Pith. sign in

REVIEW 4 major objections 5 minor 62 references

Sem-RaDiff: Diffusion-Based 3D Radar Semantic Perception in Cluttered Agricultural Environments

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Sem-RaDiff claims that a coarse-to-fine diffusion pipeline can generate LiDAR-like semantic 3D point clouds from mmWave radar cubes, recovering thin structures such as poles and wires even when optical sensors are contaminated.

desk verdict Competent, novel radar semantic perception pipeline for agriculture; the "complete reconstruction" claim outruns the evidence, but the core idea deserves referee time. read the letter →

arxiv 2509.02283 v2 pith:EPQE5A2R submitted 2025-09-02 cs.RO

Ruibin Zhang , Fei Gao This is my paper
classification cs.RO
keywords radarsemanticperceptiondiffusionmodelsmmWave3Dpointcloudgenerationagriculturalroboticssensorcontaminationsparsenetworksframeaccumulation
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

Sem-RaDiff is an attempt to make mmWave radar a complete replacement for optical sensing in agricultural robots, whose cameras and LiDAR are often blinded by mud, dust, or crop debris. The paper claims that by accumulating five radar frames, filtering sidelobe artifacts with a learned coarse-mask stage, and then refining those masks with a diffusion model, the system can produce LiDAR-like 3D point clouds with semantic labels directly from radar data. On a self-built agricultural dataset it reports better geometric and semantic quality than re-implemented baselines, with the clearest gains on thin structures—poles and wires—that other methods miss. It also reports large efficiency gains: 51.3% fewer GFLOPs and 27.5% less memory than the best baseline, plus a one-step distilled version that removes iterative diffusion sampling. If these results hold, radar alone would give robots dense, labeled 3D perception exactly when optical sensors are unusable.

What carries the argument

The argument is carried by three coupled mechanisms. Parallel frame accumulation aligns five radar spherical cubes to the current frame, voxelizes them, and sums echo power non-coherently on the GPU, raising the signal-to-noise ratio of weak targets while keeping the data in a sparse 3D representation. Stage I is a discriminative sparse 3D network that learns to filter sidelobe artifacts by regressing Gaussian-blurred structural masks and classifying greyscale-dilated semantic masks; the blurring and dilation are what turn thin, low-echo structures like wires into dense, learnable targets. Stage II is a conditional diffusion model that takes those coarse masks as conditioning and generates f

What would settle it

Measure the accumulated radar echo at the known range, azimuth, and elevation bins of a single isolated wire or pole before feeding data to the network; if the echo is below the noise floor yet the pipeline still outputs a dense wire or pole, the prediction is hallucinated. A controlled field test with the wire removed from the scene would distinguish genuine sensing from learned scene priors.

Watch

Extended reading notes

Core claim

The central claim is that the modality gap between bulky, low-SNR radar cubes and fine-grained LiDAR point clouds is bridgeable by a coarse-to-fine generative pipeline rather than by direct discriminative mapping or single-stage diffusion. Stage I treats radar preprocessing as discriminative filtering: using Gaussian-blurred and greyscale-dilated LiDAR masks as supervision, it predicts coarse structural and semantic masks that suppress sidelobes and amplify weak targets such as wires. Stage II then treats the remaining task as conditional generation: a continuous-time diffusion-based denoiser, later distilled into a consistency model, reconstructs the LiDAR-like semantic point cloud from noi

Load-bearing premise

The load-bearing premise is that after five frames of non-coherent accumulation and top-qth intensity filtering, the radar cube contains real echo energy from thin low-radar-cross-section objects like wires and poles, so the Stage I coarse masks are grounded in radar evidence rather than filled in from scene priors.

Editorial extensions

If this is right

  • A soiled camera or LiDAR no longer forces mission abort: radar-only inference supplies dense semantic 3D structure, including safety-critical thin obstacles.
  • The one-step consistency version achieves competitive semantic IoU with a single network forward pass, so generative quality is not tied to expensive iterative sampling.
  • The reported efficiency gains—51.3% lower GFLOPs and 27.5% lower memory than the best re-implemented baseline—bring the pipeline closer to real-time onboard execution.
  • Because pole and wire IoU are roughly twice the best baseline at the 0.5 m threshold, the method specifically targets the thin-structure class that makes agricultural fields dangerous.

