Pith. sign in

REVIEW 4 major objections 5 minor 58 references

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

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Converting LiDAR point clouds to spherical coordinates and adding normalized intensity lets a sparse-convolution U-Net match or beat prior state-of-the-art place recognition.

desk verdict Solid incremental follow-up with public code, but the 'surpasses SOTA' claim is undercut by the authors' own USyd tables and a plausible test-set tuning leak. read the letter →

arxiv 2505.17591 v1 pith:EUEVZZAP submitted 2025-05-23 cs.LG cs.AIcs.CVcs.RO

Judith Vilella-Cantos , Juan José Cabrera , Luis Payá , Mónica Ballesta , David Valiente This is my paper
classification cs.LGcs.AIcs.CVcs.RO
keywords placerecognitionLiDARpointcloudssphericalcoordinatesintensitynormalizationMinkowskiconvolutionsU-Netarchitecturesparsetensorsdescriptorlearning
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 a LiDAR-only place recognition network can be pushed past current benchmarks simply by changing how each point is represented: spherical coordinates instead of Cartesian, plus per-point intensity normalized to [0,1] by histogram equalization. The modified network, MinkUNeXt-SI, produces one global descriptor per scan and is evaluated with the standard retrieval metrics Recall@1% and Recall@1. The reported numbers show it close to the closest prior method MinkLoc3D-SI on USyd, above it on Oxford, and above it on KITTI generalization, with additional transfer results on the NCLT dataset and a new AR VC dataset. This matters because place recognition underpins loop closure in SLAM, so a preprocessing change that improves recall without new sensors or added latency is directly useful for real-time robot localization.

What carries the argument

The central object is a four-channel sparse tensor per LiDAR scan. Each point is mapped from Cartesian $(x,y,z)$ to spherical $(r,\theta,\phi)$ using $r=\sqrt{x^2+y^2+z^2}$, $\theta=\arctan2(y,x)$, $\phi=\arccos(z/r)$, and its intensity is histogram-equalized and scaled to $[0,1]$. These channels are quantized into a sparse voxel grid and passed through the MinkUNeXt U-Net of Minkowski (sparse) convolutions with skip connections, trained with the Truncated Smooth AP loss to produce one global descriptor. This input representation is what carries the argument: it puts the network in the sensor's native geometry while adding material information that geometry alone does not contain.

What would settle it

Run the same MinkUNeXt-SI training on USyd with hyperparameters chosen by a held-out validation split or cross-validation, then recompute Recall@1% and Recall@1 on the official USyd test split; if the numbers fall clearly below 98.44% and 92.02%, the paper's claim of being on par with MinkLoc3D-SI on USyd is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that converting LiDAR points to spherical coordinates and adding normalized intensity improves the descriptors of an already strong sparse-convolution U-Net. Recall@1% is the fraction of query scans whose true match appears in the top 1 percent of retrieved candidates, and Recall@1 is the fraction whose top-ranked candidate is correct. The claimed numbers: on USyd, MinkUNeXt-SI reaches 98.44% Recall@1% and 92.02% Recall@1, below MinkLoc3D-SI's 99.0% and 94.7%; on Oxford it reaches 97.93% and 93.22%, above MinkLoc3D-SI's 93.4% and 82.2%; and on KITTI after USyd training it reaches 84.14% and 83.33%, above the baseline's 81.0% and 78.6%. The authors conclude that the spherical-plus-intensity input is a powerful addition that keeps the descriptor robust to seasonal and dynamic changes and transferable across LiDAR sensors.

Load-bearing premise

The reported recalls stand on the assumption that the hyperparameters (batch size, split size, quantization size, and voxel size) were tuned on a validation split separate from the USyd test set used in the final tables; the paper does not describe such a split.

