// Ceres Solver - A fast non-linear least squares minimizer // Copyright 2010, 2011, 2012 Google Inc. All rights reserved. // http://code.google.com/p/ceres-solver/ // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // // * Redistributions of source code must retain the above copyright notice, // this list of conditions and the following disclaimer. // * Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution. // * Neither the name of Google Inc. nor the names of its contributors may be // used to endorse or promote products derived from this software without // specific prior written permission. // // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE // ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE // LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR // CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF // SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS // INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN // CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) // ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE // POSSIBILITY OF SUCH DAMAGE. // // Author: sameeragarwal@google.com (Sameer Agarwal) // // The LossFunction interface is the way users describe how residuals // are converted to cost terms for the overall problem cost function. // For the exact manner in which loss functions are converted to the // overall cost for a problem, see problem.h. // // For least squares problem where there are no outliers and standard // squared loss is expected, it is not necessary to create a loss // function; instead passing a NULL to the problem when adding // residuals implies a standard squared loss. // // For least squares problems where the minimization may encounter // input terms that contain outliers, that is, completely bogus // measurements, it is important to use a loss function that reduces // their associated penalty. // // Consider a structure from motion problem. The unknowns are 3D // points and camera parameters, and the measurements are image // coordinates describing the expected reprojected position for a // point in a camera. For example, we want to model the geometry of a // street scene with fire hydrants and cars, observed by a moving // camera with unknown parameters, and the only 3D points we care // about are the pointy tippy-tops of the fire hydrants. Our magic // image processing algorithm, which is responsible for producing the // measurements that are input to Ceres, has found and matched all // such tippy-tops in all image frames, except that in one of the // frame it mistook a car's headlight for a hydrant. If we didn't do // anything special (i.e. if we used a basic quadratic loss), the // residual for the erroneous measurement will result in extreme error // due to the quadratic nature of squared loss. This results in the // entire solution getting pulled away from the optimimum to reduce // the large error that would otherwise be attributed to the wrong // measurement. // // Using a robust loss function, the cost for large residuals is // reduced. In the example above, this leads to outlier terms getting // downweighted so they do not overly influence the final solution. // // What cost function is best? // // In general, there isn't a principled way to select a robust loss // function. The authors suggest starting with a non-robust cost, then // only experimenting with robust loss functions if standard squared // loss doesn't work. #ifndef CERES_PUBLIC_LOSS_FUNCTION_H_ #define CERES_PUBLIC_LOSS_FUNCTION_H_ #include "ceres/internal/macros.h" #include "ceres/internal/scoped_ptr.h" #include "ceres/types.h" #include "glog/logging.h" namespace ceres { class LossFunction { public: virtual ~LossFunction() {} // For a residual vector with squared 2-norm 'sq_norm', this method // is required to fill in the value and derivatives of the loss // function (rho in this example): // // out[0] = rho(sq_norm), // out[1] = rho'(sq_norm), // out[2] = rho''(sq_norm), // // Here the convention is that the contribution of a term to the // cost function is given by 1/2 rho(s), where // // s = ||residuals||^2. // // Calling the method with a negative value of 's' is an