devbox
人工智能uu
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

COLMAP后端优化-代码阅读_rigbundleadjuster

作者：人工智能uu | 2024-06-29 07:19:50

赞

踩

rigbundleadjuster

转载于

【colmap】COLMAP/src/exe/colmap.cc - 知乎

  2021-07-13_我想静静，的博客-CSDN博客

稀疏重建基本流程

  1. 特征提取
  2. 增量式SfM选择无序影像进行特征匹配，
  3. 并进行几何纠正、三角测量恢复稀疏点云结构，
  4. 通过已有点云重新估计相对姿态，
  5. 再进行局部和全局的BA优化。
  6. 之后逐步向已有的结构中增加视角或影像，进行三角测量和姿态估计，再进行BA优化修正结构数据，最后输出全部的相机参数和稀疏三维点云。

代码结构

源码位于colmap/src路径下，包括下图所示几个子文件夹：

  1. 特征提取与匹配的基本操作在feature文件夹下，
  2. 最基本的数据缓存存储格式定义外还有整个SFM流程中的最底层操作在 base/util文件夹下，其中最关键的是correspondence_graph的建立，对于后续的最优初始化图像对的选取以及Next best view selection；
  3. 增量式重建的最基本操作在sfm文件夹下；
  4. 光束法平差BA的最基本操作在optim文件夹下；
  5. 最底层增量式重建和光束法平差BA操作的进一步封装在controllers文件夹下，其中automatic_reconstruction将前面分析的三步数据处理流程合并，一键式重建。
  6. ui文件夹与软件界面相关。

代码主体

代码入口在src/exe文件夹下的colmap.cc中：

  1. int main(int argc, char** argv) {
  2.   using namespace colmap;
  3.  
  4.   InitializeGlog(argv);
  5. #ifdef GUI_ENABLED
  6.   Q_INIT_RESOURCE(resources);
  7. #endif
  8.  
  9.   std::vector<std::pair<std::string, command_func_t>> commands;
  10.   commands.emplace_back("gui", &RunGraphicalUserInterface);
  11.   commands.emplace_back("automatic_reconstructor", &RunAutomaticReconstructor);
  12.   commands.emplace_back("bundle_adjuster", &RunBundleAdjuster);
  13.   commands.emplace_back("color_extractor", &RunColorExtractor);
  14.   commands.emplace_back("database_cleaner", &RunDatabaseCleaner);
  15.   commands.emplace_back("database_creator", &RunDatabaseCreator);
  16.   commands.emplace_back("database_merger", &RunDatabaseMerger);
  17.   commands.emplace_back("delaunay_mesher", &RunDelaunayMesher);
  18.   commands.emplace_back("exhaustive_matcher", &RunExhaustiveMatcher);
  19.   commands.emplace_back("feature_extractor", &RunFeatureExtractor);
  20.   commands.emplace_back("feature_importer", &RunFeatureImporter);
  21.   commands.emplace_back("hierarchical_mapper", &RunHierarchicalMapper);
  22.   commands.emplace_back("image_deleter", &RunImageDeleter);
  23.   commands.emplace_back("image_filterer", &RunImageFilterer);
  24.   commands.emplace_back("image_rectifier", &RunImageRectifier);
  25.   commands.emplace_back("image_registrator", &RunImageRegistrator);
  26.   commands.emplace_back("image_undistorter", &RunImageUndistorter);
  27.   commands.emplace_back("image_undistorter_standalone",
  28.                         &RunImageUndistorterStandalone);
  29.   commands.emplace_back("mapper", &RunMapper);
  30.   commands.emplace_back("matches_importer", &RunMatchesImporter);
  31.   commands.emplace_back("model_aligner", &RunModelAligner);
  32.   commands.emplace_back("model_analyzer", &RunModelAnalyzer);
  33.   commands.emplace_back("model_comparer", &RunModelComparer);
  34.   commands.emplace_back("model_converter", &RunModelConverter);
  35.   commands.emplace_back("model_cropper", &RunModelCropper);
  36.   commands.emplace_back("model_merger", &RunModelMerger);
  37.   commands.emplace_back("model_orientation_aligner",
  38.                         &RunModelOrientationAligner);
  39.   commands.emplace_back("model_splitter", &RunModelSplitter);
  40.   commands.emplace_back("model_transformer", &RunModelTransformer);
  41.   commands.emplace_back("patch_match_stereo", &RunPatchMatchStereo);
  42.   commands.emplace_back("point_filtering", &RunPointFiltering);
  43.   commands.emplace_back("point_triangulator", &RunPointTriangulator);
  44.   commands.emplace_back("poisson_mesher", &RunPoissonMesher);
  45.   commands.emplace_back("project_generator", &RunProjectGenerator);
  46.   commands.emplace_back("rig_bundle_adjuster", &RunRigBundleAdjuster);
  47.   commands.emplace_back("sequential_matcher", &RunSequentialMatcher);
  48.   commands.emplace_back("spatial_matcher", &RunSpatialMatcher);
  49.   commands.emplace_back("stereo_fusion", &RunStereoFuser);
  50.   commands.emplace_back("transitive_matcher", &RunTransitiveMatcher);
  51.   commands.emplace_back("vocab_tree_builder", &RunVocabTreeBuilder);
  52.   commands.emplace_back("vocab_tree_matcher", &RunVocabTreeMatcher);
  53.   commands.emplace_back("vocab_tree_retriever", &RunVocabTreeRetriever);
  54.  
  55.   if (argc == 1) {
  56.     return ShowHelp(commands);
  57.   }
  58.  
  59.   const std::string command = argv[1];
  60.   if (command == "help" || command == "-h" || command == "--help") {
  61.     return ShowHelp(commands);
  62.   } else {
  63.     command_func_t matched_command_func = nullptr;
  64.     for (const auto& command_func : commands) {
  65.       if (command == command_func.first) {
  66.         matched_command_func = command_func.second;
  67.         break;
  68.       }
  69.     }
  70.     if (matched_command_func == nullptr) {
  71.       std::cerr << StringPrintf(
  72.                        "ERROR: Command `%s` not recognized. To list the "
  73.                        "available commands, run `colmap help`.",
  74.                        command.c_str())
  75.                 << std::endl;
  76.       return EXIT_FAILURE;
  77.     } else {
  78.       int command_argc = argc - 1;
  79.       char** command_argv = &argv[1];
  80.       command_argv[0] = argv[0];
  81.       return matched_command_func(command_argc, command_argv);
  82.     }
  83.   }
  84.  
  85.   return ShowHelp(commands);
  86. }

