1 // Ceres Solver - A fast non-linear least squares minimizer
2 // Copyright 2014 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: richie.stebbing@gmail.com (Richard Stebbing)
30 //
31 // This fits points randomly distributed on an ellipse with an approximate
32 // line segment contour. This is done by jointly optimizing the control points
33 // of the line segment contour along with the preimage positions for the data
34 // points. The purpose of this example is to show an example use case for
35 // dynamic_sparsity, and how it can benefit problems which are numerically
36 // dense but dynamically sparse.
37 
38 #include <cmath>
39 #include <vector>
40 #include "ceres/ceres.h"
41 #include "glog/logging.h"
42 
43 // Data generated with the following Python code.
44 //   import numpy as np
45 //   np.random.seed(1337)
46 //   t = np.linspace(0.0, 2.0 * np.pi, 212, endpoint=False)
47 //   t += 2.0 * np.pi * 0.01 * np.random.randn(t.size)
48 //   theta = np.deg2rad(15)
49 //   a, b = np.cos(theta), np.sin(theta)
50 //   R = np.array([[a, -b],
51 //                 [b, a]])
52 //   Y = np.dot(np.c_[4.0 * np.cos(t), np.sin(t)], R.T)
53 
54 const int kYRows = 212;
55 const int kYCols = 2;
56 const double kYData[kYRows * kYCols] = {
57   +3.871364e+00, +9.916027e-01,
58   +3.864003e+00, +1.034148e+00,
59   +3.850651e+00, +1.072202e+00,
60   +3.868350e+00, +1.014408e+00,
61   +3.796381e+00, +1.153021e+00,
62   +3.857138e+00, +1.056102e+00,
63   +3.787532e+00, +1.162215e+00,
64   +3.704477e+00, +1.227272e+00,
65   +3.564711e+00, +1.294959e+00,
66   +3.754363e+00, +1.191948e+00,
67   +3.482098e+00, +1.322725e+00,
68   +3.602777e+00, +1.279658e+00,
69   +3.585433e+00, +1.286858e+00,
70   +3.347505e+00, +1.356415e+00,
71   +3.220855e+00, +1.378914e+00,
72   +3.558808e+00, +1.297174e+00,
73   +3.403618e+00, +1.343809e+00,
74   +3.179828e+00, +1.384721e+00,
75   +3.054789e+00, +1.398759e+00,
76   +3.294153e+00, +1.366808e+00,
77   +3.247312e+00, +1.374813e+00,
78   +2.988547e+00, +1.404247e+00,
79   +3.114508e+00, +1.392698e+00,
80   +2.899226e+00, +1.409802e+00,
81   +2.533256e+00, +1.414778e+00,
82   +2.654773e+00, +1.415909e+00,
83   +2.565100e+00, +1.415313e+00,
84   +2.976456e+00, +1.405118e+00,
85   +2.484200e+00, +1.413640e+00,
86   +2.324751e+00, +1.407476e+00,
87   +1.930468e+00, +1.378221e+00,
88   +2.329017e+00, +1.407688e+00,
89   +1.760640e+00, +1.360319e+00,
90   +2.147375e+00, +1.396603e+00,
91   +1.741989e+00, +1.358178e+00,
92   +1.743859e+00, +1.358394e+00,
93   +1.557372e+00, +1.335208e+00,
94   +1.280551e+00, +1.295087e+00,
95   +1.429880e+00, +1.317546e+00,
96   +1.213485e+00, +1.284400e+00,
97   +9.168172e-01, +1.232870e+00,
98   +1.311141e+00, +1.299839e+00,
99   +1.231969e+00, +1.287382e+00,
100   +7.453773e-01, +1.200049e+00,
101   +6.151587e-01, +1.173683e+00,
102   +5.935666e-01, +1.169193e+00,
103   +2.538707e-01, +1.094227e+00,
104   +6.806136e-01, +1.187089e+00,
105   +2.805447e-01, +1.100405e+00,
106   +6.184807e-01, +1.174371e+00,
107   +1.170550e-01, +1.061762e+00,
108   +2.890507e-01, +1.102365e+00,
109   +3.834234e-01, +1.123772e+00,
