Pith. sign in

REVIEW 3 major objections 5 minor 41 references

Co-learning radar antenna selection with camera/LiDAR fusion yields smaller, cheaper arrays that match or beat full-array detection.

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 ↗

Learning which MIMO radar receivers to activate jointly with camera–LiDAR fusion lets fewer receivers match or exceed full-array 3D detection on RADIal, with the best budget depending on the sensor stack.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Worth a serious referee: the radar-camera sparse-layout result is credible, but the 'optimal budget' claims need a λ ablation and matched statistics. the 3 major comments →

arxiv 2607.17351 v1 pith:WBZQGAKK submitted 2026-07-19 cs.AI cs.RO

DeeperRadar: End-to-End MIMO Radar Design and Multi-Modal Fusion for Autonomous Vehicle Perception

Eli Goldenshluger , Barak Pinkovich , Chaim Baskin This is my paper
classification cs.AI cs.RO
keywords MIMO radarsparse antenna selectionmulti-modal fusion3D object detectionraw radar ADCdifferentiable Top-KBEV perceptionautonomous driving
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.

The reading

This paper tries to show that the optimal MIMO radar configuration is not a standalone hardware problem but a perception problem shaped by the other sensors in the vehicle. The authors build a pipeline that learns, end-to-end, which radar receiver antennas to activate and how many, jointly with a BEV-based 3D detector that fuses raw radar with cameras and LiDAR. On the RADIal dataset, the learned sparse receiver masks match or exceed the full 16-receiver array across radar-only, radar-camera, and radar-LiDAR stacks, while using only 8, 7, and 3 receivers respectively. A sympathetic reader would care because the finding suggests radar cost and integration complexity can drop substantially without hurting detection—if the array is designed for the whole sensor stack.

Core claim

The central discovery is that the perception task itself can choose a smaller radar array: jointly learned receiver budgets converge to 8 antennas for radar-only, 7 for radar+camera, and 3 for radar+LiDAR on the RADIal dataset, matching or beating the full 16-antenna array in each case (radar+camera even gains about 2.6 AP points). Learned antenna placements also beat fixed heuristic layouts at every budget and stack. The interpretation is that the camera provides angular structure when radar aperture shrinks, radar keeps its range information, and LiDAR further lowers radar's marginal value.

What carries the argument

The load-bearing object is the budget-aware MIMO design module: a multivariate relaxed Bernoulli distribution (parameterized by a Gaussian copula over antennas) produces correlated channel scores, and a fractional relaxed Top-K operator converts them into a soft mask whose total mass equals a learnable continuous budget κ. The mask is applied to raw radar ADC data before beamforming, so the network can differentiate through antenna selection. The full training loss adds a sparsity penalty λκ to the task loss, and at deployment the module is replaced by the learned mask, leaving the downstream architecture unchanged. This machinery makes it possible to ask, for the first time, what radar conf

Load-bearing premise

The load-bearing assumption is that the learned receiver budgets are the true task-optimal counts, but they are actually the minimizers of a regularized loss whose sparsity strength λ was set by hand; a different λ would yield different 'optimal' budgets, so the conclusion that optimal design depends on the fusion stack is not separated from the choice of regularization.

What would settle it

Run the same co-design for each sensor stack with a sweep of the sparsity weight λ (including λ=0 and values an order of magnitude above and below 5×10^-4) and compare the resulting learned budgets and AP. If the reduced budgets (n_R ≈ 8, 7, 3) do not remain optimal across a wide λ range, or if at λ=0 the full 16-receiver array wins in every stack, then the stack-dependence claim is an artifact of regularization rather than a property of the perception task.

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

If this is right

  • Receiver count can be cut by 50 to 80 percent without loss of 3D detection accuracy when radar is fused with camera or LiDAR, according to the learned budgets in each stack.
  • Learned antenna placements consistently beat fixed heuristic layouts at every tested budget and every sensor stack, with gains up to roughly 6 percent relative AP in radar+camera.
  • The optimal radar design depends on the sensor suite: the model's attention shifts angular reasoning to images as the radar aperture shrinks, while radar continues to supply range.
  • Radar+LiDAR fusion underperforms a strong LiDAR-only baseline, consistent with the finding that complementary sensors reduce radar's marginal detection value.

