Pith. sign in

REVIEW 3 major objections 4 minor 38 references

Dense volume sampling along LiDAR rays is what makes neural radiance fields able to jointly refine sensor poses and the map.

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

2026-07-11 21:12 UTC pith:RDTZRJ33

load-bearing objection Solid LiDAR-specific NeRF-BA with a real sampling insight; cube normalization is a known scale limit, not a hidden contradiction. the 3 major comments →

arxiv 2607.04169 v1 pith:RDTZRJ33 submitted 2026-07-05 cs.RO

Neural LiDAR Bundle Adjustment

Chin Yung Anson Hon , Kaicheng Zhang , Sen Wang This is my paper
classification cs.RO
keywords LiDAR NeRFbundle adjustmentvolume samplingpoint cloud registrationneural radiance fieldspose optimization3D mapping
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.

LiDAR point clouds give accurate ranges but almost no texture, so ordinary NeRF bundle adjustment fails: pose gradients vanish unless samples sit densely around the surface that actually terminates each ray. The authors show, both mathematically and with ablation, that sampling density is the decisive difference from RGB NeRFs. They therefore redesign hierarchical sampling (fine samples placed before coarse ones), clamp every ray to the unit cube, replace the usual Fourier gradient with a non-amplifying surrogate, and use a near-Dirac termination loss. The resulting NeLD-BA algorithm starts from noisy odometry, jointly optimizes the neural map and all poses, and produces cleaner multi-view registrations and rendered maps than classical LiDAR bundle adjustment or other neural mappers on two outdoor datasets. A reader who builds maps or odometry cares because the method turns raw range measurements into a differentiable, correspondence-free pose-and-map optimizer that works with ordinary LiDAR scans.

Core claim

The density of volume samples along each LiDAR ray controls whether the gradient of estimated range with respect to sensor pose remains non-zero. When fine samples are deliberately placed just before the density spike, rays are bounded inside the unit cube, Fourier positional encodings use a surrogate gradient that removes the 2^k pi amplification, and a pseudo-Dirac termination loss is applied, a single neural field can be trained to refine an entire sequence of noisy LiDAR poses and reconstruct the scene at once, outperforming both classical and learning-based multi-view registration methods.

What carries the argument

Modified hierarchical volume sampling: fine samples are allocated immediately before the coarse samples that straddle the surface density spike, and each ray’s interval is clamped to the intersections with the normalized cube [-1,1]^3 (with fallback bounds). This keeps the partial derivative of the ray-termination weight with respect to sample position from vanishing, which is the term that multiplies the pose Jacobian in the range gradient.

Load-bearing premise

Clamping every outdoor LiDAR ray inside one shared unit cube, plus putting the fine samples just before the surface, is enough to keep pose gradients alive for the whole sequence.

What would settle it

Run the identical pipeline on a long outdoor trajectory whose points no longer fit comfortably inside a single unit cube after normalization; if the final absolute trajectory error and map Chamfer distance become worse than the raw input odometry, the sampling-density claim fails.

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

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

3 major / 4 minor

Summary. The paper argues that LiDAR NeRF bundle adjustment differs fundamentally from RGB NeRF-BA because accurate range measurements require non-vanishing pose gradients through the ray-termination distribution ĥ; this in turn depends on sufficiently dense volume sampling around the density spike (Eq. 5, Fig. 2). It therefore introduces NeLD-BA, which (i) places fine samples before rather than around coarse samples, (ii) clamps each ray’s sampling interval to the intersections with the normalized cube C=[-1,1]³ (Eq. 6), (iii) replaces the Fourier positional-encoding gradient by a surrogate that removes the 2^k π amplification (Eq. 8), and (iv) supervises the termination distribution with a pseudo-Dirac target derived from the measured range (Eq. 11). Joint optimization of the NeRF MLP and SE(3) poses is shown to improve multi-view registration and both raw and rendered maps relative to HBA, BALM, SHINE-Mapping, PIN-SLAM and 4dNDF on the Newer College and FusionPortable sequences.

