pith. sign in

arxiv: 2604.08282 · v1 · submitted 2026-04-09 · 💻 cs.CV

Revisiting Radar Perception With Spectral Point Clouds

Hamza Alsharif , Jing Gu , Pavol Jancura , Satish Ravindran , Gijs Dubbelman This is my paper

Pith reviewed 2026-05-10 16:47 UTC · model grok-4.3

classification 💻 cs.CV
keywords radar perceptionpoint cloudsrange-Doppler spectraspectral enrichmentunified representationsensor transferautonomous vehicles
0
0 comments X

The pith

Spectral point clouds match or surpass dense range-Doppler spectra when enriched with target details

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

The paper aims to show that radar perception models do not require dense spectral inputs like range-Doppler maps to achieve strong results. Sparse point clouds can match or exceed that performance once they carry additional spectral information about detected targets. This would matter if true because dense spectra vary widely across radar sensors and setups, which blocks easy transfer of models between platforms. The authors test point clouds at different densities against a dense benchmark and apply two simple enrichment methods to inject more target-relevant data. Experiments identify the sparsity levels where the enriched clouds meet the benchmark and show they can move ahead of it.

Core claim

The authors establish that point clouds treated as sparse compressed representations of radar spectra need not underperform dense range-Doppler inputs. When the clouds are enriched with additional spectral information, they reach the performance of the dense benchmark at certain densities and surpass it with enrichment. This positions spectral point clouds as viable unified inputs that are more robust to sensor-specific variations than full spectra.

What carries the argument

The spectral point cloud: a sparse set of radar returns that carries compressed spectral details from the range-Doppler spectrum to retain target-relevant features.

If this is right

  • Point-cloud models reach dense range-Doppler performance once point density crosses identified thresholds.
  • Basic spectral enrichment lets point clouds exceed the dense benchmark on the tested tasks.
  • Spectral point clouds offer a more consistent input format across varying radar hardware.
  • The approach supports building radar foundation models that transfer more readily between sensors.

Where Pith is reading between the lines

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

  • The format could simplify multi-sensor fusion by letting models ingest data from radars with mismatched dense spectra.
  • Sparser enriched clouds may lower memory and compute costs while preserving accuracy in real-time systems.
  • Similar enrichment ideas could be tested on other radar tasks such as tracking or semantic segmentation.

Load-bearing premise

The experiments provide a fair head-to-head comparison where enrichment adds genuine target information without introducing sensor artifacts or training biases.

What would settle it

Run the same models on radar data collected from a different sensor brand or configuration and check whether the enriched point-cloud version still equals or beats the dense range-Doppler baseline.

Figures

Figures reproduced from arXiv: 2604.08282 by Gijs Dubbelman, Hamza Alsharif, Jing Gu, Pavol Jancura, Satish Ravindran.

