• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Ceres Solver - A fast non-linear least squares minimizer
2 // Copyright 2013 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: sameeragarwal@google.com (Sameer Agarwal)
30 
31 #include "ceres/covariance.h"
32 
33 #include <algorithm>
34 #include <cmath>
35 #include "ceres/compressed_row_sparse_matrix.h"
36 #include "ceres/cost_function.h"
37 #include "ceres/covariance_impl.h"
38 #include "ceres/local_parameterization.h"
39 #include "ceres/map_util.h"
40 #include "ceres/problem_impl.h"
41 #include "gtest/gtest.h"
42 
43 namespace ceres {
44 namespace internal {
45 
TEST(CovarianceImpl,ComputeCovarianceSparsity)46 TEST(CovarianceImpl, ComputeCovarianceSparsity) {
47   double parameters[10];
48 
49   double* block1 = parameters;
50   double* block2 = block1 + 1;
51   double* block3 = block2 + 2;
52   double* block4 = block3 + 3;
53 
54   ProblemImpl problem;
55 
56   // Add in random order
57   problem.AddParameterBlock(block1, 1);
58   problem.AddParameterBlock(block4, 4);
59   problem.AddParameterBlock(block3, 3);
60   problem.AddParameterBlock(block2, 2);
61 
62   // Sparsity pattern
63   //
64   //  x 0 0 0 0 0 x x x x
65   //  0 x x x x x 0 0 0 0
66   //  0 x x x x x 0 0 0 0
67   //  0 0 0 x x x 0 0 0 0
68   //  0 0 0 x x x 0 0 0 0
69   //  0 0 0 x x x 0 0 0 0
70   //  0 0 0 0 0 0 x x x x
71   //  0 0 0 0 0 0 x x x x
72   //  0 0 0 0 0 0 x x x x
73   //  0 0 0 0 0 0 x x x x
74 
75   int expected_rows[] = {0, 5, 10, 15, 18, 21, 24, 28, 32, 36, 40};
76   int expected_cols[] = {0, 6, 7, 8, 9,
77                          1, 2, 3, 4, 5,
78                          1, 2, 3, 4, 5,
79                          3, 4, 5,
80                          3, 4, 5,
81                          3, 4, 5,
82                          6, 7, 8, 9,
83                          6, 7, 8, 9,
84                          6, 7, 8, 9,
85                          6, 7, 8, 9};
86 
87 
88   vector<pair<const double*, const double*> > covariance_blocks;
89   covariance_blocks.push_back(make_pair(block1, block1));
90   covariance_blocks.push_back(make_pair(block4, block4));
91   covariance_blocks.push_back(make_pair(block2, block2));
92   covariance_blocks.push_back(make_pair(block3, block3));
93   covariance_blocks.push_back(make_pair(block2, block3));
94   covariance_blocks.push_back(make_pair(block4, block1));  // reversed
95 
96   Covariance::Options options;
97   CovarianceImpl covariance_impl(options);
98   EXPECT_TRUE(covariance_impl
99               .ComputeCovarianceSparsity(covariance_blocks, &problem));
100 
101   const CompressedRowSparseMatrix* crsm = covariance_impl.covariance_matrix();
102 
103   EXPECT_EQ(crsm->num_rows(), 10);
104   EXPECT_EQ(crsm->num_cols(), 10);
105   EXPECT_EQ(crsm->num_nonzeros(), 40);
106 
107   const int* rows = crsm->rows();
108   for (int r = 0; r < crsm->num_rows() + 1; ++r) {
109     EXPECT_EQ(rows[r], expected_rows[r])
110         << r << " "
111         << rows[r] << " "
112         << expected_rows[r];
113   }
114 
115   const int* cols = crsm->cols();
116   for (int c = 0; c < crsm->num_nonzeros(); ++c) {
117     EXPECT_EQ(cols[c], expected_cols[c])
118         << c << " "
119         << cols[c] << " "
120         << expected_cols[c];
121   }
122 }
123 
124 
125 class UnaryCostFunction: public CostFunction {
126  public:
UnaryCostFunction(const int num_residuals,const int32 parameter_block_size,const double * jacobian)127   UnaryCostFunction(const int num_residuals,
128                     const int32 parameter_block_size,
129                     const double* jacobian)
130       : jacobian_(jacobian, jacobian + num_residuals * parameter_block_size) {
131     set_num_residuals(num_residuals);
132     mutable_parameter_block_sizes()->push_back(parameter_block_size);
