Research map / Image generation / Autoregressive and latent generation
On the Estimation of Image-matching Uncertainty in Visual Place Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Compares three families of image-matching uncertainty estimators for visual place recognition and proposes SUE, a training-free score based on the spatial spread of the poses of the K nearest reference images.
Problem
Visual place recognition retrieves a reference image for pose estimation, but highly confident false matches (e.g., from perceptual aliasing) can cause catastrophic localization failures; there is no consensus on how best to estimate matching uncertainty.
Approach
Use the spatial spread (trace of covariance) of the physical poses of the top-K retrieved reference images, weighted by exp(-λ·feature distance), as an image-matching uncertainty score; it exploits reference poses that other efficient methods ignore.
Key results
- SUE achieves average AUC-PR 0.77 across six VPR datasets, outperforming L2-distance (0.69), PA-Score (0.68), BTL (0.28) and STUN (0.54) among efficient methods.
- SUE takes about 1.08 ms per query, roughly three orders of magnitude faster than DELF-RANSAC (1587 ms), which averages 0.93 AUC-PR.
- Combining DELF geometric verification with SUE in a linear SVM improves accuracy on Sanfrancisco from 86.6 to 89.6 and on Pittsburgh from 86.0 to 87.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SUE instead of L2-distance uncertainty | VPR uncertainty estimation with ResNet-50 + GeM · Pittsburgh-250k | AUC-PR: 0.87 -> 0.94 | Helped |
| Use SUE instead of L2-distance uncertainty | VPR uncertainty estimation with ResNet-50 + GeM · Sanfrancisco | AUC-PR: 0.76 -> 0.84 | Helped |
| Use SUE instead of L2-distance uncertainty | VPR uncertainty estimation with ResNet-50 + GeM · Stlucia | AUC-PR: 0.79 -> 0.88 | Helped |
| Use SUE instead of L2-distance uncertainty | VPR uncertainty estimation with ResNet-50 + GeM · Eysham | AUC-PR: 0.87 -> 0.93 | Helped |
| Use SUE instead of L2-distance uncertainty | VPR uncertainty estimation with ResNet-50 + GeM · MSLS | AUC-PR: 0.64 -> 0.77 | Helped |
| Use SUE instead of L2-distance uncertainty | VPR uncertainty estimation with ResNet-50 + GeM · Nordland | AUC-PR: 0.18 -> 0.26 | Helped |
| Combine SUE with DELF-RANSAC uncertainty for classification | VPR uncertainty estimation · Pittsburgh-250k | accuracy: 86.0 -> 87.1 | Helped |
| Combine SUE with DELF-RANSAC uncertainty for classification | VPR uncertainty estimation · Sanfrancisco | accuracy: 86.6 -> 89.6 | Helped |
| Combine SUE with DELF-RANSAC uncertainty for classification | VPR uncertainty estimation · Stlucia | accuracy: 85.3 -> 88.7 | Helped |
| Combine SUE with DELF-RANSAC uncertainty for classification | VPR uncertainty estimation · Eysham | accuracy: 78.3 -> 82.1 | Helped |
| Combine SUE with DELF-RANSAC uncertainty for classification | VPR uncertainty estimation · MSLS | accuracy: 80.2 -> 73.4 | Hurt |
| Vary the number of nearest neighbors K (ablation) | SUE · Six VPR datasets | AUC-PR | Helped |
Surprising findings
- A simple L2-distance between query and reference descriptors is a better matching-uncertainty estimate than the recent data-driven methods BTL and STUN.
- Learned aleatoric uncertainty from training data often does not generalize to the test-time reference map.
Limitations
- SUE may underestimate uncertainty when K is too small to retrieve aliased references from multiple locations; choosing K across maps of mixed scene depth is hard.
- Unequal reference coverage of perceptually aliased locations creates a confidence bias for SUE.
- BTL was reimplemented since the code is not open-source.
Experimental setup
Datasets: Pittsburgh-250k Sanfrancisco Stlucia Eysham MSLS Nordland
Benchmarks: Pittsburgh-250k Sanfrancisco Stlucia Eysham MSLS Nordland
Metrics: AUC-PR binary classification accuracy inference time (msec)
Baselines: L2-distance PA-Score BTL STUN SIFT-RANSAC DELF-RANSAC Superpoint-RANSAC
Builds on: Bayesian Triplet Loss Self-Teaching Uncertainty Estimation perceptual aliasing score SIFT DELF SuperPoint
Closest papers by meaning
- Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · similarity 0.88)
- Towards Seamless Adaptation of Pre-trained Models for Visual Place Recognition (ICLR 2024 · similarity 0.87)
- Semantic Ambiguity Modeling and Propagation for Fine-Grained Visual Cross View Geo-Localization (AAAI 2025 · similarity 0.87)
- Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era (NeurIPS 2025 · similarity 0.87)
- MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering (NeurIPS 2025 · similarity 0.87)
- Optimal Transport Aggregation for Visual Place Recognition (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.