这里以colmap 框架为准,主要对其重要环节代码结合自己的想法进行逐一讲解,colmap 作为目前state-of-the-art 的视觉重建pipeline,本人将其代码分为两个大环节:前端和后端.前端主要是特征提取和匹配,后端包括三角化,Register,BA等环节.
特征提取
特征匹配
三角化或前方交会
运动恢复
光束法平差及算法改进
sift-GPU 算法
待写
DSP-SIFT算法
待写
Domain-size pooling SIFT 是从多个尺度的SIFT 描述子进行一个平均,
参考论文:Domain-Size Pooling in Local Descriptors and Network Architectures,
其已被证明优于其他SIFT算法的变种及一些深度学习算子.
匹配方法
待写
几何验证-剔除outliers
1 . 对于标定相机,利用E/F 、H/F 、H/E的内点个数比值来决定使用哪种模型来剔除误点
代码最终 inlier_matches = ExtractInlierMatches(matches,模型内点数,模型mask)
// 对于标定相机 利用 E/F H/F H/E 的内点比值来决定使用那种模型来剔除outliers
const double E_F_inlier_ratio =
static_cast<double>(E_report.support.num_inliers) /
F_report.support.num_inliers;
const double H_F_inlier_ratio =
static_cast<double>(H_report.support.num_inliers) /
F_report.support.num_inliers;
const double H_E_inlier_ratio =
static_cast<double>(H_report.support.num_inliers) /
E_report.support.num_inliers;
2 . 对于非标相机,利用F 矩阵模型
// 非标相机估计 H/F 的比值即可,E 矩阵无法得到
const double H_F_inlier_ratio =
static_cast<double>(H_report.support.num_inliers) /
F_report.support.num_inliers;
if (H_F_inlier_ratio > options.max_H_inlier_ratio) {
config = ConfigurationType::PLANAR_OR_PANORAMIC;
} else {
config = ConfigurationType::UNCALIBRATED;
}
inlier_matches = ExtractInlierMatches(matches, F_report.support.num_inliers,
F_report.inlier_mask);
对于外点的剔除,除了利用ransac 估计模型来剔除outliers ,
还有可改进的思路吗?众所周知,ransac 是每次进行随机抽样,重新计算模型,这样就比较耗时.
colmap 代码中,三角化分为以下几个函数:
1 . 两帧三角化 ,即使用SVD 分解求解P(M*P=m),代码如下:
// proj_matrix1 ,proj_matrix2 分别是世界系到图像系的投影矩阵
Eigen::Vector3d TriangulatePoint(const Eigen::Matrix3x4d& proj_matrix1,
const Eigen::Matrix3x4d& proj_matrix2,
const Eigen::Vector2d& point1,
const Eigen::Vector2d& point2) {
Eigen::Matrix4d A;
A.row(0) = point1(0) * proj_matrix1.row(2) - proj_matrix1.row(0);
A.row(1) = point1(1) * proj_matrix1.row(2) - proj_matrix1.row(1);
A.row(2) = point2(0) * proj_matrix2.row(2) - proj_matrix2.row(0);
A.row(3) = point2(1) * proj_matrix2.row(2) - proj_matrix2.row(1);
Eigen::JacobiSVD svd(A, Eigen::ComputeFullV);
return svd.matrixV().col(3).hnormalized();
}
2 . 多帧三角化 ,即是求解超定方程组.从多个观测值恢复3D 场景结构,三角化需要满足以下两个条件:
(1) 所有观测值需要满足 cheirality constraint (什么是cheirality constraint ,即是正景深约束,
多视图几何书中分解E 矩阵的时候,会得到四种结果-R,t,但是重构点在相机前方的结果只有一个)
(2) 观测对之间有足够的三角化角度
对于求解的结果需要满足以上两个条件,代码如下
条件1 : xyz cheirality constraint
for (const auto& pose : pose_data) {
if (!HasPointPositiveDepth(pose.proj_matrix, xyz)) {
return std::vector();
}
}
条件2: 像对之间需足够的三角化角度(阈值为min_tri_angle)
for (size_t i = 0; i < pose_data.size(); ++i) {
for (size_t j = 0; j < i; ++j) {
const double tri_angle = CalculateTriangulationAngle(
pose_data[i].proj_center, pose_data[j].proj_center, xyz);
if (tri_angle >= min_tri_angle_) {
return std::vector<M_t>{xyz};
}
}
}
3 . 鲁棒性LO-RANSAC 多帧三角化,其主要是剔除一些错误的匹配点,
这个也是colmap 框架三角化一个特色,代码如下:
bool EstimateTriangulation(
const EstimateTriangulationOptions& options,
const std::vector<TriangulationEstimator::PointData>& point_data,
const std::vector<TriangulationEstimator::PoseData>& pose_data,
std::vector<char>* inlier_mask, Eigen::Vector3d* xyz) {
CHECK_NOTNULL(inlier_mask);
CHECK_NOTNULL(xyz);
CHECK_GE(point_data.size(), 2);
CHECK_EQ(point_data.size(), pose_data.size());
options.Check();
// Robustly estimate track using LORANSAC.
