PinNet: Keypoint-Aware Learned Local Descriptors with Geometric Embedding for Loop Closure in LiDAR SLAM

Brett T. Lopez; Christa S. Robison; Nakul S. Joshi; Philip R. Osteen; Yanlong Ma

REVIEW 3 major objections 6 minor 25 references

PinNet: Keypoint-Aware Learned Local Descriptors with Geometric Embedding for Loop Closure in LiDAR SLAM

T0 review · 3 major / 6 minor · reviewed 2026-07-12 · grok-4.5

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

Pith's one-line read PinNet turns LiDAR point clouds into sparse keypoints and local geometric descriptors that both find loop closures and estimate relative poses accurately enough for map alignment and single-shot localization.

desk verdict Competent local-descriptor stack that wins on KITTI averages and registration, but cross-sensor claims are pictures-only. read the letter →

arxiv 2606.28637 v2 pith:3MCYLCBV submitted 2026-06-26 cs.RO

Yanlong Ma , Nakul S. Joshi , Christa S. Robison , Philip R. Osteen , Brett T. Lopez This is my paper · ORCID

classification cs.RO

keywords loopclosureLiDARSLAMlocaldescriptorskeypointdetectiongeometricself-attentionplacerecognitionpoint-cloudregistrationsingle-shotlocalization

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

Loop closure is the step that stops a robot’s map from drifting as it travels. Purely geometric LiDAR data make that step hard: the same place can look different from a new viewpoint, and similar structures produce false matches. PinNet attacks both problems by first detecting consistent keypoints and local descriptors from the raw cloud, then enriching those descriptors with a plane-based geometric self-attention layer that encodes distances and angles among the keypoints. The resulting descriptors are discriminative enough to retrieve previous places and accurate enough to recover the rigid transform between two scans. On the standard KITTI sequences the method leads average place-recognition scores and yields near-perfect registration recall after a light refinement step. The same network, without retraining, also aligns maps collected with different LiDAR sensors and different viewpoints, and can localize a single scan inside a large prebuilt map.

What carries the argument

Plane-based geometric self-attention: after keypoint detection, each pair of keypoints is embedded by Mahalanobis distance (using local covariance), surface-normal angle, Euclidean distance and triplet angles; these embeddings are injected into a transformer self-attention layer so that every descriptor carries explicit inter-keypoint geometry.

What would settle it

Run the released model on a held-out sequence from an Ouster or Livox sensor with known ground-truth poses; if place-recognition F1-max falls below 0.85 or registration recall drops below 90 percent under the same 5 m / 5° thresholds used on KITTI, the cross-sensor claim fails.

Watch

Extended reading notes

Core claim

A keypoint-aware network that extracts sparse, view-consistent local descriptors and then augments them with plane-to-plane geometric embeddings can simultaneously solve place recognition and scan-to-scan registration for LiDAR SLAM, outperforming prior handcrafted and learned descriptors on KITTI while generalizing across sensors and environments without retraining.

Load-bearing premise

Descriptors trained only on KITTI Velodyne data remain sufficiently distinctive and geometrically consistent on other LiDAR sensors and multi-level outdoor scenes without any retraining or quantitative failure analysis.

Editorial extensions

If this is right

A single forward pass of PinNet can replace separate handcrafted place-recognition and registration modules inside a LiDAR SLAM pipeline.
Loop-closure constraints become accurate enough that a simple pose-graph optimizer can stitch multi-session maps collected with different sensors.
Single-shot localization inside a prebuilt map becomes practical for robots that lack continuous odometry.
The same descriptors can be reused for other geometric tasks that need sparse, view-stable correspondences.

Reading between the lines

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

Because the descriptors already encode plane-level geometry, they may remain usable under moderate seasonal foliage change without explicit semantic labels.
The keypoint layer could be frozen and the attention head fine-tuned on a few dozen frames from a new sensor, offering a cheap domain-adaptation path the paper does not explore.
If the same architecture is applied to solid-state spinning LiDARs with non-uniform density, the random-dilation grouping step may need density-aware sampling to keep keypoint consistency.

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.