Where Pith is reading between the lines

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

  • The reported 'optimal' budgets (8, 7, 3) are the minimizers of a regularized loss with a fixed λ = 5×10^-4; sweeping λ would likely trace a Pareto curve of receiver count versus AP, and the stack-dependence conclusion should be re-examined at λ = 0.
  • A natural extension is to co-learn not just receiver selection but transmit (Tx) selection and Doppler-aware objectives, which the paper names as future work and which could further cut hardware cost.
  • The method suggests a hardware-testing protocol: before fixing a MIMO array for a production vehicle, train the fusion model with the learnable mask to let the stack vote on the antenna budget.
  • The R+C+L result (71.95 AP) is reported as best observed across seeds due to high variance; if that variance persists, the claim for the full three-sensor stack is the least robust and needs more seeds.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes DeeperRadar, an end-to-end framework that jointly learns a sparse receiver activation mask for MIMO radar and a BEV-based multi-modal 3D detection network operating on raw radar ADC, camera images, and LiDAR. A fractional relaxed Top-K selector learns both the receiver subset and the effective receiver budget, and at deployment the learned mask is applied to the radar front-end. Experiments on the RADIal dataset cover radar-only, radar+camera, radar+LiDAR, and radar+camera+LiDAR stacks. The central claims are that learned sparse radar designs match or exceed full-array baselines with fewer receivers, and that the optimal MIMO design depends on the fusion stack and downstream task.

Significance. If the claimed results are established, the paper would make a valuable contribution: it is, to my knowledge, the first framework to co-design MIMO radar receiver selection with multi-modal fusion, and it addresses a practical cost/complexity knob for automotive radar. The reconstruction experiments provide a useful validation of learnable antenna layouts on automotive MIMO radar, and the idea of conditioning radar front-end design on the fusion stack is interesting and timely. However, the headline detection claims currently rest on an unmatched statistical comparison and on a fixed regularization weight, so the significance is conditional on the authors supplying additional evidence.

major comments (3)
  1. [§IV-C, Tables I and II] The central 'match or exceed full-array baselines' claim rests on an unmatched statistic. Table I reports mean±SD over 9 seeds for fixed budgets, while Table II reports 'best overall AP' for DeeperRadar (the caption says this for all configurations, and the text explicitly says R+C+L is reported as the best observed result across seeds). Comparing a best-of-K value to a mean is biased upward. Concretely, for R+L, Table I gives n_R=3 learned mean AP 50.28±1.17 vs. full-array mean 50.25±1.24 (no advantage), yet Table II reports 50.50 for n*_R=3. For R+C, 47.81 vs. 45.22±1.69 requires a paired per-seed comparison; no such test is reported. The headline claim is therefore not statistically supported.
  2. [§IV-A, Eq. (7)] The learned budget n*_R is not task-optimal; it is the minimizer of the regularized loss L = L_task + λκ with λ fixed at 5e-4 for all stacks. λ is a hand-set free parameter, and no λ ablation is reported. The values n*_R = 8, 7, 3 are points on one regularization path, not inherently optimal budgets. Different λ values could shift or erase the stack-dependence conclusion. The authors acknowledge this trade-off for the reconstruction task ('By tuning λ... one can select...'), and an analogous ablation is needed for the detection experiments before calling these budgets 'optimal'.
  3. [§IV-C, Table II] The full-stack headline result, 71.95 AP at n*_R=7, has no same-architecture full n_R=16 baseline. The R+C+L stack is explicitly excluded from the fixed-budget ablation study because of high run-to-run variance, and only the best observed result is reported. Therefore the statement that 'in all fusion settings the best operating point occurs at a reduced budget n*_R, matching or slightly surpassing the full n_R=16 array' is unsupported for the full fusion stack. A full-array R+C+L baseline with comparable seed statistics is required.
minor comments (5)
  1. [§II, Related Work] The sentence mentioning the Gumbel-Softmax distribution cites [24] and [25]; [25] is the U-Net paper and appears irrelevant here. Please correct the citation.
  2. [§III-B, Eq. (5)] Clarify how the soft mask m∈[0,1]^{N_R} used during training is converted to the deployed hard mask. The deployment description says 'zero out' removed antennas, but the training mask is fractional; the thresholding/rounding procedure should be specified.
  3. [§IV-C, Table I] The radar-only learned optimum n*_R=8 is not an evaluated row in Table I; the highlighted row is n_R=9. The text reports 32.95 AP for n*_R=8, but Table I has no row for n_R=8. This makes it hard to compare the learned budget result with the fixed-budget table.
  4. [Abstract / §IV-C] The abstract says sparse configurations 'match or exceed full-array baselines,' but for radar-only the learned n*_R=8 result (32.95 AP) is below the full-array mean (33.67±0.98). Please qualify the claim as 'comparable or better in fusion settings' or report corrected statistics.
  5. [§IV-B] The simulated radar setup is described as 'closely matching' the RADIal configuration, but the simulation's Tx/Rx geometry and waveform parameters are not given. Please add these details for reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the learned budget is explicitly the minimizer of the stated regularized objective and the headline claims are empirical comparisons, not identity-level derivations.