error and // the implementations are not required to handle that case. // // Most sane choices of rho() satisfy: // // rho(0) = 0, // rho'(0) = 1, // rho'(s) < 1 in outlier region, // rho''(s) < 0 in outlier region, // // so that they mimic the least squares cost for small residuals. virtual void Evaluate(double sq_norm, double out[3]) const = 0; }; // Some common implementations follow below. // // Note: in the region of interest (i.e. s < 3) we have: // TrivialLoss >= HuberLoss >= SoftLOneLoss >= CauchyLoss // This corresponds to no robustification. // // rho(s) = s // // At s = 0: rho = [0, 1, 0]. // // It is not normally necessary to use this, as passing NULL for the // loss function when building the problem accomplishes the same // thing. class TrivialLoss : public LossFunction { public: virtual void Evaluate(double, double*) const; }; // Scaling // ------- // Given one robustifier // s -> rho(s) // one can change the length scale at which robustification takes // place, by adding a scale factor 'a' as follows: // // s -> a^2 rho(s / a^2). // // The first and second derivatives are: // // s -> rho'(s / a^2), // s -> (1 / a^2) rho''(s / a^2), // // but the behaviour near s = 0 is the same as the original function, // i.e. // // rho(s) = s + higher order terms, // a^2 rho(s / a^2) = s + higher order terms. // // The scalar 'a' should be positive. // // The reason for the appearance of squaring is that 'a' is in the // units of the residual vector norm whereas 's' is a squared // norm. For applications it is more convenient to specify 'a' than // its square. The commonly used robustifiers below are described in // un-scaled format (a = 1) but their implementations work for any // non-zero value of 'a'. // Huber. // // rho(s) = s for s <= 1, // rho(s) = 2 sqrt(s) - 1 for s >= 1. // // At s = 0: rho = [0, 1, 0]. // // The scaling parameter 'a' corresponds to 'delta' on this page: // http://en.wikipedia.org/wiki/Huber_Loss_Function class HuberLoss : public LossFunction { public: explicit HuberLoss(double a) : a_(a), b_(a * a) { } virtual void Evaluate(double, double*) const; private: const double a_; // b = a^2. const double b_; }; // Soft L1, similar to Huber but smooth. // // rho(s) = 2 (sqrt(1 + s) - 1). // // At s = 0: rho = [0, 1, -1/2]. class SoftLOneLoss : public LossFunction { public: explicit SoftLOneLoss(double a) : b_(a * a), c_(1 / b_) { } virtual void Evaluate(double, double*) const; private: // b = a^2. const double b_; // c = 1 / a^2. const double c_; }; // Inspired by the Cauchy distribution // // rho(s) = log(1 + s). // // At s = 0: rho = [0, 1, -1]. class CauchyLoss : public LossFunction { public: explicit CauchyLoss(double a) : b_(a * a), c_(1 / b_) { } virtual void Evaluate(double, double*) const; private: // b = a^2. const double b_; // c = 1 / a^2. const double c_; }; // Loss that is capped beyond a certain level using the arc-tangent function. // The scaling parameter 'a' determines the level where falloff occurs. // For costs much smaller than 'a', the loss function is linear and behaves like // TrivialLoss, and for values much larger than 'a' the value asymptotically // approaches the constant value of a * PI / 2. // // rho(s) = a atan(s / a). // // At s = 0: rho = [0, 1, 0]. class ArctanLoss : public LossFunction { public: explicit ArctanLoss(double a) : a_(a), b_(1 / (a * a)) { } virtual void Evaluate(double, double*) const; private: const double a_; // b = 1 / a^2. const double b_; }; // Loss function that maps to approximately zero cost in a range around the // origin, and reverts to linear in error (quadratic in cost) beyond this range. // The tolerance parameter 'a' sets the nominal point at which the // transition occurs, and the transition size parameter 'b' sets the nominal // distance over which most of the transition occurs. Both a and b must be // greater than zero, and typically b will be set to a fraction of a. // The slope rho'[s] varies smoothly from about 0 at s <= a - b to // about 1 at s >= a + b. // // The term is computed as: // // rho(s) = b log(1 + exp((s - a) / b)) - c0. // // where c0 is chosen so that rho(0) == 0 // // c0 = b log(1 + exp(-a / b) // // This has the following useful