pith. sign in

arxiv: 2605.19435 · v1 · pith:XNFG4AMPnew · submitted 2026-05-19 · 💻 cs.CV · cs.AI

KappaPlace: Learning Hyperspherical Uncertainty for Visual Place Recognition via Prototype-Anchored Supervision

Maya Yanko , Yoli Shavit This is my paper

Pith reviewed 2026-05-20 05:43 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords visual place recognitionuncertainty estimationvon Mises-Fisher distributionprototype supervisionaleatoric uncertaintycalibration errorhyperspherical embeddings
0
0 comments X

The pith

KappaPlace learns calibrated uncertainty for visual place recognition by supervising von Mises-Fisher descriptors against class prototypes.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Visual place recognition systems must not only retrieve correct matches but also flag when a query is ambiguous or a match is unreliable. KappaPlace introduces prototype-anchored supervision that treats latent class representatives as targets in a probabilistic objective. Image descriptors are modeled as von Mises-Fisher variables on the hypersphere, and a lightweight module predicts the concentration parameter to serve as an aleatoric uncertainty signal. This yields a match-level reliability score rather than a query-only view. On five benchmarks the approach cuts expected calibration error by up to half while preserving or raising retrieval recall, and it works both in joint training and as a post-hoc addition to frozen backbones.

Core claim

The central claim is that anchoring supervision to class prototypes while modeling descriptors as von Mises-Fisher distributions produces concentration parameters that directly quantify match-level aleatoric uncertainty, delivering substantially lower calibration error than prior VPR methods without any post-hoc recalibration or access to ground-truth match labels at inference time.

What carries the argument

Prototype-anchored supervision, which uses latent class representatives as targets for a probabilistic objective to train a module that predicts the concentration parameter of von Mises-Fisher image descriptors.

If this is right

  • VPR pipelines can now reject or re-query ambiguous matches in real time using the learned uncertainty signal.
  • A match-level rather than query-only uncertainty formulation becomes available for downstream planning.
  • Existing frozen VPR backbones can be extended with the lightweight concentration predictor without retraining the entire model.
  • Calibration improvements hold across indoor, outdoor, and seasonal benchmark shifts while recall is maintained.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same prototype-anchored scheme could be applied to other hyperspherical embedding tasks such as face recognition or metric learning.
  • Robotics systems might use the uncertainty value to modulate exploration or fallback localization strategies.
  • If concentration parameters prove stable under domain shift, the method could reduce the need for frequent model recalibration in deployed navigation.

Load-bearing premise

Prototype-anchored supervision produces concentration parameters that are already well-calibrated measures of aleatoric uncertainty and require neither post-hoc recalibration nor ground-truth match correctness at inference.

What would settle it

Measure whether the predicted concentration parameters correlate with actual match error rates on held-out query-reference pairs; if the correlation is absent or expected calibration error stays comparable to baselines, the claim is falsified.

Figures

Figures reproduced from arXiv: 2605.19435 by Maya Yanko, Yoli Shavit.

Figure 1
Figure 1. Figure 1: Aggregated ECE@K calibration results across benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Aggregated Recall@K results across uncertainty-sorted bins. The diagonal represents [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: For three representative queries, we display the top-3 references ( [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: Match-level Confidence Analysis. Top-3 retrievals for queries with high, medium, and low [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the KappaPlace architecture during training and inference. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
read the original abstract

Visual Place Recognition (VPR) is critical for autonomous navigation, yet state-of-the-art methods lack well-calibrated uncertainty estimation. Standard pipelines cannot reliably signal when a query is ambiguous or a match is likely incorrect, posing risks in safety-critical robotics. We propose KappaPlace, a principled framework for learning uncertainty-aware VPR representations. Our core contribution is a Prototype-Anchored supervision strategy that leverages latent class representatives as targets for a probabilistic objective. By modeling image descriptors as von Mises-Fisher (vMF) variables, we learn a lightweight module to predict the concentration parameter as a direct proxy for aleatoric uncertainty. While existing VPR uncertainty methods are typically restricted to a query-centric view, we derive a novel match-level formulation to quantify the reliability of specific query-reference pairs. Across five diverse benchmarks, KappaPlace reduces Expected Calibration Error (ECE@K) by up to 50% compared to existing methods while maintaining or improving retrieval recall. We provide both a joint-training variant and a post-training extension for frozen backbones. Our results demonstrate that KappaPlace provides a robust, stable, and well-calibrated signal that enables reliable decision-making within the VPR pipeline. Our code is available at: https://github.com/mayayank95/UncertaintyAwareVPR

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces KappaPlace, a framework for uncertainty-aware visual place recognition (VPR). It models descriptors as von Mises-Fisher (vMF) distributions on the hypersphere and proposes a prototype-anchored supervision strategy to train a lightweight module that predicts the concentration parameter kappa as a direct proxy for aleatoric uncertainty. A novel match-level formulation is derived to assess reliability of specific query-reference pairs. Experiments across five benchmarks report up to 50% reduction in Expected Calibration Error (ECE@K) relative to prior methods, while maintaining or improving retrieval recall; both joint-training and post-training (frozen-backbone) variants are presented, with code released.

