SAFEVPR: Patch-Based Verification with Empirical FDR Control for Cross-Condition Sequence Visual Place Recognition
SAFEVPR: Patch-Based Verification with Empirical FDR Control for Cross-Condition Sequence Visual Place Recognition
This paper presents SAFEVPR, a non-trainable pipeline for sequence visual place recognition that replaces backbone cosine similarity with mutual-nearest-neighbour patch scores from frozen DINOv2 features and employs flat Learn-Then-Test calibration with per-bin Clopper-Pearson thresholds. The method empirically satisfies target false-discovery rates across 23 cross-condition ground-vehicle setups without formal guarantees under condition shift, reducing erroneous loop-closure constraints by 29.6x and cutting median trajectory error from 95.5 m to 9.7 m in relocalization studies, though performance degrades on texture-poor scenery where the system abstains.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
percent
29.6
View original evidence
-one-dataset-out calibration, validity drops to 20/23 with 14 retaining useful coverage (strongest cosine baseline: 16/23 and 10), and on texture-poor scenery (Nordland) the verifier is uninformative and the system abstains entirely, as it does everywhere at alpha = 0.01; we report these as limitations. In a pose-graph relocalization study, gating loop-closure constraints reduces erroneous constraints by 29.6x and cu
Open source S5
- code: Available Artifact link Open source S6
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
-Test (LTT) calibration with per-bin Clopper-Pearson thresholds. Across 23 cross-condition ground-vehicle setups (Oxford RobotCar, NCLT, St Lucia; three frozen backbones), SAFEVPR meets the FDR target on 23/23 setups at alpha = 0.10 (mean accepted FDR 0.014, mean TPR 0.75) and on 22/23 at alpha = 0.05 with 22 retaining non-trivial coverage (best cosine variant, cosine+LTT: 22/23 valid but 17 non-trivial); under leave
Open source S4
-one-dataset-out calibration, validity drops to 20/23 with 14 retaining useful coverage (strongest cosine baseline: 16/23 and 10), and on texture-poor scenery (Nordland) the verifier is uninformative and the system abstains entirely, as it does everywhere at alpha = 0.01; we report these as limitations. In a pose-graph relocalization study, gating loop-closure constraints reduces erroneous constraints by 29.6x and cu
Open source S5
ts the median trajectory error from 95.5 m to 9.7 m. Code is available at https://github.com/TIERS/SafeVPR.
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org