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: keir@google.com (Keir Mierle)
30//         sameeragarwal@google.com (Sameer Agarwal)
31//
32// Templated functions for manipulating rotations. The templated
33// functions are useful when implementing functors for automatic
34// differentiation.
35//
36// In the following, the Quaternions are laid out as 4-vectors, thus:
37//
38//   q[0]  scalar part.
39//   q[1]  coefficient of i.
40//   q[2]  coefficient of j.
41//   q[3]  coefficient of k.
42//
43// where: i*i = j*j = k*k = -1 and i*j = k, j*k = i, k*i = j.
44
45#ifndef CERES_PUBLIC_ROTATION_H_
46#define CERES_PUBLIC_ROTATION_H_
47
48#include <algorithm>
49#include <cmath>
50#include "glog/logging.h"
51
52namespace ceres {
53
54// Trivial wrapper to index linear arrays as matrices, given a fixed
55// column and row stride. When an array "T* array" is wrapped by a
56//
57//   (const) MatrixAdapter<T, row_stride, col_stride> M"
58//
59// the expression  M(i, j) is equivalent to
60//
61//   arrary[i * row_stride + j * col_stride]
62//
63// Conversion functions to and from rotation matrices accept
64// MatrixAdapters to permit using row-major and column-major layouts,
65// and rotation matrices embedded in larger matrices (such as a 3x4
66// projection matrix).
67template <typename T, int row_stride, int col_stride>
68struct MatrixAdapter;
69
70// Convenience functions to create a MatrixAdapter that treats the
71// array pointed to by "pointer" as a 3x3 (contiguous) column-major or
72// row-major matrix.
73template <typename T>
74MatrixAdapter<T, 1, 3> ColumnMajorAdapter3x3(T* pointer);
75
76template <typename T>
77MatrixAdapter<T, 3, 1> RowMajorAdapter3x3(T* pointer);
78
79// Convert a value in combined axis-angle representation to a quaternion.
80// The value angle_axis is a triple whose norm is an angle in radians,
81// and whose direction is aligned with the axis of rotation,
82// and quaternion is a 4-tuple that will contain the resulting quaternion.
83// The implementation may be used with auto-differentiation up to the first
84// derivative, higher derivatives may have unexpected results near the origin.
85template<typename T>
86void AngleAxisToQuaternion(const T* angle_axis, T* quaternion);
87
88// Convert a quaternion to the equivalent combined axis-angle representation.
89// The value quaternion must be a unit quaternion - it is not normalized first,
90// and angle_axis will be filled with a value whose norm is the angle of
91// rotation in radians, and whose direction is the axis of rotation.
92// The implemention may be used with auto-differentiation up to the first
93// derivative, higher derivatives may have unexpected results near the origin.
94template<typename T>
95void QuaternionToAngleAxis(const T* quaternion, T* angle_axis);
96
97// Conversions between 3x3 rotation matrix (in column major order) and
98// axis-angle rotation representations.  Templated for use with
99// autodifferentiation.
100template <typename T>
101void RotationMatrixToAngleAxis(const T* R, T* angle_axis);
102
103template <typename T, int row_stride, int col_stride>
104void RotationMatrixToAngleAxis(
105    const MatrixAdapter<const T, row_stride, col_stride>& R,
106    T* angle_axis);
107
108template <typename T>
109void AngleAxisToRotationMatrix(const T* angle_axis, T* R);
110
111template <typename T, int row_stride, int col_stride>
112void AngleAxisToRotationMatrix(
113    const T* angle_axis,
114    const MatrixAdapter<T, row_stride, col_stride>& R);
115
116// Conversions between 3x3 rotation matrix (in row major order) and
117// Euler angle (in degrees) rotation representations.
118//
119// The {pitch,roll,yaw} Euler angles are rotations around the {x,y,z}
120// axes, respectively.  They are applied in that same order, so the
121// total rotation R is Rz * Ry * Rx.
