Pith. sign in

REVIEW 5 major objections 6 minor 2 cited by

HeLiOS: Heterogeneous LiDAR Place Recognition via Overlap-based Learning and Local Spherical Transformer

T0 review · 5 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read HeLiOS claims a single shared network can recognize places across heterogeneous LiDAR sensors and beats prior models on public benchmarks.

desk verdict A credible first deep network for heterogeneous LiDAR place recognition, with a real evaluation bias to fix: the same overlap function supervises HeLiOS and judges all methods. read the letter →

arxiv 2501.18943 v2 pith:RW3DYYIQ submitted 2025-01-31 cs.RO

classification cs.RO
keywords LiDARplacerecognitionheterogeneoussphericaltransformeroptimaltransportaggregationoverlap-basedmetriclearningguidedtripletlossglobaldescriptorsrobotics
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

This paper tries to establish that one deep network, HeLiOS, can recognize places from single LiDAR scans regardless of the sensor type, whether a 360-degree spinning LiDAR or a narrow-field-of-view solid-state unit. Existing LiDAR place recognition methods are built and tuned for homogeneous sensors, and the authors argue they fail when scanning patterns, resolutions, and fields of view differ. HeLiOS counters this by restricting attention to small local spherical windows, aggregating local features through optimal-transport clustering, and supervising retrieval with a 3D overlap measure instead of raw positional distance. If the central claim holds, place recognition no longer requires a separate model per LiDAR type, and a robot can match a current scan against maps built years earlier by a different sensor.

What carries the argument

The central machinery is the pairing of a local spherical transformer with overlap-guided metric learning. The transformer splits 3D space into small windows in spherical coordinates $(r,\theta,\phi)$, expanding the window scale by a factor of 1.5 after strided convolutions, so attention operates on locally similar distributions rather than whole scenes that look very different across sensors. The overlap measure of Eq. (3), which counts nearest-neighbor voxel matches within a threshold $\tau = 1.5\delta$ on a $\delta = 4\,\mathrm{m}$ grid, replaces raw distance as the ground truth for mining positives, semi-positives, and negatives and for setting adaptive triplet margins; the Sinkhorn algorithm then carries out the optimal-transport assignment that turns local features into a compact global descriptor.

What would settle it

On a held-out set of HeLiPR scan pairs, compute the Eq. (3) overlap with the paper's $\delta = 4\,\mathrm{m}$ and $\tau = 1.5\delta$, compare it against overlap estimated by dense registration or ray-casting, and then re-run retrieval with an independently verified ground truth; if the $0.5$ threshold misclassifies many sparse-versus-dense pairs, the reported gains rest on the same criterion used for both training and testing.

Watch

Extended reading notes

Core claim

HeLiOS is presented as the first deep network tailored to heterogeneous LiDAR place recognition. It voxelizes a scan, extracts features with a shared U-Net sparse-convolution encoder, and applies multi-head attention inside small spherical windows of 10 m by 1.8 degrees by 1.8 degrees rather than full-radius windows, so that attention learns distributions that are similar across sensors. Local features are aggregated into a global descriptor by optimal-transport cluster assignment with a dustbin column that filters uninformative points, combined with GeM pooling and an MLP. Training uses a 3D overlap score between point clouds, classifying pairs as positive, semi-positive, or negative, and optimizes a truncated smooth-AP loss together with overlap-guided triplet losses that use adaptive margins. On the HeLiPR benchmark, HeLiOS reports higher average recall at rank 1 and 5 than prior methods for both narrow-FOV and wide-FOV database LiDARs, and in a long-term experiment it matches MulRan scans against HeLiPR scans taken four years later by a different LiDAR.

Load-bearing premise

The load-bearing premise is that the overlap score of Eq. (3), counting nearest-neighbor voxel matches within $1.5\delta$ on a $4\,\mathrm{m}$ grid, is a trustworthy measure of real physical overlap for every sensor pair, including very sparse versus very dense LiDARs; if that measure is miscalibrated, the training labels and the retrieval criterion are miscalibrated together.

