Pith. sign in

REVIEW 4 major objections 6 minor 45 references

DSFormer: A Dual-Scale Cross-Learning Transformer for Visual Place Recognition

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper proposes DSFormer, a dual-scale cross-attention Transformer paired with an HDBSCAN block-clustering repartition of the SF-XL training set, and claims state-of-the-art visual place recognition with 512-dimensional global…

desk verdict Solid VPR paper with a clever dual-scale transformer and a data-efficient SF-XL repartitioning, but the SOTA claim rests on self-reimplemented baselines and a thin 0.5-point margin—worth reviewing, not yet fully convincing. read the letter →

arxiv 2507.18444 v1 pith:LK3NOQVG submitted 2025-07-24 cs.CV cs.RO

Haiyang Jiang , Songhao Piao , Chao Gao , Lei Yu , Liguo Chen This is my paper
classification cs.CVcs.RO
keywords visualplacerecognitionglobaldescriptorstransformercross-attentiondual-scalefeaturesblockclusteringHDBSCANSF-XLdatasetDINOv2
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

The paper tries to establish that a single-stage global retrieval pipeline can match or beat two-stage reranking systems for visual place recognition. It proposes DSFormer, a Transformer module that lets features from the last two CNN layers exchange information through self-attention and shared cross-attention, and pairs it with a block-clustering repartition of the SF-XL training set based on HDBSCAN. On five benchmark datasets and on SF-XL test suites, the authors report state-of-the-art Recall@1 with 512-dimensional descriptors, surpassing reranking methods such as DELG, Patch-NetVLAD, TransVPR, and R2Former. A careful reader would care because this would make accurate place recognition cheaper in memory and latency while also cutting the required training data by roughly 30%.

What carries the argument

The central object is DSFormer, a dual-scale cross-learning Transformer: two feature patch sequences, one from each of the final two backbone layers, are processed by self-attention encoders with Improving Relative Position Encoding and a shared cross-attention encoder that lets each scale attend to the other before GeM pooling. The supporting mechanism is the block-clustering partition of SF-XL, which groups geotagged training locations by block width, HDBSCAN density clusters, KNN radius filtering, and SVD-defined focal points so that each class has coherent spatial proximity and viewpoint coverage. Together these carry the claim because the cross-attention gives the descriptor its discriminative power and the clustering gives the model better classes from fewer images.

What would settle it

Run DELG, Patch-NetVLAD, TransVPR, and R2Former with their official released code and checkpoints on the five benchmark datasets and compare Recall@1 with DSFormer's reported numbers; if DSFormer no longer surpasses the reranking methods, the paper's central state-of-the-art claim would not hold. A cheaper check is whether the block-clustering gain persists when a GeM model is trained on a random subsample of SF-XL of the same 4.2 million size; if random subsampling matches block clustering, the gain would come from data volume rather than clustering quality.

Watch

Extended reading notes

Core claim

DSFormer takes feature maps from the final two layers of a backbone (ResNet-50 or DINOv2), flattens them into patch embeddings, and passes them through transformer blocks that combine multi-head self-attention with relative position encoding inside each scale and a shared cross-attention module across scales. The two streams are then pooled by GeM into one compact 512-dimensional global descriptor. On the paper's reproductions, this descriptor achieves the best average Recall@1 among compared ResNet-50 retrieval methods and beats two-stage reranking methods on most benchmarks, with the largest gains on appearance-change-heavy datasets like Tokyo24/7 and Nordland. The companion block-clustering strategy uses UTM coordinates, HDBSCAN density clustering, KNN radius pruning, and SVD-derived focal directions to repartition SF-XL into balanced classes, shrinking the training set from about 5.6-6.2 million images to 4.2 million while improving accuracy. With a DINOv2 backbone, DSFormer reports the highest average Recall@1 among DINOv2-based retrieval methods, exceeding SALAD by 3.0 percentage points on the five benchmarks and by 10.7 points on the SF-XL test suites.

Load-bearing premise

The load-bearing premise is that the authors' own re-implementations of the compared methods are faithful, well-tuned versions of the published pipelines, since every comparison in Tables I and III comes from those reproductions rather than official results.

