Source code for pytransform3d.batch_rotations

"""Batch operations on rotations in three dimensions - SO(3).

Conversions from this module operate on batches of orientations or rotations
and can be orders of magnitude faster than a loop of individual conversions.

All functions operate on nd arrays, where the last dimension (vectors) or
the last two dimensions (matrices) contain individual rotations.
"""
import numpy as np
from .rotations import angle_between_vectors, slerp_weights


[docs]def norm_vectors(V, out=None): """Normalize vectors. Parameters ---------- V : array-like, shape (..., n) nd vectors out : array, shape (..., n), optional (default: new array) Output array to which we write the result Returns ------- V_unit : array, shape (..., n) nd unit vectors with norm 1 or zero vectors """ V = np.asarray(V) norms = np.linalg.norm(V, axis=-1) if out is None: out = np.empty_like(V) # Avoid division by zero with np.maximum(..., smallest positive float). # The norm is zero only when the vector is zero so this case does not # require further processing. out[...] = V / np.maximum(norms[..., np.newaxis], np.finfo(float).tiny) return out
[docs]def angles_between_vectors(A, B): """Compute angle between two vectors. Parameters ---------- A : array-like, shape (..., n) nd vectors B : array-like, shape (..., n) nd vectors Returns ------- angles : array, shape (...) Angles between pairs of vectors from A and B """ A = np.asarray(A) B = np.asarray(B) n_dims = A.shape[-1] A_norms = np.linalg.norm(A, axis=-1) B_norms = np.linalg.norm(B, axis=-1) AdotB = np.einsum( "ni,ni->n", A.reshape(-1, n_dims), B.reshape(-1, n_dims) ).reshape(A.shape[:-1]) return np.arccos(np.clip(AdotB / (A_norms * B_norms), -1.0, 1.0))
[docs]def active_matrices_from_angles(basis, angles, out=None): """Compute active rotation matrices from rotation about basis vectors. Parameters ---------- basis : int from [0, 1, 2] The rotation axis (0: x, 1: y, 2: z) angles : array-like, shape (...) Rotation angles out : array, shape (..., 3, 3), optional (default: new array) Output array to which we write the result Returns ------- Rs : array, shape (..., 3, 3) Rotation matrices """ angles = np.asarray(angles) c = np.cos(angles) s = np.sin(angles) R_shape = angles.shape + (3, 3) if out is None: out = np.empty(R_shape) out[..., basis, :] = 0.0 out[..., :, basis] = 0.0 out[..., basis, basis] = 1.0 basisp1 = (basis + 1) % 3 basisp2 = (basis + 2) % 3 out[..., basisp1, basisp1] = c out[..., basisp2, basisp2] = c out[..., basisp1, basisp2] = -s out[..., basisp2, basisp1] = s return out
[docs]def active_matrices_from_intrinsic_euler_angles( basis1, basis2, basis3, e, out=None): """Compute active rotation matrices from intrinsic Euler angles. Parameters ---------- basis1 : int Basis vector of first rotation. 0 corresponds to x axis, 1 to y axis, and 2 to z axis. basis2 : int Basis vector of second rotation. 0 corresponds to x axis, 1 to y axis, and 2 to z axis. basis3 : int Basis vector of third rotation. 0 corresponds to x axis, 1 to y axis, and 2 to z axis. e : array-like, shape (..., 3) Euler angles out : array, shape (..., 3, 3), optional (default: new array) Output array to which we write the result Returns ------- Rs : array, shape (..., 3, 3) Rotation matrices """ e = np.asarray(e) R_shape = e.shape + (3,) R_alpha = active_matrices_from_angles(basis1, e[..., 0].flat) R_beta = active_matrices_from_angles(basis2, e[..., 1].flat) R_gamma = active_matrices_from_angles(basis3, e[..., 2].flat) if out is None: out = np.empty(R_shape) out[:] = np.einsum( "nij,njk->nik", np.einsum("nij,njk->nik", R_alpha, R_beta), R_gamma).reshape(R_shape) return out