Editorial extensions

If this is right

  • A robot with a narrow-FOV solid-state LiDAR can retrieve places from a map built by a 360-degree spinning LiDAR, and vice versa, without retraining a separate model per sensor.
  • The same overlap-based labels can supervise training across any sensor pair in the HeLiPR suite, removing the false positives and false negatives that distance-based mining produces when fields of view differ.
  • Retrieval remains real-time at under 10 Hz for descriptor dimensions tested up to $(m,l,e)=(64,128,256)$, and users can trade descriptor size against recall.
  • The lightweight HeLiOS-S variant, at 256 descriptor dimensions, beats larger specialized models on most heterogeneous benchmarks and stays competitive on homogeneous LiDAR data.
  • Long-term recognition across a four-year gap and different LiDAR types succeeds for almost all queries except where the scene appearance changed entirely.

Reading between the lines

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

  • If the overlap measure is accepted as transferable, HeLiOS could be trained on a new LiDAR pair without any manual positive-label annotation, since Eq. (3) computes labels directly from the scans.
  • The descriptor-dimension dial suggests a deployment cascade the paper does not test: use a small descriptor for fast prefiltering and a larger one for reranking, trading a little recall for much lower retrieval time.
  • Because the encoder is shared across sensors, the same network could be fed by a multi-robot fleet carrying different LiDAR models; the paper's single-robot evaluations leave that fleet-scale heterogeneity as an open extension.
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

5 major / 6 minor

Summary. The paper introduces HeLiOS, a deep network for heterogeneous LiDAR place recognition that combines a U-Net with sparse convolution, a local spherical transformer, optimal-transport-based feature aggregation, and an overlap-guided metric learning scheme. The overlap measure of Eq. (3) is used both to mine positive/semi-positive/negative training pairs and to define retrieval correctness in evaluation. The authors report state-of-the-art AR@k results on the HeLiPR heterogeneous benchmark, competitive performance on homogeneous NCLT and HeLiPR settings, a qualitative long-term recognition study, and ablations of the loss, transformer, and descriptor dimension. The code is released.

Significance. If the reported results hold under a fair evaluation protocol, HeLiOS would be a useful contribution to an under-studied problem: matching place recognition across LiDARs with different FOVs, resolutions, and scanning patterns. The paper's strengths include a clearly specified architecture, an interesting use of optimal transport aggregation adapted from visual place recognition, and a loss design that explicitly models a semi-positive class. The manuscript also ships its code, which supports reproducibility. However, the central empirical claim is currently weakened by the coupling of training supervision and evaluation through the same overlap criterion, the absence of baseline retraining details, and the lack of any uncertainty quantification in the reported tables.

