Pith. sign in

REVIEW 3 major objections 6 minor 39 references

CCNDF: Curvature Constrained Neural Distance Fields from 3D LiDAR Sequences

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Supervising a neural distance field with a curvature-derived signed-distance estimate improves LiDAR mapping and localization over ray-distance and first-order normal baselines.

desk verdict The curvature-supervision idea is new, but Eq. (8) does not compute distance to the nearest surface for general geometry, and the reported gains are not verifiable. read the letter →

arxiv 2412.15909 v1 pith:MRAHBKPX submitted 2024-12-20 cs.CV cs.GR

Akshit Singh , Karan Bhakuni , Rajendra Nagar This is my paper · ORCID
classification cs.CVcs.GR
keywords neuraldistancefieldscurvatureestimationsignedsupervisionLiDARmappingself-supervisedlearningimplicitrepresentationsMonteCarlolocalization
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

Neural distance fields (NDFs) — neural networks that map a 3D point to its signed distance from the nearest surface — are attractive for LiDAR mapping but normally need ground-truth distance values that do not exist for outdoor scenes. The paper's claim is that a better self-supervised training signal can be derived from the field's own curvature: at any query point along a LiDAR ray, the signed distance to the surface is approximated as the difference between the radius of curvature of the NDF isoline at that point and the radius of the osculating circle through the ray's surface endpoint. The proposed estimate, $\hat{d}_\ell = R_\ell - \sqrt{d_\ell^2 + R_\ell^2 - 2 R_\ell n_\ell^\top (e_i - x_\ell)}$, is shown on a toy example to land at 0.98 where the true signed distance is 1, while ray distance gives 1.41 and LocNDF's normal-projected distance gives 1.75. If the claim holds, the same raw LiDAR scan yields sharper maps (cars, trees, building edges) and more reliable Monte Carlo localization without any new sensors or ground-truth meshes. The paper's experiments support this with qualitative reconstructions on KITTI and Apollo ColumbiaPark-3 and consistently lower RMSE/MAE localization numbers than the compared baselines.

What carries the argument

The load-bearing object is the curvature-constrained distance estimate of Equation (8), $\hat{d}_\ell = R_\ell - \sqrt{d_\ell^2 + R_\ell^2 - 2 R_\ell n_\ell^\top (e_i - x_\ell)}$. The radius of curvature $R_\ell$ of the NDF isoline is computed from the divergence of the normalized gradient of the field, $R_\ell = \left\|\nabla \cdot \left(\frac{\nabla D}{\|\nabla D\|}\right)\right\|^{-1}$; the cosine rule in the triangle with the ray endpoint $e_i$ and the curvature center $c_\ell$ then yields the radius $r_\ell$ of the circle through the endpoint. Its role is to turn the nearest-surface direction $n_\ell$ (a first-order quantity) into a scalar distance that respects the curvature of the surface's level sets, and that scalar supervises the MLP through the weighted loss over log-linearly sampled ray points.

What would settle it

On a synthetic scene with known geometry (a sphere of radius 1 and a flat plane), simulate LiDAR rays that hit the surface at points different from the nearest point to samples along the ray; compute the proposed estimate $\hat{d}_\ell$ at each sample and compare with the analytic signed distance. If the estimate's error on off-center sphere rays or on flat regions is not smaller than the errors of raw ray distance and LocNDF's normal projection, the central geometric advantage asserted by the paper would be refuted.

