| 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 SmartProjectionFactor.h |
| 14 | * @brief Smart factor on cameras (pose + calibration) |
| 15 | * @author Luca Carlone |
| 16 | * @author Zsolt Kira |
| 17 | * @author Frank Dellaert |
| 18 | */ |
| 19 | |
| 20 | #pragma once |
| 21 | |
| 22 | #include <gtsam/slam/SmartFactorBase.h> |
| 23 | #include <gtsam/slam/SmartFactorParams.h> |
| 24 | |
| 25 | #include <gtsam/geometry/triangulation.h> |
| 26 | #include <gtsam/inference/Symbol.h> |
| 27 | #include <gtsam/slam/dataset.h> |
| 28 | |
| 29 | #include <optional> |
| 30 | #include <vector> |
| 31 | |
| 32 | namespace gtsam { |
| 33 | |
| 34 | /** |
| 35 | * SmartProjectionFactor: triangulates point and keeps an estimate of it around. |
| 36 | * This factor operates with monocular cameras, where a camera is expected to |
| 37 | * behave like PinholeCamera or PinholePose. This factor is intended |
| 38 | * to be used directly with PinholeCamera, which optimizes the camera pose |
| 39 | * and calibration. This also requires that values contains the involved |
| 40 | * cameras (instead of poses and calibrations separately). |
| 41 | * If the calibration is fixed use SmartProjectionPoseFactor instead! |
| 42 | */ |
| 43 | template<class CAMERA> |
| 44 | class SmartProjectionFactor: public SmartFactorBase<CAMERA> { |
| 45 | |
| 46 | public: |
| 47 | |
| 48 | private: |
| 49 | typedef SmartFactorBase<CAMERA> Base; |
| 50 | typedef SmartProjectionFactor<CAMERA> This; |
| 51 | typedef SmartProjectionFactor<CAMERA> SmartProjectionCameraFactor; |
| 52 | |
| 53 | protected: |
| 54 | |
| 55 | /// @name Parameters |
| 56 | /// @{ |
| 57 | SmartProjectionParams params_; |
| 58 | /// @} |
| 59 | |
| 60 | /// @name Caching triangulation |
| 61 | /// @{ |
| 62 | mutable TriangulationResult result_; ///< result from triangulateSafe |
| 63 | mutable std::vector<Pose3, Eigen::aligned_allocator<Pose3> > |
| 64 | cameraPosesTriangulation_; ///< current triangulation poses |
| 65 | /// @} |
| 66 | |
| 67 | public: |
| 68 | |
| 69 | /// shorthand for a smart pointer to a factor |
| 70 | typedef std::shared_ptr<This> shared_ptr; |
| 71 | |
| 72 | /// shorthand for a set of cameras |
| 73 | typedef CAMERA Camera; |
| 74 | typedef CameraSet<CAMERA> Cameras; |
| 75 | |
| 76 | /** |
| 77 | * Default constructor, only for serialization |
| 78 | */ |
| 79 | SmartProjectionFactor() {} |
| 80 | |
| 81 | /** |
| 82 | * Constructor |
| 83 | * @param sharedNoiseModel isotropic noise model for the 2D feature measurements |
| 84 | * @param params parameters for the smart projection factors |
| 85 | */ |
| 86 | SmartProjectionFactor( |
| 87 | const SharedNoiseModel& sharedNoiseModel, |
| 88 | const SmartProjectionParams& params = SmartProjectionParams()) |
| 89 | : Base(sharedNoiseModel), |
| 90 | params_(params), |
| 91 | result_(TriangulationResult::Degenerate()) {} |
| 92 | |
| 93 | /** Virtual destructor */ |
| 94 | ~SmartProjectionFactor() override { |
| 95 | } |
| 96 | |
| 97 | /** |
| 98 | * print |
| 99 | * @param s optional string naming the factor |
| 100 | * @param keyFormatter optional formatter useful for printing Symbols |
| 101 | */ |
| 102 | void print(const std::string& s = "" , const KeyFormatter& keyFormatter = |
