File size: 41,643 Bytes
fe2b563 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 |
///////////////////////////////////////////////////////////////////////////////
/// "Sparse Iterative Closest Point"
/// by Sofien Bouaziz, Andrea Tagliasacchi, Mark Pauly
/// Copyright (C) 2013 LGG, EPFL
///////////////////////////////////////////////////////////////////////////////
/// 1) This file contains different implementations of the ICP algorithm.
/// 2) This code requires EIGEN and NANOFLANN.
/// 3) If OPENMP is activated some part of the code will be parallelized.
/// 4) This code is for now designed for 3D registration
/// 5) Two main input types are Eigen::Matrix3Xd or Eigen::Map<Eigen::Matrix3Xd>
///////////////////////////////////////////////////////////////////////////////
/// namespace nanoflann: NANOFLANN KD-tree adaptor for EIGEN
/// namespace RigidMotionEstimator: functions to compute the rigid motion
/// namespace SICP: sparse ICP implementation
/// namespace ICP: reweighted ICP implementation
///////////////////////////////////////////////////////////////////////////////
#ifndef ICP_H
#define ICP_H
#include <nanoflann.hpp>
#include <AndersonAcceleration.h>
#include <time.h>
#include <fstream>
#include <algorithm>
#include <median.h>
#include <iostream>
#define TUPLE_SCALE 0.95
#define TUPLE_MAX_CNT 1000
///////////////////////////////////////////////////////////////////////////////
namespace nanoflann {
/// KD-tree adaptor for working with data directly stored in an Eigen Matrix, without duplicating the data storage.
/// This code is adapted from the KDTreeEigenMatrixAdaptor class of nanoflann.hpp
template <class MatrixType, int DIM = -1, class Distance = nanoflann::metric_L2, typename IndexType = int>
struct KDTreeAdaptor {
typedef KDTreeAdaptor<MatrixType, DIM, Distance> self_t;
typedef typename MatrixType::Scalar num_t;
typedef typename Distance::template traits<num_t, self_t>::distance_t metric_t;
typedef KDTreeSingleIndexAdaptor< metric_t, self_t, DIM, IndexType> index_t;
index_t* index;
KDTreeAdaptor(const MatrixType &mat, const int leaf_max_size = 10) : m_data_matrix(mat) {
const size_t dims = mat.rows();
index = new index_t(dims, *this, nanoflann::KDTreeSingleIndexAdaptorParams(leaf_max_size, dims));
index->buildIndex();
}
~KDTreeAdaptor() { delete index; }
const MatrixType &m_data_matrix;
/// Query for the num_closest closest points to a given point (entered as query_point[0:dim-1]).
inline void query(const num_t *query_point, const size_t num_closest, IndexType *out_indices, num_t *out_distances_sq) const {
nanoflann::KNNResultSet<typename MatrixType::Scalar, IndexType> resultSet(num_closest);
resultSet.init(out_indices, out_distances_sq);
index->findNeighbors(resultSet, query_point, nanoflann::SearchParams());
}
/// Query for the closest points to a given point (entered as query_point[0:dim-1]).
inline IndexType closest(const num_t *query_point) const {
IndexType out_indices;
num_t out_distances_sq;
query(query_point, 1, &out_indices, &out_distances_sq);
return out_indices;
}
const self_t & derived() const { return *this; }
self_t & derived() { return *this; }
inline size_t kdtree_get_point_count() const { return m_data_matrix.cols(); }
/// Returns the distance between the vector "p1[0:size-1]" and the data point with index "idx_p2" stored in the class:
inline num_t kdtree_distance(const num_t *p1, const size_t idx_p2, size_t size) const {
num_t s = 0;
for (size_t i = 0; i<size; i++) {
num_t d = p1[i] - m_data_matrix.coeff(i, idx_p2);
s += d*d;
}
return s;
}
/// Returns the dim'th component of the idx'th point in the class:
inline num_t kdtree_get_pt(const size_t idx, int dim) const {
return m_data_matrix.coeff(dim, idx);
}
/// Optional bounding-box computation: return false to default to a standard bbox computation loop.
