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: sameeragarwal@google.com (Sameer Agarwal)
30 //
31 // The LossFunction interface is the way users describe how residuals
32 // are converted to cost terms for the overall problem cost function.
33 // For the exact manner in which loss functions are converted to the
34 // overall cost for a problem, see problem.h.
35 //
36 // For least squares problem where there are no outliers and standard
37 // squared loss is expected, it is not necessary to create a loss
38 // function; instead passing a NULL to the problem when adding
39 // residuals implies a standard squared loss.
40 //
41 // For least squares problems where the minimization may encounter
42 // input terms that contain outliers, that is, completely bogus
43 // measurements, it is important to use a loss function that reduces
44 // their associated penalty.
45 //
46 // Consider a structure from motion problem. The unknowns are 3D
47 // points and camera parameters, and the measurements are image
48 // coordinates describing the expected reprojected position for a
49 // point in a camera. For example, we want to model the geometry of a
50 // street scene with fire hydrants and cars, observed by a moving
51 // camera with unknown parameters, and the only 3D points we care
52 // about are the pointy tippy-tops of the fire hydrants. Our magic
53 // image processing algorithm, which is responsible for producing the
54 // measurements that are input to Ceres, has found and matched all
55 // such tippy-tops in all image frames, except that in one of the
56 // frame it mistook a car's headlight for a hydrant. If we didn't do
57 // anything special (i.e. if we used a basic quadratic loss), the
58 // residual for the erroneous measurement will result in extreme error
59 // due to the quadratic nature of squared loss. This results in the
60 // entire solution getting pulled away from the optimimum to reduce
61 // the large error that would otherwise be attributed to the wrong
62 // measurement.
63 //
64 // Using a robust loss function, the cost for large residuals is
65 // reduced. In the example above, this leads to outlier terms getting
66 // downweighted so they do not overly influence the final solution.
67 //
68 // What cost function is best?
69 //
70 // In general, there isn't a principled way to select a robust loss
71 // function. The authors suggest starting with a non-robust cost, then
72 // only experimenting with robust loss functions if standard squared
73 // loss doesn't work.
74 
75 #ifndef CERES_PUBLIC_LOSS_FUNCTION_H_
76 #define CERES_PUBLIC_LOSS_FUNCTION_H_
77 
78 #include "glog/logging.h"
79 #include "ceres/internal/macros.h"
80 #include "ceres/internal/scoped_ptr.h"
81 #include "ceres/types.h"
82 #include "ceres/internal/disable_warnings.h"
83 
84 namespace ceres {
85 
86 class CERES_EXPORT LossFunction {
87  public:
~LossFunction()88   virtual ~LossFunction() {}
89 
90   // For a residual vector with squared 2-norm 'sq_norm', this method
91   // is required to fill in the value and derivatives of the loss
92   // function (rho in this example):
93   //
94   //   out[0] = rho(sq_norm),
95   //   out[1] = rho'(sq_norm),
96   //   out[2] = rho''(sq_norm),
97   //
98   // Here the convention is that the contribution of a term to the
99   // cost function is given by 1/2 rho(s),  where
100   //
101   //   s = ||residuals||^2.
102   //
103   // Calling the method with a negative value of 's' is an error and
104   // the implementations are not required to handle that case.
105   //
106   // Most sane choices of rho() satisfy:
107   //
108   //   rho(0) = 0,
109   //   rho'(0) = 1,
110   //   rho'(s) < 1 in outlier region,
111   //   rho''(s) < 0 in outlier region,
112   //
113   // so that they mimic the least squares cost for small residuals.
114   virtual void Evaluate(double sq_norm, double out[3]) const = 0;
115 };
116 
117 // Some common implementations follow below.
118 //
119 // Note: in the region of interest (i.e. s < 3) we have:
120 //   TrivialLoss >= HuberLoss >= SoftLOneLoss >= CauchyLoss
121 
122 
123 // This corresponds to no robustification.
124 //
125 //   rho(s) = s
126 //
127 // At s = 0: rho = [0, 1, 0].
128 //
129 // It is not normally necessary to use this, as passing NULL for the
130 // loss function when building the problem accomplishes the same
131 // thing.
132 class CERES_EXPORT TrivialLoss : public LossFunction {
133  public:
134   virtual void Evaluate(double, double*) const;
135 };
136 
137 // Scaling
138 // -------
139 // Given one robustifier
140 //   s -> rho(s)
141 // one can change the length scale at which robustification takes
142 // place, by adding a scale factor 'a' as follows:
143 //
144 //   s -> a^2 rho(s / a^2).
145 //
146 // The first and second derivatives are:
147 //
148 //   s -> rho'(s / a^2),
149 //   s -> (1 / a^2) rho''(s / a^2),
150 //
151 // but the behaviour near s = 0 is the same as the original function,
152 // i.e.
153 //
154 //   rho(s) = s + higher order terms,
155 //   a^2 rho(s / a^2) = s + higher order terms.
