Pith. sign in

REVIEW 2 major objections 5 minor 31 references

IBPA turns streaming point clouds into orientable manifold meshes while flagging real holes in real time.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 04:16 UTC pith:3HVHY3FG

load-bearing objection Solid engineering conversion of classic BPA into a streaming, topologically clean mesher with online hole flags; usable now for underwater coverage, with one acknowledged heuristic soft spot. the 2 major comments →

arxiv 2607.11627 v1 pith:3HVHY3FG submitted 2026-07-13 cs.GR cs.RO

IBPA: Real-time Free-form Manifold Mesh Reconstruction via Incremental Ball Pivoting with Integrated Hole Detection

Mauhing Yip , Mohit Singh , Kostas Alexis , Christian Schellewald , Annette Stahl This is my paper
classification cs.GR cs.RO
keywords orientable manifold triangle meshhole detectionincremental surface reconstructionunderwater roboticsmultibeam sonarBall Pivoting Algorithmfree-form mesh
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Underwater surveys often finish only to reveal gaps in coverage, and the usual height-map models cannot even represent overhangs or walls. This paper shows that the classic Ball Pivoting Algorithm can be rewritten as a streaming process that keeps a single, consistently oriented, manifold triangle mesh as new points arrive, with no assumptions about overlap or density. At the same time it extracts closed boundary loops and marks those longer than a user threshold as missing data that still needs to be scanned. The result is a free-form surface the operator can trust during the mission rather than after it, so expensive ROV and AUV time is spent closing real holes instead of re-surveying the whole site. Experiments on multibeam and vision data confirm that the mesh stays manifold, rejects outliers, and correctly highlights incomplete regions both offline and online.

Core claim

The authors demonstrate that Ball Pivoting can be made fully incremental while preserving edge- and vertex-manifoldness and global orientability: an expandable octree admits new points, empty-ball violations delete only the offending triangles, non-manifold vertices are pruned after each batch, and orientation is enforced on every accepted facet. The same boundary-edge set that drives expansion also yields closed loops that are classified as true holes once their length exceeds a user threshold.

What carries the argument

Incremental Ball Pivoting (IBPA): dynamic octree + empty-ball facet deletion + post-batch removal of 0-disk and 1-disk non-manifold vertices + orientation consistency check, all feeding a boundary-loop extractor that isolates holes without a full-mesh scan.

Load-bearing premise

That periodically keeping only the single longest edge-connected component correctly separates genuine holes from temporary fragmentation caused by empty-ball deletions, without throwing away valid surface that happens to be disconnected for a moment.

What would settle it

Run the algorithm on a known watertight free-form surface that is deliberately split into two large edge-connected pieces by empty-ball deletions; if the longer component is retained and the shorter legitimate patch is permanently discarded, or if a true hole is misclassified as a fragment, the claim fails.

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

If this is right

  • Operators can replan ROV/AUV paths mid-mission using only the red hole boundaries, without random next-best-view sampling.
  • Free-form manifold meshes replace DTMs for any bathymetry that contains overhangs or vertical faces.
  • Outlier scans are rejected by the empty-ball test, so navigation spikes do not corrupt the growing surface.
  • The same pipeline works offline on logged multibeam data and online on vision-based stereo streams.
  • Detected holes supply the exact perimeter needed by subsequent hole-filling pose planners.

Where Pith is reading between the lines

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

  • Because the mesh is already an orientable 2-manifold with boundary, any later smoothing step yields a domain on which Stokes’ theorem can be applied directly for integral geometric quantities.
  • The same boundary loops that flag missing data could be used as hard constraints for online active-view planning, turning detection into closed-loop coverage.
  • If leaf-node orphan limits are made adaptive to local point density, the method could maintain roughly constant reconstruction time even as the map grows.
  • Extending the empty-ball test with a line-of-sight filter would further suppress multipath sonar artifacts common in confined underwater spaces.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper adapts the classical Ball Pivoting Algorithm into Incremental BPA (IBPA) that constructs an orientable, edge- and vertex-manifold, non-self-intersecting triangle mesh from a stream of oriented point clouds with no assumptions on overlap or spatial distribution. It maintains the empty-ball property via dynamic octree expansion and selective facet deletion (retaining vertices), removes 0-disk and 1-disk non-manifold vertices via Algorithms 2–3, enforces orientation consistency, and flags holes whose boundary length exceeds a user threshold. The method is demonstrated on three offline multibeam AUV datasets and two online vision-based underwater experiments, with open-source ROS-integrated code.

