gtsam/gtsam/slam/EssentialMatrixFactor.h

526 lines
18 KiB
C++

/* ----------------------------------------------------------------------------
* GTSAM Copyright 2010-2014, Georgia Tech Research Corporation,
* Atlanta, Georgia 30332-0415
* All Rights Reserved
* Authors: Frank Dellaert, et al. (see THANKS for the full author list)
* See LICENSE for the license information
* -------------------------------------------------------------------------- */
/*
* @file EssentialMatrixFactor.h
* @brief EssentialMatrixFactor class
* @author Frank Dellaert
* @author Ayush Baid
* @author Akshay Krishnan
* @date December 17, 2013
*/
#pragma once
#include <gtsam/geometry/EssentialMatrix.h>
#include <gtsam/geometry/PinholeCamera.h>
#include <gtsam/nonlinear/NonlinearFactor.h>
#include <iostream>
#include <cassert>
namespace gtsam {
/**
* Factor that evaluates epipolar error p'Ep for given essential matrix
*/
class EssentialMatrixFactor : public NoiseModelFactorN<EssentialMatrix> {
Vector3 vA_, vB_; ///< Homogeneous versions, in ideal coordinates
typedef NoiseModelFactorN<EssentialMatrix> Base;
typedef EssentialMatrixFactor This;
public:
// Provide access to the Matrix& version of evaluateError:
using Base::evaluateError;
/**
* Constructor
* @param key Essential Matrix variable key
* @param pA point in first camera, in calibrated coordinates
* @param pB point in second camera, in calibrated coordinates
* @param model noise model is about dot product in ideal, homogeneous
* coordinates
*/
EssentialMatrixFactor(Key key, const Point2& pA, const Point2& pB,
const SharedNoiseModel& model)
: Base(model, key) {
vA_ = EssentialMatrix::Homogeneous(pA);
vB_ = EssentialMatrix::Homogeneous(pB);
}
/**
* Constructor
* @param key Essential Matrix variable key
* @param pA point in first camera, in pixel coordinates
* @param pB point in second camera, in pixel coordinates
* @param model noise model is about dot product in ideal, homogeneous
* coordinates
* @param K calibration object, will be used only in constructor
*/
template <class CALIBRATION>
EssentialMatrixFactor(Key key, const Point2& pA, const Point2& pB,
const SharedNoiseModel& model,
std::shared_ptr<CALIBRATION> K)
: Base(model, key) {
assert(K);
vA_ = EssentialMatrix::Homogeneous(K->calibrate(pA));
vB_ = EssentialMatrix::Homogeneous(K->calibrate(pB));
}
/// @return a deep copy of this factor
gtsam::NonlinearFactor::shared_ptr clone() const override {
return std::static_pointer_cast<gtsam::NonlinearFactor>(
gtsam::NonlinearFactor::shared_ptr(new This(*this)));
}
/// print
void print(
const std::string& s = "",
const KeyFormatter& keyFormatter = DefaultKeyFormatter) const override {
Base::print(s);
std::cout << " EssentialMatrixFactor with measurements\n ("
<< vA_.transpose() << ")' and (" << vB_.transpose() << ")'"
<< std::endl;
}
/// vector of errors returns 1D vector
Vector evaluateError(const EssentialMatrix& E,
OptionalMatrixType H) const override {
Vector error(1);
error << E.error(vA_, vB_, H);
return error;
}
public:
GTSAM_MAKE_ALIGNED_OPERATOR_NEW
};
/**
* Binary factor that optimizes for E and inverse depth d: assumes measurement
* in image 2 is perfect, and returns re-projection error in image 1
*/
class EssentialMatrixFactor2
: public NoiseModelFactorN<EssentialMatrix, double> {
Point3 dP1_; ///< 3D point corresponding to measurement in image 1
Point2 pn_; ///< Measurement in image 2, in ideal coordinates