Significance. If the sampling-density analysis and the accompanying design choices hold, the work supplies a principled, correspondence-free alternative to classical LiDAR BA that can refine noisy odometry poses while simultaneously producing a high-fidelity neural map. The gradient derivation, the concrete sampling modifications, the ablation table that isolates each component, and the promise of open-source code constitute concrete engineering contributions that are immediately usable by the robotics community. The limitation to compact scenes is acknowledged, so the result is best viewed as a solid advance for the class of environments already covered by the two evaluation datasets rather than a universal large-scale solution.

major comments (3)
  1. [III-B, Eq. (6)] Sec. III-B / Eq. (6): The non-vanishing-gradient claim of Eq. (5) rests on the assertion that clamping every ray to the cube C=[-1,1]³ keeps sampling density high enough around the σ spike. All reported sequences are compact enough that most ray origins lie near the cube; the paper itself notes (Sec. IV-E) that large-scale scenes under-utilize the model precisely because of this normalization. Without an experiment or analysis that stresses far-origin rays, the theoretical guarantee and the SOTA registration claim remain untested outside the compact regime.
  2. [Table I, Sec. IV-B] Table I and Sec. IV-B: Raw maps generated from the provided ground-truth poses frequently yield higher L1-Chamfer distances than the optimized poses (e.g., cloister, math_hard, quad_hard). The authors attribute this to imperfect survey poses, yet the same ground-truth is used both for map evaluation and (implicitly) for trajectory alignment on FusionPortable. This circularity weakens the absolute claims of accuracy and should be quantified more carefully—e.g., by reporting residual alignment error after SE(3) registration of the GT map itself.
  3. [III-D, Eq. (11)] Sec. III-D / Eq. (11): The pseudo-Dirac termination target is presented as a natural consequence of having exact range measurements, yet it is an ad-hoc modeling choice whose sensitivity to the precise location of the first sample after the measured range is never examined. Because the termination loss is shown by ablation (Table IV) to be load-bearing, a short sensitivity study or comparison against a narrow Gaussian (as in DS-NeRF) would strengthen the claim.
minor comments (4)
  1. [throughout] Numerous typographical errors remain: “differnece” (Fig. 2 caption), “partiularly”, “Benifiting”, “cmoparable”, “substentially”, “V olume” (with spurious space), “F ormulation”, etc. A careful proof-reading pass is needed.
  2. [Fig. 3] Fig. 3 caption and surrounding text refer to an “actual LiDAR ray” but never state which sequence or which ray index; adding that information would make the figure reproducible.
  3. [III-F] The hyper-parameter schedule (sample counts growing from 32/16 to 64/32, frequency unfreezing at k*=6, etc.) is described only in prose; a short table would improve clarity and reproducibility.
  4. [II-B, IV-A] GeoNLF is cited as related work but omitted from experiments solely for lack of open-source support; a brief qualitative discussion of architectural differences would still be useful.

Circularity Check

0 steps flagged

No circularity: gradient analysis and sampling design are independent of the external Chamfer/ATE benchmarks used for validation.

full rationale

The paper's central derivation (Eq. 5 and the accompanying argument that sampling density keeps ∂ĥ/∂xi non-vanishing) is a first-principles differentiation of the alpha-compositing range estimator; it does not define any quantity in terms of the later experimental metrics, nor does it fit a free parameter that is then re-labeled as a prediction. The two concrete sampling modifications (fine samples placed before coarse samples; ray intervals clamped to the normalized cube via Eq. 6) are design choices whose efficacy is tested, not assumed, against independent baselines (HBA, BALM, SHINE-Mapping, PIN-SLAM, 4dNDF, GeoTrans, SGHR) and external ground-truth maps/trajectories from Newer College and FusionPortable. Loss terms (L1 range + pseudo-Dirac KL) simply use the measured LiDAR ranges that any BA method is entitled to use; they do not force the reported ranking by construction. No self-citation supplies a uniqueness theorem or load-bearing premise, and no known empirical pattern is merely renamed. The acknowledged limitation on large-scale scenes is an empirical scope restriction, not a circular reduction. Consequently the derivation chain is self-contained and the score is zero.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 3 invented entities

