Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Breaking D\'ej\`a Vu: Independent Auditing of Visual Place Recognition through Vision-Language Reasoning

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A vision-language model can independently audit visual place recognition, replacing hand-tuned matching thresholds with a zero-shot accept/reject decision.

desk verdict The auditor idea is worth taking seriously, but the headline recall@1 gain is an artifact of an undefined conditional metric. read the letter →

arxiv 2607.12818 v2 pith:T7G76TZN submitted 2026-07-14 cs.CV

Sania Waheed , Michael Milford , Sarvapali D. Ramchurn , Shoaib Ehsan This is my paper
classification cs.CV
keywords visualplacerecognitionloopclosuredetectionvision-languagemodelspost-retrievalverificationthreshold-freeauditingprecision-coverage-FARrisk-coverageanalysisfalseacceptancerate
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a vision-language model (VLM) can act as an independent, threshold-free auditor for visual place recognition (VPR). After a standard VPR pipeline retrieves a top candidate, the VLM reasons jointly over the query and candidate images and decides whether they show the same physical place. The paper claims this post-retrieval verification improves recall@1 by 13.6% on average over state-of-the-art verification methods, cuts false acceptance to 12%, and holds precision above 95% with coverage above 75%, all without dataset-specific calibration or ground truth. A sympathetic reader should care because robots and autonomous vehicles use VPR for loop closure, and false matches can silently corrupt their maps and trajectories. If this works, VPR becomes easier to deploy safely in unfamiliar environments.

What carries the argument

The central mechanism is the VLM verifier V(q, r) that maps a query–candidate image pair to a binary accept/reject decision, prompted with a conservative instruction to assume different places unless strong structural evidence supports a match. The framework's independence from the retrieval representation is what enables architecture-agnostic operation without retraining or environment-specific thresholds. The paper also introduces the precision-coverage-FAR triplet and risk-coverage curves as alternatives to AUC-PR for evaluating such verifiers.

What would settle it

Recompute recall@1 over all query images, counting rejected matches as incorrect (or as misses). If the 13.6% average improvement over baselines vanishes or becomes negative under that definition, the headline claim does not hold as stated.

Watch

Extended reading notes

Core claim

On the paper's own terms, VLM-based auditing provides a verification signal that is independent of the descriptor space used for retrieval. Instead of thresholding similarity scores or uncertainty estimates derived from the VPR model, the framework prompts a VLM to check whether the query and retrieved image depict the same location, using structural and semantic cues such as layout, geometry, and landmarks while ignoring dynamic content. The paper reports that this approach outperforms embedding-based and geometry-based verifiers across six datasets and five VPR backbones, and that its new precision-coverage-FAR evaluation triplet reveals failure modes that AUC-PR hides.

Load-bearing premise

The central claim rests on the assumption that 'Recall@1 after filtering' is a meaningful measure comparable to unfiltered recall@1; if recall is computed only over accepted queries, the reported 13.6% improvement is largely a by-product of the accept/reject trade-off rather than a genuine retrieval gain.

Editorial extensions

If this is right

  • VPR pipelines could add a zero-shot verification layer without retraining or threshold tuning, simplifying deployment in unknown environments.
  • Published verification results that rely only on AUC-PR may overstate false-positive protection; reporting FAR alongside precision and coverage would expose indiscriminate acceptance.
  • Because the auditor is independent of the retrieval model, it can be attached to any VPR backbone, as demonstrated across five architectures.
  • Rejecting a valid loop closure is recoverable, whereas accepting a false one corrupts the map, so the reported precision and coverage trade-off favors safety-critical use.
  • VLM false positives are almost always a subset of conventional verifier failures, suggesting the approach does not introduce a new class of catastrophic errors.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's Recall@1 comparison appears to be computed after filtering, meaning recall may be measured only over accepted queries; if so, the 13.6% improvement reflects selectivity rather than improved ranking, and a reader should verify the metric definition before interpreting it.
  • The same auditing idea could extend beyond a single top candidate to re-ranking several candidates, letting the VLM score multiple retrievals instead of making one binary decision.
  • The paper's failure analysis suggests that Qwen's false rejections often arise from ignoring prompt instructions about dynamic scene elements, so better prompt adherence or a lightweight fine-tuned VLM could recover much of the coverage loss at lower compute cost.
  • The precision-coverage-FAR triplet is a general evaluation recipe for any classifier with a reject option, not just VPR, and could become a standard reporting practice in retrieval verification.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes VPR Auditing, a post-retrieval verification layer in which a vision-language model (VLM) receives the query image and the top-1 candidate retrieved by an existing VPR pipeline and produces a binary accept/reject decision, replacing manually calibrated thresholds. The framework is evaluated on six benchmark datasets, five VPR methods, and four VLMs, and the authors propose precision/coverage/FAR as deployment-oriented evaluation metrics alongside risk-coverage curves. The central claim is that VLM auditing improves recall@1 by 13.6% on average while reducing FAR to 12%, keeping precision above 95%, and coverage above 75%. The verification idea is interesting and the experimental breadth is a strength, but the headline recall@1 claim is not supported as written because the metric used for the +VLM results is never defined and appears to be conditional on accepted queries.