| 103 | DefaultKeyFormatter) const override { |
| 104 | std::cout << s << "SmartProjectionFactor\n" ; |
| 105 | std::cout << "linearizationMode: " << params_.linearizationMode |
| 106 | << std::endl; |
| 107 | std::cout << "triangulationParameters:\n" << params_.triangulation |
| 108 | << std::endl; |
| 109 | std::cout << "result:\n" << result_ << std::endl; |
| 110 | Base::print("" , keyFormatter); |
| 111 | } |
| 112 | |
| 113 | /// equals |
| 114 | bool equals(const NonlinearFactor& p, double tol = 1e-9) const override { |
| 115 | const This *e = dynamic_cast<const This*>(&p); |
| 116 | return e && params_.linearizationMode == e->params_.linearizationMode |
| 117 | && Base::equals(p, tol); |
| 118 | } |
| 119 | |
| 120 | /** |
| 121 | * @brief Check if the new linearization point is the same as the one used for |
| 122 | * previous triangulation. |
| 123 | * |
| 124 | * @param cameras |
| 125 | * @return true if we need to re-triangulate. |
| 126 | */ |
| 127 | bool decideIfTriangulate(const Cameras& cameras) const { |
| 128 | // Several calls to linearize will be done from the same linearization |
| 129 | // point, hence it is not needed to re-triangulate. Note that this is not |
| 130 | // yet "selecting linearization", that will come later, and we only check if |
| 131 | // the current linearization is the "same" (up to tolerance) w.r.t. the last |
| 132 | // time we triangulated the point. |
| 133 | |
| 134 | size_t m = cameras.size(); |
| 135 | |
| 136 | bool retriangulate = false; |
| 137 | |
| 138 | // Definitely true if we do not have a previous linearization point or the |
| 139 | // new linearization point includes more poses. |
| 140 | if (cameraPosesTriangulation_.empty() |
| 141 | || cameras.size() != cameraPosesTriangulation_.size()) |
| 142 | retriangulate = true; |
| 143 | |
| 144 | // Otherwise, check poses against cache. |
| 145 | if (!retriangulate) { |
| 146 | for (size_t i = 0; i < cameras.size(); i++) { |
| 147 | if (!cameras[i].pose().equals(cameraPosesTriangulation_[i], |
| 148 | params_.retriangulationThreshold)) { |
| 149 | retriangulate = true; // at least two poses are different, hence we retriangulate |
| 150 | break; |
| 151 | } |
| 152 | } |
| 153 | } |
| 154 | |
| 155 | // Store the current poses used for triangulation if we will re-triangulate. |
| 156 | if (retriangulate) { |
| 157 | cameraPosesTriangulation_.clear(); |
| 158 | cameraPosesTriangulation_.reserve(n: m); |
| 159 | for (size_t i = 0; i < m; i++) |
| 160 | // cameraPosesTriangulation_[i] = cameras[i].pose(); |
| 161 | cameraPosesTriangulation_.push_back(cameras[i].pose()); |
| 162 | } |
| 163 | |
| 164 | return retriangulate; |
| 165 | } |
| 166 | |
| 167 | /** |
| 168 | * @brief Call gtsam::triangulateSafe iff we need to re-triangulate. |
| 169 | * |
| 170 | * @param cameras |
| 171 | * @return TriangulationResult |
| 172 | */ |
| 173 | TriangulationResult triangulateSafe(const Cameras& cameras) const { |
| 174 | |
| 175 | size_t m = cameras.size(); |
| 176 | if (m < 2) // if we have a single pose the corresponding factor is uninformative |
| 177 | return TriangulationResult::Degenerate(); |
| 178 | |
| 179 | bool retriangulate = decideIfTriangulate(cameras); |