major comments (5)
  1. [§IV-C] The overlap measure of Eq. (3), with voxel size δ=4 m and threshold τ=1.5δ=6 m, is used both to label HeLiOS's training pairs and to decide whether a retrieval is correct (overlap > 0.5) in evaluation. HeLiOS is therefore directly optimized for the same function that judges its success, whereas the baselines are trained with other objectives. The paper provides no calibration evidence that this voxelized nearest-neighbor count is a faithful proxy for physical co-visibility across heterogeneous sensor pairs, e.g., sparse VLP-16C versus dense OS2-128. I request: (a) a validation study of Eq. (3) against hand-labeled or pose-based ground truth on representative pairs, and (b) an evaluation using an independent retrieval criterion (e.g., a distance-based protocol or a different overlap parameterization) to show the Table I ranking is not an artifact of training and testing on the same function.
  2. [§IV-C] The long-term heterogeneous place recognition experiment in §IV-D is qualitative only: it reports matched and unmatched examples in Fig. 7 but gives no retrieval recall or comparison to baselines. Since the abstract and conclusion explicitly claim robustness to unseen LiDAR types over long-term changes, this claim needs a quantitative counterpart, such as AR@1/AR@5 under the §IV-C protocol or at least a matching rate over all queries. Without numbers, the long-term contribution is not supported.
  3. [§IV-D] The long-term heterogeneous place recognition experiment in §IV-D is qualitative only: it reports matched and unmatched examples in Fig. 7 but gives no retrieval recall or comparison to baselines. Since the abstract and conclusion explicitly claim robustness to unseen LiDAR types over long-term changes, this claim needs a quantitative counterpart, such as AR@1/AR@5 under the §IV-C protocol or at least a matching rate over all queries. Without numbers, the long-term contribution is not supported.
  4. [Tables I–V] All reported metrics are single numbers with no error bars or multiple-run statistics. Several differences between HeLiOS, HeLiOS-S, and the best baseline are small (e.g., Table I, Roundabout Ouster DB, HeLiOS-S AR@1 0.637 vs MinkLoc3Dv2 0.620; Town Bridge02 Aeva DB, HeLiOS AR@1 0.857 vs HeLiOS-S 0.808). It is therefore unclear which gaps are statistically significant. Please provide mean ± standard deviation over at least three training runs with different random seeds, or explicitly state that results are from a single run and discuss the expected variance.
  5. [§IV-B] The homogeneous evaluation uses a different correctness threshold (overlap > 0.8) than the heterogeneous evaluation (overlap > 0.5), and it is not explained why this threshold is appropriate or whether the same Eq. (3) parameters (δ=4 m, τ=6 m) are used. The choice of threshold directly affects the reported AR@k values in Table II, so the sensitivity of the conclusions to this threshold should be documented.
minor comments (6)
  1. [§III-A, Eq. (1)] Equation (1) is typeset incorrectly: the symbols "D(xq, xi) D(xq, xj) = ⇒" do not form a readable implication. The intended condition should be rewritten, for example as: if D(xq, xi) < D(xq, xj), then dg(gq, gi) < dg(gq, gj).
  2. [§III-D, Eq. (3)] Equation (3) defines an asymmetric overlap measure, and the paper later takes the maximum over the two directions. Please clarify why max is used rather than the average or min, and whether the asymmetric raw values were ever inspected for consistency.
  3. [§IV-A] The notation for the overlap voxel size is inconsistent: §III-D calls it δ and §IV-A calls it "voxel size d". Please use a single symbol throughout.
  4. [Table III] The row and column labels in the ablation table are hard to parse; in particular, the distinction between rows marked "LT SAP", "LT", and "LGT" would benefit from a legend or a more explicit description of which loss is applied to which pair type. The current formatting makes the ablation logic difficult to follow.
  5. [§IV-D] The alignment of the MulRan and HeLiPR sequences is said to be done with ELite [39], which is the authors' own earlier work. Please provide a brief description of the alignment accuracy or error, since the correctness of the long-term experiment depends on the quality of this alignment.
  6. [§IV-B] The phrase "applying our overlap criteria across all benchmarks" is ambiguous: it is not clear whether the same δ and τ values from the heterogeneous setting were also used for the NCLT homogeneous evaluation, and whether any dataset-specific tuning was performed.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; the closest concern is that the same overlap criterion (Eq. 3) supervises HeLiOS and defines retrieval correctness, which is an evaluation-consistency risk rather than a mathematical reduction.

full rationale

I walked the claimed derivation chain and found no step where a predicted result is equivalent by construction to its input. The network weights are learned from training sequences and tested on separate public sequences, so there is no fitted parameter renamed as a prediction and no statistical forcing of the headline numbers. The overlap measure in Eq. (3) is used both to define training labels ('Scans are categorized as positive if overlap is over 0.5, semi-positive between 0 and 0.5, and negative if zero') and to judge retrieval correctness ('A retrieval is correct if the overlap between the query and retrieval exceeds 0.5'). This means HeLiOS is optimized for the same criterion that decides its success, which can bias comparisons against baselines trained with distance-based or ranking objectives; however, this is a benchmark-calibration and fairness concern, not a circular derivation. The paper also cites its own HeLiPR dataset [12] and ELite alignment [39], but these are used as data and alignment tools, not as load-bearing mathematical evidence for the architecture or loss design, and no uniqueness theorem is imported from the authors' prior work. The spherical-transformer and optimal-transport components are adapted from external sources [16, 17] with explicit citation, and the ablations test component contributions on held-out splits. Overall, the central claim is externally evaluated, and the identified issues are minor consistency risks rather than circular steps.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper introduces no physical or mathematical entities, but it relies on a hand-designed overlap definition, a set of tuned thresholds and loss weights, and domain assumptions about the validity of overlap as place similarity. These are the main extra assumptions beyond standard machine learning practice.