122template <typename T>
123void EulerAnglesToRotationMatrix(const T* euler, int row_stride, T* R);
124
125template <typename T, int row_stride, int col_stride>
126void EulerAnglesToRotationMatrix(
127    const T* euler,
128    const MatrixAdapter<T, row_stride, col_stride>& R);
129
130// Convert a 4-vector to a 3x3 scaled rotation matrix.
131//
132// The choice of rotation is such that the quaternion [1 0 0 0] goes to an
133// identity matrix and for small a, b, c the quaternion [1 a b c] goes to
134// the matrix
135//
136//         [  0 -c  b ]
137//   I + 2 [  c  0 -a ] + higher order terms
138//         [ -b  a  0 ]
139//
140// which corresponds to a Rodrigues approximation, the last matrix being
141// the cross-product matrix of [a b c]. Together with the property that
142// R(q1 * q2) = R(q1) * R(q2) this uniquely defines the mapping from q to R.
143//
144// The rotation matrix is row-major.
145//
146// No normalization of the quaternion is performed, i.e.
147// R = ||q||^2 * Q, where Q is an orthonormal matrix
148// such that det(Q) = 1 and Q*Q' = I
149template <typename T> inline
150void QuaternionToScaledRotation(const T q[4], T R[3 * 3]);
151
152template <typename T, int row_stride, int col_stride> inline
153void QuaternionToScaledRotation(
154    const T q[4],
155    const MatrixAdapter<T, row_stride, col_stride>& R);
156
157// Same as above except that the rotation matrix is normalized by the
158// Frobenius norm, so that R * R' = I (and det(R) = 1).
159template <typename T> inline
160void QuaternionToRotation(const T q[4], T R[3 * 3]);
161
162template <typename T, int row_stride, int col_stride> inline
163void QuaternionToRotation(
164    const T q[4],
165    const MatrixAdapter<T, row_stride, col_stride>& R);
166
167// Rotates a point pt by a quaternion q:
168//
169//   result = R(q) * pt
170//
171// Assumes the quaternion is unit norm. This assumption allows us to
172// write the transform as (something)*pt + pt, as is clear from the
173// formula below. If you pass in a quaternion with |q|^2 = 2 then you
174// WILL NOT get back 2 times the result you get for a unit quaternion.
175template <typename T> inline
176void UnitQuaternionRotatePoint(const T q[4], const T pt[3], T result[3]);
177
178// With this function you do not need to assume that q has unit norm.
179// It does assume that the norm is non-zero.
180template <typename T> inline
181void QuaternionRotatePoint(const T q[4], const T pt[3], T result[3]);
182
183// zw = z * w, where * is the Quaternion product between 4 vectors.
184template<typename T> inline
185void QuaternionProduct(const T z[4], const T w[4], T zw[4]);
186
187// xy = x cross y;
188template<typename T> inline
189void CrossProduct(const T x[3], const T y[3], T x_cross_y[3]);
190
191template<typename T> inline
192T DotProduct(const T x[3], const T y[3]);
193
194// y = R(angle_axis) * x;
195template<typename T> inline
196void AngleAxisRotatePoint(const T angle_axis[3], const T pt[3], T result[3]);
197
198// --- IMPLEMENTATION
199
200template<typename T, int row_stride, int col_stride>
201struct MatrixAdapter {
202  T* pointer_;
203  explicit MatrixAdapter(T* pointer)
204    : pointer_(pointer)
205  {}
206
207  T& operator()(int r, int c) const {
208    return pointer_[r * row_stride + c * col_stride];
209  }
210};
211
212template <typename T>
213MatrixAdapter<T, 1, 3> ColumnMajorAdapter3x3(T* pointer) {
214  return MatrixAdapter<T, 1, 3>(pointer);
215}
216
217template <typename T>
218MatrixAdapter<T, 3, 1> RowMajorAdapter3x3(T* pointer) {
219  return MatrixAdapter<T, 3, 1>(pointer);
220}
221
222template<typename T>
223inline void AngleAxisToQuaternion(const T* angle_axis, T* quaternion) {
224  const T& a0 = angle_axis[0];
225  const T& a1 = angle_axis[1];
226  const T& a2 = angle_axis[2];
227  const T theta_squared = a0 * a0 + a1 * a1 + a2 * a2;
228
229  // For points not at the origin, the full conversion is numerically stable.