Watch

Extended reading notes

Core claim

The central discovery is that the signed distance from a query point to the nearest surface can be read off the curvature structure of the learned distance field itself. Because the field's level sets are claimed to be concentric with the surface, the true distance is the difference between the radius of curvature $R_\ell$ of the isoline through the query point and the radius $r_\ell$ of the circle through the ray endpoint on the surface, both referred to the same curvature center; the paper derives $r_\ell$ from the cosine rule in the triangle formed by the query point, the ray endpoint, and that center (Eq. 7), giving $\hat{d}_\ell = R_\ell - r_\ell$ (Eq. 8). The paper argues that this second-order supervision is more faithful than raw ray distance, which ignores the geometry of the rest of the surface, and than LocNDF's first-order normal projection, which assumes locally linear surfaces and overshoots on curved ones. It then trains the same SIREN-based MLP with the new distance as the supervision target and reports that the resulting maps capture finer detail and that 2D Monte Carlo localization on five sequences achieves lower RMSE/MAE than the LocNDF, IR-MCL, SRRG, and AMCL baselines.

Load-bearing premise

The derivation assumes the LiDAR ray endpoint lies on the same curvature circle as the nearest surface point to the query point, which is exact for spheres and other circular geometries but not for general surfaces.

Editorial extensions

If this is right

  • Mapping pipelines that currently supervise neural fields with raw ray distance could collect the same LiDAR scan and reconstruct finer geometry — cars, trees, building edges — without any new sensors or ground-truth meshes.
  • Because the estimate is expressed through the field's own curvature, the supervision becomes a self-consistent second-order constraint: better fields give better distances, which feed better fields, instead of relying on a fixed first-order projection.
  • The improved map representation transfers to downstream tasks: 2D Monte-Carlo localization reports lower RMSE and MAE on five sequences, and the error values saturate rather than growing with the threshold from 5 to 20 cm.
  • The approach extends self-supervised NDF supervision from first-order properties (direction to nearest surface) to second-order properties (curvature), opening the same machinery to other range sensors such as depth cameras.

Reading between the lines

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

  • If the curvature principle carries over to signed distance functions of general surfaces, the same Equation (8) could in principle be applied to any range sensor — RGB-D cameras, sonar — where ray endpoints and field curvature are available, not just LiDAR point clouds.
  • The geometric derivation suggests a testable scaling law: the error of ray-distance supervision should grow with local curvature, while the proposed estimate should stay bounded, so sorting scenes by curvature could predict which scenes benefit most.
  • A natural extension the authors do not develop is online incremental training: since the curvature estimate depends on the currently learned field, streamed LiDAR would need a stabilization scheme, though the log-linear ray sampling they use may already provide one.
  • The concentric-isolines argument is exact for spheres and cylinders; for general surfaces it is an approximation whose quality likely degrades with the ratio of local curvature radius to the distance between the query point and the ray endpoint, a quantity that could be computed from the learned field itself.
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

3 major / 6 minor

Summary. The manuscript proposes CCNDF, a self-supervised method for training neural distance fields (NDF) from LiDAR sequences. The central idea is to compute, for each sampled ray point, a 'curvature-constrained' signed distance estimate that uses the second-order differential structure of the current NDF, and to supervise the network with this estimate. The authors claim that this yields more accurate mapping and localization than ray distance and LocNDF's first-order estimate. The evaluation consists of qualitative mapping results on Apollo and KITTI datasets and quantitative 2D Monte Carlo localization results on five sequences, compared with LocNDF and several classic baselines.

Significance. If the proposed geometric estimator were correct, using curvature information to supervise neural distance fields would be a meaningful extension of first-order methods, and the localization benchmark could provide practical value. The paper is clearly written in parts and the authors identify a genuine limitation of existing supervision signals. However, the derivation of the estimator is not valid for general surfaces, and the paper's own limitation section does not resolve the issue. The idea is worth exploring, but the current formulation and evaluation do not support the central claim.