Editorial extensions

If this is right

  • On the USyd-trained model, the spherical-plus-intensity input gives the highest Recall@1% of the four input variants (98.44%), and on Oxford the SI variant beats the MinkLoc3D-SI baseline by 4.5 points in Recall@1%.
  • Generalization to KITTI improves over the baseline for both training datasets: USyd-trained MinkUNeXt-SI reaches 84.14% Recall@1% and 83.33% Recall@1, versus MinkLoc3D-SI's 81.0% and 78.6%.
  • The descriptor separates route locations into clear clusters on all five datasets, indicating the representation retains enough geometric information for fine-grained retrieval.
  • Inference time stays below 25 ms for clouds of up to 23,000 points, so the method can run in real time inside a SLAM loop.
  • The USyd-trained model also reaches 88.84% Recall@1% on NCLT and 97.98% on the custom AR VC dataset, supporting the claim that the descriptor generalizes to long-term environments not seen in training.

Reading between the lines

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

  • The authors train one model per dataset, but their generalization table suggests a single model trained jointly on Oxford and USyd could produce a descriptor that transfers across 2D-derived and 3D LiDAR geometries; that experiment is a natural next step they do not run.
  • Because histogram equalization removes the sensor-specific intensity scale, the preprocessing should make the descriptor portable across LiDAR models with raw intensity ranges that differ by orders of magnitude; this can be tested by evaluating a model trained on one dataset against a different sensor without retraining.
  • The clean cluster separation in the descriptor-distance maps implies the same descriptor could serve not only top-1 retrieval but also coarse localization or loop-closure hypothesis ranking; the paper does not develop that use.
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 presents MinkUNeXt-SI, a point-cloud place recognition method that feeds a MinkUNeXt backbone with spherical coordinates and normalized (histogram-equalized) LiDAR intensity. The authors evaluate on Oxford, USyd, KITTI, NCLT, and a newly collected AR VC dataset, reporting Recall@1% and Recall@1. The abstract and conclusions claim the method reaches and surpasses state-of-the-art performance and generalizes well across datasets. The contribution is primarily empirical: a preprocessing combination applied to a previously published backbone, plus a new benchmark dataset.

Significance. If the reported numbers are reliable, the paper would provide a useful data point on the value of spherical coordinates and intensity normalization for LiDAR place recognition, and the public release of code and the AR VC dataset would aid reproducibility. The method inherits a strong backbone (MinkUNeXt) and the comparison to MinkLoc3D-SI is the most relevant baseline. However, the headline claim of surpassing state-of-the-art is not supported by the paper's own tables on USyd, and the hyperparameter-tuning protocol raises a risk of optimistic bias in all reported recall values. The significance is therefore conditional on the authors clarifying the evaluation protocol and correcting the overstatements.

major comments (4)
  1. [Section 4.2 / Figure 7] The hyperparameter selection for batch size, split size, voxel size, and quantization size is described as being based on Recall@1% results 'with the USyd dataset' (Figure 7), but the paper does not state that this sweep used a held-out validation split. Section 4.1.1.2 mentions training and validation sets but only gives counts for training (19,138) and testing (8,797), with no separate validation count. If the Recall@1% values in Figure 7 were computed on the same USyd test split later reported in Tables 4 and 5, the reported recalls are optimistically biased, and the 'reaches and surpasses state-of-the-art' conclusion is not supported. Please specify the exact split used for tuning, and if the test set was used, re-run the experiments with a proper validation/tuning protocol.
  2. [Abstract, Section 5, Tables 4 and 5] The abstract states that MinkUNeXt-SI 'reaches and surpasses state-of-the-art performance,' but the paper's own Tables 4 and 5 show that on USyd, MinkUNeXt-SI (98.44% Recall@1%, 92.02% Recall@1) is lower than MinkLoc3D-SI (99.0% and 94.7%). The text in Section 5 acknowledges this ('it does not surpass the method to which we compare, it is capable of matching its performance'), which contradicts the abstract and the concluding statements in Section 6. The abstract and conclusion should be revised to describe the results accurately, distinguishing datasets and configurations where the method matches or exceeds the baseline from those where it does not.
  3. [Section 5.1 / Table 7] Table 7 reports NCLT and AR VC generalizations with Recall@1% values of 88.84% and 97.98% (USyd-trained) and states these results are 'outperforming the state-of-the-art,' yet no comparison methods are given for these datasets. Without any baseline numbers on NCLT and AR VC, the claim of outperforming the state-of-the-art is unsupported. Either add existing published results on these datasets or explicitly state that no direct comparison is currently available and remove the superiority claim.
  4. [Section 3.2 / Section 4.1.1] The method's contribution includes histogram equalization of intensity, but the text says Oxford and USyd already had normalized intensity and that equalization was applied only to AR VC and NCLT. This raises a reproducibility question: does the proposed pipeline include histogram equalization for the training datasets, or is it only applied to the generalization datasets? Please clarify the exact preprocessing applied to each dataset, since the current description mixes the general method with dataset-specific procedures.
minor comments (5)
  1. [Eq. (3)] Equation (3) is written as arccos(zi, ri); the intended formula is presumably arccos(zi/ri) (or arccos(zi/r)). Please correct the notation.
  2. [Throughout] There are several typos and inconsistent terms: 'casuistic' should be 'case,' 'whithin' should be 'within,' 'T able' should be 'Table,' and the network name appears both as 'MinkUNeXt-SI' and 'MinKUNeXt-SI.' Please proofread.
  3. [Table 6] In Table 6, the comparison rows for PointNetVLAD, LPD-Net, AugNet, and Coral-VLAD are cited from [56] and trained on Oxford, while the MinkLoc3D/MinkLoc3D-SI rows are trained on USyd. Please state explicitly which rows are directly comparable to the MinkUNeXt-SI rows, since mixing training datasets and result sources makes the 'improvement' claim harder to judge.
  4. [Section 4.1.1.2] The text says the USyd dataset was 'split into training and validation sets,' but only training (19,138) and testing (8,797) counts are given. Please provide the validation split size or explain the relationship between validation and test sets.
  5. [Section 5.2 / Figure 9] The inference-time figure is referenced but not described in detail; the claim of 'less than 25 ms for 23,000 points' would be more useful if the hardware and the exact preprocessing overhead were reported, and if the figure axis labels and units were clearly visible.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation found: the headline recalls are external benchmark measurements, not quantities reduced from fitted parameters or self-citations.