template <class BBOX> bool kdtree_get_bbox(BBOX&) const { return false; }
};
}
///////////////////////////////////////////////////////////////////////////////
/// Compute the rigid motion for point-to-point and point-to-plane distances
namespace RigidMotionEstimator {
/// @param Source (one 3D point per column)
/// @param Target (one 3D point per column)
/// @param Confidence weights
template <typename Derived1, typename Derived2, typename Derived3>
Eigen::Affine3d point_to_point(Eigen::MatrixBase<Derived1>& X,
Eigen::MatrixBase<Derived2>& Y,
const Eigen::MatrixBase<Derived3>& w) {
int dim = X.rows();
/// Normalize weight vector
Eigen::VectorXd w_normalized = w / w.sum();
/// De-mean
Eigen::VectorXd X_mean(dim), Y_mean(dim);
for (int i = 0; i<dim; ++i) {
X_mean(i) = (X.row(i).array()*w_normalized.transpose().array()).sum();
Y_mean(i) = (Y.row(i).array()*w_normalized.transpose().array()).sum();
}
X.colwise() -= X_mean;
Y.colwise() -= Y_mean;
/// Compute transformation
Eigen::Affine3d transformation;
MatrixXX sigma = X * w_normalized.asDiagonal() * Y.transpose();
Eigen::JacobiSVD<MatrixXX> svd(sigma, Eigen::ComputeFullU | Eigen::ComputeFullV);
if (svd.matrixU().determinant()*svd.matrixV().determinant() < 0.0) {
VectorX S = VectorX::Ones(dim); S(dim-1) = -1.0;
transformation.linear() = svd.matrixV()*S.asDiagonal()*svd.matrixU().transpose();
}
else {
transformation.linear() = svd.matrixV()*svd.matrixU().transpose();
}
transformation.translation() = Y_mean - transformation.linear()*X_mean;
/// Re-apply mean
X.colwise() += X_mean;
Y.colwise() += Y_mean;
/// Apply transformation
// X = transformation*X;
/// Return transformation
return transformation;
}
/// @param Source (one 3D point per column)
/// @param Target (one 3D point per column)
template <typename Derived1, typename Derived2>
inline Eigen::Affine3d point_to_point(Eigen::MatrixBase<Derived1>& X,
Eigen::MatrixBase<Derived2>& Y) {
return point_to_point(X, Y, Eigen::VectorXd::Ones(X.cols()));
}
/// @param Source (one 3D point per column)
/// @param Target (one 3D point per column)
/// @param Target normals (one 3D normal per column)
/// @param Confidence weights
/// @param Right hand side
template <typename Derived1, typename Derived2, typename Derived3, typename Derived4, typename Derived5>
Eigen::Affine3d point_to_plane(Eigen::MatrixBase<Derived1>& X,
Eigen::MatrixBase<Derived2>& Y,
Eigen::MatrixBase<Derived3>& N,
const Eigen::MatrixBase<Derived4>& w,
const Eigen::MatrixBase<Derived5>& u) {
typedef Eigen::Matrix<double, 6, 6> Matrix66;
typedef Eigen::Matrix<double, 6, 1> Vector6;
typedef Eigen::Block<Matrix66, 3, 3> Block33;
/// Normalize weight vector
Eigen::VectorXd w_normalized = w / w.sum();
/// De-mean
Eigen::Vector3d X_mean;
for (int i = 0; i<3; ++i)
X_mean(i) = (X.row(i).array()*w_normalized.transpose().array()).sum();
X.colwise() -= X_mean;
Y.colwise() -= X_mean;
/// Prepare LHS and RHS
Matrix66 LHS = Matrix66::Zero();
Vector6 RHS = Vector6::Zero();
Block33 TL = LHS.topLeftCorner<3, 3>();
Block33 TR = LHS.topRightCorner<3, 3>();
Block33 BR = LHS.bottomRightCorner<3, 3>();
Eigen::MatrixXd C = Eigen::MatrixXd::Zero(3, X.cols());
#pragma omp parallel
{
#pragma omp for
for (int i = 0; i<X.cols(); i++) {
C.col(i) = X.col(i).cross(N.col(i));
}
#pragma omp sections nowait
{
#pragma omp section
for (int i = 0; i<X.cols(); i++) TL.selfadjointView<Eigen::Upper>().rankUpdate(C.col(i), w(i));
#pragma omp section
for (int i = 0; i<X.cols(); i++) TR += (C.col(i)*N.col(i).transpose())*w(i);
#pragma omp section
for (int i = 0; i<X.cols(); i++) BR.selfadjointView<Eigen::Upper>().rankUpdate(N.col(i), w(i));
#pragma omp section
for (int i = 0; i<C.cols(); i++) {
double dist_to_plane = -((X.col(i) - Y.col(i)).dot(N.col(i)) - u(i))*w(i);
RHS.head<3>() += C.col(i)*dist_to_plane;
RHS.tail<3>() += N.col(i)*dist_to_plane;
}
}
}
LHS = LHS.selfadjointView<Eigen::Upper>();
/// Compute transformation
