Pith. sign in

REVIEW 4 major objections 4 minor 30 references

MSG-Loc shows that preserving multi-label hypotheses for both map landmarks and query detections, then propagating the best neighbor likelihoods over a semantic graph, yields more reliable object-level global localization than single-label

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 18:45 UTC pith:6FKEAXYF

load-bearing objection Solid, modest extension of single-label semantic graph matching; the exact-match likelihood in Eq. (3) is the main thing to probe, and the evaluation lacks variance, but the core idea works in tested setups. the 4 major comments →

arxiv 2512.03522 v4 pith:6FKEAXYF submitted 2025-12-03 cs.RO cs.CV

MSG-Loc: Multi-Label Likelihood-based Semantic Graph Matching for Object-Level Global Localization

Gihyeon Lee , Jungwoo Lee , Juwon Kim , Young-Sik Shin , Younggun Cho This is my paper
classification cs.RO cs.CV
keywords multi-label likelihoodsemantic graph matchingglobal localizationobject-level SLAMopen-set detectiondata associationsemantic ambiguitycontext-aware propagation
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 proposes a way for a robot to relocalize itself on a prior object map when the detector is ambiguous — when a single object could, say, be a cup, mug, or glass. Its central claim is that instead of forcing one label per object, preserving the full top-K label distribution for both map landmarks and live detections — and then propagating these likelihoods through a k-nearest-neighbor semantic graph — yields more accurate object associations and pose estimates than single-label graph descriptors. The method is detector-agnostic: it works with a supervised closed-set detector and with an open-set zero-shot detector, and it scales from an 80-class vocabulary to a 1,203-class vocabulary without retraining. A sympathetic reader would care because semantic ambiguity is a main obstacle to object-level global localization in real environments, and the reported gains — roughly 10.6% higher F1 and 33.4% higher MOTA over the best single-label baseline, plus higher success rates at strict pose thresholds — indicate that preserving uncertainty, rather than just improving features, is what carries the improvement.

Core claim

The method preserves multi-label hypotheses on both sides of matching: each landmark keeps a frequency distribution over labels seen across keyframes, and each query detection keeps a normalized top-K confidence distribution. The likelihood of a correspondence is the sum, over exactly matching label strings, of landmark frequency times query confidence — a score that does not depend on a fixed label vocabulary and so spans closed-set and open-set detectors with 80 to 1,203 classes. This score is augmented by the maximum weighted likelihood among one-hop neighbors, so a correspondence is trusted only when its neighborhood agrees. The paper reports average relative gains of 10.6% in F1 and 33.

What carries the argument

The load-bearing object is the multi-label likelihood f(X_ij | θ^p_i, θ^q_j) = Σ_u Σ_v 1[l^p_{i,u}=l^q_{j,v}] · ν_{i,u} · ĉ_{j,v} — a sum that pairs each label the map has seen on a landmark with each label the detector currently entertains, weighted by how often that label appeared in the map and how confident the detector is. The second mechanism, Context-Aware Likelihood Propagation (CALP), adds to this root likelihood the maximum weighted likelihood among the one-hop neighbors, with weights that decay with relative distance mismatch. CALP is what lets a sparse, ambiguous query graph borrow confidence from its neighborhood, suppressing the isolated misclassifications that sink single-labe

Load-bearing premise

The matching signal only contributes when the exact same label string appears in both the landmark's accumulated label set and the query's top-K list; if the detector words never coincide (cup versus mug), the likelihood term is zero and the method falls back on geometry alone.

What would settle it

Build a sequence where the query detector consistently uses a synonym absent from the prior map's accumulated labels (the map says 'cup', every query detection says 'mug'), then run the full pipeline. If multi-label likelihood is the source of the reported gains, data-association F1 should drop to the level of the geometric baseline on those frames; if it does not drop, the improvements must be coming from the graph-propagation or pose-estimation stages rather than from label matching.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A robot using this matcher can relocalize even when its detector is uncertain: on the evaluated sequences, F1 and MOTA improve over single-label graph descriptors in both closed-set and open-set configurations.
  • Because the likelihood is not tied to a fixed label vocabulary, expanding the class set from 80 to 1,203 categories does not degrade the method; the paper reports it improves or holds performance, while single-label baselines tend to worsen.
  • The pipeline runs in real time (over 36 Hz), so preserving multi-label uncertainty is cheap enough for online loop closure and relocalization.
  • Adding the semantic signal to a feature-based SLAM pipeline raises the all-frame pose success rate on ambiguous sequences, indicating the two information sources are complementary.

