Pith. sign in

REVIEW 4 major objections 5 minor 28 references

TOPO-Bench: An Open-Source Topological Mapping Evaluation Framework with Quantifiable Perceptual Aliasing

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read Localization accuracy can stand in for topological map quality, the paper argues, and a new benchmark with calibrated perceptual aliasing is released to make the claim testable.

desk verdict Useful benchmark and a genuinely new ambiguity taxonomy, but the proof that localization accuracy is a surrogate for topological consistency has a hole in exactly the loop-closure cases the benchmark targets. read the letter →

arxiv 2510.04100 v2 pith:L5RBIEID submitted 2025-10-05 cs.CV cs.AI

classification cs.CVcs.AI
keywords topologicalmappinglocalizationaccuracyconsistencyperceptualaliasingbenchmarkvisualplacerecognitionSLAM-freenavigation
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

Topological maps represent an environment as a graph, but until now there has been no standard way to say whether a particular graph is a good map. This paper's proposal is that one number—how often the system correctly identifies its current node, i.e., localization accuracy—can stand in for the harder-to-measure property of topological consistency: that graph edges correspond to places that are actually close along a traversable route. To make that claim testable, the authors define topological consistency through two conditions, Edge Precision and policy-conditioned Edge Recall, and argue that correct localization decisions preserve both conditions as the graph grows. They further argue that perceptual aliasing is a property of the test data, not just of the tested system, and propose a quantitative measure that labels each test sequence as an unambiguous revisit, an ambiguous revisit, or an ambiguous novel place. The accompanying benchmark, built from real and simulated indoor and outdoor scenes with calibrated ambiguity levels, is used to show that current mapping systems—both classical and learned—only achieve safety on ambiguous novel places by rejecting nearly all revisits.

What carries the argument

The load-bearing object is the pair of graph-vs-environment conditions the paper calls topological consistency: Edge Precision, stating that nodes within n hops in the map are within route distance d in the world, and policy-conditioned Edge Recall, stating that route-close places that the update policy could have connected are indeed within n hops. The inductive link to localization accuracy is what carries the argument: correct accept/reject decisions are shown to preserve both invariants under an edge-length-regular policy with a hop budget n = max(1, floor(epsilon d / mu_e)). The benchmark's second mechanism is the ambiguity ratio, which classifies a test sequence by comparing the simila

What would settle it

Compute Edge Precision and policy-conditioned Edge Recall on a map where edge lengths are deliberately heterogeneous (e.g., alternating very short and very long edges) and compare them against localization accuracy on the same trajectories. Concretely, simulate an ambiguous revisit where the true loop closure is correctly identified but a second route-close place within distance d is never connected; if recall drops while every localization decision is correct, the surrogate claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that topological consistency is the fundamental property of a topological map, and that it can be measured through localization accuracy. Consistency is defined by two invariants: Edge Precision, meaning any nodes within n hops in the graph are within route distance d in the world, and policy-conditioned Edge Recall, meaning any route-close pair of places that the update policy could have connected is connected within n hops. The paper sketches a proof by induction that, under an edge-length-regular policy where every new edge is at most a constant multiple of the median edge length and the hop budget is chosen accordingly, correct accept and correct reject decis

Load-bearing premise

The argument that localization accuracy measures topological consistency rests on the assumption that every new edge is no longer than a fixed multiple of the median edge length and that a correct localization decision at one step does not leave route-close places unconnected; the paper sketches this step but does not prove it in the ambiguous-revisit case.

Editorial extensions

If this is right

  • If localization accuracy is a faithful surrogate, researchers can compare topological mapping systems using a single interpretable number instead of bespoke graph metrics.
  • The ambiguity-ratio classification makes cross-dataset comparisons meaningful, because environments can be matched on measured alias difficulty before scores are compared.
  • The balanced score (geometric mean over unambiguous revisit, ambiguous revisit, and ambiguous novel) rewards systems that neither over-accept nor over-reject, rather than systems tuned to one regime.
  • The reported failure mode—high safety only at the price of near-zero revisit accuracy—becomes a concrete baseline target: a system that disambiguates rather than abstains would be the first to separate the balanced accuracy curve.
  • Open-sourced dataset and toolkit let future methods be tested under the same calibrated ambiguity conditions.

