gtsam/gtsam/sfm/ShonanAveraging.h

446 lines
15 KiB
C++

/* ----------------------------------------------------------------------------
* GTSAM Copyright 2010-2019, 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 ShonanAveraging.h
* @date March 2019 - August 2020
* @author Frank Dellaert, David Rosen, and Jing Wu
* @brief Shonan Averaging algorithm
*/
#pragma once
#include <gtsam/base/Matrix.h>
#include <gtsam/base/Vector.h>
#include <gtsam/dllexport.h>
#include <gtsam/geometry/Rot2.h>
#include <gtsam/geometry/Rot3.h>
#include <gtsam/linear/VectorValues.h>
#include <gtsam/nonlinear/LevenbergMarquardtParams.h>
#include <gtsam/sfm/BinaryMeasurement.h>
#include <gtsam/linear/PowerMethod.h>
#include <gtsam/linear/AcceleratedPowerMethod.h>
#include <gtsam/slam/dataset.h>
#include <Eigen/Sparse>
#include <map>
#include <string>
#include <type_traits>
#include <utility>
#include <vector>
namespace gtsam {
class NonlinearFactorGraph;
class LevenbergMarquardtOptimizer;
/// Parameters governing optimization etc.
template <size_t d>
struct GTSAM_EXPORT ShonanAveragingParameters {
// Select Rot2 or Rot3 interface based template parameter d
using Rot = typename std::conditional<d == 2, Rot2, Rot3>::type;
using Anchor = std::pair<size_t, Rot>;
// Parameters themselves:
LevenbergMarquardtParams lm; ///< LM parameters
double optimalityThreshold; ///< threshold used in checkOptimality
Anchor anchor; ///< pose to use as anchor if not Karcher
double alpha; ///< weight of anchor-based prior (default 0)
double beta; ///< weight of Karcher-based prior (default 1)
double gamma; ///< weight of gauge-fixing factors (default 0)
/// if enabled, the Huber loss is used (default false)
bool useHuber;
/// if enabled solution optimality is certified (default true)
bool certifyOptimality;
ShonanAveragingParameters(const LevenbergMarquardtParams &lm =
LevenbergMarquardtParams::CeresDefaults(),
const std::string &method = "JACOBI",
double optimalityThreshold = -1e-4,
double alpha = 0.0, double beta = 1.0,
double gamma = 0.0);
LevenbergMarquardtParams getLMParams() const { return lm; }
void setOptimalityThreshold(double value) { optimalityThreshold = value; }
double getOptimalityThreshold() const { return optimalityThreshold; }
void setAnchor(size_t index, const Rot &value) { anchor = {index, value}; }
std::pair<size_t, Rot> getAnchor() const { return anchor; }
void setAnchorWeight(double value) { alpha = value; }
double getAnchorWeight() const { return alpha; }
void setKarcherWeight(double value) { beta = value; }
double getKarcherWeight() const { return beta; }
void setGaugesWeight(double value) { gamma = value; }
double getGaugesWeight() const { return gamma; }
void setUseHuber(bool value) { useHuber = value; }
bool getUseHuber() const { return useHuber; }
void setCertifyOptimality(bool value) { certifyOptimality = value; }
bool getCertifyOptimality() const { return certifyOptimality; }
/// Print the parameters and flags used for rotation averaging.
void print() const {
std::cout << " ShonanAveragingParameters: " << std::endl;
std::cout << " alpha: " << alpha << std::endl;
std::cout << " beta: " << beta << std::endl;
std::cout << " gamma: " << gamma << std::endl;
std::cout << " useHuber: " << useHuber << std::endl;
}
};
using ShonanAveragingParameters2 = ShonanAveragingParameters<2>;
using ShonanAveragingParameters3 = ShonanAveragingParameters<3>;
/**
* Class that implements Shonan Averaging from our ECCV'20 paper.
* Note: The "basic" API uses all Rot values (Rot2 or Rot3, depending on value
* of d), whereas the different levels and "advanced" API at SO(p) needs Values
* of type SOn<Dynamic>.
*
* The template parameter d can be 2 or 3.
* Both are specialized in the .cpp file.
*
* If you use this code in your work, please consider citing our paper:
* Shonan Rotation Averaging, Global Optimality by Surfing SO(p)^n
* Frank Dellaert, David M. Rosen, Jing Wu, Robert Mahony, and Luca Carlone,
* European Computer Vision Conference, 2020.
* You can view our ECCV spotlight video at https://youtu.be/5ppaqMyHtE0
*/
template <size_t d>
class GTSAM_EXPORT ShonanAveraging {
public:
using Sparse = Eigen::SparseMatrix<double>;
// Define the Parameters type and use its typedef of the rotation type:
using Parameters = ShonanAveragingParameters<d>;
using Rot = typename Parameters::Rot;
// We store SO(d) BetweenFactors to get noise model
using Measurements = std::vector<BinaryMeasurement<Rot>>;
private:
Parameters parameters_;
Measurements measurements_;
size_t nrUnknowns_;
Sparse D_; // Sparse (diagonal) degree matrix
Sparse Q_; // Sparse measurement matrix, == \tilde{R} in Eriksson18cvpr
Sparse L_; // connection Laplacian L = D - Q, needed for optimality check
/**
* Build 3Nx3N sparse matrix consisting of rotation measurements, arranged as
* (i,j) and (j,i) blocks within a sparse matrix.
