Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 1 | // This file is part of Eigen, a lightweight C++ template library |
| 2 | // for linear algebra. |
| 3 | // |
| 4 | // Copyright (C) 2009 Hauke Heibel <hauke.heibel@gmail.com> |
| 5 | // |
| 6 | // This Source Code Form is subject to the terms of the Mozilla |
| 7 | // Public License v. 2.0. If a copy of the MPL was not distributed |
| 8 | // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. |
| 9 | |
| 10 | #ifndef EIGEN_UMEYAMA_H |
| 11 | #define EIGEN_UMEYAMA_H |
| 12 | |
| 13 | // This file requires the user to include |
| 14 | // * Eigen/Core |
| 15 | // * Eigen/LU |
| 16 | // * Eigen/SVD |
| 17 | // * Eigen/Array |
| 18 | |
| 19 | namespace Eigen { |
| 20 | |
| 21 | #ifndef EIGEN_PARSED_BY_DOXYGEN |
| 22 | |
| 23 | // These helpers are required since it allows to use mixed types as parameters |
| 24 | // for the Umeyama. The problem with mixed parameters is that the return type |
| 25 | // cannot trivially be deduced when float and double types are mixed. |
| 26 | namespace internal { |
| 27 | |
| 28 | // Compile time return type deduction for different MatrixBase types. |
| 29 | // Different means here different alignment and parameters but the same underlying |
| 30 | // real scalar type. |
| 31 | template<typename MatrixType, typename OtherMatrixType> |
| 32 | struct umeyama_transform_matrix_type |
| 33 | { |
| 34 | enum { |
| 35 | MinRowsAtCompileTime = EIGEN_SIZE_MIN_PREFER_DYNAMIC(MatrixType::RowsAtCompileTime, OtherMatrixType::RowsAtCompileTime), |
| 36 | |
| 37 | // When possible we want to choose some small fixed size value since the result |
| 38 | // is likely to fit on the stack. So here, EIGEN_SIZE_MIN_PREFER_DYNAMIC is not what we want. |
| 39 | HomogeneousDimension = int(MinRowsAtCompileTime) == Dynamic ? Dynamic : int(MinRowsAtCompileTime)+1 |
| 40 | }; |
| 41 | |
| 42 | typedef Matrix<typename traits<MatrixType>::Scalar, |
| 43 | HomogeneousDimension, |
| 44 | HomogeneousDimension, |
| 45 | AutoAlign | (traits<MatrixType>::Flags & RowMajorBit ? RowMajor : ColMajor), |
| 46 | HomogeneousDimension, |
| 47 | HomogeneousDimension |
| 48 | > type; |
| 49 | }; |
| 50 | |
| 51 | } |
| 52 | |
| 53 | #endif |
| 54 | |
| 55 | /** |
| 56 | * \geometry_module \ingroup Geometry_Module |
| 57 | * |
| 58 | * \brief Returns the transformation between two point sets. |
| 59 | * |
| 60 | * The algorithm is based on: |
| 61 | * "Least-squares estimation of transformation parameters between two point patterns", |
| 62 | * Shinji Umeyama, PAMI 1991, DOI: 10.1109/34.88573 |
| 63 | * |
| 64 | * It estimates parameters \f$ c, \mathbf{R}, \f$ and \f$ \mathbf{t} \f$ such that |
| 65 | * \f{align*} |
| 66 | * \frac{1}{n} \sum_{i=1}^n \vert\vert y_i - (c\mathbf{R}x_i + \mathbf{t}) \vert\vert_2^2 |
| 67 | * \f} |
| 68 | * is minimized. |
| 69 | * |
| 70 | * The algorithm is based on the analysis of the covariance matrix |
| 71 | * \f$ \Sigma_{\mathbf{x}\mathbf{y}} \in \mathbb{R}^{d \times d} \f$ |
| 72 | * of the input point sets \f$ \mathbf{x} \f$ and \f$ \mathbf{y} \f$ where |
| 73 | * \f$d\f$ is corresponding to the dimension (which is typically small). |
| 74 | * The analysis is involving the SVD having a complexity of \f$O(d^3)\f$ |
| 75 | * though the actual computational effort lies in the covariance |
| 76 | * matrix computation which has an asymptotic lower bound of \f$O(dm)\f$ when |
| 77 | * the input point sets have dimension \f$d \times m\f$. |
| 78 | * |
| 79 | * Currently the method is working only for floating point matrices. |
| 80 | * |
| 81 | * \todo Should the return type of umeyama() become a Transform? |
| 82 | * |
| 83 | * \param src Source points \f$ \mathbf{x} = \left( x_1, \hdots, x_n \right) \f$. |
| 84 | * \param dst Destination points \f$ \mathbf{y} = \left( y_1, \hdots, y_n \right) \f$. |
| 85 | * \param with_scaling Sets \f$ c=1 \f$ when <code>false</code> is passed. |
| 86 | * \return The homogeneous transformation |
