Hi all, I’m an MEng Robotics & AI student at UCL working on perception/tracking pipelines (recently built a person re-ID and tracking evaluation pipeline, and a Gaussian-splat reconstruction quality evaluator). I’m trying to understand a specific problem better: how do teams currently notice when a perception or sensor-fusion stack has silently degraded in production, before it causes a visible failure? Do you rely on manual spot-checks, logging + alerts, a formal calibration schedule, or something else? Genuinely trying to learn what’s actually painful here vs. what’s already a solved problem, not selling anything. Would appreciate any war stories, even short ones.
What we did is running a pose estimator for each detection, which quickly filters out single-frame false positives. When we knew what we’re looking for, we also verified the estimated size of the object (based on its estimated distance). You can also look at the softmax values from the detector and if the chosen class does not win by a big margin, you can say you don’t trust the detection that much.