free parameters (6)
  • Overlap thresholds (positive, semi-positive, negative) = 0.5, (0, 0.5), 0
    Used to define training classes and to label retrieval correctness in Section IV-B; chosen by hand, not derived.
  • Voxel size δ and NN threshold τ for overlap = δ = 4 m, τ = 6 m
    Set in Section IV-A to standardize resolution and reduce computation; τ = 1.5δ for misalignment robustness.
  • Spherical window size and expansion factor = (10 m, 1.8°, 1.8°), expansion 1.5
    Defines local attention windows in Section III-B; ablated in Table IV.
  • Guided-triplet hyperparameters (m1, m2, β, ω1, ω2) = 0.02, 0.19, e-1, 0.1, 0.1
    Scale factors and weights in Eq. (5)-(6), tuned on the benchmark.
  • Descriptor dimensions (m, l, e) = HeLiOS (64, 128, 256); HeLiOS-S (8, 32, 0)
    Chosen to balance accuracy and runtime; ablated in Table V.
  • Input preprocessing (max range, point count) = 100 m, 8192 points, normalized [-1,1]
    Standard input normalization in Section IV-A.
assumptions (4)
  • standard math Sinkhorn algorithm converges to a valid optimal transport assignment for the score matrix.
    Used in Section III-C without proof; standard result.
  • domain assumption Voxelized nearest-neighbor overlap Eq. (3) is an unbiased measure of physical place similarity across LiDAR types.
    Load-bearing for both training labels and evaluation; not validated per sensor pair.
  • domain assumption Single-scan global descriptors are sufficient for place recognition, and overlap > 0.5 is a correct retrieval.
    Problem definition in Section III-A and evaluation in Section IV-B.
  • ad hoc to paper Semi-positive class (0 < overlap < 0.5) with guided-triplet margins improves embedding geometry.
    Novel heuristic in Section III-D, validated only by ablation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HeLiOS: Heterogeneous LiDAR Place Recognition via Overlap-based Learning and Local Spherical Transformer." pith.science (2026). https://pith.science/paper/RW3DYYIQ

@misc{pith2026250118943,
  author       = {Pith},
  title        = {Pith review of: HeLiOS: Heterogeneous LiDAR Place Recognition via Overlap-based Learning and Local Spherical Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RW3DYYIQ}},
  note         = {Machine review of arXiv:2501.18943}
}
read the original abstract

LiDAR place recognition is a crucial module in localization that matches the current location with previously observed environments. Most existing approaches in LiDAR place recognition dominantly focus on the spinning type LiDAR to exploit its large FOV for matching. However, with the recent emergence of various LiDAR types, the importance of matching data across different LiDAR types has grown significantly-a challenge that has been largely overlooked for many years. To address these challenges, we introduce HeLiOS, a deep network tailored for heterogeneous LiDAR place recognition, which utilizes small local windows with spherical transformers and optimal transport-based cluster assignment for robust global descriptors. Our overlap-based data mining and guided-triplet loss overcome the limitations of traditional distance-based mining and discrete class constraints. HeLiOS is validated on public datasets, demonstrating performance in heterogeneous LiDAR place recognition while including an evaluation for long-term recognition, showcasing its ability to handle unseen LiDAR types. We release the HeLiOS code as an open source for the robotics community at https://github.com/minwoo0611/HeLiOS.

