Conference PaperPDF Available

Orientation averaging using spatial extension


Abstract and Figures

Orientation and rotation averaging are considered as minimisers of a novel and generalisable error metric based on the displacement of points within the object or space being rotated. The resulting metrics and O(N) averaging algorithm are formally equivalent to approaches currently recommended in the literature. The new approach justifies the use of chordal L2 mean in robotics applications and promises useful generalisations.
Content may be subject to copyright.
Orientation averaging using spatial extension
Damien Jade Duff
Department of Computer Engineering
Istanbul Technical University
Istanbul, Turkey
Abstract—Orientation and rotation averaging are considered
as minimisers of a novel and generalisable error metric based
on the displacement of points within the object or space being
rotated. The resulting metrics and O(N)averaging algorithm are
formally equivalent to approaches currently recommended in the
literature. The new approach justifies the use of chordal L2mean
in robotics applications and promises useful generalisations.
Keywordsorientation, rotation, orientation averaging
We often come across cases where it would be nice to be
able to average orientations or rotations: designing a controller
that smoothly tracks multiple target orientations, smoothing the
results of a particle filter to reduce jitter [1], combining the
output of multiple forward models [2], or creating exemplars
from low-level states for reasoning [3].
Here, I give an approach to averaging orientations and
rotations that should be intuitive to roboticists and is well-
justified by a concept of what it might mean to average
orientations. I consider orientations primarily in their form
as transforms or motion of the constitutive points of some
generic spatially extended object or volume. To average a set
of component orientations1I ask for the orientation that gets
the object’s constitutive points closest to the locations they
assume when the object is oriented according to the component
orientations - see Fig. 1 for an intuition about what is meant
by closeness.
For the case of symmetric objects/spaces considered here,
this intuition coincides with recent approaches in the literature
based apon abstract geometry, and unlike those approaches can
be easily extended to deal with non-symmetric shapes.
A. Related Work
By analogy to the process of averaging Euclidean vectors,
we quickly get efficient algorithms [4] but when not guided
by obvious criteria it is easy to get lost in details (e.g. [5]). On
the other hand, starting with some principled measure of what
constitutes a good average, two good definitions of rotation
averages have been proposed [6,7,8,9]. They correspond to
finding an orientation that minimises either the sum of squared
matrix-distances to component orientations, called the chordal
L2-mean [8] or Euclidean mean [7,9], or finding an orientation
that minimises the sum of squared geodesic distances through
the space of rotations, called the geodesic L2-mean [8] or the
Riemannian mean [7,9].
The approach followed in the current paper is also a
measure-based approach but the measure is the sum of the
displacement of points across an object. It grows out of the use
of error measures evaluating the success of forward models and
for optimizing trajectories against pose observations [2,10].
1Here, component orientations means the orientations to be averaged.
a. Chordal Distance: b. Geodesic Distance:
2+··· dg=d2
Fig. 1. Two orientations of an object and the spatially extended (left) chordal
and (right) geodesic orientation metrics between these two orientations. We can
use these metrics to find an orientation that is close to component orientations.
An alternative approach is to combine both process-analogy
and measure-derived approaches by examining efficient prac-
tical approaches with respect to how close they get to more
well-motivated alternatives, so as to bound the performance of
algorithms. Examples include arithmetic renormalised averag-
ing of rotation matrices and quaternions [4,8]. However, when
an efficient and well-bounded algorithm exists for rotation
averaging we should not need this approach.
When considered as a transform of points of an object (that
is, a displacement of all its constituent points), a rotation is
a rigid-body (distance and orientation preserving) transform
with a fixed point. It is also a linear transform and can be
represented as a 3×3orthogonal matrix with determinant +1.
An orientation is a particular kind of configuration of
an abstract rigid body with a fixed point and as long as a
reference frame for the object is assumed, an orientation can
be expressed as a rotation, so for the purposes of this paper,
rotations and orientations can be treated the same.
A rotation can also be thought of, not just as a transform
(displacement of constituent points), but also as a motion in
which there is a fixed point and distance is preserved over
every part of the motion. Similarly, an orientation in this view
is the motion of an object from its reference frame into the
oriented frame.
This difference in conception of a rotation as a transform
or a motion results in two different approaches to averaging
them, though the conception as a transform is more intuitive
in general given that in the motion-based formulation there is
also the path of motion to consider.
The arithmetic mean ¯p=1
NPpiof a set of Npoints in
3D can be justified as the unique point that minimises the sum
of squared distances between itself and the component points
(differentiate and set to 0to solve):
¯p= argmin
d(p, pi)2= argmin
urkiye Otonom Robotlar Konferansı, 2015
How to define the distance d(R1, R2)between two rotations?