Significance. If reframed properly, the paper makes a useful contribution to VPR verification. It provides a plausible demonstration that a VLM, operating independently of the retrieval descriptor space, can reject many false top-1 matches without environment-specific calibration. The precision/coverage/FAR triplet and the risk-coverage analysis are sensible complements to AUC-PR and are well motivated by the paper's own examples (e.g., methods with high AUC-PR but FAR near 1.0). The evaluation across six datasets, five backbones, and four VLM families is broad, and the open-weight Qwen experiments are reproducible from cited checkpoints. The paper does not engage in circular reasoning: the VLM decision is external to the retrieval function. However, the current headline claim is misleading and the abstract overstates the safety/availability numbers reported in Table III.

major comments (3)
  1. [Section V-A, Fig. 2, Table II] The metric 'Recall@1 after filtering' is never defined. Under the problem formulation in Section III-A, V:(q,r)->{accept,reject} operates on the fixed top-1 candidate and there is no re-ranking or fallback. With standard Recall@1 = (#correct top-1 matches)/(#queries), rejecting a false positive cannot increase the numerator; it can only keep recall unchanged or decrease it when a true positive is rejected. The +VLM entries in Table II exceed the baseline in nearly every row, so the reported values must be computed over the accepted queries only, i.e., (#correct accepted)/(#accepted queries), which is a conditional or precision-like measure and is not comparable to the baseline recall@1. No equation for this metric is given. The 13.6% average improvement in the abstract and Section V-A is therefore unsupported. The authors should either define the metric explicitly, report standard recall
  2. [Abstract and Table III] The abstract states that the method reduces 'false acceptance rates to 12%, maintaining precision above 95% and coverage above 75%.' These numbers are not consistent with Table III for the VLM rows. For example, Gemini has FAR of 0.59 on Pitts250k, 0.29 on StLucia, and 0.28 on AmsterTime; GPT-4.1 has precision 0.72 on AmsterTime; Gemini and GPT-4.1 have coverage below 0.75 on AmsterTime (0.55/0.54) and Nordland (0.68/0.64). At best, '12% FAR' is the minimum per-dataset value, not a general result. The abstract and Section V need to state the actual aggregate statistics or specify the datasets and conditions under which each claim holds.
  3. [Section IV-D] The paper states that in 'rare occurrences where the VLM refuses to answer or produces an invalid response, the retrieved match is retained and treated as accepted by the auditor.' For a safety-critical post-retrieval auditor, this is exactly the failure mode the framework aims to prevent: an invalid or refused VLM output results in the false positive being accepted. The paper provides no measurement of how often this occurs across the four VLMs and dozens of dataset/backbone combinations. The authors should report the frequency of invalid/refused outputs per VLM and, if the rate is non-negligible, revise the acceptance policy or qualify the safety claim accordingly.
minor comments (5)
  1. [Section V-A] 'PA score' is named in the list of baseline verification methods but is not introduced or described anywhere in Section IV or in the related work. Please define the acronym and the score at first use.
  2. [Table II] The pairs of Base/+VLM columns under each dataset are visually difficult to separate; adding explicit 'Base' and '+VLM' subheaders or using two panels would prevent the reader from mistaking the +VLM value for the baseline.
  3. [Fig. 4] The VLM curves are generated from the 'token-level logit probability of true' as a confidence score, which is not part of the proposed binary decision protocol described in Section IV-D. Please clarify whether this is a supplementary analysis and, if so, how the log-probability is obtained for the open-weight models.
  4. [Section IV-B / Section V-A] The geometric verification baselines (SIFT, SuperPoint) are listed without implementation details (e.g., matching strategy, inlier threshold). For reproducibility, provide the parameters used or cite a reference implementation.
  5. [Throughout] Minor writing issues: 'a query image q, f returns' in Section III-A is fragmented; 'Amstertime' is inconsistently capitalized in Section V-A; and the reference list contains future-dated preprints (e.g., [9], [26], [37]) that should be checked for final publication status.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline recall@1 gain is precision relabeled: binary accept/reject cannot raise standard recall@1, and Table II +VLM values match Table III precision.

