Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Radar-only neural implicit mapping reconstructs 3D surfaces and view-dependent intensities

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 · deepseek-v4-flash

2026-08-02 17:21 UTC pith:77XDY56R

load-bearing objection First radar-only neural implicit surface reconstruction with view-dependent intensity, but the headline accuracy claim is undermined by the paper's own truncation metric. the 4 major comments →

arxiv 2603.25623 v3 pith:77XDY56R submitted 2026-03-26 cs.RO

Neural Surface and Reflectance Modelling from 3D Radar Data

Judith Treffler , Vladim\'ir Kubelka , Henrik Andreasson , Martin Magnusson This is my paper
classification cs.RO
keywords neural implicit surfaceradar point cloudsigned distance fieldview-dependent intensityradar reflectivity3D mapping3QFPIsparse data reconstruction
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.

This paper tries to establish that a single neural implicit model can reconstruct accurate 3D surfaces from sparse, noisy 3D radar point clouds while also predicting the radar intensity returned from each surface point as a function of viewing direction. If correct, this would be the first method to jointly model geometry and view-dependent radar reflectivity from radar alone, without relying on cameras or lidar. The authors argue that radar's resilience to fog, smoke, and dust makes such a capability valuable for autonomous systems operating in low-visibility conditions. They show on two outdoor radar datasets that their method, 3QFPI, produces smoother and more locally planar surfaces than lidar-oriented and classical baselines, and that it can reconstruct view-dependent intensity patterns such as the falloff near a corner reflector.

Core claim

The central claim is that 3QFPI—a neural implicit scene representation built on the memory-efficient 3QFP tri-quadtree feature encoding, extended with a separate intensity network inspired by NeuS2—can jointly learn a signed distance field (SDF) for surface geometry and a view-dependent radar intensity field from radar point clouds alone. The method assigns SDF labels to sampled points near each radar detection and trains an intensity network conditioned on geometry, SDF value, optional normals, and viewing direction encoded with spherical harmonics. On the Radar Forest and SNAIL-Radar datasets, 3QFPI achieves the lowest accuracy error and highest accuracy ratio against lidar ground truth, a

What carries the argument

The central object is the 3QFPI architecture: a hybrid feature encoding that combines tri-quadtree vertex features (stored in hash tables) with Fourier feature positional encoding, feeding an SDF network that produces both an SDF value and optional geometry features and normals. A separate intensity network, conditioned on these geometric outputs and on a spherical-harmonics-encoded viewing direction, predicts radar intensity. The radar equation is used as a conceptual guide: transmitted power, gain, and effective area are treated as constant or implicitly learned, while the radar cross section is inferred from geometry and view direction. This joint SDF-plus-intensity network is trained wit

Load-bearing premise

The method assumes that every radar range detection is a true surface point, treating each detection as the location of a real surface for SDF supervision, so a substantial fraction of multi-path returns or clutter would bias the learned geometry.

What would settle it

Run 3QFPI on a radar dataset with known ground-truth geometry and a high proportion of multi-path returns (e.g., a corner reflector or reflective building facade). If the reconstructed surface consistently bulges or shifts toward the false detections, the central claim that radar points can be trusted as surface supervision is falsified.

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

If this is right

  • If the central claim holds, radar-only 3D mapping can yield dense, smooth surfaces suitable for navigation and scene understanding in conditions where cameras and lidar degrade.
  • Modeling view-dependent radar intensities could improve localisation and object recognition by providing material and reflection-angle information beyond pure geometry.
  • Neural implicit representations appear more robust to sparse radar input than classical explicit SDFs and meshing, which could reduce the number of radar scans needed for reliable mapping.
  • The joint network can implicitly account for sensor-specific radar equation constants, potentially easing calibration when manufacturer parameters are unknown.
  • The proposed evaluation additions (local planarity metrics, accuracy outlier ratio) offer a more faithful way to compare reconstructions from sparse radar data than conventional F-score or completion alone.