The two best candidates in the literature are the squared
Froebenius norm2of rotation matrix differences R1R2
(the chordal metric) or the minimal squared angle of rotation
needed to align the two rotations R1RT
2(the geodesic metric):
dC(R1, R2) = kR1R2k2(1)
dG(R1, R2) = kΘ(R1RT
Θreturns the angle rotated through by a rotation matrix. In
abstract geometry, the chordal metric is the distance between
rotations in the 9D Euclidean space in which the manifold of
rotations is embedded and the geodesic metric is the shortest
distance along the surface of the rotation manifold [8,9].
In my approach I offer two definitions of distance between
rotations. I call these dcand dgbecause they turn out to be
equivalent, up to a multiplication by a constant, to the chordal
and geodesic metrics in the restricted case considered here, in
which I assume that RppTp= 0 and RpppT=Ifor all points
pin the object or volume3. In both of the below definitions,
our error metric is the sum of the distances by which all points
in a generic or abstract object are either displaced or moved
during a rotation.
A. Spatially extended chordal metric
The spatially extended chordal metric is derived from the
sum of squared displacement distances in 3D Euclidean space
between two orientations R1and R2:
dc(R1, R2)2=Zp
kR1(p)R2(p)k2dxdydz (3)
Here, the integral is over all points pin the object. See Fig.
1(a) for an example of this metric applied to an arbitrarily
shaped object. This corresponds to 2cos(θ)where θis the
angle between the two orientations - the chordal metric is also
known to maximise cos(θ)[8].
This general approach can apply to any transformation
or configuration, not just rotations. It is not necesssary to
consider the paths that points must take during rotation, only
their displacement; this is a metric that considers rotation as
a pure transform and not as a motion. This, along with its
ammenability to analysis, means that it should be the metric
of choice in most robot applications.
B. Spatially extended geodesic metric
The sum of the squared shortest possible distances in 3D
Euclidean space during rotation is the same, up to a constant,
as the geodesic distance:
dg(R1, R2)2=Zp
2(p)2dxdydz (4)
The problems with this approach, apart from the non-
intuitiveness of applying the concept of motion-paths to define
orientation, are that we need to apply each motion constraint
(e.g. obstacles, differential constraints), when defining the
metric rather than just when analysing it in the context of the
relevant transform, and no closed-form is yet known for the
averaging solution [8,9]. Moreover, it generally gives similar
results to the chordal metric when averaging rotations [9].
2The Euclidean norm applied directly to the entries of a matrix.
3Imagine the space being rotated as the limit of an expanding sphere of
unit density.
Fig. 2. Visualisation of a 3D slice of the 4D quaternion sphere. Orientations
are shown as line segments passing through the sphere origin. The contour
corresponds to the sum of squared chordal L2distances to the component
orientations. Left: a random set of quaternions. Right: a set of quaternions
chosen to be balanced across two rotation axes.
C. Averaging with the spatially extended chordal metric
Applying the above principals, integrating over all the
points in the presumed symmetric object/space, we obtain:
dc(R1, R2)2=Zp
= 2 2 cos(Θ(RT
1R2)) (5)
Candidate and components expressed as quaternions q,qi:
¯q= argmin
qSO(3) X
dc(q, qi)2= argmax
Where A=PiqiqT
i(symmetric). See Fig. 2 for a visualisa-
tion of the objective function. To get the mean, we normalise
the first eigenvector of A, a theoretically O(1) operation for
a4×4matrix. Multiple solutions can occur when component
rotations are perfectly balanced but in practice this happens
infinitesimally often, and we can return a random solution from
the Eigenspace.
To average orientations, I use new error metrics invariant to
order change, inversion and rotation, describe their relationship
with other metrics in the literature, and emphasise the intuition
of rotations as either transforms or motion in arguing for using
the chordal mean in robot applications.
[1] T. M¨
orwald, M. Zillich, and M. Vincze, “Edge Tracking of Textured
Objects with a Recursive Particle Filter,” in Intl. Conf. on Computer
Graphics and Vision, Moscow, Russia, Oct. 2009, pp. 96–103.
[2] M. Kopicki, R. Stolkin, S. Zurek, T. M¨
orwald, and J. Wyatt, “Predicting
workpiece motions under pushing manipulations using the principle of
minimum energy,” in RSS Workshop: Representation for Obj. Grasping
& Manipulation in Single & Dual Arm Tasks, Zaragoza, Spain, 2010.