full rationale

The paper's derivation chain is not circular. The learned receiver budgets are defined explicitly as minimizers of the regularized loss L = L_task + lambda*kappa (Eq. 7), with a stated fixed lambda = 5e-4 (Section IV-A); Figure 4/5 even label n*_R as 'predicted by the budget-regularized objective.' Because the reported quantity is exactly the argmin of the stated objective, no fitted parameter is silently renamed as an independent prediction. The learned-subset vs. heuristic-subset comparisons are genuine empirical comparisons at fixed budgets, and the central 'match or exceed full-array baselines' claim is a measured comparison against a full-array baseline, not a consequence of an equation identity. The paper builds on prior work ([9], [16], [33]) but does not rely on any load-bearing self-citation; no uniqueness theorem or author-imported result is invoked to force the conclusion. The paper is also transparent about the limitations that could otherwise look like hidden assumptions: the Radar+Camera+LiDAR row is reported as the 'best observed result across seeds' (Section IV-A/Table II), and the fixed lambda determines the trade-off point. Those are validity/robustness concerns (lambda-ablation and matched-seed statistics would strengthen the paper), not circularity. No circular step can be exhibited from the paper's own equations or citations.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The paper's central claim rests primarily on a regularized loss with a hand-set sparsity trade-off (λ), a geometric alignment assumption, and an unvalidated fractional Top-K relaxation. No new physical entities are introduced. The learned mask is an optimization variable, not an invented entity.

free parameters (2)
  • Regularization weight λ = 5×10^-4
    Hand-fixed in Eq. (7) for all modality configurations. It directly biases the learned receiver budget toward smaller n_R; the reported n*_R values (8, 7, 3) are minimizers of this regularized objective, not unregularized task-optimal budgets.
  • Simulated radar configuration for reconstruction = 16 Rx, 12 Tx (closely matching RADIal)
    The paper states the exact RADIal radar configuration is not publicly available, so reconstruction experiments use a simulated approximation. This choice affects the PSNR results and their transferability to the real hardware.
axioms (3)
  • ad hoc to paper The fractional relaxed Top-K construction (decompose κ into floor and fractional part, iterate relaxed Top-1) yields a valid differentiable subset-selection relaxation.
    Described in Section III-B, Equations (5)-(6). No proof, derivation, or ablation is provided that this decomposition faithfully approximates the intended relaxed Bernoulli/top-k distribution or preserves gradient quality.
  • domain assumption Polar-aligned geometry: each 3D point projects to one image column and one radar row per BEV query.
    Borrowed from EchoFusion [16], used in Eqs. (11)-(13). Underlies the fusion mechanism and the interpretation that the camera compensates for reduced radar angular resolution. Depends on calibration and sensor geometry assumptions.
  • standard math The Gaussian copula / multivariate relaxed Bernoulli with learnable Σ and α (Eqs. 1-3) adequately models antenna subset selection.
    Copied from Weiss et al. [9] and Wang-Yin [32]. It is a modeling choice whose realism for physical antenna arrays is assumed, not validated against hardware constraints such as mutual coupling or switch limitations.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DeeperRadar: End-to-End MIMO Radar Design and Multi-Modal Fusion for Autonomous Vehicle Perception." pith.science (2026). https://pith.science/paper/WBZQGAKK

@misc{pith2026260717351,
  author       = {Pith},
  title        = {Pith review of: DeeperRadar: End-to-End MIMO Radar Design and Multi-Modal Fusion for Autonomous Vehicle Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBZQGAKK}},
  note         = {Machine review of arXiv:2607.17351}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

DeeperRadar is a radar-centric, sensor-stack-conditioned framework that co-designs radar sensing and multi-modal 3D detection for autonomous mobility by learning a sparse acquisition pattern end-to-end with the fusion model. A learnable MIMO design module is trained end-to-end within a fusion network that operates directly on raw radar ADC data together with camera images and LiDAR point clouds. During training, the design module is supervised by the other sensors, enabling the system to learn both which receiver antennas to activate and the effective number of them. At deployment, the design module is removed and replaced by the learned sparse subsampling mask, leaving the downstream model architecture unchanged. Evaluated on the RADIal dataset, DeeperRadar discovers sparse, task-aware radar configurations that match or exceed full-array baselines while using fewer receivers, potentially reducing radar cost and integration complexity. These results show that learned optimal MIMO radar design depends on the fusion stack and the downstream perception task.

