Home | History | Annotate | Download | only in examples 
      1 // Ceres Solver - A fast non-linear least squares minimizer
      2 // Copyright 2010, 2011, 2012 Google Inc. All rights reserved.
      3 // http://code.google.com/p/ceres-solver/
      4 //
      5 // Redistribution and use in source and binary forms, with or without
      6 // modification, are permitted provided that the following conditions are met:
      7 //
      8 // * Redistributions of source code must retain the above copyright notice,
      9 //   this list of conditions and the following disclaimer.
     10 // * Redistributions in binary form must reproduce the above copyright notice,
     11 //   this list of conditions and the following disclaimer in the documentation
     12 //   and/or other materials provided with the distribution.
     13 // * Neither the name of Google Inc. nor the names of its contributors may be
     14 //   used to endorse or promote products derived from this software without
     15 //   specific prior written permission.
     16 //
     17 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
     18 // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     19 // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     20 // ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
     21 // LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     22 // CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     23 // SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     24 // INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     25 // CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     26 // ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     27 // POSSIBILITY OF SUCH DAMAGE.
     28 //
     29 // Author: keir (at) google.com (Keir Mierle)
     30 //
     31 // A minimal, self-contained bundle adjuster using Ceres, that reads
     32 // files from University of Washington' Bundle Adjustment in the Large dataset:
     33 // http://grail.cs.washington.edu/projects/bal
     34 //
     35 // This does not use the best configuration for solving; see the more involved
     36 // bundle_adjuster.cc file for details.
     37 
     38 #include <cmath>
     39 #include <cstdio>
     40 #include <iostream>
     41 
     42 #include "ceres/ceres.h"
     43 #include "ceres/rotation.h"
     44 
     45 // Read a Bundle Adjustment in the Large dataset.
     46 class BALProblem {
     47  public:
     48   ~BALProblem() {
     49     delete[] point_index_;
     50     delete[] camera_index_;
     51     delete[] observations_;
     52     delete[] parameters_;
     53   }
     54 
     55   int num_observations()       const { return num_observations_;               }
     56   const double* observations() const { return observations_;                   }
     57   double* mutable_cameras()          { return parameters_;                     }
     58   double* mutable_points()           { return parameters_  + 9 * num_cameras_; }
     59 
     60   double* mutable_camera_for_observation(int i) {
     61     return mutable_cameras() + camera_index_[i] * 9;
     62   }
     63   double* mutable_point_for_observation(int i) {
     64     return mutable_points() + point_index_[i] * 3;
     65   }
     66 
     67   bool LoadFile(const char* filename) {
     68     FILE* fptr = fopen(filename, "r");
     69     if (fptr == NULL) {
     70       return false;
     71     };
     72 
     73     FscanfOrDie(fptr, "%d", &num_cameras_);
     74     FscanfOrDie(fptr, "%d", &num_points_);
     75     FscanfOrDie(fptr, "%d", &num_observations_);
     76 
     77     point_index_ = new int[num_observations_];
     78     camera_index_ = new int[num_observations_];
     79     observations_ = new double[2 * num_observations_];
     80 
     81     num_parameters_ = 9 * num_cameras_ + 3 * num_points_;
     82     parameters_ = new double[num_parameters_];
     83 
     84     for (int i = 0; i < num_observations_; ++i) {
     85       FscanfOrDie(fptr, "%d", camera_index_ + i);
     86       FscanfOrDie(fptr, "%d", point_index_ + i);
     87       for (int j = 0; j < 2; ++j) {
     88         FscanfOrDie(fptr, "%lf", observations_ + 2*i + j);
     89       }
     90     }
     91 
     92     for (int i = 0; i < num_parameters_; ++i) {
     93       FscanfOrDie(fptr, "%lf", parameters_ + i);
     94     }
     95     return true;
     96   }
     97 
     98  private:
     99   template<typename T>
    100   void FscanfOrDie(FILE *fptr, const char *format, T *value) {
    101     int num_scanned = fscanf(fptr, format, value);
    102     if (num_scanned != 1) {
    103       LOG(FATAL) << "Invalid UW data file.";
    104     }
    105   }
    106 
    107   int num_cameras_;
    108   int num_points_;
    109   int num_observations_;
    110   int num_parameters_;
    111 
    112   int* point_index_;
    113   int* camera_index_;
    114   double* observations_;
    115   double* parameters_;
    116 };
    117 
    118 // Templated pinhole camera model for used with Ceres.  The camera is
    119 // parameterized using 9 parameters: 3 for rotation, 3 for translation, 1 for
    120 // focal length and 2 for radial distortion. The principal point is not modeled
    121 // (i.e. it is assumed be located at the image center).