major comments (3)
  1. [Section 3.1, Eqs. (6)-(8) and Fig. 2(a)] The derivation sets ||e_i - c_l|| = r_l, where e_i is the LiDAR ray endpoint, c_l is the curvature center of the NDF isoline at the query point x_l, and r_l is the radius of curvature at the nearest surface point F. This equality holds only if the ray endpoint lies on the same circle as F, i.e., only for circular (or spherical) surfaces. For a general surface, e_i is not equal to F, and the distance from e_i to c_l is not r_l; therefore the cosine rule in Eq. (7) does not produce the radius of curvature at the nearest surface point, and Eq. (8) is not an approximation of the true signed distance to the nearest surface. The toy example in Fig. 2(b) is a circle, where the assumption happens to hold, so it cannot validate the general claim. The 'Geometric Limitation' paragraph in Section 5 acknowledges that the ray may miss the curvature circle at F but asserts that near-surface sampling mitigates the error; however, expanding Eq. (8) for small ray distance d_l gives d_hat_l ≈ d_l cosθ = n_l^T(e_i - x_l), which is exactly LocNDF's first-order estimate. Thus, in the near-surface regime where sampling weight is highest, the proposed method reduces to the method it claims to improve, and the claimed advantage is not supported.
  2. [Section 3.1, Eqs. (4)-(5)] The formulas for mean curvature and for the 'radius of curvature' of the NDF are not correct as stated. The standard mean curvature of an implicit surface is κ = (∇f^T H(f) ∇f - ||∇f||^2 trace(H)) / (2||∇f||^3), but Eq. (4) contains an extra factor of 2 in the '2trace(H)' term. Equation (5) writes the divergence of the normalized gradient as a norm of a 'gradient' with unclear notation; the divergence is a scalar, and the formula as printed is not the divergence. Moreover, for a general surface the mean curvature is not the reciprocal of a single radius of curvature; the 'radius of curvature of an isoline' is only well defined for circles or spheres. Since R_l is a load-bearing quantity in Eqs. (7)-(8), the geometric interpretation of the method is not established for surfaces with anisotropic or saddle curvature.
  3. [Section 4, Tables 1 and 2] The evaluation protocol is described only at a high level, and the reported numbers need more support. The paper states that metrics are averaged over 5 runs, but no variances are given; Table 2 shows identical RMSE values for the proposed method at 5 cm and 10 cm thresholds for Seq-1 and Seq-2 (1.4 and 1.9), which is unusual and not explained. In addition, no ablation study isolates the contribution of the curvature term from the regularizers λ1, λ2, λ3, or from the weighting scheme, so it is not clear that the empirical improvement, if any, is due to the proposed distance estimate. Since the paper's central claim is state-of-the-art performance, this lack of support is problematic.
minor comments (6)
  1. [Eq. (1)] The positional encoding definition does not specify the frequencies ω_i; please define them explicitly.
  2. [Eq. (2)] The log-linear sampling formula is not typeset clearly; the intended expression for t_l is difficult to parse and should be written with unambiguous parentheses and exponents.
  3. [Section 3.1] The text alternates between 'isolines' and 'surfaces' without clarifying that the NDF is a 3D scalar field and that the relevant objects are level surfaces, not 2D curves.
  4. [Fig. 2] The acronym 'ROC' is used extensively but is only spelled out in the caption; please define it at first use in the text.
  5. [Tables 1 and 2] Reporting mean plus/minus standard deviation and the number of runs that converged would improve the interpretability of the localization results.
  6. [Section 4] The claim 'state-of-the-art' should be tempered because the comparison set is limited to LocNDF and classical baselines; no comparison with other recent neural mapping approaches is provided.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the curvature-based supervision is an acknowledged self-referential bootstrap, but the proposed distance is not equivalent to its inputs by construction and the MCL benchmarks provide independent external grounding.

