Pith. sign in

REVIEW 3 major objections 5 minor 4 cited by

A visual place recognition model that fuses 3D scene geometry tokens with 2D texture tokens outperforms single- and multi-view baselines across standard benchmarks, the paper argues.

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 · deepseek-v4-flash

2026-08-03 14:11 UTC pith:UX2BEFXE

load-bearing objection Solid VPR systems contribution with strong reported numbers, but the 3D-token mechanism is confounded by a larger, fine-tuned backbone and higher-dimensional descriptor. the 3 major comments →

arxiv 2512.21078 v3 pith:UX2BEFXE submitted 2025-12-24 cs.CV

UniPR-3D: Towards Universal Visual Place Recognition with Visual Geometry Grounded Transformer

Tianchen Deng , Xun Chen , Ziming Li , Hongming Shen , Shuhao Zhai , Danwei Wang , Javier Civera , Hesheng Wang This is my paper
classification cs.CV
keywords visual place recognitionmulti-view retrieval3D tokensVGGToptimal transport aggregationsequence matchingDINOv2place descriptor
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.

The paper argues that visual place recognition should use 3D scene structure, not just 2D texture, and that a transformer pretrained for multi-view geometry (VGGT) can supply both. It proposes the first VPR descriptor built from 3D tokens, fused with 2D tokens, with different aggregation for each token type. Reported results show it outperforms single- and multi-view baselines on several benchmarks, and that its sequence descriptor handles variable-length inputs. If right, this shifts VPR from appearance matching to geometry-grounded retrieval.

Core claim

UniPR-3D claims that combining 2D and 3D tokens from a VGGT backbone yields a place descriptor that beats existing single-frame and sequence-level baselines. It retains 3D register and patch tokens (dropping the camera token) for viewpoint invariance, aggregates CLS and register tokens with GeM pooling and patch tokens with optimal transport, and concatenates five descriptor components. For sequences, an anchor frame plus support frames is aggregated with GeM+MLP for register tokens and Sinkhorn optimal transport for patch tokens, supporting arbitrary sequence length. The paper reports, for example, single-frame R@1 of 76.2 on Nordland versus 69.6 for DINOv2-SALAD, and sequence R@1 of 93.7/8

What carries the argument

The central object is the VGGT transformer, which, given an image sequence, outputs both 2D tokens (CLS, register, patch) and 3D tokens (camera, register, patch) via alternating frame attention and global attention. The paper's contribution is a routing: drop the 3D camera token, GeM-pool the small CLS and register tokens, and use optimal transport (Sinkhorn) with a dustbin to aggregate patch tokens into local descriptors. Sequence matching adds a GeM+MLP projector that aligns tokens across frames, making the descriptor agnostic to input length.

Load-bearing premise

The reported gains come from VGGT's 3D tokens carrying genuine scene geometry, rather than from simply having a larger backbone and a much longer descriptor than the baselines.

What would settle it

An ablation that fixes backbone architecture (24 layers), descriptor dimension (17,152), and training data, and compares (a) 2D tokens only, (b) 3D tokens only, (c) both. If (a) matches (c) within noise on Nordland and Oxford, the geometry claim is falsified. A second check: feed the model two identical images with artificially swapped camera intrinsics; if 3D tokens change the descriptor significantly, they encode camera parameters rather than stable scene structure.

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

If this is right

  • If the claims hold, a single architecture can serve both frame-to-frame and sequence-to-sequence place recognition without separate temporal post-processing.
  • Geometry-grounded tokens should be more robust to appearance changes (season, day/night) than texture-only features, as the Nordland and Oxford results suggest.
  • Variable-length sequence retrieval means a robot can query with whatever number of frames it has, without retraining for a fixed window.
  • The concatenated descriptor of roughly 17k dimensions trades latency (about 8 ms for a single frame, 140 ms for a sequence) for accuracy.

Where Pith is reading between the lines

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

  • A matched-backbone ablation (same 24-layer ViT, same 17,152-dimensional descriptor, same training data) comparing 2D-tokens-only, 3D-tokens-only, and both would isolate whether the gains come from 3D geometry or from backbone scale and descriptor size; without that ablation, the '3D helps' claim is plausible but unproven.
  • The same geometric tokens could benefit other retrieval tasks where appearance changes but structure persists, such as loop closure in SLAM, re-localization, and map-based navigation, though the paper does not test these.
  • One could probe the geometric grounding directly by feeding two identical images with artificially swapped camera intrinsics or shuffled viewpoints: if the 3D token descriptor changes significantly, the tokens encode camera parameters rather than stable scene structure.
  • The variable-length aggregator points toward lifelong mapping scenarios where a robot's memory grows over time and retrieval uses arbitrarily long egocentric clips, an extension the paper leaves implicit.

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 / 5 minor

Summary. UniPR-3D proposes a visual place recognition (VPR) architecture built on VGGT, using both VGGT's multi-view 3D tokens and its intermediate 2D tokens. The 2D CLS/register tokens and 3D register tokens are aggregated with GeM pooling and MLPs, while 2D and 3D patch tokens are aggregated through an optimal-transport / Sinkhorn module, and the five descriptors are concatenated. The method supports both single-frame retrieval and variable-length sequence-level retrieval, with the first frame as anchor. Training is done in two stages: first the aggregation heads, then LoRA fine-tuning of parts of VGGT/DINOv2 and the alternating-attention blocks. The paper reports state-of-the-art results on standard single-frame benchmarks (MSLS, Nordland, Pittsburgh, SPED) and sequence benchmarks (MSLS, Nordland, Oxford1, Oxford2), and ablates the contribution of the different token types.