LORANSAC<TriangulationEstimator, TriangulationEstimator,
InlierSupportMeasurer, CombinationSampler>
ransac(options.ransac_options);
ransac.estimator.SetMinTriAngle(options.min_tri_angle);
ransac.estimator.SetResidualType(options.residual_type);
ransac.local_estimator.SetMinTriAngle(options.min_tri_angle);
ransac.local_estimator.SetResidualType(options.residual_type);
const auto report = ransac.Estimate(point_data, pose_data);
if (!report.success) {
return false;
}
*inlier_mask = report.inlier_mask;
*xyz = report.model;
return report.success;
}
基本矩阵FundamentalMatrix
基本矩阵的求解从以下公式出发:
x ′ T F x = 0 \mathbf{x}^{\prime T} F \mathbf{x}=0 x′TFx=0
以上公式可以写成包含9个未知数的线性齐次方程,如下:
u i T f = 0 u i = [ u i u i ′ , v i u i ′ , u i ′ , u i v i ′ , v i v i ′ , v i ′ , u i , v i , 1 ] T f = [ F 11 , F 12 , F 13 , F 21 , F 22 , F 23 , F 31 , F 32 , F 33 ] T \begin{aligned} &\mathbf{u}_{i}^{T} \mathbf{f}=0\\ &\begin{aligned} \mathbf{u}_{i} &=\left[u_{i} u_{i}^{\prime}, v_{i} u_{i}^{\prime}, u_{i}^{\prime}, u_{i} v_{i}^{\prime}, v_{i} v_{i}^{\prime}, v_{i}^{\prime}, u_{i}, v_{i}, 1\right]^{T} \\ \mathbf{f} &=\left[F_{11}, F_{12}, F_{13}, F_{21}, F_{22}, F_{23}, F_{31}, F_{32}, F_{33}\right]^{T} \end{aligned} \end{aligned} uiTf=0uif=[uiui′,viui′,ui′,uivi′,vivi′,vi′,ui,vi,1]T=[F11,F12,F13,F21,F22,F23,F31,F32,F33]T
读过MVG的人都知道基本矩阵由于一个尺度不确定和其秩等于2,故其DOF 等于7.所以估计基本矩阵的算法根据不忽略rank-2 constraint 和忽略rank-2 constraint,求解方法可分为7点法和8点法或者more points 法
F = α F 1 + ( 1 − α ) F 2 \mathbf{F}=\alpha \mathbf{F}_{1}+(1-\alpha) \mathbf{F}_{2} F=αF1+(1−α)F2
det [ α F 1 + ( 1 − α ) F 2 ] = 0 \operatorname{det}\left[\alpha \mathbf{F}_{1}+(1-\alpha) \mathbf{F}_{2}\right]=0 det[αF1+(1−α)F2]=0
代码如下:
std::vector<FundamentalMatrixSevenPointEstimator::M_t>
FundamentalMatrixSevenPointEstimator::Estimate(
const std::vector<X_t>& points1, const std::vector<Y_t>& points2) {
CHECK_EQ(points1.size(), 7);
CHECK_EQ(points2.size(), 7);
Eigen::Matrix<double, 7, 9> A;
for (size_t i = 0; i < 7; ++i) {
const double x0 = points1[i](0);
const double y0 = points1[i](1);
const double x1 = points2[i](0);
const double y1 = points2[i](1);
A(i, 0) = x1 * x0;
A(i, 1) = x1 * y0;
A(i, 2) = x1;
A(i, 3) = y1 * x0;
A(i, 4) = y1 * y0;
A(i, 5) = y1;
A(i, 6) = x0;
A(i, 7) = y0;
A(i, 8) = 1;
}
// 9 个未知数,7个方程, 所以我们有两个null space .