full rationale

The paper's central claim is empirical: MinkUNeXt-SI descriptors reach state-of-the-art Recall@1% and Recall@1 on Oxford, USyd, KITTI, NCLT, and AR VC. These numbers are measured against fixed dataset splits and compared with external baselines such as MinkLoc3D-SI [13], PointNetVLAD [6], and CORAL [56]. The preprocessing steps are explicitly specified: spherical coordinate formulas (Eqs. 1-3) are standard geometry, and intensity equalization (Eqs. 4-6) is the standard scikit histogram-equalization procedure; neither is defined in terms of the reported recalls. The network is the authors' prior MinkUNeXt [16], so there is a self-citation, but the benchmark result does not reduce by construction to that citation: the same claims are tested against independent comparisons. Two methodological weaknesses deserve attention but are not circularity under the required definition. First, Section 4.2 tunes batch size, split size, quantization size, and voxel size using Recall@1% 'based on training with the USyd dataset' and Figure 7, while Section 4.1.1.2 reports only 19,138 training and 8,797 test scans with no separate validation count; if the sweep evaluated the same USyd split later reported in Tables 4-5, the USyd recalls and the reused hyperparameters would be optimistically biased. That is a leakage and statistical-validity concern, not a case where a prediction equals its input by construction. Second, Table 7 reports NCLT and AR VC 'outperforming the state-of-the-art' without listing comparator methods, so those claims are unsupported but not circular. Overall, no step in the paper exhibits the quoted equation-level reduction or fitted-parameter-renamed-as-prediction pattern that would justify a circularity score above 2.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim is an empirical benchmark result; it rests on tuned hyperparameters and preprocessing assumptions rather than on new physical entities or derived constants. No new entities are introduced.

