• 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 //         keir@google.com (Keir Mierle)
31 //
32 // The Problem object is used to build and hold least squares problems.
33 
34 #ifndef CERES_PUBLIC_PROBLEM_H_
35 #define CERES_PUBLIC_PROBLEM_H_
36 
37 #include <cstddef>
38 #include <map>
39 #include <set>
40 #include <vector>
41 
42 #include "glog/logging.h"
43 #include "ceres/internal/macros.h"
44 #include "ceres/internal/port.h"
45 #include "ceres/internal/scoped_ptr.h"
46 #include "ceres/types.h"
47 #include "ceres/internal/disable_warnings.h"
48 
49 
50 namespace ceres {
51 
52 class CostFunction;
53 class LossFunction;
54 class LocalParameterization;
55 class Solver;
56 struct CRSMatrix;
57 
58 namespace internal {
59 class Preprocessor;
60 class ProblemImpl;
61 class ParameterBlock;
62 class ResidualBlock;
63 }  // namespace internal
64 
65 // A ResidualBlockId is an opaque handle clients can use to remove residual
66 // blocks from a Problem after adding them.
67 typedef internal::ResidualBlock* ResidualBlockId;
68 
69 // A class to represent non-linear least squares problems. Such
70 // problems have a cost function that is a sum of error terms (known
71 // as "residuals"), where each residual is a function of some subset
72 // of the parameters. The cost function takes the form
73 //
74 //    N    1
75 //   SUM  --- loss( || r_i1, r_i2,..., r_ik ||^2  ),
76 //   i=1   2
77 //
78 // where
79 //
80 //   r_ij     is residual number i, component j; the residual is a
81 //            function of some subset of the parameters x1...xk. For
82 //            example, in a structure from motion problem a residual
83 //            might be the difference between a measured point in an
84 //            image and the reprojected position for the matching
85 //            camera, point pair. The residual would have two
86 //            components, error in x and error in y.
87 //
88 //   loss(y)  is the loss function; for example, squared error or
89 //            Huber L1 loss. If loss(y) = y, then the cost function is
90 //            non-robustified least squares.
91 //
92 // This class is specifically designed to address the important subset
93 // of "sparse" least squares problems, where each component of the
94 // residual depends only on a small number number of parameters, even
95 // though the total number of residuals and parameters may be very
96 // large. This property affords tremendous gains in scale, allowing
97 // efficient solving of large problems that are otherwise
98 // inaccessible.
99 //
100 // The canonical example of a sparse least squares problem is
101 // "structure-from-motion" (SFM), where the parameters are points and
102 // cameras, and residuals are reprojection errors. Typically a single
103 // residual will depend only on 9 parameters (3 for the point, 6 for
104 // the camera).
105 //
106 // To create a least squares problem, use the AddResidualBlock() and
107 // AddParameterBlock() methods, documented below. Here is an example least
108 // squares problem containing 3 parameter blocks of sizes 3, 4 and 5
109 // respectively and two residual terms of size 2 and 6:
110 //
111 //   double x1[] = { 1.0, 2.0, 3.0 };
112 //   double x2[] = { 1.0, 2.0, 3.0, 5.0 };
113 //   double x3[] = { 1.0, 2.0, 3.0, 6.0, 7.0 };
114 //
115 //   Problem problem;
116 //
117 //   problem.AddResidualBlock(new MyUnaryCostFunction(...), x1);
118 //   problem.AddResidualBlock(new MyBinaryCostFunction(...), x2, x3);
119 //
120 // Please see cost_function.h for details of the CostFunction object.
121 class CERES_EXPORT Problem {
122  public:
123   struct CERES_EXPORT Options {
OptionsOptions124     Options()
125         : cost_function_ownership(TAKE_OWNERSHIP),
126           loss_function_ownership(TAKE_OWNERSHIP),
127           local_parameterization_ownership(TAKE_OWNERSHIP),
128           enable_fast_removal(false),
129           disable_all_safety_checks(false) {}
130 
131     // These flags control whether the Problem object owns the cost
132     // functions, loss functions, and parameterizations passed into
133     // the Problem. If set to TAKE_OWNERSHIP, then the problem object
134     // will delete the corresponding cost or loss functions on
135     // destruction. The destructor is careful to delete the pointers
136     // only once, since sharing cost/loss/parameterizations is
137     // allowed.
138     Ownership cost_function_ownership;
139     Ownership loss_function_ownership;
140     Ownership local_parameterization_ownership;
141 
142     // If true, trades memory for faster RemoveResidualBlock() and
143     // RemoveParameterBlock() operations.