double f_; ///< approximate conversion factor for error scaling
typedef NoiseModelFactorN<EssentialMatrix, double> Base;
typedef EssentialMatrixFactor2 This;
public:
// Provide access to the Matrix& version of evaluateError:
using Base::evaluateError;
/**
* Constructor
* @param key1 Essential Matrix variable key
* @param key2 Inverse depth variable key
* @param pA point in first camera, in calibrated coordinates
* @param pB point in second camera, in calibrated coordinates
* @param model noise model should be in pixels, as well
*/
EssentialMatrixFactor2(Key key1, Key key2, const Point2& pA, const Point2& pB,
const SharedNoiseModel& model)
: Base(model, key1, key2),
dP1_(EssentialMatrix::Homogeneous(pA)),
pn_(pB) {
f_ = 1.0;
}
/**
* Constructor
* @param key1 Essential Matrix variable key
* @param key2 Inverse depth variable key
* @param pA point in first camera, in pixel coordinates
* @param pB point in second camera, in pixel coordinates
* @param K calibration object, will be used only in constructor
* @param model noise model should be in pixels, as well
*/
template <class CALIBRATION>
EssentialMatrixFactor2(Key key1, Key key2, const Point2& pA, const Point2& pB,
const SharedNoiseModel& model,
std::shared_ptr<CALIBRATION> K)
: Base(model, key1, key2),
dP1_(EssentialMatrix::Homogeneous(K->calibrate(pA))),
pn_(K->calibrate(pB)) {
f_ = 0.5 * (K->fx() + K->fy());
}
/// @return a deep copy of this factor
gtsam::NonlinearFactor::shared_ptr clone() const override {
return std::static_pointer_cast<gtsam::NonlinearFactor>(
gtsam::NonlinearFactor::shared_ptr(new This(*this)));
}
/// print
void print(
const std::string& s = "",
const KeyFormatter& keyFormatter = DefaultKeyFormatter) const override {
Base::print(s);
std::cout << " EssentialMatrixFactor2 with measurements\n ("
<< dP1_.transpose() << ")' and (" << pn_.transpose() << ")'"
<< std::endl;
}
/*
* Vector of errors returns 2D vector
* @param E essential matrix
* @param d inverse depth d
*/
Vector evaluateError(const EssentialMatrix& E, const double& d,
OptionalMatrixType DE,
OptionalMatrixType Dd) const override {
// We have point x,y in image 1
// Given a depth Z, the corresponding 3D point P1 = Z*(x,y,1) = (x,y,1)/d
// We then convert to second camera by P2 = 1R2'*(P1-1T2)
// The homogeneous coordinates of can be written as
// 2R1*(P1-1T2) == 2R1*d*(P1-1T2) == 2R1*((x,y,1)-d*1T2)
// where we multiplied with d which yields equivalent homogeneous
// coordinates. Note that this is just the homography 2R1 for d==0 The point
// d*P1 = (x,y,1) is computed in constructor as dP1_
// Project to normalized image coordinates, then uncalibrate
Point2 pn(0, 0);
if (!DE && !Dd) {
Point3 _1T2 = E.direction().point3();
Point3 d1T2 = d * _1T2;
Point3 dP2 = E.rotation().unrotate(dP1_ - d1T2); // 2R1*((x,y,1)-d*1T2)
pn = PinholeBase::Project(dP2);
} else {
// Calculate derivatives. TODO if slow: optimize with Mathematica
// 3*2 3*3 3*3
Matrix D_1T2_dir, DdP2_rot, DP2_point;
Point3 _1T2 = E.direction().point3(D_1T2_dir);
Point3 d1T2 = d * _1T2;
Point3 dP2 = E.rotation().unrotate(dP1_ - d1T2, DdP2_rot, DP2_point);
Matrix23 Dpn_dP2;
pn = PinholeBase::Project(dP2, Dpn_dP2);
if (DE) {
Matrix DdP2_E(3, 5);
DdP2_E << DdP2_rot, -DP2_point * d * D_1T2_dir; // (3*3), (3*3) * (3*2)
*DE = f_ * Dpn_dP2 * DdP2_E; // (2*3) * (3*5)
}
if (Dd) // efficient backwards computation:
// (2*3) * (3*3) * (3*1)
*Dd = -f_ * (Dpn_dP2 * (DP2_point * _1T2));
}