[docs]def active_matrices_from_extrinsic_euler_angles( basis1, basis2, basis3, e, out=None): """Compute active rotation matrices from extrinsic Euler angles. Parameters ---------- basis1 : int Basis vector of first rotation. 0 corresponds to x axis, 1 to y axis, and 2 to z axis. basis2 : int Basis vector of second rotation. 0 corresponds to x axis, 1 to y axis, and 2 to z axis. basis3 : int Basis vector of third rotation. 0 corresponds to x axis, 1 to y axis, and 2 to z axis. e : array-like, shape (..., 3) Euler angles out : array, shape (..., 3, 3), optional (default: new array) Output array to which we write the result Returns ------- Rs : array, shape (..., 3, 3) Rotation matrices """ e = np.asarray(e) R_shape = e.shape + (3,) R_alpha = active_matrices_from_angles(basis1, e[..., 0].flat) R_beta = active_matrices_from_angles(basis2, e[..., 1].flat) R_gamma = active_matrices_from_angles(basis3, e[..., 2].flat) if out is None: out = np.empty(R_shape) out[:] = np.einsum( "nij,njk->nik", np.einsum("nij,njk->nik", R_gamma, R_beta), R_alpha).reshape(R_shape) return out
[docs]def matrices_from_compact_axis_angles( A=None, axes=None, angles=None, out=None): """Compute rotation matrices from compact axis-angle representations. This is called exponential map or Rodrigues' formula. This typically results in an active rotation matrix. Parameters ---------- A : array-like, shape (..., 3) Axes of rotation and rotation angles in compact representation: angle * (x, y, z) axes : array, shape (..., 3) If the unit axes of rotation have been precomputed, you can pass them here. angles : array, shape (...) If the angles have been precomputed, you can pass them here. out : array, shape (..., 3, 3), optional (default: new array) Output array to which we write the result Returns ------- Rs : array, shape (..., 3, 3) Rotation matrices """ if angles is None: thetas = np.linalg.norm(A, axis=-1) else: thetas = np.asarray(angles) if axes is None: omega_unit = norm_vectors(A) else: omega_unit = axes c = np.cos(thetas) s = np.sin(thetas) ci = 1.0 - c ux = omega_unit[..., 0] uy = omega_unit[..., 1] uz = omega_unit[..., 2] uxs = ux * s uys = uy * s uzs = uz * s ciux = ci * ux ciuy = ci * uy ciuxuy = ciux * uy ciuxuz = ciux * uz ciuyuz = ciuy * uz if out is None: out = np.empty(A.shape[:-1] + (3, 3)) out[..., 0, 0] = ciux * ux + c out[..., 0, 1] = ciuxuy - uzs out[..., 0, 2] = ciuxuz + uys out[..., 1, 0] = ciuxuy + uzs out[..., 1, 1] = ciuy * uy + c out[..., 1, 2] = ciuyuz - uxs out[..., 2, 0] = ciuxuz - uys out[..., 2, 1] = ciuyuz + uxs out[..., 2, 2] = ci * uz * uz + c return out