Reading between the lines

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

  • The surrogate claim, if accepted, implies that map-evaluation effort can shift from expensive ground-truth route-distance computation to labeled localization trials, which are cheaper to collect.
  • The ambiguity-ratio definition depends on a chosen visual place recognition model; the paper validates inter-model agreement, but a testable extension would be to compute ambiguity with multiple model families and report variance as part of dataset calibration.
  • The proof's edge-length-regular assumption suggests a natural stress test: construct maps where edge lengths vary widely (short edges inside rooms, long edges across corridors) and see empirically how well localization accuracy tracks consistency beyond the assumption's domain.
  • If the zero-revisit-accuracy finding generalizes, it re-frames perceptual aliasing as a disambiguation problem rather than a rejection problem, pointing toward methods that fuse geometry, semantics, or temporal context explicitly.
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 / 5 minor

Summary. The paper proposes TOPO-Bench, an evaluation framework for SLAM-free topological mapping. It formalizes topological consistency via Edge Precision and policy-conditioned Edge Recall, argues that localization accuracy is a faithful surrogate for these structural properties, introduces a quantitative ambiguity measure based on visual place recognition (VPR) similarity, and curates a multi-dataset benchmark with labeled test cases (A+P, P.O., A.O.). The authors implement several baselines (FAB-MAP, RatSLAM, greedy matching, sequence matching, probabilistic belief update) and report that all methods fail to resolve ambiguous revisits without sacrificing safe rejection. The datasets, baselines, and evaluation tools are intended to be open-sourced.

Significance. If the theoretical claim holds, this would provide the first standardized, ambiguity-aware benchmark for SLAM-free topological mapping, with a concise surrogate metric that avoids expensive structural evaluation. The empirical finding that current methods cannot disambiguate aliased revisits is a useful baseline result. The paper ships a concrete benchmark, baseline implementations, and an ambiguity-labeling protocol, which are valuable community resources. However, the central surrogate theorem is not established for the systems actually evaluated, so the conceptual foundation is weaker than claimed.