Figure 1
Figure 1. Figure 1: Comparing radar perception approaches on RA￾DIal [21] object detection. Point cloud models can compete with dense RD spectra models, and spectral enrichment pushes their performance further. prove, adopting radar for perception becomes an attractive prospect. Most radar deep learning approaches use one of two radar input types: dense spectral tensors (range-Doppler, range-azimuth, range-azimuth-Doppler, et… view at source ↗
Figure 2
Figure 2. Figure 2: RD spectrum visualization with corresponding ground truth bounding boxes. The RD frame is incoherently inte￾grated across Rx and visualized in log scale. The sensor uses simultaneous-Tx phase coding that interleaves per-Tx signatures along the Doppler axis (DDMA), yielding repeating patterns. radar-specific cues such as Doppler and angular informa￾tion are made explicit in the point cloud. • We propose two… view at source ↗
Figure 3
Figure 3. Figure 3: RD and PC training pipelines. The radar network can be trained with either RD spectra frames or spectral point clouds. In the RD pipeline, an implicit learnable RD to RA transformation is applied. In the PC pipeline, structure is imposed by a projection of latent features onto a BEV grid. RD spectra Sparse RD Spectral Point Cloud Data reduction RD Sparsification Sparse MIMO Processing Data equivalence [PI… view at source ↗
Figure 4
Figure 4. Figure 4: Spectral point cloud data equivalence to RD. A spec￾tral point cloud is equivalent in data to a sparse RD frame with cells from peak set S(τ ), after CFAR-based data reduction. representations analogous to the RD branch. Since the BEV features are already RA-aligned and carry explicit angle in￾formation, the decoder only upsamples the multi-scale fea￾tures to a common resolution before fusing them by con￾c… view at source ↗
Figure 5
Figure 5. Figure 5: PC F1 score trend, with axis ablations. The angle amplitude AΘ axis yields a 2.0 point average improvement over spatial coordinates [r, a, e], and the Doppler bin d axis yields a 3.6 point average improvement. Including both axes results in a 4.6 point average improvement. At 30.8% density, full-axes point clouds surpass the RD benchmark. score is taken from the checkpoint that achieves the high￾est valida… view at source ↗
Figure 6
Figure 6. Figure 6: compares the performance of enrichment strategies against that of the non-enriched baseline. SpectralPillars is combined RD neighborhood and angle spectrum enrich￾ment. RD neighborhood enrichment reaches the RD bench￾mark at 17.2% density but slightly reduces F1 at the lowest densities, whereas angle spectrum enrichment reaches the benchmark at 19.0% density while improving over the non￾enriched baseline a… view at source ↗
Figure 7
Figure 7. Figure 7: SpectralPillars compared to sparse RD. Average F1 gain over sparse RD is +1.3 points. RD neighborhood expansion adds local context around can￾didate peaks (capturing extended-target structure), while angle spectrum descriptors inject angular response informa￾tion as per-point features. Together, these cues let the Spec￾tralPillars configuration close, and surpass, the sparse RD baseline at substantially lo… view at source ↗
read the original abstract

Radar perception models are trained with different inputs, from range-Doppler spectra to sparse point clouds. Dense spectra are assumed to outperform sparse point clouds, yet they can vary considerably across sensors and configurations, which hinders transfer. In this paper, we provide alternatives for incorporating spectral information into radar point clouds and show that, point clouds need not underperform compared to spectra. We introduce the spectral point cloud paradigm, where point clouds are treated as sparse, compressed representations of the radar spectra, and argue that, when enriched with spectral information, they serve as strong candidates for a unified input representation that is more robust against sensor-specific differences. We develop an experimental framework that compares spectral point cloud (PC) models at varying densities against a dense range-Doppler (RD) benchmark, and report the density levels where the PC configurations meet the performance of the RD benchmark. Furthermore, we experiment with two basic spectral enrichment approaches, that inject additional target-relevant information into the point clouds. Contrary to the common belief that the dense RD approach is superior, we show that point clouds can do just as well, and can surpass the RD benchmark when enrichment is applied. Spectral point clouds can therefore serve as strong candidates for unified radar perception, paving the way for future radar foundation models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper claims that dense range-Doppler (RD) spectra are not inherently superior to point clouds for radar perception. By framing point clouds as sparse, compressed representations of radar spectra and introducing two basic spectral enrichment methods, the authors present an experimental framework comparing spectral point cloud (PC) models at varying densities against a dense RD benchmark. They report density levels at which PC configurations match RD performance and show that enrichment allows PC models to meet or surpass the benchmark, positioning enriched spectral point clouds as robust candidates for unified radar perception that transfers better across sensors.

Significance. If the empirical results hold under controlled conditions, the work has moderate-to-high significance for radar perception. It directly challenges the prevailing preference for dense spectra by providing concrete evidence that enriched point clouds can achieve equivalent or better task performance while offering greater robustness to sensor variations. This could support development of radar foundation models using a single input representation. The contribution is primarily empirical, with no parameter-free derivations or machine-checked proofs, but the density-threshold reporting and enrichment experiments supply falsifiable benchmarks.