144     //
145     // By default, RemoveParameterBlock() and RemoveResidualBlock() take time
146     // proportional to the size of the entire problem.  If you only ever remove
147     // parameters or residuals from the problem occassionally, this might be
148     // acceptable.  However, if you have memory to spare, enable this option to
149     // make RemoveParameterBlock() take time proportional to the number of
150     // residual blocks that depend on it, and RemoveResidualBlock() take (on
151     // average) constant time.
152     //
153     // The increase in memory usage is twofold: an additonal hash set per
154     // parameter block containing all the residuals that depend on the parameter
155     // block; and a hash set in the problem containing all residuals.
156     bool enable_fast_removal;
157 
158     // By default, Ceres performs a variety of safety checks when constructing
159     // the problem. There is a small but measurable performance penalty to
160     // these checks, typically around 5% of construction time. If you are sure
161     // your problem construction is correct, and 5% of the problem construction
162     // time is truly an overhead you want to avoid, then you can set
163     // disable_all_safety_checks to true.
164     //
165     // WARNING: Do not set this to true, unless you are absolutely sure of what
166     // you are doing.
167     bool disable_all_safety_checks;
168   };
169 
170   // The default constructor is equivalent to the
171   // invocation Problem(Problem::Options()).
172   Problem();
173   explicit Problem(const Options& options);
174 
175   ~Problem();
176 
177   // Add a residual block to the overall cost function. The cost
178   // function carries with it information about the sizes of the
179   // parameter blocks it expects. The function checks that these match
180   // the sizes of the parameter blocks listed in parameter_blocks. The
181   // program aborts if a mismatch is detected. loss_function can be
182   // NULL, in which case the cost of the term is just the squared norm
183   // of the residuals.
184   //
185   // The user has the option of explicitly adding the parameter blocks
186   // using AddParameterBlock. This causes additional correctness
187   // checking; however, AddResidualBlock implicitly adds the parameter
188   // blocks if they are not present, so calling AddParameterBlock
189   // explicitly is not required.
190   //
191   // The Problem object by default takes ownership of the
192   // cost_function and loss_function pointers. These objects remain
193   // live for the life of the Problem object. If the user wishes to
194   // keep control over the destruction of these objects, then they can
195   // do this by setting the corresponding enums in the Options struct.
196   //
197   // Note: Even though the Problem takes ownership of cost_function
198   // and loss_function, it does not preclude the user from re-using
199   // them in another residual block. The destructor takes care to call
200   // delete on each cost_function or loss_function pointer only once,
201   // regardless of how many residual blocks refer to them.
202   //
203   // Example usage:
204   //
205   //   double x1[] = {1.0, 2.0, 3.0};
206   //   double x2[] = {1.0, 2.0, 5.0, 6.0};
207   //   double x3[] = {3.0, 6.0, 2.0, 5.0, 1.0};
208   //
209   //   Problem problem;
210   //
211   //   problem.AddResidualBlock(new MyUnaryCostFunction(...), NULL, x1);
212   //   problem.AddResidualBlock(new MyBinaryCostFunction(...), NULL, x2, x1);
213   //
214   ResidualBlockId AddResidualBlock(CostFunction* cost_function,
215                                    LossFunction* loss_function,
216                                    const vector<double*>& parameter_blocks);
217 
218   // Convenience methods for adding residuals with a small number of
219   // parameters. This is the common case. Instead of specifying the
220   // parameter block arguments as a vector, list them as pointers.
221   ResidualBlockId AddResidualBlock(CostFunction* cost_function,
222                                    LossFunction* loss_function,
223                                    double* x0);
224   ResidualBlockId AddResidualBlock(CostFunction* cost_function,
225                                    LossFunction* loss_function,
226                                    double* x0, double* x1);
227   ResidualBlockId AddResidualBlock(CostFunction* cost_function,
228                                    LossFunction* loss_function,
229                                    double* x0, double* x1, double* x2);
230   ResidualBlockId AddResidualBlock(CostFunction* cost_function,
231                                    LossFunction* loss_function,
232                                    double* x0, double* x1, double* x2,
233                                    double* x3);
234   ResidualBlockId AddResidualBlock(CostFunction* cost_function,
235                                    LossFunction* loss_function,
236                                    double* x0, double* x1, double* x2,
237                                    double* x3, double* x4);
238   ResidualBlockId AddResidualBlock(CostFunction* cost_function,
239                                    LossFunction* loss_function,
240                                    double* x0, double* x1, double* x2,
241                                    double* x3, double* x4, double* x5);
242   ResidualBlockId AddResidualBlock(CostFunction* cost_function,
243                                    LossFunction* loss_function,
244                                    double* x0, double* x1, double* x2,
245                                    double* x3, double* x4, double* x5,
246                                    double* x6);
247   ResidualBlockId AddResidualBlock(CostFunction* cost_function,
248                                    LossFunction* loss_function,
249                                    double* x0, double* x1, double* x2,
250                                    double* x3, double* x4, double* x5,
251                                    double* x6, double* x7);
252   ResidualBlockId AddResidualBlock(CostFunction* cost_function,
253                                    LossFunction* loss_function,
254                                    double* x0, double* x1, double* x2,
255                                    double* x3, double* x4, double* x5,
256                                    double* x6, double* x7, double* x8);
257   ResidualBlockId AddResidualBlock(CostFunction* cost_function,
258                                    LossFunction* loss_function,
259                                    double* x0, double* x1, double* x2,
260                                    double* x3, double* x4, double* x5,
261                                    double* x6, double* x7, double* x8,
262                                    double* x9);
263 
264   // Add a parameter block with appropriate size to the problem.