| 87 | * \f{align*} |
| 88 | * T = \begin{bmatrix} c\mathbf{R} & \mathbf{t} \\ \mathbf{0} & 1 \end{bmatrix} |
| 89 | * \f} |
Austin Schuh | c55b017 | 2022-02-20 17:52:35 -0800 | [diff] [blame^] | 90 | * minimizing the residual above. This transformation is always returned as an |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 91 | * Eigen::Matrix. |
| 92 | */ |
| 93 | template <typename Derived, typename OtherDerived> |
| 94 | typename internal::umeyama_transform_matrix_type<Derived, OtherDerived>::type |
| 95 | umeyama(const MatrixBase<Derived>& src, const MatrixBase<OtherDerived>& dst, bool with_scaling = true) |
| 96 | { |
| 97 | typedef typename internal::umeyama_transform_matrix_type<Derived, OtherDerived>::type TransformationMatrixType; |
| 98 | typedef typename internal::traits<TransformationMatrixType>::Scalar Scalar; |
| 99 | typedef typename NumTraits<Scalar>::Real RealScalar; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 100 | |
| 101 | EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsComplex, NUMERIC_TYPE_MUST_BE_REAL) |
| 102 | EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename internal::traits<OtherDerived>::Scalar>::value), |
| 103 | YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) |
| 104 | |
| 105 | enum { Dimension = EIGEN_SIZE_MIN_PREFER_DYNAMIC(Derived::RowsAtCompileTime, OtherDerived::RowsAtCompileTime) }; |
| 106 | |
| 107 | typedef Matrix<Scalar, Dimension, 1> VectorType; |
| 108 | typedef Matrix<Scalar, Dimension, Dimension> MatrixType; |
| 109 | typedef typename internal::plain_matrix_type_row_major<Derived>::type RowMajorMatrixType; |
| 110 | |
| 111 | const Index m = src.rows(); // dimension |
| 112 | const Index n = src.cols(); // number of measurements |
| 113 | |
| 114 | // required for demeaning ... |
| 115 | const RealScalar one_over_n = RealScalar(1) / static_cast<RealScalar>(n); |
| 116 | |
| 117 | // computation of mean |
| 118 | const VectorType src_mean = src.rowwise().sum() * one_over_n; |
| 119 | const VectorType dst_mean = dst.rowwise().sum() * one_over_n; |
| 120 | |
| 121 | // demeaning of src and dst points |
| 122 | const RowMajorMatrixType src_demean = src.colwise() - src_mean; |
| 123 | const RowMajorMatrixType dst_demean = dst.colwise() - dst_mean; |
| 124 | |
| 125 | // Eq. (36)-(37) |
| 126 | const Scalar src_var = src_demean.rowwise().squaredNorm().sum() * one_over_n; |
| 127 | |
| 128 | // Eq. (38) |
| 129 | const MatrixType sigma = one_over_n * dst_demean * src_demean.transpose(); |
| 130 | |
| 131 | JacobiSVD<MatrixType> svd(sigma, ComputeFullU | ComputeFullV); |
| 132 | |
| 133 | // Initialize the resulting transformation with an identity matrix... |
| 134 | TransformationMatrixType Rt = TransformationMatrixType::Identity(m+1,m+1); |
| 135 | |
| 136 | // Eq. (39) |
| 137 | VectorType S = VectorType::Ones(m); |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame] | 138 | |
| 139 | if ( svd.matrixU().determinant() * svd.matrixV().determinant() < 0 ) |
| 140 | S(m-1) = -1; |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 141 | |
| 142 | // Eq. (40) and (43) |
Austin Schuh | 189376f | 2018-12-20 22:11:15 +1100 | [diff] [blame] | 143 | Rt.block(0,0,m,m).noalias() = svd.matrixU() * S.asDiagonal() * svd.matrixV().transpose(); |
Brian Silverman | 72890c2 | 2015-09-19 14:37:37 -0400 | [diff] [blame] | 144 | |
| 145 | if (with_scaling) |
| 146 | { |
| 147 | // Eq. (42) |
| 148 | const Scalar c = Scalar(1)/src_var * svd.singularValues().dot(S); |
| 149 | |
| 150 | // Eq. (41) |
| 151 | Rt.col(m).head(m) = dst_mean; |
| 152 | Rt.col(m).head(m).noalias() -= c*Rt.topLeftCorner(m,m)*src_mean; |
| 153 | Rt.block(0,0,m,m) *= c; |
| 154 | } |
| 155 | else |
| 156 | { |
| 157 | Rt.col(m).head(m) = dst_mean; |
| 158 | Rt.col(m).head(m).noalias() -= Rt.topLeftCorner(m,m)*src_mean; |
| 159 | } |
| 160 | |
| 161 | return Rt; |
| 162 | } |
| 163 | |
| 164 | } // end namespace Eigen |
| 165 | |
| 166 | #endif // EIGEN_UMEYAMA_H |