230  if (theta_squared > T(0.0)) {
231    const T theta = sqrt(theta_squared);
232    const T half_theta = theta * T(0.5);
233    const T k = sin(half_theta) / theta;
234    quaternion[0] = cos(half_theta);
235    quaternion[1] = a0 * k;
236    quaternion[2] = a1 * k;
237    quaternion[3] = a2 * k;
238  } else {
239    // At the origin, sqrt() will produce NaN in the derivative since
240    // the argument is zero.  By approximating with a Taylor series,
241    // and truncating at one term, the value and first derivatives will be
242    // computed correctly when Jets are used.
243    const T k(0.5);
244    quaternion[0] = T(1.0);
245    quaternion[1] = a0 * k;
246    quaternion[2] = a1 * k;
247    quaternion[3] = a2 * k;
248  }
249}
250
251template<typename T>
252inline void QuaternionToAngleAxis(const T* quaternion, T* angle_axis) {
253  const T& q1 = quaternion[1];
254  const T& q2 = quaternion[2];
255  const T& q3 = quaternion[3];
256  const T sin_squared_theta = q1 * q1 + q2 * q2 + q3 * q3;
257
258  // For quaternions representing non-zero rotation, the conversion
259  // is numerically stable.
260  if (sin_squared_theta > T(0.0)) {
261    const T sin_theta = sqrt(sin_squared_theta);
262    const T& cos_theta = quaternion[0];
263
264    // If cos_theta is negative, theta is greater than pi/2, which
265    // means that angle for the angle_axis vector which is 2 * theta
266    // would be greater than pi.
267    //
268    // While this will result in the correct rotation, it does not
269    // result in a normalized angle-axis vector.
270    //
271    // In that case we observe that 2 * theta ~ 2 * theta - 2 * pi,
272    // which is equivalent saying
273    //
274    //   theta - pi = atan(sin(theta - pi), cos(theta - pi))
275    //              = atan(-sin(theta), -cos(theta))
276    //
277    const T two_theta =
278        T(2.0) * ((cos_theta < 0.0)
279                  ? atan2(-sin_theta, -cos_theta)
280                  : atan2(sin_theta, cos_theta));
281    const T k = two_theta / sin_theta;
282    angle_axis[0] = q1 * k;
283    angle_axis[1] = q2 * k;
284    angle_axis[2] = q3 * k;
285  } else {
286    // For zero rotation, sqrt() will produce NaN in the derivative since
287    // the argument is zero.  By approximating with a Taylor series,
288    // and truncating at one term, the value and first derivatives will be
289    // computed correctly when Jets are used.
290    const T k(2.0);
291    angle_axis[0] = q1 * k;
292    angle_axis[1] = q2 * k;
293    angle_axis[2] = q3 * k;
294  }
295}
296
297// The conversion of a rotation matrix to the angle-axis form is
298// numerically problematic when then rotation angle is close to zero
299// or to Pi. The following implementation detects when these two cases
300// occurs and deals with them by taking code paths that are guaranteed
301// to not perform division by a small number.
302template <typename T>
303inline void RotationMatrixToAngleAxis(const T* R, T* angle_axis) {
304  RotationMatrixToAngleAxis(ColumnMajorAdapter3x3(R), angle_axis);
305}
306
307template <typename T, int row_stride, int col_stride>
308void RotationMatrixToAngleAxis(
309    const MatrixAdapter<const T, row_stride, col_stride>& R,
310    T* angle_axis) {
311  // x = k * 2 * sin(theta), where k is the axis of rotation.