[docs]def axis_angles_from_matrices(Rs, traces=None, out=None): """Compute compact axis-angle representations from rotation matrices. This is called logarithmic map. Parameters ---------- Rs : array-like, shape (..., 3, 3) Rotation matrices traces : array, shape (..., 3) If the traces of rotation matrices been precomputed, you can pass them here. out : array, shape (..., 4), optional (default: new array) Output array to which we write the result Returns ------- A : array, shape (..., 4) Axes of rotation and rotation angles: (x, y, z, angle) """ Rs = np.asarray(Rs) instances_shape = Rs.shape[:-2] if traces is None: traces = np.einsum("nii", Rs.reshape(-1, 3, 3)) if instances_shape: traces = traces.reshape(*instances_shape) else: # this works because indX will be a single boolean and # out[True, n] = value will assign value to out[n], while # out[False, n] = value will not assign value to out[n] traces = traces[0] angles = np.arccos((traces - 1.0) / 2.0) if out is None: out = np.empty(instances_shape + (4,)) out[..., 0] = Rs[..., 2, 1] - Rs[..., 1, 2] out[..., 1] = Rs[..., 0, 2] - Rs[..., 2, 0] out[..., 2] = Rs[..., 1, 0] - Rs[..., 0, 1] # The threshold is a result from this discussion: # https://github.com/rock-learning/pytransform3d/issues/43 # The standard formula becomes numerically unstable, however, # Rodrigues' formula reduces to R = I + 2 (ee^T - I), with the # rotation axis e, that is, ee^T = 0.5 * (R + I) and we can find the # squared values of the rotation axis on the diagonal of this matrix. # We can still use the original formula to reconstruct the signs of # the rotation axis correctly. angle_close_to_pi = np.abs(angles - np.pi) < 1e-4 angle_not_zero = np.abs(angles) != 0.0 Rs_diag = np.einsum("nii->ni", Rs.reshape(-1, 3, 3)) if instances_shape: Rs_diag = Rs_diag.reshape(*(instances_shape + (3,))) else: Rs_diag = Rs_diag[0] out[angle_close_to_pi, :3] = ( np.sqrt(0.5 * (Rs_diag[angle_close_to_pi] + 1.0)) * np.sign(out[angle_close_to_pi, :3])) out[angle_not_zero, :3] /= np.linalg.norm( out[angle_not_zero, :3], axis=-1)[..., np.newaxis] out[..., 3] = angles return out
[docs]def cross_product_matrices(V): """Generate the cross-product matrices of vectors. The cross-product matrix :math:`\\boldsymbol{V}` satisfies the equation .. math:: \\boldsymbol{V} \\boldsymbol{w} = \\boldsymbol{v} \\times \\boldsymbol{w} It is a skew-symmetric (antisymmetric) matrix, i.e. :math:`-\\boldsymbol{V} = \\boldsymbol{V}^T`. Parameters ---------- V : array-like, shape (..., 3) 3d vectors Returns ------- V_cross_product_matrices : array, shape (..., 3, 3) Cross-product matrices of V """ V = np.asarray(V) instances_shape = V.shape[:-1] V_matrices = np.empty(instances_shape + (3, 3)) V_matrices[..., 0, 0] = 0.0 V_matrices[..., 0, 1] = -V[..., 2] V_matrices[..., 0, 2] = V[..., 1] V_matrices[..., 1, 0] = V[..., 2] V_matrices[..., 1, 1] = 0.0 V_matrices[..., 1, 2] = -V[..., 0] V_matrices[..., 2, 0] = -V[..., 1] V_matrices[..., 2, 1] = V[..., 0] V_matrices[..., 2, 2] = 0.0 return V_matrices
[docs]def matrices_from_quaternions(Q, normalize_quaternions=True, out=None): """Compute rotation matrices from quaternions. Parameters ---------- Q : array-like, shape (..., 4) Unit quaternions to represent rotations: (w, x, y, z) normalize_quaternions : bool, optional (default: True) Normalize quaternions before conversion out : array, shape (..., 3, 3), optional (default: new array) Output array to which we write the result Returns ------- Rs : array, shape (..., 3, 3) Rotation matrices """ Q = np.asarray(Q) if normalize_quaternions: Q = norm_vectors(Q) w = Q[..., 0] x = Q[..., 1] y = Q[..., 2] z = Q[..., 3] x2 = 2.0 * x * x y2 = 2.0 * y * y z2 = 2.0 * z * z xy = 2.0 * x * y xz = 2.0 * x * z yz = 2.0 * y * z xw = 2.0 * x * w yw = 2.0 * y * w zw = 2.0 * z * w if out is None: out = np.empty(w.shape + (3, 3)) out[..., 0, 0] = 1.0 - y2 - z2 out[..., 0, 1] = xy - zw out[..., 0, 2] = xz + yw out[..., 1, 0] = xy + zw out[..., 1, 1] = 1.0 - x2 - z2 out[..., 1, 2] = yz - xw out[..., 2, 0] = xz - yw out[..., 2, 1] = yz + xw out[..., 2, 2] = 1.0 - x2 - y2 return out