Editorial extensions

If this is right

  • As a global retrieval solution, DSFormer with ResNet-50 beats the compared reranking methods DELG, Patch-NetVLAD, TransVPR, and R2Former on average Recall@1 across MSLS, Pittsburgh30k, Tokyo24/7, and Nordland, without any reranking step.
  • DSFormer with a DINOv2 backbone outperforms DINOv2-based retrieval methods such as SALAD and BoQ on the same benchmarks, with the largest gain on the large-scale SF-XL test sets.
  • The block-clustering partition reduces SF-XL training data by roughly 25-32% relative to the CosPlace and EigenPlaces partitions while improving or matching accuracy, implying more efficient data utilization.
  • Because only a 512-dimensional global descriptor is needed, DSFormer removes the local-feature storage and geometric verification of two-stage methods, reporting memory 132 times smaller than R2Former and no reranking latency.
  • Ablations removing the self-encoder, cross-encoder, IRPE, or block clustering all lower Recall@1 on at least some benchmarks, attributing the gain to the combination rather than any single component.

Reading between the lines

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

  • The same HDBSCAN block-clustering recipe could likely be applied to other dense geotagged training collections, not just SF-XL, wherever grid partitioning creates class imbalance.
  • The especially large gains on Tokyo24/7 and Nordland suggest dual-scale cross-attention mainly helps when appearance changes break low-level feature matches, a hypothesis that could be tested by stress-testing on synthetic appearance perturbations.
  • Because all baseline numbers come from the authors' own re-implementations, the specific margins should be re-checked with the baselines' official released checkpoints; if the margins survive, the ~30% data cut implies current dense VPR training sets contain substantial redundancy.
  • One testable next step would be training DSFormer on progressively more aggressively clustered versions of SF-XL to find the data-efficiency ceiling before accuracy starts to drop.
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

4 major / 6 minor

Summary. DSFormer proposes a dual-scale cross-learning Transformer module that fuses features from the last two ResNet-50 (or DINOv2) layers via self- and cross-attention, applied to visual place recognition. The paper also introduces a block-clustering strategy based on HDBSCAN to repartition the SF-XL dataset, reducing training data by roughly 30%. The method is evaluated on MSLS, Pittsburgh30k, Tokyo24/7, Nordland, and SF-XL test sets, with reported state-of-the-art results both with a ResNet-50 backbone and with DINOv2, and a speed/memory comparison against reranking baselines.

Significance. If verified, the contributions are notable: a global-descriptor method that rivals or exceeds two-stage reranking methods at a fraction of the memory and latency, plus a data-partitioning scheme that improves accuracy while reducing training data. The ablations are informative and the core architectural ablation (removing cross-encoder, self-encoder, or block clustering) degrades performance on the challenging Tokyo24/7 and Nordland datasets. The paper, however, provides no code, checkpoints, or official baseline numbers, and all comparisons are against the authors' own re-implementations, which makes the headline state-of-the-art claim unverifiable as written.