Significance. If the claims hold, IBPA supplies a practical free-form alternative to DTM/DSM height fields for real-time coverage visualization on ROVs/AUVs, correctly representing overhangs and vertical structures while giving operators immediate feedback on incomplete regions. Strengths include a fully constructive pipeline with explicit topological invariants, an expandable octree formulation, open-source reference implementation, and both offline and online real-robot validation that shows outlier resilience and real-time speed. These elements make the work immediately usable for underwater mapping and next-best-view planning.

major comments (2)
  1. Section 5.3 (and Fig. 13): the policy of retaining only the single longest edge-connected component after every n batches (or on key-press ‘t’) is load-bearing for the hole-detection claim, yet empty-ball deletions can split a valid surface into multiple components. A temporarily shorter but geometrically correct patch is discarded and later reported as missing data. The authors note the event is “rare” and mitigated by a minimum hole-length threshold, but provide no quantification of fragmentation frequency or false-positive rate on the experimental streams; without that evidence the completeness guarantee remains an unmeasured engineering heuristic rather than a topological invariant.
  2. Section 6: reconstruction quality is assessed only by visual inspection, timing, and qualitative outlier rejection. No quantitative surface-error metrics (Hausdorff distance, normal consistency, or comparison against offline BPA/Poisson on the same point clouds) are reported. Because the central claim includes “geometric fidelity preservation,” the absence of such numbers leaves open whether the incremental deletions and non-manifold culling introduce systematic geometric degradation.
minor comments (5)
  1. Abstract and Introduction: “DSM/DSM” is a typographical duplication; the second occurrence should be “DTM”.
  2. Section 4.1.1, Eq. (3): the ceiling-log expression for Δn is correct for the 1-D illustration, yet the subsequent multi-axis example (vertex at (13,−3,2)) would benefit from an explicit statement that the final level is taken as the coordinate-wise maximum before re-coding all existing points.
  3. Figure 11 pipeline diagram: the decision diamond “keep isolated vertices?” is never referenced in the surrounding text; a one-sentence clarification of when vertices are retained versus deleted would improve readability.
  4. Section 2.1: the discussion of Voxblox and incremental Poisson correctly notes the lack of manifold guarantees, but a short remark on whether those methods can be post-processed into manifold meshes would help readers gauge the practical advantage of IBPA.
  5. Table 1: the column header “R hole tolerance” is ambiguous; expanding it to “ball radius R / hole-length tolerance” would remove the need to consult the caption.

Circularity Check

0 steps flagged

No circularity: IBPA is a constructive algorithmic adaptation of BPA whose manifold/orientability guarantees and hole flags are enforced by explicit local checks, not by fitting or self-definitional reduction.

full rationale

The paper's central claims are engineering constructions, not predictions derived from fitted parameters or uniqueness theorems. Sections 4.1–4.3 adapt classic BPA (Bernardini et al., Digne) by (i) making the octree expandable, (ii) deleting facets that violate the empty-ball condition when new points arrive, (iii) culling 0-disk and 1-disk non-manifold vertices via Algs. 2–3, and (iv) accepting a new triangle only when its orientation matches all neighboring facets. These steps are local geometric predicates; success is verified by the resulting mesh satisfying the stated topological invariants (edge- and vertex-manifold, orientable, non-self-intersecting) and by visual hole highlighting on real MBES and vision data. Self-citations to the authors' prior hole-detection / NBV work ([21], [22]) appear only in the related-work discussion of planning strategies and are not used to justify the reconstruction algorithm itself. No quantity is fitted to data and then re-presented as a prediction; no uniqueness result is imported from the same authors to force the method. The only acknowledged engineering compromise (periodic retention of the longest edge-connected component) is an explicit heuristic, not a circular derivation. Consequently the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claims rest on a small set of user-chosen geometric parameters and standard topological assumptions about surfaces; no new physical entities are postulated. The free parameters control resolution and hole sensitivity and must be set by the operator for each environment.