110   +3.980161e-04, +1.033061e+00,
111   -3.651680e-01, +9.370367e-01,
112   -8.386351e-01, +7.987201e-01,
113   -8.105704e-01, +8.073702e-01,
114   -8.735139e-01, +7.878886e-01,
115   -9.913836e-01, +7.506100e-01,
116   -8.784011e-01, +7.863636e-01,
117   -1.181440e+00, +6.882566e-01,
118   -1.229556e+00, +6.720191e-01,
119   -1.035839e+00, +7.362765e-01,
120   -8.031520e-01, +8.096470e-01,
121   -1.539136e+00, +5.629549e-01,
122   -1.755423e+00, +4.817306e-01,
123   -1.337589e+00, +6.348763e-01,
124   -1.836966e+00, +4.499485e-01,
125   -1.913367e+00, +4.195617e-01,
126   -2.126467e+00, +3.314900e-01,
127   -1.927625e+00, +4.138238e-01,
128   -2.339862e+00, +2.379074e-01,
129   -1.881736e+00, +4.322152e-01,
130   -2.116753e+00, +3.356163e-01,
131   -2.255733e+00, +2.754930e-01,
132   -2.555834e+00, +1.368473e-01,
133   -2.770277e+00, +2.895711e-02,
134   -2.563376e+00, +1.331890e-01,
135   -2.826715e+00, -9.000818e-04,
136   -2.978191e+00, -8.457804e-02,
137   -3.115855e+00, -1.658786e-01,
138   -2.982049e+00, -8.678322e-02,
139   -3.307892e+00, -2.902083e-01,
140   -3.038346e+00, -1.194222e-01,
141   -3.190057e+00, -2.122060e-01,
142   -3.279086e+00, -2.705777e-01,
143   -3.322028e+00, -2.999889e-01,
144   -3.122576e+00, -1.699965e-01,
145   -3.551973e+00, -4.768674e-01,
146   -3.581866e+00, -5.032175e-01,
147   -3.497799e+00, -4.315203e-01,
148   -3.565384e+00, -4.885602e-01,
149   -3.699493e+00, -6.199815e-01,
150   -3.585166e+00, -5.061925e-01,
151   -3.758914e+00, -6.918275e-01,
152   -3.741104e+00, -6.689131e-01,
153   -3.688331e+00, -6.077239e-01,
154   -3.810425e+00, -7.689015e-01,
155   -3.791829e+00, -7.386911e-01,
156   -3.789951e+00, -7.358189e-01,
157   -3.823100e+00, -7.918398e-01,
158   -3.857021e+00, -8.727074e-01,
159   -3.858250e+00, -8.767645e-01,
160   -3.872100e+00, -9.563174e-01,
161   -3.864397e+00, -1.032630e+00,
162   -3.846230e+00, -1.081669e+00,
163   -3.834799e+00, -1.102536e+00,
164   -3.866684e+00, -1.022901e+00,
165   -3.808643e+00, -1.139084e+00,
166   -3.868840e+00, -1.011569e+00,
167   -3.791071e+00, -1.158615e+00,
168   -3.797999e+00, -1.151267e+00,
169   -3.696278e+00, -1.232314e+00,
170   -3.779007e+00, -1.170504e+00,
171   -3.622855e+00, -1.270793e+00,
172   -3.647249e+00, -1.259166e+00,
173   -3.655412e+00, -1.255042e+00,
174   -3.573218e+00, -1.291696e+00,
175   -3.638019e+00, -1.263684e+00,
176   -3.498409e+00, -1.317750e+00,
177   -3.304143e+00, -1.364970e+00,
178   -3.183001e+00, -1.384295e+00,
179   -3.202456e+00, -1.381599e+00,
180   -3.244063e+00, -1.375332e+00,
181   -3.233308e+00, -1.377019e+00,
182   -3.060112e+00, -1.398264e+00,
183   -3.078187e+00, -1.396517e+00,
184   -2.689594e+00, -1.415761e+00,
185   -2.947662e+00, -1.407039e+00,
186   -2.854490e+00, -1.411860e+00,
187   -2.660499e+00, -1.415900e+00,
188   -2.875955e+00, -1.410930e+00,
189   -2.675385e+00, -1.415848e+00,
190   -2.813155e+00, -1.413363e+00,
191   -2.417673e+00, -1.411512e+00,
192   -2.725461e+00, -1.415373e+00,
193   -2.148334e+00, -1.396672e+00,
194   -2.108972e+00, -1.393738e+00,
195   -2.029905e+00, -1.387302e+00,