The central claim rests on the standard NeRF volume-rendering model plus several hand-chosen sampling and loss design decisions. Free parameters are the usual NeRF training knobs plus the new λ weights and activation schedule; axioms are the classical alpha-compositing range estimator and SE(3) pose parameterization; invented entities are the specific sampling schedule, surrogate PE gradient, and Dirac-style termination distribution.

free parameters (5)
  • λ_coarse, λ_d, λ_h (loss weights) = λ_coarse=0.1, λ_d=1, λ_h=1
    Set by grid search to 0.1 / 1 / 1 and frozen for all sequences; directly scale the objective that produces the reported maps.
  • positional encoding frequencies L and unfreeze threshold k* = L_pos=15, L_dir=4, k*=6
    L=15 for position, L=4 for direction; poses unfrozen only after k*=6 is activated (BARF schedule).
  • coarse/fine sample counts and growth schedule = train 64/32, render 128/128
    Increased from 32/16 to 64/32 over 120k iterations; final map rendering uses 128/128.
  • voxel size and sensing range filter = 0.05 m, 0.5–80 m
    Scans voxelized at 0.05 m; points kept only in 0.5–80 m.
  • Adam learning-rate schedule = 5e-5 to 2e-5
    Exponential decay 5e-5 → 2e-5 over 160k iterations.
axioms (4)
  • domain assumption Alpha-compositing range estimator (Eqs. 2–3) correctly recovers expected hit distance from density samples.
    Taken from Mildenhall et al. NeRF; used as the differentiable measurement model for all pose gradients.
  • standard math Sensor poses live in SE(3) and are optimized in the Lie algebra se(3).
    Standard rigid-body parameterization; appears in problem formulation (Sec. III-A).
  • domain assumption Range (not photometric) loss is the primary signal for LiDAR pose optimization.
    Stated as the key difference from RGB NeRF-BA (Sec. III-B); underpins the entire sampling redesign.
  • ad hoc to paper A pseudo-Dirac termination target (Eq. 11) is a valid supervision signal for the ray termination distribution.
    Replaces the Gaussian used in DS-NeRF; introduced without external theoretical guarantee.
invented entities (3)
  • Fine-before-coarse hierarchical sampling for LiDAR rays no independent evidence
    purpose: Ensure dense samples sit on the rising edge of the density spike so ∂ĥ/∂xi does not vanish.
    Core algorithmic novelty of Sec. III-B; no independent evidence outside the paper’s ablations.
  • Surrogate gradient for Fourier positional encoding (Eq. 8) no independent evidence
    purpose: Remove 2^k π amplification that destabilizes pose updates at high L.
    Inspired by spiking-network surrogates but newly applied to NeRF PE; validated only by ablation.
  • NeLD-BA joint pose-and-map optimizer no independent evidence
    purpose: End-to-end LiDAR bundle adjustment via the modified NeRF.
    The system itself; performance is the paper’s empirical claim.

pith-pipeline@v1.1.0-grok45 · 16250 in / 3309 out tokens · 34133 ms · 2026-07-11T21:12:19.108938+00:00 · methodology

0 comments
read the original abstract

Recent research has achieved remarkable novel view rendering and scene reconstruction results with Neural Radiance Field (NeRF), including extensions to the LiDAR modality. Few studies have, however, explored the key design differences between RGB NeRFs and LiDAR NeRFs, particularly considering their underlying working principles. In this work, we provide both theoretical and empirical evidence suggesting that the density of volume sampling plays a significant role in LiDAR NeRF. Based on this finding, we propose a novel Neural LiDAR Bundle Adjustment (NeLD-BA) algorithm, which is tailored using efficient volume sampling of LiDAR rays for joint optimization of LiDAR map and poses. Extensive experiments are performed using the Newer College and FusionPortable datasets to demonstrate the proposed NeLD-BA's state-of-the-art performance in multi-view point cloud registration and 3D mapping. We will open-source our code for the community.

Figures

Figures reproduced from arXiv: 2607.04169 by Chin Yung Anson Hon, Kaicheng Zhang, Sen Wang.