Significance. If the results hold, this is a useful and timely contribution: it is, to my knowledge, the first VPR method that explicitly leverages VGGT's 3D tokens, and the variable-length sequence aggregation addresses a real limitation of fixed-length multi-frame methods. The paper evaluates on several standard benchmarks and includes ablation of the token types (Tab. 4) and of test-time sequence length (Tab. 5 / Suppl. Tab. 7). The reported single-frame margins over SALAD are often small (0.1–2.4 points on MSLS, 0.4–1.5 on Pitts/SPED) but the Nordland improvement (+6.6 R@1) and the sequence-level improvements over CaseVPR (2.5–7.8 points R@1) are substantial. The main weakness is that the central mechanistic claim — that the gains come from 3D geometry rather than from a larger, fine-tuned backbone and a doubled descriptor size — is not established by the current experimental design.

major comments (3)
  1. [Sec. 4.2, Tab. 1] The headline single-frame comparison against SALAD is confounded. UniPR-3D LoRA-fine-tunes the VGGT/DINOv2 backbone (Sec. 3.4, Suppl. C), whereas the SALAD baseline uses a frozen DINOv2 with only its head trained. UniPR-3D also uses a 17,152-D descriptor vs. SALAD's 8192+256-D descriptor. Thus the +6.6 R@1 on Nordland and the smaller gains elsewhere cannot be attributed to 3D tokens. A matched control is needed: same LoRA fine-tuning regime and same descriptor dimension, but using only 2D tokens (e.g., 2D patch/cls/register with expanded clusters or duplicated register tokens to match 17,152-D). Without this control, the paper's claim that geometry-grounded tokens are the source of the gains is unsupported.
  2. [Sec. 4.3, Tab. 4] The ablation of token types is not sufficient to isolate the 3D-geometry contribution. The all-2D row (2D cls + 2D reg + 2D patch) has a descriptor dimension of roughly 8,704, while the full descriptor is 17,152-D; the +4.5 R@1 from adding 3D tokens could be partly a descriptor-size effect. Also, the table's '3D pose' column is never defined in the method and the best row omits it, so the contribution ordering is unclear. Please (i) clarify what '3D pose' refers to, and (ii) add a 2D-only condition with matched descriptor dimension, e.g., by increasing the number of 2D patch clusters or adding a second 2D register stream of equal dimension.
  3. [Sec. 3.4 / Suppl. C] Training details are insufficient for reproducibility, and no code is currently provided despite the abstract's promise. Please report LoRA rank/alpha, which specific DINOv2/VGGT blocks are fine-tuned, total number of epochs/steps, batch size, Sinkhorn iterations and temperature, dustbin initialization, and the exact two-stage schedule. These details are load-bearing because the method is a complex fine-tuned system and the reported gains are often just a few points.
minor comments (5)
  1. [Sec. 4.3] The main text says 'The results are presented in Tab. 7' but the table in the main paper is numbered Tab. 5, while Tab. 7 is in the supplementary. Please fix the cross-reference.
  2. [Eq. (1)] Equation (1) for the 2D CLS descriptor uses 'fp_cam' (camera features) in the GeM formula; this appears to be a typo for the CLS token f_cls. Also, the text says the 3D camera token is discarded, but Fig. 2 and Eq. (3) do not make this distinction clear.
  3. [Tab. 1] The descriptor size '8192 + 256' for DINOv2 SALAD is ambiguous; clarify whether this is the concatenated dimension of patch and register/cls descriptors or something else.
  4. [Fig. 2 caption] The caption says 'we use DINOv2 as our visual feature extractor and then utilize alternating attention blocks of VGGT', but the method section describes VGGT as containing a DINOv2 encoder. Please align the wording to avoid implying two separate networks.
  5. [Sec. 4.2, Tab. 2] The caption for Tab. 2 says 'Oxford datasets, the retrieval distance is set to 2 m', but the main text and Tab. 3 report both 2 m and 25 m. Clarify whether Tab. 2 uses only 2 m.

Circularity Check

0 steps flagged

No significant circularity: the method is trained and evaluated on external benchmarks, and none of the reported predictions reduce by construction to fitted inputs or self-citations.

full rationale

UniPR-3D is a learned retrieval system whose descriptor is constructed from VGGT/DINOv2 tokens via GeM pooling and optimal-transport aggregation (Eqs. 1-8), trained with a multi-similarity loss on GSV-Cities and MSLS, and then evaluated on held-out public benchmarks (MSLS, Nordland, Pittsburgh250k, SPED, Oxford). No fitted parameter is defined in terms of the target recall, and no claimed 'prediction' is a rewritten version of its training signal. The component ablation in Table 4 is a within-model removal study, not a fit-then-predict cycle, and it does not define the 3D token contribution as the benchmark result. The citations to SALAD (co-authored by J. Civera) and CaseVPR (co-authored by D. Wang) are self-citations, but they are not load-bearing in a circular sense: SALAD supplies an aggregation technique (optimal transport with a dustbin) that is independently published and externally evaluated, and CaseVPR is a baseline whose numbers are compared rather than used to define the method. No uniqueness theorem or ansatz is imported from the authors' prior work to force the architecture. The skeptical concern about a missing matched 2D-only ablation (larger fine-tuned backbone and 17,152-D descriptor versus frozen baselines) is an attribution/confound issue about explaining the gains, not a circularity issue: even if the mechanistic claim that 3D tokens cause the improvements is under-supported, the empirical evaluation remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or mathematical entities; it relies on pre-trained VGGT tokens and standard OT aggregation. The free parameters are learned or hand-chosen aggregation hyperparameters. The key uncharged assumption is that the VGGT geometry tokens, rather than raw backbone capacity and descriptor size, drive the improvements.