Figures

Figures reproduced from arXiv: 2607.17351 by Barak Pinkovich, Chaim Baskin, Eli Goldenshluger.

Figure 1
Figure 1. Figure 1: High-level overview of our radar-centric pipeline: during training, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Effect of receiver reduction on MIMO radar range–azimuth (RA) [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed DeeperRadar pipeline. LiDAR point clouds, raw radar data, and camera images serve as input. (a) A MIMO design [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overall 3D detection accuracy (AP@0.5) versus radar receiver budget [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Reconstruction PSNR versus number of antennas [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Attention analysis for varying radar budgets [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 13 linked inside Pith

  1. [1]

    Deepfusion: A robust and modular 3d object detector for lidars, cam- eras and radars,

    F. Drews, D. Feng, F. Faion, L. Rosenbaum, M. Ulrich, and C. Gl ¨aser, “Deepfusion: A robust and modular 3d object detector for lidars, cam- eras and radars,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022

    2022

  2. [2]

    Experimental validation of lidar sensors used in vehicular applications by using a mobile platform for distance and speed measurements,

    I. Vasile, E. Tudor, I.-C. Sburlan, M.-A. Ghet ,i, and G. Popa, “Experimental validation of lidar sensors used in vehicular applications by using a mobile platform for distance and speed measurements,” Sensors, vol. 21, no. 23, 2021. [Online]. Available: https://www.mdpi. com/1424-8220/21/23/8147

    2021

  3. [3]

    An overview of autonomous vehicles sensors and their vulnerability to weather conditions,

    J. Vargas, S. Alsweiss, O. Toker, R. Razdan, and J. Santos, “An overview of autonomous vehicles sensors and their vulnerability to weather conditions,”Sensors, vol. 21, no. 16, 2021. [Online]. Available: https://www.mdpi.com/1424-8220/21/16/5397

    2021

  4. [4]

    Mimo radar: an idea whose time has come,

    E. Fishler, A. Haimovich, R. Blum, D. Chizhik, L. Cimini, and R. Valenzuela, “Mimo radar: an idea whose time has come,” inProceedings of the 2004 IEEE Radar Conference (IEEE Cat. No.04CH37509), 2004, pp. 71–78

    2004

  5. [5]

    Richards,Fundamentals of Radar Signal Processing, Second Edition

    M. Richards,Fundamentals of Radar Signal Processing, Second Edition. McGraw-Hill Education, 2014. [Online]. Available: https: //books.google.co.il/books?id=i5KAMAEACAAJ

    2014

  6. [6]

    Compressed sensing,

    D. Donoho, “Compressed sensing,”IEEE Transactions on Information Theory, vol. 52, no. 4, pp. 1289–1306, 2006

    2006

  7. [7]

    Mimo radar using compressive sampling,

    Y . Yu, A. P. Petropulu, and H. V . Poor, “Mimo radar using compressive sampling,”IEEE Journal of Selected Topics in Signal Processing, vol. 4, no. 1, p. 146–163, Feb. 2010. [Online]. Available: http://dx.doi.org/10.1109/JSTSP.2009.2038973

    arXiv 2010

  8. [8]

    Spatial compressive sensing for mimo radar,

    M. Rossi, A. M. Haimovich, and Y . C. Eldar, “Spatial compressive sensing for mimo radar,”IEEE Transactions on Signal Processing, vol. 62, no. 2, p. 419–430, Jan. 2014. [Online]. Available: http://dx.doi.org/10.1109/TSP.2013.2289875

    arXiv 2014

  9. [9]

    Joint optimization of system design and reconstruction in mimo radar imaging,

    T. Weiss, N. Peretz, S. Vedula, A. Feuer, and A. Bronstein, “Joint optimization of system design and reconstruction in mimo radar imaging,” 2021. [Online]. Available: https://arxiv.org/abs/2110.03218

    Pith/arXiv arXiv 2021

  10. [10]

    A review of multi-sensor fusion in autonomous driving,

    H. Qian, M. Wang, M. Zhu, and H. Wang, “A review of multi-sensor fusion in autonomous driving,”Sensors, vol. 25, no. 19, 2025. [Online]. Available: https://www.mdpi.com/1424-8220/25/19/6033