*/
Sparse buildQ() const;
/// Build 3Nx3N sparse degree matrix D
Sparse buildD() const;
public:
/// @name Standard Constructors
/// @{
/// Construct from set of relative measurements (given as BetweenFactor<Rot3>
/// for now) NoiseModel *must* be isotropic.
ShonanAveraging(const Measurements &measurements,
const Parameters &parameters = Parameters());
/// @}
/// @name Query properties
/// @{
/// Return number of unknowns
size_t nrUnknowns() const { return nrUnknowns_; }
/// Return number of measurements
size_t nrMeasurements() const { return measurements_.size(); }
/// k^th binary measurement
const BinaryMeasurement<Rot> &measurement(size_t k) const {
return measurements_[k];
}
/**
* Update factors to use robust Huber loss.
*
* @param measurements Vector of BinaryMeasurements.
* @param k Huber noise model threshold.
*/
Measurements makeNoiseModelRobust(const Measurements &measurements,
double k = 1.345) const {
Measurements robustMeasurements;
for (auto &measurement : measurements) {
auto model = measurement.noiseModel();
const auto &robust =
boost::dynamic_pointer_cast<noiseModel::Robust>(model);
SharedNoiseModel robust_model;
// Check if the noise model is already robust
if (robust) {
robust_model = model;
} else {
// make robust
robust_model = noiseModel::Robust::Create(
noiseModel::mEstimator::Huber::Create(k), model);
}
BinaryMeasurement<Rot> meas(measurement.key1(), measurement.key2(),
measurement.measured(), robust_model);
robustMeasurements.push_back(meas);
}
return robustMeasurements;
}
/// k^th measurement, as a Rot.
const Rot &measured(size_t k) const { return measurements_[k].measured(); }
/// Keys for k^th measurement, as a vector of Key values.
const KeyVector &keys(size_t k) const { return measurements_[k].keys(); }
/// @}
/// @name Matrix API (advanced use, debugging)
/// @{
Sparse D() const { return D_; } ///< Sparse version of D
Matrix denseD() const { return Matrix(D_); } ///< Dense version of D
Sparse Q() const { return Q_; } ///< Sparse version of Q
Matrix denseQ() const { return Matrix(Q_); } ///< Dense version of Q
Sparse L() const { return L_; } ///< Sparse version of L
Matrix denseL() const { return Matrix(L_); } ///< Dense version of L
/// Version that takes pxdN Stiefel manifold elements
Sparse computeLambda(const Matrix &S) const;
/// Dense versions of computeLambda for wrapper/testing
Matrix computeLambda_(const Values &values) const {
return Matrix(computeLambda(values));
}
/// Dense versions of computeLambda for wrapper/testing
Matrix computeLambda_(const Matrix &S) const {
return Matrix(computeLambda(S));
}
/// Compute A matrix whose Eigenvalues we will examine
Sparse computeA(const Values &values) const;
/// Version that takes pxdN Stiefel manifold elements
Sparse computeA(const Matrix &S) const;
/// Dense version of computeA for wrapper/testing
Matrix computeA_(const Values &values) const {
return Matrix(computeA(values));
}
/// Project to pxdN Stiefel manifold
static Matrix StiefelElementMatrix(const Values &values);
/**
* Compute minimum eigenvalue for optimality check.
* @param values: should be of type SOn
*/
double computeMinEigenValue(const Values &values,
Vector *minEigenVector = nullptr) const;
/**
* Compute minimum eigenvalue with accelerated power method.
* @param values: should be of type SOn
*/
double computeMinEigenValueAP(const Values &values,
Vector *minEigenVector = nullptr) const;
/// Project pxdN Stiefel manifold matrix S to Rot3^N
Values roundSolutionS(const Matrix &S) const;
/// Create a VectorValues with eigenvector v_i
static VectorValues TangentVectorValues(size_t p, const Vector &v);
/// Calculate the riemannian gradient of F(values) at values
Matrix riemannianGradient(size_t p, const Values &values) const;
/**
* Lift up the dimension of values in type SO(p-1) with descent direction
* provided by minEigenVector and return new values in type SO(p)
*/
static Values LiftwithDescent(size_t p, const Values &values,
const Vector &minEigenVector);
/**
* Given some values at p-1, return new values at p, by doing a line search
* along the descent direction, computed from the minimum eigenvector at p-1.
