Introduction
Evaluation of Simultaneous Localization and Mapping (SLAM) and Visual Odometry (VO) systems typically relies on two primary metrics: Absolute Trajectory Error (ATE), which assesses global consistency, and Relative Error (RE), which measures local drift. While empirical evidence suggests these metrics are highly correlated, a formal theoretical link has been missing. Furthermore, standard evaluation tools often struggle with temporal association—matching SLAM estimates to ground truth when sampling rates differ or data gaps exist. This paper addresses these issues by reframing trajectory evaluation within a probabilistic, continuous-time framework, proving the theoretical equivalence of ATE and RE and providing a robust method for handling timing mismatches.
Probabilistic Error Modeling
The authors propose treating ground truth not as fixed points, but as a distribution of random variables. By redefining ATE and RE as likelihoods and utilizing Lie Group theory (specifically SE(3)) for 3D poses, they demonstrate a fundamental connection between the two. Specifically, by marginalizing out the unobservable global alignment (latent variables), the absolute error formula mathematically transforms into the relative error formula to the first order.
Continuous-Time Gaussian Processes
To resolve temporal association issues, the ground truth trajectory is represented as a piecewise Gaussian Process (GP). Unlike standard interpolation methods, a GP provides a continuous estimate of position along with a corresponding uncertainty. This uncertainty acts as a natural buffer: if a SLAM estimate occurs during a gap in ground truth data, the GP reports higher uncertainty, which "softens" the penalty in the likelihood-based error metric. This approach ensures that SLAM algorithms are not unfairly penalized for errors arising from ground truth interpolation artifacts.
Key Insights and Results
The core contribution of this work is the first theoretical proof showing that generalized versions of RE and ATE are equivalent to the first order. This "Aha!" moment reveals that ATE is essentially RE when the global reference frame is treated as a latent variable to be integrated out. Additionally, the continuous-time framework allows for evaluation that is agnostic to sensor sampling rates, providing a more principled approach to temporal association than naive "closest timestamp" matching. While primarily theoretical, the framework's utility is demonstrated using synthetic examples and contextualized within standard benchmarks like KITTI and TUM RGB-D.
Limitations and Future Work
While providing a significant theoretical advancement, the current proof of equivalence relies on first-order Taylor expansions; investigating higher-order terms is a necessary next step. The piecewise nature of the GP can also lead to slight discontinuities at segment boundaries, suggesting a need for smoother continuity constraints. Finally, the computational complexity of GPs may require sparse approximations for very long trajectories, and further experimental validation is needed to compare this probabilistic approach against traditional tools in diverse real-world scenarios.
Summary updated