*Jf_dx[8]+temp2*Jf_dx[17];
temp=Jf_dx[3]; temp2=Jf_dx[12];
JtJ[30]+=temp*Jf_dx[3]+temp2*Jf_dx[12];
JtJ[31]+=temp*Jf_dx[4]+temp2*Jf_dx[13];
JtJ[32]+=temp*Jf_dx[5]+temp2*Jf_dx[14];
JtJ[33]+=temp*Jf_dx[6]+temp2*Jf_dx[15];
JtJ[34]+=temp*Jf_dx[7]+temp2*Jf_dx[16];
JtJ[35]+=temp*Jf_dx[8]+temp2*Jf_dx[17];
temp=Jf_dx[4]; temp2=Jf_dx[13];
JtJ[40]+=temp*Jf_dx[4]+temp2*Jf_dx[13];
JtJ[41]+=temp*Jf_dx[5]+temp2*Jf_dx[14];
JtJ[42]+=temp*Jf_dx[6]+temp2*Jf_dx[15];
JtJ[43]+=temp*Jf_dx[7]+temp2*Jf_dx[16];
JtJ[44]+=temp*Jf_dx[8]+temp2*Jf_dx[17];
temp=Jf_dx[5]; temp2=Jf_dx[14];
JtJ[50]+=temp*Jf_dx[5]+temp2*Jf_dx[14];
JtJ[51]+=temp*Jf_dx[6]+temp2*Jf_dx[15];
JtJ[52]+=temp*Jf_dx[7]+temp2*Jf_dx[16];
JtJ[53]+=temp*Jf_dx[8]+temp2*Jf_dx[17];
temp=Jf_dx[6]; temp2=Jf_dx[15];
JtJ[60]+=temp*Jf_dx[6]+temp2*Jf_dx[15];
JtJ[61]+=temp*Jf_dx[7]+temp2*Jf_dx[16];
JtJ[62]+=temp*Jf_dx[8]+temp2*Jf_dx[17];
temp=Jf_dx[7]; temp2=Jf_dx[16];
JtJ[70]+=temp*Jf_dx[7]+temp2*Jf_dx[16];
JtJ[71]+=temp*Jf_dx[8]+temp2*Jf_dx[17];
temp=Jf_dx[8]; temp2=Jf_dx[17];
JtJ[80]+=temp*Jf_dx[8]+temp2*Jf_dx[17];
/*Add square-sum to cost*/
back+=db_sqr(f[0])+db_sqr(f[1]);
}
return(back);
}
/*Compute min_Jtf and upper right of JtJ. Return cost*/
inline double db_RobCamRotation_Jacobians(double JtJ[9],double min_Jtf[3],double H[9],int point_count,double *x_i,double *xp_i,double one_over_scale2)
{
double back,Jf_dx[6],f[2];
int i,j;
db_Zero(JtJ,9);
db_Zero(min_Jtf,3);
for(back=0.0,i=0;i<point_count;i++)
{
/*Compute reprojection error vector and its Jacobian
for this point*/
j=(i<<1);
db_DerivativeCauchyInhomRotationReprojection(Jf_dx,f,xp_i+j,H,x_i+j,one_over_scale2);
/*Perform
min_Jtf-=Jf_dx*f[0] and
min_Jtf-=(Jf_dx+3)*f[1] to accumulate -Jt%f*/
db_RowOperation3(min_Jtf,Jf_dx,f[0]);
db_RowOperation3(min_Jtf,Jf_dx+3,f[1]);
/*Accumulate upper right of JtJ with outer product*/
JtJ[0]+=Jf_dx[0]*Jf_dx[0]+Jf_dx[3]*Jf_dx[3];
JtJ[1]+=Jf_dx[0]*Jf_dx[1]+Jf_dx[3]*Jf_dx[4];
JtJ[2]+=Jf_dx[0]*Jf_dx[2]+Jf_dx[3]*Jf_dx[5];
JtJ[4]+=Jf_dx[1]*Jf_dx[1]+Jf_dx[4]*Jf_dx[4];
JtJ[5]+=Jf_dx[1]*Jf_dx[2]+Jf_dx[4]*Jf_dx[5];
JtJ[8]+=Jf_dx[2]*Jf_dx[2]+Jf_dx[5]*Jf_dx[5];
/*Add square-sum to cost*/
back+=db_sqr(f[0])+db_sqr(f[1]);
}
return(back);
}
void db_RobCamRotation_Polish(double H[9],int point_count,double *x_i,double *xp_i,double one_over_scale2,
int max_iterations,double improvement_requirement)
{
int i,update,stop;
double lambda,cost,current_cost;
double JtJ[9],min_Jtf[3],dx[3],H_p_dx[9];
lambda=0.001;
for(update=1,stop=0,i=0;(stop<2) && (i<max_iterations);i++)
{
/*if first time since improvement, compute Jacobian and residual*/
if(update)
{
current_cost=db_RobCamRotation_Jacobians(JtJ,min_Jtf,H,point_count,x_i,xp_i,one_over_scale2);
update=0;
}
#ifdef _VERBOSE_
/*std::cout << "Cost:" << current_cost << " ";*/
#endif /*_VERBOSE_*/
/*Come up with a hypothesis dx
based on the current lambda*/
db_Compute_dx_3x3(dx,JtJ,min_Jtf,lambda);
/*Compute Cost(x+dx)*/
db_UpdateRotation(H_p_dx,H,dx);
cost=db_RobImageHomography_Cost(H_p_dx,point_count,x_i,xp_i,one_over_scale2);
/*Is there an improvement?*/
if(cost<current_cost)