Significance. If the calibration results hold under independent verification, the work addresses a practical gap in safety-critical VPR by supplying a stable, query-reference-specific uncertainty signal without post-hoc recalibration. The open-source code release is a clear strength that aids reproducibility and follow-up work.

major comments (2)
  1. [Evaluation section] Evaluation section (results on ECE@K): the reported reductions of up to 50% are load-bearing for the central claim, yet the manuscript provides no explicit statement on whether an independent calibration hold-out set was withheld from the kappa-prediction training or whether the positive/negative pairs used to compute ECE@K overlap with those seen during prototype-anchored supervision. This leaves open the possibility that kappa is optimized for the surrogate vMF or contrastive objective rather than true match-probability calibration.
  2. [§3 and §4] §3 (Prototype-Anchored Supervision) and §4 (Match-Level Formulation): the claim that kappa serves as a direct, well-calibrated proxy for aleatoric uncertainty rests on the modeling assumption that anchoring to class prototypes during training produces concentration values that generalize to query ambiguity and descriptor noise. The paper should supply a concrete diagnostic (e.g., correlation between predicted kappa and observed match error rates on a distribution-shifted test set, or comparison against isotonic recalibration) to rule out the alternative that kappa primarily encodes intra-class compactness of the training places.
minor comments (2)
  1. [Abstract / Introduction] The abstract states results on 'five diverse benchmarks' but does not name them; the introduction or experimental setup should list the datasets explicitly for immediate context.
  2. [Method] Notation for the vMF density and the match-level probability should be introduced once with a clear equation reference rather than scattered across sections.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the work's significance. We address each major comment below with clarifications based on the experimental setup and indicate the specific revisions planned for the next manuscript version.

read point-by-point responses

  1. Referee: [Evaluation section] Evaluation section (results on ECE@K): the reported reductions of up to 50% are load-bearing for the central claim, yet the manuscript provides no explicit statement on whether an independent calibration hold-out set was withheld from the kappa-prediction training or whether the positive/negative pairs used to compute ECE@K overlap with those seen during prototype-anchored supervision. This leaves open the possibility that kappa is optimized for the surrogate vMF or contrastive objective rather than true match-probability calibration.

    Authors: We thank the referee for this observation. The kappa-prediction module (whether trained jointly or post-hoc) uses only the training splits for prototype-anchored supervision. All ECE@K metrics are computed exclusively on the standard test splits of the five benchmarks, which are disjoint from training data and contain no pairs seen during supervision. To remove any ambiguity, we will add an explicit paragraph in the revised Evaluation section (Section 5) stating the data partitioning and confirming that calibration evaluation uses completely held-out test pairs. This is a clarification rather than a change to the results. revision: yes

  2. Referee: [§3 and §4] §3 (Prototype-Anchored Supervision) and §4 (Match-Level Formulation): the claim that kappa serves as a direct, well-calibrated proxy for aleatoric uncertainty rests on the modeling assumption that anchoring to class prototypes during training produces concentration values that generalize to query ambiguity and descriptor noise. The paper should supply a concrete diagnostic (e.g., correlation between predicted kappa and observed match error rates on a distribution-shifted test set, or comparison against isotonic recalibration) to rule out the alternative that kappa primarily encodes intra-class compactness of the training places.

    Authors: We agree that an explicit diagnostic would strengthen the justification for kappa as a generalizable uncertainty signal. While the existing multi-benchmark results already span datasets with distribution shifts, we will add the requested analyses in the revised manuscript: (i) Pearson correlation between predicted kappa and observed match error rates on held-out test sets, and (ii) a direct comparison against isotonic recalibration applied to the same descriptors. These will appear in an expanded Section 5.3 with accompanying figures. The additions are based on additional post-hoc analysis we have conducted and support the original claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper introduces a learning framework that models image descriptors as von Mises-Fisher variables and trains a lightweight module via prototype-anchored supervision to output the concentration parameter kappa, presented as a proxy for aleatoric uncertainty. A novel match-level formulation is derived from this probabilistic model to assess query-reference reliability. The reported ECE@K reductions are empirical results measured on five external benchmarks rather than quantities forced by construction from the training objective or any self-citation chain. No load-bearing step reduces a claimed prediction or first-principles result to an input by definitional equivalence, and the method includes both joint-training and post-training variants that remain falsifiable against held-out data.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that von Mises-Fisher concentration can be learned as a direct uncertainty proxy and that prototype anchors provide sufficient supervision without introducing bias in the calibration metric.