free parameters (4)
  • GeM pooling exponent p = not reported (learned)
    Controls the selectivity of GeM pooling for CLS/register descriptors; learned during training (Eqs. 1-3, 9).
  • OT dustbin logit z = not reported (learned)
    A single learnable score for the dustbin column in the Sinkhorn assignment (Eq. 6), inherited from SuperGlue/SALAD.
  • Number of OT clusters m = 64
    Hand-chosen cluster count; the patch descriptor dimension is 128 x 64 = 8192 (Sec. 4.1).
  • Descriptor dimensionalities = 17,152 total; cls/reg 256, patch 8192
    Design choices that make the descriptor larger than all baselines; no matched-size comparison is provided, so this is a confound for the SOTA claim.
axioms (5)
  • domain assumption VGGT 3D tokens retain place-discriminative geometric information after LoRA fine-tuning on MSLS
    Core premise of Sec. 3.1; the paper only validates it through a full-system ablation and not through a matched-capacity 2D-only control.
  • domain assumption Multi-similarity contrastive training on GSV-Cities/MSLS transfers to Nordland, Pittsburgh, SPED, and Oxford
    The evaluation protocol assumes that cross-dataset retrieval accuracy measures generalization (Sec. 3.4, 4.2).
  • domain assumption The 25m/10-frame/2m positive definitions used for training and evaluation approximate true place identity
    Standard VPR convention; retrieval labels depend on these thresholds (Sec. 4.1, D).
  • standard math Sinkhorn optimal transport converges and yields a valid soft assignment matrix
    Used for patch aggregation; standard result referenced to Cuturi and SuperGlue (Sec. 3.2).
  • domain assumption LoRA fine-tuning of only the last VGGT blocks preserves the pretrained geometry representation
    The two-stage protocol freezes most of the backbone for stability; if LoRA disrupts geometry encoding, the 3D-token gains would vanish (Sec. C).

pith-pipeline@v1.3.0-alltime-deepseek · 19300 in / 12709 out tokens · 120162 ms · 2026-08-03T14:11:36.771934+00:00 · methodology

0 comments
read the original abstract

Visual Place Recognition (VPR) has been traditionally formulated as a single-image retrieval task. Using multiple views offers clear advantages, yet this setting remains relatively underexplored and existing methods often struggle to generalize across diverse environments. In this work we introduce UniPR-3D, the first VPR architecture that effectively integrates information from multiple views. UniPR-3D builds on a VGGT backbone capable of encoding multi-view 3D representations, which we adapt by designing feature aggregators and fine-tune for the place recognition task. To construct our descriptor, we jointly leverage the 3D tokens and intermediate 2D tokens produced by VGGT. Based on their distinct characteristics, we design dedicated aggregation modules for 2D and 3D features, allowing our descriptor to capture fine-grained texture cues while also reasoning across viewpoints. To further enhance generalization, we incorporate both single- and multi-frame aggregation schemes, along with a variable-length sequence retrieval strategy. Our experiments show that UniPR-3D sets a new state of the art, outperforming both single- and multi-view baselines and highlighting the effectiveness of geometry-grounded tokens for VPR. Our code and models will be made publicly available on Github https://github.com/dtc111111/UniPR-3D.

Figures

Figures reproduced from arXiv: 2512.21078 by Danwei Wang, Hesheng Wang, Hongming Shen, Javier Civera, Shuhao Zhai, Tianchen Deng, Xun Chen, Ziming Li.

