00001 #include <iostream>
00002 using namespace std;
00003
00004 #include "epnp.h"
00005
00006 epnp::epnp(void)
00007 {
00008 maximum_number_of_correspondences = 0;
00009 number_of_correspondences = 0;
00010
00011 pws = 0;
00012 us = 0;
00013 alphas = 0;
00014 pcs = 0;
00015 }
00016
00017 epnp::~epnp()
00018 {
00019 delete [] pws;
00020 delete [] us;
00021 delete [] alphas;
00022 delete [] pcs;
00023 }
00024
00025 void epnp::set_internal_parameters(double uc, double vc, double fu, double fv)
00026 {
00027 this->uc = uc;
00028 this->vc = vc;
00029 this->fu = fu;
00030 this->fv = fv;
00031 }
00032
00033 void epnp::set_maximum_number_of_correspondences(int n)
00034 {
00035 if (maximum_number_of_correspondences < n) {
00036 if (pws != 0) delete [] pws;
00037 if (us != 0) delete [] us;
00038 if (alphas != 0) delete [] alphas;
00039 if (pcs != 0) delete [] pcs;
00040
00041 maximum_number_of_correspondences = n;
00042 pws = new double[3 * maximum_number_of_correspondences];
00043 us = new double[2 * maximum_number_of_correspondences];
00044 alphas = new double[4 * maximum_number_of_correspondences];
00045 pcs = new double[3 * maximum_number_of_correspondences];
00046 }
00047 }
00048
00049 void epnp::reset_correspondences(void)
00050 {
00051 number_of_correspondences = 0;
00052 }
00053
00054 void epnp::add_correspondence(double X, double Y, double Z, double u, double v)
00055 {
00056 pws[3 * number_of_correspondences ] = X;
00057 pws[3 * number_of_correspondences + 1] = Y;
00058 pws[3 * number_of_correspondences + 2] = Z;
00059
00060 us[2 * number_of_correspondences ] = u;
00061 us[2 * number_of_correspondences + 1] = v;
00062
00063 number_of_correspondences++;
00064 }
00065
00066 void epnp::choose_control_points(void)
00067 {
00068
00069 cws[0][0] = cws[0][1] = cws[0][2] = 0;
00070 for(int i = 0; i < number_of_correspondences; i++)
00071 for(int j = 0; j < 3; j++)
00072 cws[0][j] += pws[3 * i + j];
00073
00074 for(int j = 0; j < 3; j++)
00075 cws[0][j] /= number_of_correspondences;
00076
00077
00078
00079 CvMat * PW0 = cvCreateMat(number_of_correspondences, 3, CV_64F);
00080
00081 double pw0tpw0[3 * 3], dc[3], uct[3 * 3];
00082 CvMat PW0tPW0 = cvMat(3, 3, CV_64F, pw0tpw0);
00083 CvMat DC = cvMat(3, 1, CV_64F, dc);
00084 CvMat UCt = cvMat(3, 3, CV_64F, uct);
00085
00086 for(int i = 0; i < number_of_correspondences; i++)
00087 for(int j = 0; j < 3; j++)
00088 PW0->data.db[3 * i + j] = pws[3 * i + j] - cws[0][j];
00089
00090 cvMulTransposed(PW0, &PW0tPW0, 1);
00091 cvSVD(&PW0tPW0, &DC, &UCt, 0, CV_SVD_MODIFY_A | CV_SVD_U_T);
00092
00093 cvReleaseMat(&PW0);
00094