Where Pith is reading between the lines

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

  • A likely extension is to explicitly model multi-path reflections and wide-beam effects, which the paper identifies as future work; doing so could correct the surface bias that false radar returns introduce.
  • The intensity network's ability to predict view-dependent reflectivity suggests it could be used for radar-specific novel view synthesis or for cross-sensor fusion, but the paper does not demonstrate these uses.
  • The observed robustness to sparse input suggests a testable hypothesis: 3QFPI's performance degrades gracefully with even sparser data than tested, potentially enabling mapping from a handful of radar scans.
  • The local-planarity metric may transfer to other noisy 3D sensors (e.g., sonar or low-cost lidar) as a more meaningful quality measure than global F-score.

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

4 major / 5 minor

Summary. The paper proposes 3QFPI, a neural implicit surface and reflectance model for 3D radar point clouds. It extends the 3QFP hybrid feature encoding with a NeuS2-inspired intensity network, jointly optimizing an SDF for geometry and view-dependent radar intensity predictions. Experiments on the Radar Forest and SNAIL-Radar datasets compare 3QFPI against classical meshing methods (α-shapes, BPA, Poisson), an explicit TSDF method (VDBFusion), and a lidar-oriented neural implicit method (SHINE-Mapping). The paper reports that 3QFPI achieves lower accuracy error and higher accuracy ratio on both datasets, smoother locally planar surfaces according to a gamma-distribution fit, lower memory usage, and reasonable held-out intensity prediction. The authors claim this is the first radar-only method to reconstruct 3D surfaces while modelling view-dependent radar intensities.

Significance. If the claims are substantiated, this would be a useful contribution to radar-based mapping for low-visibility autonomy. The paper addresses a real gap: most neural implicit mapping methods target camera or lidar data, and radar-specific surface reconstruction with reflectance modelling is sparse. The held-out intensity evaluation is a genuine predictive test, and the memory comparison is informative. The authors also deserve credit for explicitly discussing the limitations of standard accuracy/completion metrics for sparse radar data and for proposing an auxiliary local-planarity metric. However, the central 'most accurate surface reconstruction' claim is weakened by the very metric caveat the authors themselves raise: the accuracy error truncates distances above 0.4 m, and the reported outlier ratios are high. The sparse-input robustness claim similarly rests on completion ratio, which the paper acknowledges does not measure fidelity. These issues are load-bearing and need to be addressed before the central claims can be accepted.