Eigen::JacobiSVD<Eigen::Matrix<double, 7, 9>> svd(A, Eigen::ComputeFullV);
const Eigen::Matrix<double, 9, 9> f = svd.matrixV();
Eigen::Matrix<double, 1, 9> f1 = f.col(7);
Eigen::Matrix<double, 1, 9> f2 = f.col(8);
f1 -= f2;
// 很明显,方程个数不够,无法求解,所以我们必须增加约束
// det(F) = det(lambda * f1 + (1 - lambda) * f2)
// 其中lambda + mu = 1
const double t0 = f1(4) * f1(8) - f1(5) * f1(7);
const double t1 = f1(3) * f1(8) - f1(5) * f1(6);
const double t2 = f1(3) * f1(7) - f1(4) * f1(6);
const double t3 = f2(4) * f2(8) - f2(5) * f2(7);
const double t4 = f2(3) * f2(8) - f2(5) * f2(6);
const double t5 = f2(3) * f2(7) - f2(4) * f2(6);
Eigen::Vector4d coeffs;
coeffs(0) = f1(0) * t0 - f1(1) * t1 + f1(2) * t2;
coeffs(1) = f2(0) * t0 - f2(1) * t1 + f2(2) * t2 -
f2(3) * (f1(1) * f1(8) - f1(2) * f1(7)) +
f2(4) * (f1(0) * f1(8) - f1(2) * f1(6)) -
f2(5) * (f1(0) * f1(7) - f1(1) * f1(6)) +
f2(6) * (f1(1) * f1(5) - f1(2) * f1(4)) -
f2(7) * (f1(0) * f1(5) - f1(2) * f1(3)) +
f2(8) * (f1(0) * f1(4) - f1(1) * f1(3));
coeffs(2) = f1(0) * t3 - f1(1) * t4 + f1(2) * t5 -
f1(3) * (f2(1) * f2(8) - f2(2) * f2(7)) +
f1(4) * (f2(0) * f2(8) - f2(2) * f2(6)) -
f1(5) * (f2(0) * f2(7) - f2(1) * f2(6)) +
f1(6) * (f2(1) * f2(5) - f2(2) * f2(4)) -
f1(7) * (f2(0) * f2(5) - f2(2) * f2(3)) +
f1(8) * (f2(0) * f2(4) - f2(1) * f2(3));
coeffs(3) = f2(0) * t3 - f2(1) * t4 + f2(2) * t5;
Eigen::VectorXd roots_real;
Eigen::VectorXd roots_imag;
if (!FindPolynomialRootsCompanionMatrix(coeffs, &roots_real, &roots_imag)) {
return {};
}
std::vector<M_t> models;
models.reserve(roots_real.size());
for (Eigen::VectorXd::Index i = 0; i < roots_real.size(); ++i) {
const double kMaxRootImag = 1e-10;
if (std::abs(roots_imag(i)) > kMaxRootImag) {
continue;
}
const double lambda = roots_real(i);
const double mu = 1;
Eigen::MatrixXd F = lambda * f1 + mu * f2;
F.resize(3, 3);
const double kEps = 1e-10;
if (std::abs(F(2, 2)) < kEps) {
continue;
}
F /= F(2, 2);
models.push_back(F.transpose());
}
return models;
}
min F ∑ i ( m ~ i ′ T F m ~ i ) 2 \min _{\mathbf{F}} \sum_{i}\left(\tilde{\mathbf{m}}_{i}^{\prime T} \mathbf{F} \tilde{\mathbf{m}}_{i}\right)^{2} minF∑i(m~i′TFm~i)2
min f ∥ U n f ∥ 2 \min _{\mathrm{f}}\left\|\mathbf{U}_{n} \mathbf{f}\right\|^{2} minf∥Unf∥2