00095 for(int i = 1; i < 4; i++) {
00096 double k = sqrt(dc[i - 1] / number_of_correspondences);
00097 for(int j = 0; j < 3; j++)
00098 cws[i][j] = cws[0][j] + k * uct[3 * (i - 1) + j];
00099 }
00100 }
00101
00102 void epnp::compute_barycentric_coordinates(void)
00103 {
00104 double cc[3 * 3], cc_inv[3 * 3];
00105 CvMat CC = cvMat(3, 3, CV_64F, cc);
00106 CvMat CC_inv = cvMat(3, 3, CV_64F, cc_inv);
00107
00108 for(int i = 0; i < 3; i++)
00109 for(int j = 1; j < 4; j++)
00110 cc[3 * i + j - 1] = cws[j][i] - cws[0][i];
00111
00112 cvInvert(&CC, &CC_inv, CV_SVD);
00113 double * ci = cc_inv;
00114 for(int i = 0; i < number_of_correspondences; i++) {
00115 double * pi = pws + 3 * i;
00116 double * a = alphas + 4 * i;
00117
00118 for(int j = 0; j < 3; j++)
00119 a[1 + j] =
00120 ci[3 * j ] * (pi[0] - cws[0][0]) +
00121 ci[3 * j + 1] * (pi[1] - cws[0][1]) +
00122 ci[3 * j + 2] * (pi[2] - cws[0][2]);
00123 a[0] = 1.0f - a[1] - a[2] - a[3];
00124 }
00125 }
00126
00127 void epnp::fill_M(CvMat * M,
00128 const int row, const double * as, const double u, const double v)
00129 {
00130 double * M1 = M->data.db + row * 12;
00131 double * M2 = M1 + 12;
00132
00133 for(int i = 0; i < 4; i++) {
00134 M1[3 * i ] = as[i] * fu;
00135 M1[3 * i + 1] = 0.0;
00136 M1[3 * i + 2] = as[i] * (uc - u);
00137
00138 M2[3 * i ] = 0.0;
00139 M2[3 * i + 1] = as[i] * fv;
00140 M2[3 * i + 2] = as[i] * (vc - v);
00141 }
00142 }
00143
00144 void epnp::compute_ccs(const double * betas, const double * ut)
00145 {
00146 for(int i = 0; i < 4; i++)
00147 ccs[i][0] = ccs[i][1] = ccs[i][2] = 0.0f;
00148
00149 for(int i = 0; i < 4; i++) {
00150 const double * v = ut + 12 * (11 - i);
00151 for(int j = 0; j < 4; j++)
00152 for(int k = 0; k < 3; k++)
00153 ccs[j][k] += betas[i] * v[3 * j + k];
00154 }
00155 }
00156
00157 void epnp::compute_pcs(void)
00158 {
00159 for(int i = 0; i < number_of_correspondences; i++) {
00160 double * a = alphas + 4 * i;
00161 double * pc = pcs + 3 * i;
00162
00163 for(int j = 0; j < 3; j++)
00164 pc[j] = a[0] * ccs[0][j] + a[1] * ccs[1][j] + a[2] * ccs[2][j] + a[3] * ccs[3][j];
00165 }
00166 }
00167
00168 double epnp::compute_pose(double R[3][3], double t[3])
00169 {
00170 choose_control_points();
00171 compute_barycentric_coordinates();
00172
00173 CvMat * M = cvCreateMat(2 * number_of_correspondences, 12, CV_64F);
00174
00175 for(int i = 0; i < number_of_correspondences; i++)
00176 fill_M(M, 2 * i, alphas + 4 * i, us[2 * i], us[2 * i + 1]);
00177
00178 double mtm[12 * 12], d[12], ut[12 * 12];
00179 CvMat MtM = cvMat(12, 12, CV_64F, mtm);
00180 CvMat D = cvMat(12, 1, CV_64F, d);