156 //
157 // The scalar 'a' should be positive.
158 //
159 // The reason for the appearance of squaring is that 'a' is in the
160 // units of the residual vector norm whereas 's' is a squared
161 // norm. For applications it is more convenient to specify 'a' than
162 // its square. The commonly used robustifiers below are described in
163 // un-scaled format (a = 1) but their implementations work for any
164 // non-zero value of 'a'.
165 
166 // Huber.
167 //
168 //   rho(s) = s               for s <= 1,
169 //   rho(s) = 2 sqrt(s) - 1   for s >= 1.
170 //
171 // At s = 0: rho = [0, 1, 0].
172 //
173 // The scaling parameter 'a' corresponds to 'delta' on this page:
174 //   http://en.wikipedia.org/wiki/Huber_Loss_Function
175 class CERES_EXPORT HuberLoss : public LossFunction {
176  public:
HuberLoss(double a)177   explicit HuberLoss(double a) : a_(a), b_(a * a) { }
178   virtual void Evaluate(double, double*) const;
179 
180  private:
181   const double a_;
182   // b = a^2.
183   const double b_;
184 };
185 
186 // Soft L1, similar to Huber but smooth.
187 //
188 //   rho(s) = 2 (sqrt(1 + s) - 1).
189 //
190 // At s = 0: rho = [0, 1, -1/2].
191 class CERES_EXPORT SoftLOneLoss : public LossFunction {
192  public:
SoftLOneLoss(double a)193   explicit SoftLOneLoss(double a) : b_(a * a), c_(1 / b_) { }
194   virtual void Evaluate(double, double*) const;
195 
196  private:
197   // b = a^2.
198   const double b_;
199   // c = 1 / a^2.
200   const double c_;
201 };
202 
203 // Inspired by the Cauchy distribution
204 //
205 //   rho(s) = log(1 + s).
206 //
207 // At s = 0: rho = [0, 1, -1].
208 class CERES_EXPORT CauchyLoss : public LossFunction {
209  public:
CauchyLoss(double a)210   explicit CauchyLoss(double a) : b_(a * a), c_(1 / b_) { }
211   virtual void Evaluate(double, double*) const;
212 
213  private:
214   // b = a^2.
215   const double b_;
216   // c = 1 / a^2.
217   const double c_;
218 };
219 
220 // Loss that is capped beyond a certain level using the arc-tangent function.
221 // The scaling parameter 'a' determines the level where falloff occurs.
222 // For costs much smaller than 'a', the loss function is linear and behaves like
223 // TrivialLoss, and for values much larger than 'a' the value asymptotically
224 // approaches the constant value of a * PI / 2.
225 //
226 //   rho(s) = a atan(s / a).
227 //
228 // At s = 0: rho = [0, 1, 0].
229 class CERES_EXPORT ArctanLoss : public LossFunction {
230  public:
ArctanLoss(double a)231   explicit ArctanLoss(double a) : a_(a), b_(1 / (a * a)) { }
232   virtual void Evaluate(double, double*) const;
233 
234  private:
235   const double a_;
236   // b = 1 / a^2.
237   const double b_;
238 };
239 
240 // Loss function that maps to approximately zero cost in a range around the
241 // origin, and reverts to linear in error (quadratic in cost) beyond this range.
242 // The tolerance parameter 'a' sets the nominal point at which the
243 // transition occurs, and the transition size parameter 'b' sets the nominal
244 // distance over which most of the transition occurs. Both a and b must be
245 // greater than zero, and typically b will be set to a fraction of a.
246 // The slope rho'[s] varies smoothly from about 0 at s <= a - b to
247 // about 1 at s >= a + b.
248 //
249 // The term is computed as:
250 //
251 //   rho(s) = b log(1 + exp((s - a) / b)) - c0.
252 //
253 // where c0 is chosen so that rho(0) == 0
254 //
255 //   c0 = b log(1 + exp(-a / b)
256 //
257 // This has the following useful properties:
258 //
259 //   rho(s) == 0               for s = 0
260 //   rho'(s) ~= 0              for s << a - b
261 //   rho'(s) ~= 1              for s >> a + b
262 //   rho''(s) > 0              for all s
263 //
264 // In addition, all derivatives are continuous, and the curvature is
265 // concentrated in the range a - b to a + b.
266 //
267 // At s = 0: rho = [0, ~0, ~0].
268 class CERES_EXPORT TolerantLoss : public LossFunction {
269  public:
270   explicit TolerantLoss(double a, double b);
271   virtual void Evaluate(double, double*) const;
272 
273  private:
274   const double a_, b_, c_;
275 };
276 
277 // Composition of two loss functions.  The error is the result of first
278 // evaluating g followed by f to yield the composition f(g(s)).
279 // The loss functions must not be NULL.