free parameters (4)
  • ball radius R
    Sets both reconstruction resolution and leaf size of the octree; chosen by hand per dataset (0.4–1.0 m in experiments).
  • hole-length tolerance
    User threshold (e.g., 3.16 m) that decides which boundary loops are painted red as missing data.
  • max orphan vertices per leaf
    Hard cap on points kept inside each octree leaf to bound computational cost; excess points are randomly discarded.
  • batch size n for full connectivity scan
    How often the mesh is forced to a single edge-connected component; trade-off between real-time speed and fragmentation risk.
axioms (4)
  • domain assumption Input points either carry consistent normals or normals can be estimated as vectors from each point toward the robot pose.
    Required by BPA; stated in Section 5.1.
  • standard math A surface is non-orientable if and only if it contains a Möbius strip (Zeeman).
    Used to justify orientation-consistency checks (Section 4.3).
  • ad hoc to paper Deleting triangles whose empty ball is violated by a new point, while retaining vertices, eventually restores a valid empty-ball configuration without introducing permanent self-intersections.
    Core incremental repair step (Section 4.1.2); empirically observed but not formally proved for arbitrary streams.
  • ad hoc to paper After removal of 0-disk and 1-disk non-manifold vertices the remaining mesh is both edge- and vertex-manifold.
    Claimed after Algorithms 2–3; relies on BPA’s original edge-manifold guarantee plus the cleanup loops.

pith-pipeline@v1.1.0-grok45 · 22428 in / 2569 out tokens · 24357 ms · 2026-07-14T04:16:25.843200+00:00 · methodology

0 comments
read the original abstract

Both Remotely Operated underwater Vehicles (ROVs) and Autonomous Underwater Vehicles (AUVs) are frequently deployed to acquire geometric bathymetric data. However, it is often discovered post-survey that the acquired data coverage is incomplete. Given the high operational cost associated with underwater deployments, it is essential to incrementally visualize surface coverage in real-time to support informed decision-making by both the operators of ROVs and the AUVs during data collection. In addition, traditional incremental surface reconstruction methods, such as Digital Terrain Models (DTMs), are inherently limited in expressiveness: they represent surfaces as height fields, allows only one elevation value per $(x, y)$ coordinate and thus cannot capture overhangs or vertical structures. To overcome these limitations, we adapt the original Ball Pivoting Algorithm (BPA) into an incremental, real-time, and free-form surface reconstruction method, referred to as Incremental BPA (IBPA). Our method incrementally constructs an orientable, manifold mesh from streaming point cloud data without imposing assumptions regarding point cloud overlap or spatial distribution. Furthermore, we introduce a hole detection mechanism that identifies and highlights incomplete mesh regions. Compared to existing approaches, our method supports more complex surface topologies without prior structural assumptions. The source code of our reference implementation is available: https://github.com/Mauhing/Incremental-BPA

Figures

Figures reproduced from arXiv: 2607.11627 by Annette Stahl, Christian Schellewald, Kostas Alexis, Mauhing Yip, Mohit Singh.

Figure 1
Figure 1. Figure 1: Mesh illustrations. (a) Edge-connected mesh. (b) Three edge-connected meshes [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The first three figures (a, b, and c) illustrate a failure mode of applying the [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: 1D illustration for two cases of octree expansion. When a new point lies outside [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The incremental BPA process. (a) shows the initial surface reconstruction; (b) [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Two types of non-manifold vertices. Left: the 0-disk non-manifold vertex. Right: [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: In Fig. 7b, imagine lifting the red vertex and folding it toward the blue [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: An auxiliary figure to explain the function [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: In (a), the hole can be easily detected and is fillable. In (b), if the red vertex [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Deleting facets that contain a 0-disk non-manifold vertex may also introduce a [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: In (a), a strip is shown with two edges marked in blue and red. In (b), the two [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The orange facets are oriented oppositely to the gray facets. The suggested [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: An overview of our IBPA pipeline. During facet removal, some approaches [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Illustration of a triangle mesh and its associated incentre graph. The orange [PITH_FULL_IMAGE:figures/full_fig_p025_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: An example illustrating how the removal of facets due to the empty ball [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Left: The Eelume robot ([28]) with the MBES module. Right: A closer look [PITH_FULL_IMAGE:figures/full_fig_p027_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Incremental reconstrucion. Left: Reconstruction process at approximately [PITH_FULL_IMAGE:figures/full_fig_p028_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Hole detection. Left: Point cloud of the Heinkel dataset. Visual inspection [PITH_FULL_IMAGE:figures/full_fig_p028_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Outlier handling. Top: Point cloud of the Heinkel dataset. Bottom: Recon [PITH_FULL_IMAGE:figures/full_fig_p029_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Hole detection. Left: Point cloud of the Nyhavna dataset. Right: Recon [PITH_FULL_IMAGE:figures/full_fig_p029_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Consistency check. Left: Point cloud of Figaro dataset. Right: Reconstructed [PITH_FULL_IMAGE:figures/full_fig_p030_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: (a): The Marine Cybernetics Laboratory (MC-Lab) at NTNU. (b): The Ariel [PITH_FULL_IMAGE:figures/full_fig_p031_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Results of our IBPA method with missing information detection are shown. The [PITH_FULL_IMAGE:figures/full_fig_p032_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Trondheim Harbour experimental site. The mapping area is marked in orange. [PITH_FULL_IMAGE:figures/full_fig_p033_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Hole detection in an online outdoor experiment. The yellow point cloud repre [PITH_FULL_IMAGE:figures/full_fig_p033_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: The surface reconstruction was successfully completed with minimal missing [PITH_FULL_IMAGE:figures/full_fig_p034_24.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