00181 CvMat Ut = cvMat(12, 12, CV_64F, ut);
00182
00183 cvMulTransposed(M, &MtM, 1);
00184 cvSVD(&MtM, &D, &Ut, 0, CV_SVD_MODIFY_A | CV_SVD_U_T);
00185 cvReleaseMat(&M);
00186
00187 double l_6x10[6 * 10], rho[6];
00188 CvMat L_6x10 = cvMat(6, 10, CV_64F, l_6x10);
00189 CvMat Rho = cvMat(6, 1, CV_64F, rho);
00190
00191 compute_L_6x10(ut, l_6x10);
00192 compute_rho(rho);
00193
00194 double Betas[4][4], rep_errors[4];
00195 double Rs[4][3][3], ts[4][3];
00196
00197 find_betas_approx_1(&L_6x10, &Rho, Betas[1]);
00198 gauss_newton(&L_6x10, &Rho, Betas[1]);
00199 rep_errors[1] = compute_R_and_t(ut, Betas[1], Rs[1], ts[1]);
00200
00201 find_betas_approx_2(&L_6x10, &Rho, Betas[2]);
00202 gauss_newton(&L_6x10, &Rho, Betas[2]);
00203 rep_errors[2] = compute_R_and_t(ut, Betas[2], Rs[2], ts[2]);
00204
00205 find_betas_approx_3(&L_6x10, &Rho, Betas[3]);
00206 gauss_newton(&L_6x10, &Rho, Betas[3]);
00207 rep_errors[3] = compute_R_and_t(ut, Betas[3], Rs[3], ts[3]);
00208
00209 int N = 1;
00210 if (rep_errors[2] < rep_errors[1]) N = 2;
00211 if (rep_errors[3] < rep_errors[N]) N = 3;
00212
00213 copy_R_and_t(Rs[N], ts[N], R, t);
00214
00215 return rep_errors[N];
00216 }
00217
00218 void epnp::copy_R_and_t(const double R_src[3][3], const double t_src[3],
00219 double R_dst[3][3], double t_dst[3])
00220 {
00221 for(int i = 0; i < 3; i++) {
00222 for(int j = 0; j < 3; j++)
00223 R_dst[i][j] = R_src[i][j];
00224 t_dst[i] = t_src[i];
00225 }
00226 }
00227
00228 double epnp::dist2(const double * p1, const double * p2)
00229 {
00230 return
00231 (p1[0] - p2[0]) * (p1[0] - p2[0]) +
00232 (p1[1] - p2[1]) * (p1[1] - p2[1]) +
00233 (p1[2] - p2[2]) * (p1[2] - p2[2]);
00234 }
00235
00236 double epnp::dot(const double * v1, const double * v2)
00237 {
00238 return v1[0] * v2[0] + v1[1] * v2[1] + v1[2] * v2[2];
00239 }
00240
00241 double epnp::reprojection_error(const double R[3][3], const double t[3])
00242 {
00243 double sum2 = 0.0;
00244
00245 for(int i = 0; i < number_of_correspondences; i++) {
00246 double * pw = pws + 3 * i;
00247 double Xc = dot(R[0], pw) + t[0];
00248 double Yc = dot(R[1], pw) + t[1];
00249 double inv_Zc = 1.0 / (dot(R[2], pw) + t[2]);
00250 double ue = uc + fu * Xc * inv_Zc;
00251 double ve = vc + fv * Yc * inv_Zc;
00252 double u = us[2 * i], v = us[2 * i + 1];
00253
00254 sum2 += sqrt( (u - ue) * (u - ue) + (v - ve) * (v - ve) );
00255 }
00256
00257 return sum2 / number_of_correspondences;
00258 }
00259
00260 void epnp::estimate_R_and_t(double R[3][3], double t[3])
00261 {
00262 double pc0[3], pw0[3];
00263
00264 pc0[0] = pc0[1] = pc0[2] = 0.0;
00265 pw0[0] = pw0[1] = pw0[2] = 0.0;
00266