Desk Editor's Note

read the letter

Punchline: PinNet is a clean, usable combination of KPConv keypoints re-estimated from dense neighbors plus plane-aware (Mahalanobis + surface-angle) self-attention that delivers the best average F1/AP on the usual KITTI place-recognition sequences and near-perfect scan-to-scan recall after GICP. That is worth knowing if you care about pure-geometry loop closure.

What is actually new is modest but real: the keypoint relocation step that replaces voxel-grid inconsistency with a learned weighted average of dense neighbors, and the explicit plane-to-plane terms inside the geometric transformer. Everything else (circle loss, SVD init, GICP refine, NetVLAD ablation) is off-the-shelf. The paper does the combination carefully. Tables I–III and the ablations are transparent; removing keypoints or the plane embedding moves the numbers in the expected direction, and the local-vs-global comparison is fair. Registration without GICP already beats LCDNet by a wide margin, which is the useful part for pose-graph constraints. Citations cover the right prior art without obvious gaps.

Soft spot, in proportion: the “works on Ouster and Livox without retraining” claim is only visual overlays on ARL and UCLA multi-level data. No PR curves, no TE/RE, no failure cases. That is the weakest link, not a fatal one; the KITTI backbone is solid enough that the system still looks transferable, just unquantified. Free parameters are the usual list; nothing load-bearing is hidden. No code release is annoying but common.

This is for people building or benchmarking LiDAR SLAM loop-closure modules who want a strong local-descriptor baseline that also gives correspondences. It is not a foundational result, but it is honest engineering. Math and data look clean; I would send it to referees. Engage if you need a practical geometric place-recognition + registration pair; skip if you only care about global descriptors or radical new theory.

Referee Report

3 major / 6 minor

Summary. PinNet is a learning-based pipeline for LiDAR loop closure that extracts sparse keypoints and local geometric descriptors from voxelized point clouds (KPConv-FPN backbone with a learned keypoint aggregation step), then refines those descriptors with a plane-based geometric self-attention module (Mahalanobis distance of local covariances, surface-normal angle, plus Euclidean and triplet-angle embeddings). Place recognition uses the mean of the s smallest pairwise descriptor distances; relative pose is recovered by SVD on the corresponding keypoints and refined with GICP, with inlier-ratio and GICP-error rejection. The method is trained with contrastive circle loss, a differentiable SVD transformation loss, and Chamfer keypoint supervision (leave-one-out on KITTI 00–09). Quantitative place-recognition and registration results are reported on KITTI sequences 00/02/05/06/07 against BTC, Scan Context, OverlapTransformer, LoGG3D-Net, and LCDNet; qualitative map alignment and single-shot localization are shown on self-collected Ouster OS-1 (ARL) and Livox Mid-360 (UCLA) data without retraining.

Significance. If the claims hold, PinNet is a useful systems contribution: a single local-descriptor representation that supports both retrieval and accurate scan-to-scan registration, with competitive average F1-max/AP on KITTI (0.960 / 0.980) and near-100% registration recall after GICP (TE ~0.15 m, RE ~0.4°). The keypoint-aware downsampling and plane-level geometric embeddings are concrete design choices with ablations (Table III, Figs. 6–7) that improve consistency and discriminability relative to voxel-grid sparse points and a NetVLAD global baseline. Cross-sensor qualitative success without retraining would be practically valuable for multi-platform SLAM. Strengths include leave-one-out training, external PR/pose metrics, dual-task evaluation, and explicit local-vs-global and component ablations on KITTI.

major comments (3)