Where Pith is reading between the lines

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

  • The same multi-label likelihood could accept label distributions from any source — a captioning model, a vision-language embedding retriever, or a class-agnostic proposal network — since the formula does not care how the distribution was produced; that is a natural next step the authors flag only as future work.
  • The method's Achilles' heel is exact string equality; a cheap, testable modification is to replace the indicator in the likelihood with a soft similarity between label embeddings, which would let 'cup' and 'mug' contribute rather than zero out.
  • Read as a design principle, the paper suggests that semantic-SLAM systems should postpone hard label decisions as long as possible — carrying the full belief distribution through matching — which is likely to become more important as robots encounter open-world vocabularies.

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

4 major / 4 minor

Summary. MSG-Loc proposes an object-level global localization method that replaces single-label semantic graph descriptors with multi-label likelihoods. The prior graph accumulates per-landmark label frequencies from top-K detections across keyframes; the query graph stores normalized top-K confidence distributions. The core matching score (Eq. 3) is a label-overlap weighted sum, and Context-Aware Likelihood Propagation (CALP, Eq. 6) adds the best weighted 1-hop neighbor likelihoods. Candidate node pairs are extracted by top-τ ranking and a stochastic P3P/RANSAC pose solver is used. Experiments on TUM RGB-D, ICL-LM, and a custom resort dataset compare with SH and GOReloc under YOLOv8 (closed-set) and Grounding DINO + OVSAM/TAP (open-set) conditions; the paper reports 10.6% relative F1 and 33.4% relative MOTA improvements on average, plus ablation studies on K and τ.

Significance. The central idea of retaining multi-label hypotheses rather than committing to a single label is practical and could benefit semantic SLAM and relocalization. The paper also provides a reasonably broad evaluation: multiple datasets, closed- and open-set detectors, two label vocabularies, and quantitative association and pose metrics. Code and supplementary materials are promised, which is a strength. However, the semantic likelihood is based on exact label-string equality, so the open-set advantage is conditional on label overlap that is not measured; the stochastic pose solver is evaluated with single-run point estimates; and a few of the paper's own results contradict the 'consistent outperformance' narrative. These issues need to be addressed before the claims can be accepted.

major comments (4)
  1. [Sec. III-C, Eq. (3)] The semantic likelihood is zero unless a query top-K label is exactly string-equal to a prior label. In the open-set conditions the paper targets, synonyms or viewpoint-driven label drift (cup/mug/glass) can make the true correspondence's label sets disjoint, reducing matching to geometry. The authors themselves defer VLM embeddings to future work (Sec. V). Please report, for correct correspondences in Tables I/II, how often the top-K query labels intersect the map label set, and test a cross-vocabulary or synonym scenario; otherwise the claimed advantage over single-label descriptors is conditional on an unmeasured overlap assumption.
  2. [Algorithm 1, Sec. IV-C/D] Pose estimation is stochastic (random triple sampling, P3P, N_iter), but Tables I–III report single-run point estimates with no seeds, error bars, or significance tests. Some margins are small (e.g., Table II Fr2person Y: F1 0.894 vs 0.875; Resort2 Y: MOTA 0.620 vs 0.604). The claim that MSG-Loc 'consistently outperforms' the baselines is not supported without repeated-run statistics; please report mean/std over multiple seeds for the proposed method and, where feasible, the baselines.
  3. [Table II] Table II (Fr2person, Resort2, Y rows) shows GOReloc with higher F1 and MOTA than MSG-Loc (0.894/0.776 vs 0.875/0.730; 0.820/0.620 vs 0.815/0.604). This is in the closed-set LVIS condition, one of the configurations used to support scalability. The text in Sec. IV-C1 says MSG-Loc 'consistently outperforms' only for Table I, but Sec. IV-C2 and IV-D1 generalize the success narrative. Please discuss these exceptions explicitly and avoid 'consistent' claims that cover Table II.
  4. [Sec. IV-A3 and Eq. (8)] K, τ, C, detection thresholds, and N_iter are chosen empirically. K and τ receive ablations (Tables IV/V), but C (Wasserstein scale) and detection thresholds do not, and N_iter is not stated. Without a sensitivity analysis for these parameters, it is unclear whether the reported gains reflect a robust property of multi-label matching or tuning. Please add at least a C sweep and report N_iter/threshold settings.
minor comments (4)
  1. [Sec. III-D, Eq. (5)] The set N_r^* is defined by a union of argmax over (o_n,d_m); clarify tie-breaking and whether one prior node can be selected for multiple query neighbors.
  2. [Table III] ORB-SLAM2 rows show identical success rates at all thresholds for several sequences (e.g., Fr2 person 8.68), which needs explanation; likely only a subset of frames are evaluated.
  3. [Sec. IV-B3] The definitions of SR_succ and SR_all are ambiguous. Please define what counts as a 'success frame' before the threshold is applied.
  4. [Sec. IV-F] The runtime claim 'above 36.4 Hz' is based on total milliseconds on one CPU; include per-sequence or repeated-run statistics for reproducibility.