Point2 reprojectionError = pn - pn_;
return f_ * reprojectionError;
}
public:
GTSAM_MAKE_ALIGNED_OPERATOR_NEW
};
// EssentialMatrixFactor2
/**
* Binary factor that optimizes for E and inverse depth d: assumes measurement
* in image 2 is perfect, and returns re-projection error in image 1
* This version takes an extrinsic rotation to allow for omni-directional rigs
*/
class EssentialMatrixFactor3 : public EssentialMatrixFactor2 {
typedef EssentialMatrixFactor2 Base;
typedef EssentialMatrixFactor3 This;
Rot3 cRb_; ///< Rotation from body to camera frame
public:
// Provide access to the Matrix& version of evaluateError:
using Base::evaluateError;
/**
* Constructor
* @param key1 Essential Matrix variable key
* @param key2 Inverse depth variable key
* @param pA point in first camera, in calibrated coordinates
* @param pB point in second camera, in calibrated coordinates
* @param bRc extra rotation between "body" and "camera" frame
* @param model noise model should be in calibrated coordinates, as well
*/
EssentialMatrixFactor3(Key key1, Key key2, const Point2& pA, const Point2& pB,
const Rot3& cRb, const SharedNoiseModel& model)
: EssentialMatrixFactor2(key1, key2, pA, pB, model), cRb_(cRb) {}
/**
* Constructor
* @param key1 Essential Matrix variable key
* @param key2 Inverse depth variable key
* @param pA point in first camera, in pixel coordinates
* @param pB point in second camera, in pixel coordinates
* @param K calibration object, will be used only in constructor
* @param model noise model should be in pixels, as well
*/
template <class CALIBRATION>
EssentialMatrixFactor3(Key key1, Key key2, const Point2& pA, const Point2& pB,
const Rot3& cRb, const SharedNoiseModel& model,
std::shared_ptr<CALIBRATION> K)
: EssentialMatrixFactor2(key1, key2, pA, pB, model, K), cRb_(cRb) {}
/// @return a deep copy of this factor
gtsam::NonlinearFactor::shared_ptr clone() const override {
return std::static_pointer_cast<gtsam::NonlinearFactor>(
gtsam::NonlinearFactor::shared_ptr(new This(*this)));
}
/// print
void print(
const std::string& s = "",
const KeyFormatter& keyFormatter = DefaultKeyFormatter) const override {
Base::print(s);
std::cout << " EssentialMatrixFactor3 with rotation " << cRb_ << std::endl;
}
/*
* Vector of errors returns 2D vector
* @param E essential matrix
* @param d inverse depth d
*/
Vector evaluateError(const EssentialMatrix& E, const double& d,
OptionalMatrixType DE,
OptionalMatrixType Dd) const override {
if (!DE) {
// Convert E from body to camera frame
EssentialMatrix cameraE = cRb_ * E;
// Evaluate error
return Base::evaluateError(cameraE, d, OptionalNone, Dd);
} else {
// Version with derivatives
Matrix D_e_cameraE, D_cameraE_E; // 2*5, 5*5
EssentialMatrix cameraE = E.rotate(cRb_, D_cameraE_E);
// Using the pointer version of evaluateError since the Base class
// (EssentialMatrixFactor2) does not have the matrix reference version of
// evaluateError
Vector e = Base::evaluateError(cameraE, d, &D_e_cameraE, Dd);
*DE = D_e_cameraE * D_cameraE_E; // (2*5) * (5*5)
return e;
}
}
public:
GTSAM_MAKE_ALIGNED_OPERATOR_NEW
};
// EssentialMatrixFactor3
/**
* Binary factor that optimizes for E and calibration K using the algebraic
* epipolar error (K^-1 pA)'E (K^-1 pB). The calibration is shared between two
* images.
*
* Note: As correspondences between 2d coordinates can only recover 7 DoF,
* this factor should always be used with a prior factor on calibration.