133   }
134 
Evaluate(double const * const * parameters,double * residuals,double ** jacobians) const135   virtual bool Evaluate(double const* const* parameters,
136                         double* residuals,
137                         double** jacobians) const {
138     for (int i = 0; i < num_residuals(); ++i) {
139       residuals[i] = 1;
140     }
141 
142     if (jacobians == NULL) {
143       return true;
144     }
145 
146     if (jacobians[0] != NULL) {
147       copy(jacobian_.begin(), jacobian_.end(), jacobians[0]);
148     }
149 
150     return true;
151   }
152 
153  private:
154   vector<double> jacobian_;
155 };
156 
157 
158 class BinaryCostFunction: public CostFunction {
159  public:
BinaryCostFunction(const int num_residuals,const int32 parameter_block1_size,const int32 parameter_block2_size,const double * jacobian1,const double * jacobian2)160   BinaryCostFunction(const int num_residuals,
161                      const int32 parameter_block1_size,
162                      const int32 parameter_block2_size,
163                      const double* jacobian1,
164                      const double* jacobian2)
165       : jacobian1_(jacobian1,
166                    jacobian1 + num_residuals * parameter_block1_size),
167         jacobian2_(jacobian2,
168                    jacobian2 + num_residuals * parameter_block2_size) {
169     set_num_residuals(num_residuals);
170     mutable_parameter_block_sizes()->push_back(parameter_block1_size);
171     mutable_parameter_block_sizes()->push_back(parameter_block2_size);
172   }
173 
Evaluate(double const * const * parameters,double * residuals,double ** jacobians) const174   virtual bool Evaluate(double const* const* parameters,
175                         double* residuals,
176                         double** jacobians) const {
177     for (int i = 0; i < num_residuals(); ++i) {
178       residuals[i] = 2;
179     }
180 
181     if (jacobians == NULL) {
182       return true;
183     }
184 
185     if (jacobians[0] != NULL) {
186       copy(jacobian1_.begin(), jacobian1_.end(), jacobians[0]);
187     }
188 
189     if (jacobians[1] != NULL) {
190       copy(jacobian2_.begin(), jacobian2_.end(), jacobians[1]);
191     }
192 
193     return true;
194   }
195 
196  private:
197   vector<double> jacobian1_;
198   vector<double> jacobian2_;
199 };
200 
201 // x_plus_delta = delta * x;
202 class PolynomialParameterization : public LocalParameterization {
203  public:
~PolynomialParameterization()204   virtual ~PolynomialParameterization() {}
205 
Plus(const double * x,const double * delta,double * x_plus_delta) const206   virtual bool Plus(const double* x,
207                     const double* delta,
208                     double* x_plus_delta) const {
209     x_plus_delta[0] = delta[0] * x[0];
210     x_plus_delta[1] = delta[0] * x[1];
211     return true;
212   }
213 
ComputeJacobian(const double * x,double * jacobian) const214   virtual bool ComputeJacobian(const double* x, double* jacobian) const {
215     jacobian[0] = x[0];
216     jacobian[1] = x[1];
217     return true;
218   }
219 
GlobalSize() const220   virtual int GlobalSize() const { return 2; }
LocalSize() const221   virtual int LocalSize() const { return 1; }
222 };
223 
224 class CovarianceTest : public ::testing::Test {
225  protected:
SetUp()226   virtual void SetUp() {
227     double* x = parameters_;
228     double* y = x + 2;
229     double* z = y + 3;
230 
231     x[0] = 1;
232     x[1] = 1;
233     y[0] = 2;
234     y[1] = 2;
235     y[2] = 2;
236     z[0] = 3;
237 
238     {
239       double jacobian[] = { 1.0, 0.0, 0.0, 1.0};
240       problem_.AddResidualBlock(new UnaryCostFunction(2, 2, jacobian), NULL, x);
241     }
242 
243     {