312  angle_axis[0] = R(2, 1) - R(1, 2);
313  angle_axis[1] = R(0, 2) - R(2, 0);
314  angle_axis[2] = R(1, 0) - R(0, 1);
315
316  static const T kOne = T(1.0);
317  static const T kTwo = T(2.0);
318
319  // Since the right hand side may give numbers just above 1.0 or
320  // below -1.0 leading to atan misbehaving, we threshold.
321  T costheta = std::min(std::max((R(0, 0) + R(1, 1) + R(2, 2) - kOne) / kTwo,
322                                 T(-1.0)),
323                        kOne);
324
325  // sqrt is guaranteed to give non-negative results, so we only
326  // threshold above.
327  T sintheta = std::min(sqrt(angle_axis[0] * angle_axis[0] +
328                             angle_axis[1] * angle_axis[1] +
329                             angle_axis[2] * angle_axis[2]) / kTwo,
330                        kOne);
331
332  // Use the arctan2 to get the right sign on theta
333  const T theta = atan2(sintheta, costheta);
334
335  // Case 1: sin(theta) is large enough, so dividing by it is not a
336  // problem. We do not use abs here, because while jets.h imports
337  // std::abs into the namespace, here in this file, abs resolves to
338  // the int version of the function, which returns zero always.
339  //
340  // We use a threshold much larger then the machine epsilon, because
341  // if sin(theta) is small, not only do we risk overflow but even if
342  // that does not occur, just dividing by a small number will result
343  // in numerical garbage. So we play it safe.
344  static const double kThreshold = 1e-12;
345  if ((sintheta > kThreshold) || (sintheta < -kThreshold)) {
346    const T r = theta / (kTwo * sintheta);
347    for (int i = 0; i < 3; ++i) {
348      angle_axis[i] *= r;
349    }
350    return;
351  }
352
353  // Case 2: theta ~ 0, means sin(theta) ~ theta to a good
354  // approximation.
355  if (costheta > 0.0) {
356    const T kHalf = T(0.5);
357    for (int i = 0; i < 3; ++i) {
358      angle_axis[i] *= kHalf;
359    }
360    return;
361  }
362
363  // Case 3: theta ~ pi, this is the hard case. Since theta is large,
364  // and sin(theta) is small. Dividing by theta by sin(theta) will
365  // either give an overflow or worse still numerically meaningless
366  // results. Thus we use an alternate more complicated formula
367  // here.
368
369  // Since cos(theta) is negative, division by (1-cos(theta)) cannot
370  // overflow.
371  const T inv_one_minus_costheta = kOne / (kOne - costheta);
372
373  // We now compute the absolute value of coordinates of the axis
374  // vector using the diagonal entries of R. To resolve the sign of
375  // these entries, we compare the sign of angle_axis[i]*sin(theta)
376  // with the sign of sin(theta). If they are the same, then
377  // angle_axis[i] should be positive, otherwise negative.
378  for (int i = 0; i < 3; ++i) {
379    angle_axis[i] = theta * sqrt((R(i, i) - costheta) * inv_one_minus_costheta);
380    if (((sintheta < 0.0) && (angle_axis[i] > 0.0)) ||
381        ((sintheta > 0.0) && (angle_axis[i] < 0.0))) {
382      angle_axis[i] = -angle_axis[i];
383    }
384  }
385}
386
387template <typename T>
388inline void AngleAxisToRotationMatrix(const T* angle_axis, T* R) {
389  AngleAxisToRotationMatrix(angle_axis, ColumnMajorAdapter3x3(R));
390}
391
392template <typename T, int row_stride, int col_stride>
393void AngleAxisToRotationMatrix(
394    const T* angle_axis,
395    const MatrixAdapter<T, row_stride, col_stride>& R) {
396  static const T kOne = T(1.0);
397  const T theta2 = DotProduct(angle_axis, angle_axis);
398  if (theta2 > T(std::numeric_limits<double>::epsilon())) {
399    // We want to be careful to only evaluate the square root if the
400    // norm of the angle_axis vector is greater than zero. Otherwise
401    // we get a division by zero.
