Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Controllable Radar Simulation with Waveform Parameter Embedding

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that radar cubes can be simulated controllably through a four-parameter waveform embedding, and that the resulting synthetic data matches or beats real radar in 2D detection and semantic segmentation while improving 3D…

desk verdict Solid method with a nice waveform-parameter idea, but the headline downstream claim is confounded by 2-3x training data and needs equal-size controls before it lands. read the letter →

arxiv 2506.03134 v2 pith:T772L5XR submitted 2025-06-03 eess.SP cs.CV

Weiqing Xiao , Hao Huang , Chonghao Zhong , Yujie Lin , Nan Wang , Xiaoxue Chen , Zhaoxi Chen , Saining Zhang
show 4 more authors
Shuocheng Yang Pierre Merriaux Lei Lei Hao Zhao
This is my paper
classification eess.SPcs.CV
keywords radarsimulationrange-azimuth-Dopplercubepointspreadfunctionwaveformparameterembeddingautonomousdrivingsyntheticdataaugmentation2Dobjectdetectionsemanticsegmentation
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

Ctrl-RS is a simulator for the full range-azimuth-Doppler radar cube that makes sensor attributes controllable through four waveform parameters. The paper argues that raw radar point clouds are too sparse and stochastic to simulate well, and that the cube is the right target. Using a reflection environment tensor built from LiDAR, camera, or existing radar data plus a waveform-conditioned network, the simulator produces cubes with per-point errors around 0.27 on two public datasets. The central claim is that training perception models on these synthetic cubes matches or beats training on real cubes for 2D detection and semantic segmentation, and that mixing synthetic and real cubes improves 3D detection. If this holds, radar training data for new sensors and driving scenarios can be generated quickly without proprietary hardware specifications.

What carries the argument

The load-bearing object is the factorized standard 3D reflection signal, $R_{\mathrm{std}}(r_i,d_i,a_i) = S_R(r_i) \, S_D(d_i) \, S_A(a_i)$, with a Gaussian range response, a segmented-linear Doppler profile, and a window-function spectrum azimuth beam. The four fitted parameters become the channels of a constant attribute embedding $A$, which is concatenated with the reflection environment tensor $E$ and processed by WARP-Net, a 3D U-Net trained with an L1 loss plus a scene-reflection-point-weighted loss. The factorization is what makes different radars comparable: the authors observe that the 3D point-spread shapes of different datasets are geometrically similar, so a single network conditioned on four numbers can span many radar attributes and support attribute editing by changing those numbers.

What would settle it

Take a radar whose data was never used in training, place a single strong point reflector in an otherwise empty scene, and record its range-azimuth-Doppler cube. Fit the four waveform parameters to that measured point spread function, then have Ctrl-RS re-simulate the same scene from the same environment tensor. If the simulated cube's per-point error against the real cube stays above the roughly 0.27 level reported in Table 1, or if the measured point spread function cannot be expressed by the separable Gaussian/segmented-linear/window-spectrum product in Eq. (2), the central claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the response of an automotive radar to the environment can be captured by a single separable 3D point spread function whose shape is controlled by four waveform parameters — $\sigma$ for range resolution, $g$ for Doppler broadening, and $R_s$ and $\lambda$ for the azimuth beam's main-lobe width and sidelobe ratio — and that a network conditioned on these parameters can synthesize radar cubes for any of these attributes. The authors show that WARP-Net, trained on a mixed dataset of real, analytically synthesized, and simulator-generated cubes, generates cubes that are quantitatively close to real ones, with per-point errors around 0.27. In downstream experiments, models trained only on Ctrl-RS data reach or exceed real-data performance in 2D detection and semantic segmentation, and combining simulation with real data lifts RADDet 2D AP@0.5 by 7.3, Carrada AP@0.5 by 20.4, Carrada segmentation Dice by 11.6, and 3D detection AP@0.3 by 4.16. The paper also claims the framework supports viewpoint changes, actor removal, and attribute editing, and can simulate from LiDAR or monocular-camera inputs on unseen scenes such as nuScenes.

Load-bearing premise

The entire pipeline assumes that every real radar cube is a weighted sum of one identical, separable 3D point-spread shape, and that all radar-to-radar variation is captured by the four fitted parameters $\sigma$, $g$, $R_s$, and $\lambda$; if real phenomena such as multipath, range-Doppler coupling, or hardware-specific sidelobes break that factorization, the simulator cannot faithfully reproduce the true cube.

Editorial extensions

If this is right

  • Training on Ctrl-RS synthetic cubes alone matches or beats training on real cubes for 2D object detection and semantic segmentation, so simulated radar cubes can substitute for collected data in those tasks.
  • Adding synthetic cubes to real training data yields consistent gains (for example +7.3 AP@0.5 on RADDet 2D detection, +20.4 on Carrada, +11.6 Dice on Carrada segmentation, and +4.16 AP@0.3 on RADDet 3D detection), making the simulator a viable data-augmentation engine.
  • The simulator accepts LiDAR point clouds, monocular camera images, or existing radar data as environmental input, so radar training data can be produced for datasets that never recorded radar cubes.
  • Editing the four waveform parameters changes the simulated cube's resolution and beam shape, enabling controlled studies of how sensor attributes affect perception models.
  • At about 0.04 seconds per cube, the approach is fast enough to generate large synthetic radar datasets, unlike full electromagnetic or ray-tracing pipelines that take seconds per frame.