major comments (4)
  1. [§IV-A.2, §IV-B.1, Table I] The accuracy metric truncates distances above 0.4 m, yet 3QFPI has accuracy outlier ratios of 26.16% (Radar Forest) and 45.58% (SNAIL-Radar) in Table I. Thus the reported accuracy error is averaged only over the inlier fraction, and on SNAIL-Radar nearly half of the mesh samples are discarded before computing the headline 0.1676 m error. The paper itself warns in §IV-A.2 and §IV-B.1 that this truncation can make reconstructions with strong interpolation appear overly accurate and that high outlier ratios can coexist with high F-scores. Since the central claim is that 3QFPI is 'more accurate', the authors should report accuracy metrics without truncation (e.g., mean/median/RMSE over all samples) or at least at multiple thresholds. Note that the outlier ratio in SNAIL-Radar is not the highest in the table (VDBFusion is lower and BPA is similar), so the specific stress-test formulation over
  2. [§IV-B.2, Fig. 5, Abstract] The abstract claims that 'as input point clouds get sparser, neural implicit representations render more faithful surfaces', but the supporting evidence in Fig. 5 is only completion ratio. As the paper itself notes in §IV-B.2, high completion does not necessarily reflect reconstruction quality because watertight interpolation can inflate completion. Completion alone cannot distinguish a faithful surface from one that simply fills unobserved areas with smooth geometry. To support the robustness claim, the authors should report sparse-input accuracy errors and/or accuracy outlier ratios, and ideally qualitative comparisons, across sparsity levels. Without this, the sparse-robustness claim is not established.
  3. [§III-B.1, §V] The SDF supervision labels each radar range detection as a true surface point, with near-surface samples signed-distance to that point and free-space samples negative. Multipath reflections and clutter, acknowledged in the introduction and conclusion, are not on any physical surface. If a substantial fraction of radar detections are such returns, the learned SDF will be systematically biased. The authors do not filter or model multipath, and the evaluation does not quantify its prevalence. This is a correctness risk for the core geometry claim. A concrete test would be to train with a simple range-gating or statistical outlier filter, or to quantify the fraction of detections that have no nearby lidar support, and show that geometry metrics are stable under such processing.
  4. [Table I] Surface reconstruction metrics are reported as single numbers with no variance. For 3QFPI versus SHINE-Mapping on SNAIL-Radar, the accuracy errors differ by only about 2% (0.1676 vs 0.1713 m), which may be within run-to-run variation given random network initialization and sampling. The intensity experiments are repeated ten times with standard deviations, but the geometry experiments are not. At minimum, the neural methods should be run multiple times with mean and standard deviation reported, and the significance of the accuracy differences should be assessed. This is needed to support the claim that 3QFPI is more accurate than the baselines.
minor comments (5)
  1. [Title] The submitted title is 'Neural Surface and Reflectance Modelling from 3D Radar Data', while the main text header is 'Accurate Surface and Reflectance Modelling from 3D Radar Data with Neural Radiance Fields'. Please align the title between arXiv metadata and the manuscript.
  2. [§IV-A.2] The description of the gamma-distribution metric is clear, but the interpretation 'lower shape values correspond to distributions concentrated near zero' could be expanded; the shape parameter is not a direct planarity measure and the same shape can arise from different angle histograms. The paper acknowledges this, but a brief mathematical definition of the fitted distribution would help reproducibility.
  3. [§IV-C, Fig. 7] For the corner-reflector dataset, it would be useful to state explicitly how many held-out point clouds were used and how the view-dependent intensity ground truth was associated with query points. The description is currently qualitative.
  4. [§IV-B.1] The sentence '3QFPI preserves more radar details but also artefacts, which increase the outlier count' is important, but it raises a question: if artefacts are the cause of 3QFPI's high outlier ratio, the reader needs a way to distinguish artefacts from genuine detail. The gamma-planarity analysis helps, but a discussion of how artefacts are identified would strengthen the interpretation.
  5. [References] The radar dataset link is given as an anonymous URL; if this is an anonymized submission, this should be stated clearly. If not, the final version should provide a stable citation.

Circularity Check

0 steps flagged

No significant circularity: geometry and intensity are learned from radar data and evaluated against external lidar ground truth and held-out frames; the only self-citation (3QFP) is transparent reuse of published prior work, not a load-bearing circular premise.

full rationale

The paper is an empirical systems paper. The SDF is trained with labels computed as signed distances to input radar detections (§III-B.1); this is the intended reconstruction task, and the resulting mesh is evaluated against lidar ground truth (§IV-A.2), an external benchmark not used to define the training objective. Intensity reconstruction is tested on point clouds held out during training (§IV-C.1), so the reported MAE/MedAE is not a fitted parameter renamed as a prediction. The paper's own caveat about accuracy truncation (distances >0.4 m discarded) and outlier ratios (§IV-B.1) is an evaluation limitation, not a derivation that reduces to its inputs: the metric is defined, the caveat is stated, and the same caveat is applied to baselines. The one self-citation is to 3QFP [4], a peer-reviewed hybrid encoding used as the SDF backbone; it is referenced as prior architecture, not invoked as an unverified uniqueness/ansatz result, and the radar-specific novelty (joint geometry + view-dependent intensity) is evaluated independently. No step exhibits Eq. X = Eq. Y by construction or a fitted input presented as a prediction.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

The method rests on treating radar detections as surface hits, assigning constant intensity labels near detections, zero intensity in free space, and using lidar overlap as ground truth; plus an ad hoc planarity metric. These are reasonable engineering assumptions but are not physically guaranteed, especially for multipath radar.

