Pith. sign in

REVIEW 3 major objections 32 references

A second-stage geometric-visual network cuts ship localization error by over 60% on average across standard VPR backbones.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 12:34 UTC pith:3PE5EKPN

load-bearing objection Solid systems paper: real ship multi-floor data + a drop-in geometric second stage that cuts mean localization error ~60% across several VPR backbones; gains rest on always-available same-trajectory priors that are not ablated away. the 3 major comments →

arxiv 2606.24234 v2 pith:3PE5EKPN submitted 2026-06-23 cs.CV cs.RO

From Open Waters to Enclosed Cabins: ProteusVPR for Cross-Scene Visual Place Recognition in Maritime Perception and Cabin Inspection

Zexi Chen , Zitai Huang , Qiwen Gu , Zhiqi Li , Shengli Dong , Chenlei Wang , Junqiao Zhao , Hongdong Wang
show 1 more author
Bing Han
This is my paper
classification cs.CV cs.RO
keywords Visual Place Recognitionmaritime roboticscross-scene localizationgeometric refinementlocal affine coordinatesship cabin inspectionXHZ datasetretrieval re-ranking
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Robots that inspect ships must switch between open decks (sparse textures, harsh lighting) and enclosed cabins (repetitive corridors and multi-floor layouts that look almost identical). Standard visual place recognition systems built for cities or buildings lose both recall and precision under that domain shift. ProteusVPR leaves any existing VPR model as the first-stage retriever and adds a second-stage network that fuses the retrieved image with the two preceding frames of the query sequence. The network builds geometric descriptors among the three points, constructs a local affine coordinate frame centered on them, encodes camera azimuth, and regresses a refined position. On the new XHZ ship dataset the refinement consistently lowers mean Euclidean error by more than 60 percent across several backbones while remaining a drop-in upgrade. The work therefore claims that precise cabin-scale localization is achievable without redesigning the retrieval backbone, provided short temporal context and simple geometric structure are available.

Core claim

On the XHZ multi-deck ship dataset, ProteusVPR’s geometric-visual estimation stage, which fuses a first-stage VPR retrieval with two temporally preceding frames under a local affine coordinate system, consistently improves localization accuracy for multiple standard VPR backbones and reduces mean Euclidean localization error by over 60 percent on average (reported as 61.22 percent across models and layers).

What carries the argument

ProteusVPR second-stage geometric-visual estimation network: three reference points (retrieved image plus two preceding frames) define a centroid and a local orthonormal affine frame; geometric pair descriptors, visual attention features and azimuth encodings are fused to regress a local displacement that is then transformed back to global coordinates.

Load-bearing premise

The correction works only when the two preceding frames of continuous video almost always lie on the true deck, so their coordinates can pull the three-point centroid toward the right place even if the first-stage retrieval is wrong.

What would settle it

Remove or scramble the two temporally preceding frames (or force them onto wrong decks) while keeping Stage-1 retrieval fixed; if mean error reduction collapses below the reported ~60 percent, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. The paper proposes ProteusVPR, a two-stage retrieval-refinement framework for maritime cross-scene visual place recognition between open decks and enclosed multi-floor cabins. Stage 1 uses any standard VPR backbone for nearest-neighbor retrieval; Stage 2 is a geometric-visual estimation network that fuses the retrieved image with two temporally preceding query frames, using DINOv2 features, inter/cross attention, geometric pair descriptors, a local affine coordinate system built from the three reference points, and camera azimuth encoding to regress a local displacement that is mapped back to global coordinates. The authors introduce the XHZ ship-borne panoramic dataset with multi-deck structure, chronological splits, and strict query-database separation. On XHZ, Stage 2 consistently reduces mean Euclidean localization error across six VPR backbones (pre-trained and fine-tuned), with a reported average reduction of about 61%, and ablations show that geometric descriptors and the local affine frame are the main contributors.