    122 struct SnavelyReprojectionError {
    123   SnavelyReprojectionError(double observed_x, double observed_y)
    124       : observed_x(observed_x), observed_y(observed_y) {}
    125 
    126   template <typename T>
    127   bool operator()(const T* const camera,
    128                   const T* const point,
    129                   T* residuals) const {
    130     // camera[0,1,2] are the angle-axis rotation.
    131     T p[3];
    132     ceres::AngleAxisRotatePoint(camera, point, p);
    133 
    134     // camera[3,4,5] are the translation.
    135     p[0] += camera[3];
    136     p[1] += camera[4];
    137     p[2] += camera[5];
    138 
    139     // Compute the center of distortion. The sign change comes from
    140     // the camera model that Noah Snavely's Bundler assumes, whereby
    141     // the camera coordinate system has a negative z axis.
    142     T xp = - p[0] / p[2];
    143     T yp = - p[1] / p[2];
    144 
    145     // Apply second and fourth order radial distortion.
    146     const T& l1 = camera[7];
    147     const T& l2 = camera[8];
    148     T r2 = xp*xp + yp*yp;
    149     T distortion = T(1.0) + r2  * (l1 + l2  * r2);
    150 
    151     // Compute final projected point position.
    152     const T& focal = camera[6];
    153     T predicted_x = focal * distortion * xp;
    154     T predicted_y = focal * distortion * yp;
    155 
    156     // The error is the difference between the predicted and observed position.
    157     residuals[0] = predicted_x - T(observed_x);
    158     residuals[1] = predicted_y - T(observed_y);
    159 
    160     return true;
    161   }
    162 
    163   double observed_x;
    164   double observed_y;
    165 };
    166 
    167 int main(int argc, char** argv) {
    168   google::InitGoogleLogging(argv[0]);
    169   if (argc != 2) {
    170     std::cerr << "usage: simple_bundle_adjuster <bal_problem>\n";
    171     return 1;
    172   }
    173 
    174   BALProblem bal_problem;
    175   if (!bal_problem.LoadFile(argv[1])) {
    176     std::cerr << "ERROR: unable to open file " << argv[1] << "\n";
    177     return 1;
    178   }
    179 
    180   // Create residuals for each observation in the bundle adjustment problem. The
    181   // parameters for cameras and points are added automatically.
    182   ceres::Problem problem;
    183   for (int i = 0; i < bal_problem.num_observations(); ++i) {
    184     // Each Residual block takes a point and a camera as input and outputs a 2
    185     // dimensional residual. Internally, the cost function stores the observed
    186     // image location and compares the reprojection against the observation.
    187     ceres::CostFunction* cost_function =
    188         new ceres::AutoDiffCostFunction<SnavelyReprojectionError, 2, 9, 3>(
    189             new SnavelyReprojectionError(
    190                 bal_problem.observations()[2 * i + 0],
    191                 bal_problem.observations()[2 * i + 1]));
    192 
    193     problem.AddResidualBlock(cost_function,
    194                              NULL /* squared loss */,
    195                              bal_problem.mutable_camera_for_observation(i),
    196                              bal_problem.mutable_point_for_observation(i));
    197   }
    198 
    199   // Make Ceres automatically detect the bundle structure. Note that the
    200   // standard solver, SPARSE_NORMAL_CHOLESKY, also works fine but it is slower
    201   // for standard bundle adjustment problems.
    202   ceres::Solver::Options options;
    203   options.linear_solver_type = ceres::DENSE_SCHUR;
    204   options.minimizer_progress_to_stdout = true;
    205 
    206   ceres::Solver::Summary summary;
    207   ceres::Solve(options, &problem, &summary);
    208   std::cout << summary.FullReport() << "\n";
    209   return 0;
    210 }
    211