Skip to main content
Version: 1.0 (Latest)

Uncertainty in Raptor Guide

Before using a refined pose in a sensor-fusion framework, always check that the Confidence field of the result is above your acceptance threshold.

The covariance alone is not sufficient — it only describes uncertainty assuming the pose was estimated correctly.

What the SDK returns​

updatePose() / updatePosition() return a std::variant of either a successful PoseOutput / PositionOutput or a PoseEstimationError. On success the output carries, in decreasing order of practical importance, a Confidence score and an optional output covariance: a 6×6 matrix on PoseOutput::covariance or a 3×3 matrix on PositionOutput::covariance, always expressed in the camera frame (DIN 9300: x forward, y right, z down).

An error is returned when no pose could be produced at all — for example when large portions of the image are sky, contrast is too low, or the prior pose lies outside the loaded map.

Confidence score​

Confidence reflects how well the algorithm matched the camera image to the rendered map. It accounts for the number of established correspondences and how well those correspondences predict the position of image features after the refined pose is applied.

A low Confidence indicates few correspondences were used, or the correspondences had a high probability of being wrong. Results below ~0.5 typically have very large position error and should be rejected.

Output covariance​

The output covariance is derived by estimating an image correspondence covariance from the matcher's correlation scores and propagating it to the pose parameters via the Jacobian of the pose estimate. It therefore reflects:

  • the uncertainty of each correspondence (large in low-contrast or motion-blurred regions, smaller in high-contrast regions), and
  • the geometric configuration of correspondences in the image. Points clustered in one image region produce a large position covariance; points on a near-collinear arrangement produce uncertainty primarily orthogonal to that line.

Limitations​

The covariance is a best-effort Gaussian approximation conditioned on the pose being roughly correct. It assumes:

  • the residual error is Gaussian,
  • the camera is a perfect pinhole, and
  • all point correspondences are correctly matched.

In practice, none of these hold exactly. In particular, calibration error (lens distortion, principal-point offset, inaccurate field of view) is not modelled by either the covariance or the Confidence score; the further the real camera deviates from a calibrated pinhole, the larger the unmodelled component of the position error.