* Even with a prior, we can only optimize 2 DoF in the calibration. So the
* prior should have a noise model with very low sigma in the remaining
* dimensions. This has been tested to work on Cal3_S2. With Cal3Bundler, it
* works only with a strong prior (low sigma noise model) on all degrees of
* freedom.
*/
template <class CALIBRATION>
class EssentialMatrixFactor4
: public NoiseModelFactorN<EssentialMatrix, CALIBRATION> {
private:
Point2 pA_, pB_; ///< points in pixel coordinates
typedef NoiseModelFactorN<EssentialMatrix, CALIBRATION> Base;
typedef EssentialMatrixFactor4 This;
static constexpr int DimK = FixedDimension<CALIBRATION>::value;
typedef Eigen::Matrix<double, 2, DimK> JacobianCalibration;
public:
// Provide access to the Matrix& version of evaluateError:
using Base::evaluateError;
/**
* Constructor
* @param keyE Essential Matrix aEb variable key
* @param keyK Calibration variable key (common for both cameras)
* @param pA point in first camera, in pixel coordinates
* @param pB point in second camera, in pixel coordinates
* @param model noise model is about dot product in ideal, homogeneous
* coordinates
*/
EssentialMatrixFactor4(Key keyE, Key keyK, const Point2& pA, const Point2& pB,
const SharedNoiseModel& model = nullptr)
: Base(model, keyE, keyK), pA_(pA), pB_(pB) {}
/// @return a deep copy of this factor
gtsam::NonlinearFactor::shared_ptr clone() const override {
return std::static_pointer_cast<gtsam::NonlinearFactor>(
gtsam::NonlinearFactor::shared_ptr(new This(*this)));
}
/// print
void print(
const std::string& s = "",
const KeyFormatter& keyFormatter = DefaultKeyFormatter) const override {
Base::print(s);
std::cout << " EssentialMatrixFactor4 with measurements\n ("
<< pA_.transpose() << ")' and (" << pB_.transpose() << ")'"
<< std::endl;
}
/**
* @brief Calculate the algebraic epipolar error pA' (K^-1)' E K pB.
*
* @param E essential matrix for key keyE
* @param K calibration (common for both images) for key keyK
* @param H1 optional jacobian of error w.r.t E
* @param H2 optional jacobian of error w.r.t K
* @return * Vector 1D vector of algebraic error
*/
Vector evaluateError(const EssentialMatrix& E, const CALIBRATION& K,
OptionalMatrixType HE,
OptionalMatrixType HK) const override {
// converting from pixel coordinates to normalized coordinates cA and cB
JacobianCalibration cA_H_K; // dcA/dK
JacobianCalibration cB_H_K; // dcB/dK
Point2 cA = K.calibrate(pA_, HK ? &cA_H_K : 0, OptionalNone);
Point2 cB = K.calibrate(pB_, HK ? &cB_H_K : 0, OptionalNone);
// convert to homogeneous coordinates
Vector3 vA = EssentialMatrix::Homogeneous(cA);
Vector3 vB = EssentialMatrix::Homogeneous(cB);
if (HK) {
// compute the jacobian of error w.r.t K
// error function f = vA.T * E * vB
// H2 = df/dK = vB.T * E.T * dvA/dK + vA.T * E * dvB/dK
// where dvA/dK = dvA/dcA * dcA/dK, dVB/dK = dvB/dcB * dcB/dK
// and dvA/dcA = dvB/dcB = [[1, 0], [0, 1], [0, 0]]
*HK = vB.transpose() * E.matrix().transpose().leftCols<2>() * cA_H_K +
vA.transpose() * E.matrix().leftCols<2>() * cB_H_K;
}
Vector error(1);
error << E.error(vA, vB, HE);
return error;
}
public:
GTSAM_MAKE_ALIGNED_OPERATOR_NEW
};
// EssentialMatrixFactor4
/**
* Binary factor that optimizes for E and two calibrations Ka and Kb using the
* algebraic epipolar error (Ka^-1 pA)'E (Kb^-1 pB). The calibrations are
* assumed different for the two images, but if you use the same key for Ka and
* Kb, the sum of the two K Jacobians equals that of the K Jacobian for
* EssentialMatrix4. If you know there is a single global calibration, use
* that factor instead.