Eigen::Affine3d transformation;
Eigen::LDLT<Matrix66> ldlt(LHS);
RHS = ldlt.solve(RHS);
transformation = Eigen::AngleAxisd(RHS(0), Eigen::Vector3d::UnitX()) *
Eigen::AngleAxisd(RHS(1), Eigen::Vector3d::UnitY()) *
Eigen::AngleAxisd(RHS(2), Eigen::Vector3d::UnitZ());
transformation.translation() = RHS.tail<3>();
/// Apply transformation
X = transformation*X;
/// Re-apply mean
X.colwise() += X_mean;
Y.colwise() += X_mean;
transformation.translation() += -transformation.linear() * X_mean + X_mean;
/// Return transformation
return transformation;
}
/// @param Source (one 3D point per column)
/// @param Target (one 3D point per column)
/// @param Target normals (one 3D normal per column)
/// @param Confidence weights
template <typename Derived1, typename Derived2, typename Derived3, typename Derived4>
inline Eigen::Affine3d point_to_plane(Eigen::MatrixBase<Derived1>& X,
Eigen::MatrixBase<Derived2>& Yp,
Eigen::MatrixBase<Derived3>& Yn,
const Eigen::MatrixBase<Derived4>& w) {
return point_to_plane(X, Yp, Yn, w, Eigen::VectorXd::Zero(X.cols()));
}
}
///////////////////////////////////////////////////////////////////////////////
/// ICP implementation using ADMM/ALM/Penalty method
namespace SICP {
struct Parameters {
bool use_penalty = false; /// if use_penalty then penalty method else ADMM or ALM (see max_inner)
double p = 1.0; /// p norm
double mu = 10.0; /// penalty weight
double alpha = 1.2; /// penalty increase factor
double max_mu = 1e5; /// max penalty
int max_icp = 100; /// max ICP iteration
int max_outer = 100; /// max outer iteration
int max_inner = 1; /// max inner iteration. If max_inner=1 then ADMM else ALM
double stop = 1e-5; /// stopping criteria
bool print_icpn = false; /// (debug) print ICP iteration
Eigen::Matrix4d init_trans = Eigen::Matrix4d::Identity();
Eigen::Matrix4d gt_trans = Eigen::Matrix4d::Identity();
bool has_groundtruth = false;
int convergence_iter = 0;
double convergence_mse = 0.0;
double convergence_gt_mse = 0.0;
Eigen::Matrix4d res_trans = Eigen::Matrix4d::Identity();
std::string file_err = "";
std::string out_path = "";
int total_iters = 0;
};
/// Shrinkage operator (Automatic loop unrolling using template)
template<unsigned int I>
inline double shrinkage(double mu, double n, double p, double s) {
return shrinkage<I - 1>(mu, n, p, 1.0 - (p / mu)*std::pow(n, p - 2.0)*std::pow(s, p - 1.0));
}
template<>
inline double shrinkage<0>(double, double, double, double s) { return s; }
/// 3D Shrinkage for point-to-point
template<unsigned int I>
inline void shrink(Eigen::Matrix3Xd& Q, double mu, double p) {
double Ba = std::pow((2.0 / mu)*(1.0 - p), 1.0 / (2.0 - p));
double ha = Ba + (p / mu)*std::pow(Ba, p - 1.0);
#pragma omp parallel for
for (int i = 0; i<Q.cols(); ++i) {
double n = Q.col(i).norm();
double w = 0.0;
if (n > ha) w = shrinkage<I>(mu, n, p, (Ba / n + 1.0) / 2.0);
Q.col(i) *= w;
}
}
/// 1D Shrinkage for point-to-plane
template<unsigned int I>
inline void shrink(Eigen::VectorXd& y, double mu, double p) {
double Ba = std::pow((2.0 / mu)*(1.0 - p), 1.0 / (2.0 - p));
double ha = Ba + (p / mu)*std::pow(Ba, p - 1.0);
#pragma omp parallel for
for (int i = 0; i<y.rows(); ++i) {
double n = std::abs(y(i));
double s = 0.0;
if (n > ha) s = shrinkage<I>(mu, n, p, (Ba / n + 1.0) / 2.0);
y(i) *= s;
}
}
/// Sparse ICP with point to point
/// @param Source (one 3D point per column)
/// @param Target (one 3D point per column)
/// @param Parameters
template <typename Derived1, typename Derived2>
void point_to_point(Eigen::MatrixBase<Derived1>& X,
Eigen::MatrixBase<Derived2>& Y, Eigen::Vector3d& source_mean, Eigen::Vector3d& target_mean,
Parameters& par) {
/// Build kd-tree
nanoflann::KDTreeAdaptor<Eigen::MatrixBase<Derived2>, 3, nanoflann::metric_L2_Simple> kdtree(Y);
/// Buffers
Eigen::Matrix3Xd Q = Eigen::Matrix3Xd::Zero(3, X.cols());
Eigen::Matrix3Xd Z = Eigen::Matrix3Xd::Zero(3, X.cols());