196   -2.046214e+00, -1.388687e+00,
197   -2.057402e+00, -1.389621e+00,
198   -1.650250e+00, -1.347160e+00,
199   -1.806764e+00, -1.365469e+00,
200   -1.206973e+00, -1.283343e+00,
201   -8.029259e-01, -1.211308e+00,
202   -1.229551e+00, -1.286993e+00,
203   -1.101507e+00, -1.265754e+00,
204   -9.110645e-01, -1.231804e+00,
205   -1.110046e+00, -1.267211e+00,
206   -8.465274e-01, -1.219677e+00,
207   -7.594163e-01, -1.202818e+00,
208   -8.023823e-01, -1.211203e+00,
209   -3.732519e-01, -1.121494e+00,
210   -1.918373e-01, -1.079668e+00,
211   -4.671988e-01, -1.142253e+00,
212   -4.033645e-01, -1.128215e+00,
213   -1.920740e-01, -1.079724e+00,
214   -3.022157e-01, -1.105389e+00,
215   -1.652831e-01, -1.073354e+00,
216   +4.671625e-01, -9.085886e-01,
217   +5.940178e-01, -8.721832e-01,
218   +3.147557e-01, -9.508290e-01,
219   +6.383631e-01, -8.591867e-01,
220   +9.888923e-01, -7.514088e-01,
221   +7.076339e-01, -8.386023e-01,
222   +1.326682e+00, -6.386698e-01,
223   +1.149834e+00, -6.988221e-01,
224   +1.257742e+00, -6.624207e-01,
225   +1.492352e+00, -5.799632e-01,
226   +1.595574e+00, -5.421766e-01,
227   +1.240173e+00, -6.684113e-01,
228   +1.706612e+00, -5.004442e-01,
229   +1.873984e+00, -4.353002e-01,
230   +1.985633e+00, -3.902561e-01,
231   +1.722880e+00, -4.942329e-01,
232   +2.095182e+00, -3.447402e-01,
233   +2.018118e+00, -3.768991e-01,
234   +2.422702e+00, -1.999563e-01,
235   +2.370611e+00, -2.239326e-01,
236   +2.152154e+00, -3.205250e-01,
237   +2.525121e+00, -1.516499e-01,
238   +2.422116e+00, -2.002280e-01,
239   +2.842806e+00, +9.536372e-03,
240   +3.030128e+00, +1.146027e-01,
241   +2.888424e+00, +3.433444e-02,
242   +2.991609e+00, +9.226409e-02,
243   +2.924807e+00, +5.445844e-02,
244   +3.007772e+00, +1.015875e-01,
245   +2.781973e+00, -2.282382e-02,
246   +3.164737e+00, +1.961781e-01,
247   +3.237671e+00, +2.430139e-01,
248   +3.046123e+00, +1.240014e-01,
249   +3.414834e+00, +3.669060e-01,
250   +3.436591e+00, +3.833600e-01,
251   +3.626207e+00, +5.444311e-01,
252   +3.223325e+00, +2.336361e-01,
253   +3.511963e+00, +4.431060e-01,
254   +3.698380e+00, +6.187442e-01,
255   +3.670244e+00, +5.884943e-01,
256   +3.558833e+00, +4.828230e-01,
257   +3.661807e+00, +5.797689e-01,
258   +3.767261e+00, +7.030893e-01,
259   +3.801065e+00, +7.532650e-01,
260   +3.828523e+00, +8.024454e-01,
261   +3.840719e+00, +8.287032e-01,
262   +3.848748e+00, +8.485921e-01,
263   +3.865801e+00, +9.066551e-01,
264   +3.870983e+00, +9.404873e-01,
265   +3.870263e+00, +1.001884e+00,
266   +3.864462e+00, +1.032374e+00,
267   +3.870542e+00, +9.996121e-01,
268   +3.865424e+00, +1.028474e+00
269 };
270 ceres::ConstMatrixRef kY(kYData, kYRows, kYCols);
271 
272 class PointToLineSegmentContourCostFunction : public ceres::CostFunction {
273  public:
PointToLineSegmentContourCostFunction(const int num_segments,const Eigen::Vector2d y)274   PointToLineSegmentContourCostFunction(const int num_segments,
275                                         const Eigen::Vector2d y)
276       : num_segments_(num_segments), y_(y) {
277     // The first parameter is the preimage position.