    2025

  11. [11]

    Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,

    Z. Liu, H. Tang, A. Amini, X. Yang, H. Mao, D. Rus, and S. Han, “Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,” 2024. [Online]. Available: https://arxiv.org/abs/2205.13542

    Pith/arXiv arXiv 2024

  12. [12]

    Metabev: Solving sensor failures for bev detection and map segmentation,

    C. Ge, J. Chen, E. Xie, Z. Wang, L. Hong, H. Lu, Z. Li, and P. Luo, “Metabev: Solving sensor failures for bev detection and map segmentation,” 2023. [Online]. Available: https://arxiv.org/abs/2304. 09801

    2023

  13. [13]

    Centerfusion: Center-based radar and camera fusion for 3d object detection,

    R. Nabati and H. Qi, “Centerfusion: Center-based radar and camera fusion for 3d object detection,” in2021 IEEE Winter Conference on Applications of Computer Vision (WACV). IEEE, Jan. 2021, p. 1526–1535. [Online]. Available: http://dx.doi.org/10.1109/ W ACV48630.2021.00157

    arXiv 2021

  14. [14]

    Ralibev: Radar and lidar bev fusion learning for anchor box free object detection systems,

    Y . Yang, J. Liu, T. Huang, Q.-L. Han, G. Ma, and B. Zhu, “Ralibev: Radar and lidar bev fusion learning for anchor box free object detection systems,” 2024. [Online]. Available: https: //arxiv.org/abs/2211.06108

    Pith/arXiv arXiv 2024

  15. [15]

    Bev-guided multi-modality fusion for driving perception,

    Y . Man, L.-Y . Gui, and Y .-X. Wang, “Bev-guided multi-modality fusion for driving perception,” in2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 21 960– 21 969

    2023

  16. [16]

    Echoes Beyond Points: Unleashing the Power of Raw Radar Data in Multi-modality Fusion,

    Y . Liu, F. Wang, N. Wang, and Z. Zhang, “Echoes Beyond Points: Unleashing the Power of Raw Radar Data in Multi-modality Fusion,” inNeurIPS, 2023

    2023

  17. [17]

    Raw high- definition radar for multi-task learning,

    J. Rebut, A. Ouaknine, W. Malik, and P. P ´erez, “Raw high- definition radar for multi-task learning,” 2022. [Online]. Available: https://arxiv.org/abs/2112.10646

    Pith/arXiv arXiv 2022

  18. [18]

    Pointpillars: Fast encoders for object detection from point clouds,

    A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,”

  19. [19]

    Multi-class road user detection with 3+1d radar in the view-of-delft dataset,

    A. Palffy, E. Pool, S. Baratam, J. Kooij, and D. Gavrila, “Multi-class road user detection with 3+1d radar in the view-of-delft dataset,”IEEE Robotics and Automation Letters, vol. 7, pp. 4961–4968, 04 2022

    2022

  20. [20]

    Ratrack: Moving object detection and tracking with 4d radar point cloud,

    Z. Pan, F. Ding, H. Zhong, and C. X. Lu, “Ratrack: Moving object detection and tracking with 4d radar point cloud,” 2024. [Online]. Available: https://arxiv.org/abs/2309.09737

    Pith/arXiv arXiv 2024

  21. [21]

    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, p. 1–8, 2024. [Online]. Available: http://dx.doi.org/10.1109/LRA.2024. 3502058

    doi:10.1109/lra.2024 2024

  22. [22]

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

    J. Giroux, M. Bouchard, and R. Laganiere, “T-fftradnet: Object detection with swin vision transformers from raw adc radar signals,”

  23. [23]

    Adcnet: Learning from raw radar data via distillation,

    B. Yang, I. Khatri, M. Happold, and C. Chen, “Adcnet: Learning from raw radar data via distillation,” 2023. [Online]. Available: https://arxiv.org/abs/2303.11420

    Pith/arXiv arXiv 2023

  24. [24]

    Categorical reparameterization with gumbel-softmax,

    E. Jang, S. Gu, and B. Poole, “Categorical reparameterization with gumbel-softmax,” inInternational Conference on Learning Represen- tations (ICLR), 2017

    2017

  25. [25]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” 2015. [Online]. Available: https://arxiv.org/abs/1505.04597

    Pith/arXiv arXiv 2015

  26. [26]

    Joint optimization of sparse MIMO arrays and imaging methods,

    O. Oral and F. S. ¨Oktem, “Joint optimization of sparse MIMO arrays and imaging methods,” inProc. European Signal Processing Conference (EUSIPCO), 2025, pp. 740–744