Eigen::Matrix3Xd C = Eigen::Matrix3Xd::Zero(3, X.cols());
Eigen::Matrix3Xd ori_X = X;
Eigen::Affine3d T(par.init_trans);
Eigen::Matrix3Xd X_gt;
int nPoints = X.cols();
X = T * X;
Eigen::Matrix3Xd Xo1 = X;
Eigen::Matrix3Xd Xo2 = X;
double gt_mse = 0.0, run_time;
double begin_time, end_time;
std::vector<double> gt_mses, times;
if(par.has_groundtruth)
{
Eigen::Vector3d temp_trans = par.gt_trans.block(0, 3, 3, 1);
X_gt = ori_X;
X_gt.colwise() += source_mean;
X_gt = par.gt_trans.block(0, 0, 3, 3) * X_gt;
X_gt.colwise() += temp_trans - target_mean;
}
begin_time = omp_get_wtime();
/// ICP
int icp;
for (icp = 0; icp<par.max_icp; ++icp) {
/// Find closest point
#pragma omp parallel for
for (int i = 0; i<X.cols(); ++i) {
Q.col(i) = Y.col(kdtree.closest(X.col(i).data()));
}
end_time = omp_get_wtime();
run_time = end_time - begin_time;
///calc mse and gt_mse
if(par.has_groundtruth)
{
gt_mse = (X - X_gt).squaredNorm() / nPoints;
}
times.push_back(run_time);
gt_mses.push_back(gt_mse);
// if(par.print_icpn)
// std::cout << "iter = " << icp << ", time = " << run_time << ", mse = " << mse << ", gt_mse = " << gt_mse << std::endl;
/// Computer rotation and translation
double mu = par.mu;
for (int outer = 0; outer<par.max_outer; ++outer) {
double dual = 0.0;
for (int inner = 0; inner<par.max_inner; ++inner) {
/// Z update (shrinkage)
Z = X - Q + C / mu;
shrink<3>(Z, mu, par.p);
/// Rotation and translation update
Eigen::Matrix3Xd U = Q + Z - C / mu;
Eigen::Affine3d cur_T = RigidMotionEstimator::point_to_point(X, U);
X = cur_T * X;
T = cur_T * T;
/// Stopping criteria
dual = pow((X - Xo1).norm(),2) / nPoints;
Xo1 = X;
if (dual < par.stop) break;
}
/// C update (lagrange multipliers)
Eigen::Matrix3Xd P = X - Q - Z;
if (!par.use_penalty) C.noalias() += mu*P;
/// mu update (penalty)
if (mu < par.max_mu) mu *= par.alpha;
/// Stopping criteria
double primal = P.colwise().norm().maxCoeff();
if (primal < par.stop && dual < par.stop) break;
}
/// Stopping criteria
double stop = (X-Xo2).colwise().norm().maxCoeff();
Xo2 = X;
if (stop < par.stop) break;
}
if(par.has_groundtruth)
gt_mse = (X-X_gt).squaredNorm()/nPoints;
if(par.print_icpn)
{
std::ofstream out_res(par.out_path);
for(int i = 0; i<times.size(); i++)
{
out_res << times[i] << " " << gt_mses[i] << std::endl;
}
out_res.close();
}
T.translation().noalias() += -T.rotation()*source_mean + target_mean;
X.colwise() += target_mean;
///save convergence result
par.convergence_gt_mse = gt_mse;
par.convergence_iter = icp;
par.res_trans = T.matrix();
}
/// Sparse ICP with point to plane
/// @param Source (one 3D point per column)
/// @param Target (one 3D point per column)
/// @param Target normals (one 3D normal per column)
/// @param Parameters
template <typename Derived1, typename Derived2, typename Derived3>
void point_to_plane(Eigen::MatrixBase<Derived1>& X,
Eigen::MatrixBase<Derived2>& Y,
Eigen::MatrixBase<Derived3>& N, Eigen::Vector3d source_mean, Eigen::Vector3d target_mean,
Parameters &par) {
/// Build kd-tree
nanoflann::KDTreeAdaptor<Eigen::MatrixBase<Derived2>, 3, nanoflann::metric_L2_Simple> kdtree(Y);
/// Buffers
Eigen::Matrix3Xd Qp = Eigen::Matrix3Xd::Zero(3, X.cols());
Eigen::Matrix3Xd Qn = Eigen::Matrix3Xd::Zero(3, X.cols());
Eigen::VectorXd Z = Eigen::VectorXd::Zero(X.cols());
Eigen::VectorXd C = Eigen::VectorXd::Zero(X.cols());
Eigen::Matrix3Xd ori_X = X;
Eigen::Affine3d T(par.init_trans);
Eigen::Matrix3Xd X_gt;
int nPoints = X.cols();
X = T*X;
Eigen::Matrix3Xd Xo1 = X;
Eigen::Matrix3Xd Xo2 = X;
double gt_mse = 0.0, run_time;
double begin_time, end_time;
std::vector<double> gt_mses, times;
if(par.has_groundtruth)
{
Eigen::Vector3d temp_trans = par.gt_trans.block(0, 3, 3, 1);
X_gt = ori_X;
X_gt.colwise() += source_mean;
X_gt = par.gt_trans.block(0, 0, 3, 3) * X_gt;
X_gt.colwise() += temp_trans - target_mean;
}
begin_time = omp_get_wtime();
/// ICP
int icp;
int total_iters = 0;
for (icp = 0; icp<par.max_icp; ++icp) {