major comments (4)
  1. [Section IV.A.3, Tables I and III] The state-of-the-art claim is supported only by the authors' own re-implementations of all baselines, as stated in Section IV.A.3: 'All comparison results reported here are from our own reproductions of the respective methods.' The paper provides no code, checkpoints, or per-baseline tuning details. The margin over R2Former is only 0.5 average R@1 (83.8 vs 83.3) and DSFormer is lower than R2Former on both MSLS rows by 0.8 and 4.9 points, so a modest under-tuning of a reproduction could flip the ranking. Please provide the official published baseline numbers alongside the re-implementations, or release the code/checkpoints and a detailed reproduction protocol.
  2. [Section IV.C] The ablations for the number of DSFormer layers (Table IV) and for the focal distance (Table V) are performed on the same benchmark datasets (MSLS Val, Pitts30k, Tokyo24/7, Nordland) that appear in the final comparison tables. This constitutes a test-set selection effect: the hyperparameters are chosen to maximize performance on the test sets, making the final reported numbers optimistic. Please select hyperparameters on a held-out validation split or use cross-validation, and report the sensitivity of the final results to these choices.
  3. [Section III.B] The block-clustering pipeline has four free parameters (block width M, number of groups N, KNN radius r, and minimum peak-distance l) plus the focal distance F, but only F is ablated. The claim that the partition reduces training data by about 30% while improving accuracy depends on these parameters. The paper does not report how many classes are produced by HDBSCAN or how many are removed by the filtering steps, nor does it provide a sensitivity analysis for M, N, r, and l. Please provide these statistics and a sensitivity study.
  4. [Section IV.B] All results are reported as single runs with no error bars, variance, or significance tests. Given the small margins (e.g., 0.5 average R@1 over R2Former in Table I) and the selection effect described above, the reported numbers do not demonstrate that DSFormer is consistently superior. Please report multiple seed runs with mean±std or a significance analysis.
minor comments (6)
  1. [Section III.A.3, Eq. (6)] The cross-attention equation is ambiguous: it defines Attention(q_i, k_m, v_m) but then says 'i ≠ m' without explaining how queries, keys, and values are drawn from the two scales, and the 'shared' nature of the cross-encoder is not explicit. Please clarify the assignment of inputs and the weight-sharing scheme.
  2. [Section IV.B] The sentence 'we add a novel large-scale benchmark, SF-XL' is misleading: SF-XL originates from [24] and the test sets (v1, v2, night, occlusion) were introduced in [44]. Please rephrase to credit the original source.
  3. [Related Work] The paragraph 'Recent advances [18]–[20] leverage the capabilities of DINOv2...' is repeated verbatim twice in the introduction/related work; one copy should be deleted.
  4. [Table II] The latency and memory comparison does not state the input resolutions; the two-stage methods use 640×480 while DSFormer uses 320×320, which may partly explain the efficiency advantage. Please report a same-resolution comparison or explicitly discuss the resolution difference.
  5. [Section IV.A.1] The training description says 'up to 40 epochs' but does not state the actual number of epochs used for the final model or the early-stopping criterion.
  6. [Eq. (5)] The IRPE attention formula applies rpe_v to the output of softmax; standard formulations add the relative-position bias to the attention logits before softmax and multiply the value term differently. Please verify the formula and correct any typo.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the architecture and data-partition contributions are evaluated on external benchmarks; the main caveats (own baseline re-implementations, benchmark-based hyperparameter choice) are verification risks, not construction-equivalence.

full rationale

This paper is an empirical systems paper rather than a theoretical derivation, so the classic derivation-equals-input failure mode is absent. The DSFormer mechanism (Eqs. 5-6) is a standard multi-head self- and cross-attention fusion of two backbone feature maps; nothing in the formulation is defined in terms of the Recall@1 metric it later reports. The block-clustering partition is constructed from UTM geometry via HDBSCAN, KNN radius filtering, SVD principal directions, and a focal distance; it does not use benchmark query/reference labels or retrieval scores as inputs. Training uses LMCL on SF-XL, and evaluation is on MSLS Val, MSLS Challenge, Pittsburgh30k, Tokyo24/7, Nordland, and SF-XL test subsets, i.e., the reported quantities are not fitted parameters renamed as predictions. The only self-reference is citation [17] in the related-work sentence contrasting prior multi-layer concatenation with the proposed cross-attention; that citation is not load-bearing for any stated result. The two legitimate concerns are the statement in Sec. IV.A.3 that 'All comparison results reported here are from our own reproductions of the respective methods' and the selection of layer count and focal distance via ablations on benchmark datasets (Sec. IV.C). These are reproducibility and data-snooping risks that could weaken the SOTA claims, but they do not make any result equivalent to its inputs by construction. No uniqueness theorem, no ansatz smuggled through self-citation, and no renaming of a known result as a new organization are present. Therefore the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central empirical claims rest on six manually chosen or ablated hyperparameters in the data partition and architecture, plus assumptions about dataset validity, pretrained feature transferability, cluster coherence, and baseline fidelity. No new physical entities are introduced.