Figure 1
Figure 1. Figure 1: Illustration of a typical single-view VPR model (top) and our multi-view model (bottom). Single-view VPR extracts image features with a deep backbone and aggregates them into an image descriptor. In contrast, our multi-view VPR model em￾ploys a VGGT backbone to jointly extract both 2D and 3D tokens from multiple views, followed by tailored aggregation strategies for each token type. Our framework supports … view at source ↗
Figure 2
Figure 2. Figure 2: Overview. We propose the first VPR method that supports both single-frame and sequence-level place retrieval. Specifically, we use DINOv2 as our visual feature extractor and then utilize the alternating attention blocks of VGGT to derive 3D tokens. The resulting 3D tokens are divided into different groups, each processed with a dedicated aggregation strategy to form the final descriptor. In particular, cls… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative sequence matching results on the Oxford dataset [41]. The left column shows two query sequence images, while the three right columns present the top-3 candidates retrieved by our UniPR-3D and the baseline CaseVPR [34]. Successful retrievals are framed in green, while erroneous ones are shown in red. UniPR-3D retrieves the correct place even under challenging seasonal, weather, viewpoint, and da… view at source ↗
Figure 4
Figure 4. Figure 4: Activation heatmaps of 3D and 2D features in the Oxford dataset [41]. The left column shows reference images, and the two right columns their corresponding 3D and 2D feature heatmaps. Note how the different feature types fire at different locations in the image, illustrating their complementary nature. #frames at test time Oxford2 R@1 R@5 R@10 3 70.3 87.2 91.7 5 (training setup) 80.6 90.3 93.9 10 89.1 93.1… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative sequence matching results on the MSLS dataset [66]. The left column shows two query sequence images, while the three right columns present the top-3 candidates retrieved by our UniPR-3D and the baseline CaseVPR [34]. Successful retrievals are framed in green, while erroneous ones are shown in red. UniPR-3D retrieves the correct place even under challenging seasonal, weather, viewpoint, and day-… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative sequence matching results on the MSLS dataset [66]. The left column shows two query sequence images, while the three right columns present the top-3 candidates retrieved by our UniPR-3D and the baseline CaseVPR [34]. Successful retrievals are framed in green, while erroneous ones are shown in red. UniPR-3D retrieves the correct place even under challenging seasonal, weather, viewpoint, and day-… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative sequence matching results on the Nordland dataset [58]. The left column shows two query sequence images, while the three right columns present the top-3 candidates retrieved by our UniPR-3D and the baseline CaseVPR [34]. Successful retrievals are framed in green, while erroneous ones are shown in red. UniPR-3D retrieves the correct place even under challenging seasonal, weather, viewpoint, and … view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative sequence matching results on the Nordland dataset [58]. The left column shows two query sequence images, while the three right columns present the top-3 candidates retrieved by our UniPR-3D and the baseline CaseVPR [34]. Successful retrievals are framed in green, while erroneous ones are shown in red. UniPR-3D retrieves the correct place even under challenging seasonal, weather, viewpoint, and … view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative sequence matching results on the Oxford dataset [58]. The left column shows two query sequence images, while the three right columns present the top-3 candidates retrieved by our UniPR-3D and the baseline CaseVPR [34]. Successful retrievals are framed in green, while erroneous ones are shown in red. UniPR-3D retrieves the correct place even under challenging seasonal, weather, viewpoint, and da… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative sequence matching results on the Oxford dataset [58]. The left column shows two query sequence images, while the three right columns present the top-3 candidates retrieved by our UniPR-3D and the baseline CaseVPR [34]. Successful retrievals are framed in green, while erroneous ones are shown in red. UniPR-3D retrieves the correct place even under challenging seasonal, weather, viewpoint, and d… view at source ↗
Figure 11
Figure 11. Figure 11: t-SNE visualization results. We additionally provide a t-SNE visualization comparing our descriptor with that of CaseVPR across different datasets. As shown in [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Keep It CALM: Toward Calibration-Free Kilometer-Level SLAM with Visual Geometry Foundation Models via an Assistant Eye

    cs.RO 2026-04 unverdicted novelty 7.0

    CAL2M achieves calibration-free kilometer-level SLAM by using an assistant eye for scale, epipolar-guided intrinsic correction, and anchor propagation for nonlinear sub-map alignment.

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

    cs.CV 2026-06 conditional novelty 6.0

    A geometry-aware second-stage network plus the XHZ ship dataset reduces mean VPR localization error by over 60% across multiple backbones in multi-floor maritime cabins.

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

    cs.CV 2026-06 unverdicted novelty 5.0

    ProteusVPR is a retrieval-plus-geometric-refinement framework that cuts mean localization error by over 60% on average for cross-scene maritime VPR and is evaluated on the new XHZ ship-borne panoramic dataset.

  4. VGGT-Occ: Geometry-Grounded and Density-Aware Gated Fusion for 3D Occupancy Prediction

    cs.CV 2026-05 unverdicted novelty 5.0

    VGGT-Occ embeds geometric tokens via PA-DA and uses sequential coarse-to-fine gated fusion to reach 33.00% IoU and 21.08% mIoU on SurroundOcc-nuScenes while using only ~41M parameters in the occupancy head.

Reference graph

Works this paper leans on

72 extracted references · 9 linked inside Pith · cited by 3 Pith papers

  1. [1]

    Gsv-cities: Toward appropriate supervised visual place recognition.Neurocomputing, 513:194–203, 2022

    Amar Ali-bey, Brahim Chaib-draa, and Philippe Giguere. Gsv-cities: Toward appropriate supervised visual place recognition.Neurocomputing, 513:194–203, 2022. 5, 6

    2022

  2. [2]

    Mixvpr: Feature mixing for visual place recognition

    Amar Ali-Bey, Brahim Chaib-Draa, and Philippe Giguere. Mixvpr: Feature mixing for visual place recognition. InPro- ceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2998–3007, 2023. 1, 2, 5, 6, 9

    2023

  3. [3]

    Netvlad: Cnn architecture for weakly supervised place recognition

    Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pa- jdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5297–5307, 2016. 1, 2, 5, 6

    2016

  4. [4]

    Towards life-long visual localization us- ing an efficient matching of binary sequences from images

    Roberto Arroyo, Pablo F Alcantarilla, Luis M Bergasa, and Eduardo Romera. Towards life-long visual localization us- ing an efficient matching of binary sequences from images. In2015 IEEE international conference on robotics and au- tomation (ICRA), pages 6328–6335. IEEE, 2015. 2

    2015

  5. [5]

    Megaloc: One retrieval to place them all

    Gabriele Berton and Carlo Masone. Megaloc: One retrieval to place them all. InProceedings of the Computer Vision and Pattern Recognition Conference Workshops, pages 2861– 2867, 2025. 1

    2025

  6. [6]

    Re- thinking visual geo-localization for large-scale applications

    Gabriele Berton, Carlo Masone, and Barbara Caputo. Re- thinking visual geo-localization for large-scale applications. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4878–4888, 2022. 5, 6

    2022

  7. [7]

    Eigenplaces: Training viewpoint robust models for visual place recognition

    Gabriele Berton, Gabriele Trivigno, Barbara Caputo, and Carlo Masone. Eigenplaces: Training viewpoint robust models for visual place recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 11080–11090, 2023. 5, 6

    2023

  8. [8]