free parameters (1)
  • kappa prediction module weights
    The lightweight module that outputs the concentration parameter is trained and therefore contains fitted parameters.
axioms (1)
  • domain assumption Image descriptors can be modeled as vMF random variables on the hypersphere.
    Stated in the abstract as the modeling choice for descriptors.

pith-pipeline@v0.9.0 · 5763 in / 1338 out tokens · 29935 ms · 2026-05-20T05:43:40.999083+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    We model image descriptors as von Mises-Fisher (vMF) variables... learn a lightweight module to predict the concentration parameter κ as a direct proxy for aleatoric uncertainty... match-level uncertainty UQ↔R = 1 / sqrt(κQ² + κR² + 2 κQ κR (zQ⊤ zR))

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages

  1. [1]

    Gsv-cities: Toward appropriate supervised visual place recognition.Neurocomputing, 2022

    Amar Ali-bey, Brahim Chaib-draa, and Philippe Giguere. Gsv-cities: Toward appropriate supervised visual place recognition.Neurocomputing, 2022

    work page 2022

  2. [2]

    Mixvpr: Feature mixing for visual place recognition

    Amar Ali-Bey, Brahim Chaib-Draa, and Philippe Giguere. Mixvpr: Feature mixing for visual place recognition. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2998–3007, 2023

    work page 2023

  3. [3]

    Donald E. Amos. Computation of modified Bessel functions and their ratios.Mathematics of Computation, 28(125):239–251, 1974

    work page 1974

  4. [4]

    Netvlad: Cnn architecture for weakly supervised place recognition

    Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5297–5307, 2016

    work page 2016

  5. [5]

    Rethinking visual geo-localization for large-scale applications

    Gabriele Berton, Carlo Masone, and Barbara Caputo. Rethinking visual geo-localization for large-scale applications. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4878–4888, 2022

    work page 2022

  6. [6]

    Eigenplaces: Train- ing viewpoint robust models for visual place recognition

    Gabriele Berton, Gabriele Trivigno, Barbara Caputo, and Carlo Masone. Eigenplaces: Train- ing viewpoint robust models for visual place recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 11080–11090, 2023

    work page 2023

  7. [7]

    Stun: Self-teaching uncertainty estimation for place recognition

    Kaiwen Cai, Chris Xiaoxuan Lu, and Xiaowei Huang. Stun: Self-teaching uncertainty estimation for place recognition. In2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6614–6621. IEEE, 2022

    work page 2022

  8. [8]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. Ininternational conference on machine learning, pages 1050–1059. PMLR, 2016

    work page 2016

  9. [9]

    Deep metric learning using triplet network

    Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. InInternational workshop on similarity-based pattern recognition, pages 84–92. Springer, 2015

    work page 2015

  10. [10]

    Optimal transport aggregation for visual place recognition

    Sergio Izquierdo and Javier Civera. Optimal transport aggregation for visual place recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17658–17668, 2024

    work page 2024

  11. [11]

    Functional clustering on a circle using von mises mixtures.Journal of Statistical Theory and Practice, 15(2):38, 2021

    S Rao Jammalamadaka, Brian Wainwright, and Qianyu Jin. Functional clustering on a circle using von mises mixtures.Journal of Statistical Theory and Practice, 15(2):38, 2021

    work page 2021

  12. [12]

    Class-relational approach for visual place recognition under extreme appearance changes.IEEE Robotics and Automation Letters, 2026

    Jaeyoon Kim, Yoonki Cho, and Sung-Eui Yoon. Class-relational approach for visual place recognition under extreme appearance changes.IEEE Robotics and Automation Letters, 2026

    work page 2026

  13. [13]

    Cricavpr: Cross-image correlation-aware representation learning for visual place recognition

    Feng Lu, Xiangyuan Lan, Lijun Zhang, Dongmei Jiang, Yaowei Wang, and Chun Yuan. Cricavpr: Cross-image correlation-aware representation learning for visual place recognition. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16772–16782, 2024

    work page 2024

  14. [14]

    To- wards seamless adaptation of pre-trained models for visual place recognition

    Feng Lu, Lijun Zhang, Xiangyuan Lan, Shuting Dong, Yaowei Wang, and Chun Yuan. To- wards seamless adaptation of pre-trained models for visual place recognition. InThe Twelfth International Conference on Learning Representations, 2024. 10

    work page 2024

  15. [15]