free parameters (5)
  • Batch size = 2048
    Chosen from the Recall@1% sweep in Figure 7 using the USyd dataset; the final USyd evaluation then uses this value.
  • Batch split size = 16
    Chosen in the same Figure 7 sweep to manage memory for the Truncated Smooth AP loss with large batches.
  • Voxel size = None (no downsampling)
    Figure 7 shows that applying voxel downsampling reduces Recall@1%, so it was disabled in the final configuration.
  • Quantization size = USyd: 0.1 or [0.1, 2.0, 1.875]; Oxford: 0.01 or [0.01, 2.0, 1.875] per Table 3
    Set separately for Cartesian vs spherical coordinates and for each training dataset; these values control the voxel grid and are not derived.
  • Optimizer and loss hyperparameters = Initial LR 1e-3, weight decay 1e-4, sigmoid temperature 0.01, scheduler steps 250 and 350
    Taken from Table 2 as a fixed training configuration; the paper does not derive these values.
assumptions (4)
  • domain assumption The Cartesian to spherical transformation in Eqs. (1)-(3) preserves the geometric information needed for place recognition across all sensors used.
    Section 3.1 states that the FoV must be considered for each sensor, but no explicit per-sensor transformation is given; the success of the method depends on this being valid for 2D-derived Oxford clouds and 3D LiDAR clouds alike.
  • domain assumption Per-scan histogram equalization of intensity is a stable normalization that transfers across different LiDAR sensors.
    Section 3.2 applies equalization to NCLT and AR VC while training datasets are used as provided; the paper does not test whether equalized intensities from different sensors are comparable.
  • domain assumption The evaluation splits and preprocessing used for USyd and Oxford are identical or equivalent to the splits used for MinkLoc3D-SI and CORAL in the cited works.
    Section 4.1.1 follows [13] and Section 5 cites [56] for KITTI, but exact split lists and preprocessing scripts are not included in this preprint.
  • domain assumption The MinkUNeXt U-Net backbone behaves as described in the authors' earlier paper [16].
    The architecture and its feature extraction ability are taken from prior work and are not re-derived here; all reported results inherit this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MinkUNeXt-SI: Improving point cloud-based place recognition including spherical coordinates and LiDAR intensity." pith.science (2026). https://pith.science/paper/EUEVZZAP

@misc{pith2026250517591,
  author       = {Pith},
  title        = {Pith review of: MinkUNeXt-SI: Improving point cloud-based place recognition including spherical coordinates and LiDAR intensity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EUEVZZAP}},
  note         = {Machine review of arXiv:2505.17591}
}
read the original abstract

In autonomous navigation systems, the solution of the place recognition problem is crucial for their safe functioning. But this is not a trivial solution, since it must be accurate regardless of any changes in the scene, such as seasonal changes and different weather conditions, and it must be generalizable to other environments. This paper presents our method, MinkUNeXt-SI, which, starting from a LiDAR point cloud, preprocesses the input data to obtain its spherical coordinates and intensity values normalized within a range of 0 to 1 for each point, and it produces a robust place recognition descriptor. To that end, a deep learning approach that combines Minkowski convolutions and a U-net architecture with skip connections is used. The results of MinkUNeXt-SI demonstrate that this method reaches and surpasses state-of-the-art performance while it also generalizes satisfactorily to other datasets. Additionally, we showcase the capture of a custom dataset and its use in evaluating our solution, which also achieves outstanding results. Both the code of our solution and the runs of our dataset are publicly available for reproducibility purposes.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 50 canonical work pages

  1. [1]

    European Conference on Computer Vision, 748–761 (2010)

    Knopp, J., Sivic, J., Pajdla, T.: Avoiding confusing features in place recognition. European Conference on Computer Vision, 748–761 (2010). Springer

    work page 2010

  2. [2]

    2019 16th International Conference on Ubiquitous Robots (UR), 189–193 (2019)

    Lee, J., Kim, A.: Neural network-based long-term place recognition from omni- images. 2019 16th International Conference on Ubiquitous Robots (UR), 189–193 (2019). IEEE

    work page 2019

  3. [3]

    arXiv preprint arXiv:2205.11567 (2022)

    Schleiss, M., Rouatbi, F., Cremers, D.: VPAIR–aerial visual place recognition and localization in large-scale outdoor environments. arXiv preprint arXiv:2205.11567 (2022)

    arXiv 2022

  4. [4]

    2024 IEEE Radar Conference (RadarConf24), 1–6 (2024)

    Gadd, M., Newman, P.: Open-RadVLAD: Fast and robust RADAR place recognition. 2024 IEEE Radar Conference (RadarConf24), 1–6 (2024). IEEE

    work page 2024

  5. [5]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 17595–17605 (2024)