properties: // // rho(s) == 0 for s = 0 // rho'(s) ~= 0 for s << a - b // rho'(s) ~= 1 for s >> a + b // rho''(s) > 0 for all s // // In addition, all derivatives are continuous, and the curvature is // concentrated in the range a - b to a + b. // // At s = 0: rho = [0, ~0, ~0]. class TolerantLoss : public LossFunction { public: explicit TolerantLoss(double a, double b); virtual void Evaluate(double, double*) const; private: const double a_, b_, c_; }; // Composition of two loss functions. The error is the result of first // evaluating g followed by f to yield the composition f(g(s)). // The loss functions must not be NULL. class ComposedLoss : public LossFunction { public: explicit ComposedLoss(const LossFunction* f, Ownership ownership_f, const LossFunction* g, Ownership ownership_g); virtual ~ComposedLoss(); virtual void Evaluate(double, double*) const; private: internal::scoped_ptr<const LossFunction> f_, g_; const Ownership ownership_f_, ownership_g_; }; // The discussion above has to do with length scaling: it affects the space // in which s is measured. Sometimes you want to simply scale the output // value of the robustifier. For example, you might want to weight // different error terms differently (e.g., weight pixel reprojection // errors differently from terrain errors). // // If rho is the wrapped robustifier, then this simply outputs // s -> a * rho(s) // // The first and second derivatives are, not surprisingly // s -> a * rho'(s) // s -> a * rho''(s) // // Since we treat the a NULL Loss function as the Identity loss // function, rho = NULL is a valid input and will result in the input // being scaled by a. This provides a simple way of implementing a // scaled ResidualBlock. class ScaledLoss : public LossFunction { public: // Constructs a ScaledLoss wrapping another loss function. Takes // ownership of the wrapped loss function or not depending on the // ownership parameter. ScaledLoss(const LossFunction* rho, double a, Ownership ownership) : rho_(rho), a_(a), ownership_(ownership) { } virtual ~ScaledLoss() { if (ownership_ == DO_NOT_TAKE_OWNERSHIP) { rho_.release(); } } virtual void Evaluate(double, double*) const; private: internal::scoped_ptr<const LossFunction> rho_; const double a_; const Ownership ownership_; CERES_DISALLOW_COPY_AND_ASSIGN(ScaledLoss); }; // Sometimes after the optimization problem has been constructed, we // wish to mutate the scale of the loss function. For example, when // performing estimation from data which has substantial outliers, // convergence can be improved by starting out with a large scale, // optimizing the problem and then reducing the scale. This can have // better convergence behaviour than just using a loss function with a // small scale. // // This templated class allows the user to implement a loss function // whose scale can be mutated after an optimization problem has been // constructed. // // Example usage // // Problem problem; // // // Add parameter blocks // // CostFunction* cost_function = // new AutoDiffCostFunction < UW_Camera_Mapper, 2, 9, 3>( // new UW_Camera_Mapper(feature_x, feature_y)); // // LossFunctionWrapper* loss_function(new HuberLoss(1.0), TAKE_OWNERSHIP); // // problem.AddResidualBlock(cost_function, loss_function, parameters); // // Solver::Options options; // Solger::Summary summary; // // Solve(options, &problem, &summary) // // loss_function->Reset(new HuberLoss(1.0), TAKE_OWNERSHIP); // // Solve(options, &problem, &summary) // class LossFunctionWrapper : public LossFunction { public: LossFunctionWrapper(LossFunction* rho, Ownership ownership) : rho_(rho), ownership_(ownership) { } virtual ~LossFunctionWrapper() { if (ownership_ == DO_NOT_TAKE_OWNERSHIP) { rho_.release(); } } virtual void Evaluate(double sq_norm, double out[3]) const { CHECK_NOTNULL(rho_.get()); rho_->Evaluate(sq_norm, out); } void Reset(LossFunction* rho, Ownership ownership) { if (ownership_ == DO_NOT_TAKE_OWNERSHIP) { rho_.release(); } rho_.reset(rho); ownership_ = ownership; } private: internal::scoped_ptr<const LossFunction> rho_; Ownership ownership_; CERES_DISALLOW_COPY_AND_ASSIGN(LossFunctionWrapper); }; } // namespace ceres #endif // CERES_PUBLIC_LOSS_FUNCTION_H_