{
/*improvement*/
if(current_cost-cost<current_cost*improvement_requirement) stop++;
else stop=0;
lambda*=0.1;
/*Move to the hypothesised position x+dx*/
current_cost=cost;
db_Copy9(H,H_p_dx);
db_OrthonormalizeRotation(H);
update=1;
#ifdef _VERBOSE_
std::cout << "Step" << i << "Imp,Lambda=" << lambda << "Cost:" << current_cost << std::endl;
#endif /*_VERBOSE_*/
}
else
{
/*no improvement*/
lambda*=10.0;
stop=0;
}
}
}
inline void db_RobImageHomographyFetchJacobian(double **JtJ_ref,double *min_Jtf,double **JtJ_temp_ref,double *min_Jtf_temp,int n,int *fetch_vector)
{
int i,j,t;
double *t1,*t2;
for(i=0;i<n;i++)
{
t=fetch_vector[i];
min_Jtf[i]=min_Jtf_temp[t];
t1=JtJ_ref[i];
t2=JtJ_temp_ref[t];
for(j=i;j<n;j++)
{
t1[j]=t2[fetch_vector[j]];
}
}
}
inline void db_RobImageHomographyMultiplyJacobian(double **JtJ_ref,double *min_Jtf,double **JtJ_temp_ref,double *min_Jtf_temp,double **JE_dx_ref,int n)
{
double JtJ_JE[72],*JtJ_JE_ref[9];
db_SetupMatrixRefs(JtJ_JE_ref,9,8,JtJ_JE);
db_SymmetricExtendUpperToLower(JtJ_temp_ref,9,9);
db_MultiplyMatricesAB(JtJ_JE_ref,JtJ_temp_ref,JE_dx_ref,9,9,n);
db_UpperMultiplyMatricesAtB(JtJ_ref,JE_dx_ref,JtJ_JE_ref,n,9,n);
db_MultiplyMatrixVectorAtb(min_Jtf,JE_dx_ref,min_Jtf_temp,n,9);
}
inline void db_RobImageHomographyJH_Js(double **JE_dx_ref,int j,double H[9])
{
/*Update of upper 2x2 is multiplication by
[s 0][ cos(theta) sin(theta)]
[0 s][-sin(theta) cos(theta)]*/
JE_dx_ref[0][j]=H[0];
JE_dx_ref[1][j]=H[1];
JE_dx_ref[2][j]=0;
JE_dx_ref[3][j]=H[2];
JE_dx_ref[4][j]=H[3];
JE_dx_ref[5][j]=0;
JE_dx_ref[6][j]=0;
JE_dx_ref[7][j]=0;
JE_dx_ref[8][j]=0;
}
inline void db_RobImageHomographyJH_JR(double **JE_dx_ref,int j,double H[9])
{
/*Update of upper 2x2 is multiplication by
[s 0][ cos(theta) sin(theta)]
[0 s][-sin(theta) cos(theta)]*/
JE_dx_ref[0][j]= H[3];
JE_dx_ref[1][j]= H[4];
JE_dx_ref[2][j]=0;
JE_dx_ref[3][j]= -H[0];
JE_dx_ref[4][j]= -H[1];
JE_dx_ref[5][j]=0;
JE_dx_ref[6][j]=0;
JE_dx_ref[7][j]=0;
JE_dx_ref[8][j]=0;
}
inline void db_RobImageHomographyJH_Jt(double **JE_dx_ref,int j,int k,double H[9])
{
JE_dx_ref[0][j]=0;
JE_dx_ref[1][j]=0;
JE_dx_ref[2][j]=1.0;
JE_dx_ref[3][j]=0;
JE_dx_ref[4][j]=0;
JE_dx_ref[5][j]=0;
JE_dx_ref[6][j]=0;
JE_dx_ref[7][j]=0;
JE_dx_ref[8][j]=0;
JE_dx_ref[0][k]=0;
JE_dx_ref[1][k]=0;
JE_dx_ref[2][k]=0;
JE_dx_ref[3][k]=0;
JE_dx_ref[4][k]=0;
JE_dx_ref[5][k]=1.0;
JE_dx_ref[6][k]=0;
JE_dx_ref[7][k]=0;
JE_dx_ref[8][k]=0;
}
inline void db_RobImageHomographyJH_dRotFocal(double **JE_dx_ref,int j,int k,int l,int m,double H[9])
{
double f,fi,fi2;
double R[9],J[9];
/*Updated matrix is diag(f+df,f+df)*dR*R*diag(1/(f+df),1/(f+df),1)*/
f=db_FocalAndRotFromCamRotFocalHomography(R,H);
fi=db_SafeReciprocal(f);
fi2=db_sqr(fi);
db_JacobianOfRotatedPointStride(J,R,3);
JE_dx_ref[0][j]= J[0];
JE_dx_ref[1][j]= J[1];
JE_dx_ref[2][j]=f* J[2];
JE_dx_ref[3][j]= J[3];
JE_dx_ref[4][j]= J[4];
JE_dx_ref[5][j]=f* J[5];
JE_dx_ref[6][j]=fi*J[6];
JE_dx_ref[7][j]=fi*J[7];
JE_dx_ref[8][j]= J[8];
db_JacobianOfRotatedPointStride(J,R+1,3);
JE_dx_ref[0][k]= J[0];
JE_dx_ref[1][k]= J[1];
JE_dx_ref[2][k]=f* J[2];