Circularity Check

0 steps flagged

No significant circularity: MSG-Loc's likelihoods are constructed from observed detector outputs and validated against external baselines and ground truth.

full rationale

The paper's derivation chain is empirical rather than definitional. The prior multi-label frequencies (Eq. 1) and query confidence distributions (Eq. 2) are computed directly from detector outputs, not from the target association or pose. The matching likelihood (Eq. 3) is a defined scoring function, and the propagated similarity (Eq. 6) and candidate selection (Eq. 7) rely on that score, but the final pose is estimated by P3P and then independently verified by projected bounding-box Wasserstein alignment (Eqs. 8-9), with success measured against ground-truth poses (Eq. 12). No equation reintroduces the result being explained, and no parameter is fitted to the reported F1, MOTA, or TE metrics in a way that forces the outcome. The only self-references are ordinary citations to external prior work such as GOReloc [11], whose authors do not overlap with the present paper, and no uniqueness theorem or load-bearing premise is imported from the authors' own prior work. The skeptical concern about exact label-string equality in Eq. (3) is a robustness and correctness limitation for synonym/vocabulary-mismatch scenarios, not a circularity: the method could fail there, but it is not true by construction. Overall, the central claim is supported by comparison against external baselines on held-out sequences, so the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on treating multi-label detector outputs as likelihoods, on the geometric reliability of the object SLAM map, and on stability of KNN graph topology. No new physical entities are introduced. The main unresolved ledger items are tuning constants (K, tau, C), unspecified graph/iterative parameters (k, N_iter), and the top-K renormalization that converts raw confidences into pseudo-likelihoods.

free parameters (6)
  • Top-K labels (K) = 5
    Number of labels retained per query detection and for prior label aggregation; set 'empirically based on preliminary experiments' (Sec. IV-A3) and ablated in Table V; central to the multi-label representation.
  • Candidate pair limit (tau) = 3
    Top-tau prior nodes per query node used in Eq. (7); empirically set (Sec. IV-A3) and ablated in Table IV.
  • Wasserstein scale factor C = 100
    Scale in normalized Wasserstein distance Eq. (8); set empirically for all experiments (Sec. IV-A3).
  • KNN neighbor count k = not stated
    Both prior and query graphs are built as k-nearest-neighbor graphs (Sec. III-B), but the value of k is never given; it controls graph topology for CALP and the edge-consistency constraint in Algorithm 1.
  • RANSAC iterations N_iter = not stated
    Input to Algorithm 1, not specified in the implementation details; affects pose estimation stability and runtime.
  • Detection confidence thresholds / NMS overlap = 0.1 (YOLOv8), 0.14 (Grounding DINO), NMS 0.6
    Set in Sec. IV-A3; control which objects enter the query graph and thus affect all downstream metrics.
axioms (5)
  • domain assumption Node-pair likelihoods are pairwise independent ('Under the assumption of pairwise independence for tractability')
    Sec. III-C, Eq. (3): the method treats each object-pair score independently before CALP adds neighbor context; no joint probabilistic model is defined.
  • domain assumption Exact label-string equality is semantically meaningful
    The indicator function in Eq. (3) counts only identical label strings; no word/semantic embedding is used, so synonyms or cross-vocabulary aliases can zero out the likelihood.
  • domain assumption Prior map dual quadrics and graph positions are accurate enough for KNN and Euclidean distance comparisons
    The prior graph is inherited from QuadricSLAM [17] (Sec. IV-A3); all matching assumes positions and object shapes from the map are reliable.
  • domain assumption Normalized top-K confidences are commensurable with accumulated detection frequencies
    Eq. (2) renormalizes confidences over the top-K set; Eq. (3) multiplies these with frequencies as if both are calibrated likelihoods across detector types and label sets.
  • domain assumption Corresponding KNN subgraphs exist between sparse query and dense prior graphs
    Algorithm 1 requires the sampled three pairs to preserve 'identical edge connectivity', and CALP assumes 1-hop neighbor sets can be matched; this can fail with very sparse or differently distributed detections.