/// Find closest point
#pragma omp parallel for
for (int i = 0; i<X.cols(); ++i) {
int id = kdtree.closest(X.col(i).data());
Qp.col(i) = Y.col(id);
Qn.col(i) = N.col(id);
}
end_time = omp_get_wtime();
run_time = end_time - begin_time;
///calc mse and gt_mse
if(par.has_groundtruth)
{
gt_mse = (X - X_gt).squaredNorm() / nPoints;
}
times.push_back(run_time);
gt_mses.push_back(gt_mse);
if(par.print_icpn)
std::cout << "iter = " << icp << ", time = " << run_time << ", gt_mse = " << gt_mse << std::endl;
/// Computer rotation and translation
double mu = par.mu;
for (int outer = 0; outer<par.max_outer; ++outer) {
double dual = 0.0;
for (int inner = 0; inner<par.max_inner; ++inner) {
total_iters++;
/// Z update (shrinkage)
Z = (Qn.array()*(X - Qp).array()).colwise().sum().transpose() + C.array() / mu;
shrink<3>(Z, mu, par.p);
/// Rotation and translation update
Eigen::VectorXd U = Z - C / mu;
T = RigidMotionEstimator::point_to_plane(X, Qp, Qn, Eigen::VectorXd::Ones(X.cols()), U)*T;
/// Stopping criteria
dual = (X - Xo1).colwise().norm().maxCoeff();
Xo1 = X;
if (dual < par.stop) break;
}
/// C update (lagrange multipliers)
Eigen::VectorXd P = (Qn.array()*(X - Qp).array()).colwise().sum().transpose() - Z.array();
if (!par.use_penalty) C.noalias() += mu*P;
/// mu update (penalty)
if (mu < par.max_mu) mu *= par.alpha;
/// Stopping criteria
double primal = P.array().abs().maxCoeff();
if (primal < par.stop && dual < par.stop) break;
}
/// Stopping criteria
double stop = (X - Xo2).colwise().norm().maxCoeff();
Xo2 = X;
if (stop < par.stop) break;
}
if(par.has_groundtruth)
{
gt_mse = (X-X_gt).squaredNorm()/nPoints;
}
if(par.print_icpn)
{
std::ofstream out_res(par.out_path);
for(int i = 0; i<times.size(); i++)
{
out_res << times[i] << " " <<gt_mses[i] << std::endl;
}
out_res.close();
}
T.translation() += - T.rotation() * source_mean + target_mean;
X.colwise() += target_mean;
///save convergence result
par.convergence_gt_mse = gt_mse;
par.convergence_iter = icp;
par.res_trans = T.matrix();
par.total_iters = total_iters;
}
}
///////////////////////////////////////////////////////////////////////////////
/// ICP implementation using iterative reweighting
namespace ICP {
enum Function {
PNORM,
TUKEY,
FAIR,
LOGISTIC,
TRIMMED,
WELSCH,
AUTOWELSCH,
NONE
};
enum AAElementType {
EULERANGLE,
QUATERNION,
LOG_TRANS,
ROTATION,
FPFH,
DUAL_QUATERNION
};
class Parameters {
public:
Parameters() : f(NONE),
p(0.1),
max_icp(100),
max_outer(1),
stop(1e-5),
use_AA(false),
print_energy(false),
print_output(false),
anderson_m(5),
beta_(1.0),
error_overflow_threshold_(0.05),
has_groundtruth(false),
gt_trans(Eigen::Matrix4d::Identity()),
convergence_energy(0.0),
convergence_iter(0),
convergence_gt_mse(0.0),
nu_begin_k(3),
nu_end_k(1.0/(3.0*sqrt(3.0))),
use_init(false),
nu_alpha(1.0/2) {}
/// Parameters
Function f; /// robust function type
double p; /// paramter of the robust function/// para k
int max_icp; /// max ICP iteration
int max_outer; /// max outer iteration
double stop; /// stopping criteria
bool use_AA; /// whether using anderson acceleration
std::string out_path;
bool print_energy;///whether print energy
bool print_output; ///whether write result to txt
int anderson_m;
double beta_;
double error_overflow_threshold_;
MatrixXX init_trans;
MatrixXX gt_trans;
bool has_groundtruth;
double convergence_energy;
int convergence_iter;
double convergence_gt_mse;
MatrixXX res_trans;
double nu_begin_k;
double nu_end_k;
bool use_init;
double nu_alpha;
};
/// Weight functions
/// @param Residuals
/// @param Parameter
void uniform_weight(Eigen::VectorXd& r) {
r = Eigen::VectorXd::Ones(r.rows());
}
/// @param Residuals
/// @param Parameter
void pnorm_weight(Eigen::VectorXd& r, double p, double reg = 1e-8) {
for (int i = 0; i<r.rows(); ++i) {
r(i) = p / (std::pow(r(i), 2 - p) + reg);
}
}
/// @param Residuals
/// @param Parameter
void tukey_weight(Eigen::VectorXd& r, double p) {