Figure 1
Figure 1. Figure 1: 3D mapping results with HBA [5] (left) and the pro [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of RGB NeRF-BA and LiDAR NeRF-BA in Volume Sampling along Depth/Range. (a) In RGB NeRF-BA, depth is only implicitly constrained via colors and multi-view geometry (orange rays), thus volume samples are loosely constrained. In contrast, LiDAR provides range measurements which require explicit volume sampling to constrain. (b) Sampling around sharp transition of σ(z) (highlighted pink region) is c… view at source ↗
Figure 3
Figure 3. Figure 3: Ray Termination Distributions of a LiDAR Ray in canteen_day sequence Our sampling method better approximates the baseline shifted distribution (ϵ = 0.01) constructed with 8192 uniform samples, which is critical for providing an accurate gradients in LiDAR NeRF-BA. positional encoding function γ(·) for feature x is [29]: γ(x) = [x, γ0(x), γ1(x), . . . , γL−1(x)] ∈ R 1+2L (7) where γk(x) = cos 2 kπx ,sin 2 … view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative Results of Raw Point Cloud Maps for Newer College (NC) and FusionPortable (FP) Sequences. TABLE I: Evaluation of Raw Point Cloud Maps Using L1-Chamfer Distance [cm] Newer College FusionPortable Method cloister math_easy math_hard quad_easy quad_mid quad_hard canteen_ day canteen_ night garden_ day garden_ night FAST-LIO2 [32] (initial) 41.45 34.74 37.70 27.85 31.71 37.74 35.76 35.03 34.89 30.84… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Results for Rendered Maps of Newer College (NC) and FusionPortable (FP) Sequences [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Error Histogram on Raw Map. From the error distribution of garden_day raw map points, it can be seen that our method result in lesser outlier points against BALM. clouds effectively, while BALM failed to handle outlier frames, which is also evident in [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative Results for Trajectory for FusionPortable Sequences. Circle markers represent the start of trajectories, cross markers represent the end of trajectories. TABLE III: ATE Mean Values for Trajectory Evaluation for FusionPortable Dataset [meter] Method canteen_ day canteen_ night garden_ day garden_ night FAST-LIO2 [32] (initial) 0.070 0.078 0.073 0.080 SGHR [10] 4.025 3.781 3.663 — GeoTrans [11] 0… view at source ↗
Figure 8
Figure 8. Figure 8: Ablation Study Qualitative Results. NeRF’s [1] sampling method result in an outlier frame (left) while our method successfully registered the outlier frame (right) [PITH_FULL_IMAGE:figures/full_fig_p007_8.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

38 extracted references · 4 linked inside Pith

  1. [1]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  2. [2]

    Barf: Bundle- adjusting neural radiance fields,

    C.-H. Lin, W.-C. Ma, A. Torralba, and S. Lucey, “Barf: Bundle- adjusting neural radiance fields,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 5741–5751

    2021

  3. [3]

    Nerf–: Neural radiance fields without known camera parameters,

    Z. Wang, S. Wu, W. Xie, M. Chen, and V . A. Prisacariu, “Nerf–: Neural radiance fields without known camera parameters,” 2021

    2021

  4. [4]

    Local-to-global registration for bundle-adjusting neural radiance fields,

    Y . Chenet al., “Local-to-global registration for bundle-adjusting neural radiance fields,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 8264–8273

    2023

  5. [5]

    Large-scale lidar consistent mapping using hierarchical lidar bundle adjustment,

    X. Liu, Z. Liu, F. Kong, and F. Zhang, “Large-scale lidar consistent mapping using hierarchical lidar bundle adjustment,”IEEE Robotics and Automation Letters, vol. 8, no. 3, pp. 1523–1530, 2023

    2023

  6. [6]

    The newer college dataset: Handheld lidar, inertial and vision with ground truth,

    M. Ramezani, Y . Wang, M. Camurri, D. Wisth, M. Mattamala, and M. Fallon, “The newer college dataset: Handheld lidar, inertial and vision with ground truth,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 4353–4360

    2020

  7. [7]

    Loam: Lidar odometry and mapping in real-time

    J. Zhang, S. Singhet al., “Loam: Lidar odometry and mapping in real-time.” inRobotics: Science and Systems, vol. 2, no. 9. Berkeley, CA, 2014, pp. 1–9