244       double jacobian[] = { 2.0, 0.0, 0.0, 0.0, 2.0, 0.0, 0.0, 0.0, 2.0 };
245       problem_.AddResidualBlock(new UnaryCostFunction(3, 3, jacobian), NULL, y);
246     }
247 
248     {
249       double jacobian = 5.0;
250       problem_.AddResidualBlock(new UnaryCostFunction(1, 1, &jacobian), NULL, z);
251     }
252 
253     {
254       double jacobian1[] = { 1.0, 2.0, 3.0 };
255       double jacobian2[] = { -5.0, -6.0 };
256       problem_.AddResidualBlock(
257           new BinaryCostFunction(1, 3, 2, jacobian1, jacobian2),
258           NULL,
259           y,
260           x);
261     }
262 
263     {
264       double jacobian1[] = {2.0 };
265       double jacobian2[] = { 3.0, -2.0 };
266       problem_.AddResidualBlock(
267           new BinaryCostFunction(1, 1, 2, jacobian1, jacobian2),
268           NULL,
269           z,
270           x);
271     }
272 
273     all_covariance_blocks_.push_back(make_pair(x, x));
274     all_covariance_blocks_.push_back(make_pair(y, y));
275     all_covariance_blocks_.push_back(make_pair(z, z));
276     all_covariance_blocks_.push_back(make_pair(x, y));
277     all_covariance_blocks_.push_back(make_pair(x, z));
278     all_covariance_blocks_.push_back(make_pair(y, z));
279 
280     column_bounds_[x] = make_pair(0, 2);
281     column_bounds_[y] = make_pair(2, 5);
282     column_bounds_[z] = make_pair(5, 6);
283   }
284 
ComputeAndCompareCovarianceBlocks(const Covariance::Options & options,const double * expected_covariance)285   void ComputeAndCompareCovarianceBlocks(const Covariance::Options& options,
286                                          const double* expected_covariance) {
287     // Generate all possible combination of block pairs and check if the
288     // covariance computation is correct.
289     for (int i = 1; i <= 64; ++i) {
290       vector<pair<const double*, const double*> > covariance_blocks;
291       if (i & 1) {
292         covariance_blocks.push_back(all_covariance_blocks_[0]);
293       }
294 
295       if (i & 2) {
296         covariance_blocks.push_back(all_covariance_blocks_[1]);
297       }
298 
299       if (i & 4) {
300         covariance_blocks.push_back(all_covariance_blocks_[2]);
301       }
302 
303       if (i & 8) {
304         covariance_blocks.push_back(all_covariance_blocks_[3]);
305       }
306 
307       if (i & 16) {
308         covariance_blocks.push_back(all_covariance_blocks_[4]);
309       }
310 
311       if (i & 32) {
312         covariance_blocks.push_back(all_covariance_blocks_[5]);
313       }
314 
315       Covariance covariance(options);
316       EXPECT_TRUE(covariance.Compute(covariance_blocks, &problem_));
317 
318       for (int i = 0; i < covariance_blocks.size(); ++i) {
319         const double* block1 = covariance_blocks[i].first;
320         const double* block2 = covariance_blocks[i].second;
321         // block1, block2
322         GetCovarianceBlockAndCompare(block1, block2, covariance, expected_covariance);
323         // block2, block1
324         GetCovarianceBlockAndCompare(block2, block1, covariance, expected_covariance);
325       }
326     }
327   }
328 
GetCovarianceBlockAndCompare(const double * block1,const double * block2,const Covariance & covariance,const double * expected_covariance)329   void GetCovarianceBlockAndCompare(const double* block1,
330                                     const double* block2,
331                                     const Covariance& covariance,
332                                     const double* expected_covariance) {
333     const int row_begin = FindOrDie(column_bounds_, block1).first;
334     const int row_end = FindOrDie(column_bounds_, block1).second;
335     const int col_begin = FindOrDie(column_bounds_, block2).first;
336     const int col_end = FindOrDie(column_bounds_, block2).second;