31 extracted references · 1 canonical work pages

  1. [1]

    Sheinin, Y

    M. Sheinin, Y. Y. Schechner, The next best underwater view, in: Pro- ceedings of the IEEE conference on computer vision and pattern recog- nition, 2016, pp. 3764–3773. 38

    2016

  2. [2]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, R. Ng, Nerf: Representing scenes as neural radiance fields for view synthesis, Communications of the ACM 65 (2021) 99–106

    2021

  3. [3]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimk¨ uhler, G. Drettakis, 3d gaussian splat- ting for real-time radiance field rendering., ACM Trans. Graph. 42 (2023) 139–1

    2023

  4. [4]

    Campos, R

    R. Campos, R. Garcia, P. Alliez, M. Yvinec, A surface reconstruction method for in-detail underwater 3d optical mapping, The International Journal of Robotics Research 34 (2015) 64–89

    2015

  5. [5]

    Bernardini, J

    F. Bernardini, J. Mittleman, H. Rushmeier, C. Silva, G. Taubin, The ball-pivoting algorithm for surface reconstruction, IEEE transactions on visualization and computer graphics 5 (2002) 349–359

    2002

  6. [6]

    Digne, An analysis and implementation of a parallel ball pivoting algorithm, Image Processing On Line 4 (2014) 149–168

    J. Digne, An analysis and implementation of a parallel ball pivoting algorithm, Image Processing On Line 4 (2014) 149–168

    2014

  7. [7]

    R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, A. J. Davison, P. Kohi, J. Shotton, S. Hodges, A. Fitzgibbon, Kinectfusion: Real-time dense surface mapping and tracking, in: 2011 10th IEEE International Symposium on Mixed and Augmented Reality, 2011, pp. 127–136. doi:10.1109/ISMAR.2011.6092378

    doi:10.1109/ismar.2011.6092378 2011

  8. [8]

    Curless, M

    B. Curless, M. Levoy, A volumetric method for building complex models from range images, in: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996, pp. 303–312

    1996

  9. [9]

    Oleynikova, Z

    H. Oleynikova, Z. Taylor, M. Fehr, R. Siegwart, J. Nieto, Voxblox: Incremental 3D Euclidean Signed Distance Fields for on-board MAV planning, in: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE, Vancouver, BC, 2017, pp. 1366–

    2017

  10. [10]

    doi:10.1109/IROS.2017.8202315

    doi:10.1109/iros.2017.8202315 2017

  11. [11]

    Romanoni, M

    A. Romanoni, M. Matteucci, Efficient moving point handling for in- cremental 3d manifold reconstruction, in: International Conference on Image Analysis and Processing, Springer, 2015, pp. 489–499. 39

    2015

  12. [12]

    Piazza, A

    E. Piazza, A. Romanoni, M. Matteucci, Real-time cpu-based large-scale three-dimensional mesh reconstruction, IEEE Robotics and Automation Letters 3 (2018) 1584–1591

    2018

  13. [13]

    Sch¨ ops, T

    T. Sch¨ ops, T. Sattler, M. Pollefeys, Surfelmeshing: Online surfel-based mesh reconstruction, IEEE transactions on pattern analysis and ma- chine intelligence 42 (2019) 2494–2507

    2019

  14. [14]

    Kazhdan, M

    M. Kazhdan, M. Bolitho, H. Hoppe, Poisson surface reconstruction, in: Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006

    2006

  15. [15]