| 180 | if (retriangulate) |
| 181 | result_ = gtsam::triangulateSafe(cameras, this->measured_, |
| 182 | params_.triangulation); |
| 183 | return result_; |
| 184 | } |
| 185 | |
| 186 | /** |
| 187 | * @brief Possibly re-triangulate before calculating Jacobians. |
| 188 | * |
| 189 | * @param cameras |
| 190 | * @return true if we could safely triangulate |
| 191 | */ |
| 192 | bool triangulateForLinearize(const Cameras& cameras) const { |
| 193 | triangulateSafe(cameras); // imperative, might reset result_ |
| 194 | return bool(result_); |
| 195 | } |
| 196 | |
| 197 | /// Create a Hessianfactor that is an approximation of error(p). |
| 198 | std::shared_ptr<RegularHessianFactor<Base::Dim> > createHessianFactor( |
| 199 | const Cameras& cameras, const double lambda = 0.0, |
| 200 | bool diagonalDamping = false) const { |
| 201 | size_t numKeys = this->keys_.size(); |
| 202 | // Create structures for Hessian Factors |
| 203 | KeyVector js; |
| 204 | std::vector<Matrix> Gs(numKeys * (numKeys + 1) / 2); |
| 205 | std::vector<Vector> gs(numKeys); |
| 206 | |
| 207 | if (this->measured_.size() != cameras.size()) |
| 208 | throw std::runtime_error( |
| 209 | "SmartProjectionHessianFactor: this->measured_" |
| 210 | ".size() inconsistent with input" ); |
| 211 | |
| 212 | triangulateSafe(cameras); |
| 213 | |
| 214 | if (params_.degeneracyMode == ZERO_ON_DEGENERACY && !result_) { |
| 215 | // failed: return"empty" Hessian |
| 216 | for (Matrix& m : Gs) m = Matrix::Zero(Base::Dim, Base::Dim); |
| 217 | for (Vector& v : gs) v = Vector::Zero(Base::Dim); |
| 218 | return std::make_shared<RegularHessianFactor<Base::Dim> >(this->keys_, |
| 219 | Gs, gs, 0.0); |
| 220 | } |
| 221 | |
| 222 | // Jacobian could be 3D Point3 OR 2D Unit3, difference is E.cols(). |
| 223 | typename Base::FBlocks Fs; |
| 224 | Matrix E; |
| 225 | Vector b; |
| 226 | computeJacobiansWithTriangulatedPoint(Fs, E, b, cameras); |
| 227 | |
| 228 | // Whiten using noise model |
| 229 | Base::whitenJacobians(Fs, E, b); |
| 230 | |
| 231 | // build augmented hessian |
| 232 | SymmetricBlockMatrix augmentedHessian = // |
| 233 | Cameras::SchurComplement(Fs, E, b, lambda, diagonalDamping); |
| 234 | |
| 235 | return std::make_shared<RegularHessianFactor<Base::Dim> >( |
| 236 | this->keys_, augmentedHessian); |
| 237 | } |
| 238 | |
| 239 | // Create RegularImplicitSchurFactor factor. |
| 240 | std::shared_ptr<RegularImplicitSchurFactor<CAMERA> > createRegularImplicitSchurFactor( |
| 241 | const Cameras& cameras, double lambda) const { |
| 242 | if (triangulateForLinearize(cameras)) |
| 243 | return Base::createRegularImplicitSchurFactor(cameras, *result_, lambda); |
| 244 | else |
| 245 | // failed: return empty |
| 246 | return std::shared_ptr<RegularImplicitSchurFactor<CAMERA> >(); |
| 247 | } |
| 248 | |
| 249 | /// Create JacobianFactorQ factor. |
| 250 | std::shared_ptr<JacobianFactorQ<Base::Dim, 2> > createJacobianQFactor( |
| 251 | const Cameras& cameras, double lambda) const { |
| 252 | if (triangulateForLinearize(cameras)) |
| 253 | return Base::createJacobianQFactor(cameras, *result_, lambda); |
| 254 | else |
| 255 | // failed: return empty |
| 256 | return std::make_shared<JacobianFactorQ<Base::Dim, 2> >(this->keys_); |