    Peng, G., Li, H., Zhao, Y., Zhang, J., Wu, Z., Zheng, P., Wang, D.: TransLoc4D: Transformer-based 4D RADAR place recognition. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 17595–17605 (2024)

    work page 2024

  6. [6]

    Proceedings of the IEEE conference on computer vision and pattern recognition, 4470–4479 (2018)

    Uy, M.A., Lee, G.H.: PointNetVLAD: Deep point cloud based retrieval for large- scale place recognition. Proceedings of the IEEE conference on computer vision and pattern recognition, 4470–4479 (2018)

    work page 2018

  7. [7]

    Journal of Field Robotics 36(2), 355–369 (2019)

    Santos, M.M., Zaffari, G.B., Ribeiro, P.O., Drews-Jr, P.L., Botelho, S.S.: Under- water place recognition using forward-looking SONAR images: A topological approach. Journal of Field Robotics 36(2), 355–369 (2019)

    work page 2019

  8. [8]

    Journal of Marine Science and Engineering 11(11), 2198 (2023)

    Gaspar, A.R., Matos, A.: Feature-based place recognition using forward-looking SONAR. Journal of Marine Science and Engineering 11(11), 2198 (2023)

    work page 2023

Show all 58 references
  1. [9]

    Pro- ceedings of the IEEE/CVF winter conference on applications of computer vision, 22 1790–1799 (2021)

    Komorowski, J.: MinkLoc3D: Point cloud based large-scale place recognition. Pro- ceedings of the IEEE/CVF winter conference on applications of computer vision, 22 1790–1799 (2021)

    2021

  2. [10]

    arXiv preprint arXiv:2403.13188 (2024)

    Viswanath, K., Jiang, P., Saripalli, S.: Reflectivity is all you need!: Advancing LiDAR semantic segmentation. arXiv preprint arXiv:2403.13188 (2024)

    2024 arXiv

  3. [11]

    2021 IEEE/RSJ international conference on intelligent robots and systems (IROS), 4467–4473 (2021)

    Deng, F., Feng, H., Liang, M., Wang, H., Yang, Y., Gao, Y., Chen, J., Hu, J., Guo, X., Lam, T.L.: FEANet: Feature-enhanced attention network for RGB-thermal real-time semantic segmentation. 2021 IEEE/RSJ international conference on intelligent robots and systems (IROS), 4467–4...

    2021

  4. [12]

    Remote Sensing 17(6) (2025)

    Heredia-Aguado, E., Cabrera, J.J., Jim´ enez, L.M., Valiente, D., Gil, A.: Static early fusion techniques for visible and thermal images to enhance convolutional neural network detection: A performance analysis. Remote Sensing 17(6) (2025)

    2025

  5. [13]

    IEEE Robotics and Automation Letters 7(2), 1079–1086 (2022) https://doi.org/10.1109/lra.2021.3136863

    ˙Zywanowski, K., Banaszczyk, A., Nowicki, M.R., Komorowski, J.: MinkLoc3D- SI: 3D LiDAR place recognition with sparse convolutions, spherical coordinates, and intensity. IEEE Robotics and Automation Letters 7(2), 1079–1086 (2022) https://doi.org/10.1109/lra.2021.3136863

    2022

  6. [14]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Rao, Y., Lu, J., Zhou, J.: Spherical fractal convolutional neural networks for point cloud recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 452–460 (2019)

    2019

  7. [15]

    In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp

    Chang, M.Y., Yeon, S., Ryu, S., Lee, D.: SpoxelNet: Spherical voxel-based deep place recognition for 3d point clouds of crowded indoor spaces. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 8564– 8570 (2020). IEEE

    2020

  8. [16]

    https: //arxiv.org/abs/2403.07593

    Cabrera, J.J., Santo, A., Gil, A., Viegas, C., Pay´ a, L.: MinkUNeXt: Point Cloud- based Large-scale Place Recognition using 3D Sparse Convolutions (2024). https: //arxiv.org/abs/2403.07593

    2024 arXiv

  9. [17]

    Neural Computing and Applications 35(22), 16487–16508 (2023)

    Cebollada, S., Pay´ a, L., Peidr´ o, A., Mayol, W., Reinoso, O.: Environment mod- eling and localization from datasets of omnidirectional scenes using machine learning techniques. Neural Computing and Applications 35(22), 16487–16508 (2023)