free parameters (6)
  • Block width M = 10 m
    Manual hyperparameter for SF-XL partition; controls class granularity, no ablation.
  • Number of groups N = 5 (10 across east/north)
    Manual hyperparameter for SF-XL partition; no ablation.
  • KNN radius r = 7.5 m
    Manual retention radius around density peaks; controls intra-class spread, no ablation.
  • Minimum peak-distance l = 40 m
    Manual threshold to remove overlapping classes; no ablation.
  • Focal distance F = 15 m
    Selected by ablation on MSLS Val and Pitts30K (Table V), which are also final benchmarks.
  • Number of DSFormer layers L = 3 for ResNet-50, 1 for DINOv2
    Selected by ablation on MSLS, Pittsburgh, Tokyo, and Nordland (Table IV), the same datasets used for final scores.
assumptions (5)
  • domain assumption SF-XL is a valid dense dataset for VPR training and evaluation.
    All training and most benchmarking build on SF-XL and its test queries (Section IV.A).
  • domain assumption Features of ImageNet or LVD-142M pretrained backbones are transferable to VPR after fine-tuning.
    The method fine-tunes only the last two layers of ResNet-50 or DINOv2 (Section IV.A.1).
  • domain assumption HDBSCAN clusters over UTM coordinates create coherent classes for metric learning.
    The block clustering contribution depends on density clusters corresponding to places that should be learned together (Section III.B).
  • ad hoc to paper Self-implemented baselines faithfully reproduce published methods.
    Section IV.A.3 states all comparison results come from the authors' own reproductions; no external code or checkpoints are cited for these numbers.
  • standard math Standard benchmark protocols are correctly implemented.
    Recall@N with 25m tolerance and Nordland with +/-2 frames are used (Section IV.A.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of DSFormer: A Dual-Scale Cross-Learning Transformer for Visual Place Recognition." pith.science (2026). https://pith.science/paper/LK3NOQVG

@misc{pith2026250718444,
  author       = {Pith},
  title        = {Pith review of: DSFormer: A Dual-Scale Cross-Learning Transformer for Visual Place Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LK3NOQVG}},
  note         = {Machine review of arXiv:2507.18444}
}
read the original abstract

Visual Place Recognition (VPR) is crucial for robust mobile robot localization, yet it faces significant challenges in maintaining reliable performance under varying environmental conditions and viewpoints. To address this, we propose a novel framework that integrates Dual-Scale-Former (DSFormer), a Transformer-based cross-learning module, with an innovative block clustering strategy. DSFormer enhances feature representation by enabling bidirectional information transfer between dual-scale features extracted from the final two CNN layers, capturing both semantic richness and spatial details through self-attention for long-range dependencies within each scale and shared cross-attention for cross-scale learning. Complementing this, our block clustering strategy repartitions the widely used San Francisco eXtra Large (SF-XL) training dataset from multiple distinct perspectives, optimizing data organization to further bolster robustness against viewpoint variations. Together, these innovations not only yield a robust global embedding adaptable to environmental changes but also reduce the required training data volume by approximately 30\% compared to previous partitioning methods. Comprehensive experiments demonstrate that our approach achieves state-of-the-art performance across most benchmark datasets, surpassing advanced reranking methods like DELG, Patch-NetVLAD, TransVPR, and R2Former as a global retrieval solution using 512-dim global descriptors, while significantly improving computational efficiency.

Figures

Figures reproduced from arXiv: 2507.18444 by the authors.

Figure 1
Figure 1. Feature heatmap visualizations of both the pre-trained foundation [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (a) DSFormer module: The inputs consists of dual-scale feature maps from ResNet-50, linearly mapped into two embedded patches of the same [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Sketch of block clustering’s principle. The region is initially partitioned along the east or north direction into blocks of width M meters, where sampling locations truncated by street boundaries are grouped into a single class. The KNN method retains sampling locations within a radius r centered on the peak density location, discarding others. SVD is applied to identify two principal directions, and from each cent… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 38 canonical work pages

  1. [1]

    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

    work page 2015

  2. [2]

    Multires-netvlad: Augmenting place recognition training with low-resolution imagery,

    A. Khaliq, M. Milford, and S. Garg, “Multires-netvlad: Augmenting place recognition training with low-resolution imagery,” IEEE Robotics and Automation Letters , vol. 7, no. 2, pp. 3882–3889, 2022

    work page 2022

  3. [3]

    From coarse to fine: Robust hierarchical localization at large scale,

    P.-E. Sarlin, C. Cadena, R. Siegwart, and M. Dymczyk, “From coarse to fine: Robust hierarchical localization at large scale,” in CVPR, 2019

    work page 2019

  4. [4]

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

    J. Revaud, J. Almazan, R. Rezende, and C. de Souza, “Learning with average precision: Training image retrieval with a listwise loss,” in ICCV, 2019

    work page 2019

  5. [5]

