Monocular SLAM Under Corruption: Learned Trackers Trade Catastrophic Failure for Drift

zhenjun_zhao · x · 2026-09-01

Paper: Failure or Drift? Evaluating Monocular SLAM under Synthetic and Real-World Corruptions (3rd NeuSLAM workshop, ECCV 2026).

Evaluates classical ORB-SLAM2 and learned trackers DPVO and DROID-SLAM under image-space, geometry-aware, and compound corruptions, compared with real adverse conditions from 4Seasons.
Rather than reducing robustness to a single trajectory error, the evaluation separates explicit tracking failure from drift accumulated by methods that remain active.
Findings: learned trackers largely replace catastrophic loss with sustained, sometimes severe drift; the apparent ordering of learned systems changes with the physical fidelity of the corruption — structured rain/fog proxies preserve real-world ordering, a simple illumination proxy does not. Code is released.

More from Embodied