402    const T theta = sqrt(theta2);
403    const T wx = angle_axis[0] / theta;
404    const T wy = angle_axis[1] / theta;
405    const T wz = angle_axis[2] / theta;
406
407    const T costheta = cos(theta);
408    const T sintheta = sin(theta);
409
410    R(0, 0) =     costheta   + wx*wx*(kOne -    costheta);
411    R(1, 0) =  wz*sintheta   + wx*wy*(kOne -    costheta);
412    R(2, 0) = -wy*sintheta   + wx*wz*(kOne -    costheta);
413    R(0, 1) =  wx*wy*(kOne - costheta)     - wz*sintheta;
414    R(1, 1) =     costheta   + wy*wy*(kOne -    costheta);
415    R(2, 1) =  wx*sintheta   + wy*wz*(kOne -    costheta);
416    R(0, 2) =  wy*sintheta   + wx*wz*(kOne -    costheta);
417    R(1, 2) = -wx*sintheta   + wy*wz*(kOne -    costheta);
418    R(2, 2) =     costheta   + wz*wz*(kOne -    costheta);
419  } else {
420    // Near zero, we switch to using the first order Taylor expansion.
421    R(0, 0) =  kOne;
422    R(1, 0) =  angle_axis[2];
423    R(2, 0) = -angle_axis[1];
424    R(0, 1) = -angle_axis[2];
425    R(1, 1) =  kOne;
426    R(2, 1) =  angle_axis[0];
427    R(0, 2) =  angle_axis[1];
428    R(1, 2) = -angle_axis[0];
429    R(2, 2) = kOne;
430  }
431}
432
433template <typename T>
434inline void EulerAnglesToRotationMatrix(const T* euler,
435                                        const int row_stride_parameter,
436                                        T* R) {
437  CHECK_EQ(row_stride_parameter, 3);
438  EulerAnglesToRotationMatrix(euler, RowMajorAdapter3x3(R));
439}
440
441template <typename T, int row_stride, int col_stride>
442void EulerAnglesToRotationMatrix(
443    const T* euler,
444    const MatrixAdapter<T, row_stride, col_stride>& R) {
445  const double kPi = 3.14159265358979323846;
446  const T degrees_to_radians(kPi / 180.0);
447
448  const T pitch(euler[0] * degrees_to_radians);
449  const T roll(euler[1] * degrees_to_radians);
450  const T yaw(euler[2] * degrees_to_radians);
451
452  const T c1 = cos(yaw);
453  const T s1 = sin(yaw);
454  const T c2 = cos(roll);
455  const T s2 = sin(roll);
456  const T c3 = cos(pitch);
457  const T s3 = sin(pitch);
458
459  R(0, 0) = c1*c2;
460  R(0, 1) = -s1*c3 + c1*s2*s3;
461  R(0, 2) = s1*s3 + c1*s2*c3;
462
463  R(1, 0) = s1*c2;
464  R(1, 1) = c1*c3 + s1*s2*s3;
465  R(1, 2) = -c1*s3 + s1*s2*c3;
466
467  R(2, 0) = -s2;
468  R(2, 1) = c2*s3;
469  R(2, 2) = c2*c3;
470}
471
472template <typename T> inline
473void QuaternionToScaledRotation(const T q[4], T R[3 * 3]) {
474  QuaternionToScaledRotation(q, RowMajorAdapter3x3(R));
475}
476
477template <typename T, int row_stride, int col_stride> inline
478void QuaternionToScaledRotation(
479    const T q[4],
480    const MatrixAdapter<T, row_stride, col_stride>& R) {
481  // Make convenient names for elements of q.