278     mutable_parameter_block_sizes()->push_back(1);
279     // The next parameters are the control points for the line segment contour.
280     for (int i = 0; i < num_segments_; ++i) {
281       mutable_parameter_block_sizes()->push_back(2);
282     }
283     set_num_residuals(2);
284   }
285 
Evaluate(const double * const * x,double * residuals,double ** jacobians) const286   virtual bool Evaluate(const double* const* x,
287                         double* residuals,
288                         double** jacobians) const {
289     // Convert the preimage position `t` into a segment index `i0` and the
290     // line segment interpolation parameter `u`. `i1` is the index of the next
291     // control point.
292     const double t = ModuloNumSegments(*x[0]);
293     CHECK_GE(t, 0.0);
294     CHECK_LT(t, num_segments_);
295     const int i0 = floor(t), i1 = (i0 + 1) % num_segments_;
296     const double u = t - i0;
297 
298     // Linearly interpolate between control points `i0` and `i1`.
299     residuals[0] = y_[0] - ((1.0 - u) * x[1 + i0][0] + u * x[1 + i1][0]);
300     residuals[1] = y_[1] - ((1.0 - u) * x[1 + i0][1] + u * x[1 + i1][1]);
301 
302     if (jacobians == NULL) {
303       return true;
304     }
305 
306     if (jacobians[0] != NULL) {
307       jacobians[0][0] = x[1 + i0][0] - x[1 + i1][0];
308       jacobians[0][1] = x[1 + i0][1] - x[1 + i1][1];
309     }
310     for (int i = 0; i < num_segments_; ++i) {
311       if (jacobians[i + 1] != NULL) {
312         ceres::MatrixRef(jacobians[i + 1], 2, 2).setZero();
313         if (i == i0) {
314           jacobians[i + 1][0] = -(1.0 - u);
315           jacobians[i + 1][3] = -(1.0 - u);
316         } else if (i == i1) {
317           jacobians[i + 1][0] = -u;
318           jacobians[i + 1][3] = -u;
319         }
320       }
321     }
322     return true;
323   }
324 
Create(const int num_segments,const Eigen::Vector2d y)325   static ceres::CostFunction* Create(const int num_segments,
326                                      const Eigen::Vector2d y) {
327     return new PointToLineSegmentContourCostFunction(num_segments, y);
328   }
329 
330  private:
ModuloNumSegments(const double & t) const331   inline double ModuloNumSegments(const double& t) const {
332     return t - num_segments_ * floor(t / num_segments_);
333   }
334 
335   const int num_segments_;
336   const Eigen::Vector2d y_;
337 };
338 
339 struct EuclideanDistanceFunctor {
EuclideanDistanceFunctorEuclideanDistanceFunctor340   EuclideanDistanceFunctor(const double& sqrt_weight)
341       : sqrt_weight_(sqrt_weight) {}
342 
343   template <typename T>
operator ()EuclideanDistanceFunctor344   bool operator()(const T* x0, const T* x1, T* residuals) const {
345     residuals[0] = T(sqrt_weight_) * (x0[0] - x1[0]);
346     residuals[1] = T(sqrt_weight_) * (x0[1] - x1[1]);
347     return true;
348   }
349 
CreateEuclideanDistanceFunctor350   static ceres::CostFunction* Create(const double& sqrt_weight) {
351     return new ceres::AutoDiffCostFunction<EuclideanDistanceFunctor, 2, 2, 2>(
352         new EuclideanDistanceFunctor(sqrt_weight));
353   }
354 
355  private:
356   const double sqrt_weight_;
357 };
358 
SolveWithFullReport(ceres::Solver::Options options,ceres::Problem * problem,bool dynamic_sparsity)359 bool SolveWithFullReport(ceres::Solver::Options options,
360                          ceres::Problem* problem,
361                          bool dynamic_sparsity) {
362   options.dynamic_sparsity = dynamic_sparsity;
363 
364   ceres::Solver::Summary summary;
365   ceres::Solve(options, problem, &summary);
366 
367   std::cout << "####################" << std::endl;
368   std::cout << "dynamic_sparsity = " << dynamic_sparsity << std::endl;
369   std::cout << "####################" << std::endl;
370   std::cout << summary.FullReport() << std::endl;
371 
372   return summary.termination_type == ceres::CONVERGENCE;
373 }
374 
main(int argc,char ** argv)375 int main(int argc, char** argv) {
376   google::InitGoogleLogging(argv[0]);
377 
378   // Problem configuration.