    2023

  10. [18]

    IEEE access11, 128421–128428 (2023)

    Juneja, S., Daniuˇ sis, P., Marcinkeviˇ cius, V.: Visual place recognition pre-training for end-to-end trained autonomous driving agent. IEEE access11, 128421–128428 (2023)

    2023

  11. [19]

    IEEE Access 9, 19516–19547 (2021)

    Masone, C., Caputo, B.: A survey on deep visual place recognition. IEEE Access 9, 19516–19547 (2021)

    2021

  12. [20]

    Journal Computer Vision and Image Understanding (CVIU) 72(2), 173–196 (2008) 23

    Bay, H., Ess, A., Tuytelaars, T., Van Gool, L.: Speeded-up robust features (SURF). Journal Computer Vision and Image Understanding (CVIU) 72(2), 173–196 (2008) 23

    2008

  13. [21]

    Interna- tional journal of computer vision 60, 91–110 (2004)

    Lowe, D.G.: Distinctive image features from scale-invariant keypoints. Interna- tional journal of computer vision 60, 91–110 (2004)

    2004

  14. [22]

    Proceedings of the IEEE conference on computer vision and pattern recognition, 5297–5307 (2016)

    Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: NetVLAD: CNN architecture for weakly supervised place recognition. Proceedings of the IEEE conference on computer vision and pattern recognition, 5297–5307 (2016)

    2016

  15. [23]

    IEEE Transactions on Robotics 28(5), 1188–1197 (2012)

    G´ alvez-L´ opez, D., Tard´ os, J.D.: Bags of binary words for fast place recognition in image sequences. IEEE Transactions on Robotics 28(5), 1188–1197 (2012)

    2012

  16. [24]

    arXiv preprint arXiv:2009.09255 (2020)

    Le, D.C., Youn, C.H.: City-scale visual place recognition with deep local features based on multi-scale ordered VLAD pooling. arXiv preprint arXiv:2009.09255 (2020)

    2020 arXiv

  17. [25]

    European Conference on Computer Vision, 196–215 (2022)

    Pramanick, S., Nowara, E.M., Gleason, J., Castillo, C.D., Chellappa, R.: Where in the world is this image? transformer-based geo-localization in the wild. European Conference on Computer Vision, 196–215 (2022). Springer

    2022

  18. [26]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 23182–23190 (2023)

    Clark, B., Kerrigan, A., Kulkarni, P.P., Cepeda, V.V., Shah, M.: Where we are and what we’re looking at: Query based worldwide image geo-localization using hierarchies and scenes. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 23182–23190 (2023)

    2023

  19. [27]

    Advances in Neural Information Processing Systems 34, 29009–29020 (2021)

    Yang, H., Lu, X., Zhu, Y.: Cross-view geo-localization with layer-to-layer trans- former. Advances in Neural Information Processing Systems 34, 29009–29020 (2021)

    2021

  20. [28]

    Waheed, S., Ferrarini, B., Milford, M., Ramchurn, S.D., Ehsan, S.: Image-based geo-localization for robotics: Are black-box vision-language models there yet? arXiv preprint arXiv:2501.16947 (2025)

    2025 arXiv

  21. [29]

    Proceedings of the IEEE conference on computer vision and pattern recognition, 652–660 (2017)

    Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep learning on point sets for 3D classification and segmentation. Proceedings of the IEEE conference on computer vision and pattern recognition, 652–660 (2017)

    2017

  22. [30]

    2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), 922–928 (2015)

    Maturana, D., Scherer, S.: VoxNet: A 3D convolutional neural network for real- time object recognition. 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), 922–928 (2015). IEEE

    2015

  23. [31]

    https://arxiv.org/abs/2105.11605

    Xu, T.X., Guo, Y.C., Li, Z., Yu, G., Lai, Y.K., H., Z.S.: TransLoc3D: Point cloud based large-scale place recognition using adaptive receptive fields (2021). https://arxiv.org/abs/2105.11605

    2021 arXiv

  24. [32]

    2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2831–2840 (2019) 24

    Liu, Z., Zhou, S., Suo, C., Yin, P., Chen, W., Wang, H., Li, H., Liu, Y.: LPD- Net: 3d point cloud learning for large-scale place recognition and environment analysis. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2831–2840 (2019) 24