现在f只 受一个未知尺度的影响,为了避免f=0,我们必须增加约束,即使SVD 中V 的最后一个特征值等于0(详见<<计算机视觉中的多视图几何>>书本281页),代码如下:
std::vector<FundamentalMatrixEightPointEstimator::M_t>
FundamentalMatrixEightPointEstimator::Estimate(
const std::vector<X_t>& points1, const std::vector<Y_t>& points2) {
CHECK_EQ(points1.size(), points2.size());
// 中心归一化是为了数值稳定性
std::vector<X_t> normed_points1;
std::vector<Y_t> normed_points2;
Eigen::Matrix3d points1_norm_matrix;
Eigen::Matrix3d points2_norm_matrix;
CenterAndNormalizeImagePoints(points1, &normed_points1, &points1_norm_matrix);
CenterAndNormalizeImagePoints(points2, &normed_points2, &points2_norm_matrix);
//解决线性方程 x2' * F * x1 = 0.
Eigen::Matrix<double, Eigen::Dynamic, 9> cmatrix(points1.size(), 9);
for (size_t i = 0; i < points1.size(); ++i) {
cmatrix.block<1, 3>(i, 0) = normed_points1[i].homogeneous();
cmatrix.block<1, 3>(i, 0) *= normed_points2[i].x();
cmatrix.block<1, 3>(i, 3) = normed_points1[i].homogeneous();
cmatrix.block<1, 3>(i, 3) *= normed_points2[i].y();
cmatrix.block<1, 3>(i, 6) = normed_points1[i].homogeneous();
}
// SVD 分解
Eigen::JacobiSVD<Eigen::Matrix<double, Eigen::Dynamic, 9>> cmatrix_svd(
cmatrix, Eigen::ComputeFullV);
const Eigen::VectorXd cmatrix_nullspace = cmatrix_svd.matrixV().col(8);
const Eigen::Map<const Eigen::Matrix3d> ematrix_t(cmatrix_nullspace.data());
// 增加约束,即SVD 中的V 的特征向量最后一个等于0
Eigen::JacobiSVD<Eigen::Matrix3d> fmatrix_svd(
ematrix_t.transpose(), Eigen::ComputeFullU | Eigen::ComputeFullV);
Eigen::Vector3d singular_values = fmatrix_svd.singularValues();
singular_values(2) = 0.0;
const Eigen::Matrix3d F = fmatrix_svd.matrixU() *
singular_values.asDiagonal() *
fmatrix_svd.matrixV().transpose();
const std::vector<M_t> models = {points2_norm_matrix.transpose() * F *
points1_norm_matrix};
return models;
}
本质矩阵
( 1) 5点法求解
待写
( 2) 8点法求解
待写
单应矩阵
单应矩阵的求解相对于本质和基本就相对简单,在colmap 代码中直接利用DLT算法直接求解,理论公式如下:
( u v 1 ) = H ( x y 1 ) \left(\begin{array}{l}u \\ v \\ 1\end{array}\right)=H\left(\begin{array}{l}x \\ y \\ 1\end{array}\right) ⎝⎛uv1⎠⎞=H⎝⎛xy1⎠⎞
其中H:
H = [ h 1 h 2 h 3 h 4 h 5 h 6 h 7 h 8 h 9 ] H=\left[\begin{array}{lll}h_{1} & h_{2} & h_{3} \\ h_{4} & h_{5} & h_{6} \\ h_{7} & h_{8} & h_{9}\end{array}\right] H=⎣⎡h1h4h7h2h5h8h3h6h9⎦⎤
线性化展开得到:
− h 1 x − h 2 y − h 3 + ( h 7 x + h 8 y + h 9 ) u = 0 -h_{1} x-h_{2} y-h_{3}+\left(h_{7} x+h_{8} y+h_{9}\right) u=0 −h1x−h2y−h3+(h7x+h8y+h9)u=0