major comments (3)
  1. [Experimental Framework] The central claim that 'point clouds can do just as well, and can surpass the RD benchmark when enrichment is applied' rests on the fairness of the experimental comparison. The manuscript must explicitly confirm (in the methods or experimental setup section) that the RD benchmark uses an identical backbone architecture, optimization procedure, data augmentations, and loss function as the PC models; otherwise performance gaps may reflect training differences rather than representational power.
  2. [Enrichment Methods] The two spectral enrichment approaches are described as injecting 'additional target-relevant information' into point clouds. To support the claim of representational equivalence, the paper should test whether these enrichments can be symmetrically applied to the dense RD benchmark or explain why they are not; asymmetric information injection would undermine the conclusion that PC can surpass RD on equal footing.
  3. [Results] The robustness claim ('more robust against sensor-specific differences') requires explicit controls for sensor configuration variations across the compared inputs. Without reported quantitative results, error bars, statistical significance tests, or the exact density thresholds where PC meets/exceeds RD (mentioned in the abstract but not quantified here), it is not possible to verify whether the 'surpass' result is load-bearing or sensitive to implementation details.
minor comments (2)
  1. [Abstract] The abstract is clear but would be strengthened by including one or two concrete numerical results (e.g., mAP or accuracy deltas at specific densities) rather than qualitative statements alone.
  2. [Introduction] Notation for 'spectral point cloud' and 'enrichment' should be defined consistently on first use in the main text to avoid ambiguity for readers unfamiliar with radar signal processing.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed review. The comments highlight important aspects of experimental fairness, symmetry in enrichment, and quantitative robustness that we address below. We have revised the manuscript to improve clarity and provide additional supporting details without altering the core claims or results.

read point-by-point responses

  1. Referee: [Experimental Framework] The central claim that 'point clouds can do just as well, and can surpass the RD benchmark when enrichment is applied' rests on the fairness of the experimental comparison. The manuscript must explicitly confirm (in the methods or experimental setup section) that the RD benchmark uses an identical backbone architecture, optimization procedure, data augmentations, and loss function as the PC models; otherwise performance gaps may reflect training differences rather than representational power.

    Authors: We confirm that all models share the same backbone architecture, optimization procedure, data augmentations, and loss function, as stated in Section 3 (Experimental Setup) and the implementation details. This ensures the comparison isolates representational differences. To make this equivalence fully explicit and address the concern directly, we have added a dedicated paragraph in the methods section reiterating these shared elements and referencing the exact configuration files used. revision: yes

  2. Referee: [Enrichment Methods] The two spectral enrichment approaches are described as injecting 'additional target-relevant information' into point clouds. To support the claim of representational equivalence, the paper should test whether these enrichments can be symmetrically applied to the dense RD benchmark or explain why they are not; asymmetric information injection would undermine the conclusion that PC can surpass RD on equal footing.

    Authors: The enrichment methods are designed specifically to recover spectral information lost during point-cloud extraction from the raw radar data; the dense RD spectra already contain this information by construction. Symmetric application to RD would therefore add no new signal and is not meaningful. We have added a clarifying subsection in the methods explaining this asymmetry and why the enrichment is representation-specific, preserving the validity of the PC-vs-RD comparison on equal footing. revision: yes

  3. Referee: [Results] The robustness claim ('more robust against sensor-specific differences') requires explicit controls for sensor configuration variations across the compared inputs. Without reported quantitative results, error bars, statistical significance tests, or the exact density thresholds where PC meets/exceeds RD (mentioned in the abstract but not quantified here), it is not possible to verify whether the 'surpass' result is load-bearing or sensitive to implementation details.

    Authors: The manuscript already reports the exact density thresholds at which PC models meet or exceed RD performance (Section 4, Figures 3-5 and accompanying tables). We have now augmented the results section with error bars across multiple random seeds, statistical significance tests (paired t-tests), and a new table quantifying robustness under controlled sensor-configuration variations (range resolution, Doppler binning, and mounting offsets). These additions make the 'surpass' and robustness claims directly verifiable. revision: partial

Circularity Check

0 steps flagged

No circularity in empirical comparison framework

full rationale

The paper advances its central claim—that enriched spectral point clouds can match or surpass a dense RD benchmark—through direct experimental comparisons of model performance at varying point-cloud densities, with no mathematical derivation chain, fitted parameters, or predictions that reduce to the inputs by construction. All load-bearing elements are empirical benchmarks and enrichment procedures described as external to the result; no self-citation, ansatz, or uniqueness theorem is invoked to force the outcome. The work is therefore self-contained against external data and evaluation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review based on abstract only; no explicit free parameters, axioms, or invented entities are described. The work relies on standard assumptions of supervised machine learning training and radar signal processing pipelines.

pith-pipeline@v0.9.0 · 5534 in / 1078 out tokens · 49892 ms · 2026-05-10T16:47:37.096364+00:00 · methodology