Reading between the lines

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

  • The paper's own qualitative analysis admits that slanting wires are missed and sometimes misclassified as trees; before claiming general thin-structure perception, one would need to test wire geometries absent from the training set.
  • A sharp falsification experiment is to remove a known wire from the scene and see whether the diffusion stage still hallucinates it from context; if it does, predictions rely on scene priors rather than radar echoes.
  • The dilated-LiDAR coarse masks could serve as an interpretable intermediate product for downstream planning, independent of whether the diffusion refinement is used.
  • The coarse-to-fine conditioning scheme should transfer to other low-RCS structures such as cables, fences, and branches, and to other degraded sensing modalities, but the paper does not demonstrate that transfer.
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

4 major / 5 minor

Summary. The paper proposes Sem-RaDiff, a two-stage diffusion-based framework that generates LiDAR-like 3D semantic point clouds from mmWave radar spherical cubes in agricultural scenes. Stage I filters sidelobe artifacts and predicts coarse structural/semantic masks supervised by dilated LiDAR point clouds; Stage II uses a conditional EDM/consistency model to refine these masks into fine-grained semantic point clouds. A parallel frame accumulation module fuses K=5 radar frames into a sparse Cartesian cube (RCC), used together with CFAR points (RPC). Experiments on a self-built agricultural drone dataset (training and test in different fields) report improved mIoU (up to 45.98 at 0.5 m) and lower GFLOPs/memory than re-implemented baselines. The abstract claims complete reconstruction and accurate classification of thin structures such as poles and wires.

Significance. If the results hold, the work is a useful step toward radar-only dense semantic perception for agricultural robots. The strengths are concrete: a coarse-to-fine design that reduces the radar-to-LiDAR modality gap, one-step consistency sampling, a held-out test set from a different field, and ablations supporting frame accumulation, Stage I, and dual input modality. I found no circularity in the derivations; the free parameters (K, qth, kernel sizes, class weights) are training choices, not outputs of the derivation. However, the headline generalization about thin structures is stronger than the evidence, and the radar grounding of wire/pole predictions is not directly demonstrated. The paper is publishable after the empirical support is strengthened and the claims calibrated.

major comments (4)
  1. [Abstract, Section V-C, Fig. 8, Table IV] The abstract's phrase 'complete reconstruction and accurate classification of thin structures such as poles and wires' is not supported by the paper's own results. In Fig. 8 and Section V-C, slanting wires are both missed and misclassified as trees; the authors attribute this to the training set containing only horizontal overhead wires and to RPC failing to detect the slanting wires. Table IV gives wire IoU only 18.89% (Ours-M, EDM, 0.5 m threshold). 'Complete reconstruction' is therefore an overstatement. Please either limit the claim to the trained wire geometry/configuration or add an evaluation covering varied wire orientations and report per-class precision/recall for thin structures.
  2. [Section IV-C, V-C, V-D, Table V] The claim that generated thin structures are grounded in genuine radar returns is not directly verified. The paper reports no SNR statistics for wire/pole returns after K=5 accumulation, no per-class detection recall on the accumulated RCC, and no analysis of whether predicted wire/pole voxels coincide with accumulated intensity peaks. The 'w/o SNR Enhancing' ablation (Table V) shows that accumulation improves wire IoU (from 8.44 to 18.89 at 0.5 m), which is indirect evidence, but it does not establish that five frames provide sufficient low-RCS signal; the slanting-wire failure is consistent with completion from scene priors. Stage I mask quality is also not separately evaluated. Please add direct evidence (e.g., alignment of predicted voxels with RCC peaks above a noise floor, or a false-cause analysis) or temper the radar-grounding claim.
  3. [Table IV, Table V, Section V-B] All quantitative conclusions rest on a single training run and on baselines that the authors re-implemented and modified (e.g., RadarHD* as a 3D U-Net, Sun* with a different loss, Luan* with a different 3D backbone). No error bars, confidence intervals, or significance tests are provided. This is particularly important for close geometric comparisons (e.g., Ours-M (CD) vs Luan*: Chamfer distance 0.79 vs 0.67 m and geometric IoU 62.16% vs 67.23% at 0.5 m) and for the claim of 'superior' performance. Please report results over multiple seeds or otherwise quantify variability, and describe the baseline tuning budget.
  4. [Section I, Section V-A, Fig. 3] The motivating scenario is sensor contamination (soiled camera/LiDAR), but the experiments do not include any contaminated-sensor condition. Since the method is radar-only at inference, clean-sensor results may transfer, but the abstract's 'viable alternative' claim, and the visual argument of Fig. 3, are not empirically validated. Please add a contamination or occlusion experiment, or explicitly scope the claim as an assumption about radar robustness rather than a demonstrated result.
minor comments (5)
  1. [Section V-E] 'The whole pipeline includes parallel frame accumulation, Stage I, and Stage I' should read 'Stage II'.
  2. [Section III-B] 'The reserve diffusion process' is a typo for 'reverse diffusion process'. Also, Eq. (4) is the denoising objective, not the reverse SDE; the text near Section IV-D referring to 'solving the reverse SDE (Eq. 4)' should cite Eq. (3).
  3. [Section IV-C] The Gaussian blur and greyscale dilation kernel sizes and sigma are not reported. Please specify them, as they define the Stage I supervision target and affect the coarse mask quality.
  4. [Fig. 2 caption] 'radar CFAR PC' is undefined; use 'RPC' consistently with the main text.
  5. [Section V-B] RadarHD* is described as having 100M parameters while Table III reports 106.4 MB; clarify whether this is parameter count or memory footprint.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Sem-RaDiff is an empirically supervised pipeline validated on held-out data; the only self-citation is a baseline, not a load-bearing premise.