    Jist: Joint image and sequence training for sequential visual place recognition.IEEE Robotics and Au- tomation Letters, 9(2):1310–1317, 2023

    Gabriele Berton, Gabriele Trivigno, Barbara Caputo, and Carlo Masone. Jist: Joint image and sequence training for sequential visual place recognition.IEEE Robotics and Au- tomation Letters, 9(2):1310–1317, 2023. 6

    2023

  9. [9]

    A survey of optimal trans- port for computer graphics and computer vision

    Nicolas Bonneel and Julie Digne. A survey of optimal trans- port for computer graphics and computer vision. InCom- puter Graphics Forum, pages 439–460. Wiley Online Li- brary, 2023. 2

    2023

  10. [10]

    Learning context flexible attention model for long-term visual place recognition.IEEE Robotics and Au- tomation Letters, 3(4):4015–4022, 2018

    Zetao Chen, Lingqiao Liu, Inkyu Sa, Zongyuan Ge, and Mar- garita Chli. Learning context flexible attention model for long-term visual place recognition.IEEE Robotics and Au- tomation Letters, 3(4):4015–4022, 2018. 5, 10

    2018

  11. [11]

    Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information pro- cessing systems, 26, 2013

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information pro- cessing systems, 26, 2013. 4

    2013

  12. [12]

    Flashattention-2: Faster attention with bet- ter parallelism and work partitioning.arXiv preprint arXiv:2307.08691, 2023

    Tri Dao. Flashattention-2: Faster attention with bet- ter parallelism and work partitioning.arXiv preprint arXiv:2307.08691, 2023. 5, 9

    Pith/arXiv arXiv 2023

  13. [13]

    Prosgnerf: Progressive dynamic neural scene graph with frequency modulated auto-encoder in urban scenes.arXiv preprint arXiv:2312.09076, 2023

    Tianchen Deng, Siyang Liu, Xuan Wang, Yejia Liu, Danwei Wang, and Weidong Chen. Prosgnerf: Progressive dynamic neural scene graph with frequency modulated auto-encoder in urban scenes.arXiv preprint arXiv:2312.09076, 2023. 2

    Pith/arXiv arXiv 2023

  14. [14]

    Long-term visual simultaneous localization and map- ping: Using a bayesian persistence filter-based global map prediction.IEEE Robotics & Automation Magazine, 30(1): 36–49, 2023

    Tianchen Deng, Hongle Xie, Jingchuan Wang, and Weidong Chen. Long-term visual simultaneous localization and map- ping: Using a bayesian persistence filter-based global map prediction.IEEE Robotics & Automation Magazine, 30(1): 36–49, 2023. 2

    2023

  15. [15]

    Plgslam: Progressive neural scene represenation with local to global bundle adjustment

    Tianchen Deng, Guole Shen, Tong Qin, Jianyu Wang, Wen- tao Zhao, Jingchuan Wang, Danwei Wang, and Weidong Chen. Plgslam: Progressive neural scene represenation with local to global bundle adjustment. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19657–19666, 2024. 1

    2024

  16. [16]

    What is the best 3d scene representation for robotics? from geometric to foundation models.arXiv preprint arXiv:2512.03422, 2025

    Tianchen Deng, Yue Pan, Shenghai Yuan, Dong Li, Chen Wang, Mingrui Li, Long Chen, Lihua Xie, Danwei Wang, Jingchuan Wang, Javier Civera, Hesheng Wang, and Wei- dong Chen. What is the best 3d scene representation for robotics? from geometric to foundation models.arXiv preprint arXiv:2512.03422, 2025. 2

    arXiv 2025

  17. [17]

    Mcn-slam: Multi-agent collaborative neural slam with hybrid implicit neural scene representation.arXiv preprint arXiv:2506.18678, 2025

    Tianchen Deng, Guole Shen, Xun Chen, Shenghai Yuan, Hongming Shen, Guohao Peng, Zhenyu Wu, Jingchuan Wang, Lihua Xie, Danwei Wang, Hesheng Wang, and Wei- dong Chen. Mcn-slam: Multi-agent collaborative neural slam with hybrid implicit neural scene representation.arXiv preprint arXiv:2506.18678, 2025. 2

    Pith/arXiv arXiv 2025

  18. [18]

    Mne-slam: Multi-agent neural slam for mobile robots

    Tianchen Deng, Guole Shen, Chen Xun, Shenghai Yuan, Tongxin Jin, Hongming Shen, Yanbo Wang, Jingchuan Wang, Hesheng Wang, Danwei Wang, et al. Mne-slam: Multi-agent neural slam for mobile robots. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 1485–1494, 2025. 1

    2025

  19. [19]

    Tianchen Deng, Nailin Wang, Chongdi Wang, Shenghai Yuan, Jingchuan Wang, Hesheng Wang, Danwei Wang, and Weidong Chen. Incremental joint learning of depth, pose and implicit scene representation on monocular camera in large- scale scenes.IEEE Transactions on Automation Science and Engineering, pages 1–1, 2025. 2

    2025

  20. [20]

    Neslam: Neural implicit mapping and self-supervised feature tracking with depth completion and denoising.IEEE Transactions on Automation Science and Engineering, pages 1–1, 2025

    Tianchen Deng, Yanbo Wang, Hongle Xie, Hesheng Wang, Rui Guo, Jingchuan Wang, Danwei Wang, and Weidong Chen. Neslam: Neural implicit mapping and self-supervised feature tracking with depth completion and denoising.IEEE Transactions on Automation Science and Engineering, pages 1–1, 2025. 2

    2025

  21. [21]