pith-pipeline@v1.3.0-alltime-deepseek · 13789 in / 16452 out tokens · 142043 ms · 2026-08-03T18:45:15.847696+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of MSG-Loc: Multi-Label Likelihood-based Semantic Graph Matching for Object-Level Global Localization." pith.science (2026). https://pith.science/paper/6FKEAXYF

@misc{pith2026251203522,
  author       = {Pith},
  title        = {Pith review of: MSG-Loc: Multi-Label Likelihood-based Semantic Graph Matching for Object-Level Global Localization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6FKEAXYF}},
  note         = {Machine review of arXiv:2512.03522}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Robots are often required to localize in environments with unknown object classes and semantic ambiguity. However, when performing global localization using semantic objects, high semantic ambiguity intensifies object misclassification and increases the likelihood of incorrect associations, which in turn can cause significant errors in the estimated pose. Thus, in this letter, we propose a multi-label likelihood-based semantic graph matching framework for object-level global localization. The key idea is to exploit multi-label graph representations, rather than single-label alternatives, to capture and leverage the inherent semantic context of object observations. Based on these representations, our approach enhances semantic correspondence across graphs by combining the likelihood of each node with the maximum likelihood of its neighbors via context-aware likelihood propagation. For rigorous validation, data association and pose estimation performance are evaluated under both closed-set and open-set detection configurations. In addition, we demonstrate the scalability of our approach to large-vocabulary object categories in both real-world indoor scenes and synthetic environments. Project Page: https://sparolab.github.io/research/msg-loc/.

Figures

Figures reproduced from arXiv: 2512.03522 by Gihyeon Lee, Jungwoo Lee, Juwon Kim, Younggun Cho, Young-Sik Shin.

Figure 1
Figure 1. Figure 1: MSG-Loc. Our method preserves multi-label hypotheses for both landmarks and observed objects, and estimates their likelihoods. It propa￾gates context-aware likelihoods and calculates similarity across the semantic graph to mitigate viewpoint-dependent misclassifications. This enables reliable object-level global localization even in scenes of severe semantic ambiguity. variations. In contrast, object-level… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our method. In the Multi-Label Graph Construction phase, we generate a k-nearest neighbor (KNN)-based semantic graph, embedding multi-label detection frequencies and normalized confidence scores via the Prior Graph Management (PGM) and Query Graph Management (QGM) modules. Multi-Label Likelihood Estimation (M-LLE) then computes semantic likelihoods from these attributes for both landmarks and o… view at source ↗
Figure 3
Figure 3. Figure 3: Multi-label likelihood estimation algorithm. This figure illustrates how M-LLE computes semantic likelihoods between landmark and observa￾tion nodes by aggregating detection frequencies and normalized confidence scores. The probabilistic estimation of likelihoods enables data association that is agnostic to the object classification paradigm within our framework. where ηi is the total number of detections … view at source ↗
Figure 4
Figure 4. Figure 4: Frame-wise semantic uncertainty analysis. Shannon entropy is calculated for each object using the normalized confidence scores of its top-5 predicted LVIS [27] labels. The closed-set detector (YOLOv8 [25]) consistently produces low entropy, as it tends to allocate most of its confidence to a single predicted label. Conversely, the open-set detector with a zero-shot classifier (TAP [30]) produces a confiden… view at source ↗
Figure 5
Figure 5. Figure 5: The qualitative results of data association and pose estimation. (a) For each pair, the query frame image is presented on the left, and the corresponding landmark rendering is on the right. Green lines represent correct associations, whereas red lines indicate incorrect ones. (b) The color bar on the right denotes the magnitude of translation error. Scatter points represent successful pose estimates and ar… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

30 extracted references · 3 linked inside Pith

  1. [1]

