#include #include "precomp.hpp" #include "epnp.h" namespace cv { epnp::epnp(const Mat& cameraMatrix, const Mat& opoints, const Mat& ipoints) { if (cameraMatrix.depth() == CV_32F) init_camera_parameters(cameraMatrix); else init_camera_parameters(cameraMatrix); number_of_correspondences = std::max(opoints.checkVector(3, CV_32F), opoints.checkVector(3, CV_64F)); pws.resize(3 * number_of_correspondences); us.resize(2 * number_of_correspondences); if (opoints.depth() == ipoints.depth()) { if (opoints.depth() == CV_32F) init_points(opoints, ipoints); else init_points(opoints, ipoints); } else if (opoints.depth() == CV_32F) init_points(opoints, ipoints); else init_points(opoints, ipoints); alphas.resize(4 * number_of_correspondences); pcs.resize(3 * number_of_correspondences); max_nr = 0; A1 = NULL; A2 = NULL; } epnp::~epnp() { if (A1) delete[] A1; if (A2) delete[] A2; } void epnp::choose_control_points(void) { // Take C0 as the reference points centroid: cws[0][0] = cws[0][1] = cws[0][2] = 0; for(int i = 0; i < number_of_correspondences; i++) for(int j = 0; j < 3; j++) cws[0][j] += pws[3 * i + j]; for(int j = 0; j < 3; j++) cws[0][j] /= number_of_correspondences; // Take C1, C2, and C3 from PCA on the reference points: CvMat * PW0 = cvCreateMat(number_of_correspondences, 3, CV_64F); double pw0tpw0[3 * 3], dc[3], uct[3 * 3]; CvMat PW0tPW0 = cvMat(3, 3, CV_64F, pw0tpw0); CvMat DC = cvMat(3, 1, CV_64F, dc); CvMat UCt = cvMat(3, 3, CV_64F, uct); for(int i = 0; i < number_of_correspondences; i++) for(int j = 0; j < 3; j++) PW0->data.db[3 * i + j] = pws[3 * i + j] - cws[0][j]; cvMulTransposed(PW0, &PW0tPW0, 1); cvSVD(&PW0tPW0, &DC, &UCt, 0, CV_SVD_MODIFY_A | CV_SVD_U_T); cvReleaseMat(&PW0); for(int i = 1; i < 4; i++) { double k = sqrt(dc[i - 1] / number_of_correspondences); for(int j = 0; j < 3; j++) cws[i][j] = cws[0][j] + k * uct[3 * (i - 1) + j]; } } void epnp::compute_barycentric_coordinates(void) { double cc[3 * 3], cc_inv[3 * 3]; CvMat CC = cvMat(3, 3, CV_64F, cc); CvMat CC_inv = cvMat(3, 3, CV_64F, cc_inv); for(int i = 0; i < 3; i++) for(int j = 1; j < 4; j++) cc[3 * i + j - 1] = cws[j][i] - cws[0][i]; cvInvert(&CC, &CC_inv, CV_SVD); double * ci = cc_inv; for(int i = 0; i < number_of_correspondences; i++) { double * pi = &pws[0] + 3 * i; double * a = &alphas[0] + 4 * i; for(int j = 0; j < 3; j++) a[1 + j] = ci[3 * j ] * (pi[0] - cws[0][0]) + ci[3 * j + 1] * (pi[1] - cws[0][1]) + ci[3 * j + 2] * (pi[2] - cws[0][2]); a[0] = 1.0f - a[1] - a[2] - a[3]; } } void epnp::fill_M(CvMat * M, const int row, const double * as, const double u, const double v) { double * M1 = M->data.db + row * 12; double * M2 = M1 + 12; for(int i = 0; i < 4; i++) { M1[3 * i ] = as[i] * fu; M1[3 * i + 1] = 0.0; M1[3 * i + 2] = as[i] * (uc - u); M2[3 * i ] = 0.0; M2[3 * i + 1] = as[i] * fv; M2[3 * i + 2] = as[i] * (vc - v); } } void epnp::compute_ccs(const