free parameters (8)
  • leaf node resolution = 0.2 m
    Spatial grid resolution for 3QFPI and SHINE-Mapping; hand-set in §IV-A.4.
  • mesh extraction voxel size = 0.1 m
    Voxel size used for mesh extraction/rendering of VDBFusion, SHINE, and 3QFPI; hand-set in §IV-A.4.
  • near-surface and free-space sample counts = Ns = Nf = 6
    Increased from the 3QFP default to improve completion (§IV-A.4).
  • training iterations = 4000
    Hand-set; SDF network frozen after 1000 iterations (§IV-A.4).
  • learning rate = 1e-3
    Hand-set in §IV-A.4.
  • intensity normalization range = rounded dataset-wide min/max
    Used to scale intensities to [0,1]; affects the intensity loss scale (§IV-A.4).
  • geometry feature dimension = 15
    g in R^15, output of the SDF network used for intensity conditioning (§III-A.1).
  • baseline hyperparameters = alpha=0.2, BPA radius=0.1, Poisson depth=12, VDB voxel=0.2
    Classical baseline parameters hand-set for 'reasonable performance' (§IV-A.4); can bias comparisons.
axioms (5)
  • domain assumption Every radar detection is treated as a true surface point for SDF supervision.
    Training assigns signed-distance labels relative to each input point x (§III-B.1); multipath reflections and clutter are not modelled, so many radar returns may not lie on physical surfaces.
  • domain assumption Radar intensity is approximately constant near a detection, so nearby sampled points inherit the measured intensity.
    Explicit assumption in §III-B.1: 'assuming local intensities are approximately constant due to the low spatial resolution of radar.'
  • domain assumption Free-space samples have zero radar intensity.
    Loss sets free-space labels to 0 (§III-B.1); ignores sidelobes and multipath energy in free space.
  • domain assumption Lidar point clouds, restricted to a manually chosen overlap box, are valid ground truth for radar surface reconstruction.
    Evaluation uses lidar as denser/less noisy reference and manually restricts to overlapping high-density regions (§IV-A.2); this can disadvantage reconstructions that fill radar-only regions.
  • ad hoc to paper Lower mean/variance of the Gamma distribution fitted to adjacent-triangle angles indicates better reconstruction in the test scenes.
    Introduced as a supplementary smoothness metric (§IV-A.2); authors acknowledge interpretation depends on scene planarity, and both datasets are predominantly planar, favoring smooth methods.

pith-pipeline@v1.3.0-alltime-deepseek · 11279 in / 14973 out tokens · 145515 ms · 2026-08-02T17:21:00.689673+00:00 · methodology

0 comments
read the original abstract

Robust scene representation is essential for autonomous systems to safely operate in challenging low-visibility environments. In these conditions, radar has a clear advantage over cameras and lidars due to its resilience to environmental factors such as fog, smoke, or dust. However, radar data is inherently sparse and noisy, making reliable 3D surface reconstruction challenging. To address this, we propose a neural implicit approach for 3D mapping from radar point clouds that jointly models scene geometry and view-dependent radar intensities. Our method leverages a memory-efficient hybrid feature encoding to learn a continuous Signed Distance Field (SDF) for surface reconstruction, while also capturing radar-specific reflective properties. We show that our approach produces smoother, more accurate 3D surface reconstructions compared to existing lidar-based reconstruction methods applied to radar data and can reconstruct view-dependent radar intensities. We also show that, in general, as input point clouds get sparser, neural implicit representations render more faithful surfaces than traditional explicit SDFs and meshing techniques.

Figures

Figures reproduced from arXiv: 2603.25623 by Henrik Andreasson, Judith Treffler, Martin Magnusson, Vladim\'ir Kubelka.

Figure 1
Figure 1. Figure 1: Accurate surface reconstruction (right) produced by [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Network Architecture: Given a 3D point x, we concatenate its tri-quadtree feature and Fourier feature posi￾tional encoding and pass them to the SDF network. The SDF network predicts an SDF value and, optionally, a learned geometry feature and/or approximated SDF normals. These outputs, along with the spherical harmonics-encoded viewing direction and the Fourier feature encoding of x, are concatenated and f… view at source ↗
Figure 3
Figure 3. Figure 3: Surface reconstruction quality of different methods on the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Outlines of the histograms of angles between adjacent [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of completion ratios using every [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of memory usage for the Radar Forest dataset. 3QFPI has the lowest memory usage, needing only about 20 − 30% of the map size of SHINE￾Mapping; however, SHINE-Mapping captures more fine details. in fig. 5, Poisson consistently achieves the highest completion ratios across all sparsity levels, due to its watertight surfaces that interpolate unobserved areas. However, as this interpo￾lation can ove… view at source ↗
Figure 7
Figure 7. Figure 7: Reconstruction of view-dependent intensities. A corner reflector on a wall (a) is observed from multiple viewing [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