Figures

Figures reproduced from arXiv: 2501.18943 by the authors.

Figure 1
Figure 1. HeLiOS utilizes overlap for mining and the loss function. Conventional distance [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall pipeline of HeLiOS. HeLiOS voxelizes point clouds [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Local spherical window for applying multi-head attention in heterogeneous [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: (a) Overlap calculation to illustrate (3) as a diagram. (b) HeLiOS computes overlap for different LiDARs in 3D space. In contrast, OverlapNet misrepresents real￾world overlap even if the LiDARs are in the same location, as their overlap occurs only when each point fall…
Figure 5
Figure 5. Figure 5: PR curves with heterogeneous LiDARs. The title of each curve represents [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. KISS-IMU: Self-supervised Inertial Odometry with Motion-balanced Learning and Uncertainty-aware Inference

    cs.RO 2026-03 conditional novelty 6.0 of 10

    Self-supervised IMU odometry learns only from LiDAR ICP/PGO pseudo-labels with GMM motion reweighting and uncertainty-aware inference, without ground truth or joint multi-modal networks.

  2. MinkUNeXt-SI: Improving point cloud-based place recognition including spherical coordinates and LiDAR intensity

    cs.LG 2025-05 conditional novelty 3.0 of 10

    MinkUNeXt-SI feeds spherical coordinates and normalized LiDAR intensity into a Minkowski U-Net and reports competitive place recognition recall on Oxford, USyd, KITTI, NCLT, and a new campus dataset.

Reference graph

Works this paper leans on

44 extracted references · 42 canonical work pages · cited by 2 Pith papers

  1. [1]

    Scan context: Egocentric spatial de- scriptor for place recognition within 3d point cloud map,

    G. Kim and A. Kim, “Scan context: Egocentric spatial de- scriptor for place recognition within 3d point cloud map,” in Proc. IEEE/RSJ Intl. Conf. on Intell. Robots and Sys. , 2018, pp. 4802–4809

  2. [2]

    Scan context++: Structural place recognition robust to rotation and lateral variations in urban environments,

    G. Kim, S. Choi, and A. Kim, “Scan context++: Structural place recognition robust to rotation and lateral variations in urban environments,” IEEE Trans. Robot. , vol. 38, no. 3, pp. 1856–1874, 2021

  3. [3]

    Ring++: Roto-translation-invariant gram for global localization on a sparse scan map,

    X. Xu, S. Lu, J. Wu, H. Lu, Q. Zhu, Y . Liao, R. Xiong, and Y . Wang, “Ring++: Roto-translation-invariant gram for global localization on a sparse scan map,” IEEE Trans. Robot., 2023

  4. [4]

    Bevplace: Learning lidar-based place recognition using bird’s eye view images,

    L. Luo, S. Zheng, Y . Li, Y . Fan, B. Yu, S.-Y . Cao, J. Li, and H.-L. Shen, “Bevplace: Learning lidar-based place recognition using bird’s eye view images,” in Proc. IEEE Intl. Conf. on Comput. Vision, 2023

  5. [5]

    Std: Stable triangle descriptor for 3d place recognition,

    C. Yuan, J. Lin, Z. Zou, X. Hong, and F. Zhang, “Std: Stable triangle descriptor for 3d place recognition,” in Proc. IEEE Intl. Conf. on Robot. and Automat. IEEE, 2023, pp. 1897– 1903

  6. [6]

    Btc: A binary and triangle combined descriptor for 3d place recognition,

    C. Yuan, J. Lin, Z. Liu, H. Wei, X. Hong, and F. Zhang, “Btc: A binary and triangle combined descriptor for 3d place recognition,” IEEE Trans. Robot. , 2024

  7. [7]

    Narrowing your fov with solid: Spatially organized and lightweight global descriptor for fov-constrained lidar place recognition,

    H. Kim, J. Choi, T. Sim, G. Kim, and Y . Cho, “Narrowing your fov with solid: Spatially organized and lightweight global descriptor for fov-constrained lidar place recognition,” IEEE Robot. and Automat. Lett. , 2024

