6D Pose Uncertainty in Robotic Perception

Tracking the spin on a ping pong ball with the quaternion Bingham filter

“…The solid, thin, blue line in each plot marked "oracle prior" shows results from running QBF-2 with a best-case-scenario prior, centered on the average ground truth spin for that spin type, with Λ = (−10, −10, −10). We show mean orientation and spin errors (to regressed ground truth), and also spin classification accuracy using the MAP estimate of spin type (out of 12) given the current spin belief 6 . The results clearly show that QBF-2 does the best job of identifying and tracking the ball rotations on this extremely challenging dataset, achieving a classification rate of 91% after just 30 video frames, and a mean spin (quaternion) error of 0.17 radians (10 degrees), with an average of 6.1 degrees of logo axis error and 6.8 degrees of logo angle error.…”

“…In Figure 8 we show an example of the output of the second-order QBF we used to track the orientation and 6 Spin was classified into one of the 12 spin types by taking the average ground truth spin for each spin type and choosing the one with the highest likelihood with respect to the current spin belief. spin on the ball through one of the underpin/left-sidespin trajectories.…”

“…Putting a Gaussian distribution directly in quaternion space does not respect the underlying topology of 3-D rotations; however, this approach has been the basis of tracking methods based on approximations of the Kalman filter [13], [4], [15], [5], [10], [11]. A more appropriate method is to represent distributions in an R 3 space that is tangent to the quaternion hypersphere at the mean of the distribution [6]; but such a tangent-space approach will be unable to effectively represent distributions that have large variances. In many perceptual problems, it may be possible to make observations that provide significant information about only one or two dimensions, yielding highvariance estimates.…”

An experience-driven robotic assistant acquiring human knowledge to improve haptic cooperation

“…• the 6-dimensional wrench 0 u in inertial coordinates As unified Gaussian computations on 6-D poses remain a computationally extensive problem [21], we decided to use angular velocities as unambiguous training input. The online learning algorithm is parameterized as follows: Each segment and each node is encoded in an HMM with 15 states and 1 Gaussian per state.…”

Active perception and scene modeling by planning with probabilistic 6D object poses

“…In [10] we discussed several approaches of probabilistic 6D pose representations. In this work we decided to use a Rodrigues vector (or axis/angle) representation to describe the orientation because of its good applicability and simple comprehensibility.…”