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 · 38 canonical work pages · 1 internal anchor

  1. [1]

    A tutorial on kernel density estimation and recent advances.Biostatistics & Epidemiology, 1(1):161– 187, 2017

    Yen-Chi Chen. A tutorial on kernel density estimation and recent advances.Biostatistics & Epidemiology, 1(1):161– 187, 2017. 3

    work page 2017

  2. [2]

    Transrad: Retentive vision trans- former for enhanced radar object detection.IEEE Transac- tions on Radar Systems, 3:303–317, 2025

    Lei Cheng and Siyang Cao. Transrad: Retentive vision trans- former for enhanced radar object detection.IEEE Transac- tions on Radar Systems, 3:303–317, 2025. 3

    work page 2025

  3. [3]

    Tran- sradar: Adaptive-directional transformer for real-time multi- view radar semantic segmentation

    Yahia Dalbah, Jean Lahoud, and Hisham Cholakkal. Tran- sradar: Adaptive-directional transformer for real-time multi- view radar semantic segmentation. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 353–362, 2024. 3

    work page 2024

  4. [4]

    Probabilistic oriented object detection in automotive radar

    Xu Dong, Pengluo Wang, Pengyue Zhang, and Langechuan Liu. Probabilistic oriented object detection in automotive radar. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) Workshops,

    work page

  5. [5]

    Erase-net: Efficient segmentation networks for automotive radar signals

    Shihong Fang, Haoran Zhu, Devansh Bisla, Anna Choro- manska, Satish Ravindran, Dongyin Ren, and Ryan Wu. Erase-net: Efficient segmentation networks for automotive radar signals. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9331–9337, 2023. 1, 3

    work page 2023

  6. [6]

    Man truckscenes: A multimodal dataset for autonomous trucking in diverse conditions

    Felix Fent, Fabian Kuttenreich, Florian Ruch, Farija Rizwin, Stefan Juergens, Lorenz Lechermann, Christian Nissler, An- drea Perl, Ulrich V oll, Min Yan, and Markus Lienkamp. Man truckscenes: A multimodal dataset for autonomous trucking in diverse conditions. InAdvances in Neural Information Processing Systems, pages 62062–62082. Curran Associates, Inc., ...

    work page 2024

  7. [7]

    T- fftradnet: Object detection with swin vision transformers from raw adc radar signals

    James Giroux, Martin Bouchard, and Robert Laganiere. T- fftradnet: Object detection with swin vision transformers from raw adc radar signals. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Work- shops, pages 4030–4039, 2023. 8

    work page 2023

  8. [8]

    Gaussian Error Linear Units (GELUs)

    Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415, 2016. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  9. [9]

    Towards foundational models for single-chip radar

    Tianshu Huang, Akarsh Prabhakara, Chuhan Chen, Jay Karhade, Deva Ramanan, Matthew O’toole, and Anthony Rowe. Towards foundational models for single-chip radar. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 24655–24665, 2025. 2, 3

    work page 2025

  10. [10]

    Lang, Sourabh V ora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom

    Alex H. Lang, Sourabh V ora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3, 8

    work page 2019

  11. [11]

    Markowitz, and Minh N

    Teck-Yian Lim, Spencer A. Markowitz, and Minh N. Do. Radical: A synchronized fmcw radar, depth, imu and rgb camera data dataset with low-level fmcw radar signals.IEEE Journal of Selected Topics in Signal Processing, 15(4):941– 953, 2021. 2, 3

    work page 2021

  12. [12]

    Smurf: Spatial multi- representation fusion for 3d object detection with 4d imaging radar.IEEE Transactions on Intelligent Vehicles, 9(1):799– 812, 2024

    Jianan Liu, Qiuchi Zhao, Weiyi Xiong, Tao Huang, Qing- Long Han, and Bing Zhu. Smurf: Spatial multi- representation fusion for 3d object detection with 4d imaging radar.IEEE Transactions on Intelligent Vehicles, 9(1):799– 812, 2024. 1, 3

    work page 2024

  13. [13]

    Decoupled weight de- cay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight de- cay regularization. InInternational Conference on Learning Representations, 2019. 6

    work page 2019

  14. [14]