full rationale

The paper's central claim is that a two-stage radar-to-LiDAR semantic point cloud generator, trained with LiDAR-derived supervision and evaluated on a held-out test set, outperforms baselines. No equation reduces to a fitted parameter or to a self-citation. Stage I's supervision (Gaussian-blurred, greyscale-dilated LiDAR masks, Section IV-C) is an external training target, not a quantity derived from the model's own predictions. Stage II's diffusion objective (Eq. 7) is a standard denoising loss conditioned on Stage I's output; the conditional input C is not a re-statement of the prediction target. Evaluation metrics (Table IV) are computed against held-out LiDAR ground truth with fixed thresholds, and the best-performing variant is selected from test-set measurements; this is empirical benchmarking, not circular derivation. The only self-citation is Radar-Diffusion [33], used as a baseline and in related work; it is not used to justify the method's correctness or uniqueness. The paper also honestly reports limitations: slanting wires are missed and misclassified as trees because the training set contains only horizontal wires and because the model relies on RPC that fails to detect them (Section V-C, Fig. 8). This is an acknowledged generalization failure, not a circularity. The reader's concern about radar grounding of thin-structure predictions is a sound empirical caveat, but it concerns whether the learned mapping is physically grounded or relies on scene priors; it does not make the derivation circular. Overall, the derivation chain is self-contained and empirically falsifiable, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper contains no new physical entities. Its numerical claims depend on several hand-chosen hyperparameters (K, qth, kernel size, class weights) and on dataset/calibration assumptions that are not independently verified.

free parameters (4)
  • K (frame accumulation depth) = 5
    Number of frames accumulated in Algorithm 1; chosen by hand, no sensitivity analysis reported. Directly controls SNR gain and latency.
  • qth (intensity filter percentile) = not specified
    Top-qth% intensity threshold applied in Algorithm 1 and in RCC sparsification. The percentile value is not given in the paper, yet it determines which radar voxels survive and therefore shapes all downstream inputs.
  • Stage I label dilation kernel and Gaussian sigma = not specified
    The coarse labels in Section IV-C are produced by Gaussian blurring and greyscale dilation with 'the same kernel size', but the kernel size/sigma are not reported; this hyperparameter determines how much wire/pole structure is injected into supervision.
  • Class weights for WCE and diffusion loss = not specified
    Weighted cross-entropy (WCE) for Stage I and class weight wc in the diffusion loss (Eq. 7) are mentioned but values are not given; these balance the rare pole/wire classes and materially affect the reported IoU numbers.
assumptions (4)
  • domain assumption LiDAR point clouds, after cropping and voxelization, define the complete and correct ground truth for radar-based reconstruction.
    Used in Sections IV-C and V to supervise both stages. If LiDAR misses thin or distant targets, or if calibration between radar and LiDAR is imperfect, the learned targets are biased and the reported IoU numbers are misleading.
  • domain assumption RTK poses and radar-LiDAR calibration are accurate to sub-voxel level for aligning K=5 frames and labels.
    Algorithm 1 relies on COORD_TRANSFORM to align past frames; Section V-A relies on spatial calibration between radar and LiDAR. Any residual error smears accumulated radar energy and mismatches labels.
  • domain assumption The environment is quasi-static across the accumulated frames (5 frames at 5 Hz) and the rotating scanning radar can be treated as a fixed Cartesian volume after ego-compensation.
    Parallel frame accumulation sums intensities in a fixed voxel grid; moving objects, sensor rotation, and object motion are not explicitly compensated, which could blur weak targets such as wires.
  • standard math EDM, consistency models, and Point Transformer V3 work as described in their source papers.
    Sections III-B and IV-D adopt the EDM framework, consistency distillation, and PTv3 without re-deriving them. This is standard practice for machine learning papers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sem-RaDiff: Diffusion-Based 3D Radar Semantic Perception in Cluttered Agricultural Environments." pith.science (2026). https://pith.science/paper/EPQE5A2R