其中各种指令分布在如下文件中：

  1. #include "exe/database.h"
  2. #include "exe/feature.h"
  3. #include "exe/gui.h"
  4. #include "exe/image.h"
  5. #include "exe/model.h"
  6. #include "exe/mvs.h"
  7. #include "exe/sfm.h"
  8. #include "exe/vocab_tree.h"
  9. #include "util/version.h"

colmap.cc是可执行文件的主文件，从app或者终端传入的命令会由该文件处理。

(比如打开UI界面的命令colmap gui，就是由这个文件处理)

从该文件中可以看出，colmap重建整体流程为：

colmap在automatic_reconstruct模式下，需要传入的参数为image_path以及images所在的文件夹。

colmap在不使用自动重建时，在不同阶段需要传入的参数分别为：

  • feature_extractor，需要指定image_path和一个空的database_path，在特征提取结束后，提取完的特征以及其他信息被记录在database_path指定的.db文件中
  • exhaustive_matcher，需要给入上一轮生成的database文件，这一步详尽匹配的结果会继续记录在该.db文件中
  • colmap mapper，需要给入image_path以及上一轮生成的.db文件，以及一个output_path用来存入生成的model

colmap.cc的主要函数有两个，main函数：接收传入参数，根据参数执行命令；ShowHelp函数：当参数出现错误或者参数为-h时输出帮助命令。

基于ceres的RunBundleAdjuster

以commands.emplace_back("bundle_adjuster", &RunBundleAdjuster);为例，它定义在sfm文件里

  1. namespace colmap {
  2.  
  3. int RunAutomaticReconstructor(int argc, char** argv);
  4. int RunBundleAdjuster(int argc, char** argv);
  5. int RunColorExtractor(int argc, char** argv);
  6. int RunMapper(int argc, char** argv);
  7. int RunHierarchicalMapper(int argc, char** argv);
  8. int RunPointFiltering(int argc, char** argv);
  9. int RunPointTriangulator(int argc, char** argv);
  10. int RunRigBundleAdjuster(int argc, char** argv);
  11.  
  12. } 
  1. int RunBundleAdjuster(int argc, char** argv) {
  2.   std::string input_path;
  3.   std::string output_path;
  4.  
  5.   OptionManager options;
  6.   options.AddRequiredOption("input_path", &input_path);
  7.   options.AddRequiredOption("output_path", &output_path);
  8.   options.AddBundleAdjustmentOptions();
  9.   options.Parse(argc, argv);
  10.  
  11.   if (!ExistsDir(input_path)) {
  12.     std::cerr << "ERROR: `input_path` is not a directory" << std::endl;
  13.     return EXIT_FAILURE;
  14.   }
  15.  
  16.   if (!ExistsDir(output_path)) {
  17.     std::cerr << "ERROR: `output_path` is not a directory" << std::endl;
  18.     return EXIT_FAILURE;
  19.   }
  20.  
  21.   Reconstruction reconstruction;
  22.   reconstruction.Read(input_path);
  23.  
  24.   BundleAdjustmentController ba_controller(options, &reconstruction); //   BA
  25.   ba_controller.Start();
  26.   ba_controller.Wait();
  27.  
  28.   reconstruction.Write(output_path);
  29.  
  30.   return EXIT_SUCCESS;
  31. }

可以看到，BundleAdjustmentController为关键

控制全局BA的BundleAdjustmentController

RunBundleAdjuster中 BundleAdjustmentController定义在bundle_adjustment.h里，它是控制全局BA的类，该类的run方法定义了BA过程。

  1. // Class that controls the global bundle adjustment procedure.
  2. class BundleAdjustmentController : public Thread {
  3.  public:
  4.   BundleAdjustmentController(const OptionManager& options,
  5.                              Reconstruction* reconstruction);
  6.  
  7.  private:
  8.   void Run();
  9.  
  10.   const OptionManager options_;
  11.   Reconstruction* reconstruction_;
  12. };

1.构造函数

  1. BundleAdjustmentController::BundleAdjustmentController(