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

    C. Cadena, L. Carlone, H. Carrillo, Y . Latif, D. Scaramuzza, J. Neira, I. Reid, and J. J. Leonard, “Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age,”IEEE Transactions on robotics, vol. 32, no. 6, pp. 1309–1332, 2016

    2016

  2. [2]

    Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras,

    R. Mur-Artal and J. D. Tard ´os, “Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras,”IEEE Transactions on Robotics, vol. 33, no. 5, pp. 1255–1262, 2017

    2017

  3. [3]

    Oa-slam: Leveraging objects for camera relocalization in visual slam,

    M. Zins, G. Simon, and M.-O. Berger, “Oa-slam: Leveraging objects for camera relocalization in visual slam,” in2022 IEEE international symposium on mixed and augmented reality (ISMAR). IEEE, 2022, pp. 720–728

    2022

  4. [4]

    LoFTR: Detector-free local feature matching with transformers,

    J. Sun, Z. Shen, Y . Wang, H. Bao, and X. Zhou, “LoFTR: Detector-free local feature matching with transformers,”CVPR, 2021

    2021

  5. [5]

    Not all views are created equal: Analyzing viewpoint instabilities in vision foundation models,

    M. Michalkiewicz, S. Bai, M. Baktashmotlagh, V . Jampani, and G. Balakrishnan, “Not all views are created equal: Analyzing viewpoint instabilities in vision foundation models,”arXiv preprint arXiv:2412.19920, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Object-based loop closure with directional histogram descriptor,

    B. Zhou, Y . Meng, and F. Kai, “Object-based loop closure with directional histogram descriptor,” in2022 IEEE 18th International Conference on Automation Science and Engineering (CASE). IEEE, 2022, pp. 1346–1351

    2022

  7. [7]

    Semantictopoloop: Semantic loop closure with 3d topological graph based on quadric-level object map,

    Z. Cao, Q. Zhang, J. Guang, S. Wu, Z. Hu, and J. Liu, “Semantictopoloop: Semantic loop closure with 3d topological graph based on quadric-level object map,”IEEE Robotics and Automation Letters, 2024

    2024

  8. [8]

    An object slam framework for association, mapping, and high-level tasks,

    Y . Wu, Y . Zhang, D. Zhu, Z. Deng, W. Sun, X. Chen, and J. Zhang, “An object slam framework for association, mapping, and high-level tasks,”IEEE Transactions on Robotics, vol. 39, no. 4, pp. 2912–2932, 2023

    2023

  9. [9]

    X-view: Graph-based semantic multi-view localization,

    A. Gawel, C. D. Don, R. Siegwart, J. Nieto, and C. Cadena, “X-view: Graph-based semantic multi-view localization,”IEEE Robotics and Automation Letters, vol. 3, no. 3, pp. 1687–1694, 2018

    2018

  10. [10]

    Semantic histogram based graph matching for real-time multi-robot global localization in large scale environment,

    X. Guo, J. Hu, J. Chen, F. Deng, and T. L. Lam, “Semantic histogram based graph matching for real-time multi-robot global localization in large scale environment,” IEEE Robotics and Automation Letters, vol. 6, no. 4, pp. 8349–8356, 2021

    2021

  11. [11]

    Goreloc: Graph-based object-level relocalization for visual slam,

    Y . Wang, C. Jiang, and X. Chen, “Goreloc: Graph-based object-level relocalization for visual slam,”IEEE Robotics and Automation Letters, vol. 9, no. 10, pp. 8234– 8241, 2024

    2024

  12. [12]

    Clip-loc: Multi-modal landmark association for global localization in object-based maps,

    S. Matsuzaki, T. Sugino, K. Tanaka, Z. Sha, S. Nakaoka, S. Yoshizawa, and K. Shintani, “Clip-loc: Multi-modal landmark association for global localization in object-based maps,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 13 673–13 679

    2024

  13. [13]

    Clip-clique: Graph-based correspon- dence matching augmented by vision language models for object-based global localization,

    S. Matsuzaki, K. Tanaka, and K. Shintani, “Clip-clique: Graph-based correspon- dence matching augmented by vision language models for object-based global localization,”IEEE Robotics and Automation Letters, 2024

    2024

  14. [14]

    Slam++: Simultaneous localisation and mapping at the level of objects,

    R. F. Salas-Moreno, R. A. Newcombe, H. Strasdat, P. H. Kelly, and A. J. Davison, “Slam++: Simultaneous localisation and mapping at the level of objects,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2013, pp. 1352–1359