@misc{pith2026250902283,
  author       = {Pith},
  title        = {Pith review of: Sem-RaDiff: Diffusion-Based 3D Radar Semantic Perception in Cluttered Agricultural Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EPQE5A2R}},
  note         = {Machine review of arXiv:2509.02283}
}
read the original abstract

Accurate and robust environmental perception is crucial for robot autonomous navigation. While current methods typically adopt optical sensors (e.g., camera, LiDAR) as primary sensing modalities, their susceptibility to visual occlusion often leads to degraded performance or complete system failure. In this paper, we focus on agricultural scenarios where robots are exposed to the risk of onboard sensor contamination. Leveraging radar's strong penetration capability, we introduce a radar-based 3D environmental perception framework as a viable alternative. It comprises three core modules designed for dense and accurate semantic perception: 1) Parallel frame accumulation to enhance signal-to-noise ratio of radar raw data. 2) A diffusion model-based hierarchical learning framework that first filters radar sidelobe artifacts then generates fine-grained 3D semantic point clouds. 3) A specifically designed sparse 3D network optimized for processing large-scale radar raw data. We conducted extensive benchmark comparisons and experimental evaluations on a self-built dataset collected in real-world agricultural field scenes. Results demonstrate that our method achieves superior structural and semantic prediction performance compared to existing methods, while simultaneously reducing computational and memory costs by 51.3% and 27.5%, respectively. Furthermore, our approach achieves complete reconstruction and accurate classification of thin structures such as poles and wires-which existing methods struggle to perceive-highlighting its potential for dense and accurate 3D radar perception.

Figures

Figures reproduced from arXiv: 2509.02283 by the authors.

Figure 1
Figure 1. Comparison of different approaches to cross-modal radar per￾ception. (a) Discriminative approaches attempt to make radar resemble high￾performance sensors through supervised learning. (b) Generative approaches take radar data as conditional input to achieve perception enhancing utilizing generative learning strategies such as diffusion models. (c) We integrate the above approaches into a coarse-to-fine paradigm that… view at source ↗
Figure 2
Figure 2. Overall framework of our Sem-RaDiff. It consists of three modules: (1) Radar data pre-processing that transforms multi-frame radar spherical cubes [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of different sensing modalities under normal and soiled [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Illustration of the radar signal processing pipeline, as well as the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Illustration of the proposed radar cube pre-processing method which [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Overall architecture of the proposed sparse network for both Stage I and Stage II. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results of single-frame 3D radar point cloud reconstruction and semantic prediction on the self-built dataset (test set). [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results of multi-frame accumulated 3D radar point cloud reconstruction and semantic prediction on the self-built dataset (test set). . [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 2
Figure 2. Figure 2: Fig.2. The results verify the effectiveness of Stage I within [PITH_FULL_IMAGE:figures/full_fig_p012_2.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 47 canonical work pages

  1. [1]

    Application of ai techniques and robotics in agriculture: A review,

    M. Wakchaure, B. Patle, and A. Mahindrakar, “Application of ai techniques and robotics in agriculture: A review,” Artificial Intelligence in the Life Sciences , vol. 3, p. 100057, 2023

    work page 2023

  2. [2]

    Xag reveals new-generation drones and robots for agrifuture,

    XAG, “Xag reveals new-generation drones and robots for agrifuture,” https://www.xa.com/en/news/official/xag/144, 2021

    work page 2021

  3. [3]