* @param values should be of type SO(p-1)
* @param minEigenVector corresponding to minEigenValue at level p-1
* @return values of type SO(p)
*/
Values initializeWithDescent(
size_t p, const Values &values, const Vector &minEigenVector,
double minEigenValue, double gradienTolerance = 1e-2,
double preconditionedGradNormTolerance = 1e-4) const;
/// @}
/// @name Advanced API
/// @{
/**
* Build graph for SO(p)
* @param p the dimensionality of the rotation manifold to optimize over
*/
NonlinearFactorGraph buildGraphAt(size_t p) const;
/**
* Create initial Values of type SO(p)
* @param p the dimensionality of the rotation manifold
*/
Values initializeRandomlyAt(size_t p, std::mt19937 &rng) const;
/// Version of initializeRandomlyAt with fixed random seed.
Values initializeRandomlyAt(size_t p) const;
/**
* Calculate cost for SO(p)
* Values should be of type SO(p)
*/
double costAt(size_t p, const Values &values) const;
/**
* Given an estimated local minimum Yopt for the (possibly lifted)
* relaxation, this function computes and returns the block-diagonal elements
* of the corresponding Lagrange multiplier.
*/
Sparse computeLambda(const Values &values) const;
/**
* Compute minimum eigenvalue for optimality check.
* @param values: should be of type SOn
* @return minEigenVector and minEigenValue
*/
std::pair<double, Vector> computeMinEigenVector(const Values &values) const;
/**
* Check optimality
* @param values: should be of type SOn
*/
bool checkOptimality(const Values &values) const;
/**
* Try to create optimizer at SO(p)
* @param p the dimensionality of the rotation manifold to optimize over
* @param initial initial SO(p) values
* @return lm optimizer
*/
boost::shared_ptr<LevenbergMarquardtOptimizer> createOptimizerAt(
size_t p, const Values &initial) const;
/**
* Try to optimize at SO(p)
* @param p the dimensionality of the rotation manifold to optimize over
* @param initial initial SO(p) values
* @return SO(p) values
*/
Values tryOptimizingAt(size_t p, const Values &initial) const;
/**
* Project from SO(p) to Rot2 or Rot3
* Values should be of type SO(p)
*/
Values projectFrom(size_t p, const Values &values) const;
/**
* Project from SO(p)^N to Rot2^N or Rot3^N
* Values should be of type SO(p)
*/
Values roundSolution(const Values &values) const;
/// Lift Values of type T to SO(p)
template <class T>
static Values LiftTo(size_t p, const Values &values) {
Values result;
for (const auto it : values.filter<T>()) {
result.insert(it.key, SOn::Lift(p, it.value.matrix()));
}
return result;
}
/// @}
/// @name Basic API
/// @{
/**
* Calculate cost for SO(3)
* Values should be of type Rot3
*/
double cost(const Values &values) const;
/**
* Initialize randomly at SO(d)
* @param rng random number generator
* Example:
* std::mt19937 rng(42);
* Values initial = initializeRandomly(rng, p);
*/
Values initializeRandomly(std::mt19937 &rng) const;
/// Random initialization for wrapper, fixed random seed.
Values initializeRandomly() const;
/**
* Optimize at different values of p until convergence.
* @param initial initial Rot3 values
* @param pMin value of p to start Riemanian staircase at (default: d).
* @param pMax maximum value of p to try (default: 10)
* @return (Rot3 values, minimum eigenvalue)
*/
std::pair<Values, double> run(const Values &initialEstimate, size_t pMin = d,
size_t pMax = 10) const;
/// @}
/**
* Helper function to convert measurements to robust noise model
* if flag is set.
*
* @tparam T the type of measurement, e.g. Rot3.
* @param measurements vector of BinaryMeasurements of type T.
* @param useRobustModel flag indicating whether use robust noise model
* instead.
*/
template <typename T>
inline std::vector<BinaryMeasurement<T>> maybeRobust(
const std::vector<BinaryMeasurement<T>> &measurements,
bool useRobustModel = false) const {
return useRobustModel ? makeNoiseModelRobust(measurements) : measurements;
}
};
// Subclasses for d=2 and d=3 that explicitly instantiate, as well as provide a
// convenience interface with file access.
class GTSAM_EXPORT ShonanAveraging2 : public ShonanAveraging<2> {
public:
ShonanAveraging2(const Measurements &measurements,
const Parameters &parameters = Parameters());
explicit ShonanAveraging2(std::string g2oFile,
const Parameters &parameters = Parameters());
};
class GTSAM_EXPORT ShonanAveraging3 : public ShonanAveraging<3> {
public:
ShonanAveraging3(const Measurements &measurements,
const Parameters &parameters = Parameters());
explicit ShonanAveraging3(std::string g2oFile,
const Parameters &parameters = Parameters());
// TODO(frank): Deprecate after we land pybind wrapper
ShonanAveraging3(const BetweenFactorPose3s &factors,
const Parameters &parameters = Parameters());
};
} // namespace gtsam