    Through the lens of doubt: Robust and efficient uncertainty estimation for visual place recognition.IEEE Robotics and Automation Letters, 2026

    Emily Miller, Michael Milford, Muhammad Burhan Hafez, Sarvapali Ramchurn, and Shoaib Ehsan. Through the lens of doubt: Robust and efficient uncertainty estimation for visual place recognition.IEEE Robotics and Automation Letters, 2026

    work page 2026

  16. [16]

    Fine-tuning cnn image retrieval with no human annotation.IEEE transactions on pattern analysis and machine intelligence, 41(7):1655– 1668, 2018

    Filip Radenovi´c, Giorgos Tolias, and Ondˇrej Chum. Fine-tuning cnn image retrieval with no human annotation.IEEE transactions on pattern analysis and machine intelligence, 41(7):1655– 1668, 2018

    work page 2018

  17. [17]

    A short note on parameter approximation for von mises-fisher distributions: and a fast implementation of i s (x).Computational Statistics, 27(1):177–190, 2012

    Suvrit Sra. A short note on parameter approximation for von mises-fisher distributions: and a fast implementation of i s (x).Computational Statistics, 27(1):177–190, 2012

    work page 2012

  18. [18]

    Visual place recognition with repetitive structures

    Akihiko Torii, Josef Sivic, Tomas Pajdla, and Masatoshi Okutomi. Visual place recognition with repetitive structures. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 883–890, 2013

    work page 2013

  19. [19]

    EffoVPR: Ef- fective foundation model utilization for visual place recognition

    Issar Tzachor, Boaz Lerner, Matan Levy, Michael Green, Tal Berkovitz Shalev, Gavriel Habib, Dvir Samuel, Noam Korngut Zailer, Or Shimshi, Nir Darshan, and Rami Ben-Ari. EffoVPR: Ef- fective foundation model utilization for visual place recognition. InThe Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  20. [20]

    Cosface: Large margin cosine loss for deep face recognition

    Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5265–5274, 2018

    work page 2018

  21. [21]

    Multi-similarity loss with general pair weighting for deep metric learning

    Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R Scott. Multi-similarity loss with general pair weighting for deep metric learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5022–5030, 2019

    work page 2019

  22. [22]

    Mapillary street-level sequences: A dataset for lifelong place recognition

    Frederik Warburg, Soren Hauberg, Manuel Lopez-Antequera, Pau Gargallo, Yubin Kuang, and Javier Civera. Mapillary street-level sequences: A dataset for lifelong place recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2626–2635, 2020

    work page 2020

  23. [23]

    Bayesian triplet loss: Uncertainty quantification in image retrieval

    Frederik Warburg, Martin Jørgensen, Javier Civera, and Søren Hauberg. Bayesian triplet loss: Uncertainty quantification in image retrieval. InProceedings of the IEEE/CVF International conference on Computer Vision, pages 12158–12168, 2021

    work page 2021

  24. [24]

    Lh2face: Loss function for hard high-quality face.arXiv preprint arXiv:2506.23555, 2025

    Fan Xie, Yang Wang, Yikang Jiao, Zhenyu Yuan, Congxi Chen, and Chuanxin Zhao. Lh2face: Loss function for hard high-quality face.arXiv preprint arXiv:2506.23555, 2025

    work page arXiv 2025

  25. [25]

    Amstertime: A visual place recognition benchmark dataset for severe domain shift

    Burak Yildiz, Seyran Khademi, Ronald Maria Siebes, and Jan Van Gemert. Amstertime: A visual place recognition benchmark dataset for severe domain shift. In2022 26th International Conference on Pattern Recognition (ICPR), pages 2749–2755. IEEE, 2022

    work page 2022

  26. [26]

    On the estimation of image-matching uncertainty in visual place recognition

    Mubariz Zaffar, Liangliang Nan, and Julian FP Kooij. On the estimation of image-matching uncertainty in visual place recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17743–17753, 2024. A Appendix We provide additional details regarding our experimental setup, as well as extended results that supplemen...

    work page arXiv 2024

  27. [27]

    and late (Epoch 11) training stages of the student model in STUN. Method amstertime msls-val Recall@K↑ECE@K↓Recall@K↑ECE@K↓ 1 5 10 1 5 10 1 5 10 1 5 10 Epoch 643.7 64.7 71.60.191 0.2220.25784.790.992.30.344 0.395 0.409 Epoch 11 43.6 64.4 71.00.1830.222 0.260 84.7 90.892.70.357 0.415 0.434 16

    work page arXiv