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 (at) google.com (Sameer Agarwal) 30 31 #include "ceres/corrector.h" 32 33 #include <cstddef> 34 #include <cmath> 35 #include "glog/logging.h" 36 37 namespace ceres { 38 namespace internal { 39 40 Corrector::Corrector(double sq_norm, const double rho[3]) { 41 CHECK_GE(sq_norm, 0.0); 42 CHECK_GT(rho[1], 0.0); 43 sqrt_rho1_ = sqrt(rho[1]); 44 45 // If sq_norm = 0.0, the correction becomes trivial, the residual 46 // and the jacobian are scaled by the squareroot of the derivative 47 // of rho. Handling this case explicitly avoids the divide by zero 48 // error that would occur below. 49 // 50 // The case where rho'' < 0 also gets special handling. Technically 51 // it shouldn't, and the computation of the scaling should proceed 52 // as below, however we found in experiments that applying the 53 // curvature correction when rho'' < 0, which is the case when we 54 // are in the outlier region slows down the convergence of the 55 // algorithm significantly. 56 // 57 // Thus, we have divided the action of the robustifier into two 58 // parts. In the inliner region, we do the full second order 59 // correction which re-wights the gradient of the function by the 60 // square root of the derivative of rho, and the Gauss-Newton 61 // Hessian gets both the scaling and the rank-1 curvature 62 // correction. Normaly, alpha is upper bounded by one, but with this 63 // change, alpha is bounded above by zero. 64 // 65 // Empirically we have observed that the full Triggs correction and 66 // the clamped correction both start out as very good approximations 67 // to the loss function when we are in the convex part of the 68 // function, but as the function starts transitioning from convex to 69 // concave, the Triggs approximation diverges more and more and 70 // ultimately becomes linear. The clamped Triggs model however 71 // remains quadratic. 72 // 73 // The reason why the Triggs approximation becomes so poor is 74 // because the curvature correction that it applies to the gauss 75 // newton hessian goes from being a full rank correction to a rank 76 // deficient correction making the inversion of the Hessian fraught 77 // with all sorts of misery and suffering. 78 // 79 // The clamped correction retains its quadratic nature and inverting it 80 // is always well formed. 81 if ((sq_norm == 0.0) || (rho[2] <= 0.0)) { 82 residual_scaling_ = sqrt_rho1_; 83 alpha_sq_norm_ = 0.0; 84 return; 85 } 86 87 // Calculate the smaller of the two solutions to the equation 88 // 89 // 0.5 * alpha^2 - alpha - rho'' / rho' * z'z = 0. 90 // 91 // Start by calculating the discriminant D. 92 const double D = 1.0 + 2.0 * sq_norm * rho[2] / rho[1]; 93 94 // Since both rho[1] and rho[2] are guaranteed to be positive at 95 // this point, we know that D > 1.0. 96 97 const double alpha = 1.0 - sqrt(D); 98 99 // Calculate the constants needed by the correction routines. 100 residual_scaling_ = sqrt_rho1_ / (1 - alpha); 101 alpha_sq_norm_ = alpha / sq_norm; 102 } 103 104 void Corrector::CorrectResiduals(int num_rows, double* residuals) { 105 DCHECK(residuals != NULL); 106 // Equation 11 in BANS. 107 for (int r = 0; r < num_rows; ++r) { 108 residuals[r] *= residual_scaling_; 109 } 110 } 111 112 void Corrector::CorrectJacobian(int num_rows, 113 int num_cols, 114 double* residuals, 115 double* jacobian) { 116 DCHECK(residuals != NULL); 117 DCHECK(jacobian != NULL); 118 // Equation 11 in BANS. 119 // 120 // J = sqrt(rho) * (J - alpha^2 r * r' J) 121 // 122 // In days gone by this loop used to be a single Eigen expression of 123 // the form 124 // 125 // J = sqrt_rho1_ * (J - alpha_sq_norm_ * r* (r.transpose() * J)); 126 // 127 // Which turns out to about 17x slower on bal problems. The reason 128 // is that Eigen is unable to figure out that this expression can be 129 // evaluated columnwise and ends up creating a temporary. 130 for (int c = 0; c < num_cols; ++c) { 131 double r_transpose_j = 0.0; 132 for (int r = 0; r < num_rows; ++r) { 133 r_transpose_j += jacobian[r * num_cols + c] * residuals[r]; 134 } 135 136 for (int r = 0; r < num_rows; ++r) { 137 jacobian[r * num_cols + c] = sqrt_rho1_ * 138 (jacobian[r * num_cols + c] - 139 alpha_sq_norm_ * residuals[r] * r_transpose_j); 140 } 141 } 142 } 143 144 } // namespace internal 145 } // namespace ceres 146