  2.     const OptionManager& options, Reconstruction* reconstruction)
  3.     : options_(options), reconstruction_(reconstruction) {}
  4.

2.Run()函数

  1. void BundleAdjustmentController::Run() {
  2.   CHECK_NOTNULL(reconstruction_);
  3.  
  4.   PrintHeading1("Global bundle adjustment");
  5.  
  6.   const std::vector<image_t>& reg_image_ids = reconstruction_->RegImageIds();
  7.  
  8.   if (reg_image_ids.size() < 2) {
  9.     std::cout << "ERROR: Need at least two views." << std::endl;
  10.     return;
  11.   }
  12.  
  13.   // Avoid degeneracies in bundle adjustment.
  14.   reconstruction_->FilterObservationsWithNegativeDepth();
  15.  
  16.   BundleAdjustmentOptions ba_options = *options_.bundle_adjustment;
  17.   ba_options.solver_options.minimizer_progress_to_stdout = true;
  18.  
  19.   BundleAdjustmentIterationCallback iteration_callback(this);
  20.   ba_options.solver_options.callbacks.push_back(&iteration_callback);
  21.  
  22.   // Configure bundle adjustment.
  23.   BundleAdjustmentConfig ba_config;
  24.   for (const image_t image_id : reg_image_ids) {
  25.     ba_config.AddImage(image_id);
  26.   }
  27.   ba_config.SetConstantPose(reg_image_ids[0]);
  28.   ba_config.SetConstantTvec(reg_image_ids[1], {0});
  29.  
  30.   // Run bundle adjustment.
  31.   BundleAdjuster bundle_adjuster(ba_options, ba_config);
  32.   bundle_adjuster.Solve(reconstruction_);
  33.  
  34.   GetTimer().PrintMinutes();
  35. }

Run()函数里调用了BundleAdjuster，BundleAdjuster是利用Ceres-Solver解决问题

  1. BundleAdjuster bundle_adjuster(ba_options, ba_config);
  2. bundle_adjuster.Solve(reconstruction_);

基于Ceres-Solver的BundleAdjuster

  1. class BundleAdjuster {
  2.  public:
  3.   BundleAdjuster(const BundleAdjustmentOptions& options,
  4.                  const BundleAdjustmentConfig& config);
  5.  
  6.   bool Solve(Reconstruction* reconstruction);
  7.  
  8.   // Get the Ceres solver summary for the last call to `Solve`.
  9.   const ceres::Solver::Summary& Summary() const;
  10.  
  11.  private:
  12.   void SetUp(Reconstruction* reconstruction,
  13.              ceres::LossFunction* loss_function);
  14.   void TearDown(Reconstruction* reconstruction);
  15.  
  16.   void AddImageToProblem(const image_t image_id, Reconstruction* reconstruction,
  17.                          ceres::LossFunction* loss_function);
  18.  
  19.   void AddPointToProblem(const point3D_t point3D_id,
  20.                          Reconstruction* reconstruction,
  21.                          ceres::LossFunction* loss_function);
  22.  
  23.  protected:
  24.   void ParameterizeCameras(Reconstruction* reconstruction);
  25.   void ParameterizePoints(Reconstruction* reconstruction);
  26.  
  27.   const BundleAdjustmentOptions options_;
  28.   BundleAdjustmentConfig config_;
  29.   std::unique_ptr<ceres::Problem> problem_;
  30.   ceres::Solver::Summary summary_;
  31.   std::unordered_set<camera_t> camera_ids_;
  32.   std::unordered_map<point3D_t, size_t> point3D_num_observations_;
  33. };