    Learned contextual feature reweighting for image geo-localization

    H. J. Kim, E. Dunn, and J.-M. Frahm, “Learned contextual feature reweighting for image geo-localization.” in CVPR, vol. 1, no. 2, 2017, p. 3

    work page 2017

  6. [6]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2016, pp. 770–778

    2016

  7. [7]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556 , 2014

    arXiv 2014

  8. [8]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition . Ieee, 2009, pp. 248–255. 8

    work page 2009

Show all 45 references
  1. [9]

    Fine-tuning cnn image retrieval with no human annotation,

    F. Radenovi ´c, G. Tolias, and O. Chum, “Fine-tuning cnn image retrieval with no human annotation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 41, no. 7, pp. 1655–1668, 2018

    2018

  2. [10]

    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,” in Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 5297–5307

    2016

  3. [11]

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

    F. Warburg, S. Hauberg, M. Lopez-Antequera, P. Gargallo, Y . Kuang, and J. Civera, “Mapillary street-level sequences: A dataset for lifelong place recognition,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 2626–2635

    2020

  4. [12]

    Visual place recognition with repetitive structures,

    A. Torii, J. Sivic, T. Pajdla, and M. Okutomi, “Visual place recognition with repetitive structures,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2013, pp. 883–890

    2013

  5. [13]

    Unifying deep local and global features for image search,

    B. Cao, A. Araujo, and J. Sim, “Unifying deep local and global features for image search,” in European Conference on Computer Vision. Springer, 2020, pp. 726–743

    2020

  6. [14]

    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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 14 141–14 152

    2021

  7. [15]

    Transvpr: Transformer-based place recognition with multi-level attention aggrega- tion,

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

    2022

  8. [16]

    R2former: Unified retrieval and reranking transformer for place recognition,

    S. Zhu, L. Yang, C. Chen, M. Shah, X. Shen, and H. Wang, “R2former: Unified retrieval and reranking transformer for place recognition,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 19 370–19 380

    2023

  9. [17]

    Robust visual place recog- nition for severe appearance changes,

    H. Jiang, S. Piao, H. Yu, W. Li, and L. Yu, “Robust visual place recog- nition for severe appearance changes,” IEEE Robotics and Automation Letters, 2024

    2024

  10. [18]

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

    F. Lu, L. Zhang, X. Lan, S. Dong, Y . Wang, and C. Yuan, “Towards seamless adaptation of pre-trained models for visual place recognition,” in The Twelfth International Conference on Learning Representations ,

  11. [19]

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

    F. Lu, X. Lan, L. Zhang, D. Jiang, Y . Wang, and C. Yuan, “Cricavpr: Cross-image correlation-aware representation learning for visual place recognition,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 16 772–16 782

    2024

  12. [20]

    Optimal transport aggregation for visual place recognition,

    S. Izquierdo and J. Civera, “Optimal transport aggregation for visual place recognition,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 17 658–17 668

    2024

  13. [21]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby et al. , “Dinov2: Learning robust visual features without supervision,” arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  14. [22]

    Rethinking visual geo- localization for large-scale applications,

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

    2022

  15. [23]

    Eigenplaces: Training viewpoint robust models for visual place recognition,

    G. Berton, G. Trivigno, B. Caputo, and C. Masone, “Eigenplaces: Training viewpoint robust models for visual place recognition,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 11 080–11 090

    2023

  16. [24]

    City- scale landmark identification on mobile devices,

    D. M. Chen, G. Baatz, K. K ¨oser, S. S. Tsai, R. Vedantham, T. Pylv¨an¨ainen, K. Roimela, X. Chen, J. Bach, M. Pollefeys et al., “City- scale landmark identification on mobile devices,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)...

    2011

  17. [25]

    hdbscan: Hierarchical density based clustering,

    L. McInnes, J. Healy, and S. Astels, “hdbscan: Hierarchical density based clustering,” The Journal of Open Source Software , vol. 2, no. 11, p. 205, 2017

    2017

  18. [26]

    Bags of binary words for fast place recognition in image sequences,

    D. G ´alvez-L´opez and J. D. Tardos, “Bags of binary words for fast place recognition in image sequences,” IEEE Transactions on Robotics , vol. 28, no. 5, pp. 1188–1197, 2012

    2012

  19. [27]

    Aggregating local descriptors into a compact image representation,

    H. J ´egou, M. Douze, C. Schmid, and P. P ´erez, “Aggregating local descriptors into a compact image representation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . IEEE, 2010, pp. 3304–3311

    2010

  20. [28]