Review history (2 revisions) →

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

36 extracted references

  1. [1]

    NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis”, Commun. ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  2. [2]

    LONER: LiDAR Only Neural Representations for Real-Time SLAM

    S. Isaacson, P.-C. Kung, M. Ramanagopal, R. Vasude- van, and K. A. Skinner, “LONER: LiDAR Only Neural Representations for Real-Time SLAM”,IEEE Robot. Automat. Lett., vol. 8, no. 12, pp. 8042–8049, 2023

    2023

  3. [3]

    SHINE- Mapping: Large-Scale 3D Mapping Using Sparse Hi- erarchical Implicit Neural Representations

    X. Zhong, Y . Pan, J. Behley, and C. Stachniss, “SHINE- Mapping: Large-Scale 3D Mapping Using Sparse Hi- erarchical Implicit Neural Representations”, inProc. IEEE Int. Conf. Robot. Automat., 2023, pp. 8371–8377

    2023

  4. [4]

    3QFP: Efficient neural implicit surface reconstruction using Tri-Quadtrees and Fourier feature Positional encoding

    S. Sun, M. Mielle, A. J. Lilienthal, and M. Magnusson, “3QFP: Efficient neural implicit surface reconstruction using Tri-Quadtrees and Fourier feature Positional encoding”, inProc. IEEE Int. Conf. Robot. Automat., 2024, pp. 4036–4044

    2024

  5. [5]

    Radar Fields: Frequency-Space Neural Scene Representations for FMCW Radar

    D. Borts et al., “Radar Fields: Frequency-Space Neural Scene Representations for FMCW Radar”, inProc. ACM SIGGRAPH, 2024, pp. 1–10

    2024

  6. [6]

    DART: Implicit Doppler Tomography for Radar Novel View Synthesis

    T. Huang et al., “DART: Implicit Doppler Tomography for Radar Novel View Synthesis”, inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2024, pp. 24 118– 24 129

    2024

  7. [7]

    RF4D: Neural Radar Fields for Novel View Synthesis in Outdoor Dynamic Scenes

    J. Zhang, Z. Li, C. Wang, and B. Wen. “RF4D: Neural Radar Fields for Novel View Synthesis in Outdoor Dynamic Scenes”. arXiv:2505.20967

    arXiv

  8. [8]

    GeoRF: Geomet- ric Constrained RaDAR Fields

    W. Sheng, H. Liu, K. Fan, and P. Su, “GeoRF: Geomet- ric Constrained RaDAR Fields”,IEEE Access, vol. 13, pp. 78 391–78 402, 2025

    2025

  9. [9]

    NeuRadar: Neural Radi- ance Fields for Automotive Radar Point Clouds

    M. Rafidashti, J. Lan, M. Fatemi, J. Fu, L. Ham- marstrand, and L. Svensson, “NeuRadar: Neural Radi- ance Fields for Automotive Radar Point Clouds”, in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops, 2025, pp. 2479–2489

    2025

  10. [10]

    NeuS2: Fast Learning of Neural Implicit Surfaces for Multi-view Reconstruction

    Y . Wang, Q. Han, M. Habermann, K. Daniilidis, C. Theobalt, and L. Liu, “NeuS2: Fast Learning of Neural Implicit Surfaces for Multi-view Reconstruction”, in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2023, pp. 3272–3283

    2023

  11. [11]

    OctoMap: An efficient probabilistic 3D mapping framework based on octrees

    A. Hornung, K. M. Wurm, M. Bennewitz, C. Stachniss, and W. Burgard, “OctoMap: An efficient probabilistic 3D mapping framework based on octrees”,Auton. Robots, vol. 34, no. 3, pp. 189–206, 2013

    2013

  12. [12]

    Radar-Inertial State Estimation and Obstacle Detection for Micro-Aerial Vehicles in Dense Fog