| 257 | } |
| 258 | |
| 259 | /// Create JacobianFactorQ factor, takes values. |
| 260 | std::shared_ptr<JacobianFactorQ<Base::Dim, 2> > createJacobianQFactor( |
| 261 | const Values& values, double lambda) const { |
| 262 | return createJacobianQFactor(this->cameras(values), lambda); |
| 263 | } |
| 264 | |
| 265 | /// Different (faster) way to compute a JacobianFactorSVD factor. |
| 266 | std::shared_ptr<JacobianFactor> createJacobianSVDFactor( |
| 267 | const Cameras& cameras, double lambda) const { |
| 268 | if (triangulateForLinearize(cameras)) |
| 269 | return Base::createJacobianSVDFactor(cameras, *result_, lambda); |
| 270 | else |
| 271 | // failed: return empty |
| 272 | return std::make_shared<JacobianFactorSVD<Base::Dim, 2> >(this->keys_); |
| 273 | } |
| 274 | |
| 275 | /// Linearize to a Hessianfactor. |
| 276 | virtual std::shared_ptr<RegularHessianFactor<Base::Dim> > linearizeToHessian( |
| 277 | const Values& values, double lambda = 0.0) const { |
| 278 | return createHessianFactor(cameras: this->cameras(values), lambda); |
| 279 | } |
| 280 | |
| 281 | /// Linearize to an Implicit Schur factor. |
| 282 | virtual std::shared_ptr<RegularImplicitSchurFactor<CAMERA> > linearizeToImplicit( |
| 283 | const Values& values, double lambda = 0.0) const { |
| 284 | return createRegularImplicitSchurFactor(cameras: this->cameras(values), lambda); |
| 285 | } |
| 286 | |
| 287 | /// Linearize to a JacobianfactorQ. |
| 288 | virtual std::shared_ptr<JacobianFactorQ<Base::Dim, 2> > linearizeToJacobian( |
| 289 | const Values& values, double lambda = 0.0) const { |
| 290 | return createJacobianQFactor(this->cameras(values), lambda); |
| 291 | } |
| 292 | |
| 293 | /** |
| 294 | * Linearize to Gaussian Factor |
| 295 | * @param values Values structure which must contain camera poses for this factor |
| 296 | * @return a Gaussian factor |
| 297 | */ |
| 298 | std::shared_ptr<GaussianFactor> linearizeDamped(const Cameras& cameras, |
| 299 | const double lambda = 0.0) const { |
| 300 | // depending on flag set on construction we may linearize to different linear factors |
| 301 | switch (params_.linearizationMode) { |
| 302 | case HESSIAN: |
| 303 | return createHessianFactor(cameras, lambda); |
| 304 | case IMPLICIT_SCHUR: |
| 305 | return createRegularImplicitSchurFactor(cameras, lambda); |
| 306 | case JACOBIAN_SVD: |
| 307 | return createJacobianSVDFactor(cameras, lambda); |
| 308 | case JACOBIAN_Q: |
| 309 | return createJacobianQFactor(cameras, lambda); |
| 310 | default: |
| 311 | throw std::runtime_error("SmartFactorlinearize: unknown mode" ); |
| 312 | } |
| 313 | } |
| 314 | |
| 315 | /** |
| 316 | * Linearize to Gaussian Factor |
| 317 | * @param values Values structure which must contain camera poses for this factor |
| 318 | * @return a Gaussian factor |
| 319 | */ |
| 320 | std::shared_ptr<GaussianFactor> linearizeDamped(const Values& values, |
| 321 | const double lambda = 0.0) const { |
| 322 | // depending on flag set on construction we may linearize to different linear factors |
| 323 | Cameras cameras = this->cameras(values); |
| 324 | return linearizeDamped(cameras, lambda); |
| 325 | } |
| 326 | |
| 327 | /// linearize |