337 
338     Matrix actual(row_end - row_begin, col_end - col_begin);
339     EXPECT_TRUE(covariance.GetCovarianceBlock(block1,
340                                               block2,
341                                               actual.data()));
342 
343     ConstMatrixRef expected(expected_covariance, 6, 6);
344     double diff_norm = (expected.block(row_begin,
345                                        col_begin,
346                                        row_end - row_begin,
347                                        col_end - col_begin) - actual).norm();
348     diff_norm /= (row_end - row_begin) * (col_end - col_begin);
349 
350     const double kTolerance = 1e-5;
351     EXPECT_NEAR(diff_norm, 0.0, kTolerance)
352         << "rows: " << row_begin << " " << row_end << "  "
353         << "cols: " << col_begin << " " << col_end << "  "
354         << "\n\n expected: \n " << expected.block(row_begin,
355                                                   col_begin,
356                                                   row_end - row_begin,
357                                                   col_end - col_begin)
358         << "\n\n actual: \n " << actual
359         << "\n\n full expected: \n" << expected;
360   }
361 
362   double parameters_[10];
363   Problem problem_;
364   vector<pair<const double*, const double*> > all_covariance_blocks_;
365   map<const double*, pair<int, int> > column_bounds_;
366 };
367 
368 
TEST_F(CovarianceTest,NormalBehavior)369 TEST_F(CovarianceTest, NormalBehavior) {
370   // J
371   //
372   //   1  0  0  0  0  0
373   //   0  1  0  0  0  0
374   //   0  0  2  0  0  0
375   //   0  0  0  2  0  0
376   //   0  0  0  0  2  0
377   //   0  0  0  0  0  5
378   //  -5 -6  1  2  3  0
379   //   3 -2  0  0  0  2
380 
381   // J'J
382   //
383   //   35  24 -5 -10 -15  6
384   //   24  41 -6 -12 -18 -4
385   //   -5  -6  5   2   3  0
386   //  -10 -12  2   8   6  0
387   //  -15 -18  3   6  13  0
388   //    6  -4  0   0   0 29
389 
390   // inv(J'J) computed using octave.
391   double expected_covariance[] = {
392      7.0747e-02,  -8.4923e-03,   1.6821e-02,   3.3643e-02,   5.0464e-02,  -1.5809e-02,  // NOLINT
393     -8.4923e-03,   8.1352e-02,   2.4758e-02,   4.9517e-02,   7.4275e-02,   1.2978e-02,  // NOLINT
394      1.6821e-02,   2.4758e-02,   2.4904e-01,  -1.9271e-03,  -2.8906e-03,  -6.5325e-05,  // NOLINT
395      3.3643e-02,   4.9517e-02,  -1.9271e-03,   2.4615e-01,  -5.7813e-03,  -1.3065e-04,  // NOLINT
396      5.0464e-02,   7.4275e-02,  -2.8906e-03,  -5.7813e-03,   2.4133e-01,  -1.9598e-04,  // NOLINT
397     -1.5809e-02,   1.2978e-02,  -6.5325e-05,  -1.3065e-04,  -1.9598e-04,   3.9544e-02,  // NOLINT
398   };
399 
400   Covariance::Options options;
401 
402 #ifndef CERES_NO_SUITESPARSE
403   options.algorithm_type = SUITE_SPARSE_QR;
404   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
405 #endif
406 
407   options.algorithm_type = DENSE_SVD;
408   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
409 
410   options.algorithm_type = EIGEN_SPARSE_QR;
411   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
412 }
413 
414 #ifdef CERES_USE_OPENMP
415 
TEST_F(CovarianceTest,ThreadedNormalBehavior)416 TEST_F(CovarianceTest, ThreadedNormalBehavior) {
417   // J
418   //
419   //   1  0  0  0  0  0
420   //   0  1  0  0  0  0
421   //   0  0  2  0  0  0
422   //   0  0  0  2  0  0
423   //   0  0  0  0  2  0
424   //   0  0  0  0  0  5
425   //  -5 -6  1  2  3  0
426   //   3 -2  0  0  0  2
427 
428   // J'J
429   //
430   //   35  24 -5 -10 -15  6
431   //   24  41 -6 -12 -18 -4
432   //   -5  -6  5   2   3  0
433   //  -10 -12  2   8   6  0
434   //  -15 -18  3   6  13  0
435   //    6  -4  0   0   0 29
436 
437   // inv(J'J) computed using octave.
