Confidence-Gated Pose Propagation for Scene Coordinate Regression under Incomplete Coverage

Hang Yu

Abstract: LiDAR relocalization systems based on Scene Coordinate Regression typically assume dense coverage of the target route during training. When deployed in city-scale environments, however, training logs often miss substantial portions of the route, and SCR networks can produce overconfident but inaccurate absolute pose estimates in unseen segments. This paper studies global localization under incomplete scene coverage and introduces a confidence-gated pose propagation strategy. Our model jointly predicts an absolute pose hypothesis and scene-level features from each LiDAR scan, and uses a temporal reliability estimator to detect when the SCR output is likely to fail. During such intervals, a decision policy replaces direct coordinate regression with relative motion propagation from previously reliable poses, preserving global consistency until the system re-enters familiar regions. We evaluate the approach on UrbanUnseen, together with controlled known-to-unknown protocols on Oxford and NCLT. The results demonstrate consistent gains under partial training coverage and competitive accuracy when the full route is available.