Abstract and §§IV-D–IV-E claim comprehensive multi-sensor evaluation and successful single-shot localization / map alignment on ARL (Ouster OS-1) and UCLA (Livox Mid-360) without retraining. Support is only qualitative overlays (Figs. 1, 4, 5) and statements that maps “align correctly/accurately.” No PR curves, F1/AP, TE/RE, recall, or inlier-ratio statistics are given for these domains, nor any failure analysis for low-overlap or FOV/density mismatch. The quantitative backbone (Tables I–III, Figs. 3, 6–7) is KITTI-only. This underpins a load-bearing part of the contribution; either add held-out quantitative metrics (or controlled failure cases) on the non-KITTI sets, or clearly scope the multi-sensor claims as qualitative only.
§III-E and §IV-A: outlier rejection relies on “predefined thresholds” for GICP alignment error (Mahalanobis) and post-alignment inlier ratio, and place recognition uses a distance threshold on d(P,Q) in Eq. (2). Numerical values of these thresholds (and sensitivity) are not reported. Because false-positive rejection is described as essential to map integrity, the reported high recall/precision and “successful” alignments cannot be fully reproduced or stress-tested without them; please state the operating thresholds used for Tables I–II and the qualitative experiments.
Table II / §IV-C: registration comparison removes RANSAC from LCDNet and GICP from PinNet for a “descriptor-only” setting, then separately reports PinNet with GICP. That is reasonable, but LCDNet’s published pipeline includes a robust estimator; without a like-for-like full-pipeline comparison (or reporting LCDNet+GICP / LCDNet+RANSAC under the same TE/RE/recall criteria), the large gap in recall and TE/RE may partly reflect estimator choice rather than descriptor quality alone. A short full-pipeline column or note would strengthen the claim that the local descriptors themselves drive the registration gains.

minor comments (6)

Abstract: “scanto-scan” → “scan-to-scan”.
§III-C Eq. (1) and following: β_d / β_m and β_n are used inconsistently in text vs. implementation (§IV-A uses β_m, β_n); unify notation.
Table I: PinNet is second on KITTI 05 F1/AP while best on average; a brief discussion of sequence-05 failure modes (structure, loop density) would help interpret when plane embeddings help or hurt.
§IV-A: loss combination weights for circle / L_T / Chamfer are not listed among the free parameters; please report them for reproducibility.
Fig. 2 caption and pipeline: “smallest n pairwise distances” vs. body text “s”; keep a single symbol.
Related work could briefly position against other recent local-descriptor / keypoint LiDAR PR methods beyond the five baselines to clarify novelty of the plane-based embedding relative to Geometric Transformer [18].

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical architecture + contrastive training evaluated on held-out KITTI metrics and external qualitative datasets; self-cites are tooling only.

full rationale

PinNet proposes a KPConv-based keypoint detector, plane-augmented geometric self-attention (Mahalanobis + surface-angle embeddings plus Euclidean/angular terms from prior non-overlapping work), and contrastive circle + transformation + Chamfer losses. Positive/negative pairs are defined by geometric patch overlap thresholds independent of the network outputs; the network is then trained to minimize descriptor distances on those pairs. Place-recognition and registration claims are scored with standard external metrics (F1-max, AP, TE, RE, recall) on leave-one-out KITTI sequences against published baselines; ablations remove components and re-measure the same metrics. Self-citations ([3],[20],[25]) supply only the keyframe/map generator used in qualitative ARL/UCLA experiments and do not define the descriptors, losses, or evaluation criteria. No equation reduces a claimed prediction to a fitted input by construction, no uniqueness theorem is imported from the authors, and no known empirical pattern is merely renamed. The derivation chain is therefore self-contained and non-circular.

Assumptions & free parameters 9 free parameters · 5 assumptions · 2 invented entities

The central empirical claim rests on standard deep point-cloud learning assumptions plus several hand-chosen architectural and training hyperparameters. No new physical entities are postulated; 'PinNet' and the plane-based attention module are engineered constructs whose value is only the reported metrics. Free parameters are the usual ML knobs that affect reported F1/TE/RE.

free parameters (9)

descriptor dimension = 256
Fixed at 256; controls capacity and matching cost; not derived.
number of matched keypoint pairs s = 256
s=256 used for inter-scan distance and SVD; chosen by authors.
voxel leaf size = 0.3 m
0.3 m input voxelization; affects sparsity and runtime.
k nearest dense points for keypoint aggregation = 64
k=64 (with 2k random dilation); design choice for keypoint consistency.
self-attention depth N_sa = 3
Set to 3 plane-based geometric self-attention layers.
embedding scales β_d, β_n = 1/4.8, 1/15°
β_d=1/4.8 and β_n=1/15° set sensitivity of Mahalanobis and normal-angle embeddings.
positive-pair overlap threshold = 10%
Local patches with ≥10% overlap labeled positive for contrastive training; defines supervision.
loop-closure and GICP rejection thresholds
Descriptor-distance, inlier-ratio, and Mahalanobis/GICP error cutoffs used to accept/reject loops; values not fully tabulated but load-bearing for reported precision.
loss combination weights
Overall loss is a weighted sum of circle, transformation, and Chamfer terms; weights not numerically specified.