    Thor (ea-20x),

    Eavision, “Thor (ea-20x),” https://www.eavisionag.com/thor-ea-20x p6. html, 2021

    work page 2021

  4. [4]

    M. A. Richards et al. , Fundamentals of radar signal processing . Mcgraw-hill New York, 2005, vol. 1

    2005

  5. [5]

    Multiple emitter location and signal parameter estimation,

    R. Schmidt, “Multiple emitter location and signal parameter estimation,” IEEE transactions on antennas and propagation, vol. 34, no. 3, pp. 276– 280, 1986

    1986

  6. [6]

    Precise ego-motion estimation with millimeter-wave radar under diverse and challenging conditions,

    S. H. Cen and P. Newman, “Precise ego-motion estimation with millimeter-wave radar under diverse and challenging conditions,” in 2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2018, pp. 6045–6052

    work page 2018

  7. [7]

    Radar-inertial ego-velocity estimation for visually degraded environments,

    A. Kramer, C. Stahoviak, A. Santamaria-Navarro, A.-A. Agha- Mohammadi, and C. Heckman, “Radar-inertial ego-velocity estimation for visually degraded environments,” in2020 IEEE International Confer- ence on Robotics and Automation (ICRA). IEEE, 2020, pp. 5739–5746

    work page 2020

  8. [8]

    Milli- rio: Ego-motion estimation with low-cost millimetre-wave radar,

    Y . Almalioglu, M. Turan, C. X. Lu, N. Trigoni, and A. Markham, “Milli- rio: Ego-motion estimation with low-cost millimetre-wave radar,” IEEE Sensors Journal, vol. 21, no. 3, pp. 3314–3323, 2020

    work page 2020

Show all 62 references
  1. [9]

    Ramp-cnn: A novel neural net- work for enhanced automotive radar object recognition,

    X. Gao, G. Xing, S. Roy, and H. Liu, “Ramp-cnn: A novel neural net- work for enhanced automotive radar object recognition,” IEEE Sensors Journal, vol. 21, no. 4, pp. 5119–5132, 2020

    2020

  2. [10]

    Rodnet: A real-time radar object detection network cross-supervised by camera- radar fused object 3d localization,

    Y . Wang, Z. Jiang, Y . Li, J.-N. Hwang, G. Xing, and H. Liu, “Rodnet: A real-time radar object detection network cross-supervised by camera- radar fused object 3d localization,” IEEE Journal of Selected Topics in Signal Processing, vol. 15, no. 4, pp. 954–967, 2021

    2021

  3. [11]

    Robust small object detection on the water surface through fusion of camera and millimeter wave radar,

    Y . Cheng, H. Xu, and Y . Liu, “Robust small object detection on the water surface through fusion of camera and millimeter wave radar,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 15 263–15 272

    2021

  4. [12]

    Semantic seg- mentation on radar point clouds,

    O. Schumann, M. Hahn, J. Dickmann, and C. W ¨ohler, “Semantic seg- mentation on radar point clouds,” in 2018 21st International Conference on Information Fusion (FUSION) . IEEE, 2018, pp. 2179–2186

    2018

  5. [13]

    Rss-net: Weakly- supervised multi-class semantic segmentation with fmcw radar,

    P. Kaul, D. De Martini, M. Gadd, and P. Newman, “Rss-net: Weakly- supervised multi-class semantic segmentation with fmcw radar,” in 2020 IEEE Intelligent Vehicles Symposium (IV) . IEEE, 2020, pp. 431–436

    2020

  6. [14]

    Radar instance transformer: Reliable moving instance segmentation in sparse radar point clouds,

    M. Zeller, V . S. Sandhu, B. Mersch, J. Behley, M. Heidingsfeld, and C. Stachniss, “Radar instance transformer: Reliable moving instance segmentation in sparse radar point clouds,” IEEE Transactions on Robotics, vol. 40, pp. 2357–2372, 2023

    2023

  7. [15]

    Transradar: Adaptive- directional transformer for real-time multi-view radar semantic seg- mentation,

    Y . Dalbah, J. Lahoud, and H. Cholakkal, “Transradar: Adaptive- directional transformer for real-time multi-view radar semantic seg- mentation,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2024, pp. 353–362

    2024

  8. [16]

