Pose-conditioned and uncertainty-aware scene representation for visual relocalization

Visual relocalization requires compact scene representations that capture global spatial context while preserving local geometric precision. Existing regression-based methods often emphasize one aspect at the expense of the other. Global image representations encode scene-level spatial information but provide limited geometric accuracy, whereas locally regressed scene coordinates preserve explicit 2D–3D geometry yet remain vulnerable to visual ambiguity and unreliable predictions. We propose a pose-conditioned and uncertainty-aware scene representation framework that integrates these complementary cues within a unified relocalization pipeline. Global pose supervision is used to learn a pose-aware image representation that conditions local features for scene-coordinate regression while preserving their spatial structure. Scene coordinates are further modeled with heteroscedastic uncertainty to characterize spatially varying prediction reliability. The estimated uncertainty is propagated into geometric pose estimation to guide correspondence sampling, pose-hypothesis evaluation, and inlier selection. On the 7-Scenes benchmark, the proposed method achieves an average localization success rate of 88.5% under the 5 c m / 5 ° criterion with a compact 47 MB model. Qualitative analysis further suggests that high-uncertainty predictions are often associated with geometrically inconsistent scene coordinates, supporting the use of uncertainty as a reliability cue for pose estimation. These results demonstrate the effectiveness of jointly exploiting pose-aware global context, local geometric correspondence, and predictive uncertainty for compact visual relocalization.