265   // Repeated calls with the same arguments are ignored. Repeated
266   // calls with the same double pointer but a different size results
267   // in undefined behaviour.
268   void AddParameterBlock(double* values, int size);
269 
270   // Add a parameter block with appropriate size and parameterization
271   // to the problem. Repeated calls with the same arguments are
272   // ignored. Repeated calls with the same double pointer but a
273   // different size results in undefined behaviour.
274   void AddParameterBlock(double* values,
275                          int size,
276                          LocalParameterization* local_parameterization);
277 
278   // Remove a parameter block from the problem. The parameterization of the
279   // parameter block, if it exists, will persist until the deletion of the
280   // problem (similar to cost/loss functions in residual block removal). Any
281   // residual blocks that depend on the parameter are also removed, as
282   // described above in RemoveResidualBlock().
283   //
284   // If Problem::Options::enable_fast_removal is true, then the
285   // removal is fast (almost constant time). Otherwise, removing a parameter
286   // block will incur a scan of the entire Problem object.
287   //
288   // WARNING: Removing a residual or parameter block will destroy the implicit
289   // ordering, rendering the jacobian or residuals returned from the solver
290   // uninterpretable. If you depend on the evaluated jacobian, do not use
291   // remove! This may change in a future release.
292   void RemoveParameterBlock(double* values);
293 
294   // Remove a residual block from the problem. Any parameters that the residual
295   // block depends on are not removed. The cost and loss functions for the
296   // residual block will not get deleted immediately; won't happen until the
297   // problem itself is deleted.
298   //
299   // WARNING: Removing a residual or parameter block will destroy the implicit
300   // ordering, rendering the jacobian or residuals returned from the solver
301   // uninterpretable. If you depend on the evaluated jacobian, do not use
302   // remove! This may change in a future release.
303   void RemoveResidualBlock(ResidualBlockId residual_block);
304 
305   // Hold the indicated parameter block constant during optimization.
306   void SetParameterBlockConstant(double* values);
307 
308   // Allow the indicated parameter block to vary during optimization.
309   void SetParameterBlockVariable(double* values);
310 
311   // Set the local parameterization for one of the parameter blocks.
312   // The local_parameterization is owned by the Problem by default. It
313   // is acceptable to set the same parameterization for multiple
314   // parameters; the destructor is careful to delete local
315   // parameterizations only once. The local parameterization can only
316   // be set once per parameter, and cannot be changed once set.
317   void SetParameterization(double* values,
318                            LocalParameterization* local_parameterization);
319 
320   // Get the local parameterization object associated with this
321   // parameter block. If there is no parameterization object
322   // associated then NULL is returned.
323   const LocalParameterization* GetParameterization(double* values) const;
324 
325   // Set the lower/upper bound for the parameter with position "index".
326   void SetParameterLowerBound(double* values, int index, double lower_bound);
327   void SetParameterUpperBound(double* values, int index, double upper_bound);
328 
329   // Number of parameter blocks in the problem. Always equals
330   // parameter_blocks().size() and parameter_block_sizes().size().
331   int NumParameterBlocks() const;
332 
333   // The size of the parameter vector obtained by summing over the
334   // sizes of all the parameter blocks.
335   int NumParameters() const;
336 
337   // Number of residual blocks in the problem. Always equals
338   // residual_blocks().size().
339   int NumResidualBlocks() const;
340 
341   // The size of the residual vector obtained by summing over the
342   // sizes of all of the residual blocks.
343   int NumResiduals() const;
344 
345   // The size of the parameter block.
346   int ParameterBlockSize(const double* values) const;
347 
348   // The size of local parameterization for the parameter block. If
349   // there is no local parameterization associated with this parameter
350   // block, then ParameterBlockLocalSize = ParameterBlockSize.
351   int ParameterBlockLocalSize(const double* values) const;
352 
353   // Is the given parameter block present in this problem or not?
354   bool HasParameterBlock(const double* values) const;
355 
356   // Fills the passed parameter_blocks vector with pointers to the
357   // parameter blocks currently in the problem. After this call,
358   // parameter_block.size() == NumParameterBlocks.