    2013

  15. [15]

    Cubeslam: Monocular 3-d object slam,

    S. Yang and S. Scherer, “Cubeslam: Monocular 3-d object slam,”IEEE Transac- tions on Robotics, vol. 35, no. 4, pp. 925–938, 2019

    2019

  16. [16]

    Monocular object and plane slam in structured envi- ronments,

    S. Yang and S. Scherer, “Monocular object and plane slam in structured envi- ronments,”IEEE Robotics and Automation Letters, vol. 4, no. 4, pp. 3145–3152, 2019

    2019

  17. [17]

    Quadricslam: Dual quadrics from object detections as landmarks in object-oriented slam,

    L. Nicholson, M. Milford, and N. S ¨underhauf, “Quadricslam: Dual quadrics from object detections as landmarks in object-oriented slam,”IEEE Robotics and Automation Letters, vol. 4, no. 1, pp. 1–8, 2019

    2019

  18. [18]

    Eao-slam: Monocular semi-dense object slam based on ensemble data association,

    Y . Wu, Y . Zhang, D. Zhu, Y . Feng, S. Coleman, and D. Kerr, “Eao-slam: Monocular semi-dense object slam based on ensemble data association,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 4966–4973

    2020

  19. [19]

    Qiso-slam: Object-oriented slam using dual quadrics as landmarks based on instance segmentation,

    Y . Wang, B. Xu, W. Fan, and C. Xiang, “Qiso-slam: Object-oriented slam using dual quadrics as landmarks based on instance segmentation,”IEEE Robotics and Automation Letters, vol. 8, no. 4, pp. 2253–2260, 2023

    2023

  20. [20]

    V oom: Robust visual object odometry and mapping using hierarchical landmarks,

    Y . Wang, C. Jiang, and X. Chen, “V oom: Robust visual object odometry and mapping using hierarchical landmarks,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 10 298–10 304

    2024

  21. [21]

    Robust visual place recognition with graph kernels,

    E. Stumm, C. Mei, S. Lacroix, J. Nieto, M. Hutter, and R. Siegwart, “Robust visual place recognition with graph kernels,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 4535–4544

    2016

  22. [22]

    A normalized gaussian wasserstein distance for tiny object detection,

    J. Wang, C. Xu, W. Yang, and L. Yu, “A normalized gaussian wasserstein distance for tiny object detection,”arXiv preprint arXiv:2110.13389, 2021

    Pith/arXiv arXiv 2021

  23. [23]

    A benchmark for the evaluation of rgb-d slam systems,

    J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers, “A benchmark for the evaluation of rgb-d slam systems,” in2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2012, pp. 573–580

    2012

  24. [24]

    Characterizing visual localization and mapping datasets,

    S. Saeedi, E. D. C. Carvalho, W. Li, D. Tzoumanikas, S. Leutenegger, P. H. J. Kelly, and A. J. Davison, “Characterizing visual localization and mapping datasets,” in2019 International Conference on Robotics and Automation (ICRA), 2019, pp. 6699–6705

    2019

  25. [25]

    Ultralytics yolov8,

    G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics yolov8,” 2023. [Online]. Available: https://github.com/ultralytics/ultralytics

    2023

  26. [26]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” inComputer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. Springer, 2014, pp. 740–755

    2014

  27. [27]

    Lvis: A dataset for large vocabulary instance segmentation,

    A. Gupta, P. Dollar, and R. Girshick, “Lvis: A dataset for large vocabulary instance segmentation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 5356–5364

    2019

  28. [28]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, C. Li, J. Yang, H. Su, J. Zhu et al., “Grounding dino: Marrying dino with grounded pre-training for open-set object detection,”arXiv preprint arXiv:2303.05499, 2023

    Pith/arXiv arXiv 2023

  29. [29]

    Open-vocabulary sam: Segment and recognize twenty-thousand classes interactively,

    H. Yuan, X. Li, C. Zhou, Y . Li, K. Chen, and C. C. Loy, “Open-vocabulary sam: Segment and recognize twenty-thousand classes interactively,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 419–437

    2024

  30. [30]

    Tokenize anything via prompting,

    T. Pan, L. Tang, X. Wang, and S. Shan, “Tokenize anything via prompting,” in European Conference on Computer Vision. Springer, 2024, pp. 330–348

    2024