full rationale

The paper's central construction, Eq. (8), computes a proposed signed distance d_hat_l from the current NDF's gradient and curvature together with the LiDAR ray distance; it is a geometric formula, not a fitted parameter or an algebraic identity with the quantity it supervises. The training loop is indeed self-referential: the authors state 'the learning of NDF is supervised from estimated d_hat_l and d_hat_l is estimated from the learned NDF. This is a chicken-and-egg problem to solve.' That is a fixed-point/bootstrap update, but minimizing the loss in Eq. (10) does not force D to equal a precomputed constant; it drives D toward a moving target that depends on D, which is a nontrivial self-consistency condition rather than a definitional tautology. The localization evaluation against AMCL, SRRG, IR-MCL, and LocNDF provides independent, externally grounded evidence outside the fitted pseudo-labels. No load-bearing self-citations appear: the curvature formula is attributed to Goldman (an external reference) and the concentricity intuition to an external dataset/explorer paper. The acknowledged 'Geometric Limitation' (the ray endpoint e_i may not lie on the curvature circle of the closest surface point) is a correctness/robustness concern about the geometric derivation, not a circularity in the reasoning. Thus, while the method is self-supervised and bootstrap-like, it does not reduce by construction to its own inputs, and no circular step meeting the required standard is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are introduced; c_l and r_l are standard geometric constructions. The central claim rests on five stated or implied assumptions, the most fragile being that the ray endpoint can replace the nearest surface point in the curvature geometry.

free parameters (6)
  • loss weight lambda_1 = 0.1
    Weight on the SDF loss at ray endpoints; chosen empirically and affects map thickness.
  • loss weight lambda_2 = 1e-4
    Weight on the eikonal regularizer that encourages unit gradient; chosen empirically.
  • loss weight lambda_3 = 1e-3
    Weight on the normal-consistency regularizer; chosen empirically.
  • gamma = 3
    Exponent in w_l = (d_max - D)^gamma that weights near-surface samples more heavily; controls training stability and is chosen by hand.
  • positional encoding order h = 30
    Number of frequency bands in positional encoding; architectural hyperparameter.
  • number of ray intervals n_i = 40
    Number of sample points per LiDAR ray; sampling density hyperparameter.
assumptions (5)
  • domain assumption NDF isolines are concentric with the nearest surface point, so the radius of curvature at an inquiry point and at the corresponding surface point share a center.
    Stated in Section 3.1 as a property of NDF; true only for special surfaces such as spheres and cylinders, not for general signed distance fields.
  • ad hoc to paper The LiDAR ray endpoint e_i can be substituted for the nearest surface point F in the cosine rule, so ||e_i - c_l|| = r_l.
    Equation (7) uses e_i in place of F, but e_i is defined as the ray hit point and is generally not the nearest surface point.
  • standard math The mean curvature of an implicit surface is given by the norm of the gradient of the normalized gradient, as in Eq. (5).
    This curvature formula is attributed to Goldman [8], though the paper's notation is nonstandard and appears to omit a divergence; the formula is standard for level-set curvature when written correctly.
  • domain assumption The network gradient n_l = -grad D(x_l) points toward the nearest surface.
    True only if D satisfies the eikonal equation; the loss includes a weakly weighted eikonal regularizer but does not guarantee this property.
  • domain assumption The chicken-and-egg training loop converges because errors approach zero for points sampled near the surface.
    Argued in Section 3.1 after Eq. (9); this is an empirical claim about optimization dynamics, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CCNDF: Curvature Constrained Neural Distance Fields from 3D LiDAR Sequences." pith.science (2026). https://pith.science/paper/MRAHBKPX

@misc{pith2026241215909,
  author       = {Pith},
  title        = {Pith review of: CCNDF: Curvature Constrained Neural Distance Fields from 3D LiDAR Sequences},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MRAHBKPX}},
  note         = {Machine review of arXiv:2412.15909}
}
read the original abstract

Neural distance fields (NDF) have emerged as a powerful tool for addressing challenges in 3D computer vision and graphics downstream problems. While significant progress has been made to learn NDF from various kind of sensor data, a crucial aspect that demands attention is the supervision of neural fields during training as the ground-truth NDFs are not available for large-scale outdoor scenes. Previous works have utilized various forms of expected signed distance to guide model learning. Yet, these approaches often need to pay more attention to critical considerations of surface geometry and are limited to small-scale implementations. To this end, we propose a novel methodology leveraging second-order derivatives of the signed distance field for improved neural field learning. Our approach addresses limitations by accurately estimating signed distance, offering a more comprehensive understanding of underlying geometry. To assess the efficacy of our methodology, we conducted comparative evaluations against prevalent methods for mapping and localization tasks, which are primary application areas of NDF. Our results demonstrate the superiority of the proposed approach, highlighting its potential for advancing the capabilities of neural distance fields in computer vision and graphics applications.