00267 for(int i = 0; i < number_of_correspondences; i++) {
00268 const double * pc = pcs + 3 * i;
00269 const double * pw = pws + 3 * i;
00270
00271 for(int j = 0; j < 3; j++) {
00272 pc0[j] += pc[j];
00273 pw0[j] += pw[j];
00274 }
00275 }
00276 for(int j = 0; j < 3; j++) {
00277 pc0[j] /= number_of_correspondences;
00278 pw0[j] /= number_of_correspondences;
00279 }
00280
00281 double abt[3 * 3], abt_d[3], abt_u[3 * 3], abt_v[3 * 3];
00282 CvMat ABt = cvMat(3, 3, CV_64F, abt);
00283 CvMat ABt_D = cvMat(3, 1, CV_64F, abt_d);
00284 CvMat ABt_U = cvMat(3, 3, CV_64F, abt_u);
00285 CvMat ABt_V = cvMat(3, 3, CV_64F, abt_v);
00286
00287 cvSetZero(&ABt);
00288 for(int i = 0; i < number_of_correspondences; i++) {
00289 double * pc = pcs + 3 * i;
00290 double * pw = pws + 3 * i;
00291
00292 for(int j = 0; j < 3; j++) {
00293 abt[3 * j ] += (pc[j] - pc0[j]) * (pw[0] - pw0[0]);
00294 abt[3 * j + 1] += (pc[j] - pc0[j]) * (pw[1] - pw0[1]);
00295 abt[3 * j + 2] += (pc[j] - pc0[j]) * (pw[2] - pw0[2]);
00296 }
00297 }
00298
00299 cvSVD(&ABt, &ABt_D, &ABt_U, &ABt_V, CV_SVD_MODIFY_A);
00300
00301 for(int i = 0; i < 3; i++)
00302 for(int j = 0; j < 3; j++)
00303 R[i][j] = dot(abt_u + 3 * i, abt_v + 3 * j);
00304
00305 const double det =
00306 R[0][0] * R[1][1] * R[2][2] + R[0][1] * R[1][2] * R[2][0] + R[0][2] * R[1][0] * R[2][1] -
00307 R[0][2] * R[1][1] * R[2][0] - R[0][1] * R[1][0] * R[2][2] - R[0][0] * R[1][2] * R[2][1];
00308
00309 if (det < 0) {
00310 R[2][0] = -R[2][0];
00311 R[2][1] = -R[2][1];
00312 R[2][2] = -R[2][2];
00313 }
00314
00315 t[0] = pc0[0] - dot(R[0], pw0);
00316 t[1] = pc0[1] - dot(R[1], pw0);
00317 t[2] = pc0[2] - dot(R[2], pw0);
00318 }
00319
00320 void epnp::print_pose(const double R[3][3], const double t[3])
00321 {
00322 cout << R[0][0] << " " << R[0][1] << " " << R[0][2] << " " << t[0] << endl;
00323 cout << R[1][0] << " " << R[1][1] << " " << R[1][2] << " " << t[1] << endl;
00324 cout << R[2][0] << " " << R[2][1] << " " << R[2][2] << " " << t[2] << endl;
00325 }
00326
00327 void epnp::solve_for_sign(void)
00328 {
00329 if (pcs[2] < 0.0) {
00330 for(int i = 0; i < 4; i++)
00331 for(int j = 0; j < 3; j++)
00332 ccs[i][j] = -ccs[i][j];
00333
00334 for(int i = 0; i < number_of_correspondences; i++) {
00335 pcs[3 * i ] = -pcs[3 * i];
00336 pcs[3 * i + 1] = -pcs[3 * i + 1];
00337 pcs[3 * i + 2] = -pcs[3 * i + 2];
00338 }
00339 }
00340 }
00341
00342 double epnp::compute_R_and_t(const double * ut, const double * betas,
00343 double R[3][3], double t[3])
00344 {
00345 compute_ccs(betas, ut);
00346 compute_pcs();
00347
00348 solve_for_sign();
00349
00350 estimate_R_and_t(R, t);