    2025

  27. [27]

    Grif-net: Gated region of interest fusion network for robust 3d object detection from radar point cloud and monocular image,

    Y . Kim, J. W. Choi, and D. Kum, “Grif-net: Gated region of interest fusion network for robust 3d object detection from radar point cloud and monocular image,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020

    2020

  28. [28]

    Interfusion: Interaction-based 4d radar and lidar fusion for 3d object detection,

    L. Wang, X. Zhang, B. Xvet al., “Interfusion: Interaction-based 4d radar and lidar fusion for 3d object detection,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022

    2022

  29. [29]

    Towards robust 3d object detection with lidar and 4d radar fusion in various weather conditions,

    Y . Chae, H. Kim, and K.-J. Yoon, “Towards robust 3d object detection with lidar and 4d radar fusion in various weather conditions,” in2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 15 162–15 172

    2024

  30. [30]

    Delving into the devils of bird’s-eye-view perception: A review, evaluation and recipe,

    H. Li, C. Sima, J. Dai, W. Wang, L. Lu, H. Wang, J. Zeng, Z. Li, J. Yang, H. Deng, H. Tian, E. Xie, J. Xie, L. Chen, T. Li, Y . Li, Y . Gao, X. Jia, S. Liu, J. Shi, D. Lin, and Y . Qiao, “Delving into the devils of bird’s-eye-view perception: A review, evaluation and recipe,” 2023. [Online]. Available: https://arxiv.org/abs/2209.05324

    Pith/arXiv arXiv 2023

  31. [31]

    Exploring mmwave radar and camera fusion for high-resolution and long-range depth imaging,

    A. Prabhakara, D. Zhang, C. Li, S. Munir, A. C. Sankaranarayanan, A. Rowe, and S. Kumar, “Exploring mmwave radar and camera fusion for high-resolution and long-range depth imaging,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022

    2022

  32. [32]

    Relaxed multivariate bernoulli distribution and its applications to deep generative models,

    X. Wang and J. Yin, “Relaxed multivariate bernoulli distribution and its applications to deep generative models,” inProceedings of the 36th Conference on Uncertainty in Artificial Intelligence (UAI), ser. Proceedings of Machine Learning Research, J. Peters and D. Sontag, Eds., vol. 124. PMLR, 03–06 Aug 2020, pp. 500–509. [Online]. Available: https://proce...

    2020

  33. [33]

    Reparameterizable subset sampling via continuous relaxations,

    S. M. Xie and S. Ermon, “Reparameterizable subset sampling via continuous relaxations,” 2021. [Online]. Available: https: //arxiv.org/abs/1901.10517

    Pith/arXiv arXiv 2021

  34. [34]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Doll ´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” 2017. [Online]. Available: https://arxiv.org/abs/1612.03144

    Pith/arXiv arXiv 2017

  35. [35]

    Polarformer: Multi-camera 3d object detection with polar transformer,

    Y . Jiang, L. Zhang, Z. Miao, X. Zhu, J. Gao, W. Hu, and Y .-G. Jiang, “Polarformer: Multi-camera 3d object detection with polar transformer,” 2023. [Online]. Available: https://arxiv.org/abs/2206. 15398

    2023

  36. [36]

    Radarsimx: The tool chain for radar simulation,

    AStuff, LLC, “Radarsimx: The tool chain for radar simulation,” https: //radarsimx.com, 2024, accessed: 2025-01-01

    2024

  37. [37]

    Shapenet: An information-rich 3d model repository,

    A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su, J. Xiao, L. Yi, and F. Yu, “Shapenet: An information-rich 3d model repository,” 2015. [Online]. Available: https://arxiv.org/abs/1512.03012

    Pith/arXiv arXiv 2015

  38. [38]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” 2017. [Online]. Available: https://arxiv.org/abs/1412. 6980

    2017

  39. [39]

    Dick and F

    J. Dick and F. Pillichshammer,Digital Nets and Sequences: Dis- crepancy Theory and Quasi-Monte Carlo Integration. Cambridge University Press, 2010

    2010

  40. [2019]

    Available: https://arxiv.org/abs/1812.05784

    [Online]. Available: https://arxiv.org/abs/1812.05784

    Pith/arXiv arXiv

  41. [2023]

    Available: https://arxiv.org/abs/2303.16940

    [Online]. Available: https://arxiv.org/abs/2303.16940

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.