    Mixvpr: Feature mixing for visual place recognition,

    A. Ali-bey, B. Chaib-draa, and P. Gigu `ere, “Mixvpr: Feature mixing for visual place recognition,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , January 2023, pp. 2998–3007

    2023

  21. [29]

    Transvlad: Multi-scale attention-based global descriptors for visual geo-localization,

    Y . Xu, P. Shamsolmoali, E. Granger, C. Nicodeme, L. Gardes, and J. Yang, “Transvlad: Multi-scale attention-based global descriptors for visual geo-localization,” in Proceedings of the IEEE/CVF Winter Con- ference on Applications of Computer Vision , 2023, pp. 2840–2849

    2023

  22. [30]

    Anyloc: Towards universal visual place recognition,

    N. Keetha, A. Mishra, J. Karhade, K. M. Jatavallabhula, S. Scherer, M. Krishna, and S. Garg, “Anyloc: Towards universal visual place recognition,” IEEE Robotics and Automation Letters , vol. 9, no. 2, pp. 1286–1293, 2023

    2023

  23. [31]

    An image is worth 16x16 words: Trans- formers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Trans- formers for image recognition at scale,” in International Conference on Learni...

    2021

  24. [32]

    Rethinking and improv- ing relative position encoding for vision transformer,

    K. Wu, H. Peng, M. Chen, J. Fu, and H. Chao, “Rethinking and improv- ing relative position encoding for vision transformer,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 10 033–10 041

    2021

  25. [33]

    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,” in Proceedings of the IEEE International Conference on Computer Vision , 2017, pp. 3456–3465

    2017

  26. [34]

    Superpoint: Self- supervised interest point detection and description,

    D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superpoint: Self- supervised interest point detection and description,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2018, pp. 224–236

    2018

  27. [35]

    Superglue: Learning feature matching with graph neural networks,

    P.-E. Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superglue: Learning feature matching with graph neural networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2020, pp. 4938–4947

    2020

  28. [36]

    Gsv-cities: Toward appro- priate supervised visual place recognition,

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

    2022

  29. [37]

    BoQ: A place is worth a bag of learnable queries,

    A. Ali-bey, B. Chaib-draa, and P. Gigu `ere, “BoQ: A place is worth a bag of learnable queries,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , June 2024, pp. 17 794–17 803

    2024

  30. [38]

    Eigenplaces: Training viewpoint robust models for visual place recognition,

    G. Berton, G. Trivigno, B. Caputo, and C. Masone, “Eigenplaces: Training viewpoint robust models for visual place recognition,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023, pp. 11 080–11 090

    2023

  31. [39]

    Cosface: Large margin cosine loss for deep face recognition,

    H. Wang, Y . Wang, Z. Zhou, X. Ji, D. Gong, J. Zhou, Z. Li, and W. Liu, “Cosface: Large margin cosine loss for deep face recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 5265–5274

    2018

  32. [40]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in 3rd International Conference on Learning Representations, ICLR , 2015

    2015

  33. [41]

    24/7 place recognition by view synthesis,

    A. Torii, R. Arandjelovic, J. Sivic, M. Okutomi, and T. Pajdla, “24/7 place recognition by view synthesis,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2015, pp. 1808–1817

    2015

  34. [42]

    Are we there yet? challenging seqslam on a 3000 km journey across all four seasons,

    N. S ¨underhauf, P. Neubert, and P. Protzel, “Are we there yet? challenging seqslam on a 3000 km journey across all four seasons,” Proc. of Workshop on Long-Term Autonomy, IEEE International Conference on Robotics and Automation (ICRA) , p. 2013, 01 2013

    2013

  35. [43]

    Single-view place recognition under seasonal changes,

    D. Olid, J. M. F ´acil, and J. Civera, “Single-view place recognition under seasonal changes,” in PPNIV Workshop at IROS 2018 , 2018

    2018

  36. [44]

    Are local features all you need for cross- domain visual place recognition?

    G. Barbarani, M. Mostafa, H. Bayramov, G. Trivigno, G. Berton, C. Masone, and B. Caputo, “Are local features all you need for cross- domain visual place recognition?” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 6154–6164

    2023

  37. [2024]

    Available: https://openreview.net/forum?id=TVg6hlfsKa

    [Online]. Available: https://openreview.net/forum?id=TVg6hlfsKa

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.