00351
00352 return reprojection_error(R, t);
00353 }
00354
00355
00356
00357
00358 void epnp::find_betas_approx_1(const CvMat * L_6x10, const CvMat * Rho,
00359 double * betas)
00360 {
00361 double l_6x4[6 * 4], b4[4];
00362 CvMat L_6x4 = cvMat(6, 4, CV_64F, l_6x4);
00363 CvMat B4 = cvMat(4, 1, CV_64F, b4);
00364
00365 for(int i = 0; i < 6; i++) {
00366 cvmSet(&L_6x4, i, 0, cvmGet(L_6x10, i, 0));
00367 cvmSet(&L_6x4, i, 1, cvmGet(L_6x10, i, 1));
00368 cvmSet(&L_6x4, i, 2, cvmGet(L_6x10, i, 3));
00369 cvmSet(&L_6x4, i, 3, cvmGet(L_6x10, i, 6));
00370 }
00371
00372 cvSolve(&L_6x4, Rho, &B4, CV_SVD);
00373
00374 if (b4[0] < 0) {
00375 betas[0] = sqrt(-b4[0]);
00376 betas[1] = -b4[1] / betas[0];
00377 betas[2] = -b4[2] / betas[0];
00378 betas[3] = -b4[3] / betas[0];
00379 } else {
00380 betas[0] = sqrt(b4[0]);
00381 betas[1] = b4[1] / betas[0];
00382 betas[2] = b4[2] / betas[0];
00383 betas[3] = b4[3] / betas[0];
00384 }
00385 }
00386
00387
00388
00389
00390 void epnp::find_betas_approx_2(const CvMat * L_6x10, const CvMat * Rho,
00391 double * betas)
00392 {
00393 double l_6x3[6 * 3], b3[3];
00394 CvMat L_6x3 = cvMat(6, 3, CV_64F, l_6x3);
00395 CvMat B3 = cvMat(3, 1, CV_64F, b3);
00396
00397 for(int i = 0; i < 6; i++) {
00398 cvmSet(&L_6x3, i, 0, cvmGet(L_6x10, i, 0));
00399 cvmSet(&L_6x3, i, 1, cvmGet(L_6x10, i, 1));
00400 cvmSet(&L_6x3, i, 2, cvmGet(L_6x10, i, 2));
00401 }
00402
00403 cvSolve(&L_6x3, Rho, &B3, CV_SVD);
00404
00405 if (b3[0] < 0) {
00406 betas[0] = sqrt(-b3[0]);
00407 betas[1] = (b3[2] < 0) ? sqrt(-b3[2]) : 0.0;
00408 } else {
00409 betas[0] = sqrt(b3[0]);
00410 betas[1] = (b3[2] > 0) ? sqrt(b3[2]) : 0.0;
00411 }
00412
00413 if (b3[1] < 0) betas[0] = -betas[0];
00414
00415 betas[2] = 0.0;
00416 betas[3] = 0.0;
00417 }
00418
00419
00420
00421
00422 void epnp::find_betas_approx_3(const CvMat * L_6x10, const CvMat * Rho,
00423 double * betas)
00424 {
00425 double l_6x5[6 * 5], b5[5];
00426 CvMat L_6x5 = cvMat(6, 5, CV_64F, l_6x5);
00427 CvMat B5 = cvMat(5, 1, CV_64F, b5);
00428
00429 for(int i = 0; i < 6; i++) {
00430 cvmSet(&L_6x5, i, 0, cvmGet(L_6x10, i, 0));
00431 cvmSet(&L_6x5, i, 1, cvmGet(L_6x10, i, 1));
00432 cvmSet(&L_6x5, i, 2, cvmGet(L_6x10, i, 2));
00433 cvmSet(&L_6x5, i, 3, cvmGet(L_6x10, i, 3));
00434 cvmSet(&L_6x5, i, 4, cvmGet(L_6x10, i, 4));
00435 }
00436
00437 cvSolve(&L_6x5, Rho, &B5, CV_SVD);
00438
00439 if (b5[0] < 0) {
00440 betas[0] = sqrt(-b5[0]);
00441 betas[1] = (b5[2] < 0) ? sqrt(-b5[2]) : 0.0;
00442 } else {
00443 betas[0] = sqrt(b5[0]);
00444 betas[1] = (b5[2] > 0) ? sqrt(b5[2]) : 0.0;
00445 }
00446 if (b5[1] < 0) betas[0] = -betas[0];