  8. [8]

    Cvtnet: A cross- view transformer network for lidar-based place recognition in autonomous driving environments,

    J. Ma, G. Xiong, J. Xu, and X. Chen, “Cvtnet: A cross- view transformer network for lidar-based place recognition in autonomous driving environments,” IEEE Trans. Ind. In- formatics, 2023

Show all 44 references
  1. [9]

    Overlap- transformer: An efficient and yaw-angle-invariant transformer network for lidar-based place recognition,

    J. Ma, J. Zhang, J. Xu, R. Ai, W. Gu, and X. Chen, “Overlap- transformer: An efficient and yaw-angle-invariant transformer network for lidar-based place recognition,” IEEE Robot. and Automat. Lett., vol. 7, no. 3, pp. 6958–6965, 2022

  2. [10]

    Robust place recognition using an imaging lidar,

    T. Shan, B. Englot, F. Duarte, C. Ratti, and D. Rus, “Robust place recognition using an imaging lidar,” in Proc. IEEE Intl. Conf. on Robot. and Automat. , 2021, pp. 5469–5475

  3. [11]

    A decentralized framework for simultaneous calibration, localization and mapping with multiple lidars,

    J. Lin, X. Liu, and F. Zhang, “A decentralized framework for simultaneous calibration, localization and mapping with multiple lidars,” in Proc. IEEE/RSJ Intl. Conf. on Intell. Robots and Sys. , 2020, pp. 4870–4877

  4. [12]

    Helipr: Heterogeneous lidar dataset for inter-lidar place recognition under spatiotemporal variations,

    M. Jung, W. Yang, D. Lee, H. Gil, G. Kim, and A. Kim, “Helipr: Heterogeneous lidar dataset for inter-lidar place recognition under spatiotemporal variations,” Intl. J. of Robot. Research, p. 02783649241242136, 2023

  5. [13]

    Cross- loc3d: Aerial-ground cross-source 3d place recognition,

    T. Guan, A. Muthuselvam, M. Hoover, X. Wang, J. Liang, A. J. Sathyamoorthy, D. Conover, and D. Manocha, “Cross- loc3d: Aerial-ground cross-source 3d place recognition,” in Proc. IEEE Intl. Conf. on Comput. Vision , 2023, pp. 11 335– 11 344

  6. [14]

    (lc) 2: Lidar-camera loop constraints for cross-modal place recog- nition,

    A. J. Lee, S. Song, H. Lim, W. Lee, and H. Myung, “(lc) 2: Lidar-camera loop constraints for cross-modal place recog- nition,” IEEE Robot. and Automat. Lett. , vol. 8, no. 6, pp. 3589–3596, 2023

  7. [15]

    Min- kloc++: lidar and monocular image fusion for place recog- nition,

    J. Komorowski, M. Wysocza ´nska, and T. Trzcinski, “Min- kloc++: lidar and monocular image fusion for place recog- nition,” in Intl. Joint Conf. on Neural Networks , 2021, pp. 1–8

  8. [16]

    Spherical trans- former for lidar-based 3d recognition,

    X. Lai, Y . Chen, F. Lu, J. Liu, and J. Jia, “Spherical trans- former for lidar-based 3d recognition,” in Proc. IEEE Conf. on Comput. Vision and Pattern Recog. , pp. 17 545–17 555

  9. [17]

    Optimal transport aggregation for visual place recognition,

    S. Izquierdo and J. Civera, “Optimal transport aggregation for visual place recognition,” in Proc. IEEE Conf. on Comput. Vision and Pattern Recog., 2024, pp. 17 658–17 668

  10. [18]

    Pointnetvlad: Deep point cloud based retrieval for large-scale place recognition,

    M. A. Uy and G. H. Lee, “Pointnetvlad: Deep point cloud based retrieval for large-scale place recognition,” in Proc. IEEE Conf. on Comput. Vision and Pattern Recog. , 2018, pp. 4470–4479