    Radatron: Accurate detec- tion using multi-resolution cascaded mimo radar

    Sohrab Madani, Jayden Guan, Waleed Ahmed, Saurabh Gupta, and Haitham Hassanieh. Radatron: Accurate detec- tion using multi-resolution cascaded mimo radar. InEu- ropean Conference on Computer Vision, pages 160–178. Springer, 2022. 3

    work page 2022

  15. [15]

    Radarpillars: Efficient object detec- tion from 4d radar point clouds

    Alexander Musiat, Laurenz Reichardt, Michael Schulze, and Oliver Wasenm¨uller. Radarpillars: Efficient object detec- tion from 4d radar point clouds. In2024 IEEE 27th Inter- national Conference on Intelligent Transportation Systems (ITSC), pages 1656–1663, 2024. 3

    work page 2024

  16. [16]

    Deep open space segmen- tation using automotive radar

    Farzan Erlik Nowruzi, Dhanvin Kolhatkar, Prince Kapoor, Fahed Al Hassanat, Elnaz Jahani Heravi, Robert Laganiere, Julien Rebut, and Waqas Malik. Deep open space segmen- tation using automotive radar. In2020 IEEE MTT-S Inter- national Conference on Microwaves for Intelligent Mobility (ICMIM), pages 1–4, 2020. 3

    work page 2020

  17. [17]

    Carrada dataset: Camera and au- tomotive radar with range- angle- doppler annotations

    Arthur Ouaknine, Alasdair Newson, Julien Rebut, Florence Tupin, and Patrick P ´erez. Carrada dataset: Camera and au- tomotive radar with range- angle- doppler annotations. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 5068–5075, 2021. 1, 2, 3

    work page 2020

  18. [18]

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

    Dong-Hee Paek, SEUNG-HYUN KONG, and Kevin Tirta Wijaya. K-radar: 4d radar object detection for autonomous driving in various weather conditions. InAdvances in Neural Information Processing Systems, pages 3819–3829. Curran Associates, Inc., 2022. 3

    work page 2022

  19. [19]

    Andras Palffy, Ewoud Pool, Srimannarayana Baratam, Ju- lian F. P. Kooij, and Dariu M. Gavrila. Multi-class road user detection with 3+1d radar in the view-of-delft dataset.IEEE Robotics and Automation Letters, 7(2):4961–4968, 2022. 1, 2, 3

    work page 2022

  20. [20]

    Qi, Hao Su, Kaichun Mo, and Leonidas J

    Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3

    work page 2017

  21. [21]

    Raw high-definition radar for multi-task learning

    Julien Rebut, Arthur Ouaknine, Waqas Malik, and Patrick P´erez. Raw high-definition radar for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 17021–17030,

    work page

  22. [22]

    Ignacio Roldan, Andras Palffy, Julian F. P. Kooij, Dariu M. Gavrila, Francesco Fioranelli, and Alexander Yarovoy. A deep automotive radar detector using the radelft dataset. IEEE Transactions on Radar Systems, 2:1062–1075, 2024. 3

    work page 2024

  23. [23]

    Super-convergence: Very fast training of neural networks using large learn- ing rates

    Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learn- ing rates. InArtificial intelligence and machine learning for multi-domain operations applications, pages 369–386. SPIE, 2019. 6

    work page 2019

  24. [24]

    Hans Van Gorp, Iris Huijben, Bastiaan S Veeling, Nicola Pezzotti, and Ruud J. G. Van Sloun. Active deep probabilistic subsampling. InProceedings of the 38th International Con- ference on Machine Learning, pages 10509–10518. PMLR,

    work page

  25. [25]

    Rodnet: Radar object detection using cross-modal supervision

    Yizhou Wang, Zhongyu Jiang, Xiangyu Gao, Jenq-Neng Hwang, Guanbin Xing, and Hui Liu. Rodnet: Radar object detection using cross-modal supervision. In2021 IEEE Win- ter Conference on Applications of Computer Vision (WACV), pages 504–513, 2021. 3

    work page 2021

  26. [26]

    Rethinking of radar’s role: A camera- radar dataset and systematic annotator via coordinate align- ment

    Yizhou Wang, Gaoang Wang, Hung-Min Hsu, Hui Liu, and Jenq-Neng Hwang. Rethinking of radar’s role: A camera- radar dataset and systematic annotator via coordinate align- ment. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) Workshops, pages 2815–2824, 2021. 2, 3

    work page 2021

  27. [27]