  1. renaming known result [Abstract; Section V-A; Fig. 2 caption]
    "Results show that VLM-based auditing improves recall@1 by 13.6% on average as compared to state-of-the-art methods while reducing false acceptance rates to 12%, maintaining precision above 95% and coverage above 75%. Fig. 2. Recall@1 after filtering for every combination of dataset, VPR backbone, and verification method, including the unfiltered baseline (leftmost column) for reference."

    The problem formulation gives a fixed top-1 candidate r=f(q;D) and a binary verifier V:(q,r)->{accept,reject}. Under standard recall@1, rejecting false positives cannot increase recall: the numerator (#correct top-1 matches) is unchanged or reduced, and the denominator (#queries) is fixed. The only way Table II '+VLM' entries can exceed the baseline is to compute recall over accepted queries only, i.e., TP/(TP+FP) = P_tau, the paper's own precision metric in Eq. (1). Numerically, Table II '+VLM' values reproduce Table III precision values (e.g., AmsterTime GPT-4.1: 0.72 vs P=0.72; Qwen-7B: 0.74 vs 0.74; Qwen-32B: 0.62 vs 0.62). Thus the headline 'recall@1 improvement by 13.6%' is precision relabeled as recall after filtering, not a retrieval gain; the metric is equivalent to the paper's pr

full rationale

The derivation chain contains one load-bearing metric substitution that reduces the central headline claim to a previously reported quantity. The paper's method itself is largely self-contained: VLM accept/reject decisions come from external zero-shot models, no parameter is fitted to test labels, and the precision/coverage/FAR evaluation is defined by explicit equations (1)-(3). The self-citations to the authors' earlier VLM papers (refs. 33, 34) are contextual and not load-bearing for the auditing framework. However, the claim that auditing 'improves recall@1 by 13.6%' is not standard recall@1. With a fixed top-1 candidate and a binary accept/reject decision, rejecting false positives cannot increase recall under the standard definition; the only quantity that can increase is the proportion of correct matches among accepted queries, which is precisely the paper's precision metric. The numerical match between Table II '+VLM' entries and Table III precision values confirms that 'Recall@1 after filtering' is conditional recall over accepted queries, i.e., precision. This is a renaming of a known metric rather than an independent retrieval improvement. Because it affects the paper's headline result, the circularity score is 6; the underlying VLM auditing results may still be valid as precision/FAR/coverage findings, but the recall improvement claim is not supported by a well-defined standard recall metric.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The method introduces no physical or mathematical entities; it is an empirical pipeline. The VLM auditing layer is a procedure, not an entity. Main load-bearing assumptions are about VLM generalization and evaluation metric validity.

assumptions (4)
  • domain assumption A VLM can perform zero-shot instance-level same-place verification from two images without environment-specific training.
    Central premise of the method; Section III-B relies on VLM reasoning over semantic/structural consistency.
  • domain assumption The top-1 candidate returned by the VPR retriever is the only candidate considered and is either accepted or rejected; no re-ranking or fallback is used.
    Section IV-B fixes candidate as top-1; this makes standard recall@1 unable to increase, exposing the metric inconsistency.
  • ad hoc to paper Invalid or refused VLM outputs are treated as accepted matches.
    Section IV-D states the decision policy; it biases evaluation toward coverage and is not defended as safe.
  • domain assumption The six benchmark datasets are sufficiently representative to support claims of generalization across unseen environments.
    Section IV-A claims diverse deployment challenges; no closed-world guarantees.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Breaking D\'ej\`a Vu: Independent Auditing of Visual Place Recognition through Vision-Language Reasoning." pith.science (2026). https://pith.science/paper/T7G76TZN