assumptions (5)

domain assumption KPConv-FPN multi-scale features plus learned weighted aggregation of dense neighbors yield more viewpoint-consistent keypoints than voxel-grid downsampling alone.
Section III-B; supported by ablation residual figure but not proved.
ad hoc to paper Plane-level relations (Mahalanobis distance of local covariances and surface-normal angle) plus Euclidean/triplet angle embeddings improve local descriptor discriminability for place recognition.
Section III-C design; ablation shows modest F1/AP gains when removed.
domain assumption Contrastive circle loss on patch-overlap positives/negatives plus differentiable SVD transform loss and Chamfer keypoint loss produce descriptors suitable for both retrieval and registration.
Section III-D; standard in learned registration literature.
domain assumption Inter-scan distance as mean of s smallest pairwise descriptor distances is a valid place-similarity score, and SVD+GICP with inlier/error checks rejects false loops adequately.
Section III-E pipeline definition.
domain assumption True-positive loop if ground-truth distance <5 m; exclude last 50 frames; RE<5° and TE<2 m define successful registration.
Section IV-B/C evaluation protocol aligned with prior place-recognition papers.

invented entities (2)

PinNet (keypoint-aware local descriptor network)
purpose: End-to-end extract sparse keypoints and descriptors for LiDAR loop closure and registration.
Named system; value is empirical performance only; no independent physical existence.
Plane-based geometric self-attention module
purpose: Inject Mahalanobis and surface-angle relations among keypoints into local descriptors.
Architectural module built on prior geometric transformers; validated only by ablations in this paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of PinNet: Keypoint-Aware Learned Local Descriptors with Geometric Embedding for Loop Closure in LiDAR SLAM." pith.science (2026). https://pith.science/paper/3MCYLCBV

@misc{pith2026260628637,
  author       = {Pith},
  title        = {Pith review of: PinNet: Keypoint-Aware Learned Local Descriptors with Geometric Embedding for Loop Closure in LiDAR SLAM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3MCYLCBV}},
  note         = {Machine review of arXiv:2606.28637}
}

read the original abstract

Loop closure is essential to reduce drift and build globally consistent maps in large-scale environments. However, reliable loop closure with only geometric information from, e.g., a LiDAR sensor, remains challenging due to the difficulty of constructing discriminative geometric features. We present PinNet, a neural network that produces local geometric descriptors from point clouds for place recognition and scanto-scan registration. PinNet incorporates a neural network that generates keypoints and their corresponding descriptors, together with a plane-based geometric self-attention module that models inter-keypoint spatial relationships to enhance descriptor discriminability for loop-closure detection and point-cloud registration. The approach is comprehensively evaluated on multiple datasets collected with different LiDAR sensors. Experimental results demonstrate strong place-recognition performance, precise relative pose estimation, and successful single-shot localization in different environments.

Figures

Figures reproduced from arXiv: 2606.28637 by the authors.

**Figure 1.** PinNet provides a robust and accurate place recognition and map alignment method based on learned local geometric descriptors for point clouds. A point cloud acquired by the robot used in this work (bottom right) is fed into the PinNet, which generates a set of discriminative local descriptors (top middle). A database of scans from the current or previous mapping session is then queried (left half) for potential mat… view at source ↗

**Figure 2.** The proposed framework consists of three modules: (i) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Precision-Recall curves for KITTI dataset. The proposed method outperforms the baselines on sequences 00, 02, and 07, while LoGG3D-Net achieves the best performance on sequence 05. and the second-best is underlined. Our method achieved the highest F1-max score on sequences 00, 02, 06, and 07, while LoGG3D-Net performed the best on sequence 05. Regarding Average Precision (AP), our approach ranked first on sequences … view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Single-shot localization. The global map is visualized in white, the query scans (A2, B2, C2) in red, and the retrieved partitioned maps (A3, B3, C3) in green. The alignment results show that the query scans are accurately registered to the corresponding regions of the…

**Figure 5.** Figure 5: Accurate alignment of point clouds from different viewpoints on the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Ablation study of keypoint detection. Left: No keypoint detection. Right: Keypoint detection. Keypoint detection significantly reduces the residuals between matched points. full 6-DoF pose of a query scan directly within a prebuilt large-scale point-cloud map. This set…