[docs]def quaternions_from_matrices(Rs, out=None): """Compute quaternions from rotation matrices. Parameters ---------- Rs : array-like, shape (..., 3, 3) Rotation matrices out : array, shape (..., 4), optional (default: new array) Output array to which we write the result Returns ------- Q : array, shape (..., 4) Unit quaternions to represent rotations: (w, x, y, z) """ Rs = np.asarray(Rs) instances_shape = Rs.shape[:-2] if out is None: out = np.empty(instances_shape + (4,)) traces = np.einsum("nii", Rs.reshape(-1, 3, 3)) if instances_shape: traces = traces.reshape(*instances_shape) else: # this works because indX will be a single boolean and # out[True, n] = value will assign value to out[n], while # out[False, n] = value will not assign value to out[n] traces = traces[0] ind1 = traces > 0.0 s = 2.0 * np.sqrt(1.0 + traces[ind1]) out[ind1, 0] = 0.25 * s out[ind1, 1] = (Rs[ind1, 2, 1] - Rs[ind1, 1, 2]) / s out[ind1, 2] = (Rs[ind1, 0, 2] - Rs[ind1, 2, 0]) / s out[ind1, 3] = (Rs[ind1, 1, 0] - Rs[ind1, 0, 1]) / s ind2 = np.logical_and( np.logical_not(ind1), np.logical_and(Rs[..., 0, 0] > Rs[..., 1, 1], Rs[..., 0, 0] > Rs[..., 2, 2])) s = 2.0 * np.sqrt(1.0 + Rs[ind2, 0, 0] - Rs[ind2, 1, 1] - Rs[ind2, 2, 2]) out[ind2, 0] = (Rs[ind2, 2, 1] - Rs[ind2, 1, 2]) / s out[ind2, 1] = 0.25 * s out[ind2, 2] = (Rs[ind2, 1, 0] + Rs[ind2, 0, 1]) / s out[ind2, 3] = (Rs[ind2, 0, 2] + Rs[ind2, 2, 0]) / s ind3 = np.logical_and( np.logical_not(ind1), Rs[..., 1, 1] > Rs[..., 2, 2]) s = 2.0 * np.sqrt(1.0 + Rs[ind3, 1, 1] - Rs[ind3, 0, 0] - Rs[ind3, 2, 2]) out[ind3, 0] = (Rs[ind3, 0, 2] - Rs[ind3, 2, 0]) / s out[ind3, 1] = (Rs[ind3, 1, 0] + Rs[ind3, 0, 1]) / s out[ind3, 2] = 0.25 * s out[ind3, 3] = (Rs[ind3, 2, 1] + Rs[ind3, 1, 2]) / s ind4 = np.logical_and( np.logical_and(np.logical_not(ind1), np.logical_not(ind2)), np.logical_not(ind3)) s = 2.0 * np.sqrt(1.0 + Rs[ind4, 2, 2] - Rs[ind4, 0, 0] - Rs[ind4, 1, 1]) out[ind4, 0] = (Rs[ind4, 1, 0] - Rs[ind4, 0, 1]) / s out[ind4, 1] = (Rs[ind4, 0, 2] + Rs[ind4, 2, 0]) / s out[ind4, 2] = (Rs[ind4, 2, 1] + Rs[ind4, 1, 2]) / s out[ind4, 3] = 0.25 * s return out
[docs]def quaternion_slerp_batch(start, end, t): """Spherical linear interpolation for a batch of steps. Parameters ---------- start : array-like, shape (4,) Start unit quaternion to represent rotation: (w, x, y, z) end : array-like, shape (4,) End unit quaternion to represent rotation: (w, x, y, z) t : array-like, shape (n_steps,) Steps between start and goal, must be in interval [0, 1] Returns ------- Q : array, shape (n_steps, 4) Interpolated unit quaternions """ t = np.asarray(t) angle = angle_between_vectors(start, end) w1, w2 = slerp_weights(angle, t) return (w1[:, np.newaxis] * start[np.newaxis] + w2[:, np.newaxis] * end[np.newaxis])