    Condition-invariant multi-view place recognition

    Jose M Facil, Daniel Olid, Luis Montesano, and Javier Civera. Condition-invariant multi-view place recognition. arXiv preprint arXiv:1902.09516, 2019. 2

    Pith/arXiv arXiv 1902

  22. [22]

    Bags of binary words for fast place recognition in image sequences.IEEE Transactions on robotics, 28(5):1188–1197, 2012

    Dorian G ´alvez-L´opez and Juan D Tardos. Bags of binary words for fast place recognition in image sequences.IEEE Transactions on robotics, 28(5):1188–1197, 2012. 2

    2012

  23. [23]

    Seqnet: Learning de- scriptors for sequence-based hierarchical place recognition

    Sourav Garg and Michael Milford. Seqnet: Learning de- scriptors for sequence-based hierarchical place recognition. IEEE Robotics and Automation Letters, 6(3):4305–4312,

  24. [24]

    Delta descriptors: Change-based place representa- tion for robust visual localization.IEEE Robotics and Au- tomation Letters, 5(4):5120–5127, 2020

    Sourav Garg, Ben Harwood, Gaurangi Anand, and Michael Milford. Delta descriptors: Change-based place representa- tion for robust visual localization.IEEE Robotics and Au- tomation Letters, 5(4):5120–5127, 2020. 6

    2020

  25. [25]

    Where is your place, visual place recognition?arXiv preprint arXiv:2103.06443, 2021

    Sourav Garg, Tobias Fischer, and Michael Milford. Where is your place, visual place recognition?arXiv preprint arXiv:2103.06443, 2021. 1

    Pith/arXiv arXiv 2021

  26. [26]

    Seq- matchnet: Contrastive learning with sequence matching for 18 place recognition & relocalization

    Sourav Garg, Madhu Vankadari, and Michael Milford. Seq- matchnet: Contrastive learning with sequence matching for 18 place recognition & relocalization. InConference on Robot Learning, pages 429–443. PMLR, 2022. 2, 5, 6

    2022

  27. [27]

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

    Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14141–14152, 2021. 1

    2021

  28. [28]

    Improving visual place recognition with sequence-matching receptiveness prediction.arXiv preprint arXiv:2503.06840,

    Somayeh Hussaini, Tobias Fischer, and Michael Milford. Improving visual place recognition with sequence-matching receptiveness prediction.arXiv preprint arXiv:2503.06840,

    Pith/arXiv arXiv

  29. [29]

    Close, but not there: Boosting geographic distance sensitivity in visual place recognition

    Sergio Izquierdo and Javier Civera. Close, but not there: Boosting geographic distance sensitivity in visual place recognition. InEuropean Conference on Computer Vision, pages 240–257. Springer, 2024. 1

    2024

  30. [30]

    Optimal transport ag- gregation for visual place recognition

    Sergio Izquierdo and Javier Civera. Optimal transport ag- gregation for visual place recognition. InProceedings of the ieee/cvf conference on computer vision and pattern recogni- tion, pages 17658–17668, 2024. 1, 2, 3, 4, 5, 6, 9

    2024

  31. [31]

    Aggregating local descriptors into a compact image representation

    Herv ´e J´egou, Matthijs Douze, Cordelia Schmid, and Patrick P´erez. Aggregating local descriptors into a compact image representation. In2010 IEEE computer society conference on computer vision and pattern recognition, pages 3304–

  32. [32]

    Anyloc: Towards universal visual place recognition.IEEE Robotics and Automation Letters, 9 (2):1286–1293, 2023

    Nikhil Keetha, Avneesh Mishra, Jay Karhade, Kr- ishna Murthy Jatavallabhula, Sebastian Scherer, Madhava Krishna, and Sourav Garg. Anyloc: Towards universal visual place recognition.IEEE Robotics and Automation Letters, 9 (2):1286–1293, 2023. 1, 2, 3

    2023

  33. [33]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InICLR, 2015. 5

    2015

  34. [34]

    Casevpr: Correlation-aware sequential embedding for sequence-to-frame visual place recognition.IEEE Robotics and Automation Letters, 2025

    Heshan Li, Guohao Peng, Jun Zhang, Mingxing Wen, Ying- chong Ma, and Danwei Wang. Casevpr: Correlation-aware sequential embedding for sequence-to-frame visual place recognition.IEEE Robotics and Automation Letters, 2025. 3, 5, 6, 7, 11, 12, 13, 14, 15, 16

    2025

  35. [35]

    To- ward learning-based visuomotor navigation with neural radi- ance fields.IEEE Transactions on Industrial Informatics, 20 (6):8907–8916, 2024

    Qiming Liu, Nanxi Chen, Zhe Liu, and Hesheng Wang. To- ward learning-based visuomotor navigation with neural radi- ance fields.IEEE Transactions on Industrial Informatics, 20 (6):8907–8916, 2024. 2

    2024

  36. [36]

    In- tegrating neural radiance fields end-to-end for cognitive vi- suomotor navigation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):11200–11215, 2024

    Qiming Liu, Haoran Xin, Zhe Liu, and Hesheng Wang. In- tegrating neural radiance fields end-to-end for cognitive vi- suomotor navigation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):11200–11215, 2024. 2

    2024

  37. [37]

    Mg-slam: Structure gaussian splatting slam with manhattan world hy- pothesis.IEEE Transactions on Automation Science and En- gineering, 2025

    Shuhong Liu, Tianchen Deng, Heng Zhou, Liuzhuozheng Li, Hongyu Wang, Danwei Wang, and Mingrui Li. Mg-slam: Structure gaussian splatting slam with manhattan world hy- pothesis.IEEE Transactions on Automation Science and En- gineering, 2025. 2