Significance. If the result holds under realistic operating conditions, the work is a useful systems contribution for autonomous ship inspection: a drop-in refinement module that upgrades existing VPR pipelines for repetitive cabin scenes, plus a new multi-floor maritime benchmark with careful geo-tagging and evaluation protocol. Strengths include multi-backbone evaluation (BoQ, CosPlace, EigenPlaces, MixVPR, SALAD, MutualVPR), both mean and 90% quantile errors per deck, qualitative cross-layer correction examples, and component ablations. The practical value is real for continuous onboard inspection trajectories; the main open question is how much of the headline gain depends on always-available same-trajectory priors rather than pure visual-geometric refinement of a single retrieval.

major comments (3)
  1. The central claim of ~61% average mean-error reduction (Abstract; §5.5) is measured only under the Stage-2 protocol of §3.2 and §5.1.3: every sample is a same-layer, same-yaw, temporally adjacent 4-frame tuple, so I_r2 and I_r3 almost always lie on the correct deck and pull the three-point centroid (Eq. 8) and local affine basis (Eqs. 9–12) toward the true place even when Stage-1 retrieval is cross-layer (explicitly used in §5.3.2 and Fig. 5). No experiment reports Stage-2 error when the two preceding frames are dropped, replaced by random database images, or themselves wrong-layer. Without that ablation the headline average is conditional on continuous trajectory context that real cabin inspection may not guarantee; Limitations §7.i notes dependence on retrieval quality but does not quantify this temporal-context assumption that drives large mean reductions (e.g., MutualVPR tra L3 1.519
  2. Generalization claims for “cross-scene maritime environments” and “open waters to enclosed cabins” (title, Abstract, §1, §6) rest almost entirely on XHZ from a single vessel. Open-deck sparse-texture/illumination challenges are described but not isolated in the reported tables; results are per deck (L0–L3) on one ship under limited lighting/operational variation (§7.ii). At least one external indoor/outdoor VPR benchmark, or a held-out deck/transition split that does not share the same continuous trajectories, is needed to support the broader positioning beyond cabin-scale refinement on XHZ.
  3. Stage-2 training and evaluation use same-layer, same-angle, temporally adjacent tuples (§5.1.3), while the first-stage database mixes yaws and (for test) training+validation images. This is a reasonable continuous-inspection protocol, but it is not the same as refining an arbitrary single-image retrieval under viewpoint change. The paper should either (a) report Stage-2 performance when the two priors are not same-yaw / not adjacent, or (b) clearly scope the claim to sequence-aware refinement rather than a general second-stage upgrade for any VPR retrieval.

Circularity Check

1 steps flagged

Empirical systems paper with one non-load-bearing self-citation to MutualVPR (author overlap); central multi-backbone error reductions are measured on held-out geo-tagged splits, not forced by definition or fit.

specific steps
  1. self citation load bearing [§2.5 Summary and Positioning; also §3 intro and experiments using MutualVPR]
    "MutualVPR [18], through its mutual learning framework that adaptively resolves supervision inconsistency, provides a foundation for this study. ... Building upon the first stage of MutualVPR, we introduce a two-stage estimation network..."

    MutualVPR is prior work by overlapping authors (Gu, Zhao). It is invoked as the conceptual foundation and as one of the primary backbones whose Stage-1 errors are refined. However the central claim (consistent improvement across multiple independent VPR models, average 61% mean-error drop) does not reduce to this citation: identical Stage-2 gains appear for BoQ/CosPlace/SALAD/etc. with both pre-trained and fine-tuned weights, so the self-citation is present but not load-bearing for the reported result.

full rationale

ProteusVPR is a two-stage neural refinement (standard VPR retrieval + geometric-visual regression network using DINOv2 features, pair-wise geometric descriptors, local affine frame from three reference points, and azimuth MLP) trained with ordinary absolute + geometric-consistency losses and evaluated by Euclidean error on chronologically held-out XHZ queries against geo-tagged ground truth. The headline >60% mean-error reduction (explicitly averaged 61.22% over models/layers from Tables 1–2) is an empirical statistic, not a quantity recovered by construction from fitted parameters or self-referential definitions. Stage-2 inputs (retrieved image + two preceding frames) and the local-affine transform (Eqs. 8–18) are architectural choices whose benefit is measured, not assumed to equal the target. MutualVPR (Gu et al., overlapping co-authors) is cited as related work and used as one backbone among six (BoQ, CosPlace, EigenPlaces, MixVPR, SALAD, MutualVPR), with consistent Stage-2 gains reported for all; the self-citation therefore supplies neither uniqueness nor the sole support for the multi-backbone claim. No uniqueness theorem is imported, no ansatz is smuggled via prior work, and no known empirical pattern is merely renamed. The only minor circularity-adjacent element is the non-load-bearing self-citation; the derivation chain itself is ordinary supervised evaluation and scores low.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The central empirical claim rests on standard VPR retrieval plus a learned regressor whose training and evaluation depend on several hand-chosen loss/geometry hyperparameters, the availability of two prior frames with reliable geo-tags, artificial inter-deck coordinate offsets used in evaluation, and a single-vessel dataset. No new physical entities are postulated; the invented pieces are the method modules and the dataset itself.