@misc{pith2026260712818,
  author       = {Pith},
  title        = {Pith review of: Breaking D\'ej\`a Vu: Independent Auditing of Visual Place Recognition through Vision-Language Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T7G76TZN}},
  note         = {Machine review of arXiv:2607.12818}
}
read the original abstract

Visual place recognition (VPR) is a key enabler of accurate localization and long-term autonomous navigation in robotics applications, such as loop closure detection for simultaneous localisation and mapping (SLAM). However, real-world VPR deployment relies on selecting an image matching threshold that balances precision and recall. These thresholds are typically tuned using labeled validation data and fixed during deployment, making them unreliable under environmental changes where ground truth is unavailable. This is particularly problematic in safety-critical robotics, where accepting a false loop closure can corrupt the estimated trajectory and map. In this work, we introduce Visual Place Recognition Auditing, an independent post-retrieval verification framework that leverages Vision-Language Models (VLMs) to assess retrieved matches by reasoning jointly over query and candidate images. Unlike conventional verification methods, our approach performs instance-level verification without requiring architecture-specific confidence measures, dataset-dependent thresholds, or prior knowledge of the deployment environment. We evaluate our method on six benchmark datasets using five state-of-the-art VPR methods and four VLMs. Results show that VLM-based auditing improves recall@1 by 13.6% on average as compared to state-of-the-art methods while reducing false acceptance rates to 12%, maintaining precision above 95% and coverage above 75%.

Figures

Figures reproduced from arXiv: 2607.12818 by the authors.

Figure 1
Figure 1. Proposed VPR Auditing framework: A standard VPR pipeline [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Recall@1 after filtering for every combination of dataset, VPR backbone, and verification method, including the unfiltered baseline (leftmost [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Risk–coverage curves per dataset, averaged across VPR backbones. Lines show threshold-swept verification methods; markers show VLM operating [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Precision–recall curves for three representative datasets (AmsterTime, SanFran Small, Nordland), using EigenPlaces as the VPR backbone. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 10 linked inside Pith

  1. [1]

    Vpr-bench: An open-source visual place recognition evaluation framework with quantifiable viewpoint and appearance change,

    M. Zaffar, S. Garg, M. Milford, J. Kooij, D. Flynn, K. McDonald- Maier, and S. Ehsan, “Vpr-bench: An open-source visual place recognition evaluation framework with quantifiable viewpoint and appearance change,”International Journal of Computer Vision, vol. 129, no. 7, pp. 2136–2174, 2021

    2021

  2. [2]

    Visual place recognition: A survey from deep learning perspective,

    X. Zhang, L. Wang, and Y . Su, “Visual place recognition: A survey from deep learning perspective,”Pattern Recognition, vol. 113, p. 107760, 2021

    2021

  3. [3]

    Where is your place, visual place recognition?

    S. Garg, T. Fischer, and M. Milford, “Where is your place, visual place recognition?”arXiv preprint arXiv:2103.06443, 2021

    arXiv 2021

  4. [4]

    Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age,

    C. Cadena, L. Carlone, H. Carrillo, Y . Latif, D. Scaramuzza, J. Neira, I. Reid, and J. J. Leonard, “Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age,”IEEE Transactions on robotics, vol. 32, no. 6, pp. 1309–1332, 2017

    2017

  5. [5]

    Visual place recognition: A survey,

    S. Lowry, N. S ¨underhauf, P. Newman, J. J. Leonard, D. Cox, P. Corke, and M. J. Milford, “Visual place recognition: A survey,”ieee trans- actions on robotics, vol. 32, no. 1, pp. 1–19, 2015

    2015

  6. [6]

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

    M. Zaffar, L. Nan, and J. F. Kooij, “On the estimation of image- matching uncertainty in visual place recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2024, pp. 17 743–17 753

    2024

  7. [7]

    Beyond ann: Exploiting structural knowledge for efficient place recognition,

    S. Schubert, P. Neubert, and P. Protzel, “Beyond ann: Exploiting structural knowledge for efficient place recognition,” in2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 5861–5867

    2021

  8. [8]

    Adaptive thresholding for sequence-based place recognition,

    O. Vysotska, I. Bogoslavskyi, M. Hutter, and C. Stachniss, “Adaptive thresholding for sequence-based place recognition,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 2219–2225

    2025

Show all 37 references
  1. [9]

    Quantile transfer for reliable operating point selection in visual place recognition,

    D. M. Rajani, M. Milford, and T. Fischer, “Quantile transfer for reliable operating point selection in visual place recognition,”arXiv preprint arXiv:2602.04401, 2026

    2026 arXiv

  2. [10]

    What makes visual place recognition easy or hard?