JE_dx_ref[3][k]= J[3];
JE_dx_ref[4][k]= J[4];
JE_dx_ref[5][k]=f* J[5];
JE_dx_ref[6][k]=fi*J[6];
JE_dx_ref[7][k]=fi*J[7];
JE_dx_ref[8][k]= J[8];
db_JacobianOfRotatedPointStride(J,R+2,3);
JE_dx_ref[0][l]= J[0];
JE_dx_ref[1][l]= J[1];
JE_dx_ref[2][l]=f* J[2];
JE_dx_ref[3][l]= J[3];
JE_dx_ref[4][l]= J[4];
JE_dx_ref[5][l]=f* J[5];
JE_dx_ref[6][l]=fi*J[6];
JE_dx_ref[7][l]=fi*J[7];
JE_dx_ref[8][l]= J[8];
JE_dx_ref[0][m]=0;
JE_dx_ref[1][m]=0;
JE_dx_ref[2][m]=H[2];
JE_dx_ref[3][m]=0;
JE_dx_ref[4][m]=0;
JE_dx_ref[5][m]=H[5];
JE_dx_ref[6][m]= -fi2*H[6];
JE_dx_ref[7][m]= -fi2*H[7];
JE_dx_ref[8][m]=0;
}
inline double db_RobImageHomography_Jacobians_Generic(double *JtJ_ref[8],double min_Jtf[8],int *num_param,int *frozen_coord,double H[9],int point_count,double *x_i,double *xp_i,int homography_type,double one_over_scale2)
{
double back;
int i,j,fetch_vector[8],n;
double JtJ_temp[81],min_Jtf_temp[9],JE_dx[72];
double *JE_dx_ref[9],*JtJ_temp_ref[9];
/*Compute cost and JtJ,min_Jtf with respect to H*/
back=db_RobImageHomography_Jacobians(JtJ_temp,min_Jtf_temp,H,point_count,x_i,xp_i,one_over_scale2);
/*Compute JtJ,min_Jtf with respect to the right parameters
The formulas are
JtJ=transpose(JE_dx)*JtJ*JE_dx and
min_Jtf=transpose(JE_dx)*min_Jtf,
where the 9xN matrix JE_dx is the Jacobian of H with respect
to the update*/
db_SetupMatrixRefs(JtJ_temp_ref,9,9,JtJ_temp);
db_SetupMatrixRefs(JE_dx_ref,9,8,JE_dx);
switch(homography_type)
{
case DB_HOMOGRAPHY_TYPE_SIMILARITY:
case DB_HOMOGRAPHY_TYPE_SIMILARITY_U:
n=4;
db_RobImageHomographyJH_Js(JE_dx_ref,0,H);
db_RobImageHomographyJH_JR(JE_dx_ref,1,H);
db_RobImageHomographyJH_Jt(JE_dx_ref,2,3,H);
db_RobImageHomographyMultiplyJacobian(JtJ_ref,min_Jtf,JtJ_temp_ref,min_Jtf_temp,JE_dx_ref,n);
break;
case DB_HOMOGRAPHY_TYPE_ROTATION:
case DB_HOMOGRAPHY_TYPE_ROTATION_U:
n=1;
db_RobImageHomographyJH_JR(JE_dx_ref,0,H);
db_RobImageHomographyMultiplyJacobian(JtJ_ref,min_Jtf,JtJ_temp_ref,min_Jtf_temp,JE_dx_ref,n);
break;
case DB_HOMOGRAPHY_TYPE_SCALING:
n=1;
db_RobImageHomographyJH_Js(JE_dx_ref,0,H);
db_RobImageHomographyMultiplyJacobian(JtJ_ref,min_Jtf,JtJ_temp_ref,min_Jtf_temp,JE_dx_ref,n);
break;
case DB_HOMOGRAPHY_TYPE_S_T:
n=3;
db_RobImageHomographyJH_Js(JE_dx_ref,0,H);
db_RobImageHomographyJH_Jt(JE_dx_ref,1,2,H);
db_RobImageHomographyMultiplyJacobian(JtJ_ref,min_Jtf,JtJ_temp_ref,min_Jtf_temp,JE_dx_ref,n);
break;
case DB_HOMOGRAPHY_TYPE_R_T:
n=3;
db_RobImageHomographyJH_JR(JE_dx_ref,0,H);
db_RobImageHomographyJH_Jt(JE_dx_ref,1,2,H);
db_RobImageHomographyMultiplyJacobian(JtJ_ref,min_Jtf,JtJ_temp_ref,min_Jtf_temp,JE_dx_ref,n);
break;
case DB_HOMOGRAPHY_TYPE_R_S:
n=2;
db_RobImageHomographyJH_Js(JE_dx_ref,0,H);
db_RobImageHomographyJH_JR(JE_dx_ref,1,H);
db_RobImageHomographyMultiplyJacobian(JtJ_ref,min_Jtf,JtJ_temp_ref,min_Jtf_temp,JE_dx_ref,n);
break;
case DB_HOMOGRAPHY_TYPE_TRANSLATION:
n=2;
fetch_vector[0]=2;
fetch_vector[1]=5;
db_RobImageHomographyFetchJacobian(JtJ_ref,min_Jtf,JtJ_temp_ref,min_Jtf_temp,n,fetch_vector);
break;
case DB_HOMOGRAPHY_TYPE_AFFINE:
n=6;
fetch_vector[0]=0;
fetch_vector[1]=1;