[docs]def batch_concatenate_quaternions(Q1, Q2, out=None): """Concatenate two batches of quaternions. We use Hamilton's quaternion multiplication. Suppose we want to apply two extrinsic rotations given by quaternions q1 and q2 to a vector v. We can either apply q2 to v and then q1 to the result or we can concatenate q1 and q2 and apply the result to v. Parameters ---------- Q1 : array-like, shape (..., 4) First batch of quaternions Q2 : array-like, shape (..., 4) Second batch of quaternions out : array, shape (..., 4), optional (default: new array) Output array to which we write the result Returns ------- Q12 : array, shape (..., 4) Batch of quaternions that represents the concatenated rotations Raises ------ ValueError If the input dimensions are incorrect """ Q1 = np.asarray(Q1) Q2 = np.asarray(Q2) if Q1.ndim != Q2.ndim: raise ValueError("Number of dimensions must be the same. " "Got %d for Q1 and %d for Q2." % (Q1.ndim, Q2.ndim)) for d in range(Q1.ndim - 1): if Q1.shape[d] != Q2.shape[d]: raise ValueError( "Size of dimension %d does not match: %d != %d" % (d + 1, Q1.shape[d], Q2.shape[d])) if Q1.shape[-1] != 4: raise ValueError( "Last dimension of first argument does not match. A quaternion " "must have 4 entries, got %d" % Q1.shape[-1]) if Q2.shape[-1] != 4: raise ValueError( "Last dimension of second argument does not match. A quaternion " "must have 4 entries, got %d" % Q2.shape[-1]) if out is None: out = np.empty_like(Q1) vector_inner_products = np.sum(Q1[..., 1:] * Q2[..., 1:], axis=-1) out[..., 0] = Q1[..., 0] * Q2[..., 0] - vector_inner_products out[..., 1:] = (Q1[..., 0, np.newaxis] * Q2[..., 1:] + Q2[..., 0, np.newaxis] * Q1[..., 1:] + np.cross(Q1[..., 1:], Q2[..., 1:])) return out
[docs]def batch_q_conj(Q): """Conjugate of quaternions. The conjugate of a unit quaternion inverts the rotation represented by this unit quaternion. The conjugate of a quaternion q is often denoted as q*. Parameters ---------- Q : array-like, shape (..., 4) Unit quaternions to represent rotations: (w, x, y, z) Returns ------- Q_c : array, shape (..., 4,) Conjugates (w, -x, -y, -z) """ Q = np.asarray(Q) out = np.empty_like(Q) out[..., 0] = Q[..., 0] out[..., 1:] = -Q[..., 1:] return out
[docs]def batch_quaternion_wxyz_from_xyzw(Q_xyzw, out=None): """Converts from x, y, z, w to w, x, y, z convention. Parameters ---------- Q_xyzw : array-like, shape (..., 4) Quaternions with scalar part after vector part out : array, shape (..., 4), optional (default: new array) Output array to which we write the result Returns ------- Q_wxyz : array-like, shape (..., 4) Quaternions with scalar part before vector part """ Q_xyzw = np.asarray(Q_xyzw) if out is None: out = np.empty_like(Q_xyzw) out[..., 0] = Q_xyzw[..., 3] out[..., 1] = Q_xyzw[..., 0] out[..., 2] = Q_xyzw[..., 1] out[..., 3] = Q_xyzw[..., 2] return out
[docs]def batch_quaternion_xyzw_from_wxyz(Q_wxyz, out=None): """Converts from w, x, y, z to x, y, z, w convention. Parameters ---------- Q_wxyz : array-like, shape (..., 4) Quaternions with scalar part before vector part out : array, shape (..., 4), optional (default: new array) Output array to which we write the result Returns ------- Q_xyzw : array-like, shape (..., 4) Quaternions with scalar part after vector part """ Q_wxyz = np.asarray(Q_wxyz) if out is None: out = np.empty_like(Q_wxyz) out[..., 0] = Q_wxyz[..., 1] out[..., 1] = Q_wxyz[..., 2] out[..., 2] = Q_wxyz[..., 3] out[..., 3] = Q_wxyz[..., 0] return out