    S. Schubert and P. Neubert, “What makes visual place recognition easy or hard?”arXiv preprint arXiv:2106.12671, 2021

    2021 arXiv

  3. [11]

    A survey on visual-based localization: On the benefit of heterogeneous data,

    N. Piasco, D. Sidib ´e, C. Demonceaux, and V . Gouet-Brunet, “A survey on visual-based localization: On the benefit of heterogeneous data,” Pattern Recognition, vol. 74, pp. 90–109, 2018

    2018

  4. [12]

    Unsupervised complementary- aware multi-process fusion for visual place recognition,

    S. Hausler, T. Fischer, and M. Milford, “Unsupervised complementary- aware multi-process fusion for visual place recognition,”arXiv preprint arXiv:2112.04701, 2021

    2021 arXiv

  5. [13]

    Bayesian triplet loss: Uncertainty quantification in image retrieval,

    F. Warburg, M. Jørgensen, J. Civera, and S. Hauberg, “Bayesian triplet loss: Uncertainty quantification in image retrieval,” inProceedings of the IEEE/CVF International conference on Computer Vision, 2021, pp. 12 158–12 168

    2021

  6. [14]

    Stun: Self-teaching uncertainty estimation for place recognition,

    K. Cai, C. X. Lu, and X. Huang, “Stun: Self-teaching uncertainty estimation for place recognition,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 6614–6621

    2022

  7. [15]

    Generalized con- trastive optimization of siamese networks for place recognition,

    M. Leyva-Vallina, N. Strisciuglio, and N. Petkov, “Generalized con- trastive optimization of siamese networks for place recognition,”arXiv preprint arXiv:2103.06638, 2021

    2021 arXiv

  8. [16]

    Learning with average precision: Training image retrieval with a listwise loss,

    J. Revaud, J. Almaz ´an, R. S. Rezende, and C. R. d. Souza, “Learning with average precision: Training image retrieval with a listwise loss,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 5107–5116

    2019

  9. [17]

    Geometrically mappable image features,

    J. Thoma, D. P. Paudel, A. Chhatkuli, and L. Van Gool, “Geometrically mappable image features,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 2062–2069, 2020

  10. [18]

    Gsv-cities: Toward appropriate supervised visual place recognition,

    A. Ali-Bey, B. Chaib-Draa, and P. Giguere, “Gsv-cities: Toward appropriate supervised visual place recognition,”Neurocomputing, vol. 513, pp. 194–203, 2022

    2022

  11. [19]

    Rethinking visual geo- localization for large-scale applications,

    G. Berton, C. Masone, and B. Caputo, “Rethinking visual geo- localization for large-scale applications,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 4878–4888

    2022

  12. [20]

    Transvpr: Transformer-based place recognition with multi-level attention aggre- gation,

    R. Wang, Y . Shen, W. Zuo, S. Zhou, and N. Zheng, “Transvpr: Transformer-based place recognition with multi-level attention aggre- gation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 13 648–13 657

    2022

  13. [21]

    Spatial pyramid- enhanced netvlad with weighted triplet loss for place recognition,

    J. Yu, C. Zhu, J. Zhang, Q. Huang, and D. Tao, “Spatial pyramid- enhanced netvlad with weighted triplet loss for place recognition,” IEEE transactions on neural networks and learning systems, vol. 31, no. 2, pp. 661–674, 2019

    2019

  14. [22]

    Vector of locally and adaptively ag- gregated descriptors for image feature representation,

    J. Zhang, Y . Cao, and Q. Wu, “Vector of locally and adaptively ag- gregated descriptors for image feature representation,”Pattern Recog- nition, vol. 116, p. 107952, 2021

    2021

  15. [23]