**Figure 7.** Figure 7: Precision–Recall curves on the KITTI dataset. Results show that the local-descriptor approach consistently outperforms the global descriptor across all sequences. matched points, measured as Euclidean distances between the corresponding points used in point cloud regis…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 2 linked inside Pith

[1]

Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping,

T. Shanet al., “Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping,” inIEEE/RSJ Int. Conf. Intell. Robots Syst., 2020, pp. 5135–5142

2020
[2]

Fast-lio2: Fast direct lidar-inertial odometry,

W. Xu, Y . Cai, D. He, J. Lin, and F. Zhang, “Fast-lio2: Fast direct lidar-inertial odometry,”IEEE Trans. Robot., vol. 38, no. 4, pp. 2053– 2073, 2022
[3]

Direct lidar-inertial odome- try: Lightweight lio with continuous-time motion correction,

K. Chen, R. Nemiroff, and B. T. Lopez, “Direct lidar-inertial odome- try: Lightweight lio with continuous-time motion correction,” inIEEE Int. Conf. Robot. Autom., 2023, pp. 3983–3989

2023
[4]

Fast point feature histograms (fpfh) for 3d registration,

R. B. Rusu, N. Blodow, and M. Beetz, “Fast point feature histograms (fpfh) for 3d registration,” inIEEE Int. Conf. Robot. Autom., 2009, pp. 3212–3217

2009
[5]

M2dp: A novel 3d point cloud descriptor and its application in loop closure detection,

L. He, X. Wang, and H. Zhang, “M2dp: A novel 3d point cloud descriptor and its application in loop closure detection,” inIEEE/RSJ Int. Conf. Intell. Robots Syst., 2016, pp. 231–237

2016
[6]

Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map,

G. Kim and A. Kim, “Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map,” inIEEE/RSJ Int. Conf. Intell. Robots Syst., 2018, pp. 4802–4809

2018
[7]

Ndt-transformer: Large-scale 3d point cloud localisa- tion using the normal distribution transform representation,

Z. Zhouet al., “Ndt-transformer: Large-scale 3d point cloud localisa- tion using the normal distribution transform representation,” inIEEE Int. Conf. Robot. Autom., 2021, pp. 5654–5660

2021
[8]

Overlapnet: A siamese network for computing lidar scan similarity with applications to loop closing and localization,

X. Chenet al., “Overlapnet: A siamese network for computing lidar scan similarity with applications to loop closing and localization,” Auton. Robots, vol. 46, no. 1, pp. 61–81, 2022

2022

Show all 25 references

[9]

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

J. Maet al., “Overlaptransformer: An efficient and yaw-angle-invariant transformer network for lidar-based place recognition,”IEEE Robot. Autom. Lett., vol. 7, no. 3, pp. 6958–6965, 2022

2022
[10]

Minkloc3d: Point cloud based large-scale place recognition,

J. Komorowski, “Minkloc3d: Point cloud based large-scale place recognition,” inProc. IEEE/CVF Winter Conf. Appl. Comput. Vis., 2021, pp. 1790–1799

2021
[11]

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

K. Vidanapathiranaet al., “Logg3d-net: Locally guided global de- scriptor learning for 3d place recognition,” inIEEE Int. Conf. Robot. Autom., 2022, pp. 2215–2221

2022
[12]

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

C. Yuanet al., “Btc: A binary and triangle combined descriptor for 3-d place recognition,”IEEE Trans. Robot., vol. 40, pp. 1580–1599, 2024

2024
[13]

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

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

2021
[14]

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,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2018, pp. 4470–4479

2018
[15]

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,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2017, pp. 652–660

2017
[16]

Lcdnet: Deep loop closure detection and point cloud registration for lidar slam,

D. Cattaneo, M. Vaghi, and A. Valada, “Lcdnet: Deep loop closure detection and point cloud registration for lidar slam,”IEEE Trans. Robot., vol. 38, no. 4, pp. 2074–2093, 2022
[17]

Deep global registration,

C. Choy, W. Dong, and V . Koltun, “Deep global registration,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2020, pp. 2514– 2523

2020
[18]