    2025

  38. [38]

    Visual place recognition: A survey.IEEE Transactions on Robotics, 32(1):1–19, 2015

    Stephanie Lowry, Niko S ¨underhauf, Paul Newman, John J Leonard, David Cox, Peter Corke, and Michael J Milford. Visual place recognition: A survey.IEEE Transactions on Robotics, 32(1):1–19, 2015. 1

    2015

  39. [39]

    Neural network-based nonconservative predefined-time backstep- ping control for uncertain strict-feedback nonlinear systems

    Jixing Lv, Xiaozhe Ju, and Changhong Wang. Neural network-based nonconservative predefined-time backstep- ping control for uncertain strict-feedback nonlinear systems. IEEE Transactions on Neural Networks and Learning Sys- tems, 2023. 2

    2023

  40. [40]

    Adaptive dis- tributed observer design for nonlinear multiagent systems

    Jixing Lv, Changhong Wang, and Lihua Xie. Adaptive dis- tributed observer design for nonlinear multiagent systems. Automatica, 183:112625, 2026. 2

    2026

  41. [41]

    1 year, 1000 km: The oxford robotcar dataset

    Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, 1000 km: The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3–15,

  42. [42]

    Vggt- slam: Dense rgb slam optimized on the sl (4) manifold.arXiv preprint arXiv:2505.12549, 2025

    Dominic Maggio, Hyungtae Lim, and Luca Carlone. Vggt- slam: Dense rgb slam optimized on the sl (4) manifold.arXiv preprint arXiv:2505.12549, 2025. 1

    Pith/arXiv arXiv 2025

  43. [43]

    A survey on deep visual place recognition.IEEE Access, 9:19516–19547, 2021

    Carlo Masone and Barbara Caputo. A survey on deep visual place recognition.IEEE Access, 9:19516–19547, 2021. 1

    2021

  44. [44]

    Learning sequential descrip- tors for sequence-based visual place recognition.IEEE Robotics and Automation Letters, 7(4):10383–10390, 2022

    Riccardo Mereu, Gabriele Trivigno, Gabriele Berton, Carlo Masone, and Barbara Caputo. Learning sequential descrip- tors for sequence-based visual place recognition.IEEE Robotics and Automation Letters, 7(4):10383–10390, 2022. 2, 6

    2022

  45. [45]

    Going places: Place recognition in artificial and natural systems.Annual Review of Control, Robotics, and Autonomous Systems, 9, 2025

    Michael Milford and Tobias Fischer. Going places: Place recognition in artificial and natural systems.Annual Review of Control, Robotics, and Autonomous Systems, 9, 2025. 1

    2025

  46. [46]

    Seqslam: Visual route-based navigation for sunny summer days and stormy winter nights

    Michael J Milford and Gordon F Wyeth. Seqslam: Visual route-based navigation for sunny summer days and stormy winter nights. In2012 IEEE international conference on robotics and automation, pages 1643–1649. IEEE, 2012. 2, 6

    2012

  47. [47]

    Localization in urban environments using a panoramic gist descriptor.IEEE Transactions on Robotics, 29(1):146–160, 2012

    Ana C Murillo, Gautam Singh, Jana Kosecka, and Jos ´e Jes´us Guerrero. Localization in urban environments using a panoramic gist descriptor.IEEE Transactions on Robotics, 29(1):146–160, 2012. 2

    2012

  48. [48]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3

    Pith/arXiv arXiv 2023

  49. [49]

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

    Filip Radenovi ´c, Giorgos Tolias, and Ond ˇrej Chum. Fine- tuning cnn image retrieval with no human annotation.IEEE transactions on pattern analysis and machine intelligence, 41(7):1655–1668, 2018. 1, 2, 4, 5, 6

    2018

  50. [50]

    From coarse to fine: Robust hierarchical localization at large scale

    Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12716–12725, 2019. 1

    2019

  51. [51]

    Superglue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 4

    2020

  52. [52]

    Fast and memory efficient graph optimization via icm for visual place recognition

    Stefan Schubert, Peer Neubert, and Peter Protzel. Fast and memory efficient graph optimization via icm for visual place recognition. InRobotics: Science and Systems, 2021. 2

    2021

  53. [53]

    Visual Place Recognition: A Tuto- rial.IEEE Robotics & Automation Magazine, 2023

    Stefan Schubert, Peer Neubert, Sourav Garg, Michael Mil- ford, and Tobias Fischer. Visual Place Recognition: A Tuto- rial.IEEE Robotics & Automation Magazine, 2023. 1

    2023

  54. [54]

    Grs-slam3r: Real-time dense slam with gated recurrent state.arXiv preprint arXiv:2509.23737, 2025

    Guole Shen, Tianchen Deng, Yanbo Wang, Yongtao Chen, Yilin Shen, Jiuming Liu, and Jingchuan Wang. Grs-slam3r: Real-time dense slam with gated recurrent state.arXiv preprint arXiv:2509.23737, 2025. 2 19

    arXiv 2025

  55. [55]

    Unilgl: Learning uniform place recognition for fov- limited/panoramic lidar global localization.arXiv preprint arXiv:2507.12194, 2025

    Hongming Shen, Xun Chen, Yulin Hui, Zhenyu Wu, Wei Wang, Qiyang Lyu, Tianchen Deng, and Danwei Wang. Unilgl: Learning uniform place recognition for fov- limited/panoramic lidar global localization.arXiv preprint arXiv:2507.12194, 2025. 2

    arXiv 2025

  56. [56]

