Visual localization systems sometimes fail when the scene contains repetitive structures, such as identical building facades or corridors. In these cases, an incorrect pose candidate can receive more inlier support than the accurate one, leading the system to select the wrong result. The paper addresses this specific failure mode in repeated localization tasks.
The authors propose History-Guided Pose Reselection (HPR), which leverages coarse position forecasts from up to five previous observations. By comparing the current pose estimate against this history, the method can detect when the selected pose is likely wrong and trigger a reselection. This allows the system to recover from the failure rather than committing to the erroneous pose.
The approach is presented as a way to improve robustness in scenarios where visual localization is performed repeatedly, such as robot navigation or augmented reality. Because the method relies on existing pose history, it does not require additional sensors or changes to the underlying localization pipeline. The paper is a single source, so no independent comparisons are available here.