for (int i = 0; i<r.rows(); ++i) {
if (r(i) > p) r(i) = 0.0;
else r(i) = std::pow((1.0 - std::pow(r(i) / p, 2.0)), 2.0);
}
}
/// @param Residuals
/// @param Parameter
void fair_weight(Eigen::VectorXd& r, double p) {
for (int i = 0; i<r.rows(); ++i) {
r(i) = 1.0 / (1.0 + r(i) / p);
}
}
/// @param Residuals
/// @param Parameter
void logistic_weight(Eigen::VectorXd& r, double p) {
for (int i = 0; i<r.rows(); ++i) {
r(i) = (p / r(i))*std::tanh(r(i) / p);
}
}
struct sort_pred {
bool operator()(const std::pair<int, double> &left,
const std::pair<int, double> &right) {
return left.second < right.second;
}
};
/// @param Residuals
/// @param Parameter
void trimmed_weight(Eigen::VectorXd& r, double p) {
std::vector<std::pair<int, double> > sortedDist(r.rows());
for (int i = 0; i<r.rows(); ++i) {
sortedDist[i] = std::pair<int, double>(i, r(i));
}
std::sort(sortedDist.begin(), sortedDist.end(), sort_pred());
r.setZero();
int nbV = r.rows()*p;
for (int i = 0; i<nbV; ++i) {
r(sortedDist[i].first) = 1.0;
}
}
/// @param Residuals
/// @param Parameter
void welsch_weight(Eigen::VectorXd& r, double p) {
for (int i = 0; i<r.rows(); ++i) {
r(i) = std::exp(-r(i)*r(i)/(2*p*p));
}
}
/// @param Residuals
/// @param Parameter
void autowelsch_weight(Eigen::VectorXd& r, double p) {
double median;
igl::median(r, median);
welsch_weight(r, p*median/(std::sqrt(2)*2.3));
//welsch_weight(r,p);
}
/// Energy functions
/// @param Residuals
/// @param Parameter
double uniform_energy(Eigen::VectorXd& r) {
double energy = 0;
for (int i = 0; i<r.rows(); ++i) {
energy += r(i)*r(i);
}
return energy;
}
/// @param Residuals
/// @param Parameter
double pnorm_energy(Eigen::VectorXd& r, double p, double reg = 1e-8) {
double energy = 0;
for (int i = 0; i<r.rows(); ++i) {
energy += (r(i)*r(i))*p / (std::pow(r(i), 2 - p) + reg);
}
return energy;
}
/// @param Residuals
/// @param Parameter
double tukey_energy(Eigen::VectorXd& r, double p) {
double energy = 0;
double w;
for (int i = 0; i<r.rows(); ++i) {
if (r(i) > p) w = 0.0;
else w = std::pow((1.0 - std::pow(r(i) / p, 2.0)), 2.0);
energy += (r(i)*r(i))*w;
}
return energy;
}
/// @param Residuals
/// @param Parameter
double fair_energy(Eigen::VectorXd& r, double p) {
double energy = 0;
for (int i = 0; i<r.rows(); ++i) {
energy += (r(i)*r(i))*1.0 / (1.0 + r(i) / p);
}
return energy;
}
/// @param Residuals
/// @param Parameter
double logistic_energy(Eigen::VectorXd& r, double p) {
double energy = 0;
for (int i = 0; i<r.rows(); ++i) {
energy += (r(i)*r(i))*(p / r(i))*std::tanh(r(i) / p);
}
return energy;
}
/// @param Residuals
/// @param Parameter
double trimmed_energy(Eigen::VectorXd& r, double p) {
std::vector<std::pair<int, double> > sortedDist(r.rows());
for (int i = 0; i<r.rows(); ++i) {
sortedDist[i] = std::pair<int, double>(i, r(i));
}
std::sort(sortedDist.begin(), sortedDist.end(), sort_pred());
Eigen::VectorXd t = r;
t.setZero();
double energy = 0;
int nbV = r.rows()*p;
for (int i = 0; i<nbV; ++i) {
energy += r(i)*r(i);
}
return energy;
}
/// @param Residuals
/// @param Parameter
double welsch_energy(Eigen::VectorXd& r, double p) {
double energy = 0;
for (int i = 0; i<r.rows(); ++i) {
energy += 1.0 - std::exp(-r(i)*r(i)/(2*p*p));
}
return energy;
}
/// @param Residuals
/// @param Parameter
double autowelsch_energy(Eigen::VectorXd& r, double p) {
double energy = 0;
energy = welsch_energy(r, 0.5);
return energy;
}
/// @param Function type
/// @param Residuals
/// @param Parameter
void robust_weight(Function f, Eigen::VectorXd& r, double p) {
switch (f) {
case PNORM: pnorm_weight(r, p); break;
case TUKEY: tukey_weight(r, p); break;
case FAIR: fair_weight(r, p); break;
case LOGISTIC: logistic_weight(r, p); break;
case TRIMMED: trimmed_weight(r, p); break;
case WELSCH: welsch_weight(r, p); break;
case AUTOWELSCH: autowelsch_weight(r,p); break;
case NONE: uniform_weight(r); break;
default: uniform_weight(r); break;
}
}
//Cacl energy
