Through the Lens of Doubt: Robust and Efficient Uncertainty Estimation for Visual Place Recognition

TL;DR

The paper addresses robust uncertainty estimation for Visual Place Recognition (VPR) under varying environments by introducing three training-free metrics that operate on top-$k$ similarity scores: Ratio Spread ($RS$), Similarity Distribution ($SD$), and their combination Statistical Uncertainty ($SU$). These metrics require no retraining or architectural changes and are evaluated across nine VPR methods and six datasets, achieving a mean AUC-PR of about $0.92$ with runtime under $0.03$ ms per query, enabling real-time deployment. The approach improves end-to-end VPR precision under uncertainty and generalizes across architectures, though it may be less effective in extremely repetitive scenes without geometric cues. Future work includes integrating lightweight geometric or semantic information and adapting the weighting between RS and SD based on distribution characteristics.

Abstract

Visual Place Recognition (VPR) enables robots and autonomous vehicles to identify previously visited locations by matching current observations against a database of known places. However, VPR systems face significant challenges when deployed across varying visual environments, lighting conditions, seasonal changes, and viewpoints changes. Failure-critical VPR applications, such as loop closure detection in simultaneous localization and mapping (SLAM) pipelines, require robust estimation of place matching uncertainty. We propose three training-free uncertainty metrics that estimate prediction confidence by analyzing inherent statistical patterns in similarity scores from any existing VPR method. Similarity Distribution (SD) quantifies match distinctiveness by measuring score separation between candidates; Ratio Spread (RS) evaluates competitive ambiguity among top-scoring locations; and Statistical Uncertainty (SU) is a combination of SD and RS that provides a unified metric that generalizes across datasets and VPR methods without requiring validation data to select the optimal metric. All three metrics operate without additional model training, architectural modifications, or computationally expensive geometric verification. Comprehensive evaluation across nine state-of-the-art VPR methods and six benchmark datasets confirms that our metrics excel at discriminating between correct and incorrect VPR matches, and consistently outperform existing approaches while maintaining negligible computational overhead, making it deployable for real-time robotic applications across varied environmental conditions with improved precision-recall performance.

Figures (7)

• Figure 1: Overview of our training-free uncertainty estimation metrics integrated into the VPR pipeline, operating directly on similarity scores to provide reliable confidence estimates without requiring model modifications. Shown in Stage 3 in the diagram, examples of the initial match being wrong and having high uncertainty and the following image, which is correct, is the least uncertain image from the corresponding top-$k$ on the Pittsburgh250k dataset using Eigenplaces.
• Figure 2: Visual illustration of uncertainty metrics operating on similarity score distributions. RS captures competitive ambiguity through score clustering among top-k candidates, while SD measures match distinctiveness using the distance between maximum and median scores. SU combines both metrics to provide unified uncertainty estimation.
• Figure 3: Visual examples of the top 4 matches for two queries using APGeM on the Pittsburgh250k dataset. Each retrieved image displays its similarity score, with the RS, SD and SU values shown below each query image. The top row demonstrates a correct retrieval with low uncertainty, while the bottom row shows incorrect matches with high uncertainty.
• Figure 4: Performance difference (AUC-PR) for each VPR and uncertainty metric combination across datasets.
• Figure 5: Qualitative comparison on the Pittsburgh250k dataset. Top: correct uncertainty identification; bottom: failure cases. Query (left) and top-ranked retrieval (right) pairs shown.