free parameters (6)
  • λ_abs, λ_geo (total loss weights) = 1.0, 0.2
    Set to 1.0 and 0.2 for Eq. (20); balance absolute position vs geometric consistency and directly affect Stage-2 training.
  • λ_cons, ρ (geometric consistency hinge) = 0.3, 0.1
    λ_cons=0.3 and ρ=0.1 in Eq. (22) control how strongly predictions are pulled near the reference triangle; chosen by authors, not derived.
  • similarity temperature τ = 10
    τ=10 in Eq. (3) softens visual similarity weights used for the coarse centroid p_init.
  • Stage-2 training hyperparameters (lr, batch, epochs, warmup) = lr=2e-5, bs=16, epochs=300
    lr=2e-5, batch 16, 300 epochs, cosine warmup 30, weight decay 1e-4; control convergence of the refinement network on XHZ.
  • Inter-deck Y-axis offsets for evaluation coordinates = -75,-25,+25,+75 m
    Manual offsets L0:-75, L1:-25, L2:+25, L3:+75 m (§4.4.3) separate floors in 2D so cross-floor retrieval errors become large Euclidean distances; evaluation metrics depend on this design choice.
  • Positive-pair distance threshold for Recall = 0.5 m
    R@1/R@5 use ≤0.5 m Euclidean threshold (§5.1.1); changes which retrievals count as correct.
axioms (5)
  • domain assumption Two immediately preceding frames of the query video are available at inference and share the query’s continuous trajectory (usually same deck).
    Stage-2 inputs I_r2, I_r3 are defined this way in §3.2; case study §5.3.2 relies on them to correct wrong-layer retrievals.
  • domain assumption Database images carry accurate 2D geo-tags (and FOV-center azimuths) derived from ship CAD with 80:1 scale mapping.
    Ground truth and reference coordinates p_ri come from CAD annotation (§4.4); Limitations §7.iii notes robustness to noisy odometry/SLAM is untested.
  • domain assumption A frozen DINOv2 ViT-B/14 encoder plus learned attention/MLP heads is a sufficient visual front-end for relative displacement regression.
    §3.2.1 freezes DINOv2 and only trains fusion/geometry/azimuth/regression modules.
  • standard math Standard cosine nearest-neighbor retrieval on global descriptors is an acceptable Stage-1 interface for any VPR backbone.
    Eq. (1) and §3.1 formalize the usual VPR retrieval pipeline used as a black box.
  • ad hoc to paper Chronological 2/3–1/6–1/6 split with yaw=0 queries and other yaws as database prevents leakage and is a fair VPR protocol.
    §4.6 defines the XHZ-specific split and query–database construction used for all reported numbers.
invented entities (2)
  • ProteusVPR two-stage geometric-visual estimation network no independent evidence
    purpose: Refine any first-stage VPR retrieval into a continuous 2D pose estimate using three references, local affine coordinates, geometric descriptors, and azimuth.
    Core method of the paper (§3.2); evaluated only on XHZ in this work, no external independent validation yet.
  • XHZ (Xinhongzhuan Visual Dataset) no independent evidence
    purpose: Provide multi-floor cabin/deck panoramic data with strict query–database separation for maritime VPR evaluation.
    New dataset introduced in §4; collected on one operational vessel; public release status not confirmed in text.

pith-pipeline@v1.1.0-grok45 · 25750 in / 4121 out tokens · 38982 ms · 2026-07-12T12:34:29.243503+00:00 · methodology

0 comments
read the original abstract

Autonomous robotic inspection in maritime environments presents unique challenges for Visual Place Recognition (VPR) due to cross-scene perceptual shifts. Robots navigating ship-borne environments must transition between visually distinct domains: open decks with sparse textures and severe illumination changes, and enclosed cabins with repetitive structures and high visual ambiguity. Existing VPR methods, designed primarily for urban or indoor scenes, fail to generalize reliably across these starkly different scenarios. To address this, we propose ProteusVPR, a two-stage retrieval-refinement framework. The first stage employs any standard VPR model for initial image retrieval. The second stage introduces a geometric-visual estimation network that fuses the retrieved image with two temporally preceding frames, incorporating geometric descriptors, a local affine coordinate system, and camera azimuth encoding to achieve precise localization. To support this task, we introduce the XHZ dataset, an 8K-panoramic ship-borne dataset collected from an operational vessel, featuring multi-floor cabin structures, deck transition zones, and strict query-database separation for rigorous evaluation. Extensive experiments on the XHZ dataset demonstrate that ProteusVPR consistently improves the localization accuracy across multiple VPR backbones, reducing mean localization error by over 60\% on average and that ProteusVPR offers an effective and robust solution for precise visual localization in challenging, cross-scene maritime environments.