major comments (4)
  1. [III-C, Eq. (4), induction case (i)] The proof that correct localization preserves Edge Precision assumes every new edge—including the loop-closure edge added in case (i)—has geodesic length at most κμ_e. But the localization success criterion (Eq. 4) only requires the predicted node to be within distance d of the true node; it does not bound the length of the newly created edge. In a loop closure the revisited node can be arbitrarily far along the route, so the new edge can have length far exceeding κμ_e. Such an edge alone violates Edge Precision. The proof therefore excludes the primary scenario that the benchmark is designed to evaluate. This is load-bearing for the central claim that localization accuracy is a faithful surrogate for topological consistency.
  2. [III-C, definition of n] The hop budget is defined as n = max(1, floor(εd/μ_e)) with ε ≤ 1/κ. When d < κμ_e, floor(εd/μ_e) = 0 and n = 1, so the inequality nκμ_e ≤ d fails. The proof's assertion that any n-hop path has geodesic length ≤ d is then false. The authors note that d is 'usually larger' than μ_e, but the theorem as stated has no such condition. A corrected theorem should state explicit conditions on d, κ, and ε, or use a different definition of n (e.g., ceil) with an adjusted ε.
  3. [III-C, Edge Recall preservation in case (i)] The proof claims that in a correct accept, 'any new eligible pair (in particular (v_i, v_j)) is directly connected.' However, Eq. (4) only guarantees that the retrieved node v_j is within d of the ground-truth node v*_t; it does not require v_j = v*_t. If the pair (v_i, v*_t) is a policy-eligible opportunity, Edge Recall requires dist_G(v_i, v*_t) ≤ n, which is not implied by an edge to v_j. The proof silently identifies the 'correctly retrieved node' with the ground-truth node, contradicting the tolerance allowed by Eq. (4).
  4. [III-D, ambiguity classification] The A+P / P.O. / A.O. labels are generated using BOQ VPR similarity, the same family of machinery used by the evaluated baselines. The inter-method correlation of 0.83 over 200 image pairs is encouraging, but it is computed at the frame-pair level, not for the sequence-level ambiguity categories that define the benchmark. The threshold α and the decision threshold τ are calibrated on BOQ scores, so the dataset labels are not method-independent. This weakens the claim of 'quantifiable perceptual aliasing' as an intrinsic dataset property. Please provide sequence-level label stability across multiple VPR models or a clear argument for why the taxonomy is robust to the choice of similarity model.
minor comments (5)
  1. [III-A] The 'edge-length regular policy' is introduced informally, and the constant κ is first used in III-C. Define κ in III-A and specify whether the bound applies to all edges (including loop closures) or only to consecutive-node edges.
  2. [III-D] The symbol τ is used both for the VPR similarity threshold in the ambiguity definition and for the decision threshold in evaluation. This dual use is confusing; use different notation for the two thresholds.
  3. [Table I] The column headers 'In.' and 'Out.' are not expanded. Add a note defining them as indoor/outdoor.
  4. [IV-A] For the sequence-matching baselines, the half-window size h is not specified in the text or tables. Please report the values used and sensitivity analysis.
  5. [V] The conclusion states code 'will be released in a public GitHub repository,' while the abstract claims the resources are 'fully open-sourced.' If the code is not yet available at submission, rephrase to avoid inconsistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: metrics and ambiguity labels are externally grounded, baselines are evaluated without fitting to the labels, and the only self-citation is introductory and non-load-bearing.

full rationale

The paper's main derivation chain is not circular. The surrogate claim (Sec. III-C) is an inductive argument from correct localization decisions (Eq. 4) to Edge Precision and policy-conditioned Edge Recall (Eqs. 2-3); the conclusion is not contained in the definition of localization success, so the argument is not self-definitional. The ambiguity labels (Sec. III-D) are produced by BOQ VPR similarities, but the paper provides cross-method validation (mean inter-method correlation 0.83 over 11 VPR methods) and the evaluated baselines use a different VPR model (MegaLoc), with decision thresholds selected on a validation set and then applied to the full dataset, so the evaluation is not fitted to the labels. There is a genuine proof gap in Sec. III-C: the sketch assumes every new edge has geodesic length at most kappa*mu_e, but a correct loop-closure accept can create an arbitrarily long edge, and the choice n=max(1,floor(eps*d/mu_e)) does not guarantee n*kappa*mu_e <= d when n=1. This is a mathematical-rigor/correctness concern, not a circular reduction: the theorem's premises do not assume its conclusion. The only self-citation ([1]) appears as an example of mapless navigation in the introduction and is not load-bearing. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via self-citation. Therefore no circular step meets the evidentiary standard.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

Aside from standard math (medians, geometric mean), the framework depends heavily on domain assumptions about route-distance ground truth and the VPR proxy for ambiguity, plus a restrictive update-policy assumption. The surrogate proof adds an unstated stability assumption on µ_e. The four free parameters (τ, α, ϵ, κ) are not reported, so the quantitative claims are conditional on choices invisible in the preprint.