私有成员

problem 和 Summary

  1. std::unique_ptr<ceres::Problem> problem_;
  2. ceres::Solver::Summary summary_;
  1. const ceres::Solver::Summary& Summary() const;
  2.  
  3. const ceres::Solver::Summary& BundleAdjuster::Summary() const {
  4.   return summary_;
  5. }

构造函数

  1. BundleAdjuster::BundleAdjuster(const BundleAdjustmentOptions& options,
  2.                                const BundleAdjustmentConfig& config)
  3.     : options_(options), config_(config) {
  4.   CHECK(options_.Check());
  5. }

构造函数里赋值了options_与config_，这里看一下他们的类型

BundleAdjustmentOptions

  1. struct BundleAdjustmentOptions {
  2.   // Loss function types: Trivial (non-robust) and Cauchy (robust) loss.
  3.   enum class LossFunctionType { TRIVIAL, SOFT_L1, CAUCHY };
  4.   LossFunctionType loss_function_type = LossFunctionType::TRIVIAL;
  5.  
  6.   // Scaling factor determines residual at which robustification takes place.
  7.   double loss_function_scale = 1.0;
  8.  
  9.   // Whether to refine the focal length parameter group.
  10.   bool refine_focal_length = true;
  11.  
  12.   // Whether to refine the principal point parameter group.
  13.   bool refine_principal_point = false;
  14.  
  15.   // Whether to refine the extra parameter group.
  16.   bool refine_extra_params = true;
  17.  
  18.   // Whether to refine the extrinsic parameter group.
  19.   bool refine_extrinsics = true;
  20.  
  21.   // Whether to print a final summary.
  22.   bool print_summary = true;
  23.  
  24.   // Minimum number of residuals to enable multi-threading. Note that
  25.   // single-threaded is typically better for small bundle adjustment problems
  26.   // due to the overhead of threading.
  27.   int min_num_residuals_for_multi_threading = 50000;
  28.  
  29.   // Ceres-Solver options.
  30.   ceres::Solver::Options solver_options;
  31.  
  32.   BundleAdjustmentOptions() {
  33.     solver_options.function_tolerance = 0.0;
  34.     solver_options.gradient_tolerance = 0.0;
  35.     solver_options.parameter_tolerance = 0.0;
  36.     solver_options.minimizer_progress_to_stdout = false;
  37.     solver_options.max_num_iterations = 100;
  38.     solver_options.max_linear_solver_iterations = 200;
  39.     solver_options.max_num_consecutive_invalid_steps = 10;
  40.     solver_options.max_consecutive_nonmonotonic_steps = 10;
  41.     solver_options.num_threads = -1;
  42. #if CERES_VERSION_MAJOR < 2
  43.     solver_options.num_linear_solver_threads = -1;
  44. #endif  // CERES_VERSION_MAJOR
  45.   }
  46.  
  47.   // Create a new loss function based on the specified options. The caller
  48.   // takes ownership of the loss function.
  49.   ceres::LossFunction* CreateLossFunction() const;
  50.  
  51.   bool Check() const;
  52. };

其中 BundleAdjustmentOptions定义了诸多选项，
以及ceres::Solver::Options solver_options;和ceres::LossFunction* CreateLossFunction() const;