482  T a = q[0];
483  T b = q[1];
484  T c = q[2];
485  T d = q[3];
486  // This is not to eliminate common sub-expression, but to
487  // make the lines shorter so that they fit in 80 columns!
488  T aa = a * a;
489  T ab = a * b;
490  T ac = a * c;
491  T ad = a * d;
492  T bb = b * b;
493  T bc = b * c;
494  T bd = b * d;
495  T cc = c * c;
496  T cd = c * d;
497  T dd = d * d;
498
499  R(0, 0) = aa + bb - cc - dd; R(0, 1) = T(2) * (bc - ad);  R(0, 2) = T(2) * (ac + bd);  // NOLINT
500  R(1, 0) = T(2) * (ad + bc);  R(1, 1) = aa - bb + cc - dd; R(1, 2) = T(2) * (cd - ab);  // NOLINT
501  R(2, 0) = T(2) * (bd - ac);  R(2, 1) = T(2) * (ab + cd);  R(2, 2) = aa - bb - cc + dd; // NOLINT
502}
503
504template <typename T> inline
505void QuaternionToRotation(const T q[4], T R[3 * 3]) {
506  QuaternionToRotation(q, RowMajorAdapter3x3(R));
507}
508
509template <typename T, int row_stride, int col_stride> inline
510void QuaternionToRotation(const T q[4],
511                          const MatrixAdapter<T, row_stride, col_stride>& R) {
512  QuaternionToScaledRotation(q, R);
513
514  T normalizer = q[0]*q[0] + q[1]*q[1] + q[2]*q[2] + q[3]*q[3];
515  CHECK_NE(normalizer, T(0));
516  normalizer = T(1) / normalizer;
517
518  for (int i = 0; i < 3; ++i) {
519    for (int j = 0; j < 3; ++j) {
520      R(i, j) *= normalizer;
521    }
522  }
523}
524
525template <typename T> inline
526void UnitQuaternionRotatePoint(const T q[4], const T pt[3], T result[3]) {
527  const T t2 =  q[0] * q[1];
528  const T t3 =  q[0] * q[2];
529  const T t4 =  q[0] * q[3];
530  const T t5 = -q[1] * q[1];
531  const T t6 =  q[1] * q[2];
532  const T t7 =  q[1] * q[3];
533  const T t8 = -q[2] * q[2];
534  const T t9 =  q[2] * q[3];
535  const T t1 = -q[3] * q[3];
536  result[0] = T(2) * ((t8 + t1) * pt[0] + (t6 - t4) * pt[1] + (t3 + t7) * pt[2]) + pt[0];  // NOLINT
537  result[1] = T(2) * ((t4 + t6) * pt[0] + (t5 + t1) * pt[1] + (t9 - t2) * pt[2]) + pt[1];  // NOLINT
538  result[2] = T(2) * ((t7 - t3) * pt[0] + (t2 + t9) * pt[1] + (t5 + t8) * pt[2]) + pt[2];  // NOLINT
539}
540
541template <typename T> inline
542void QuaternionRotatePoint(const T q[4], const T pt[3], T result[3]) {
543  // 'scale' is 1 / norm(q).
544  const T scale = T(1) / sqrt(q[0] * q[0] +
545                              q[1] * q[1] +
546                              q[2] * q[2] +
547                              q[3] * q[3]);
548
549  // Make unit-norm version of q.