double get_energy(Function f, Eigen::VectorXd& r, double p) {
double energy = 0;
switch (f) {
//case PNORM: pnorm_weight(r,p); break;
case TUKEY: energy = tukey_energy(r, p); break;
case FAIR: energy = fair_energy(r, p); break;
case LOGISTIC: energy = logistic_energy(r, p); break;
case TRIMMED: energy = trimmed_energy(r, p); break;
case WELSCH: energy = welsch_energy(r, p); break;
case AUTOWELSCH: energy = autowelsch_energy(r, p); break;
case NONE: energy = uniform_energy(r); break;
default: energy = uniform_energy(r); break;
}
return energy;
}
}
namespace AAICP{
typedef Eigen::Matrix<Scalar, 3, 1> Vector3;
typedef Eigen::Matrix<Scalar, 6, 1> Vector6;
typedef Eigen::Matrix<Scalar, 3, 3> Matrix3;
typedef Eigen::Matrix<Scalar, 4, 4> Matrix4;
typedef Eigen::Matrix<Scalar, Eigen::Dynamic, 1> VectorX;
typedef Eigen::Matrix<Scalar, 6, Eigen::Dynamic> Matrix6X;
///aaicp
///////////////////////////////////////////////////////////////////////////////////////////
Vector6 Matrix42Vector6 (const Matrix4 m)
{
Vector6 v;
Matrix3 s = m.block(0,0,3,3);
v.head(3) = s.eulerAngles(0, 1, 2);
v.tail(3) = m.col(3).head(3);
return v;
}
///////////////////////////////////////////////////////////////////////////////////////////
Matrix4 Vector62Matrix4 (const Vector6 v)
{
Matrix3 s (Eigen::AngleAxis<Scalar>(v(0), Vector3::UnitX())
* Eigen::AngleAxis<Scalar>(v(1), Vector3::UnitY())
* Eigen::AngleAxis<Scalar>(v(2), Vector3::UnitZ()));
Matrix4 m = Matrix4::Zero();
m.block(0,0,3,3) = s;
m(3,3) = 1;
m.col(3).head(3) = v.tail(3);
return m;
}
///////////////////////////////////////////////////////////////////////////////////////////
int alphas_cond (VectorX alphas)
{
double alpha_limit_min_ = -10;
double alpha_limit_max_ = 10;
return alpha_limit_min_ < alphas.minCoeff() && alphas.maxCoeff() < alpha_limit_max_ && alphas(alphas.size()-1) > 0;
}
///////////////////////////////////////////////////////////////////////////////////////////
VectorX get_alphas_lstsq (const Matrix6X f)
{
Matrix6X A = f.leftCols(f.cols()-1);
A *= -1;
A += f.rightCols(1) * VectorX::Constant(f.cols()-1, 1).transpose();
VectorX sol = A.colPivHouseholderQr().solve(f.rightCols(1));
sol.conservativeResize(sol.size()+1);
sol[sol.size()-1] = 0;
sol[sol.size()-1] = 1-sol.sum();
return sol;
}
///////////////////////////////////////////////////////////////////////////////////////////
VectorX get_next_u (const Matrix6X u, const Matrix6X g, const Matrix6X f, std::vector<double> & save_alphas)
{
int i = 1;
double beta_ = 1.0;
save_alphas.clear();
Vector6 sol = ((1-beta_)*u.col(u.cols()-1) + beta_*g.col(g.cols()-1));
VectorX sol_alphas(1);
sol_alphas << 1;
i = 2;
for (; i <= f.cols(); i++)
{
VectorX alphas = get_alphas_lstsq(f.rightCols(i));
if (!alphas_cond(alphas))
{
break;
}
sol = (1-beta_)*u.rightCols(i)*alphas + beta_*g.rightCols(i)*alphas;
sol_alphas = alphas;
}
for(int i= 0; i<sol_alphas.rows(); i++)
{
save_alphas.push_back(sol_alphas[i]);
}
return sol;
}
template <typename Derived1, typename Derived2, typename Derived3>
void point_to_point_aaicp(Eigen::MatrixBase<Derived1>& X,Eigen::MatrixBase<Derived2>& Y, Eigen::MatrixBase<Derived3>& source_mean,
Eigen::MatrixBase<Derived3>& target_mean,
ICP::Parameters& par) {
/// Build kd-tree
nanoflann::KDTreeAdaptor<Eigen::MatrixBase<Derived2>, 3, nanoflann::metric_L2_Simple> kdtree(Y);
/// Buffers
Eigen::Matrix3Xd Q = Eigen::Matrix3Xd::Zero(3, X.cols());
Eigen::VectorXd W = Eigen::VectorXd::Zero(X.cols());
Eigen::Matrix3Xd ori_X = X;
double prev_energy = std::numeric_limits<double>::max(), energy = std::numeric_limits<double>::max();
Eigen::Affine3d T;
if(par.use_init)
{
T.linear() = par.init_trans.block(0,0,3,3);
T.translation() = par.init_trans.block(0,3,3,1);
}
else
T = Eigen::Affine3d::Identity();
Eigen::Matrix3Xd X_gt = X;
///stop criterion paras
MatrixXX To1 =T.matrix();
MatrixXX To2 = T.matrix();
///AA paras
Matrix6X u(6,0), g(6,0), f(6,0);