BundleAdjustmentConfig

  1. // Configuration container to setup bundle adjustment problems.
  2. class BundleAdjustmentConfig {
  3.  public:
  4.   BundleAdjustmentConfig();
  5.  
  6.   size_t NumImages() const;
  7.   size_t NumPoints() const;
  8.   size_t NumConstantCameras() const;
  9.   size_t NumConstantPoses() const;
  10.   size_t NumConstantTvecs() const;
  11.   size_t NumVariablePoints() const;
  12.   size_t NumConstantPoints() const;
  13.  
  14.   // Determine the number of residuals for the given reconstruction. The number
  15.   // of residuals equals the number of observations times two.
  16.   size_t NumResiduals(const Reconstruction& reconstruction) const;
  17.  
  18.   // Add / remove images from the configuration.
  19.   void AddImage(const image_t image_id);
  20.   bool HasImage(const image_t image_id) const;
  21.   void RemoveImage(const image_t image_id);
  22.  
  23.   // Set cameras of added images as constant or variable. By default all
  24.   // cameras of added images are variable. Note that the corresponding images
  25.   // have to be added prior to calling these methods.
  26.   void SetConstantCamera(const camera_t camera_id);
  27.   void SetVariableCamera(const camera_t camera_id);
  28.   bool IsConstantCamera(const camera_t camera_id) const;
  29.  
  30.   // Set the pose of added images as constant. The pose is defined as the
  31.   // rotational and translational part of the projection matrix.
  32.   void SetConstantPose(const image_t image_id);
  33.   void SetVariablePose(const image_t image_id);
  34.   bool HasConstantPose(const image_t image_id) const;
  35.  
  36.   // Set the translational part of the pose, hence the constant pose
  37.   // indices may be in [0, 1, 2] and must be unique. Note that the
  38.   // corresponding images have to be added prior to calling these methods.
  39.   void SetConstantTvec(const image_t image_id, const std::vector<int>& idxs);
  40.   void RemoveConstantTvec(const image_t image_id);
  41.   bool HasConstantTvec(const image_t image_id) const;
  42.  
  43.   // Add / remove points from the configuration. Note that points can either
  44.   // be variable or constant but not both at the same time.
  45.   void AddVariablePoint(const point3D_t point3D_id);
  46.   void AddConstantPoint(const point3D_t point3D_id);
  47.   bool HasPoint(const point3D_t point3D_id) const;
  48.   bool HasVariablePoint(const point3D_t point3D_id) const;
  49.   bool HasConstantPoint(const point3D_t point3D_id) const;
  50.   void RemoveVariablePoint(const point3D_t point3D_id);
  51.   void RemoveConstantPoint(const point3D_t point3D_id);
  52.  
  53.   // Access configuration data.
  54.   const std::unordered_set<image_t>& Images() const;
  55.   const std::unordered_set<point3D_t>& VariablePoints() const;
  56.   const std::unordered_set<point3D_t>& ConstantPoints() const;
  57.   const std::vector<int>& ConstantTvec(const image_t image_id) const;
  58.  
  59.  private:
  60.   std::unordered_set<camera_t> constant_camera_ids_;
  61.   std::unordered_set<image_t> image_ids_;
  62.   std::unordered_set<point3D_t> variable_point3D_ids_;
  63.   std::unordered_set<point3D_t> constant_point3D_ids_;
  64.   std::unordered_set<image_t> constant_poses_;
  65.   std::unordered_map<image_t, std::vector<int>> constant_tvecs_;
  66. };

然后我们来看BundleAdjuster的主体，BundleAdjuster::Solve()函数

1. 定义问题problem

 problem_.reset(new ceres::Problem());

2. 设置目标函数loss_function

ceres::LossFunction* loss_function = options_.CreateLossFunction();

  1. ceres::LossFunction* BundleAdjustmentOptions::CreateLossFunction() const {
  2.   ceres::LossFunction* loss_function = nullptr;
  3.   switch (loss_function_type) {
  4.     case LossFunctionType::TRIVIAL:
  5.       loss_function = new ceres::TrivialLoss();
  6.       break;
  7.     case LossFunctionType::SOFT_L1:
  8.       loss_function = new ceres::SoftLOneLoss(loss_function_scale);
  9.       break;
  10.     case LossFunctionType::CAUCHY:
  11.       loss_function = new ceres::CauchyLoss(loss_function_scale);
  12.       break;
  13.   }
  14.   CHECK_NOTNULL(loss_function);
  15.   return loss_function;
  16. }

3. 通过SetUp将图像和点加进去

  1. void BundleAdjuster::SetUp(Reconstruction* reconstruction,
  2.                            ceres::LossFunction* loss_function) {
  3.   // Warning: AddPointsToProblem assumes that AddImageToProblem is called first.
  4.   // Do not change order of instructions!
  5.   for (const image_t image_id : config_.Images()) {
  6.     AddImageToProblem(image_id, reconstruction, loss_function);
  7.   }
  8.   for (const auto point3D_id : config_.VariablePoints()) {
  9.     AddPointToProblem(point3D_id, reconstruction, loss_function);
  10.   }
  11.   for (const auto point3D_id : config_.ConstantPoints()) {
  12.     AddPointToProblem(point3D_id, reconstruction, loss_function);
  13.   }
  14.  
  15.   ParameterizeCameras(reconstruction);
  16.   ParameterizePoints(reconstruction);
  17. }