    A. Kramer and C. Heckman, “Radar-Inertial State Estimation and Obstacle Detection for Micro-Aerial Vehicles in Dense Fog”, inExp. Robot.V ol. 19, 2021, pp. 3–16

    2021

  13. [13]

    Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environ- ments

    J. Behley and C. Stachniss, “Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environ- ments”, inProc. Robot. Sci. Syst., 2018

    2018

  14. [14]

    Real-time Scalable Dense Surfel Mapping

    K. Wang, F. Gao, and S. Shen, “Real-time Scalable Dense Surfel Mapping”, inProc. IEEE Int. Conf. Robot. Automat., 2019, pp. 6919–6925

    2019

  15. [15]

    On the shape of a set of points in the plane

    H. Edelsbrunner, D. Kirkpatrick, and R. Seidel, “On the shape of a set of points in the plane”,IEEE Trans. Inf. Theory, vol. 29, no. 4, pp. 551–559, 1983

    1983

  16. [16]

    The ball-pivoting algorithm for surface reconstruction

    F. Bernardini, J. Mittleman, H. Rushmeier, C. Silva, and G. Taubin, “The ball-pivoting algorithm for surface reconstruction”,IEEE Trans. Vis. Comput. Graph., vol. 5, no. 4, pp. 349–359, 1999

    1999

  17. [17]

    Poisson Surface Reconstruction

    M. Kazhdan, M. Bolitho, and H. Hoppe, “Poisson Surface Reconstruction”, inProc. Eurographics Symp. Geom. Process., 2006, pp. 61–70

    2006

  18. [18]

    Poisson Surface Reconstruction for LiDAR Odometry and Mapping

    I. Vizzo, X. Chen, N. Chebrolu, J. Behley, and C. Stachniss, “Poisson Surface Reconstruction for LiDAR Odometry and Mapping”, inProc. IEEE Int. Conf. Robot. Automat., 2021, pp. 5624–5630

    2021

  19. [19]

    Online 3D Reconstruction Based On Lidar Point Cloud

    Z. Han, H. Fang, Q. Yang, Y . Bai, and L. Chen, “Online 3D Reconstruction Based On Lidar Point Cloud”, in Proc. Chin. Control Conf., 2023, pp. 4505–4509

    2023

  20. [20]

    V oxblox: Incremental 3D Euclidean Signed Distance Fields for on-board MA V planning

    H. Oleynikova, Z. Taylor, M. Fehr, R. Siegwart, and J. Nieto, “V oxblox: Incremental 3D Euclidean Signed Distance Fields for on-board MA V planning”, inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2017, pp. 1366– 1373

    2017

  21. [21]

    VDBFusion: Flexible and Efficient TSDF Integration of Range Sensor Data

    I. Vizzo, T. Guadagnino, J. Behley, and C. Stachniss, “VDBFusion: Flexible and Efficient TSDF Integration of Range Sensor Data”,Sensors, vol. 22, no. 3, p. 1296, 2022

    2022

  22. [22]

    Occupancy Networks: Learning 3D Reconstruction in Function Space

    L. Mescheder, M. Oechsle, M. Niemeyer, S. Nowozin, and A. Geiger, “Occupancy Networks: Learning 3D Reconstruction in Function Space”, inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2019, pp. 4455– 4465

    2019

  23. [23]

    UNISURF: Uni- fying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction

    M. Oechsle, S. Peng, and A. Geiger, “UNISURF: Uni- fying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction”, inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2021, pp. 5569–5579

    2021

  24. [24]

    Efficient Implicit Neural Reconstruction Using LiDAR

    D. Yan, X. Lyu, J. Shi, and Y . Lin, “Efficient Implicit Neural Reconstruction Using LiDAR”, inProc. IEEE Int. Conf. Robot. Automat., 2023, pp. 8407–8414

    2023

  25. [25]

    V olume Rendering of Neural Implicit Surfaces

    L. Yariv, J. Gu, Y . Kasten, and Y . Lipman, “V olume Rendering of Neural Implicit Surfaces”, inAdv. Neural Inf. Process. Syst., 2021