Vector6 u_next, u_k;
Matrix4 transformation = Matrix4::Identity();
Matrix4 final_transformation = Matrix4::Identity();
///output para
std::vector<double> times, energys, gt_mses;
double gt_mse;
double begin_time, end_time, run_time;
begin_time = omp_get_wtime();
///output coeffs
std::vector<std::vector<double>> coeffs;
coeffs.clear();
std::vector<double> alphas;
X = T * X;
///groud truth target point cloud
if(par.has_groundtruth)
{
Eigen::Vector3d temp_trans = par.gt_trans.block(0, 3, 3, 1);
X_gt = ori_X;
X_gt.colwise() += source_mean;
X_gt = par.gt_trans.block(0, 0, 3, 3) * X_gt;
X_gt.colwise() += temp_trans - target_mean;
}
///begin ICP
int icp = 0;
for (; icp<par.max_icp; ++icp)
{
bool accept_aa = false;
int nPoints = X.cols();
end_time = omp_get_wtime();
run_time = end_time - begin_time;
/// Find closest point
#pragma omp parallel for
for (int i = 0; i<nPoints; ++i) {
Q.col(i) = Y.col(kdtree.closest(X.col(i).data()));
}
///calc time
times.push_back(run_time);
if(par.has_groundtruth)
{
gt_mse = pow((X - X_gt).norm(),2) / nPoints;
}
gt_mses.push_back(gt_mse);
/// Computer rotation and translation
/// Compute weights
W = (X - Q).colwise().norm();
robust_weight(par.f, W, par.p);
/// Rotation and translation update
T = RigidMotionEstimator::point_to_point(X, Q, W) * T;
final_transformation = T.matrix();
///Anderson acceleration
if(icp)
{
Vector6 g_k = Matrix42Vector6(transformation * final_transformation);
// Calculate energy
W = (X - Q).colwise().norm();
energy = get_energy(par.f, W, par.p);
///The first heuristic
if ((energy - prev_energy)/prev_energy > par.error_overflow_threshold_) {
u_next = u_k = g.rightCols(1);
prev_energy = std::numeric_limits<double>::max();
u = u.rightCols(2);
g = g.rightCols(1);
f = f.rightCols(1);
}
else
{
prev_energy = energy;
g.conservativeResize(g.rows(),g.cols()+1);
g.col(g.cols()-1) = g_k;
Vector6 f_k = g_k - u_k;
f.conservativeResize(f.rows(),f.cols()+1);
f.col(f.cols()-1) = f_k;
u_next = get_next_u(u, g, f, alphas);
u.conservativeResize(u.rows(),u.cols()+1);
u.col(u.cols()-1) = u_next;
u_k = u_next;
accept_aa = true;
}
}
///init
else
{
// Calculate energy
W = (X - Q).colwise().norm();
prev_energy = get_energy(par.f, W, par.p);
Vector6 u0 = Matrix42Vector6(Matrix4::Identity());
u.conservativeResize(u.rows(),u.cols()+1);
u.col(0)=u0;
Vector6 u1 = Matrix42Vector6(transformation * final_transformation);
g.conservativeResize(g.rows(),g.cols()+1);
g.col(0)=u1;
u.conservativeResize(u.rows(),u.cols()+1);
u.col(1)=u1;
f.conservativeResize(f.rows(),f.cols()+1);
f.col(0)=u1 - u0;
u_next = u1;
u_k = u1;
energy = prev_energy;
}
transformation = Vector62Matrix4(u_next)*(final_transformation.inverse());
final_transformation = Vector62Matrix4(u_next);
X = final_transformation.block(0,0,3,3) * ori_X;
Vector3 trans = final_transformation.block(0,3,3,1);
X.colwise() += trans;
energys.push_back(energy);
if (par.print_energy)
std::cout << "icp iter = " << icp << ", Energy = " << energy << ", gt_mse = " << gt_mse<< std::endl;
/// Stopping criteria
double stop2 = (final_transformation - To2).norm();
To2 = final_transformation;
if (stop2 < par.stop && icp) break;
}
W = (X - Q).colwise().norm();
double last_energy = get_energy(par.f, W, par.p);
gt_mse = pow((X - X_gt).norm(),2) / X.cols();
final_transformation.block(0,3,3,1) += -final_transformation.block(0, 0, 3, 3)*source_mean + target_mean;
X.colwise() += target_mean;
///save convergence result
par.convergence_energy = last_energy;
par.convergence_gt_mse = gt_mse;
par.convergence_iter = icp;
par.res_trans = final_transformation;
///output
if (par.print_output)
{
std::ofstream out_res(par.out_path);
if (!out_res.is_open())
{
std::cout << "Can't open out file " << par.out_path << std::endl;
}
///output time and energy
out_res.precision(16);
for (int i = 0; i<icp; i++)
{
out_res << times[i] << " " << energys[i] <<" " << gt_mses[i] << std::endl;
}
out_res.close();
}
}
}
#endif
|