Figures

Figures reproduced from arXiv: 2606.24234 by Bing Han, Chenlei Wang, Hongdong Wang, Junqiao Zhao, Qiwen Gu, Shengli Dong, Zexi Chen, Zhiqi Li, Zitai Huang.

Figure 1
Figure 1. Figure 1: Introduction of ProteusVPR. Left: Main challenges in maritime cross-scene VPR; Middle: architecture of the proposed retrieval refinement framework; Right: quantitative comparisons on the XHZ dataset. feature representations are still designed for generic large￾scale place retrieval rather than repetitive indoor localiza￾tion. We therefore ask the following question: Can a general refinement framework be bu… view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of ProteusVPR. The first stage VPR model extracts a global descriptor for retrieval. The second stage fuses visual attention, geometric descriptors, a local affine coordinate system, and azimuth information to regress the query location. adaptively resolves supervision inconsistency, provides a foundation for this study. However, none of the above methods are specifically designed for … view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the XHZ dataset. The upper part shows the acquisition platform, device, and panorama examples; the lower part shows the annotation tool, exported images, annotations (position, timestamp, IMU), and folder structure. device generates equirectangular images with a 360◦ hori￾zontal field of view and 180◦ vertical field of view through dual-fisheye lens stitching, providing high-resolution source m… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative Demonstration on the correction effect of the estimation module on intra-layer retrieval errors (retrieval module: fine-tuned MutualVPR). The connecting lines illustrate the correction relationship between a mis-retrieved point and its corresponding predicted point within the same layer. The two zoom-in panels below correspond to the boxed regions on L 1 (left) and L 2 (right), respectively. • … view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Demonstration on the correction effect of the pose estimation on cross-layer retrieval errors (retrieval module: fine-tuned MutualVPR). In the case, we are locating query images in ”Layer 0”, The VPR might reach out to a mis-retrieved point (on a wrong layer) to its corresponding predicted point. Colors encode the number of layers crossed: blue for crossing one layer, purple for two layers, and… view at source ↗

Review history (2 revisions) →

discussion (0)

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

Reference graph

Works this paper leans on

32 extracted references · 4 linked inside Pith

  1. [1]

    Place recognition meet multiple modalities: a comprehensive review, current challenges and future development

    Li, Z., Shang, T., Xu, P., Deng, Z., 2025. Place recognition meet multiple modalities: a comprehensive review, current challenges and future development. Artificial Intelligence Review 58, 363

    2025

  2. [2]

    Deep visual place recognition for waterborne domains, in: 2022 IEEE International Conference on Image Processing (ICIP), IEEE

    Thomas, L., Edwards, M., Capsey, A., Rahat, A., Roach, M., 2022. Deep visual place recognition for waterborne domains, in: 2022 IEEE International Conference on Image Processing (ICIP), IEEE. pp. 3546–3550

    2022

  3. [3]

    Netvlad:Cnnarchitectureforweaklysupervisedplacerecognition,in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J., 2016. Netvlad:Cnnarchitectureforweaklysupervisedplacerecognition,in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 5297–5307

    2016

  4. [4]

    Rethinking visual geo-localization for large-scale applications, in: Proceedings of the IEEE/CVFConferenceonComputerVisionandPatternRecognition, pp

    Berton, G., Masone, C., Caputo, B., 2022. Rethinking visual geo-localization for large-scale applications, in: Proceedings of the IEEE/CVFConferenceonComputerVisionandPatternRecognition, pp. 4878–4888

    2022

  5. [5]

    Eigenplaces: Training viewpoint robust models for visual place recognition, in: ProceedingsoftheIEEE/CVFInternationalConferenceonComputer Vision, pp

    Berton,G.,Trivigno,G.,Caputo,B.,Masone,C.,2023. Eigenplaces: Training viewpoint robust models for visual place recognition, in: ProceedingsoftheIEEE/CVFInternationalConferenceonComputer Vision, pp. 11080–11090

    2023

  6. [6]

    Mixvpr: Feature mixingforvisualplacerecognition,in:ProceedingsoftheIEEE/CVF winterconferenceonapplicationsofcomputervision,pp.2998–3007

