Quantile Transfer for Reliable Operating Point Selection in Visual Place Recognition

TL;DR

This work tackles the problem of selecting a reliable image-matching threshold for Visual Place Recognition under a user-defined precision target. It introduces a quantile transfer framework that uses a small calibration traversal to estimate precision-constrained thresholds and transfers them to deployment via quantile normalization of similarity score distributions. The approach is validated across seven VPR techniques and three challenging datasets, demonstrating consistent improvements in high-precision regimes and robustness to sampling variability, with up to about 25% recall gains. The method reduces manual tuning, adapts to new environments, and generalizes across operating conditions, offering practical benefits for reliable localization in GNSS-denied scenarios.

Abstract

Visual Place Recognition (VPR) is a key component for localisation in GNSS-denied environments, but its performance critically depends on selecting an image matching threshold (operating point) that balances precision and recall. Thresholds are typically hand-tuned offline for a specific environment and fixed during deployment, leading to degraded performance under environmental change. We propose a method that, given a user-defined precision requirement, automatically selects the operating point of a VPR system to maximise recall. The method uses a small calibration traversal with known correspondences and transfers thresholds to deployment via quantile normalisation of similarity score distributions. This quantile transfer ensures that thresholds remain stable across calibration sizes and query subsets, making the method robust to sampling variability. Experiments with multiple state-of-the-art VPR techniques and datasets show that the proposed approach consistently outperforms the state-of-the-art, delivering up to 25% higher recall in high-precision operating regimes. The method eliminates manual tuning by adapting to new environments and generalising across operating conditions. Our code will be released upon acceptance.

Figures (6)

• Figure 1: Overview of the proposed quantile transfer method for operating point selection in Visual Place Recognition (VPR). A calibration traversal with known correspondences is used to estimate thresholds that satisfy a user-defined precision requirement. These thresholds are converted into quantiles of the calibration score distribution and transferred to deployment traversals, yielding operating thresholds that maximise recall while ensuring the desired precision, without requiring ground-truth labels online.
• Figure 2: Methodology of our proposed quantile transfer approach. A VPR technique is first applied to generate similarity matrices from a database traversal and both calibration and deployment queries. In the calibration stage, queries with ground-truth correspondences form a calibration similarity matrix. Each deployment query is then matched to its most similar calibration queries using a correlation matrix, yielding adapted calibration subsets. From these subsets, thresholds that maximise recall under the user-specified precision requirement are determined and expressed as quantiles of the calibration score distributions. Finally, these quantiles are transferred to the deployment similarity matrix to obtain matching thresholds that satisfy the precision constraint and maximise recall without requiring ground-truth labels during deployment.
• Figure 3: Area Under Performance Curve (AUPC; lower is better) for MegaLoc across Nordland, SFU Mountain, and Oxford RobotCar datasets. Results compare the baseline by Schubert et al. schubert2021beyond, the baseline schubert2021beyond with calibration, our method without quantile transfer, and our full (i.e. with quantile transfer) method. Across all environments, our approach consistently achieves lower AUPC, especially under higher precision requirements (i.e., Nordland and SFU Mountain), demonstrating that quantile transfer stabilises threshold selection and improves robustness. Performance matches or exceeds alternatives even in easier settings (i.e., Oxford RobotCar), highlighting that the method generalises across environments. Yellow colouring highlights the 100% precision operating region.
• Figure 4: Dynamic matching thresholds for MegaLoc on Nordland (database: Summer, query: Winter). The oracle threshold (violet) reflects the best possible operating point given ground-truth, while our method (green) tracks the oracle far more closely than the baseline (red). The baseline persistently underestimates the threshold (mean 0.24 vs. oracle mean 0.47), leading to excess false positives. In contrast, our method maintains thresholds near the oracle (mean 0.49), demonstrating that quantile transfer provides adaptive and reliable thresholding across long traversals.
• Figure 5: Qualitative examples comparing our method and the baseline across Nordland and SFU Mountain datasets. Our method accepts more true positives that the baseline incorrectly rejects (top row) while simultaneously filtering out false positives that the baseline would admit (bottom rows). These cases illustrate that our approach achieves a better balance between recall and precision, adapting thresholds to accept informative matches without sacrificing reliability.