| 1 | /* ---------------------------------------------------------------------------- |
| 2 | |
| 3 | * GTSAM Copyright 2010, Georgia Tech Research Corporation, |
| 4 | * Atlanta, Georgia 30332-0415 |
| 5 | * All Rights Reserved |
| 6 | * Authors: Frank Dellaert, et al. (see THANKS for the full author list) |
| 7 | |
| 8 | * See LICENSE for the license information |
| 9 | |
| 10 | * -------------------------------------------------------------------------- */ |
| 11 | |
| 12 | /** |
| 13 | * @file ConcurrentCalibration.cpp |
| 14 | * @brief First step towards estimating monocular calibration in concurrent |
| 15 | * filter/smoother framework. To start with, just batch LM. |
| 16 | * @date June 11, 2014 |
| 17 | * @author Chris Beall |
| 18 | */ |
| 19 | |
| 20 | |
| 21 | #include <gtsam/geometry/Pose3.h> |
| 22 | #include <gtsam/nonlinear/Values.h> |
| 23 | #include <gtsam/nonlinear/utilities.h> |
| 24 | #include <gtsam/nonlinear/NonlinearEquality.h> |
| 25 | #include <gtsam/nonlinear/NonlinearFactorGraph.h> |
| 26 | #include <gtsam/nonlinear/LevenbergMarquardtOptimizer.h> |
| 27 | #include <gtsam/inference/Symbol.h> |
| 28 | #include <gtsam/slam/ProjectionFactor.h> |
| 29 | #include <gtsam/slam/GeneralSFMFactor.h> |
| 30 | #include <gtsam/slam/dataset.h> |
| 31 | |
| 32 | #include <string> |
| 33 | #include <fstream> |
| 34 | #include <iostream> |
| 35 | |
| 36 | using namespace std; |
| 37 | using namespace gtsam; |
| 38 | using symbol_shorthand::K; |
| 39 | using symbol_shorthand::L; |
| 40 | using symbol_shorthand::X; |
| 41 | |
| 42 | int main(int argc, char** argv){ |
| 43 | |
| 44 | Values initial_estimate; |
| 45 | NonlinearFactorGraph graph; |
| 46 | const auto model = noiseModel::Isotropic::Sigma(dim: 2,sigma: 1); |
| 47 | |
| 48 | string calibration_loc = findExampleDataFile(name: "VO_calibration00s.txt" ); |
| 49 | string pose_loc = findExampleDataFile(name: "VO_camera_poses00s.txt" ); |
| 50 | string factor_loc = findExampleDataFile(name: "VO_stereo_factors00s.txt" ); |
| 51 | |
| 52 | //read camera calibration info from file |
| 53 | // focal lengths fx, fy, skew s, principal point u0, v0, baseline b |
| 54 | double fx, fy, s, u0, v0, b; |
| 55 | ifstream calibration_file(calibration_loc.c_str()); |
| 56 | cout << "Reading calibration info" << endl; |
| 57 | calibration_file >> fx >> fy >> s >> u0 >> v0 >> b; |
| 58 | |
| 59 | //create stereo camera calibration object |
| 60 | const Cal3_S2 true_K(fx,fy,s,u0,v0); |
| 61 | const Cal3_S2 noisy_K(fx*1.2,fy*1.2,s,u0-10,v0+10); |
| 62 | |
| 63 | initial_estimate.insert(j: K(j: 0), val: noisy_K); |
| 64 | |
| 65 | auto calNoise = noiseModel::Diagonal::Sigmas(sigmas: (Vector(5) << 500, 500, 1e-5, 100, 100).finished()); |
| 66 | graph.addPrior(key: K(j: 0), prior: noisy_K, model: calNoise); |
| 67 | |
| 68 | |
| 69 | ifstream pose_file(pose_loc.c_str()); |
| 70 | cout << "Reading camera poses" << endl; |
| 71 | int pose_id; |
| 72 | MatrixRowMajor m(4,4); |
| 73 | //read camera pose parameters and use to make initial estimates of camera poses |
| 74 | while (pose_file >> pose_id) { |
| 75 | for (int i = 0; i < 16; i++) { |
| 76 | pose_file >> m.data()[i]; |
| 77 | } |
| 78 | initial_estimate.insert(j: Symbol('x', pose_id), val: Pose3(m)); |
| 79 | } |
| 80 | |
| 81 | auto poseNoise = noiseModel::Isotropic::Sigma(dim: 6, sigma: 0.01); |
| 82 | graph.addPrior(key: Symbol('x', pose_id), prior: Pose3(m), model: poseNoise); |
| 83 | |
| 84 | // camera and landmark keys |
| 85 | size_t x, l; |
| 86 | |
| 87 | // pixel coordinates uL, uR, v (same for left/right images due to rectification) |