  11. [19]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 652–660

  12. [20]

    Lpd-net: 3d point cloud learning for large-scale place recognition and environment analysis,

    Z. Liu, S. Zhou, C. Suo, P. Yin, W. Chen, H. Wang, H. Li, and Y .-H. Liu, “Lpd-net: 3d point cloud learning for large-scale place recognition and environment analysis,” in Proc. IEEE Intl. Conf. on Comput. Vision , 2019, pp. 2831–2840

  13. [21]

    Local descriptor for robust place recognition using lidar intensity,

    J. Guo, P. V . Borges, C. Park, and A. Gawel, “Local descriptor for robust place recognition using lidar intensity,”IEEE Robot. and Automat. Lett. , vol. 4, no. 2, pp. 1470–1477, 2019

  14. [22]

    Improving point cloud based place recog- nition with ranking-based loss and large batch training,

    J. Komorowski, “Improving point cloud based place recog- nition with ranking-based loss and large batch training,” in Proc. Intl. Conf. Pattern Recog. , 2022, pp. 3699–3705

  15. [23]

    Logg3d-net: Locally guided global descriptor learning for 3d place recognition,

    K. Vidanapathirana, M. Ramezani, P. Moghadam, S. Srid- haran, and C. Fookes, “Logg3d-net: Locally guided global descriptor learning for 3d place recognition,” in Proc. IEEE Intl. Conf. on Robot. and Automat. , 2022, pp. 2215–2221

  16. [24]

    Casspr: Cross attention single scan place recognition,

    Y . Xia, M. Gladkova, R. Wang, Q. Li, U. Stilla, J. F. Henriques, and D. Cremers, “Casspr: Cross attention single scan place recognition,” in Proc. IEEE Intl. Conf. on Comput. Vision, pp. 8461–8472

  17. [25]

    Salsa: Swift adaptive lightweight self-attention for enhanced lidar place recognition,

    R. G. Goswami, N. Patel, P. Krishnamurthy, and F. Khorrami, “Salsa: Swift adaptive lightweight self-attention for enhanced lidar place recognition,” IEEE Robot. and Automat. Lett. , 2024

  18. [26]

    Lcpr: A multi- scale attention-based lidar-camera fusion network for place recognition,

    Z. Zhou, J. Xu, G. Xiong, and J. Ma, “Lcpr: A multi- scale attention-based lidar-camera fusion network for place recognition,” IEEE Robot. and Automat. Lett. , 2023

  19. [27]

    General- ized contrastive optimization of siamese networks for place recognition,

    M. Leyva-Vallina, N. Strisciuglio, and N. Petkov, “General- ized contrastive optimization of siamese networks for place recognition,” 2023

  20. [28]

    OverlapNet: Loop Closing for LiDAR-based SLAM,

    X. Chen, T. L ¨abe, A. Milioto, T. R ¨ohling, O. Vysotska, A. Haag, J. Behley, and C. Stachniss, “OverlapNet: Loop Closing for LiDAR-based SLAM,” in Proc. Robot.: Science & Sys. Conf. , 2020

  21. [29]

    Triplet loss in siamese network for object tracking,

    X. Dong and J. Shen, “Triplet loss in siamese network for object tracking,” in Proc. European Conf. on Comput. Vision , 2018, pp. 459–474

  22. [30]

    Understanding the behaviour of con- trastive loss,

    F. Wang and H. Liu, “Understanding the behaviour of con- trastive loss,” in Proc. IEEE Conf. on Comput. Vision and Pattern Recog., 2021, pp. 2495–2504

  23. [31]

    Deep metric learning with hierarchical triplet loss,

    W. Ge, “Deep metric learning with hierarchical triplet loss,” in Proc. European Conf. on Comput. Vision, September 2018

  24. [32]

    Revisiting self- similarity: Structural embedding for image retrieval,

    S. Lee, S. Lee, H. Seong, and E. Kim, “Revisiting self- similarity: Structural embedding for image retrieval,” in Proc. IEEE Conf. on Comput. Vision and Pattern Recog., June 2023