[3] F. Fern´
andez and M. Veloso, “Probabilistic policy reuse in a reinforce-
ment learning agent,” in Proc. 5th Intl. Joint Conf. on Autonomous
agents and Mmultiagent Systems. ACM, 2006, pp. 720–727.
[4] C. Gramkow, “On averaging rotations,Journal of Mathematical Imag-
ing and Vision, vol. 15, no. 1-2, pp. 7–16, 2001.
[5] L. Vicci, “Averages of Rotations and Orientations in 3-space,” UNC
Chapel Hill, Department of Computer Science (unpublished), 2001.
[6] F. L. Markley, Y. Cheng, J. L. Crassidis, and Y. Oshman, “Averaging
Quaternions,” in AAS/AIAA Space Flight Mech., 2007, pp. 1675–1682.
[7] M. R. Pierrynowski and K. A. Ball, “Oppugning the assumptions
of spatial averaging of segment and joint orientations,Journal of
Biomechanics, vol. 42, no. 3, pp. 375–378, Feb. 2009.
[8] R. Hartley, J. Trumpf, Y. Dai, and H. Li, “Rotation Averaging,” Intl.
Journal of Computer Vision, vol. 103, no. 3, pp. 267–305, Jul. 2013.
[9] M. Moakher, “Means & averaging in the group of rotations,SIAM
Matrix Analysis & Applications, vol. 24, no. 1, pp. 1–16, 2002.
[10] D. J. Duff, “Visual motion estimation and tracking of rigid bodies by
physical simulation,” PhD Thesis, Univ. Birmingham, Dec. 2011.
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
We are investigating the problem of predicting how objects behave under manipulative actions. In particular, we wish to predict the workpiece motions which will result from simple pushing manipulations by a single robotic fingertip. Such interactions are themselves fundamental components of multi-fingered grasping and other complex interactions. Physics simulators can be used to do this, but they model many kinds of object interactions poorly, being dependent on detailed scene descriptions and parameters, which in practice are often difficult to tune. Additionally, we have previously investigated ways of learning to predict, by employing density estimation techniques to learn, from many example pushes, a probabilistic mapping between applied pushing motions and resulting work-piece motions. In contrast, this paper presents an alternative approach to prediction, which does not rely on learning but infers the likelihood of possible workpiece motions by using the simple physics principle of minimum energy. This approach is advantageous in situations where insufficient prior knowledge is available for training our learned predictors. In such situations, possible strategies include either training learned predictors on unrealistic simulation data, or making use of the simple physics approach which requires no training. We show that the second of these strategies performs significantly better, and approaches the performance of learned predictors are trained on observations of real object motions.
Full-text available
Researchers have derived an algorithm that determines an optimal average quaternion from a set of scalar- or matrix-weighted quaternions. Another methor method has been derived to maximize likelihood estimation and develop sufficient conditionfor the uniqueness of the average quaternion and thec equivalence of the minimization problem. The solution to the problem involved performing an eigenvalue or eigenvector decomposition of a matrix that is composed of the given average quaternions and weights. It was found that the optimal average quaternion can be determined by the computationally efficient QUEST algorithm, for both the scalar- and matrix-weighted cases. The average quaternion was found to be a maximum likelihood estimate in the matrix-weighted case, when the matrix was derived by the inverse of the covariance of the small attitude vector errors.
Full-text available
This paper proposes a new approach of model-based 3D object tracking in real-time. The developed algorithm uses edges as fea-tures to track, which are easy and robust to detect. It exploits the functionality of modern highly parallel graphics boards by perform-ing hidden face removal, image processing, texturing and particle filtering. Using a standard 3D -model the tracker requires neither memory and time -consuming training nor other pre-calculations. In contrast to other approaches this tracker also works on objects where geometry edges are barely visible because of low contrast. Best to read in color.
Conference Paper
Full-text available
We contribute Policy Reuse as a technique to improve a re- inforcement learning agent with guidance from past learned similar policies. Our method relies on using the past poli- cies as a probabilistic bias where the learning agent faces three choices: the exploitation of the ongoing learned pol- icy, the exploration of random unexplored actions, and the exploitation of past policies. We introduce the algorithm and its major components: an exploration strategy to in- clude the new reuse bias, and a similarity function to esti- mate the similarity of past policies with respect to a new one. We provide empirical results demonstrating that Pol- icy Reuse improves the learning performance over different strategies that learn without reuse. Interestingly and almost as a side effect, Policy Reuse also identifies classes of similar policies revealing a basis of core policies of the domain. We demonstrate that such a basis can be built incrementally, contributing the learning of the structure of a domain.