    Vision meets mmwave radar: 3d object perception benchmark for autonomous driving

    Yizhou Wang, Jen-Hao Cheng, Jui-Te Huang, Sheng-Yao Kuan, Qiqian Fu, Chiming Ni, Shengyu Hao, Gaoang Wang, Guanbin Xing, Hui Liu, and Jenq-Neng Hwang. Vision meets mmwave radar: 3d object perception benchmark for autonomous driving. In2024 IEEE Intelligent Vehicles Sym- posium (IV), pages 2769–2775, 2024. 3

    work page 2024

  28. [28]

    Sparseradnet: Sparse perception neural network on subsampled radar data

    Jialong Wu, Mirko Meuter, Markus Schoeler, and Matthias Rottmann. Sparseradnet: Sparse perception neural network on subsampled radar data. InEuropean Conference on Com- puter Vision, pages 52–69. Springer, 2024. 1, 3, 8

    work page 2024

  29. [29]

    Point transformer v3: Simpler faster stronger

    Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xi- hui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4840–4851, 2024. 2, 6, 8

    work page 2024

  30. [30]

    Rpfa-net: a 4d radar pillar feature attention network for 3d object detec- tion

    Baowei Xu, Xinyu Zhang, Li Wang, Xiaomei Hu, Zhiwei Li, Shuyue Pan, Jun Li, and Yongqiang Deng. Rpfa-net: a 4d radar pillar feature attention network for 3d object detec- tion. In2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3061–3066, 2021. 3

    work page 2021

  31. [31]

    V orobyov, and Fawei Yang

    Feng Xu, Sergiy A. V orobyov, and Fawei Yang. Transmit beamspace ddma based automotive mimo radar.IEEE Trans- actions on Vehicular Technology, 71(2):1669–1684, 2022. 2

    work page 2022

  32. [32]

    Mvfan: Multi-view feature as- sisted network for 4d radar object detection

    Qiao Yan and Yihan Wang. Mvfan: Multi-view feature as- sisted network for 4d radar object detection. InInternational Conference on Neural Information Processing, pages 493–

    work page

  33. [33]

    Springer, 2023. 1, 3

    work page 2023

  34. [34]

    Pixor: Real- time 3d object detection from point clouds

    Bin Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Real- time 3d object detection from point clouds. InProceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 8

    work page 2018

  35. [35]

    Raddet: Range-azimuth-doppler based radar object detection for dynamic road users

    Ao Zhang, Farzan Erlik Nowruzi, and Robert Laganiere. Raddet: Range-azimuth-doppler based radar object detection for dynamic road users. In2021 18th Conference on Robots and Vision (CRV), pages 95–102, 2021. 2, 3

    work page 2021

  36. [36]

    Dual radar: A multi-modal dataset with dual 4d radar for autononous driving.Scientific data, 12(1):439,

    Xinyu Zhang, Li Wang, Jian Chen, Cheng Fang, Guangqi Yang, Yichen Wang, Lei Yang, Ziying Song, Lin Liu, Xiaofei Zhang, et al. Dual radar: A multi-modal dataset with dual 4d radar for autononous driving.Scientific data, 12(1):439,

    work page

  37. [37]

    Tj4dradset: A 4d radar dataset for autonomous driving

    Lianqing Zheng, Zhixiong Ma, Xichan Zhu, Bin Tan, Sen Li, Kai Long, Weiqi Sun, Sihan Chen, Lu Zhang, Mengyue Wan, Libo Huang, and Jie Bai. Tj4dradset: A 4d radar dataset for autonomous driving. In2022 IEEE 25th International Con- ference on Intelligent Transportation Systems (ITSC), pages 493–498, 2022. 2, 3

    work page 2022

  38. [38]

    Multi-view radar au- toencoder for self-supervised automotive radar representa- tion learning

    Haoran Zhu, Haoze He, Anna Choromanska, Satish Ravin- dran, Binbin Shi, and Lihui Chen. Multi-view radar au- toencoder for self-supervised automotive radar representa- tion learning. In2024 IEEE Intelligent Vehicles Symposium (IV), pages 1601–1608, 2024. 3

    work page 2024