Figures

Figures reproduced from arXiv: 2412.15909 by the authors.

Figure 1
Figure 1. Different methods for computing the signed distance to supervise NDF for a point on a ray beam. learning the NDF for large-scale environments by directly utilising the sensor readings, which can subsequently be applied to a range of tasks including path planning, mapping and localization, and more. This method leverages the ge￾ometric characteristics of the environment and also contributes to the current properties … view at source ↗
Figure 2
Figure 2. The figure illustrates the concentric nature of NDF; ROC of points on NDF constantly increase as we move away from the surface isoline at a given point is the ROC of that point on the NDF. As isolines are only formed to represent the distance to the nearest surface, this would ensure that the ROC of a point on the iso-line would be concentric with the ROC of [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Visual comparison of results obtained with various geometrical methods for NDF supervision on the benchmark dataset. (a) Ray Distance, (b) Ray distance along the normal pointing towards the closest surface, and (c) Proposed approach. approach, we utilized the sequence as given by Kuang et al. [11]. The particle filter transitions to pose tracking mode with 10,000 particles when the stan￾dard deviation of particles f… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Results on KITTI dataset. (a) (b) (c) [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: This figure illustrates front-view mappings derived from various geometric methods, with darker colours indicating improved normal estimation. (a) Ray Dis￾tance, (b) Ray distance along the normal pointing towards the closest surface, and (c) Proposed approach. obtained…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 38 canonical work pages

  1. [1]

    IEEE Robotics and Automation Letters7(2), 4606–4613 (2022) 3

    Adamkiewicz, M., Chen, T., Caccavale, A., Gardner, R., Culbertson, P., Bohg, J., Schwager, M.: Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters7(2), 4606–4613 (2022) 3

    work page 2022

  2. [2]

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

    Azinović, D., Martin-Brualla, R., Goldman, D.B., Nießner, M., Thies, J.: Neural rgb-d surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6290–6301 (2022) 3

    work page 2022

  3. [3]

    In: International Conference on Machine Learning

    Chaplot, D.S., Pathak, D., Malik, J.: Differentiable spatial planning using trans- formers. In: International Conference on Machine Learning. pp. 1484–1495. PMLR (2021) 3

    work page 2021

  4. [4]

    Advances in Neural Information Processing Systems33, 21638– 21652 (2020) 1, 2, 3, 4

    Chibane, J., Pons-Moll, G., et al.: Neural unsigned distance fields for implicit function learning. Advances in Neural Information Processing Systems33, 21638– 21652 (2020) 1, 2, 3, 4

    work page 2020

  5. [5]

    : A dataset and explorer for 3d signed distance functions

    Takikawa et al. : A dataset and explorer for 3d signed distance functions. Journal of Computer Graphics Techniques (2022) 7

    work page 2022

  6. [6]

    Advances in neural information processing systems 14 (2001) 12, 13

    Fox, D.: Kld-sampling: Adaptive particle filters. Advances in neural information processing systems 14 (2001) 12, 13

    work page 2001

  7. [7]

    In: 2012 IEEE conference on computer vision and pattern recognition

    Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354–3361. IEEE (2012) 10

    work page 2012

  8. [8]