fetch_vector[2]=2;
fetch_vector[3]=3;
fetch_vector[4]=4;
fetch_vector[5]=5;
db_RobImageHomographyFetchJacobian(JtJ_ref,min_Jtf,JtJ_temp_ref,min_Jtf_temp,n,fetch_vector);
break;
case DB_HOMOGRAPHY_TYPE_PROJECTIVE:
n=8;
*frozen_coord=db_MaxAbsIndex9(H);
for(j=0,i=0;i<9;i++) if(i!=(*frozen_coord))
{
fetch_vector[j]=i;
j++;
}
db_RobImageHomographyFetchJacobian(JtJ_ref,min_Jtf,JtJ_temp_ref,min_Jtf_temp,n,fetch_vector);
break;
case DB_HOMOGRAPHY_TYPE_CAMROTATION_F:
case DB_HOMOGRAPHY_TYPE_CAMROTATION_F_UD:
n=4;
db_RobImageHomographyJH_dRotFocal(JE_dx_ref,0,1,2,3,H);
db_RobImageHomographyMultiplyJacobian(JtJ_ref,min_Jtf,JtJ_temp_ref,min_Jtf_temp,JE_dx_ref,n);
break;
}
*num_param=n;
return(back);
}
inline void db_ImageHomographyUpdateGeneric(double H_p_dx[9],double H[9],double *dx,int homography_type,int frozen_coord)
{
switch(homography_type)
{
case DB_HOMOGRAPHY_TYPE_SIMILARITY:
case DB_HOMOGRAPHY_TYPE_SIMILARITY_U:
db_Copy9(H_p_dx,H);
db_MultiplyScaleOntoImageHomography(H,1.0+dx[0]);
db_MultiplyRotationOntoImageHomography(H,dx[1]);
H_p_dx[2]+=dx[2];
H_p_dx[5]+=dx[3];
break;
case DB_HOMOGRAPHY_TYPE_ROTATION:
case DB_HOMOGRAPHY_TYPE_ROTATION_U:
db_MultiplyRotationOntoImageHomography(H,dx[0]);
break;
case DB_HOMOGRAPHY_TYPE_SCALING:
db_MultiplyScaleOntoImageHomography(H,1.0+dx[0]);
break;
case DB_HOMOGRAPHY_TYPE_S_T:
db_Copy9(H_p_dx,H);
db_MultiplyScaleOntoImageHomography(H,1.0+dx[0]);
H_p_dx[2]+=dx[1];
H_p_dx[5]+=dx[2];
break;
case DB_HOMOGRAPHY_TYPE_R_T:
db_Copy9(H_p_dx,H);
db_MultiplyRotationOntoImageHomography(H,dx[0]);
H_p_dx[2]+=dx[1];
H_p_dx[5]+=dx[2];
break;
case DB_HOMOGRAPHY_TYPE_R_S:
db_Copy9(H_p_dx,H);
db_MultiplyScaleOntoImageHomography(H,1.0+dx[0]);
db_MultiplyRotationOntoImageHomography(H,dx[1]);
break;
case DB_HOMOGRAPHY_TYPE_TRANSLATION:
db_Copy9(H_p_dx,H);
H_p_dx[2]+=dx[0];
H_p_dx[5]+=dx[1];
break;
case DB_HOMOGRAPHY_TYPE_AFFINE:
db_UpdateImageHomographyAffine(H_p_dx,H,dx);
break;
case DB_HOMOGRAPHY_TYPE_PROJECTIVE:
db_UpdateImageHomographyProjective(H_p_dx,H,dx,frozen_coord);
break;
case DB_HOMOGRAPHY_TYPE_CAMROTATION_F:
case DB_HOMOGRAPHY_TYPE_CAMROTATION_F_UD:
db_UpdateRotFocalHomography(H_p_dx,H,dx);
break;
}
}
void db_RobCamRotation_Polish_Generic(double H[9],int point_count,int homography_type,double *x_i,double *xp_i,double one_over_scale2,
int max_iterations,double improvement_requirement)
{
int i,update,stop,n;
int frozen_coord = 0;
double lambda,cost,current_cost;
double JtJ[72],min_Jtf[9],dx[8],H_p_dx[9];
double *JtJ_ref[9],d[8];
lambda=0.001;
for(update=1,stop=0,i=0;(stop<2) && (i<max_iterations);i++)
{
/*if first time since improvement, compute Jacobian and residual*/
if(update)
{
db_SetupMatrixRefs(JtJ_ref,9,8,JtJ);
current_cost=db_RobImageHomography_Jacobians_Generic(JtJ_ref,min_Jtf,&n,&frozen_coord,H,point_count,x_i,xp_i,homography_type,one_over_scale2);
update=0;
}
#ifdef _VERBOSE_
/*std::cout << "Cost:" << current_cost << " ";*/
#endif /*_VERBOSE_*/
/*Come up with a hypothesis dx
based on the current lambda*/
db_Compute_dx(dx,JtJ_ref,min_Jtf,lambda,d,n);
/*Compute Cost(x+dx)*/