438   double expected_covariance[] = {
439      7.0747e-02,  -8.4923e-03,   1.6821e-02,   3.3643e-02,   5.0464e-02,  -1.5809e-02,  // NOLINT
440     -8.4923e-03,   8.1352e-02,   2.4758e-02,   4.9517e-02,   7.4275e-02,   1.2978e-02,  // NOLINT
441      1.6821e-02,   2.4758e-02,   2.4904e-01,  -1.9271e-03,  -2.8906e-03,  -6.5325e-05,  // NOLINT
442      3.3643e-02,   4.9517e-02,  -1.9271e-03,   2.4615e-01,  -5.7813e-03,  -1.3065e-04,  // NOLINT
443      5.0464e-02,   7.4275e-02,  -2.8906e-03,  -5.7813e-03,   2.4133e-01,  -1.9598e-04,  // NOLINT
444     -1.5809e-02,   1.2978e-02,  -6.5325e-05,  -1.3065e-04,  -1.9598e-04,   3.9544e-02,  // NOLINT
445   };
446 
447   Covariance::Options options;
448   options.num_threads = 4;
449 
450 #ifndef CERES_NO_SUITESPARSE
451   options.algorithm_type = SUITE_SPARSE_QR;
452   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
453 #endif
454 
455   options.algorithm_type = DENSE_SVD;
456   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
457 
458   options.algorithm_type = EIGEN_SPARSE_QR;
459   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
460 }
461 
462 #endif  // CERES_USE_OPENMP
463 
TEST_F(CovarianceTest,ConstantParameterBlock)464 TEST_F(CovarianceTest, ConstantParameterBlock) {
465   problem_.SetParameterBlockConstant(parameters_);
466 
467   // J
468   //
469   //  0  0  0  0  0  0
470   //  0  0  0  0  0  0
471   //  0  0  2  0  0  0
472   //  0  0  0  2  0  0
473   //  0  0  0  0  2  0
474   //  0  0  0  0  0  5
475   //  0  0  1  2  3  0
476   //  0  0  0  0  0  2
477 
478   // J'J
479   //
480   //  0  0  0  0  0  0
481   //  0  0  0  0  0  0
482   //  0  0  5  2  3  0
483   //  0  0  2  8  6  0
484   //  0  0  3  6 13  0
485   //  0  0  0  0  0 29
486 
487   // pinv(J'J) computed using octave.