*
* Note: see the comment about priors from EssentialMatrixFactor4: even stronger
* caveats about having priors on calibration apply here.
*/
template <class CALIBRATION>
class EssentialMatrixFactor5
: public NoiseModelFactorN<EssentialMatrix, CALIBRATION, CALIBRATION> {
private:
Point2 pA_, pB_; ///< points in pixel coordinates
typedef NoiseModelFactorN<EssentialMatrix, CALIBRATION, CALIBRATION> Base;
typedef EssentialMatrixFactor5 This;
static constexpr int DimK = FixedDimension<CALIBRATION>::value;
typedef Eigen::Matrix<double, 2, DimK> JacobianCalibration;
public:
// Provide access to the Matrix& version of evaluateError:
using Base::evaluateError;
/**
* Constructor
* @param keyE Essential Matrix aEb variable key
* @param keyKa Calibration variable key for camera A
* @param keyKb Calibration variable key for camera B
* @param pA point in first camera, in pixel coordinates
* @param pB point in second camera, in pixel coordinates
* @param model noise model is about dot product in ideal, homogeneous
* coordinates
*/
EssentialMatrixFactor5(Key keyE, Key keyKa, Key keyKb, const Point2& pA,
const Point2& pB,
const SharedNoiseModel& model = nullptr)
: Base(model, keyE, keyKa, keyKb), pA_(pA), pB_(pB) {}
/// @return a deep copy of this factor
gtsam::NonlinearFactor::shared_ptr clone() const override {
return std::static_pointer_cast<gtsam::NonlinearFactor>(
gtsam::NonlinearFactor::shared_ptr(new This(*this)));
}
/// print
void print(
const std::string& s = "",
const KeyFormatter& keyFormatter = DefaultKeyFormatter) const override {
Base::print(s);
std::cout << " EssentialMatrixFactor5 with measurements\n ("
<< pA_.transpose() << ")' and (" << pB_.transpose() << ")'"
<< std::endl;
}
/**
* @brief Calculate the algebraic epipolar error pA' (Ka^-1)' E Kb pB.
*
* @param E essential matrix for key keyE
* @param Ka calibration for camera A for key keyKa
* @param Kb calibration for camera B for key keyKb
* @param H1 optional jacobian of error w.r.t E
* @param H2 optional jacobian of error w.r.t Ka
* @param H3 optional jacobian of error w.r.t Kb
* @return * Vector 1D vector of algebraic error
*/
Vector evaluateError(const EssentialMatrix& E, const CALIBRATION& Ka,
const CALIBRATION& Kb, OptionalMatrixType HE,
OptionalMatrixType HKa,
OptionalMatrixType HKb) const override {
// converting from pixel coordinates to normalized coordinates cA and cB
JacobianCalibration cA_H_Ka; // dcA/dKa
JacobianCalibration cB_H_Kb; // dcB/dKb
Point2 cA = Ka.calibrate(pA_, HKa ? &cA_H_Ka : 0, OptionalNone);
Point2 cB = Kb.calibrate(pB_, HKb ? &cB_H_Kb : 0, OptionalNone);
// Convert to homogeneous coordinates.
Vector3 vA = EssentialMatrix::Homogeneous(cA);
Vector3 vB = EssentialMatrix::Homogeneous(cB);
if (HKa) {
// Compute the jacobian of error w.r.t Ka.
*HKa = vB.transpose() * E.matrix().transpose().leftCols<2>() * cA_H_Ka;
}
if (HKb) {
// Compute the jacobian of error w.r.t Kb.
*HKb = vA.transpose() * E.matrix().leftCols<2>() * cB_H_Kb;
}
Vector error(1);
error << E.error(vA, vB, HE);
return error;
}
public:
GTSAM_MAKE_ALIGNED_OPERATOR_NEW
};
// EssentialMatrixFactor5
} // namespace gtsam