    2021

  26. [26]

    NeuS: Learning Neural Implicit Surfaces by V olume Rendering for Multi-view Reconstruction

    P. Wang, L. Liu, Y . Liu, C. Theobalt, T. Komura, and W. Wang, “NeuS: Learning Neural Implicit Surfaces by V olume Rendering for Multi-view Reconstruction”, inAdv. Neural Inf. Process. Syst., 2021, pp. 27 171– 27 183

    2021

  27. [27]

    Ultra-NeRF: Neural Radiance Fields for Ultrasound Imaging

    M. Wysocki, M. F. Azampour, C. Eilers, B. Busam, M. Salehi, and N. Navab, “Ultra-NeRF: Neural Radiance Fields for Ultrasound Imaging”, inProc. Med. Imaging Deep Learn., 2024, pp. 382–401

    2024

  28. [28]

    Neural Im- plicit Surface Reconstruction using Imaging Sonar

    M. Qadri, M. Kaess, and I. Gkioulekas, “Neural Im- plicit Surface Reconstruction using Imaging Sonar”, in Proc. IEEE Int. Conf. Robot. Automat., 2023, pp. 1040– 1047

    2023

  29. [29]

    Bathy- metric Surveying With Imaging Sonar Using Neural V olume Rendering

    Y . Xie, G. Troni, N. Bore, and J. Folkesson, “Bathy- metric Surveying With Imaging Sonar Using Neural V olume Rendering”,IEEE Robot. Automat. Lett., vol. 9, no. 9, pp. 8146–8153, 2024

    2024

  30. [30]

    SAR-NeRF: Neural Radiance Fields for Synthetic Aperture Radar Multi-View Representation

    Z. Lei, F. Xu, J. Wei, F. Cai, F. Wang, and Y . -Q. Jin, “SAR-NeRF: Neural Radiance Fields for Synthetic Aperture Radar Multi-View Representation”,IEEE Trans. Geosci. Remote Sens., vol. 62, pp. 1–15, 2024

    2024

  31. [31]

    Neural Implicit Repre- sentations for 3D Synthetic Aperture Radar Imaging

    N. Sugavanam and E. Ertin, “Neural Implicit Repre- sentations for 3D Synthetic Aperture Radar Imaging”, IEEE Trans. Geosci. Remote Sens., vol. 62, pp. 1–15, 2024

    2024

  32. [32]

    NeRF in the Dark: High Dynamic Range View Synthesis from Noisy Raw Images

    B. Mildenhall, P. Hedman, R. Martin-Brualla, P. P. Srinivasan, and J. T. Barron, “NeRF in the Dark: High Dynamic Range View Synthesis from Noisy Raw Images”, inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2022, pp. 16 169–16 178

    2022

  33. [33]

    DehazeNeRF: Multi-image Haze Removal and 3D Shape Reconstruction using Neural Radiance Fields

    W.-T. Chen, W. Yifan, S. -Y . Kuo, and G. Wetzstein, “DehazeNeRF: Multi-image Haze Removal and 3D Shape Reconstruction using Neural Radiance Fields”, inProc. Int. Conf. 3D Vis., 2024, pp. 247–256

    2024

  34. [34]

    Marching cubes: A high resolution 3D surface construction algorithm

    W. E. Lorensen and H. E. Cline, “Marching cubes: A high resolution 3D surface construction algorithm”, SIGGRAPH Comput. Graph., vol. 21, no. 4, pp. 163– 169, 1987

    1987

  35. [35]

    SNAIL radar: A large-scale diverse benchmark for evaluating 4D-radar-based SLAM

    J. Huai, B. Wang, Y . Zhuang, Y . Chen, Q. Li, and Y . Han, “SNAIL radar: A large-scale diverse benchmark for evaluating 4D-radar-based SLAM”,Int. J. Robot. Res., vol. 44, no. 12, pp. 1941–1958, 2025

    1941

  36. [36]

    Open3D: A Modern Library for 3D Data Processing

    Q.-Y . Zhou, J. Park, and V . Koltun. “Open3D: A Modern Library for 3D Data Processing”. arXiv: 1801. 09847