    Introduction to mmwave sensing: Fmcw radars,

    S. Rao, “Introduction to mmwave sensing: Fmcw radars,” Texas Instru- ments (TI) mmWave Training Series , pp. 1–11, 2017

    2017

  9. [17]

    K-radar: 4d radar object detection for autonomous driving in various weather conditions,

    D.-H. Paek, S.-H. Kong, and K. T. Wijaya, “K-radar: 4d radar object detection for autonomous driving in various weather conditions,” Ad- vances in Neural Information Processing Systems , vol. 35, pp. 3819– 3829, 2022

    2022

  10. [18]

    A new automotive radar 4d point clouds detector by using deep learning,

    Y . Cheng, J. Su, H. Chen, and Y . Liu, “A new automotive radar 4d point clouds detector by using deep learning,” in ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 8398–8402

    2021

  11. [19]

    A novel radar point cloud gen- eration method for robot environment perception,

    Y . Cheng, J. Su, M. Jiang, and Y . Liu, “A novel radar point cloud gen- eration method for robot environment perception,” IEEE Transactions on Robotics, vol. 38, no. 6, pp. 3754–3773, 2022

    2022

  12. [20]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III

    2015

  13. [21]

    Springer, 2015, pp. 234–241

    2015

  14. [22]

    Enhancing mmwave radar point cloud via visual-inertial supervision,

    C. Fan, S. Zhang, K. Liu, S. Wang, Z. Yang, and W. Wang, “Enhancing mmwave radar point cloud via visual-inertial supervision,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 9010–9017

    2024

  15. [23]

    Dream-pcd: Deep reconstruction and enhancement of mmwave radar pointcloud,

    R. Geng, Y . Li, D. Zhang, J. Wu, Y . Gao, Y . Hu, and Y . Chen, “Dream-pcd: Deep reconstruction and enhancement of mmwave radar pointcloud,” IEEE Transactions on Image Processing , 2024

    2024

  16. [24]

    Learned depth estimation of 3d imaging radar for indoor mapping,

    R. Xu, W. Dong, A. Sharma, and M. Kaess, “Learned depth estimation of 3d imaging radar for indoor mapping,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2022, pp. 13 260–13 267

    2022

  17. [25]

    Pillargen: Enhancing radar point cloud density and quality via pillar- based point generation network,

    J. Kim, G. Bang, K. Choi, M. Seong, J. Yoo, E. Pyo, and J. W. Choi, “Pillargen: Enhancing radar point cloud density and quality via pillar- based point generation network,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 9117–9124

    2024

  18. [26]

    High resolution point clouds from mmwave radar,

    A. Prabhakara, T. Jin, A. Das, G. Bhatt, L. Kumari, E. Soltanaghai, J. Bilmes, S. Kumar, and A. Rowe, “High resolution point clouds from mmwave radar,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 4135–4142

    2023

  19. [27]

    Radcloud: Real-time high-resolution point cloud generation using low-cost radars for aerial and ground vehicles,

    D. Hunt, S. Luo, A. Khazraei, X. Zhang, S. Hallyburton, T. Chen, and M. Pajic, “Radcloud: Real-time high-resolution point cloud generation using low-cost radars for aerial and ground vehicles,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 202...

    2024

  20. [28]

    Denserradar: A 4d millimeter-wave radar point cloud detector based on dense lidar point clouds,

    Z. Han, J. Jiang, X. Ding, J. Wang, Q. Meng, S. Xu, L. He, and J. Wang, “Denserradar: A 4d millimeter-wave radar point cloud detector based on dense lidar point clouds,” in 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2024, pp. 930–936

    2024

  21. [29]

    A deep automotive radar detector using the radelft dataset,

    I. Roldan, A. Palffy, J. F. Kooij, D. M. Gavrila, F. Fioranelli, and A. Yarovoy, “A deep automotive radar detector using the radelft dataset,” IEEE Transactions on Radar Systems , 2024

    2024

  22. [30]

    See through smoke: robust indoor mapping with low-cost mmwave radar,

    C. X. Lu, S. Rosa, P. Zhao, B. Wang, C. Chen, J. A. Stankovic, N. Trigoni, and A. Markham, “See through smoke: robust indoor mapping with low-cost mmwave radar,” in Proceedings of the 18th International Conference on Mobile Systems, Applications, and Services, 2020, pp. 14–27

    2020

  23. [31]