    Ali-Bey, A., Chaib-Draa, B., Giguere, P., 2023. Mixvpr: Feature mixingforvisualplacerecognition,in:ProceedingsoftheIEEE/CVF winterconferenceonapplicationsofcomputervision,pp.2998–3007

    2023

  7. [7]

    Optimal transport aggregation for visual place recognition, in: Proceedings of the ieee/cvf conference on computer vision and pattern recognition, pp

    Izquierdo, S., Civera, J., 2024. Optimal transport aggregation for visual place recognition, in: Proceedings of the ieee/cvf conference on computer vision and pattern recognition, pp. 17658–17668

    2024

  8. [8]

    Patch- netvlad: Multi-scale fusion of locally-global descriptors for place recognition, in: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pp

    Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T., 2021. Patch- netvlad: Multi-scale fusion of locally-global descriptors for place recognition, in: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pp. 14141–14152

    2021

  9. [9]

    Transvpr: Transformer-basedplacerecognitionwithmulti-levelattentionaggre- gation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Wang, R., Shen, Y., Zuo, W., Zhou, S., Zheng, N., 2022. Transvpr: Transformer-basedplacerecognitionwithmulti-levelattentionaggre- gation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13648–13657

    2022

  10. [10]

    Cricavpr: Cross-image correlation-aware representation learning for visualplacerecognition,in:ProceedingsoftheIEEE/CVFconference on computer vision and pattern recognition, pp

    Lu, F., Lan, X., Zhang, L., Jiang, D., Wang, Y., Yuan, C., 2024. Cricavpr: Cross-image correlation-aware representation learning for visualplacerecognition,in:ProceedingsoftheIEEE/CVFconference on computer vision and pattern recognition, pp. 16772–16782

    2024

  11. [11]

    Context-based visual-language place recognition

    Woo, S., Kim, S.W., 2024. Context-based visual-language place recognition. arXiv preprint arXiv:2410.19341

    Pith/arXiv arXiv 2024

  12. [12]

    Revisit anything: Visual place recognition via image segment re- trieval, in: European Conference on Computer Vision, Springer

    Garg, K., Puligilla, S.S., Kolathaya, S., Krishna, M., Garg, S., 2024. Revisit anything: Visual place recognition via image segment re- trieval, in: European Conference on Computer Vision, Springer. pp. 326–343. First Author et al.:Preprint submitted to ElsevierPage 15 of 16 Short Title of the Article

    2024

  13. [13]

    Structvpr++: distill structural and semantic knowledge with weight- ing samples for visual place recognition

    Shen, Y., Zhou, S., Fu, J., Wang, R., Chen, S., Zheng, N., 2025. Structvpr++: distill structural and semantic knowledge with weight- ing samples for visual place recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence

    2025

  14. [14]

    Oquab,M.,Darcet,T.,Moutakanni,T.,Vo,H.,Szafraniec,M.,Khali- dov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.,

  15. [15]

    arXiv preprint arXiv:2304.07193

    Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193

    Pith/arXiv arXiv

  16. [16]

    Anyloc: Towards universal visual place recognition

    Keetha,N.,Mishra,A.,Karhade,J.,Jatavallabhula,K.M.,Scherer,S., Krishna, M., Garg, S., 2023. Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters 9, 1286–1293

    2023

  17. [17]

    Megaloc: One retrieval to place them all, in: Proceedings of the Computer Vision and Pattern Recognition Conference, pp

    Berton, G., Masone, C., 2025. Megaloc: One retrieval to place them all, in: Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 2861–2867

    2025

  18. [18]

    Boq: A place is worth a bag of learnable queries, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Ali-Bey, A., Chaib-Draa, B., Giguere, P., 2024. Boq: A place is worth a bag of learnable queries, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 17794– 17803

    2024

  19. [19]

    Mutualvpr: A mutual learning framework for resolving supervision inconsistencies via adaptive clustering

    Gu, Q., Wang, X., Zhao, J., Tao, S., Feng, T., Wang, Z., Chen, G., 2026. Mutualvpr: A mutual learning framework for resolving supervision inconsistencies via adaptive clustering. Advances in Neural Information Processing Systems 38, 2899–2922

    2026

  20. [20]

    R2former:Unifiedretrievalandrerankingtransformerforplacerecog- nition, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Zhu, S., Yang, L., Chen, C., Shah, M., Shen, X., Wang, H., 2023. R2former:Unifiedretrievalandrerankingtransformerforplacerecog- nition, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 19370–19380