    2014

  8. [8]

    Method for registration of 3-d shapes,

    P. J. Besl and N. D. McKay, “Method for registration of 3-d shapes,” inSensor fusion IV: control paradigms and data structures, vol. 1611. Spie, 1992, pp. 586–606

    1992

  9. [9]

    Object modelling by registration of multiple range images,

    Y . Chen and G. Medioni, “Object modelling by registration of multiple range images,”Image and Vision Computing, vol. 10, no. 3, pp. 145– 155, 1992

    1992

  10. [10]

    Robust multiview point cloud registration with reliable pose graph initialization and history reweighting,

    H. Wanget al., “Robust multiview point cloud registration with reliable pose graph initialization and history reweighting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9506–9515

    2023

  11. [11]

    Geotransformer: Fast and robust point cloud registration with geometric transformer,

    Z. Qinet al., “Geotransformer: Fast and robust point cloud registration with geometric transformer,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 8, pp. 9806–9821, 2023

    2023

  12. [12]

    Hregnet: A hierarchical network for large-scale outdoor lidar point cloud registration,

    F. Luet al., “Hregnet: A hierarchical network for large-scale outdoor lidar point cloud registration,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 16 014–16 023

    2021

  13. [13]

    Highly accurate 3D surface models by sparse surface adjustment,

    M. Ruhnke, R. Kümmerle, G. Grisetti, and W. Burgard, “Highly accurate 3D surface models by sparse surface adjustment,” inIEEE International Conference on Robotics and Automation. IEEE, 2012, pp. 751–757

    2012

  14. [14]

    Balm: Bundle adjustment for lidar mapping,

    Z. Liu and F. Zhang, “Balm: Bundle adjustment for lidar mapping,” IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 3184–3191, 2021

    2021

  15. [15]

    Lidar4d: Dynamic neural fields for novel space-time view lidar synthesis,

    Z. Zheng, F. Lu, W. Xue, G. Chen, and C. Jiang, “Lidar4d: Dynamic neural fields for novel space-time view lidar synthesis,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 5145–5154

    2024

  16. [16]

    Gs-lidar: Generating realistic lidar point clouds with panoramic gaussian splatting,

    J. Jiang, C. Gu, Y . Chen, and L. Zhang, “Gs-lidar: Generating realistic lidar point clouds with panoramic gaussian splatting,”arXiv preprint arXiv:2501.13971, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Nerf-loam: Neural implicit representation for large- scale incremental lidar odometry and mapping,

    J. Denget al., “Nerf-loam: Neural implicit representation for large- scale incremental lidar odometry and mapping,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 8218–8227

    2023

  18. [18]

    Loner: Lidar only neural representations for real-time slam,

    S. Isaacson, P.-C. Kung, M. Ramanagopal, R. Vasudevan, and K. A. Skinner, “Loner: Lidar only neural representations for real-time slam,” IEEE Robotics and Automation Letters, vol. 8, no. 12, pp. 8042–8049, 2023

    2023

  19. [19]

    Geonlf: Geometry guided pose-free neural lidar fields,

    W. Xue, Z. Zheng, F. Lu, H. Wei, G. Chenet al., “Geonlf: Geometry guided pose-free neural lidar fields,”Advances in Neural Information Processing Systems, vol. 37, pp. 73 672–73 692, 2024

    2024

  20. [20]

    Shine-mapping: Large-scale 3d mapping using sparse hierarchical implicit neural representations,

    X. Zhong, Y . Pan, J. Behley, and C. Stachniss, “Shine-mapping: Large-scale 3d mapping using sparse hierarchical implicit neural representations,”arXiv preprint arXiv:2210.02299, 2022

    Pith/arXiv arXiv 2022

  21. [21]

    Pin-slam: Lidar slam using a point-based implicit neural representation for achieving global map consistency,

    Y . Pan, X. Zhong, L. Wiesmann, T. Posewsky, J. Behley, and C. Stachniss, “Pin-slam: Lidar slam using a point-based implicit neural representation for achieving global map consistency,”IEEE Transactions on Robotics, vol. 40, pp. 4045–4064, 2024