db_ImageHomographyUpdateGeneric(H_p_dx,H,dx,homography_type,frozen_coord);
cost=db_RobImageHomography_Cost(H_p_dx,point_count,x_i,xp_i,one_over_scale2);
/*Is there an improvement?*/
if(cost<current_cost)
{
/*improvement*/
if(current_cost-cost<current_cost*improvement_requirement) stop++;
else stop=0;
lambda*=0.1;
/*Move to the hypothesised position x+dx*/
current_cost=cost;
db_Copy9(H,H_p_dx);
update=1;
#ifdef _VERBOSE_
std::cout << "Step" << i << "Imp,Lambda=" << lambda << "Cost:" << current_cost << std::endl;
#endif /*_VERBOSE_*/
}
else
{
/*no improvement*/
lambda*=10.0;
stop=0;
}
}
}
void db_RobImageHomography(
/*Best homography*/
double H[9],
/*2DPoint to 2DPoint constraints
Points are assumed to be given in
homogenous coordinates*/
double *im, double *im_p,
/*Nr of points in total*/
int nr_points,
/*Calibration matrices
used to normalize the points*/
double K[9],
double Kp[9],
/*Pre-allocated space temp_d
should point to at least
12*nr_samples+10*nr_points
allocated positions*/
double *temp_d,
/*Pre-allocated space temp_i
should point to at least
max(nr_samples,nr_points)
allocated positions*/
int *temp_i,
int homography_type,
db_Statistics *stat,
int max_iterations,
int max_points,
double scale,
int nr_samples,
int chunk_size,
/////////////////////////////////////////////
// regular use: set outlierremoveflagE =0;
// flag for the outlier removal
int outlierremoveflagE,
// if flag is 1, then the following variables
// need the input
//////////////////////////////////////
// 3D coordinates
double *wp,
// its corresponding stereo pair's points
double *im_r,
// raw image coordinates
double *im_raw, double *im_raw_p,
// final matches
int *finalNumE)
{
/*Random seed*/
int r_seed;
int point_count_new;
/*Counters*/
int i,j,c,point_count,hyp_count;
int last_hyp,new_last_hyp,last_corr;
int pos,point_pos,last_point;
/*Accumulator*/
double acc;
/*Hypothesis pointer*/
double *hyp_point;
/*Random sample*/
int s[4];
/*Pivot for hypothesis pruning*/
double pivot;
/*Best hypothesis position*/
int best_pos;
/*Best score*/
double lowest_cost;
/*One over the squared scale of
Cauchy distribution*/
double one_over_scale2;
/*temporary pointers*/
double *x_i_temp,*xp_i_temp;
/*Temporary space for inverse calibration matrices*/
double K_inv[9];
double Kp_inv[9];
/*Temporary space for homography*/
double H_temp[9],H_temp2[9];
/*Pointers to homogenous coordinates*/
double *x_h_point,*xp_h_point;
/*Array of pointers to inhomogenous coordinates*/
double *X[3],*Xp[3];
/*Similarity parameters*/
int orientation_preserving,allow_scaling,allow_rotation,allow_translation,sample_size;
/*Homogenous coordinates of image points in first image*/
double *x_h;
/*Homogenous coordinates of image points in second image*/
double *xp_h;
/*Inhomogenous coordinates of image points in first image*/
double *x_i;
/*Inhomogenous coordinates of image points in second image*/
double *xp_i;
/*Homography hypotheses*/
double *hyp_H_array;
/*Cost array*/
double *hyp_cost_array;
/*Permutation of the hypotheses*/
int *hyp_perm;
/*Sample of the points*/
int *point_perm;