    Computer Aided Geometric Design 22(7), 632–658 (2005) 7

    Goldman, R.: Curvature formulas for implicit curves and surfaces. Computer Aided Geometric Design 22(7), 632–658 (2005) 7

    work page 2005

Show all 39 references
  1. [9]

    Grisetti, G.: srrg-localizer2d (1.6.0)” (2018) 12, 13

    2018

  2. [10]

    arXiv preprint arXiv:2303.13615 (2023) 7

    Holmes, D.P.: Germain curvature: The case for naming the mean curvature of a surface after sophie germain. arXiv preprint arXiv:2303.13615 (2023) 7

    2023 arXiv

  3. [11]

    IEEE Robotics and Automation Letters 8(3), 1627–1634 (2023) 11, 12, 13

    Kuang, H., Chen, X., Guadagnino, T., Zimmerman, N., Behley, J., Stachniss, C.: Ir-mcl: Implicit representation-based online global localization. IEEE Robotics and Automation Letters 8(3), 1627–1634 (2023) 11, 12, 13

    2023

  4. [12]

    Journal of graphics tools8(2), 1–15 (2003) 10

    Lewiner, T., Lopes, H., Vieira, A.W., Tavares, G.: Efficient implementation of marching cubes’ cases with topological guarantees. Journal of graphics tools8(2), 1–15 (2003) 10

    2003

  5. [13]

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

    Lu, W., Zhou, Y., Wan, G., Hou, S., Song, S.: L3-net: Towards learning based lidar localization for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6389–6398 (2019) 10

    2019

  6. [14]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460–4470 (2019) 2

    2019

  7. [15]

    Commu- nications of the ACM65(1), 99–106 (2021) 2, 5 16 A

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Commu- nications of the ACM65(1), 99–106 (2021) 2, 5 16 A. Singh et al

    2021

  8. [16]

    In: 2011 10th IEEE international symposium on mixed and augmented reality

    Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: Kinectfusion: Real-time dense surface mapping and tracking. In: 2011 10th IEEE international symposium on mixed and augmented reality. pp. 127–136...

    2011

  9. [17]

    In: RSS 2016 workshop: geometry and beyond-representations, physics, and scene under- standing for robotics

    Oleynikova, H., Millane, A., Taylor, Z., Galceran, E., Nieto, J., Siegwart, R.: Signed distance fields: A natural representation for both mapping and planning. In: RSS 2016 workshop: geometry and beyond-representations, physics, and scene under- standing for robotics. Universi...

    2016

  10. [18]

    In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Oleynikova, H., Taylor, Z., Fehr, M., Siegwart, R., Nieto, J.: Voxblox: Incremental 3d euclidean signed distance fields for on-board mav planning. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1366–

    2017

  11. [19]

    system30, 39 1, 2, 3, 4

    Ortiz, J., Clegg, A., Dong, J., Sucar, E., Novotny, D., Zollhoefer, M., Mukadam, M.: isdf: Real-time neural signed distance fields for robot perception. system30, 39 1, 2, 3, 4

  12. [20]

    In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) 2

    Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learn- ing continuous signed distance functions for shape representation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) 2

    2019

  13. [21]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335–14345 (2021) 3

    2021

  14. [22]

    In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Shan, T., Englot, B.: Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4758–4765. IEEE (2018) 3

    2018

  15. [23]

    Advances in neural information processing systems 33, 7462–7473 (2020) 1, 3, 4, 10

    Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in neural information processing systems 33, 7462–7473 (2020) 1, 3, 4, 10

    2020

  16. [24]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: imap: Implicit mapping and position- ing in real-time. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6229–6238 (2021) 3

    2021

  17. [25]

    arXiv preprint arXiv:2310.09463 (2023) 3

    Vasilopoulos, V., Garg, S., Huh, J., Lee, B., Isler, V.: Hio-sdf: Hierarchical incre- mental online signed distance fields. arXiv preprint arXiv:2310.09463 (2023) 3

    2023 arXiv

  18. [26]