Geometric transformer for fast and robust point cloud registration,

Z. Qinet al., “Geometric transformer for fast and robust point cloud registration,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2022, pp. 11 143–11 152

2022
[19]

Generalized-icp

A. Segal, D. Haehnel, and S. Thrun, “Generalized-icp.” inRob.: Sci. Syst., vol. 2, no. 4, 2009, p. 435

2009
[20]

Direct lidar-inertial odometry and mapping: Perceptive and connective slam,

K. Chen, R. Nemiroff, and B. T. Lopez, “Direct lidar-inertial odometry and mapping: Perceptive and connective slam,”arXiv:2305.01843, 2023

2023 arXiv
[21]

Kpconv: Flexible and deformable convolution for point clouds,

H. Thomaset al., “Kpconv: Flexible and deformable convolution for point clouds,” inProc. IEEE/CVF Int. Conf. Comput. Vis., 2019, pp. 6411–6420

2019
[22]

Rskdd-net: Random sample-based keypoint detector and descriptor,

F. Lu, G. Chen, Y . Liu, Z. Qu, and A. Knoll, “Rskdd-net: Random sample-based keypoint detector and descriptor,”Adv. Neural Inf. Process. Syst., vol. 33, pp. 21 297–21 308, 2020

2020
[23]

Circle loss: A unified perspective of pair similarity op- timization,

Y . Sunet al., “Circle loss: A unified perspective of pair similarity op- timization,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2020, pp. 6398–6407

2020
[24]

Netvlad: Cnn architecture for weakly su- pervised place recognition,

R. Arandjelovicet al., “Netvlad: Cnn architecture for weakly su- pervised place recognition,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2016, pp. 5297–5307

2016
[25]

Optmap: Geometric map distillation via submodular maximization,

D. Thorne, N. Chan, C. S. Robison, P. R. Osteen, and B. T. Lopez, “Optmap: Geometric map distillation via submodular maximization,” arXiv:2512.07775, 2025

2025 arXiv

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.

[1] [1]

Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping,

T. Shanet al., “Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping,” inIEEE/RSJ Int. Conf. Intell. Robots Syst., 2020, pp. 5135–5142

2020

[2] [2]

Fast-lio2: Fast direct lidar-inertial odometry,

W. Xu, Y . Cai, D. He, J. Lin, and F. Zhang, “Fast-lio2: Fast direct lidar-inertial odometry,”IEEE Trans. Robot., vol. 38, no. 4, pp. 2053– 2073, 2022

[3] [3]

Direct lidar-inertial odome- try: Lightweight lio with continuous-time motion correction,

K. Chen, R. Nemiroff, and B. T. Lopez, “Direct lidar-inertial odome- try: Lightweight lio with continuous-time motion correction,” inIEEE Int. Conf. Robot. Autom., 2023, pp. 3983–3989

2023

[4] [4]

Fast point feature histograms (fpfh) for 3d registration,

R. B. Rusu, N. Blodow, and M. Beetz, “Fast point feature histograms (fpfh) for 3d registration,” inIEEE Int. Conf. Robot. Autom., 2009, pp. 3212–3217

2009

[5] [5]

M2dp: A novel 3d point cloud descriptor and its application in loop closure detection,

L. He, X. Wang, and H. Zhang, “M2dp: A novel 3d point cloud descriptor and its application in loop closure detection,” inIEEE/RSJ Int. Conf. Intell. Robots Syst., 2016, pp. 231–237

2016

[6] [6]

Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map,

G. Kim and A. Kim, “Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map,” inIEEE/RSJ Int. Conf. Intell. Robots Syst., 2018, pp. 4802–4809

2018

[7] [7]

Ndt-transformer: Large-scale 3d point cloud localisa- tion using the normal distribution transform representation,

Z. Zhouet al., “Ndt-transformer: Large-scale 3d point cloud localisa- tion using the normal distribution transform representation,” inIEEE Int. Conf. Robot. Autom., 2021, pp. 5654–5660

2021

[8] [8]

Overlapnet: A siamese network for computing lidar scan similarity with applications to loop closing and localization,

X. Chenet al., “Overlapnet: A siamese network for computing lidar scan similarity with applications to loop closing and localization,” Auton. Robots, vol. 46, no. 1, pp. 61–81, 2022

2022