    Through fog high-resolution imaging using millimeter wave radar,

    J. Guan, S. Madani, S. Jog, S. Gupta, and H. Hassanieh, “Through fog high-resolution imaging using millimeter wave radar,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 11 464–11 473

    2020

  24. [32]

    3drimr: 3d recon- struction and imaging via mmwave radar based on deep learning,

    Y . Sun, Z. Huang, H. Zhang, Z. Cao, and D. Xu, “3drimr: 3d recon- struction and imaging via mmwave radar based on deep learning,” in 2021 IEEE International Performance, Computing, and Communications Conference (IPCCC). IEEE, 2021, pp. 1–8

    2021

  25. [33]

    Deeppoint: A deep learning model for 3d reconstruction in point clouds via mmwave radar,

    Y . Sun, H. Zhang, Z. Huang, and B. Liu, “Deeppoint: A deep learning model for 3d reconstruction in point clouds via mmwave radar,” arXiv preprint arXiv:2109.09188, 2021

    2021 arXiv

  26. [34]

    Towards dense and accurate radar perception via efficient cross-modal diffusion model,

    R. Zhang, D. Xue, Y . Wang, R. Geng, and F. Gao, “Towards dense and accurate radar perception via efficient cross-modal diffusion model,” IEEE Robotics and Automation Letters , 2024

    2024

  27. [35]

    Diffradar: High-quality mmwave radar perception with diffusion prob- abilistic model,

    J. Wu, R. Geng, Y . Li, D. Zhang, Z. Lu, Y . Hu, and Y . Chen, “Diffradar: High-quality mmwave radar perception with diffusion prob- abilistic model,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 8291–8295

    2024

  28. [36]

    Diffusion- based point cloud super-resolution for mmwave radar data,

    K. Luan, C. Shi, N. Wang, Y . Cheng, H. Lu, and X. Chen, “Diffusion- based point cloud super-resolution for mmwave radar data,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 11 171–11 177

    2024

  29. [37]

    Diffusion-based mmwave radar point cloud enhancement driven by range images,

    R. Wu, Z. Li, J. Wang, X. Xu, H. Yu, Z. Zheng, K. Huang, and G. Lu, “Diffusion-based mmwave radar point cloud enhancement driven by range images,” arXiv preprint arXiv:2503.02300 , 2025

    2025

  30. [38]

    R2ldm: An efficient 4d radar super-resolution framework leveraging diffusion model,

    B. Zheng, S. Lu, R. Huang, M. Huang, F. Lu, W. Tian, G. Zhuo, and L. Xiong, “R2ldm: An efficient 4d radar super-resolution framework leveraging diffusion model,” arXiv preprint arXiv:2503.17097 , 2025

    2025 arXiv

  31. [39]

    Unsupervised radar point cloud enhancement via arbitrary lidar guided diffusion prior,

    Y . Yang, J. Liu, G. Luo, H. Li, E. Ahn, M. R. Azghadi, and T. Huang, “Unsupervised radar point cloud enhancement via arbitrary lidar guided diffusion prior,” arXiv preprint arXiv:2505.09887 , 2025

    2025 arXiv

  32. [40]

    Sddiff: Boost radar perception via spatial-doppler diffusion,

    S. Wang, X. Luo, Y . Xie, and W. Wang, “Sddiff: Boost radar perception via spatial-doppler diffusion,” arXiv preprint arXiv:2506.16936 , 2025

    2025 arXiv

  33. [41]

    Conditional generative adversarial nets,

    M. Mirza and S. Osindero, “Conditional generative adversarial nets,” arXiv preprint arXiv:1411.1784 , 2014

    2014 arXiv

  34. [42]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840– 6851, 2020

    2020

  35. [43]

    Score-based generative modeling through stochastic differ- ential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differ- ential equations,” in International Conference on Learning Representa- tions

  36. [44]

    Generative adversarial nets,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances in neural information processing systems , vol. 27, 2014

    2014

  37. [45]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  38. [46]

    Semrafiner: Panoptic segmentation in sparse and noisy radar point clouds,

    M. Zeller, D. C. Herraez, B. Ayan, J. Behley, M. Heidingsfeld, and C. Stachniss, “Semrafiner: Panoptic segmentation in sparse and noisy radar point clouds,” IEEE Robotics and Automation Letters , 2024

    2024

  39. [47]