    Mixvpr: Feature mixing for visual place recognition,

    A. Ali-Bey, B. Chaib-Draa, and P. Giguere, “Mixvpr: Feature mixing for visual place recognition,” inProceedings of the IEEE/CVF winter conference on applications of computer vision, 2023, pp. 2998–3007

    2023

  16. [24]

    Netvlad: Cnn architecture for weakly supervised place recognition,

    R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and J. Sivic, “Netvlad: Cnn architecture for weakly supervised place recognition,” inPro- ceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 5297–5307

    2016

  17. [25]

    Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition,

    S. Hausler, S. Garg, M. Xu, M. Milford, and T. Fischer, “Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 14 141–14 152

    2021

  18. [26]

    Through the lens of doubt: Robust and efficient uncertainty estimation for visual place recognition,

    E. Miller, M. Milford, M. B. Hafez, S. Ramchurn, and S. Ehsan, “Through the lens of doubt: Robust and efficient uncertainty estimation for visual place recognition,”IEEE Robotics and Automation Letters, 2026

    2026

  19. [27]

    Distinctive image features from scale-invariant key- points,

    D. G. Lowe, “Distinctive image features from scale-invariant key- points,”International journal of computer vision, vol. 60, no. 2, pp. 91–110, 2004

    2004

  20. [28]

    Large-scale image retrieval with attentive deep local features,

    H. Noh, A. Araujo, J. Sim, T. Weyand, and B. Han, “Large-scale image retrieval with attentive deep local features,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 3456–3465

    2017

  21. [29]

    Superpoint: Self- supervised interest point detection and description,

    D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superpoint: Self- supervised interest point detection and description,” inProceedings of the IEEE conference on computer vision and pattern recognition workshops, 2018, pp. 224–236

    2018

  22. [30]

    Large- scale location recognition and the geometric burstiness problem,

    T. Sattler, M. Havlena, K. Schindler, and M. Pollefeys, “Large- scale location recognition and the geometric burstiness problem,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 1582–1590

    2016

  23. [31]

    Unsupervised selection of optimal operating parameters for visual place recognition algorithms using gaussian mixture models,

    J. Mount, M. Xu, L. Dawes, and M. Milford, “Unsupervised selection of optimal operating parameters for visual place recognition algorithms using gaussian mixture models,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 343–350, 2020

    2020

  24. [32]

    Im2city: image geo-localization via multi- modal learning,

    M. Wu and Q. Huang, “Im2city: image geo-localization via multi- modal learning,” inProceedings of the 5th ACM SIGSPATIAL Inter- national Workshop on AI for Geographic Knowledge Discovery, 2022, pp. 50–61

    2022

  25. [33]

    Vlm-guided visual place recognition for planet-scale geo- localization,

    S. Waheed, N. M. An, M. Milford, S. D. Ramchurn, and S. Ehsan, “Vlm-guided visual place recognition for planet-scale geo- localization,”arXiv preprint arXiv:2507.17455, 2025

    2025 arXiv

  26. [34]

    Image-based geo-localization for robotics: Are black-box vision- language models there yet?

    S. Waheed, B. Ferrarini, M. Milford, S. D. Ramchurn, and S. Ehsan, “Image-based geo-localization for robotics: Are black-box vision- language models there yet?”arXiv preprint arXiv:2501.16947, 2025

    2025 arXiv

  27. [35]

    Do as i can, not as i say: Grounding language in robotic affordances,

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakrishnan, K. Hausman,et al., “Do as i can, not as i say: Grounding language in robotic affordances,”arXiv preprint arXiv:2204.01691, 2022

    2022 arXiv

  28. [36]

    Language to rewards for robotic skill synthesis,

    W. Yu, N. Gileadi, C. Fu, S. Kirmani, K.-H. Lee, M. G. Arenas, H.- T. L. Chiang, T. Erez, L. Hasenclever, J. Humplik,et al., “Language to rewards for robotic skill synthesis,”arXiv preprint arXiv:2306.08647, 2023

    2023 arXiv

  29. [37]

    Localnav: Distilling frontier vlms and embodied rl for on-device object goal navigation,

    N. Baumann, L. Boyle, P. Deng, E. Ghignone, B. Sun, M. Pollefeys, L. Benini, and M. Magno, “Localnav: Distilling frontier vlms and embodied rl for on-device object goal navigation,”arXiv preprint arXiv:2606.27871, 2026

    2026 arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.