    2019

  25. [33]

    IEEE Robotics and Automation Letters 8(6), 3414–3421 (2023) https://doi.org/10.1109/LRA.2023

    Chen, L., Wang, H., Kong, H., Yang, W., Ren, M.: PTC-Net: Point-wise trans- former with sparse convolution network for place recognition. IEEE Robotics and Automation Letters 8(6), 3414–3421 (2023) https://doi.org/10.1109/LRA.2023. 3267693

    2023 doi

  26. [34]

    2024 IEEE International Conference on Robotics and Automation (ICRA), 448–454 (2024) https://doi.org/10.1109/ ICRA57147.2024.10610758

    Liu, C., Chen, G., Song, R.: LPS-Net: Lightweight parameter-shared network for point cloud-based place recognition. 2024 IEEE International Conference on Robotics and Automation (ICRA), 448–454 (2024) https://doi.org/10.1109/ ICRA57147.2024.10610758

    2024

  27. [35]

    Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 11348–11357 (2021)

    Xia, Y., Xu, Y., Li, S., Wang, R., Du, J., Cremers, D., Stilla, U.: SOE-Net: A self-attention and orientation encoding network for point cloud based place recognition. Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 11348–11357 (2021)

    2021

  28. [36]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 3075–3084 (2019)

    Choy, C., Gwak, J., Savarese, S.: 4D spatio-temporal convnets: Minkowski convo- lutional neural networks. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 3075–3084 (2019)

    2019

  29. [37]

    2022 26th international conference on pattern recognition (ICPR), 3699–3705 (2022)

    Komorowski, J.: Improving point cloud based place recognition with ranking- based loss and large batch training. 2022 26th international conference on pattern recognition (ICPR), 3699–3705 (2022)

    2022

  30. [38]

    https://arxiv.org/abs/2406.01395

    Santo, A., Cabrera, J.J., Valiente, D., Viegas, C., Gil, A.: TE-NeXt: A LiDAR- Based 3D Sparse Convolutional Network for Traversability Estimation (2025). https://arxiv.org/abs/2406.01395

    2025 arXiv

  31. [39]

    Technologies 13(4), 162 (2025)

    Santo, A., Heredia, E., Viegas, C., Valiente, D., Gil, A.: Ground segmentation for lidar point clouds in structured and unstructured environments using a hybrid neural–geometric approach. Technologies 13(4), 162 (2025)

    2025

  32. [40]

    16th IEEE International Conference on Control, Automation, Robotics and Vision, ICARCV 2020., 886–891 (2020) https://doi.org/10.1109/ICARCV50220.2020.9305429

    Zywanowski, K., Banaszczyk, A., Nowicki, M.R.: Comparison of camera-based and 3D LiDAR-based place recognition across weather conditions. 16th IEEE International Conference on Control, Automation, Robotics and Vision, ICARCV 2020., 886–891 (2020) https://doi.org/10.1109/ICARCV...

    2020

  33. [41]

    2021 International Joint Conference on Neural Networks (IJCNN), 1–8 (2021)

    Komorowski, J., Wysocza´ nska, M., Trzcinski, T.: MinkLoc++: LiDAR and monocular image fusion for place recognition. 2021 International Joint Conference on Neural Networks (IJCNN), 1–8 (2021). IEEE

    2021

  34. [42]

    IEEE Robotics and Automation Letters 4(2), 1470–1477 (2019)

    Guo, J., Borges, P.V., Park, C., Gawel, A.: Local descriptor for robust place recognition using lidar intensity. IEEE Robotics and Automation Letters 4(2), 1470–1477 (2019)

    2019

  35. [43]

    Computing in Civil Engineering 2021, 1–8 (2021) 25

    Park, J., Cho, Y.K.: Laser intensity-assisted construction material classification in point cloud data using deep learning. Computing in Civil Engineering 2021, 1–8 (2021) 25

    2021

  36. [44]

    Advances in neural information processing systems 30 (2017)

    Qi, C.R., Yi, L., Su, H., Guibas, L.J.: PointNet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)

    2017

  37. [45]