  25. [33]

    4d spatio-temporal convnets: Minkowski convolutional neural networks,

    C. Choy, J. Gwak, and S. Savarese, “4d spatio-temporal convnets: Minkowski convolutional neural networks,” in Proc. IEEE Conf. on Comput. Vision and Pattern Recog. , 2019, pp. 3075–3084

  26. [34]

    A survey on vision transformer,

    K. Han, Y . Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y . Tang, A. Xiao, C. Xu, Y . Xuet al., “A survey on vision transformer,” IEEE Trans. Pattern Analysis and Machine Intell. , vol. 45, no. 1, pp. 87–110, 2022

  27. [35]

    Sinkhorn distances: Lightspeed computation of optimal transport,

    M. Cuturi, “Sinkhorn distances: Lightspeed computation of optimal transport,” in Advances in Neural Information Processing Sys. Conf. , C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Weinberger, Eds., vol. 26. Curran Associates, Inc., 2013

  28. [36]

    Fine-tuning cnn image retrieval with no human annotation,

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

  29. [37]

    University of michigan north campus long-term vision and lidar dataset,

    N. Carlevaris-Bianco, A. K. Ushani, and R. M. Eustice, “University of michigan north campus long-term vision and lidar dataset,” Intl. J. of Robot. Research , vol. 35, no. 9, pp. 1023–1035, 2016

  30. [38]

    Mulran: Multimodal range dataset for urban place recognition,

    G. Kim, Y . S. Park, Y . Cho, J. Jeong, and A. Kim, “Mulran: Multimodal range dataset for urban place recognition,” in Proc. IEEE Intl. Conf. on Robot. and Automat. , 2020, pp. 6246–6253

  31. [39]

    Ephemerality meets lidar-based lifelong mapping,

    H. Gil, D. Lee, G. Kim, and A. Kim, “Ephemerality meets lidar-based lifelong mapping,” in Proc. IEEE Intl. Conf. on Robot. and Automat. , Atlanta, May. 2025

  32. [40]

    Lidar odometry survey: recent advancements and remaining challenges,

    D. Lee, M. Jung, W. Yang, and A. Kim, “Lidar odometry survey: recent advancements and remaining challenges,” Intl. Service Robot., vol. 17, no. 2, pp. 95–118, 2024

  33. [41]

    Asynchronous multiple lidar-inertial odometry using point-wise inter-lidar uncertainty propagation,

    M. Jung, S. Jung, and A. Kim, “Asynchronous multiple lidar-inertial odometry using point-wise inter-lidar uncertainty propagation,” IEEE Robotics and Automation Letters , vol. 8, no. 7, pp. 4211–4218, 2023

  34. [42]

    Multi-lio: A lightweight multiple lidar-inertial odometry system,

    Q. Chen, G. Li, X. Xue, and J. Pu, “Multi-lio: A lightweight multiple lidar-inertial odometry system,” in Proc. IEEE Intl. Conf. on Robot. and Automat. , 2024, pp. 13 748–13 754

  35. [43]

    Disco-slam: Distributed scan context-enabled multi-robot lidar slam with two-stage global-local graph optimization,

    Y . Huang, T. Shan, F. Chen, and B. Englot, “Disco-slam: Distributed scan context-enabled multi-robot lidar slam with two-stage global-local graph optimization,” IEEE Robot. and Automat. Lett., vol. 7, no. 2, pp. 1150–1157, 2021

  36. [44]

    Lamp 2.0: A robust multi-robot slam system for op- eration in challenging large-scale underground environments,

    Y . Chang, K. Ebadi, C. E. Denniston, M. F. Ginting, A. Rosi- nol, A. Reinke, M. Palieri, J. Shi, A. Chatterjee, B. Morrell et al. , “Lamp 2.0: A robust multi-robot slam system for op- eration in challenging large-scale underground environments,” IEEE Robot. and Automat. Lett....

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.