    Licrocc: Teach radar for accurate semantic occupancy prediction using lidar and camera,

    Y . Ma, J. Mei, X. Yang, L. Wen, W. Xu, J. Zhang, X. Zuo, B. Shi, and Y . Liu, “Licrocc: Teach radar for accurate semantic occupancy prediction using lidar and camera,” IEEE Robotics and Automation Letters , 2024

    2024

  40. [48]

    Radarocc: Robust 3d occupancy prediction with 4d imaging radar,

    F. Ding, X. Wen, Y . Zhu, Y . Li, and C. X. Lu, “Radarocc: Robust 3d occupancy prediction with 4d imaging radar,” Advances in Neural Information Processing Systems , vol. 37, pp. 101 589–101 617, 2024

    2024

  41. [49]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 652–660

    2017

  42. [50]

    Automatic labelling & semantic segmentation with 4d radar tensors,

    B. Sun, I. Roldan, and F. Fioranelli, “Automatic labelling & semantic segmentation with 4d radar tensors,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  43. [51]

    Mimo-sar: A hierarchical high-resolution imaging algorithm for mmwave fmcw radar in autonomous driving,

    X. Gao, S. Roy, and G. Xing, “Mimo-sar: A hierarchical high-resolution imaging algorithm for mmwave fmcw radar in autonomous driving,” IEEE Transactions on Vehicular Technology , vol. 70, no. 8, pp. 7322– 7334, 2021

    2021

  44. [52]

    Elucidating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,” Advances in Neural Infor- mation Processing Systems , vol. 35, pp. 26 565–26 577, 2022

    2022

  45. [53]

    Estimation of non-normalized statistical models by score matching

    A. Hyv ¨arinen and P. Dayan, “Estimation of non-normalized statistical models by score matching.” Journal of Machine Learning Research , vol. 6, no. 4, 2005

    2005

  46. [54]

    Progressive distillation for fast sampling of diffusion models,

    T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,” in International Conference on Learning Represen- tations, 2022

    2022

  47. [55]

    Consistency models,

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” arXiv preprint arXiv:2303.01469 , 2023

    2023 arXiv

  48. [56]

    Point transformer v3: Simpler faster stronger,

    X. Wu, L. Jiang, P.-S. Wang, Z. Liu, X. Liu, Y . Qiao, W. Ouyang, T. He, and H. Zhao, “Point transformer v3: Simpler faster stronger,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 4840–4851

    2024

  49. [57]

    Point transformer v2: Grouped vector attention and partition-based pooling,

    X. Wu, Y . Lao, L. Jiang, X. Liu, and H. Zhao, “Point transformer v2: Grouped vector attention and partition-based pooling,” Advances in Neural Information Processing Systems , vol. 35, pp. 33 330–33 342, 2022

    2022

  50. [58]

    On the variance of the adaptive learning rate and beyond,

    L. Liu, H. Jiang, P. He, W. Chen, X. Liu, J. Gao, and J. Han, “On the variance of the adaptive learning rate and beyond,” in International Conference on Learning Representations , 2020

    2020

  51. [59]

    Unetr++: delving into efficient and accurate 3d medical image segmentation,

    A. Shaker, M. Maaz, H. Rasheed, S. Khan, M.-H. Yang, and F. S. Khan, “Unetr++: delving into efficient and accurate 3d medical image segmentation,” IEEE Transactions on Medical Imaging , vol. 43, no. 9, pp. 3377–3390, 2024

    2024

  52. [60]

    Monoscene: Monocular 3d semantic scene completion,

    A.-Q. Cao and R. De Charette, “Monoscene: Monocular 3d semantic scene completion,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 3991–4001

    2022

  53. [61]

    Diff-unet: A diffu- sion embedded network for volumetric segmentation,

    Z. Xing, L. Wan, H. Fu, G. Yang, and L. Zhu, “Diff-unet: A diffu- sion embedded network for volumetric segmentation,” arXiv preprint arXiv:2303.10326, 2023

    2023 arXiv

  54. [62]

    Real-time 3d semantic occupancy prediction for autonomous vehicles using memory-efficient sparse convolution,

    S. Sze and L. Kunze, “Real-time 3d semantic occupancy prediction for autonomous vehicles using memory-efficient sparse convolution,” in 2024 IEEE Intelligent Vehicles Symposium (IV) . IEEE, 2024, pp. 1286–1293

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.