    Video google: A text retrieval approach to object matching in videos

    Sivic and Zisserman. Video google: A text retrieval approach to object matching in videos. InProceedings ninth IEEE international conference on computer vision, pages 1470–

  57. [57]

    Brief-gist-closing the loop by simple means

    Niko S ¨underhauf and Peter Protzel. Brief-gist-closing the loop by simple means. In2011 IEEE/RSJ International Con- ference on Intelligent Robots and Systems, pages 1234–1241. IEEE, 2011. 2

    2011

  58. [58]

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

    Niko S ¨underhauf, Peer Neubert, and Peter Protzel. Are we there yet? challenging seqslam on a 3000 km journey across all four seasons. InProc. of workshop on long-term auton- omy, IEEE international conference on robotics and automa- tion (ICRA), page 2013. Citeseer, 2013. 5, 10, 13, 14, 15, 16

    2013

  59. [59]

    Inloc: Indoor visual localization with dense matching and view synthesis

    Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Ak- ihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 7199–7209, 2018. 1

    2018

  60. [60]

    Visual place recognition with repetitive structures

    Akihiko Torii, Josef Sivic, Tomas Pajdla, and Masatoshi Okutomi. Visual place recognition with repetitive structures. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 883–890, 2013. 5, 10

    2013

  61. [61]

    Vggt: Vi- sual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Vi- sual geometry grounded transformer. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 5294–5306, 2025. 2, 3, 5, 9

    2025

  62. [62]

    Transvpr: Transformer-based place recognition with multi-level attention aggregation

    Ruotong Wang, Yanqing Shen, Weiliang Zuo, Sanping Zhou, and Nanning Zheng. Transvpr: Transformer-based place recognition with multi-level attention aggregation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13648–13657, 2022. 1

    2022

  63. [63]

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

    Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R Scott. Multi-similarity loss with general pair weighting for deep metric learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5022–5030, 2019. 5

    2019

  64. [64]

    Sfpnet: Sparse fo- cal point network for semantic segmentation on general lidar point clouds

    Yanbo Wang, Wentao Zhao, Chuan Cao, Tianchen Deng, Jingchuan Wang, and Weidong Chen. Sfpnet: Sparse fo- cal point network for semantic segmentation on general lidar point clouds. InEuropean Conference on Computer Vision, pages 403–421. Springer, 2024. 2

    2024

  65. [65]

    Learning to tune like an expert: Interpretable and scene- aware navigation via mllm reasoning and cvae-based adapta- tion.arXiv preprint arXiv:2507.11001, 2025

    Yanbo Wang, Zipeng Fang, Lei Zhao, and Weidong Chen. Learning to tune like an expert: Interpretable and scene- aware navigation via mllm reasoning and cvae-based adapta- tion.arXiv preprint arXiv:2507.11001, 2025. 2

    Pith/arXiv arXiv 2025

  66. [66]

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

    Frederik Warburg, Soren Hauberg, Manuel Lopez- Antequera, Pau Gargallo, Yubin Kuang, and Javier Civera. Mapillary street-level sequences: A dataset for lifelong place recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2626–2635, 2020. 5, 10, 11, 12

    2020

  67. [67]

    Stgcnformer: Spatio-temporal dual-stream graph convolutional networks and transformers for traffic forecasting.IEEE Transactions on Vehicular Technology, 74(10):15254–15263, 2025

    Changzhi Yang, Huihui Pan, and Jue Wang. Stgcnformer: Spatio-temporal dual-stream graph convolutional networks and transformers for traffic forecasting.IEEE Transactions on Vehicular Technology, 74(10):15254–15263, 2025. 2

    2025

  68. [68]

    Trajdiff: Trajectory prediction with diffusion probabilistic models.IEEE Transactions on Image Processing, pages 1– 14, 2025

    Changzhi Yang, Huihui Pan, Jue Wang, and Yuanduo Hong. Trajdiff: Trajectory prediction with diffusion probabilistic models.IEEE Transactions on Image Processing, pages 1– 14, 2025. 2

    2025

  69. [69]

    Visual place recog- nition: A survey from deep learning perspective.Pattern Recognition, 113:107760, 2021

    Xiwu Zhang, Lei Wang, and Yan Su. Visual place recog- nition: A survey from deep learning perspective.Pattern Recognition, 113:107760, 2021. 1

    2021

  70. [70]

    Learning sequence descriptor based on spatio-temporal attention for visual place recognition.IEEE Robotics and Automation Let- ters, 9(3):2351–2358, 2024

    Junqiao Zhao, Fenglin Zhang, Yingfeng Cai, Gengxuan Tian, Wenjie Mu, Chen Ye, and Tiantian Feng. Learning sequence descriptor based on spatio-temporal attention for visual place recognition.IEEE Robotics and Automation Let- ters, 9(3):2351–2358, 2024. 6

    2024

  71. [71]

    The nerfect match: Exploring nerf features for visual localization

    Qunjie Zhou, Maxim Maximov, Or Litany, and Laura Leal- Taix´e. The nerfect match: Exploring nerf features for visual localization. InEuropean Conference on Computer Vision, pages 108–127. Springer, 2024. 1

    2024

  72. [72]

    R2former: Unified retrieval and reranking transformer for place recognition

    Sijie Zhu, Linjie Yang, Chen Chen, Mubarak Shah, Xiao- hui Shen, and Heng Wang. R2former: Unified retrieval and reranking transformer for place recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 19370–19380, 2023. 1 20

    2023