| 88 | // landmark coordinates X, Y, Z in camera frame, resulting from triangulation |
| 89 | double uL, uR, v, _X, Y, Z; |
| 90 | ifstream factor_file(factor_loc.c_str()); |
| 91 | cout << "Reading stereo factors" << endl; |
| 92 | //read stereo measurement details from file and use to create and add GenericStereoFactor objects to the graph representation |
| 93 | while (factor_file >> x >> l >> uL >> uR >> v >> _X >> Y >> Z) { |
| 94 | // graph.emplace_shared<GenericStereoFactor<Pose3, Point3> >(StereoPoint2(uL, uR, v), model, X(x), L(l), K); |
| 95 | |
| 96 | graph.emplace_shared<GeneralSFMFactor2<Cal3_S2> >(args: Point2(uL,v), args: model, args: X(j: x), args: L(j: l), args: K(j: 0)); |
| 97 | |
| 98 | |
| 99 | //if the landmark variable included in this factor has not yet been added to the initial variable value estimate, add it |
| 100 | if (!initial_estimate.exists(j: L(j: l))) { |
| 101 | Pose3 camPose = initial_estimate.at<Pose3>(j: X(j: x)); |
| 102 | //transformFrom() transforms the input Point3 from the camera pose space, camPose, to the global space |
| 103 | Point3 worldPoint = camPose.transformFrom(point: Point3(_X, Y, Z)); |
| 104 | initial_estimate.insert(j: L(j: l), val: worldPoint); |
| 105 | } |
| 106 | } |
| 107 | |
| 108 | Pose3 first_pose = initial_estimate.at<Pose3>(j: Symbol('x',1)); |
| 109 | //constrain the first pose such that it cannot change from its original value during optimization |
| 110 | // NOTE: NonlinearEquality forces the optimizer to use QR rather than Cholesky |
| 111 | // QR is much slower than Cholesky, but numerically more stable |
| 112 | graph.emplace_shared<NonlinearEquality<Pose3> >(args: Symbol('x',1),args&: first_pose); |
| 113 | |
| 114 | cout << "Optimizing" << endl; |
| 115 | LevenbergMarquardtParams params; |
| 116 | params.verbosityLM = LevenbergMarquardtParams::TRYLAMBDA; |
| 117 | params.verbosity = NonlinearOptimizerParams::ERROR; |
| 118 | |
| 119 | //create Levenberg-Marquardt optimizer to optimize the factor graph |
| 120 | LevenbergMarquardtOptimizer optimizer(graph, initial_estimate,params); |
| 121 | // Values result = optimizer.optimize(); |
| 122 | |
| 123 | string K_values_file = "K_values.txt" ; |
| 124 | ofstream stream_K(K_values_file.c_str()); |
| 125 | |
| 126 | double currentError; |
| 127 | |
| 128 | |
| 129 | stream_K << optimizer.iterations() << " " << optimizer.values().at<Cal3_S2>(j: K(j: 0)).vector().transpose() << endl; |
| 130 | |
| 131 | |
| 132 | // Iterative loop |
| 133 | do { |
| 134 | // Do next iteration |
| 135 | currentError = optimizer.error(); |
| 136 | optimizer.iterate(); |
| 137 | |
| 138 | stream_K << optimizer.iterations() << " " << optimizer.values().at<Cal3_S2>(j: K(j: 0)).vector().transpose() << endl; |
| 139 | |
| 140 | if(params.verbosity >= NonlinearOptimizerParams::ERROR) cout << "newError: " << optimizer.error() << endl; |
| 141 | } while(optimizer.iterations() < params.maxIterations && |
| 142 | !checkConvergence(relativeErrorTreshold: params.relativeErrorTol, absoluteErrorTreshold: params.absoluteErrorTol, |
| 143 | errorThreshold: params.errorTol, currentError, newError: optimizer.error(), verbosity: params.verbosity)); |
| 144 | |
| 145 | Values result = optimizer.values(); |
| 146 | |
| 147 | cout << "Final result sample:" << endl; |
| 148 | Values pose_values = utilities::allPose3s(values: result); |
| 149 | pose_values.print(str: "Final camera poses:\n" ); |
| 150 | |
| 151 | result.at<Cal3_S2>(j: K(j: 0)).print(s: "Final K\n" ); |
| 152 | |
| 153 | noisy_K.print(s: "Initial noisy K\n" ); |
| 154 | true_K.print(s: "Initial correct K\n" ); |
| 155 | |
| 156 | return 0; |
| 157 | } |
| 158 | |