− h 4 x − h 5 y − h 6 + ( h 7 x + h 8 y + h 9 ) v = 0 -h_{4} x-h_{5} y-h_{6}+\left(h_{7} x+h_{8} y+h_{9}\right) v=0 −h4x−h5y−h6+(h7x+h8y+h9)v=0
整理可得:
A i = ( − x − y − 1 0 0 0 u x u y u 0 0 0 − x − y − 1 v x v y v ) A_{i}=\left(\begin{array}{ccccccccc}-x & -y & -1 & 0 & 0 & 0 & u x & u y & u \\ 0 & 0 & 0 & -x & -y & -1 & v x & v y & v\end{array}\right) Ai=(−x0−y0−100−x0−y0−1uxvxuyvyuv)
h = ( h 1 h 2 h t 3 h 4 h 5 h 6 h 7 h 8 h 9 ) h=\left(\begin{array}{cccccc}h_{1} & h_{2} h t_{3} & h_{4} & h_{5} & h_{6} & h_{7} & h_{8} & h_{9}\end{array}\right)_{} h=(h1h2ht3h4h5h6h7h8h9)
利用最小二乘求解超定方程组便可求,代码如下:
std::vector<HomographyMatrixEstimator::M_t> HomographyMatrixEstimator::Estimate(
const std::vector<X_t>& points1, const std::vector<Y_t>& points2) {
CHECK_EQ(points1.size(), points2.size());
const size_t N = points1.size();
// 中心归一化,数值稳定性
std::vector<X_t> normed_points1;
std::vector<Y_t> normed_points2;
Eigen::Matrix3d points1_norm_matrix;
Eigen::Matrix3d points2_norm_matrix;
CenterAndNormalizeImagePoints(points1, &normed_points1, &points1_norm_matrix);
CenterAndNormalizeImagePoints(points2, &normed_points2, &points2_norm_matrix);
//构建方程
Eigen::Matrix<double, Eigen::Dynamic, 9> A = Eigen::MatrixXd::Zero(2 * N, 9);
for (size_t i = 0, j = N; i < points1.size(); ++i, ++j) {
const double s_0 = normed_points1[i](0);
const double s_1 = normed_points1[i](1);
const double d_0 = normed_points2[i](0);
const double d_1 = normed_points2[i](1);
A(i, 0) = -s_0;
A(i, 1) = -s_1;
A(i, 2) = -1;
A(i, 6) = s_0 * d_0;
A(i, 7) = s_1 * d_0;
A(i, 8) = d_0;
A(j, 3) = -s_0;
A(j, 4) = -s_1;
A(j, 5) = -1;
A(j, 6) = s_0 * d_1;
A(j, 7) = s_1 * d_1;
A(j, 8) = d_1;
}
// SVD 求解
Eigen::JacobiSVD<Eigen::Matrix<double, Eigen::Dynamic, 9>> svd(
A, Eigen::ComputeFullV);
const Eigen::VectorXd nullspace = svd.matrixV().col(8);
Eigen::Map<const Eigen::Matrix3d> H_t(nullspace.data()); // 单应矩阵
const std::vector<M_t> models = {points2_norm_matrix.inverse() *
H_t.transpose() * points1_norm_matrix};
return models;
}
Estimate Relative Pose (2D-2D)
待写
Estimate Absolute Pose(2D-3D)
待写
经典的重投影误差
待写
pose 为常量的重投影误差
待写
部分参数优化(如R or T)的重投影误差~自己改进
待写
GPS 约束的重投影误差(SfM 融合gps)~自己改进
待写
GCP 约束的重投影误差( Marker SfM)~自己改进
待写
IMU 角度约束的重投影误差~自己改进
待写