488   double expected_covariance[] = {
489               0,            0,            0,            0,            0,            0,  // NOLINT
490               0,            0,            0,            0,            0,            0,  // NOLINT
491               0,            0,      0.23611,     -0.02778,     -0.04167,     -0.00000,  // NOLINT
492               0,            0,     -0.02778,      0.19444,     -0.08333,     -0.00000,  // NOLINT
493               0,            0,     -0.04167,     -0.08333,      0.12500,     -0.00000,  // NOLINT
494               0,            0,     -0.00000,     -0.00000,     -0.00000,      0.03448   // NOLINT
495   };
496 
497   Covariance::Options options;
498 
499 #ifndef CERES_NO_SUITESPARSE
500   options.algorithm_type = SUITE_SPARSE_QR;
501   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
502 #endif
503 
504   options.algorithm_type = DENSE_SVD;
505   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
506 
507   options.algorithm_type = EIGEN_SPARSE_QR;
508   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
509 }
510 
TEST_F(CovarianceTest,LocalParameterization)511 TEST_F(CovarianceTest, LocalParameterization) {
512   double* x = parameters_;
513   double* y = x + 2;
514 
515   problem_.SetParameterization(x, new PolynomialParameterization);
516 
517   vector<int> subset;
518   subset.push_back(2);
519   problem_.SetParameterization(y, new SubsetParameterization(3, subset));
520 
521   // Raw Jacobian: J
522   //
523   //   1   0  0  0  0  0
524   //   0   1  0  0  0  0
525   //   0   0  2  0  0  0
526   //   0   0  0  2  0  0
527   //   0   0  0  0  0  0
528   //   0   0  0  0  0  5
529   //  -5  -6  1  2  0  0
530   //   3  -2  0  0  0  2
531 
532   // Global to local jacobian: A
533   //
534   //
535   //  1   0   0   0   0
536   //  1   0   0   0   0
537   //  0   1   0   0   0
538   //  0   0   1   0   0
539   //  0   0   0   1   0
540   //  0   0   0   0   1
541 
542   // A * pinv((J*A)'*(J*A)) * A'
543   // Computed using octave.
544   double expected_covariance[] = {
545     0.01766,   0.01766,   0.02158,   0.04316,   0.00000,  -0.00122,
546     0.01766,   0.01766,   0.02158,   0.04316,   0.00000,  -0.00122,
547     0.02158,   0.02158,   0.24860,  -0.00281,   0.00000,  -0.00149,
548     0.04316,   0.04316,  -0.00281,   0.24439,   0.00000,  -0.00298,
549     0.00000,   0.00000,   0.00000,   0.00000,   0.00000,   0.00000,
550    -0.00122,  -0.00122,  -0.00149,  -0.00298,   0.00000,   0.03457
551   };
552 
553   Covariance::Options options;
554 
555 #ifndef CERES_NO_SUITESPARSE
556   options.algorithm_type = SUITE_SPARSE_QR;
557   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
558 #endif
559 
560   options.algorithm_type = DENSE_SVD;
561   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
562 
563   options.algorithm_type = EIGEN_SPARSE_QR;
564   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
565 }
566 
567 
TEST_F(CovarianceTest,TruncatedRank)568 TEST_F(CovarianceTest, TruncatedRank) {
569   // J
570   //
571   //   1  0  0  0  0  0
572   //   0  1  0  0  0  0
573   //   0  0  2  0  0  0
574   //   0  0  0  2  0  0
575   //   0  0  0  0  2  0
576   //   0  0  0  0  0  5
577   //  -5 -6  1  2  3  0
578   //   3 -2  0  0  0  2
579 
580   // J'J
581   //
582   //   35  24 -5 -10 -15  6
583   //   24  41 -6 -12 -18 -4
584   //   -5  -6  5   2   3  0
585   //  -10 -12  2   8   6  0
586   //  -15 -18  3   6  13  0
587   //    6  -4  0   0   0 29
588 
589   // 3.4142 is the smallest eigen value of J'J. The following matrix
590   // was obtained by dropping the eigenvector corresponding to this
591   // eigenvalue.
592   double expected_covariance[] = {
593      5.4135e-02,  -3.5121e-02,   1.7257e-04,   3.4514e-04,   5.1771e-04,  -1.6076e-02,
594     -3.5121e-02,   3.8667e-02,  -1.9288e-03,  -3.8576e-03,  -5.7864e-03,   1.2549e-02,
595      1.7257e-04,  -1.9288e-03,   2.3235e-01,  -3.5297e-02,  -5.2946e-02,  -3.3329e-04,
596      3.4514e-04,  -3.8576e-03,  -3.5297e-02,   1.7941e-01,  -1.0589e-01,  -6.6659e-04,
597      5.1771e-04,  -5.7864e-03,  -5.2946e-02,  -1.0589e-01,   9.1162e-02,  -9.9988e-04,
598     -1.6076e-02,   1.2549e-02,  -3.3329e-04,  -6.6659e-04,  -9.9988e-04,   3.9539e-02
599   };
600 
601 
602   {
603     Covariance::Options options;
604     options.algorithm_type = DENSE_SVD;
605     // Force dropping of the smallest eigenvector.