free parameters (4)
  • τ_sim (visual-similarity threshold)
    Used to decide whether a sequence pair counts as visually similar (A.O. requires max_u Sim ≥ τ) and to define ambiguous cases; its value is not reported in the paper.
  • α (ambiguity ratio threshold)
    Separates A+P from P.O. cases via max distractor similarity / true similarity ≥ α; the value is not reported.
  • ϵ (hop-budget tolerance)
    Introduced in hop budget n = max(1, floor(ϵ d / µ_e)); required to satisfy ϵ ≤ 1/κ for the surrogate proof, but no specific value is given.
  • κ (edge-length regularity bound)
    Assumed constant bound on new edge length relative to median edge length in the edge-length-regular policy; not measured or reported.
assumptions (5)
  • domain assumption The update policy P is edge-length regular: each new edge has route length ≤ κ µ_e.
    Restricts the class of systems the surrogate theorem covers (Section III-A); real loop-closure edges can be much longer than consecutive-node edges.
  • domain assumption Route distances are approximated by odometry traversal distances, manually verified to contain no zig-zags.
    Section III-B; odometry drift and unverified loops could bias dist_R and hence all metrics.
  • domain assumption VPR similarities from BOQ are a valid proxy for perceptual aliasing, supported by mean inter-method correlation 0.83 on 200 pairs.
    Section III-D; 0.83 correlation over 11 methods is moderate, and the 200-pair validation is small.
  • domain assumption Ground-truth correspondence π(i) from GPS or motion capture is exact.
    Section III-D; errors in alignment propagate into case labels and accuracy measurements.
  • ad hoc to paper During the surrogate induction, the median edge length µ_e and hop budget n remain constant as the graph grows, and each step adds at most one edge.
    Unstated in Section III-C; if µ_e changes or multiple edges/eligible nodes appear, the invariant proof breaks.
invented entities (2)
  • A+P / P.O. / A.O. ambiguity case taxonomy
    purpose: Classifies each test sequence by whether a true revisit exists and whether a distractor is nearly as similar as the true match, enabling per-failure-mode accuracy reporting.
    The taxonomy is defined through VPR-model similarities and thresholds (τ, α) not reported in the paper; without those values the categories cannot be reconstructed, and they inherit the VPR proxy's biases.
  • Balanced Localization Accuracy (BLA)
    purpose: Aggregates per-case accuracies as a geometric mean to penalize performance imbalance across case types.
    A summary metric rather than a physical entity; its interpretation depends on the three case accuracies being commensurable, which is not established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TOPO-Bench: An Open-Source Topological Mapping Evaluation Framework with Quantifiable Perceptual Aliasing." pith.science (2026). https://pith.science/paper/L5RBIEID

@misc{pith2026251004100,
  author       = {Pith},
  title        = {Pith review of: TOPO-Bench: An Open-Source Topological Mapping Evaluation Framework with Quantifiable Perceptual Aliasing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L5RBIEID}},
  note         = {Machine review of arXiv:2510.04100}
}
read the original abstract

Topological mapping offers a compact and robust representation for navigation, but progress in the field is hindered by the lack of standardized evaluation metrics, datasets, and protocols. Existing systems are assessed using different environments and criteria, preventing fair and reproducible comparisons. Moreover, a key challenge - perceptual aliasing - remains under-quantified, despite its strong influence on system performance. We address these gaps by (1) formalizing topological consistency as the fundamental property of topological maps and showing that localization accuracy provides an efficient and interpretable surrogate metric, and (2) proposing the first quantitative measure of dataset ambiguity to enable fair comparisons across environments. To support this protocol, we curate a diverse benchmark dataset with calibrated ambiguity levels, implement and release deep-learned baseline systems, and evaluate them alongside classical methods. Our experiments and analysis yield new insights into the limitations of current approaches under perceptual aliasing. All datasets, baselines, and evaluation tools are fully open-sourced to foster consistent and reproducible research in topological mapping.

Figures

Figures reproduced from arXiv: 2510.04100 by the authors.