/*Temporary space for quick-select
2*nr_samples*/
double *temp_select;
/*Get inverse calibration matrices*/
db_InvertCalibrationMatrix(K_inv,K);
db_InvertCalibrationMatrix(Kp_inv,Kp);
/*Compute scale coefficient*/
one_over_scale2=1.0/(scale*scale);
/*Initialize random seed*/
r_seed=12345;
/*Set pointers to pre-allocated space*/
hyp_cost_array=temp_d;
hyp_H_array=temp_d+nr_samples;
temp_select=temp_d+10*nr_samples;
x_h=temp_d+12*nr_samples;
xp_h=temp_d+12*nr_samples+3*nr_points;
x_i=temp_d+12*nr_samples+6*nr_points;
xp_i=temp_d+12*nr_samples+8*nr_points;
hyp_perm=temp_i;
point_perm=temp_i;
/*Prepare a randomly permuted subset of size
point_count from the input points*/
point_count=db_mini(nr_points,(int)(chunk_size*log((double)nr_samples)/DB_LN2));
point_count_new = point_count;
for(i=0;i<nr_points;i++) point_perm[i]=i;
for(last_point=nr_points-1,i=0;i<point_count;i++,last_point--)
{
pos=db_RandomInt(r_seed,last_point);
point_pos=point_perm[pos];
point_perm[pos]=point_perm[last_point];
/*Normalize image points with calibration
matrices and move them to x_h and xp_h*/
c=3*point_pos;
j=3*i;
x_h_point=x_h+j;
xp_h_point=xp_h+j;
db_Multiply3x3_3x1(x_h_point,K_inv,im+c);
db_Multiply3x3_3x1(xp_h_point,Kp_inv,im_p+c);
db_HomogenousNormalize3(x_h_point);
db_HomogenousNormalize3(xp_h_point);
/*Dehomogenize image points and move them
to x_i and xp_i*/
c=(i<<1);
db_DeHomogenizeImagePoint(x_i+c,x_h_point); // 2-dimension
db_DeHomogenizeImagePoint(xp_i+c,xp_h_point); //2-dimension
}
/*Generate Hypotheses*/
hyp_count=0;
switch(homography_type)
{
case DB_HOMOGRAPHY_TYPE_SIMILARITY:
case DB_HOMOGRAPHY_TYPE_SIMILARITY_U:
case DB_HOMOGRAPHY_TYPE_TRANSLATION:
case DB_HOMOGRAPHY_TYPE_ROTATION:
case DB_HOMOGRAPHY_TYPE_ROTATION_U:
case DB_HOMOGRAPHY_TYPE_SCALING:
case DB_HOMOGRAPHY_TYPE_S_T:
case DB_HOMOGRAPHY_TYPE_R_T:
case DB_HOMOGRAPHY_TYPE_R_S:
switch(homography_type)
{
case DB_HOMOGRAPHY_TYPE_SIMILARITY:
orientation_preserving=1;
allow_scaling=1;
allow_rotation=1;
allow_translation=1;
sample_size=2;
break;
case DB_HOMOGRAPHY_TYPE_SIMILARITY_U:
orientation_preserving=0;
allow_scaling=1;
allow_rotation=1;
allow_translation=1;
sample_size=3;
break;
case DB_HOMOGRAPHY_TYPE_TRANSLATION:
orientation_preserving=1;
allow_scaling=0;
allow_rotation=0;
allow_translation=1;
sample_size=1;
break;
case DB_HOMOGRAPHY_TYPE_ROTATION:
orientation_preserving=1;
allow_scaling=0;
allow_rotation=1;
allow_translation=0;
sample_size=1;
break;
case DB_HOMOGRAPHY_TYPE_ROTATION_U:
orientation_preserving=0;
allow_scaling=0;
allow_rotation=1;
allow_translation=0;
sample_size=2;
break;
case DB_HOMOGRAPHY_TYPE_SCALING:
orientation_preserving=1;
allow_scaling=1;
allow_rotation=0;
allow_translation=0;
sample_size=1;
break;
case DB_HOMOGRAPHY_TYPE_S_T:
orientation_preserving=1;
allow_scaling=1;
allow_rotation=0;
allow_translation=1;
sample_size=2;
break;
case DB_HOMOGRAPHY_TYPE_R_T:
orientation_preserving=1;
allow_scaling=0;
allow_rotation=1;
allow_translation=1;
sample_size=2;
break;
case DB_HOMOGRAPHY_TYPE_R_S:
orientation_preserving=1;
allow_scaling=1;
allow_rotation=0;