606     options.null_space_rank = 1;
607     ComputeAndCompareCovarianceBlocks(options, expected_covariance);
608   }
609 
610   {
611     Covariance::Options options;
612     options.algorithm_type = DENSE_SVD;
613     // Force dropping of the smallest eigenvector via the ratio but
614     // automatic truncation.
615     options.min_reciprocal_condition_number = 0.044494;
616     options.null_space_rank = -1;
617     ComputeAndCompareCovarianceBlocks(options, expected_covariance);
618   }
619 }
620 
621 class RankDeficientCovarianceTest : public CovarianceTest {
622  protected:
SetUp()623   virtual void SetUp() {
624     double* x = parameters_;
625     double* y = x + 2;
626     double* z = y + 3;
627 
628     {
629       double jacobian[] = { 1.0, 0.0, 0.0, 1.0};
630       problem_.AddResidualBlock(new UnaryCostFunction(2, 2, jacobian), NULL, x);
631     }
632 
633     {
634       double jacobian[] = { 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 };
635       problem_.AddResidualBlock(new UnaryCostFunction(3, 3, jacobian), NULL, y);
636     }
637 
638     {
639       double jacobian = 5.0;
640       problem_.AddResidualBlock(new UnaryCostFunction(1, 1, &jacobian), NULL, z);
641     }
642 
643     {
644       double jacobian1[] = { 0.0, 0.0, 0.0 };
645       double jacobian2[] = { -5.0, -6.0 };
646       problem_.AddResidualBlock(
647           new BinaryCostFunction(1, 3, 2, jacobian1, jacobian2),
648           NULL,
649           y,
650           x);
651     }
652 
653     {
654       double jacobian1[] = {2.0 };
655       double jacobian2[] = { 3.0, -2.0 };
656       problem_.AddResidualBlock(
657           new BinaryCostFunction(1, 1, 2, jacobian1, jacobian2),
658           NULL,
659           z,
660           x);
661     }
662 
663     all_covariance_blocks_.push_back(make_pair(x, x));
664     all_covariance_blocks_.push_back(make_pair(y, y));
665     all_covariance_blocks_.push_back(make_pair(z, z));
666     all_covariance_blocks_.push_back(make_pair(x, y));
667     all_covariance_blocks_.push_back(make_pair(x, z));
668     all_covariance_blocks_.push_back(make_pair(y, z));
669 
670     column_bounds_[x] = make_pair(0, 2);
671     column_bounds_[y] = make_pair(2, 5);
672     column_bounds_[z] = make_pair(5, 6);
673   }
674 };
675 
TEST_F(RankDeficientCovarianceTest,AutomaticTruncation)676 TEST_F(RankDeficientCovarianceTest, AutomaticTruncation) {
677   // J
678   //
679   //   1  0  0  0  0  0
680   //   0  1  0  0  0  0
681   //   0  0  0  0  0  0
682   //   0  0  0  0  0  0
683   //   0  0  0  0  0  0
684   //   0  0  0  0  0  5
685   //  -5 -6  0  0  0  0
686   //   3 -2  0  0  0  2
687 
688   // J'J
689   //
690   //  35 24  0  0  0  6
691   //  24 41  0  0  0 -4
692   //   0  0  0  0  0  0
693   //   0  0  0  0  0  0
694   //   0  0  0  0  0  0
695   //   6 -4  0  0  0 29
696 
697   // pinv(J'J) computed using octave.