    2024

  22. [22]

    Neural geometric level of detail: Real-time rendering with implicit 3d shapes,

    T. Takikawaet al., “Neural geometric level of detail: Real-time rendering with implicit 3d shapes,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 11 358–11 367

    2021

  23. [23]

    Nope-nerf: Optimising neural radiance field with no pose prior,

    W. Bian, Z. Wang, K. Li, J.-W. Bian, and V . A. Prisacariu, “Nope-nerf: Optimising neural radiance field with no pose prior,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 4160–4169

    2023

  24. [24]

    Bad-nerf: Bundle adjusted deblur neural radiance fields,

    P. Wang, L. Zhao, R. Ma, and P. Liu, “Bad-nerf: Bundle adjusted deblur neural radiance fields,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 4170–4179

    2023

  25. [25]

    Mip-nerf: A multiscale representation for anti- aliasing neural radiance fields,

    J. T. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin-Brualla, and P. P. Srinivasan, “Mip-nerf: A multiscale representation for anti- aliasing neural radiance fields,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 5855–5864

    2021

  26. [26]

    Zip-nerf: Anti-aliased grid-based neural radiance fields,

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman, “Zip-nerf: Anti-aliased grid-based neural radiance fields,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19 697–19 705

    2023

  27. [27]

    Nerf-casting: Improved view-dependent appearance with consistent reflections,

    D. Verbinet al., “Nerf-casting: Improved view-dependent appearance with consistent reflections,” inSIGGRAPH Asia 2024 Conference Papers, 2024, pp. 1–10

    2024

  28. [28]

    An efficient and robust ray-box intersection algorithm,

    A. Williams, S. Barrus, R. K. Morley, and P. Shirley, “An efficient and robust ray-box intersection algorithm,” inACM SIGGRAPH 2005 Courses, 2005, pp. 9–es

    2005

  29. [29]

    Fourier features let networks learn high frequency functions in low dimensional domains,

    M. Tanciket al., “Fourier features let networks learn high frequency functions in low dimensional domains,”Advances in Neural Information Processing Systems, vol. 33, pp. 7537–7547, 2020

    2020

  30. [30]

    Nerf in the wild: Neural radiance fields for unconstrained photo collections,

    R. Martin-Brualla, N. Radwan, M. S. Sajjadi, J. T. Barron, A. Doso- vitskiy, and D. Duckworth, “Nerf in the wild: Neural radiance fields for unconstrained photo collections,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 7210–7219

    2021

  31. [31]

    The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks,

    F. Zenke and T. P. V ogels, “The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks,”Neural Computation, vol. 33, no. 4, pp. 899–925, 2021

    2021

  32. [32]

    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 Transactions on Robotics, vol. 38, no. 4, pp. 2053–2073, 2022

    2053

  33. [33]

    Depth-supervised nerf: Fewer views and faster training for free,

    K. Deng, A. Liu, J.-Y . Zhu, and D. Ramanan, “Depth-supervised nerf: Fewer views and faster training for free,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 12 882–12 891

    2022

  34. [34]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszkeet al., “Pytorch: An imperative style, high-performance deep learning library,” 2019. [Online]. Available: https://arxiv.org/abs/ 1912.01703

    Pith/arXiv arXiv 2019

  35. [35]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    Pith/arXiv arXiv 2014

  36. [36]

    3d lidar mapping in dynamic environments using a 4d implicit neural representation,

    X. Zhong, Y . Pan, C. Stachniss, and J. Behley, “3d lidar mapping in dynamic environments using a 4d implicit neural representation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 417–15 427

    2024

  37. [37]

    Fusionportablev2: A unified multi-sensor dataset for generalized slam across diverse platforms and scalable environments,

    H. Weiet al., “Fusionportablev2: A unified multi-sensor dataset for generalized slam across diverse platforms and scalable environments,” The International Journal of Robotics Research, vol. 44, no. 7, pp. 1093–1116, 2025

    2025

  38. [38]

    evo: Python package for the evaluation of odometry and slam

    M. Grupp, “evo: Python package for the evaluation of odometry and slam.” https://github.com/MichaelGrupp/evo, 2017

    2017