    Q. Yu, W. Sui, Y. Wang, S. Xiang, C. Pan, Incremental poisson sur- face reconstruction for large scale three-dimensional modeling, in: Pat- tern Recognition and Computer Vision: Second Chinese Conference, PRCV 2019, Xi’an, China, November 8–11, 2019, Proceedings, Part III 2, Springer, 2019, pp. 442–453

    2019

  16. [16]

    Vizzo, X

    I. Vizzo, X. Chen, N. Chebrolu, J. Behley, C. Stachniss, Poisson surface reconstruction for lidar odometry and mapping, in: 2021 IEEE interna- tional conference on robotics and automation (ICRA), IEEE, 2021, pp. 5624–5630

    2021

  17. [17]

    J. Ruan, B. Li, Y. Wang, Y. Sun, Slamesh: Real-time lidar simultaneous localization and meshing, in: 2023 IEEE International Conference on Robotics and Automation (ICRA), IEEE, 2023, pp. 3546–3552

    2023

  18. [18]

    Niedzwiedzki, P

    J. Niedzwiedzki, P. Lipinski, L. Podsedkowski, IDTMM: Incremental Direct Triangle Mesh Mapping, IEEE Robotics and Automation Letters (2023) 1–8. doi:10.1109/LRA.2023.3293751

    work page doi:10.1109/lra.2023.3293751 2023

  19. [19]

    Palomeras, N

    N. Palomeras, N. Hurt´ os, E. Vidal, M. Carreras, Autonomous explo- ration of complex underwater environments using a probabilistic next- best-view planner, IEEE Robotics and Automation Letters 4 (2019) 1619–1625

    2019

  20. [20]

    next-best-view

    A. Bircher, M. Kamel, K. Alexis, H. Oleynikova, R. Siegwart, Receding horizon “next-best-view” planner for 3d exploration, in: 2016 IEEE in- ternational conference on robotics and automation (ICRA), IEEE, 2016, pp. 1462–1468. 40

    2016

  21. [21]

    Hornung, K

    A. Hornung, K. M. Wurm, M. Bennewitz, C. Stachniss, W. Burgard, Octomap: An efficient probabilistic 3d mapping framework based on octrees, Autonomous robots 34 (2013) 189–206

    2013

  22. [22]

    M. Yip, C. Schellewald, T. Gambin, A. Stahl, Active pose-autonomous hole-filling techniques for 3d surface reconstruction in underwater ex- ploration, IFAC-PapersOnLine 58 (2024) 293–300

    2024

  23. [23]

    M. Yip, A. Stahl, C. Schellewald, Robust hole-detection in triangular meshes irrespective of the presence of singular vertices, Computer-Aided Design 170 (2024) 103696

    2024

  24. [24]

    R. A. Adams, C. Essex, Calculus: a complete course, 9th edition, Pear- son, 2018

    2018

  25. [25]

    Zeeman, An introduction to topology: The classification theorem for surfaces, (No Title) (1966)

    E.-C. Zeeman, An introduction to topology: The classification theorem for surfaces, (No Title) (1966)

    1966

  26. [26]

    C. H. S´ equin, From moebius bands to klein-knottles, in: Proceedings of Bridges 2012: Mathematics, Music, Art, Architecture, Culture, 2012, pp. 93–102

    2012

  27. [27]

    R. P. Grimaldi, Discrete and Combinatorial Mathematics, 5/e, Pearson Education India, 2006

    2006

  28. [28]

    Q.-Y. Zhou, J. Park, V. Koltun, Open3d: A modern library for 3d data processing, arXiv preprint arXiv:1801.09847 (2018)

    Pith/arXiv arXiv 2018

  29. [29]

    Liljeb¨ ack, R

    P. Liljeb¨ ack, R. Mills, Eelume: A flexible and subsea resident imr vehi- cle, in: Oceans 2017-Aberdeen, IEEE, 2017, pp. 1–4

    2017

  30. [30]

    Singh, K

    M. Singh, K. Alexis, Online refractive camera model calibration in visual inertial odometry, in: 2024 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS), 2024, pp. 12609–12616. doi:10.1109/IROS58592.2024.10802302

    doi:10.1109/iros58592.2024.10802302 2024

  31. [31]

    Singh, K

    M. Singh, K. Alexis, Deepvl: Dynamics and inertial measurements- based deep velocity learning for underwater odometry, in: 2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 1–7. doi:10.1109/ICRA55743.2025.11128041. 41

    doi:10.1109/icra55743.2025.11128041 2025