Figure 1
Figure 1. Illustration of topological consistency. Nodes [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Examples of map, test, and ambiguous cases across datasets. Top row: A+P case from OpenLORIS [19], where the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Examples of evaluation scenarios with quantified [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Localization accuracy for A+P, P.O., and A.O. test cases, and BLA, plotted against the decision threshold for all [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 2 linked inside Pith

  1. [1]

    Ge- nie: A generalizable navigation system for in-the-wild environments,

    J. Wang, D. Liu, J. Chen, J. Da, N. Qian, T. M. Man, and H. Soh, “Ge- nie: A generalizable navigation system for in-the-wild environments,” arXiv preprint arXiv:2506.17960, 2025

  2. [2]

    Nomad: Goal masked diffusion policies for navigation and exploration,

    A. Sridhar, D. Shah, C. Glossop, and S. Levine, “Nomad: Goal masked diffusion policies for navigation and exploration,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 63–70

  3. [3]

    Intelligence without representation,

    R. A. Brooks, “Intelligence without representation,”Artificial intelli- gence, vol. 47, no. 1-3, pp. 139–159, 1991

  4. [4]

    Topological simultane- ous localization and mapping: a survey,

    J. Boal, A. S ´anchez-Miralles, and A. Arranz, “Topological simultane- ous localization and mapping: a survey,”Robotica, vol. 32, no. 5, pp. 803–821, 2014

  5. [5]

    Mapping a suburb with a single camera using a biologically inspired slam system,

    M. J. Milford and G. F. Wyeth, “Mapping a suburb with a single camera using a biologically inspired slam system,”IEEE Transactions on Robotics, vol. 24, no. 5, pp. 1038–1053, 2008

  6. [6]

    Highly scalable appearance-only slam- fab-map 2.0,

    M. Cummins and P. Newman, “Highly scalable appearance-only slam- fab-map 2.0,”Robotics: Science and Systems V, 2009

  7. [7]

    Semi-parametric topolog- ical memory for navigation,

    N. Savinov, A. Dosovitskiy, and V . Koltun, “Semi-parametric topolog- ical memory for navigation,”arXiv preprint arXiv:1803.00653, 2018

  8. [8]

    Fab- map+ ratslam: Appearance-based slam for multiple times of day,

    A. J. Glover, W. P. Maddern, M. J. Milford, and G. F. Wyeth, “Fab- map+ ratslam: Appearance-based slam for multiple times of day,” in2010 IEEE international conference on robotics and automation. IEEE, 2010, pp. 3507–3512

Show all 28 references
  1. [9]

    Cat-slam: probabilistic local- isation and mapping using a continuous appearance-based trajectory,

    W. Maddern, M. Milford, and G. Wyeth, “Cat-slam: probabilistic local- isation and mapping using a continuous appearance-based trajectory,” The International Journal of Robotics Research, vol. 31, no. 4, pp. 429–451, 2012

  2. [10]

    Probabilistic visual place recognition for hierarchical localization,

    M. Xu, N. Snderhauf, and M. Milford, “Probabilistic visual place recognition for hierarchical localization,”IEEE Robotics and Automa- tion Letters, vol. 6, no. 2, pp. 311–318, 2020

  3. [11]

    Placenav: Topological navigation through place recognition,

    L. Suomela, J. Kalliola, H. Edelman, and J.-K. K ¨am¨ar¨ainen, “Placenav: Topological navigation through place recognition,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 5205–5213

  4. [12]

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

    A. Ali-Bey, B. Chaib-draa, and P. Giguere, “Boq: A place is worth a bag of learnable queries,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17 794–17 803

  5. [13]

    Scaling local control to large-scale topological navigation,

    X. Meng, N. Ratliff, Y . Xiang, and D. Fox, “Scaling local control to large-scale topological navigation,” in2020 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2020, pp. 672–678

  6. [14]

    Topomap: Topological mapping and navigation based on visual slam maps,

    F. Bl ¨ochliger, M. Fehr, M. Dymczyk, T. Schneider, and R. Siegwart, “Topomap: Topological mapping and navigation based on visual slam maps,” inProc. IEEE International Conference on Robotics and Automation (ICRA), 2018, pp. 3818–3825