280 class ComposedLoss : public LossFunction {
281  public:
282   explicit ComposedLoss(const LossFunction* f, Ownership ownership_f,
283                         const LossFunction* g, Ownership ownership_g);
284   virtual ~ComposedLoss();
285   virtual void Evaluate(double, double*) const;
286 
287  private:
288   internal::scoped_ptr<const LossFunction> f_, g_;
289   const Ownership ownership_f_, ownership_g_;
290 };
291 
292 // The discussion above has to do with length scaling: it affects the space
293 // in which s is measured. Sometimes you want to simply scale the output
294 // value of the robustifier. For example, you might want to weight
295 // different error terms differently (e.g., weight pixel reprojection
296 // errors differently from terrain errors).
297 //
298 // If rho is the wrapped robustifier, then this simply outputs
299 // s -> a * rho(s)
300 //
301 // The first and second derivatives are, not surprisingly
302 // s -> a * rho'(s)
303 // s -> a * rho''(s)
304 //
305 // Since we treat the a NULL Loss function as the Identity loss
306 // function, rho = NULL is a valid input and will result in the input
307 // being scaled by a. This provides a simple way of implementing a
308 // scaled ResidualBlock.
309 class CERES_EXPORT ScaledLoss : public LossFunction {
310  public:
311   // Constructs a ScaledLoss wrapping another loss function. Takes
312   // ownership of the wrapped loss function or not depending on the
313   // ownership parameter.
ScaledLoss(const LossFunction * rho,double a,Ownership ownership)314   ScaledLoss(const LossFunction* rho, double a, Ownership ownership) :
315       rho_(rho), a_(a), ownership_(ownership) { }
316 
~ScaledLoss()317   virtual ~ScaledLoss() {
318     if (ownership_ == DO_NOT_TAKE_OWNERSHIP) {
319       rho_.release();
320     }
321   }
322   virtual void Evaluate(double, double*) const;
323 
324  private:
325   internal::scoped_ptr<const LossFunction> rho_;
326   const double a_;
327   const Ownership ownership_;
328   CERES_DISALLOW_COPY_AND_ASSIGN(ScaledLoss);
329 };
330 
331 // Sometimes after the optimization problem has been constructed, we
332 // wish to mutate the scale of the loss function. For example, when
333 // performing estimation from data which has substantial outliers,
334 // convergence can be improved by starting out with a large scale,
335 // optimizing the problem and then reducing the scale. This can have
336 // better convergence behaviour than just using a loss function with a
337 // small scale.
338 //
339 // This templated class allows the user to implement a loss function
340 // whose scale can be mutated after an optimization problem has been
341 // constructed.
342 //
343 // Example usage
344 //
345 //  Problem problem;
346 //
347 //  // Add parameter blocks
348 //
349 //  CostFunction* cost_function =
350 //    new AutoDiffCostFunction < UW_Camera_Mapper, 2, 9, 3>(
351 //      new UW_Camera_Mapper(feature_x, feature_y));
352 //
353 //  LossFunctionWrapper* loss_function(new HuberLoss(1.0), TAKE_OWNERSHIP);
354 //
355 //  problem.AddResidualBlock(cost_function, loss_function, parameters);
356 //
357 //  Solver::Options options;
358 //  Solger::Summary summary;
359 //
360 //  Solve(options, &problem, &summary)
361 //
362 //  loss_function->Reset(new HuberLoss(1.0), TAKE_OWNERSHIP);
363 //
364 //  Solve(options, &problem, &summary)
365 //
366 class CERES_EXPORT LossFunctionWrapper : public LossFunction {
367  public:
LossFunctionWrapper(LossFunction * rho,Ownership ownership)368   LossFunctionWrapper(LossFunction* rho, Ownership ownership)
369       : rho_(rho), ownership_(ownership) {
370   }
371 
~LossFunctionWrapper()372   virtual ~LossFunctionWrapper() {
373     if (ownership_ == DO_NOT_TAKE_OWNERSHIP) {
374       rho_.release();
375     }
376   }
377 
Evaluate(double sq_norm,double out[3])378   virtual void Evaluate(double sq_norm, double out[3]) const {
379     CHECK_NOTNULL(rho_.get());
380     rho_->Evaluate(sq_norm, out);
381   }
382 
Reset(LossFunction * rho,Ownership ownership)383   void Reset(LossFunction* rho, Ownership ownership) {
384     if (ownership_ == DO_NOT_TAKE_OWNERSHIP) {
385       rho_.release();
386     }
387     rho_.reset(rho);
388     ownership_ = ownership;
389   }
390 
391  private:
392   internal::scoped_ptr<const LossFunction> rho_;
393   Ownership ownership_;
394   CERES_DISALLOW_COPY_AND_ASSIGN(LossFunctionWrapper);
395 };
396 
397 }  // namespace ceres
398 
399 #include "ceres/internal/disable_warnings.h"
400 
401 #endif  // CERES_PUBLIC_LOSS_FUNCTION_H_
402