00447 betas[2] = b5[3] / betas[0];
00448 betas[3] = 0.0;
00449 }
00450
00451 void epnp::compute_L_6x10(const double * ut, double * l_6x10)
00452 {
00453 const double * v[4];
00454
00455 v[0] = ut + 12 * 11;
00456 v[1] = ut + 12 * 10;
00457 v[2] = ut + 12 * 9;
00458 v[3] = ut + 12 * 8;
00459
00460 double dv[4][6][3];
00461
00462 for(int i = 0; i < 4; i++) {
00463 int a = 0, b = 1;
00464 for(int j = 0; j < 6; j++) {
00465 dv[i][j][0] = v[i][3 * a ] - v[i][3 * b];
00466 dv[i][j][1] = v[i][3 * a + 1] - v[i][3 * b + 1];
00467 dv[i][j][2] = v[i][3 * a + 2] - v[i][3 * b + 2];
00468
00469 b++;
00470 if (b > 3) {
00471 a++;
00472 b = a + 1;
00473 }
00474 }
00475 }
00476
00477 for(int i = 0; i < 6; i++) {
00478 double * row = l_6x10 + 10 * i;
00479
00480 row[0] = dot(dv[0][i], dv[0][i]);
00481 row[1] = 2.0f * dot(dv[0][i], dv[1][i]);
00482 row[2] = dot(dv[1][i], dv[1][i]);
00483 row[3] = 2.0f * dot(dv[0][i], dv[2][i]);
00484 row[4] = 2.0f * dot(dv[1][i], dv[2][i]);
00485 row[5] = dot(dv[2][i], dv[2][i]);
00486 row[6] = 2.0f * dot(dv[0][i], dv[3][i]);
00487 row[7] = 2.0f * dot(dv[1][i], dv[3][i]);
00488 row[8] = 2.0f * dot(dv[2][i], dv[3][i]);
00489 row[9] = dot(dv[3][i], dv[3][i]);
00490 }
00491 }
00492
00493 void epnp::compute_rho(double * rho)
00494 {
00495 rho[0] = dist2(cws[0], cws[1]);
00496 rho[1] = dist2(cws[0], cws[2]);
00497 rho[2] = dist2(cws[0], cws[3]);
00498 rho[3] = dist2(cws[1], cws[2]);
00499 rho[4] = dist2(cws[1], cws[3]);
00500 rho[5] = dist2(cws[2], cws[3]);
00501 }
00502
00503 void epnp::compute_A_and_b_gauss_newton(const double * l_6x10, const double * rho,
00504 double betas[4], CvMat * A, CvMat * b)
00505 {
00506 for(int i = 0; i < 6; i++) {
00507 const double * rowL = l_6x10 + i * 10;
00508 double * rowA = A->data.db + i * 4;
00509
00510 rowA[0] = 2 * rowL[0] * betas[0] + rowL[1] * betas[1] + rowL[3] * betas[2] + rowL[6] * betas[3];
00511 rowA[1] = rowL[1] * betas[0] + 2 * rowL[2] * betas[1] + rowL[4] * betas[2] + rowL[7] * betas[3];
00512 rowA[2] = rowL[3] * betas[0] + rowL[4] * betas[1] + 2 * rowL[5] * betas[2] + rowL[8] * betas[3];
00513 rowA[3] = rowL[6] * betas[0] + rowL[7] * betas[1] + rowL[8] * betas[2] + 2 * rowL[9] * betas[3];
00514
00515 cvmSet(b, i, 0, rho[i] -
00516 (
00517 rowL[0] * betas[0] * betas[0] +
00518 rowL[1] * betas[0] * betas[1] +
00519 rowL[2] * betas[1] * betas[1] +
00520 rowL[3] * betas[0] * betas[2] +
00521 rowL[4] * betas[1] * betas[2] +
00522 rowL[5] * betas[2] * betas[2] +
00523 rowL[6] * betas[0] * betas[3] +
00524 rowL[7] * betas[1] * betas[3] +
00525 rowL[8] * betas[2] * betas[3] +
00526 rowL[9] * betas[3] * betas[3]