AddImageToProblem()和AddPointToProblem()函数将图像和点加到problem里，这两个函数的具体代码就不展示了，这两个函数中设计误差项添加的代码如下：

  1. cost_function = BundleAdjustmentConstantPoseCostFunction<CameraModel>::Create( 
  2.             image.Qvec(), image.Tvec(), point2D.XY());
  3.  
  4. problem_->AddResidualBlock(cost_function, loss_function,
  5.                                  point3D.XYZ().data(), camera_params_data);

和

  1. // Set pose parameterization.
  2.     if (!constant_pose) {
  3.       ceres::LocalParameterization* quaternion_parameterization =
  4.           new ceres::QuaternionParameterization;
  5.       problem_->SetParameterization(qvec_data, quaternion_parameterization);
  6.       if (config_.HasConstantTvec(image_id)) {
  7.         const std::vector<int>& constant_tvec_idxs =
  8.             config_.ConstantTvec(image_id);
  9.         ceres::SubsetParameterization* tvec_parameterization =
  10.             new ceres::SubsetParameterization(3, constant_tvec_idxs);
  11.         problem_->SetParameterization(tvec_data, tvec_parameterization);
  12.       }
  13.     }

其中BundleAdjustmentConstantPoseCostFunction定义于base/cost_functions里，其中包括了大量其他代价函数。该文件涉及ceres的诸多方法：

  1. static ceres::CostFunction* Create(const Eigen::Vector2d& point2D) {
  2.     return (new ceres::AutoDiffCostFunction<
  3.             RigBundleAdjustmentCostFunction<CameraModel>, 2, 4, 3, 4, 3, 3,
  4.             CameraModel::kNumParams>(
  5.         new RigBundleAdjustmentCostFunction(point2D)));
  6.   }
  7.  
  8. ceres::QuaternionProduct(rel_qvec, rig_qvec, qvec);
  9.  
  10.  
  11. ceres::UnitQuaternionRotatePoint(rel_qvec, rig_tvec, tvec);

4.在options里配置各种优化的选项(接着上面的3)

  1. ceres::Solver::Options solver_options = options_.solver_options;
  2.  
  3. const bool has_sparse =
  4.       solver_options.sparse_linear_algebra_library_type != ceres::NO_SPARSE;
  5.  
  6. // Empirical choice.
  7. const size_t kMaxNumImagesDirectDenseSolver = 50;
  8. const size_t kMaxNumImagesDirectSparseSolver = 1000;
  9. const size_t num_images = config_.NumImages();
  10. if (num_images <= kMaxNumImagesDirectDenseSolver) {
  11.   solver_options.linear_solver_type = ceres::DENSE_SCHUR;
  12. } else if (num_images <= kMaxNumImagesDirectSparseSolver && has_sparse) {
  13.   solver_options.linear_solver_type = ceres::SPARSE_SCHUR;
  14. } else {  // Indirect sparse (preconditioned CG) solver.
  15.   solver_options.linear_solver_type = ceres::ITERATIVE_SCHUR;
  16.   solver_options.preconditioner_type = ceres::SCHUR_JACOBI;
  17. }

5.多线程设计

  1. if (problem_->NumResiduals() <
  2.       options_.min_num_residuals_for_multi_threading) {
  3.     solver_options.num_threads = 1;
  4. #if CERES_VERSION_MAJOR < 2
  5.     solver_options.num_linear_solver_threads = 1;
  6. #endif  // CERES_VERSION_MAJOR
  7.   } else {
  8.     solver_options.num_threads =
  9.         GetEffectiveNumThreads(solver_options.num_threads);
  10. #if CERES_VERSION_MAJOR < 2
  11.     solver_options.num_linear_solver_threads =
  12.         GetEffectiveNumThreads(solver_options.num_linear_solver_threads);
  13. #endif  // CERES_VERSION_MAJOR
  14.   }
  15.

6.求解

ceres::Solve(solver_options, problem_.get(), &summary_);

相关博客：关于Colmap中BA的Ceres源码_又决定放弃的博客-CSDN博客

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/人工智能uu/article/detail/768610
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号