    2023

  21. [21]

    Pair-vpr: Place-aware pre-training and contrastive pair classification for visual place recognition with vision transformers

    Hausler, S., Moghadam, P., 2025. Pair-vpr: Place-aware pre-training and contrastive pair classification for visual place recognition with vision transformers. IEEE Robotics and Automation Letters

    2025

  22. [22]

    Progeo: Generatingpromptsthroughimage-textcontrastivelearningforvisual geo-localization, in: International Conference on Artificial Neural Networks, Springer

    Hu, J., Mao, C., Tan, C., Li, H., Liu, H., Zheng, M., 2024. Progeo: Generatingpromptsthroughimage-textcontrastivelearningforvisual geo-localization, in: International Conference on Artificial Neural Networks, Springer. pp. 448–462

    2024

  23. [23]

    Towards seamless adaptation of pre-trained models for visual place recognition

    Lu, F., Zhang, L., Lan, X., Dong, S., Wang, Y., Yuan, C., 2024. Towards seamless adaptation of pre-trained models for visual place recognition. arXiv preprint arXiv:2402.14505

    Pith/arXiv arXiv 2024

  24. [24]

    Effovpr:Effectivefoundationmodelutilizationforvisualplacerecog- nition

    Tzachor, I., Lerner, B., Levy, M., Green, M., Shalev, T.B., Habib, G., Samuel, D., Zailer, N.K., Shimshi, O., Darshan, N., et al., 2024. Effovpr:Effectivefoundationmodelutilizationforvisualplacerecog- nition. arXiv preprint arXiv:2405.18065

    arXiv 2024

  25. [25]

    Zuo, W., Liu, L., Li, Y., Shen, Y., Xiang, F., Xin, J., Zheng, N.,

  26. [26]

    Pattern Recognition 166, 111673

    Prgs:patch-to-regiongraphsearchforvisualplacerecognition. Pattern Recognition 166, 111673

  27. [27]

    Unipr-3d: Towards universal visual place recogni- tion with visual geometry grounded transformer

    Deng, T., Chen, X., Li, Z., Shen, H., Wang, D., Civera, J., Wang, H., 2025. Unipr-3d: Towards universal visual place recogni- tion with visual geometry grounded transformer. arXiv preprint arXiv:2512.21078

    Pith/arXiv arXiv 2025

  28. [28]

    Casevpr: Correlation-aware sequential embedding for sequence-to- frame visual place recognition

    Li, H., Peng, G., Zhang, J., Wen, M., Ma, Y., Wang, D., 2025. Casevpr: Correlation-aware sequential embedding for sequence-to- frame visual place recognition. IEEE Robotics and Automation Letters

    2025

  29. [29]

    Imvpr: Implicitbev-enhancedmulti-viewaggregationforvisualplacerecog- nition, in: 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE

    Cao, X., Zhang, C., Liu, Z., Zhong, X., Fang, H., 2025. Imvpr: Implicitbev-enhancedmulti-viewaggregationforvisualplacerecog- nition, in: 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE. pp. 9669–9676

    2025

  30. [30]

    Vlad- buff:Burst-awarefastfeatureaggregationforvisualplacerecognition, in:EuropeanConferenceonComputerVision,Springer.pp.447–466

    Khaliq, A., Xu, M., Hausler, S., Milford, M., Garg, S., 2024. Vlad- buff:Burst-awarefastfeatureaggregationforvisualplacerecognition, in:EuropeanConferenceonComputerVision,Springer.pp.447–466

    2024

  31. [31]

    Visual place recognition for coastal scenes with semantic and sequential constraints

    Zhou, R., Wang, Y., Gao, Y., Zhao, X., Xu, X., 2025. Visual place recognition for coastal scenes with semantic and sequential constraints. Measurement 256, 118071

    2025

  32. [32]

    Mrs-vpr: a multi-resolution sampling based global visual place recognition method, in: 2019 International conference on robotics and automation (ICRA), IEEE

    Yin, P., Srivatsan, R.A., Chen, Y., Li, X., Zhang, H., Xu, L., Li, L., Jia, Z., Ji, J., He, Y., 2019. Mrs-vpr: a multi-resolution sampling based global visual place recognition method, in: 2019 International conference on robotics and automation (ICRA), IEEE. pp. 7137– 7142. First Author et al.:Preprint submitted to ElsevierPage 16 of 16

    2019