allow_translation=0;
sample_size=1;
break;
}
if(point_count>=sample_size) for(i=0;i<nr_samples;i++)
{
db_RandomSample(s,3,point_count,r_seed);
X[0]= &x_i[s[0]<<1];
X[1]= &x_i[s[1]<<1];
X[2]= &x_i[s[2]<<1];
Xp[0]= &xp_i[s[0]<<1];
Xp[1]= &xp_i[s[1]<<1];
Xp[2]= &xp_i[s[2]<<1];
db_StitchSimilarity2D(&hyp_H_array[9*hyp_count],Xp,X,sample_size,orientation_preserving,
allow_scaling,allow_rotation,allow_translation);
hyp_count++;
}
break;
case DB_HOMOGRAPHY_TYPE_CAMROTATION:
if(point_count>=2) for(i=0;i<nr_samples;i++)
{
db_RandomSample(s,2,point_count,r_seed);
db_StitchCameraRotation_2Points(&hyp_H_array[9*hyp_count],
&x_h[3*s[0]],&x_h[3*s[1]],
&xp_h[3*s[0]],&xp_h[3*s[1]]);
hyp_count++;
}
break;
case DB_HOMOGRAPHY_TYPE_CAMROTATION_F:
if(point_count>=3) for(i=0;i<nr_samples;i++)
{
db_RandomSample(s,3,point_count,r_seed);
hyp_count+=db_StitchRotationCommonFocalLength_3Points(&hyp_H_array[9*hyp_count],
&x_h[3*s[0]],&x_h[3*s[1]],&x_h[3*s[2]],
&xp_h[3*s[0]],&xp_h[3*s[1]],&xp_h[3*s[2]]);
}
break;
case DB_HOMOGRAPHY_TYPE_CAMROTATION_F_UD:
if(point_count>=3) for(i=0;i<nr_samples;i++)
{
db_RandomSample(s,3,point_count,r_seed);
hyp_count+=db_StitchRotationCommonFocalLength_3Points(&hyp_H_array[9*hyp_count],
&x_h[3*s[0]],&x_h[3*s[1]],&x_h[3*s[2]],
&xp_h[3*s[0]],&xp_h[3*s[1]],&xp_h[3*s[2]],NULL,0);
}
break;
case DB_HOMOGRAPHY_TYPE_AFFINE:
if(point_count>=3) for(i=0;i<nr_samples;i++)
{
db_RandomSample(s,3,point_count,r_seed);
db_StitchAffine2D_3Points(&hyp_H_array[9*hyp_count],
&x_h[3*s[0]],&x_h[3*s[1]],&x_h[3*s[2]],
&xp_h[3*s[0]],&xp_h[3*s[1]],&xp_h[3*s[2]]);
hyp_count++;
}
break;
case DB_HOMOGRAPHY_TYPE_PROJECTIVE:
default:
if(point_count>=4) for(i=0;i<nr_samples;i++)
{
db_RandomSample(s,4,point_count,r_seed);
db_StitchProjective2D_4Points(&hyp_H_array[9*hyp_count],
&x_h[3*s[0]],&x_h[3*s[1]],&x_h[3*s[2]],&x_h[3*s[3]],
&xp_h[3*s[0]],&xp_h[3*s[1]],&xp_h[3*s[2]],&xp_h[3*s[3]]);
hyp_count++;
}
}
if(hyp_count)
{
/*Count cost in chunks and decimate hypotheses
until only one remains or the correspondences are
exhausted*/
for(i=0;i<hyp_count;i++)
{
hyp_perm[i]=i;
hyp_cost_array[i]=0.0;
}
for(i=0,last_hyp=hyp_count-1;(last_hyp>0) && (i<point_count);i+=chunk_size)
{
/*Update cost with the next chunk*/
last_corr=db_mini(i+chunk_size-1,point_count-1);
for(j=0;j<=last_hyp;j++)
{
hyp_point=hyp_H_array+9*hyp_perm[j];
for(c=i;c<=last_corr;)
{
/*Take log of product of ten reprojection
errors to reduce nr of expensive log operations*/
if(c+9<=last_corr)
{
x_i_temp=x_i+(c<<1);
xp_i_temp=xp_i+(c<<1);
acc=db_ExpCauchyInhomogenousHomographyError(xp_i_temp,hyp_point,x_i_temp,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+2,hyp_point,x_i_temp+2,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+4,hyp_point,x_i_temp+4,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+6,hyp_point,x_i_temp+6,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+8,hyp_point,x_i_temp+8,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+10,hyp_point,x_i_temp+10,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+12,hyp_point,x_i_temp+12,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+14,hyp_point,x_i_temp+14,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+16,hyp_point,x_i_temp+16,one_over_scale2);
