Geometric Observability Index: An Operator-Theoretic Framework for Per-Feature Sensitivity, Weak Observability, and Dynamic Effects in SE(3) Pose Estimation

Technical Summary: Geometric Observability Index for SE(3) Pose Estimation

Problem Statement
Pose estimation on the special Euclidean group $SE(3)$ is fundamental to visual odometry, SLAM, and structure-from-motion. Current mainstream back-ends determine which measurements to trust primarily based on residual magnitude (e.g., chi-square gating, robust kernels). The paper argues that this premise fails in weakly observable geometries (e.g., low-parallax or near-pure-rotation scenarios). In such cases, a measurement with a moderate residual aligned with a poorly constrained direction can shift the estimated pose significantly more than a large residual along a well-constrained direction. The amplification factor is inversely proportional to the smallest curvature eigenvalue ($1/\lambda_{min}$), which becomes unbounded as observability degrades. Consequently, relying solely on residual magnitude conflates a measurement's information content with its potential harm, leading to the over-rejection of high-leverage inliers in weakly observable regimes.

Methodology and Framework
The authors introduce the Geometric Observability Index (GOI), a per-feature sensitivity measure derived from an operator-theoretic framework on $SE(3)$.

1. Mathematical Formulation:

   • Let $H = E[J^\top W J]$ be the curvature (Gauss-Newton Hessian) matrix and $G$ be a Riemannian metric on the Lie algebra $se(3)$.
   • The curvature operator is defined as $A = G^{-1}H$, which is self-adjoint with respect to the $G$-inner product.
   • The observable subspace is $O = \text{range}(A)$. The operator $A$ is restricted to this subspace as $A_{OO}$.
   • For a measurement $z$, the score is $\psi(z) = J^\top W r(z)$, and its gradient representative is $\phi(z) = G^{-1}\psi(z)$.
   • The GOI is defined as the $G$-norm of the influence exerted by a single measurement:
     $$\text{GOI}(z) = \| A_{OO}^{-1} P_O \phi(z) \|_G = \| A^\dagger_G \phi(z) \|_G$$
     where $P_O$ is the $G$-orthogonal projector onto $O$, and $A^\dagger_G$ is the $G$-pseudoinverse.

2. Spectral Characterization:
   In a $G$-orthonormal eigenbasis of $A$ with eigenvalues $\lambda_i$, the squared index is:
   $$\text{GOI}(z)^2 = \sum_{i: \lambda_i > 0} \frac{\psi_i(z)^2}{\lambda_i^2}$$
   This form explicitly shows that GOI amplifies score energy along weakly observable directions (small $\lambda_i$).

3. Theoretical Unification:
   The framework unifies four distinct concepts under the spectral geometry of the single operator $A$:

   • Influence Function: GOI equals the $G$-norm of the M-estimator influence function (Theorem 1).
   • Fisher Information: Under Gaussian noise, the Fisher information coincides with the curvature $H$ (Theorem 2).
   • Conditioning: The smallest eigenvalue $\lambda_{min}$ governs weak observability and the amplification of noise.
   • Finite-Sample Stability: The estimation error scales as $O(\sigma / \sqrt{n \lambda_{min}})$ (Theorem 3).

4. Standardization and Chi-Square Equivalence:
   A critical theoretical result (Proposition 2) proves that when the influence is standardized by its inlier null covariance, it collapses exactly to the classical chi-square residual statistic ($r^\top W r / \sigma^2$). This establishes that residual gating is effectively a "leverage-corrected" influence test. Conversely, using raw GOI (unstandardized) for outlier gating is shown to be theoretically flawed in weakly observable geometries because it conflates information with harm.

Key Contributions

1. Exact Attribution: The paper proves that GOI is the exact first-order pose shift caused by a measurement. It predicts the true leave-one-out pose shift with a log-correlation of $r=1.00$.
2. Unification of Heuristics: The framework grounds robust kernels, chi-square tests, and condition-number thresholds into a single operator-theoretic object, explaining their relationships and failure modes.
3. Weak Observability Law: The authors derive a closed-form spectral expression showing that translational observability collapses quadratically with parallax ($\lambda_{min} \propto \text{parallax}^2$), providing a precise mechanism for degeneracy detection.
4. Stability Radius: A finite-sample stability bound is derived, linking deterministic conditioning to statistical error scaling.
5. Rigorous Validation: Unlike purely formal treatments, the paper validates every quantitative claim on controlled synthetic $SE(3)$ problems and real-world datasets (TUM RGB-D and KITTI).

Experimental Results

• Synthetic Validation:
  • Influence Law: Confirmed $r=1.00$ correlation between predicted GOI and empirical pose shifts.
  • Fisher-Curvature Identity: Verified with relative error $1.8 \times 10^{-3}$.
  • Translational Collapse: Fitted exponent of $1.98$ for $\lambda_{min}$ vs. parallax, matching the predicted quadratic scaling.
  • Stability Law: Error scaling collapsed to a single line with a 5% coefficient of variation across noise levels, sample sizes, and conditioning.
• Real-Data Studies:
  • TUM RGB-D (Well-Conditioned): In indoor scenes with high conditioning ($\text{cond}(H) \approx 10^2$), GOI gating and residual gating performed identically (parity), as predicted by the theory when the spectrum is flat.
  • KITTI (Weakly Observable): In outdoor street scenes with low parallax ($\text{cond}(H) \approx 10^4$), raw GOI gating significantly degraded performance compared to residual gating. The raw index over-rejected high-leverage inliers aligned with weak directions, confirming the theoretical prediction that unstandardized influence is misleading in noise-dominated weak geometries.

Significance and Claims
The paper claims to provide a first-principles explanation for the empirical robustness of the classical chi-square residual test: it is the correctly standardized influence function. Simultaneously, it delineates the precise failure mode of raw-influence gating: it fails in weakly observable geometries by conflating information with harm.

The authors position GOI not as a replacement for the chi-square test in outlier detection, but as the correct metric for attribution, auditing, and sensitivity analysis (identifying which measurements move the estimate and by how much). The framework unifies conditioning analysis, Fisher information, and degeneracy detection, offering a rigorous tool for diagnosing weak observability (via $\lambda_{min}$) and alignment risks (via the ratio $\rho_1$). The work is presented as a theoretical and diagnostic advancement rather than a new SLAM system, with all code released for reproducibility.