    Proceedings of the IEEE conference on computer vision and pattern recognition, 244–253 (2018)

    Xu, D., Anguelov, D., Jain, A.: PointFusion: Deep sensor fusion for 3D bound- ing box estimation. Proceedings of the IEEE conference on computer vision and pattern recognition, 244–253 (2018)

    2018

  38. [46]

    In: Computer Graphics Forum, vol

    Guerrero, P., Kleiman, Y., Ovsjanikov, M., Mitra, N.J.: PCPNET learning local shape properties from raw point clouds. In: Computer Graphics Forum, vol. 37, pp. 75–85 (2018). Wiley Online Library

    2018

  39. [47]

    arXiv preprint arXiv:1805.07872 (2018)

    Lei, H., Akhtar, N., Mian, A.: Spherical convolutional neural network for 3D point clouds. arXiv preprint arXiv:1805.07872 (2018)

    2018 arXiv

  40. [48]

    arXiv preprint arXiv:2501.18943 (2025)

    Jung, M., Jung, S., Gil, H., Kim, A.: HeLiOS: Heterogeneous LiDAR place recog- nition via overlap-based learning and local spherical transformer. arXiv preprint arXiv:2501.18943 (2025)

    2025 arXiv

  41. [49]

    In: 2020 IEEE International Conference on Robotics and Automation (ICRA), pp

    Wang, H., Wang, C., Xie, L.: Intensity scan context: Coding intensity and geom- etry relations for loop closure detection. In: 2020 IEEE International Conference on Robotics and Automation (ICRA), pp. 2095–2101 (2020). IEEE

    2020

  42. [50]

    In: ISARC

    Moon, D., Kwon, S., Chung, S., Lee, S., Ko, T.: Pre-processing methodology of image compensation using histogram equalization for generating point-cloud of construction environment. In: ISARC. Proceedings of the International Sym- posium on Automation and Robotics in Construct...

    2017

  43. [51]

    The International Journal of Robotics Research 36(1), 3–15 (2017)

    Maddern, W., Pascoe, G., Linegar, C., Newman, P.: 1 year, 1000 km: The oxford robotcar dataset. The International Journal of Robotics Research 36(1), 3–15 (2017)

    2017

  44. [52]

    IEEE Intelligent Transportation Systems Magazine 12(4), 23–40 (2020) https://doi.org/10.1109/MITS.2020.2990183

    Zhou, W., Berrio, J.S., De Alvis, C., Shan, M., Worrall, S., Ward, J., Nebot, E.: Developing and testing robust autonomy: The University of Sydney Campus data set. IEEE Intelligent Transportation Systems Magazine 12(4), 23–40 (2020) https://doi.org/10.1109/MITS.2020.2990183

    2020

  45. [53]

    The International Journal of Robotics Research 35(9), 1023–1035 (2016)

    Carlevaris-Bianco, N., Ushani, A.K., Eustice, R.M.: University of Michigan North Campus long-term vision and LiDAR dataset. The International Journal of Robotics Research 35(9), 1023–1035 (2016)

    2016

  46. [54]

    The international journal of robotics research 32(11), 1231–1237 (2013)

    Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. The international journal of robotics research 32(11), 1231–1237 (2013)

    2013

  47. [55]

    https://arvc.umh.es/db/databases/ 26

    AR VC: The AR VC dataset (2024). https://arvc.umh.es/db/databases/ 26

    2024

  48. [56]

    2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2084–2091 (2021)

    Pan, Y., Xu, X., Li, W., Cui, Y., Wang, Y., Xiong, R.: CORAL: Colored struc- tural representation for bi-modal place recognition. 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2084–2091 (2021). IEEE

    2021

  49. [57]

    Electronics 9(8), 1295 (2020)

    Ahmed, M., Seraj, R., Islam, S.M.S.: The k-means algorithm: A comprehensive survey and performance evaluation. Electronics 9(8), 1295 (2020)

    2020

  50. [58]

    IEEE Robotics and Automation Letters 5(4), 5534–5541 (2020) 27

    Oertel, A., Cieslewski, T., Scaramuzza, D.: Augmenting visual place recognition with structural cues. IEEE Robotics and Automation Letters 5(4), 5534–5541 (2020) 27

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.