double * betas, const double * ut) { for(int i = 0; i < 4; i++) ccs[i][0] = ccs[i][1] = ccs[i][2] = 0.0f; for(int i = 0; i < 4; i++) { const double * v = ut + 12 * (11 - i); for(int j = 0; j < 4; j++) for(int k = 0; k < 3; k++) ccs[j][k] += betas[i] * v[3 * j + k]; } } void epnp::compute_pcs(void) { for(int i = 0; i < number_of_correspondences; i++) { double * a = &alphas[0] + 4 * i; double * pc = &pcs[0] + 3 * i; for(int j = 0; j < 3; j++) pc[j] = a[0] * ccs[0][j] + a[1] * ccs[1][j] + a[2] * ccs[2][j] + a[3] * ccs[3][j]; } } void epnp::compute_pose(Mat& R, Mat& t) { choose_control_points(); compute_barycentric_coordinates(); CvMat * M = cvCreateMat(2 * number_of_correspondences, 12, CV_64F); for(int i = 0; i < number_of_correspondences; i++) fill_M(M, 2 * i, &alphas[0] + 4 * i, us[2 * i], us[2 * i + 1]); double mtm[12 * 12], d[12], ut[12 * 12]; CvMat MtM = cvMat(12, 12, CV_64F, mtm); CvMat D = cvMat(12, 1, CV_64F, d); CvMat Ut = cvMat(12, 12, CV_64F, ut); cvMulTransposed(M, &MtM, 1); cvSVD(&MtM, &D, &Ut, 0, CV_SVD_MODIFY_A | CV_SVD_U_T); cvReleaseMat(&M); double l_6x10[6 * 10], rho[6]; CvMat L_6x10 = cvMat(6, 10, CV_64F, l_6x10); CvMat Rho = cvMat(6, 1, CV_64F, rho); compute_L_6x10(ut, l_6x10); compute_rho(rho); double Betas[4][4], rep_errors[4]; double Rs[4][3][3], ts[4][3]; find_betas_approx_1(&L_6x10, &Rho, Betas[1]); gauss_newton(&L_6x10, &Rho, Betas[1]); rep_errors[1] = compute_R_and_t(ut, Betas[1], Rs[1], ts[1]); find_betas_approx_2(&L_6x10, &Rho, Betas[2]); gauss_newton(&L_6x10, &Rho, Betas[2]); rep_errors[2] = compute_R_and_t(ut, Betas[2], Rs[2], ts[2]); find_betas_approx_3(&L_6x10, &Rho, Betas[3]); gauss_newton(&L_6x10, &Rho, Betas[3]); rep_errors[3] = compute_R_and_t(ut, Betas[3], Rs[3], ts[3]); int N = 1; if (rep_errors[2] < rep_errors[1]) N = 2; if (rep_errors[3] < rep_errors[N]) N = 3; Mat(3, 1, CV_64F, ts[N]).copyTo(t); Mat(3, 3, CV_64F, Rs[N]).copyTo(R); } void epnp::copy_R_and_t(const double R_src[3][3], const double t_src[3], double R_dst[3][3], double t_dst[3]) { for(int i = 0; i < 3; i++) { for(int j = 0; j < 3; j++) R_dst[i][j] = R_src[i][j]; t_dst[i] = t_src[i]; } } double epnp::dist2(const double * p1, const double * p2) { return (p1[0] - p2[0]) * (p1[0] - p2[0]) + (p1[1] - p2[1]) * (p1[1] - p2[1]) + (p1[2] - p2[2]) * (p1[2] - p2[2]); } double epnp::dot(const double * v1, const double * v2) { return v1[0] * v2[0] + v1[1] * v2[1] + v1[2] * v2[2]; } void epnp::estimate_R_and_t(double R[3][3], double t[3]) { double pc0[3], pw0[3]; pc0[0] = pc0[1] = pc0[2] = 0.0; pw0[0] = pw0[1] = pw0[2] = 0.0; for(int i = 0; i < number_of_correspondences; i++) { const double * pc = &pcs[3 * i]; const double * pw = &pws[3 * i]; for(int j = 0; j < 3; j++) { pc0[j] += pc[j]; pw0[j] += pw[j]; } } for(int j = 0; j < 3; j++) { pc0[j] /= number_of_correspondences; pw0[j] /= number_of_correspondences; } double