  7. [15]

    Sparse 3d topo- logical graphs for micro-aerial vehicle planning,

    H. Oleynikova, Z. Taylor, R. Siegwart, and J. Nieto, “Sparse 3d topo- logical graphs for micro-aerial vehicle planning,” inProc. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018

  8. [16]

    Navigation in hybrid metric–topological maps,

    K. Konolige, E. Marder-Eppstein, and B. Marthi, “Navigation in hybrid metric–topological maps,” inProc. IEEE International Con- ference on Robotics and Automation (ICRA), 2011, pp. 3041–3047

  9. [17]

    V oronoi random fields: Ex- tracting the topological structure of indoor environments via place labeling,

    S. Friedman, H. Pasula, and D. Fox, “V oronoi random fields: Ex- tracting the topological structure of indoor environments via place labeling,” inProc. Int. Joint Conf. on Artificial Intelligence (IJCAI), 2007

  10. [18]

    Towards autonomous topo- logical place detection using the extended voronoi graph,

    P. Beeson, N. K. Jong, and B. Kuipers, “Towards autonomous topo- logical place detection using the extended voronoi graph,” inProc. IEEE International Conference on Robotics and Automation (ICRA), 2005, pp. 4373–4379

  11. [19]

    Are we ready for service robots? the OpenLORIS-Scene datasets for lifelong SLAM,

    X. Shi, D. Li, P. Zhao, Q. Tian, Y . Tian, Q. Long, C. Zhu, J. Song, F. Qiao, L. Song, Y . Guo, Z. Wang, Y . Zhang, B. Qin, W. Yang, F. Wang, R. H. M. Chan, and Q. She, “Are we ready for service robots? the OpenLORIS-Scene datasets for lifelong SLAM,” in2020 International Conf...

  12. [20]

    1 Year, 1000km: The Oxford RobotCar Dataset,

    W. Maddern, G. Pascoe, C. Linegar, and P. Newman, “1 Year, 1000km: The Oxford RobotCar Dataset,”The International Journal of Robotics Research (IJRR), vol. 36, no. 1, pp. 3–15, 2017. [Online]. Available: http://dx.doi.org/10.1177/0278364916679498

  13. [21]

    Rellis-3d dataset: Data, benchmarks and analysis,

    P. Jiang, P. Osteen, M. Wigness, and S. Saripalli, “Rellis-3d dataset: Data, benchmarks and analysis,”arXiv preprint arXiv:2011.12954, 2020

  14. [22]

    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

  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

  16. [24]

    Rawseeds: Building a benchmarking toolkit for autonomous robotics,

    G. Fontana, M. Matteucci, and D. G. Sorrenti, “Rawseeds: Building a benchmarking toolkit for autonomous robotics,” inMethods and experimental techniques in computer engineering. Springer, 2013, pp. 55–68

  17. [25]

    Habitat: A Platform for Embodied AI Research,

    M. Savva, A. Kadian, O. Maksymets, Y . Zhao, E. Wijmans, B. Jain, J. Straub, J. Liu, V . Koltun, J. Malik, D. Parikh, and D. Batra, “Habitat: A Platform for Embodied AI Research,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019

  18. [26]

    Rover: A multiseason dataset for visual slam,

    F. Schmidt, J. Daubermann, M. Mitschke, C. Blessing, S. Meyer, M. Enzweiler, and A. Valada, “Rover: A multiseason dataset for visual slam,”IEEE Transactions on Robotics, vol. 41, pp. 4005–4022, 2025

  19. [27]

    Seqslam: Visual route-based naviga- tion for sunny summer days and stormy winter nights,

    M. J. Milford and G. F. Wyeth, “Seqslam: Visual route-based naviga- tion for sunny summer days and stormy winter nights,” in2012 IEEE international conference on robotics and automation. IEEE, 2012, pp. 1643–1649

  20. [28]

    Megaloc: One retrieval to place them all,

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

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.