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i_temp+18,hyp_point,x_i_temp+18,one_over_scale2);
c+=10;
}
else
{
for(acc=1.0;c<=last_corr;c++)
{
acc*=db_ExpCauchyInhomogenousHomographyError(xp_i+(c<<1),hyp_point,x_i+(c<<1),one_over_scale2);
}
}
hyp_cost_array[j]+=log(acc);
}
}
if (chunk_size<point_count){
/*Prune out half of the hypotheses*/
new_last_hyp=(last_hyp+1)/2-1;
pivot=db_LeanQuickSelect(hyp_cost_array,last_hyp+1,new_last_hyp,temp_select);
for(j=0,c=0;(j<=last_hyp) && (c<=new_last_hyp);j++)
{
if(hyp_cost_array[j]<=pivot)
{
hyp_cost_array[c]=hyp_cost_array[j];
hyp_perm[c]=hyp_perm[j];
c++;
}
}
last_hyp=new_last_hyp;
}
}
/*Find the best hypothesis*/
lowest_cost=hyp_cost_array[0];
best_pos=0;
for(j=1;j<=last_hyp;j++)
{
if(hyp_cost_array[j]<lowest_cost)
{
lowest_cost=hyp_cost_array[j];
best_pos=j;
}
}
/*Move the best hypothesis*/
db_Copy9(H_temp,hyp_H_array+9*hyp_perm[best_pos]);
// outlier removal
if (outlierremoveflagE) // no polishment needed
{
point_count_new = db_RemoveOutliers_Homography(H_temp,x_i,xp_i,wp,im,im_p,im_r,im_raw,im_raw_p,point_count,one_over_scale2);
}
else
{
/*Polish*/
switch(homography_type)
{
case DB_HOMOGRAPHY_TYPE_SIMILARITY:
case DB_HOMOGRAPHY_TYPE_SIMILARITY_U:
case DB_HOMOGRAPHY_TYPE_TRANSLATION:
case DB_HOMOGRAPHY_TYPE_ROTATION:
case DB_HOMOGRAPHY_TYPE_ROTATION_U:
case DB_HOMOGRAPHY_TYPE_SCALING:
case DB_HOMOGRAPHY_TYPE_S_T:
case DB_HOMOGRAPHY_TYPE_R_T:
case DB_HOMOGRAPHY_TYPE_R_S:
case DB_HOMOGRAPHY_TYPE_AFFINE:
case DB_HOMOGRAPHY_TYPE_PROJECTIVE:
case DB_HOMOGRAPHY_TYPE_CAMROTATION_F:
case DB_HOMOGRAPHY_TYPE_CAMROTATION_F_UD:
db_RobCamRotation_Polish_Generic(H_temp,db_mini(point_count,max_points),homography_type,x_i,xp_i,one_over_scale2,max_iterations);
break;
case DB_HOMOGRAPHY_TYPE_CAMROTATION:
db_RobCamRotation_Polish(H_temp,db_mini(point_count,max_points),x_i,xp_i,one_over_scale2,max_iterations);
break;
}
}
}
else db_Identity3x3(H_temp);
switch(homography_type)
{
case DB_HOMOGRAPHY_TYPE_PROJECTIVE:
if(stat) stat->nr_parameters=8;
break;
case DB_HOMOGRAPHY_TYPE_AFFINE:
if(stat) stat->nr_parameters=6;
break;
case DB_HOMOGRAPHY_TYPE_SIMILARITY:
case DB_HOMOGRAPHY_TYPE_SIMILARITY_U:
case DB_HOMOGRAPHY_TYPE_CAMROTATION_F:
case DB_HOMOGRAPHY_TYPE_CAMROTATION_F_UD:
if(stat) stat->nr_parameters=4;
break;
case DB_HOMOGRAPHY_TYPE_CAMROTATION:
if(stat) stat->nr_parameters=3;
break;
case DB_HOMOGRAPHY_TYPE_TRANSLATION:
case DB_HOMOGRAPHY_TYPE_S_T:
case DB_HOMOGRAPHY_TYPE_R_T:
case DB_HOMOGRAPHY_TYPE_R_S:
if(stat) stat->nr_parameters=2;
break;
case DB_HOMOGRAPHY_TYPE_ROTATION:
case DB_HOMOGRAPHY_TYPE_ROTATION_U:
case DB_HOMOGRAPHY_TYPE_SCALING:
if(stat) stat->nr_parameters=1;
break;
}
db_RobImageHomography_Statistics(H_temp,db_mini(point_count,max_points),x_i,xp_i,one_over_scale2,stat);
/*Put on the calibration matrices*/
db_Multiply3x3_3x3(H_temp2,H_temp,K_inv);
db_Multiply3x3_3x3(H,Kp,H_temp2);
if (finalNumE)
*finalNumE = point_count_new;
}