359   void GetParameterBlocks(vector<double*>* parameter_blocks) const;
360 
361   // Fills the passed residual_blocks vector with pointers to the
362   // residual blocks currently in the problem. After this call,
363   // residual_blocks.size() == NumResidualBlocks.
364   void GetResidualBlocks(vector<ResidualBlockId>* residual_blocks) const;
365 
366   // Get all the parameter blocks that depend on the given residual block.
367   void GetParameterBlocksForResidualBlock(
368       const ResidualBlockId residual_block,
369       vector<double*>* parameter_blocks) const;
370 
371   // Get all the residual blocks that depend on the given parameter block.
372   //
373   // If Problem::Options::enable_fast_removal is true, then
374   // getting the residual blocks is fast and depends only on the number of
375   // residual blocks. Otherwise, getting the residual blocks for a parameter
376   // block will incur a scan of the entire Problem object.
377   void GetResidualBlocksForParameterBlock(
378       const double* values,
379       vector<ResidualBlockId>* residual_blocks) const;
380 
381   // Options struct to control Problem::Evaluate.
382   struct EvaluateOptions {
EvaluateOptionsEvaluateOptions383     EvaluateOptions()
384         : apply_loss_function(true),
385           num_threads(1) {
386     }
387 
388     // The set of parameter blocks for which evaluation should be
389     // performed. This vector determines the order that parameter
390     // blocks occur in the gradient vector and in the columns of the
391     // jacobian matrix. If parameter_blocks is empty, then it is
392     // assumed to be equal to vector containing ALL the parameter
393     // blocks.  Generally speaking the parameter blocks will occur in
394     // the order in which they were added to the problem. But, this
395     // may change if the user removes any parameter blocks from the
396     // problem.
397     //
398     // NOTE: This vector should contain the same pointers as the ones
399     // used to add parameter blocks to the Problem. These parameter
400     // block should NOT point to new memory locations. Bad things will
401     // happen otherwise.
402     vector<double*> parameter_blocks;
403 
404     // The set of residual blocks to evaluate. This vector determines
405     // the order in which the residuals occur, and how the rows of the
406     // jacobian are ordered. If residual_blocks is empty, then it is
407     // assumed to be equal to the vector containing all the residual
408     // blocks. If this vector is empty, then it is assumed to be equal
409     // to a vector containing ALL the residual blocks. Generally
410     // speaking the residual blocks will occur in the order in which
411     // they were added to the problem. But, this may change if the
412     // user removes any residual blocks from the problem.
413     vector<ResidualBlockId> residual_blocks;
414 
415     // Even though the residual blocks in the problem may contain loss
416     // functions, setting apply_loss_function to false will turn off
417     // the application of the loss function to the output of the cost
418     // function. This is of use for example if the user wishes to
419     // analyse the solution quality by studying the distribution of
420     // residuals before and after the solve.
421     bool apply_loss_function;
422 
423     int num_threads;
424   };
425 
426   // Evaluate Problem. Any of the output pointers can be NULL. Which
427   // residual blocks and parameter blocks are used is controlled by
428   // the EvaluateOptions struct above.
429   //
430   // Note 1: The evaluation will use the values stored in the memory
431   // locations pointed to by the parameter block pointers used at the
432   // time of the construction of the problem. i.e.,
433   //
434   //   Problem problem;
435   //   double x = 1;
436   //   problem.AddResidualBlock(new MyCostFunction, NULL, &x);
437   //
438   //   double cost = 0.0;
439   //   problem.Evaluate(Problem::EvaluateOptions(), &cost, NULL, NULL, NULL);
440   //
441   // The cost is evaluated at x = 1. If you wish to evaluate the
442   // problem at x = 2, then
443   //
444   //    x = 2;
445   //    problem.Evaluate(Problem::EvaluateOptions(), &cost, NULL, NULL, NULL);
446   //
447   // is the way to do so.
448   //
449   // Note 2: If no local parameterizations are used, then the size of
450   // the gradient vector (and the number of columns in the jacobian)
451   // is the sum of the sizes of all the parameter blocks. If a
452   // parameter block has a local parameterization, then it contributes
453   // "LocalSize" entries to the gradient vector (and the number of
454   // columns in the jacobian).
455   bool Evaluate(const EvaluateOptions& options,
456                 double* cost,
457                 vector<double>* residuals,
458                 vector<double>* gradient,
459                 CRSMatrix* jacobian);
460 
461  private:
462   friend class Solver;
463   friend class Covariance;
464   internal::scoped_ptr<internal::ProblemImpl> problem_impl_;
465   CERES_DISALLOW_COPY_AND_ASSIGN(Problem);
466 };
467 
468 }  // namespace ceres
469 
470 #include "ceres/internal/reenable_warnings.h"
471 
472 #endif  // CERES_PUBLIC_PROBLEM_H_
473