698   double expected_covariance[] = {
699      0.053998,  -0.033145,   0.000000,   0.000000,   0.000000,  -0.015744,
700     -0.033145,   0.045067,   0.000000,   0.000000,   0.000000,   0.013074,
701      0.000000,   0.000000,   0.000000,   0.000000,   0.000000,   0.000000,
702      0.000000,   0.000000,   0.000000,   0.000000,   0.000000,   0.000000,
703      0.000000,   0.000000,   0.000000,   0.000000,   0.000000,   0.000000,
704     -0.015744,   0.013074,   0.000000,   0.000000,   0.000000,   0.039543
705   };
706 
707   Covariance::Options options;
708   options.algorithm_type = DENSE_SVD;
709   options.null_space_rank = -1;
710   ComputeAndCompareCovarianceBlocks(options, expected_covariance);
711 }
712 
713 class LargeScaleCovarianceTest : public ::testing::Test {
714  protected:
SetUp()715   virtual void SetUp() {
716     num_parameter_blocks_ = 2000;
717     parameter_block_size_ = 5;
718     parameters_.reset(new double[parameter_block_size_ * num_parameter_blocks_]);
719 
720     Matrix jacobian(parameter_block_size_, parameter_block_size_);
721     for (int i = 0; i < num_parameter_blocks_; ++i) {
722       jacobian.setIdentity();
723       jacobian *= (i + 1);
724 
725       double* block_i = parameters_.get() + i * parameter_block_size_;
726       problem_.AddResidualBlock(new UnaryCostFunction(parameter_block_size_,
727                                                       parameter_block_size_,
728                                                       jacobian.data()),
729                                 NULL,
730                                 block_i);
731       for (int j = i; j < num_parameter_blocks_; ++j) {
732         double* block_j = parameters_.get() + j * parameter_block_size_;
733         all_covariance_blocks_.push_back(make_pair(block_i, block_j));
734       }
735     }
736   }
737 
ComputeAndCompare(CovarianceAlgorithmType algorithm_type,int num_threads)738   void ComputeAndCompare(CovarianceAlgorithmType algorithm_type,
739                          int num_threads) {
740     Covariance::Options options;
741     options.algorithm_type = algorithm_type;
742     options.num_threads = num_threads;
743     Covariance covariance(options);
744     EXPECT_TRUE(covariance.Compute(all_covariance_blocks_, &problem_));
745 
746     Matrix expected(parameter_block_size_, parameter_block_size_);
747     Matrix actual(parameter_block_size_, parameter_block_size_);
748     const double kTolerance = 1e-16;
749 
750     for (int i = 0; i < num_parameter_blocks_; ++i) {
751       expected.setIdentity();
752       expected /= (i + 1.0) * (i + 1.0);
753 
754       double* block_i = parameters_.get() + i * parameter_block_size_;
755       covariance.GetCovarianceBlock(block_i, block_i, actual.data());
756       EXPECT_NEAR((expected - actual).norm(), 0.0, kTolerance)
757           << "block: " << i << ", " << i << "\n"
758           << "expected: \n" << expected << "\n"
759           << "actual: \n" << actual;
760 
761       expected.setZero();
762       for (int j = i + 1; j < num_parameter_blocks_; ++j) {
763         double* block_j = parameters_.get() + j * parameter_block_size_;
764         covariance.GetCovarianceBlock(block_i, block_j, actual.data());
765         EXPECT_NEAR((expected - actual).norm(), 0.0, kTolerance)
766             << "block: " << i << ", " << j << "\n"
767             << "expected: \n" << expected << "\n"
768             << "actual: \n" << actual;
769       }
770     }
771   }
772 
773   scoped_array<double> parameters_;
774   int parameter_block_size_;
775   int num_parameter_blocks_;
776 
777   Problem problem_;
778   vector<pair<const double*, const double*> > all_covariance_blocks_;
779 };
780 
781 #if !defined(CERES_NO_SUITESPARSE) && defined(CERES_USE_OPENMP)
782 
TEST_F(LargeScaleCovarianceTest,Parallel)783 TEST_F(LargeScaleCovarianceTest, Parallel) {
784   ComputeAndCompare(SUITE_SPARSE_QR, 4);
785 }
786 
787 #endif  // !defined(CERES_NO_SUITESPARSE) && defined(CERES_USE_OPENMP)
788 
789 }  // namespace internal
790 }  // namespace ceres
791