abt[3 * 3], abt_d[3], abt_u[3 * 3], abt_v[3 * 3]; CvMat ABt = cvMat(3, 3, CV_64F, abt); CvMat ABt_D = cvMat(3, 1, CV_64F, abt_d); CvMat ABt_U = cvMat(3, 3, CV_64F, abt_u); CvMat ABt_V = cvMat(3, 3, CV_64F, abt_v); cvSetZero(&ABt); for(int i = 0; i < number_of_correspondences; i++) { double * pc = &pcs[3 * i]; double * pw = &pws[3 * i]; for(int j = 0; j < 3; j++) { abt[3 * j ] += (pc[j] - pc0[j]) * (pw[0] - pw0[0]); abt[3 * j + 1] += (pc[j] - pc0[j]) * (pw[1] - pw0[1]); abt[3 * j + 2] += (pc[j] - pc0[j]) * (pw[2] - pw0[2]); } } cvSVD(&ABt, &ABt_D, &ABt_U, &ABt_V, CV_SVD_MODIFY_A); for(int i = 0; i < 3; i++) for(int j = 0; j < 3; j++) R[i][j] = dot(abt_u + 3 * i, abt_v + 3 * j); const double det = 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] - 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]; if (det < 0) { R[2][0] = -R[2][0]; R[2][1] = -R[2][1]; R[2][2] = -R[2][2]; } t[0] = pc0[0] - dot(R[0], pw0); t[1] = pc0[1] - dot(R[1], pw0); t[2] = pc0[2] - dot(R[2], pw0); } void epnp::solve_for_sign(void) { if (pcs[2] < 0.0) { for(int i = 0; i < 4; i++) for(int j = 0; j < 3; j++) ccs[i][j] = -ccs[i][j]; for(int i = 0; i < number_of_correspondences; i++) { pcs[3 * i ] = -pcs[3 * i]; pcs[3 * i + 1] = -pcs[3 * i + 1]; pcs[3 * i + 2] = -pcs[3 * i + 2]; } } } double epnp::compute_R_and_t(const double * ut, const double * betas, double R[3][3], double t[3]) { compute_ccs(betas, ut); compute_pcs(); solve_for_sign(); estimate_R_and_t(R, t); return reprojection_error(R, t); } double epnp::reprojection_error(const double R[3][3], const double t[3]) { double sum2 = 0.0; for(int i = 0; i < number_of_correspondences; i++) { double * pw = &pws[3 * i]; double Xc = dot(R[0], pw) + t[0]; double Yc = dot(R[1], pw) + t[1]; double inv_Zc = 1.0 / (dot(R[2], pw) + t[2]); double ue = uc + fu * Xc * inv_Zc; double ve = vc + fv * Yc * inv_Zc; double u = us[2 * i], v = us[2 * i + 1]; sum2 += sqrt( (u - ue) * (u - ue) + (v - ve) * (v - ve) ); } return sum2 / number_of_correspondences; } // betas10 = [B11 B12 B22 B13 B23 B33 B14 B24 B34 B44] // betas_approx_1 = [B11 B12 B13 B14] void epnp::find_betas_approx_1(const CvMat * L_6x10, const CvMat * Rho, double * betas) { double l_6x4[6 * 4], b4[4]; CvMat L_6x4 = cvMat(6, 4, CV_64F, l_6x4); CvMat B4 = cvMat(4, 1, CV_64F, b4); for(int i = 0; i < 6; i++) { cvmSet(&L_6x4, i, 0, cvmGet(L_6x10, i, 0)); cvmSet(&L_6x4, i, 1, cvmGet(L_6x10, i, 1)); cvmSet(&L_6x4, i, 2, cvmGet(L_6x10, i, 3)); cvmSet(&L_6x4, i, 3, cvmGet(L_6x10, i, 6)); } cvSolve(&L_6x4, Rho, &B4, CV_SVD); if (b4[0] < 0) { betas[0] = sqrt(-b4[0]); betas[1] = -b4[1] / betas[0]; betas[2] = -b4[2] / betas[0]; betas[3] = -b4[3] / betas[0]; } else { betas[0] = sqrt(b4[0]); betas[1] = b4[1] / betas[0]; betas[2] = b4[2] / betas[0]; betas[3] = b4[3] / betas[0]; } } // betas10 = [B11 B12 B22 B13 B23 B33 B14 B24 B34 B44] // betas_approx_2 = [B11 B12 B22 ] void epnp::find_betas_approx_2(const CvMat * L_6x10, const CvMat * Rho, double * betas) { double l_6x3[6 * 3], b3[3]; CvMat L_6x3 = cvMat(6, 3, CV_64F, l_6x3); CvMat B3 = cvMat(3, 1, CV_64F, b3); for(int i = 0; i < 6; i++) { cvmSet(&L_6x3, i, 0, cvmGet(L_6x10, i, 0)); cvmSet(&L_6x3, i, 1, cvmGet(L_6x10, i, 1)); cvmSet(&L_6x3, i, 2, cvmGet(L_6x10, i, 2)); } cvSolve(&L_6x3, Rho, &B3, CV_SVD); if (b3[0] < 0) { betas[0] = sqrt(-b3[0]); betas[1] = (b3[2] < 0) ? sqrt(-b3[2]) : 0.0; } else { betas[0] = sqrt(b3[0]); betas[1] = (b3[2] > 0) ? sqrt(b3[2]) : 0.0; } if (b3[1] < 0) betas[0] = -betas[0]; betas[2] = 0.0; betas[3] = 0.0; } // betas10 = [B11 B12 B22 B13 B23 B33 B14 B24 B34 B44] // betas_approx_3 = [B11 B12 B22 B13 B23 ] void epnp::find_betas_approx_3(const CvMat * L_6x10, const CvMat * Rho, double * betas) { double l_6x5[6 * 5], b5[5]; CvMat L_6x5 = cvMat(6, 5, CV_64F, l_6x5); CvMat B5 = cvMat(5, 1, CV_64F, b5); for(int i = 0; i < 6; i++) { cvmSet(&L_6x5, i, 0, cvmGet(L_6x10, i, 0)); cvmSet(&L_6x5, i, 1, cvmGet(L_6x10, i, 1)); cvmSet(&L_6x5, i, 2, cvmGet(L_6x10, i, 2)); cvmSet(&L_6x5, i, 3, cvmGet(L_6x10, i, 3)); cvmSet(&L_6x5, i, 4, cvmGet(L_6x10, i, 4)); } cvSolve(&L_6x5, Rho, &B5, CV_SVD); if (b5[0] < 0) { betas[0] = sqrt(-b5[0]); betas[1] = (b5[2] < 0) ? sqrt(-b5[2]) : 0.0; } else { betas[0] = sqrt(b5[0]); betas[1] = (b5[2] > 0) ? sqrt(b5[2]) : 0.0; } if (b5[1] < 0) betas[0] = -betas[0]; betas[2] = b5[3] / betas[0]; betas[3] = 0.0; } void epnp::compute_L_6x10(const double * ut, double * l_6x10) { const double * v[4]; v[0] = ut + 12 * 11; v[1] = ut + 12 * 10; v[2] = ut + 12 * 9; v[3] = ut + 12 * 8; double dv[4][6][3]; for(int i = 0; i < 4; i++) { int a = 0, b = 1; for(int j = 0; j < 6; j++) { dv[i][j][0] = v[i][3 * a ] - v[i][3 * b]; dv[i][j][1] = v[i][3 * a + 1] - v[i][3 * b + 1]; dv[i][j][2] = v[i][3 * a + 2] - v[i][3 * b + 2]; b++; if (b > 3) { a++; b = a + 1; } } } for(int i = 0; i < 6; i++) { double * row = l_6x10 + 10 * i; row[0] = dot(dv[0][i], dv[0][i]); row[1] = 2.0f * dot(dv[0][i], dv[1][i]); row[2] = dot(dv[1][i], dv[1][i]); row[3] = 2.0f * dot(dv[0][i], dv[2][i]); row[4] = 2.0f * dot(dv[1][i], dv[2][i]); row[5] = dot(dv[2][i], dv[2][i]); row[6] = 2.0f * dot(dv[0][i], dv[3][i]); row[7] = 2.0f * dot(dv[1][i], dv[3][i]); row[8] = 2.0f * dot(dv[2][i], dv[3][i]); row[9] = dot(dv[3][i], dv[3][i]); } } void epnp::compute_rho(double * rho) { rho[0] = dist2(cws[0], cws[1]); rho[1] = dist2(cws[0], cws[2]); rho[2] = dist2(cws[0], cws[3]); rho[3] = dist2(cws[1], cws[2]); rho[4] = dist2(cws[1], cws[3]); rho[5] = dist2(cws[2], cws[3]); } void epnp::compute_A_and_b_gauss_newton(const double * l_6x10, const double * rho, const double betas[4], CvMat * A, CvMat * b) { for(int i = 0; i < 6; i++) { const double * rowL = l_6x10 + i * 10; double * rowA = A->data.db + i * 4; rowA[0] = 2 * rowL[0] * betas[0] + rowL[1] * betas[1] + rowL[3] * betas[2] + rowL[6] * betas[3]; rowA[1] = rowL[1] * betas[0] + 2 * rowL[2] * betas[1] + rowL[4] * betas[2] + rowL[7] * betas[3]; rowA[2] = rowL[3] * betas[0] + rowL[4] * betas[1] + 2 * rowL[5] * betas[2] + rowL[8] * betas[3]; rowA[3] = rowL[6] * betas[0] + rowL[7] * betas[1] + rowL[8] * betas[2] + 2 * rowL[9] * betas[3]; cvmSet(b, i, 0, rho[i] - ( rowL[0] * betas[0] * betas[0] + rowL[1] * betas[0] * betas[1] + rowL[2] * betas[1] * betas[1] + rowL[3] * betas[0] * betas[2] + rowL[4] * betas[1] * betas[2] + rowL[5] * betas[2] * betas[2] + rowL[6] * betas[0] * betas[3] + rowL[7] * betas[1] * betas[3] + rowL[8] * betas[2] * betas[3] + rowL[9] * betas[3] * betas[3] )); } } void epnp::gauss_newton(const CvMat * L_6x10, const CvMat * Rho, double betas[4]) { const int iterations_number = 5; double a[6*4], b[6], x[4]; CvMat A = cvMat(6, 4, CV_64F, a); CvMat B = cvMat(6, 1, CV_64F, b); CvMat X = cvMat(4, 1, CV_64F, x); for(int k = 0; k < iterations_number; k++) { compute_A_and_b_gauss_newton(L_6x10->data.db, Rho->data.db, betas, &A, &B); qr_solve(&A, &B, &X); for(int i = 0; i < 4; i++) betas[i] += x[i]; } } void epnp::qr_solve(CvMat * A, CvMat * b, CvMat * X) { const int nr = A->rows; const int nc = A->cols; if (max_nr != 0 && max_nr < nr) { delete [] A1; delete [] A2; } if (max_nr < nr) { max_nr = nr; A1 = new double[nr]; A2 = new double[nr]; } double * pA = A->data.db, * ppAkk = pA; for(int k = 0; k < nc; k++) { double * ppAik1 = ppAkk, eta = fabs(*ppAik1); for(int i = k + 1; i < nr; i++) { double elt = fabs(*ppAik1); if (eta < elt) eta = elt; ppAik1 += nc; } if (eta == 0) { A1[k] = A2[k] = 0.0; //cerr << "God damnit, A is singular, this shouldn't happen." << endl; return; } else { double * ppAik2 = ppAkk, sum2 = 0.0, inv_eta = 1. / eta; for(int i = k; i < nr; i++) { *ppAik2 *= inv_eta; sum2 += *ppAik2 * *ppAik2; ppAik2 += nc; } double sigma = sqrt(sum2); if (*ppAkk < 0) sigma = -sigma; *ppAkk += sigma; A1[k] = sigma * *ppAkk; A2[k] = -eta * sigma; for(int j = k + 1; j < nc; j++) { double * ppAik = ppAkk, sum = 0; for(int i = k; i < nr; i++) { sum += *ppAik * ppAik[j - k]; ppAik += nc; } double tau = sum / A1[k]; ppAik = ppAkk; for(int i = k; i < nr; i++) { ppAik[j - k] -= tau * *ppAik; ppAik += nc; } } } ppAkk += nc + 1; } // b <- Qt b double * ppAjj = pA, * pb = b->data.db; for(int j = 0; j < nc; j++) { double * ppAij = ppAjj, tau = 0; for(int i = j; i < nr; i++) { tau += *ppAij * pb[i]; ppAij += nc; } tau /= A1[j]; ppAij = ppAjj; for(int i = j; i < nr; i++) { pb[i] -= tau * *ppAij; ppAij += nc; } ppAjj += nc + 1; } // X = R-1 b double * pX = X->data.db; pX[nc - 1] = pb[nc - 1] / A2[nc - 1]; for(int i = nc - 2; i >= 0; i--) { double * ppAij = pA + i * nc + (i + 1), sum = 0; for(int j = i + 1; j < nc; j++) { sum += *ppAij * pX[j]; ppAij++; } pX[i] = (pb[i] - sum) / A2[i]; } } }