| 328 | std::shared_ptr<GaussianFactor> linearize( |
| 329 | const Values& values) const override { |
| 330 | return linearizeDamped(values); |
| 331 | } |
| 332 | |
| 333 | /** |
| 334 | * Triangulate and compute derivative of error with respect to point |
| 335 | * @return whether triangulation worked |
| 336 | */ |
| 337 | bool triangulateAndComputeE(Matrix& E, const Cameras& cameras) const { |
| 338 | bool nonDegenerate = triangulateForLinearize(cameras); |
| 339 | if (nonDegenerate) { |
| 340 | cameras.project2(*result_, nullptr, &E); |
| 341 | } |
| 342 | return nonDegenerate; |
| 343 | } |
| 344 | |
| 345 | /** |
| 346 | * Triangulate and compute derivative of error with respect to point |
| 347 | * @return whether triangulation worked |
| 348 | */ |
| 349 | bool triangulateAndComputeE(Matrix& E, const Values& values) const { |
| 350 | Cameras cameras = this->cameras(values); |
| 351 | return triangulateAndComputeE(E, cameras); |
| 352 | } |
| 353 | |
| 354 | /// Compute F, E only (called below in both vanilla and SVD versions) |
| 355 | /// Assumes the point has been computed |
| 356 | /// Note E can be 2m*3 or 2m*2, in case point is degenerate |
| 357 | void computeJacobiansWithTriangulatedPoint( |
| 358 | typename Base::FBlocks& Fs, Matrix& E, Vector& b, |
| 359 | const Cameras& cameras) const { |
| 360 | |
| 361 | if (!result_) { |
| 362 | // Handle degeneracy |
| 363 | // TODO check flag whether we should do this |
| 364 | Unit3 backProjected = cameras[0].backprojectPointAtInfinity( |
| 365 | this->measured_.at(0)); |
| 366 | Base::computeJacobians(Fs, E, b, cameras, backProjected); |
| 367 | } else { |
| 368 | // valid result: just return Base version |
| 369 | Base::computeJacobians(Fs, E, b, cameras, *result_); |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | /// Version that takes values, and creates the point |
| 374 | bool triangulateAndComputeJacobians( |
| 375 | typename Base::FBlocks& Fs, Matrix& E, Vector& b, |
| 376 | const Values& values) const { |
| 377 | Cameras cameras = this->cameras(values); |
| 378 | bool nonDegenerate = triangulateForLinearize(cameras); |
| 379 | if (nonDegenerate) |
| 380 | computeJacobiansWithTriangulatedPoint(Fs, E, b, cameras); |
| 381 | return nonDegenerate; |
| 382 | } |
| 383 | |
| 384 | /// takes values |
| 385 | bool triangulateAndComputeJacobiansSVD( |
| 386 | typename Base::FBlocks& Fs, Matrix& Enull, Vector& b, |
| 387 | const Values& values) const { |
| 388 | Cameras cameras = this->cameras(values); |
| 389 | bool nonDegenerate = triangulateForLinearize(cameras); |
| 390 | if (nonDegenerate) |
| 391 | Base::computeJacobiansSVD(Fs, Enull, b, cameras, *result_); |
| 392 | return nonDegenerate; |
| 393 | } |
| 394 | |
| 395 | /// Calculate vector of re-projection errors, before applying noise model |
| 396 | Vector reprojectionErrorAfterTriangulation(const Values& values) const { |
| 397 | Cameras cameras = this->cameras(values); |
| 398 | bool nonDegenerate = triangulateForLinearize(cameras); |
| 399 | if (nonDegenerate) |
| 400 | return Base::unwhitenedError(cameras, *result_); |
| 401 | else |
| 402 | return Vector::Zero(cameras.size() * 2); |
| 403 | } |
| 404 | |
| 405 | /** |
| 406 | * Calculate the error of the factor. |