00527 ));
00528 }
00529 }
00530
00531 void epnp::gauss_newton(const CvMat * L_6x10, const CvMat * Rho,
00532 double betas[4])
00533 {
00534 const int iterations_number = 5;
00535
00536 double a[6*4], b[6], x[4];
00537 CvMat A = cvMat(6, 4, CV_64F, a);
00538 CvMat B = cvMat(6, 1, CV_64F, b);
00539 CvMat X = cvMat(4, 1, CV_64F, x);
00540
00541 for(int k = 0; k < iterations_number; k++) {
00542 compute_A_and_b_gauss_newton(L_6x10->data.db, Rho->data.db,
00543 betas, &A, &B);
00544 qr_solve(&A, &B, &X);
00545
00546 for(int i = 0; i < 4; i++)
00547 betas[i] += x[i];
00548 }
00549 }
00550
00551 void epnp::qr_solve(CvMat * A, CvMat * b, CvMat * X)
00552 {
00553 static int max_nr = 0;
00554 static double * A1, * A2;
00555
00556 const int nr = A->rows;
00557 const int nc = A->cols;
00558
00559 if (max_nr != 0 && max_nr < nr) {
00560 delete [] A1;
00561 delete [] A2;
00562 }
00563 if (max_nr < nr) {
00564 max_nr = nr;
00565 A1 = new double[nr];
00566 A2 = new double[nr];
00567 }
00568
00569 double * pA = A->data.db, * ppAkk = pA;
00570 for(int k = 0; k < nc; k++) {
00571 double * ppAik = ppAkk, eta = fabs(*ppAik);
00572 for(int i = k + 1; i < nr; i++) {
00573 double elt = fabs(*ppAik);
00574 if (eta < elt) eta = elt;
00575 ppAik += nc;
00576 }
00577
00578 if (eta == 0) {
00579 A1[k] = A2[k] = 0.0;
00580 cerr << "God damnit, A is singular, this shouldn't happen." << endl;
00581 return;
00582 } else {
00583 double * ppAik = ppAkk, sum = 0.0, inv_eta = 1. / eta;
00584 for(int i = k; i < nr; i++) {
00585 *ppAik *= inv_eta;
00586 sum += *ppAik * *ppAik;
00587 ppAik += nc;
00588 }
00589 double sigma = sqrt(sum);
00590 if (*ppAkk < 0)
00591 sigma = -sigma;
00592 *ppAkk += sigma;
00593 A1[k] = sigma * *ppAkk;
00594 A2[k] = -eta * sigma;
00595 for(int j = k + 1; j < nc; j++) {
00596 double * ppAik = ppAkk, sum = 0;
00597 for(int i = k; i < nr; i++) {
00598 sum += *ppAik * ppAik[j - k];
00599 ppAik += nc;
00600 }
00601 double tau = sum / A1[k];
00602 ppAik = ppAkk;
00603 for(int i = k; i < nr; i++) {
00604 ppAik[j - k] -= tau * *ppAik;
00605 ppAik += nc;
00606 }
00607 }
00608 }
00609 ppAkk += nc + 1;
00610 }
00611
00612
00613 double * ppAjj = pA, * pb = b->data.db;
00614 for(int j = 0; j < nc; j++) {
00615 double * ppAij = ppAjj, tau = 0;
00616 for(int i = j; i < nr; i++) {
00617 tau += *ppAij * pb[i];
00618 ppAij += nc;
00619 }
00620 tau /= A1[j];
00621 ppAij = ppAjj;
00622 for(int i = j; i < nr; i++) {
00623 pb[i] -= tau * *ppAij;
00624 ppAij += nc;
00625 }
00626 ppAjj += nc + 1;
00627 }
00628
00629
00630 double * pX = X->data.db;
00631 pX[nc - 1] = pb[nc - 1] / A2[nc - 1];