Full-text available
In this paper we give precise definitions of different, properly invariant notions of mean or average rotation. Each mean is associated with a metric in SO(3). The metric induced from the Frobenius inner product gives rise to a mean rotation that is given by the closest special orthogonal matrix to the usual arithmetic mean of the given rotation matrices. The mean rotation associated with the intrinsic metric on SO(3) is the Riemannian center of mass of the given rotation matrices. We show that the Riemannian mean rotation shares many common features with the geometric mean of positive numbers and the geometric mean of positive Hermitian operators. We give some examples with closed-form solutions of both notions of mean.
Full-text available
Introduction Finding the mean #average# of n arbitrary 3D rotations or orientations is conceptually challenging both in its algebra and its geometric interpretation. Tomy knowledge, no solutions have been found for the general case, although a number of approximate solutions have been proposed #3, 4, 5# to accurately and e#ciently handle those limiting cases typical of engineering interest. This report proposes an exact, closed form solution which is extended from the 2D complex number domain where the averaging methods are widely accepted. Representing rotations as rotation #unit# quaternions #1# is convenient in that a rotation composed of a sequence of successive rotations can be simply represented as the #non-commutative# product of the successive rotation quaternions. But this very property is problematical for averaging. A mean average is conventionally taken as a sum of averagands divided by their number. If applied to rotation quaternions however, the resul
This paper is conceived as a tutorial on rotation averaging, summarizing the research that has been carried out in this area; it discusses methods for single-view and multiple-view rotation averaging, as well as providing proofs of convergence and convexity in many cases. However, at the same time it contains many new results, which were developed to fill gaps in knowledge, answering fundamental questions such as radius of convergence of the algorithms, and existence of local minima. These matters, or even proofs of correctness have in many cases not been considered in the Computer Vision literature. We consider three main problems: single rotation averaging, in which a single rotation is computed starting from several measurements; multiple-rotation averaging, in which absolute orientations are computed from several relative orientation measurements; and conjugate rotation averaging, which relates a pair of coordinate frames. This last is related to the hand-eye coordination problem and to multiple-camera calibration.
In this paper two common approaches to averaging rotations are compared to a more advanced approach based on a Riemannian metric. Very often the barycenter of the quaternions or matrices that represent the rotations are used as an estimate of the mean. These methods neglect that rotations belong to a non-linear manifold and re-normalization or orthogonalization must be applied to obtain proper rotations. These latter steps have been viewed as ad hoc corrections for the errors introduced by assuming a vector space. The article shows that the two approximative methods can be derived from natural approximations to the Riemannian metric, and that the subsequent corrections are inherent in the least squares estimation.
Movement scientists frequently calculate "arithmetic averages" when examining body segment or joint orientations. Such calculations appear routinely, yet are fundamentally flawed. Three-dimensional orientation data are computed as matrices, yet three-ordered Euler/Cardan/Bryant angle parameters are frequently used for interpretation. These parameters are not geometrically independent; thus, the conventional process of averaging each parameter is incorrect. The process of arithmetic averaging also assumes that the distances between data are linear (Euclidean); however, for the orientation data these distances are geodesically curved (Riemannian). Therefore we question (oppugn) whether use of the conventional averaging approach is an appropriate statistic. Fortunately, exact methods of averaging orientation data have been developed which both circumvent the parameterization issue, and explicitly acknowledge the Euclidean or Riemannian distance measures. The details of these matrix-based averaging methods are presented and their theoretical advantages discussed. The Euclidian and Riemannian approaches offer appealing advantages over the conventional technique. With respect to practical biomechanical relevancy, examinations of simulated data suggest that for sets of orientation data possessing characteristics of low dispersion, an isotropic distribution, and less than 30 degrees second and third angle parameters, discrepancies with the conventional approach are less than 1.1 degrees . However, beyond these limits, arithmetic averaging can have substantive non-linear inaccuracies in all three parameterized angles. The biomechanics community is encouraged to recognize that limitations exist with the use of the conventional method of averaging orientations. Investigations requiring more robust spatial averaging over a broader range of orientations may benefit from the use of matrix-based Euclidean or Riemannian calculations.
  • F L Markley
  • Y Cheng
  • J L Crassidis
  • Y Oshman
F. L. Markley, Y. Cheng, J. L. Crassidis, and Y. Oshman, " Averaging Quaternions, " in AAS/AIAA Space Flight Mech., 2007, pp. 1675–1682.