550  const T unit[4] = {
551    scale * q[0],
552    scale * q[1],
553    scale * q[2],
554    scale * q[3],
555  };
556
557  UnitQuaternionRotatePoint(unit, pt, result);
558}
559
560template<typename T> inline
561void QuaternionProduct(const T z[4], const T w[4], T zw[4]) {
562  zw[0] = z[0] * w[0] - z[1] * w[1] - z[2] * w[2] - z[3] * w[3];
563  zw[1] = z[0] * w[1] + z[1] * w[0] + z[2] * w[3] - z[3] * w[2];
564  zw[2] = z[0] * w[2] - z[1] * w[3] + z[2] * w[0] + z[3] * w[1];
565  zw[3] = z[0] * w[3] + z[1] * w[2] - z[2] * w[1] + z[3] * w[0];
566}
567
568// xy = x cross y;
569template<typename T> inline
570void CrossProduct(const T x[3], const T y[3], T x_cross_y[3]) {
571  x_cross_y[0] = x[1] * y[2] - x[2] * y[1];
572  x_cross_y[1] = x[2] * y[0] - x[0] * y[2];
573  x_cross_y[2] = x[0] * y[1] - x[1] * y[0];
574}
575
576template<typename T> inline
577T DotProduct(const T x[3], const T y[3]) {
578  return (x[0] * y[0] + x[1] * y[1] + x[2] * y[2]);
579}
580
581template<typename T> inline
582void AngleAxisRotatePoint(const T angle_axis[3], const T pt[3], T result[3]) {
583  const T theta2 = DotProduct(angle_axis, angle_axis);
584  if (theta2 > T(std::numeric_limits<double>::epsilon())) {
585    // Away from zero, use the rodriguez formula
586    //
587    //   result = pt costheta +
588    //            (w x pt) * sintheta +
589    //            w (w . pt) (1 - costheta)
590    //
591    // We want to be careful to only evaluate the square root if the
592    // norm of the angle_axis vector is greater than zero. Otherwise
593    // we get a division by zero.
594    //
595    const T theta = sqrt(theta2);
596    const T costheta = cos(theta);
597    const T sintheta = sin(theta);
598    const T theta_inverse = 1.0 / theta;
599
600    const T w[3] = { angle_axis[0] * theta_inverse,
601                     angle_axis[1] * theta_inverse,
602                     angle_axis[2] * theta_inverse };
603
604    // Explicitly inlined evaluation of the cross product for
605    // performance reasons.
606    const T w_cross_pt[3] = { w[1] * pt[2] - w[2] * pt[1],
607                              w[2] * pt[0] - w[0] * pt[2],
608                              w[0] * pt[1] - w[1] * pt[0] };
609    const T tmp =
610        (w[0] * pt[0] + w[1] * pt[1] + w[2] * pt[2]) * (T(1.0) - costheta);
611
612    result[0] = pt[0] * costheta + w_cross_pt[0] * sintheta + w[0] * tmp;
613    result[1] = pt[1] * costheta + w_cross_pt[1] * sintheta + w[1] * tmp;
614    result[2] = pt[2] * costheta + w_cross_pt[2] * sintheta + w[2] * tmp;
615  } else {
616    // Near zero, the first order Taylor approximation of the rotation
617    // matrix R corresponding to a vector w and angle w is
618    //
619    //   R = I + hat(w) * sin(theta)
620    //
621    // But sintheta ~ theta and theta * w = angle_axis, which gives us
622    //
623    //  R = I + hat(w)
624    //
625    // and actually performing multiplication with the point pt, gives us
626    // R * pt = pt + w x pt.
627    //
628    // Switching to the Taylor expansion near zero provides meaningful
629    // derivatives when evaluated using Jets.
630    //
631    // Explicitly inlined evaluation of the cross product for
632    // performance reasons.
633    const T w_cross_pt[3] = { angle_axis[1] * pt[2] - angle_axis[2] * pt[1],
634                              angle_axis[2] * pt[0] - angle_axis[0] * pt[2],
635                              angle_axis[0] * pt[1] - angle_axis[1] * pt[0] };
636
637    result[0] = pt[0] + w_cross_pt[0];
638    result[1] = pt[1] + w_cross_pt[1];
639    result[2] = pt[2] + w_cross_pt[2];
640  }
641}
642
643}  // namespace ceres
644
645#endif  // CERES_PUBLIC_ROTATION_H_
646