| 407 | * This is the log-likelihood, e.g. \f$ 0.5(h(x)-z)^2/\sigma^2 \f$ in case of Gaussian. |
| 408 | * In this class, we take the raw prediction error \f$ h(x)-z \f$, ask the noise model |
| 409 | * to transform it to \f$ (h(x)-z)^2/\sigma^2 \f$, and then multiply by 0.5. |
| 410 | */ |
| 411 | double totalReprojectionError(const Cameras& cameras, |
| 412 | std::optional<Point3> externalPoint = {}) const { |
| 413 | |
| 414 | if (externalPoint) |
| 415 | result_ = TriangulationResult(*externalPoint); |
| 416 | else |
| 417 | result_ = triangulateSafe(cameras); |
| 418 | |
| 419 | if (result_) |
| 420 | // All good, just use version in base class |
| 421 | return Base::totalReprojectionError(cameras, *result_); |
| 422 | else if (params_.degeneracyMode == HANDLE_INFINITY) { |
| 423 | // Otherwise, manage the exceptions with rotation-only factors |
| 424 | Unit3 backprojected = cameras.front().backprojectPointAtInfinity( |
| 425 | this->measured_.at(0)); |
| 426 | return Base::totalReprojectionError(cameras, backprojected); |
| 427 | } else |
| 428 | // if we don't want to manage the exceptions we discard the factor |
| 429 | return 0.0; |
| 430 | } |
| 431 | |
| 432 | /// Calculate total reprojection error |
| 433 | double error(const Values& values) const override { |
| 434 | if (this->active(values)) { |
| 435 | return totalReprojectionError(cameras: Base::cameras(values)); |
| 436 | } else { // else of active flag |
| 437 | return 0.0; |
| 438 | } |
| 439 | } |
| 440 | |
| 441 | /** return the landmark */ |
| 442 | TriangulationResult point() const { |
| 443 | return result_; |
| 444 | } |
| 445 | |
| 446 | /** COMPUTE the landmark */ |
| 447 | TriangulationResult point(const Values& values) const { |
| 448 | Cameras cameras = this->cameras(values); |
| 449 | return triangulateSafe(cameras); |
| 450 | } |
| 451 | |
| 452 | /// Is result valid? |
| 453 | bool isValid() const { return result_.valid(); } |
| 454 | |
| 455 | /** return the degenerate state */ |
| 456 | bool isDegenerate() const { return result_.degenerate(); } |
| 457 | |
| 458 | /** return the cheirality status flag */ |
| 459 | bool isPointBehindCamera() const { return result_.behindCamera(); } |
| 460 | |
| 461 | /** return the outlier state */ |
| 462 | bool isOutlier() const { return result_.outlier(); } |
| 463 | |
| 464 | /** return the farPoint state */ |
| 465 | bool isFarPoint() const { return result_.farPoint(); } |
| 466 | |
| 467 | private: |
| 468 | |
| 469 | #if GTSAM_ENABLE_BOOST_SERIALIZATION /// |
| 470 | /// Serialization function |
| 471 | friend class boost::serialization::access; |
| 472 | template<class ARCHIVE> |
| 473 | void serialize(ARCHIVE & ar, const unsigned int version) { |
| 474 | ar & BOOST_SERIALIZATION_BASE_OBJECT_NVP(Base); |
| 475 | ar & BOOST_SERIALIZATION_NVP(params_); |
| 476 | ar & BOOST_SERIALIZATION_NVP(result_); |
| 477 | ar & BOOST_SERIALIZATION_NVP(cameraPosesTriangulation_); |
| 478 | } |
| 479 | #endif |
| 480 | } |
| 481 | ; |
| 482 | |
| 483 | /// traits |
| 484 | template<class CAMERA> |
| 485 | struct traits<SmartProjectionFactor<CAMERA> > : public Testable< |
| 486 | SmartProjectionFactor<CAMERA> > { |
| 487 | }; |
| 488 | |
| 489 | } // \ namespace gtsam |
| 490 | |