00632 for(int i = nc - 2; i >= 0; i--) {
00633 double * ppAij = pA + i * nc + (i + 1), sum = 0;
00634
00635 for(int j = i + 1; j < nc; j++) {
00636 sum += *ppAij * pX[j];
00637 ppAij++;
00638 }
00639 pX[i] = (pb[i] - sum) / A2[i];
00640 }
00641 }
00642
00643
00644
00645 void epnp::relative_error(double & rot_err, double & transl_err,
00646 const double Rtrue[3][3], const double ttrue[3],
00647 const double Rest[3][3], const double test[3])
00648 {
00649 double qtrue[4], qest[4];
00650
00651 mat_to_quat(Rtrue, qtrue);
00652 mat_to_quat(Rest, qest);
00653
00654 double rot_err1 = sqrt((qtrue[0] - qest[0]) * (qtrue[0] - qest[0]) +
00655 (qtrue[1] - qest[1]) * (qtrue[1] - qest[1]) +
00656 (qtrue[2] - qest[2]) * (qtrue[2] - qest[2]) +
00657 (qtrue[3] - qest[3]) * (qtrue[3] - qest[3]) ) /
00658 sqrt(qtrue[0] * qtrue[0] + qtrue[1] * qtrue[1] + qtrue[2] * qtrue[2] + qtrue[3] * qtrue[3]);
00659
00660 double rot_err2 = sqrt((qtrue[0] + qest[0]) * (qtrue[0] + qest[0]) +
00661 (qtrue[1] + qest[1]) * (qtrue[1] + qest[1]) +
00662 (qtrue[2] + qest[2]) * (qtrue[2] + qest[2]) +
00663 (qtrue[3] + qest[3]) * (qtrue[3] + qest[3]) ) /
00664 sqrt(qtrue[0] * qtrue[0] + qtrue[1] * qtrue[1] + qtrue[2] * qtrue[2] + qtrue[3] * qtrue[3]);
00665
00666 rot_err = min(rot_err1, rot_err2);
00667
00668 transl_err =
00669 sqrt((ttrue[0] - test[0]) * (ttrue[0] - test[0]) +
00670 (ttrue[1] - test[1]) * (ttrue[1] - test[1]) +
00671 (ttrue[2] - test[2]) * (ttrue[2] - test[2])) /
00672 sqrt(ttrue[0] * ttrue[0] + ttrue[1] * ttrue[1] + ttrue[2] * ttrue[2]);
00673 }
00674
00675 void epnp::mat_to_quat(const double R[3][3], double q[4])
00676 {
00677 double tr = R[0][0] + R[1][1] + R[2][2];
00678 double n4;
00679
00680 if (tr > 0.0f) {
00681 q[0] = R[1][2] - R[2][1];
00682 q[1] = R[2][0] - R[0][2];
00683 q[2] = R[0][1] - R[1][0];
00684 q[3] = tr + 1.0f;
00685 n4 = q[3];
00686 } else if ( (R[0][0] > R[1][1]) && (R[0][0] > R[2][2]) ) {
00687 q[0] = 1.0f + R[0][0] - R[1][1] - R[2][2];
00688 q[1] = R[1][0] + R[0][1];
00689 q[2] = R[2][0] + R[0][2];
00690 q[3] = R[1][2] - R[2][1];
00691 n4 = q[0];
00692 } else if (R[1][1] > R[2][2]) {
00693 q[0] = R[1][0] + R[0][1];
00694 q[1] = 1.0f + R[1][1] - R[0][0] - R[2][2];
00695 q[2] = R[2][1] + R[1][2];
00696 q[3] = R[2][0] - R[0][2];
00697 n4 = q[1];
00698 } else {
00699 q[0] = R[2][0] + R[0][2];
00700 q[1] = R[2][1] + R[1][2];
00701 q[2] = 1.0f + R[2][2] - R[0][0] - R[1][1];
00702 q[3] = R[0][1] - R[1][0];
00703 n4 = q[2];
00704 }
00705 double scale = 0.5f / double(sqrt(n4));
00706
00707 q[0] *= scale;
00708 q[1] *= scale;
00709 q[2] *= scale;
00710 q[3] *= scale;
00711 }