    Vizzo, I., Chen, X., Chebrolu, N., Behley, J., Stachniss, C.: Poisson surface recon- struction for lidar odometry and mapping. pp. 5624–5630. IEEE (2021) 3, 4

    2021

  19. [27]

    Sensors22(3), 1296 (2022) 5, 8

    Vizzo, I., Guadagnino, T., Behley, J., Stachniss, C.: Vdbfusion: Flexible and effi- cient tsdf integration of range sensor data. Sensors22(3), 1296 (2022) 5, 8

    2022

  20. [28]

    IEEE Robotics and Automation Letters (RA-L) 8(2), 1029–1036 (2023).https://doi.org/10.1109/LRA.2023.3236571 3

    Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: KISS-ICP: In Defense of Point-to-Point ICP – Simple, Accurate, and Robust Reg- istration If Done the Right Way. IEEE Robotics and Automation Letters (RA-L) 8(2), 1029–1036 (2023).https://doi.org/10...

    2023

  21. [29]

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

    Wang, H., Wang, J., Agapito, L.: Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13293–13302 (2023) 3

    2023

  22. [30]

    IEEE Robotics and Automation Letters (2023) 1, 2, 3, 4, 5, 8, 10, 11, 12, 13

    Wiesmann, L., Guadagnino, T., Vizzo, I., Zimmerman, N., Pan, Y., Kuang, H., Behley, J., Stachniss, C.: Locndf: Neural distance field mapping for robot localiza- tion. IEEE Robotics and Automation Letters (2023) 1, 2, 3, 4, 5, 8, 10, 11, 12, 13

    2023

  23. [31]

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

    Williams, F., Trager, M., Bruna, J., Zorin, D.: Neural splines: Fitting 3d surfaces with infinitely-wide neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9949–9958 (2021) 3 Curvature Constrained Neural Distance Fields 17

    2021

  24. [32]

    IEEE Transactions on Robotics38(4), 2053–2073 (2022) 3

    Xu, W., Cai, Y., He, D., Lin, J., Zhang, F.: Fast-lio2: Fast direct lidar-inertial odometry. IEEE Transactions on Robotics38(4), 2053–2073 (2022) 3

    2022

  25. [33]

    IEEE Robotics and Automation Letters 6(2), 3317–3324 (2021) 3

    Xu, W., Zhang, F.: Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter. IEEE Robotics and Automation Letters 6(2), 3317–3324 (2021) 3

    2021

  26. [34]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Yan, Z., Tian, Y., Shi, X., Guo, P., Wang, P., Zha, H.: Continual neural map- ping: Learning an implicit scene representation from sequential observations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15782–15792 (2021) 3

    2021

  27. [35]

    Advances in Neural Information Processing Systems34, 22483–22497 (2021) 2

    Yang, G., Belongie, S., Hariharan, B., Koltun, V.: Geometry processing with neural fields. Advances in Neural Information Processing Systems34, 22483–22497 (2021) 2

    2021

  28. [36]

    In: Robotics: Science and systems

    Zhang, J., Singh, S.: Loam: Lidar odometry and mapping in real-time. In: Robotics: Science and systems. vol. 2, pp. 1–9. Berkeley, CA (2014) 3

    2014

  29. [37]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zhang, J., Yao, Y., Quan, L.: Learning signed distance field for multi-view sur- face reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6525–6534 (2021) 2, 3

    2021

  30. [38]

    In: 2023 IEEE Inter- national Conference on Robotics and Automation (ICRA)

    Zhong, X., Pan, Y., Behley, J., Stachniss, C.: Shine-mapping: Large-scale 3d map- ping using sparse hierarchical implicit neural representations. In: 2023 IEEE Inter- national Conference on Robotics and Automation (ICRA). pp. 8371–8377. IEEE (2023) 1, 2, 3, 5

    2023

  31. [39]

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

    Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786– 12796 (2022) 3

    2022

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.