379   const int num_segments = 151;
380   const double regularization_weight = 1e-2;
381 
382   // Eigen::MatrixXd is column major so we define our own MatrixXd which is
383   // row major. Eigen::VectorXd can be used directly.
384   typedef Eigen::Matrix<double,
385                         Eigen::Dynamic, Eigen::Dynamic,
386                         Eigen::RowMajor> MatrixXd;
387   using Eigen::VectorXd;
388 
389   // `X` is the matrix of control points which make up the contour of line
390   // segments. The number of control points is equal to the number of line
391   // segments because the contour is closed.
392   //
393   // Initialize `X` to points on the unit circle.
394   VectorXd w(num_segments + 1);
395   w.setLinSpaced(num_segments + 1, 0.0, 2.0 * M_PI);
396   w.conservativeResize(num_segments);
397   MatrixXd X(num_segments, 2);
398   X.col(0) = w.array().cos();
399   X.col(1) = w.array().sin();
400 
401   // Each data point has an associated preimage position on the line segment
402   // contour. For each data point we initialize the preimage positions to
403   // the index of the closest control point.
404   const int num_observations = kY.rows();
405   VectorXd t(num_observations);
406   for (int i = 0; i < num_observations; ++i) {
407     (X.rowwise() - kY.row(i)).rowwise().squaredNorm().minCoeff(&t[i]);
408   }
409 
410   ceres::Problem problem;
411 
412   // For each data point add a residual which measures its distance to its
413   // corresponding position on the line segment contour.
414   std::vector<double*> parameter_blocks(1 + num_segments);
415   parameter_blocks[0] = NULL;
416   for (int i = 0; i < num_segments; ++i) {
417     parameter_blocks[i + 1] = X.data() + 2 * i;
418   }
419   for (int i = 0; i < num_observations; ++i) {
420     parameter_blocks[0] = &t[i];
421     problem.AddResidualBlock(
422       PointToLineSegmentContourCostFunction::Create(num_segments, kY.row(i)),
423       NULL,
424       parameter_blocks);
425   }
426 
427   // Add regularization to minimize the length of the line segment contour.
428   for (int i = 0; i < num_segments; ++i) {
429     problem.AddResidualBlock(
430       EuclideanDistanceFunctor::Create(sqrt(regularization_weight)),
431       NULL,
432       X.data() + 2 * i,
433       X.data() + 2 * ((i + 1) % num_segments));
434   }
435 
436   ceres::Solver::Options options;
437   options.max_num_iterations = 100;
438   options.linear_solver_type = ceres::SPARSE_NORMAL_CHOLESKY;
439 
440   // First, solve `X` and `t` jointly with dynamic_sparsity = true.
441   MatrixXd X0 = X;
442   VectorXd t0 = t;
443   CHECK(SolveWithFullReport(options, &problem, true));
444 
445   // Second, solve with dynamic_sparsity = false.
446   X = X0;
447   t = t0;
448   CHECK(SolveWithFullReport(options, &problem, false));
449 
450   return 0;
451 }
452