Reading between the lines

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

  • If the separable four-parameter point-spread function is a general prior for automotive radar, then equipping a new sensor becomes a calibration problem: measure four numbers from a short capture, and the same network could synthesize cubes for that sensor without retraining.
  • The paper's own numbers suggest a sharper test than the ones reported: hold the scene fixed and vary only one of $\sigma$, $g$, $R_s$, $\lambda$ while measuring detection AP, which would isolate which attribute drives the downstream gains.
  • Because the authors note that 3D detection still prefers real data, a plausible next step is extending the embedding with a fifth parameter or a non-separable correction term; the conditioning mechanism should absorb that change directly.
  • The cross-dataset swapping experiments indicate that attribute diversity, not scene duplication, is the active ingredient, which suggests a scaling rule: each additional real radar-cube dataset with different hardware should improve cross-sensor controllability.
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

3 major / 6 minor

Summary. This paper proposes Ctrl-RS, a hybrid radar cube simulator that combines an environment reflection tensor (built from LiDAR, camera, or existing radar data) with a compact waveform-parameter embedding of radar attributes (sigma, g, Rs, lambda) and a U-Net-style generator called WARP-Net. The authors claim that simulated range-azimuth-Doppler cubes can match or surpass real radar data in 2D detection and semantic segmentation and can consistently improve 3D detection when combined with real data. The paper also demonstrates scene-editing operations, novel-trajectory simulation, and cross-sensor environment construction on RADDet, Carrada, and nuScenes. The system is fast (about 0.04 s per cube), the code is released, and the scene-editing functionality is interesting. However, the headline downstream claims are not established by the current experiments because the comparisons are confounded by training-set size, and the simulation-quality metrics in Table 1 are partly circular.

Significance. If the central claim were supported, Ctrl-RS would be a practically valuable radar data engine: it is fast, controllable, supports scene editing, and offers a way to generate radar cubes from modalities that are much easier to obtain than aligned radar hardware specifications. The paper ships a public implementation, which is a strength. That said, the empirical core of the paper is the claim that simulated data 'match or surpass' real data on downstream tasks. This claim is currently undercut by the fact that every sim-only comparison uses twice as many training slices as the real-only baseline and every combined comparison uses three times as many, so the observed gains are exactly what a generic data-quantity effect would produce. The physical factorization in Eqs. (1)-(2) is a plausible modeling assumption, but no independent hardware measurement is provided to validate it, and the Limitations section itself concedes that the attribute representation is an incomplete substitute for real data in fine-distribution tasks such as 3D detection. The contributions are promising, but the present evidence is insufficient for the abstract's unconditional claims.

major comments (3)
  1. [Sec. 4.4, Tables 2-5, Table A3] The headline downstream comparisons are confounded by training-set size. For RADDet, the real-only baseline uses 22,694 RD slices, the sim-only condition uses 45,388 (2x), and the combined condition uses 68,082 (3x); for Carrada the analogous counts are 5,802, 11,604, and 17,406. Because the training schedule is fixed in epochs and the batch size is not adjusted, the sim-only and combined conditions also receive more optimizer updates. The reported gains (e.g., +7.3 AP@0.5 on RADDet, +20.4 AP@0.5 on Carrada, +11.6 Dice on Carrada) cannot therefore be attributed to simulation quality or attribute diversity. The 3D results in Tables 4-5 show exactly the pattern predicted by a data-quantity effect: sim-only is worse than real-only, while the combined condition, which has 3x the data, improves. The comparison with RadSimReal and standard augmentation in Table A5 does not repair this flaw because the Ctrl-RS rows still add more training data than the baselines. Please rerun all downstream experiments with matched numbers of training slices and matched optimizer steps, for example by subsampling the simulated sets or by adding real-data-only baselines at 2x and 3x sizes.
  2. [Sec. 4.2, Table 1, Sec. A.5] The simulation-quality metrics in Table 1 are partly circular. The environment tensor E is constructed by extracting reflection points and intensities, including noise points, from the real radar cube via CFAR, and the same real cube is then used as the ground truth for PPE, PPSE, and FID. Low per-point error therefore partly measures the network's ability to reproduce its own conditioning input rather than its ability to simulate a radar cube from an independent environment description. In addition, the waveform parameters in Sec. 3.3.1 and Sec. A.6 are fitted on the same RADDet and Carrada training sets that are used to train WARP-Net and to generate the downstream simulation data, so there is no external measurement of the attribute embedding. Please validate simulation quality with E obtained from held-out sensor sources (e.g., LiDAR or camera only) and report metrics for scenes whose reflection points were not extracted from the target cube, or clearly reinterpret Table 1 as a reconstruction experiment rather than a fidelity measure.
  3. [Sec. 3.2.1, Eqs. (1)-(2), Sec. 5] The four-parameter separable PSF model in Eqs. (1)-(2) is a strong physical assumption that is load-bearing for the controllability claim. The Gaussian range response, segmented-linear Doppler broadening, and window-spectrum azimuth beam are fitted to the training data, but no independent radar-hardware measurement is provided to show that real cubes are well represented as a weighted superposition of such separable PSFs. Multipath, target extent, range-Doppler coupling, and hardware-specific sidelobes would violate the factorization and break attribute transfer. The Limitations section explicitly says the representation 'limits the perception of more detailed radar attributes' and makes the method 'an incomplete substitute for real data' for tasks with fine data distributions such as 3D detection; this is in tension with the abstract's claim that simulated data can 'match or surpass real radar.' Please add validation on a radar dataset not used for parameter fitting, or a quantitative sensitivity analysis that shows which physical effects the four parameters do and do not capture.
minor comments (6)
  1. [Introduction] There is a typo in 'absense' (should be 'absence'), and the name 'WARP-Net' is typeset inconsistently as 'W ARP-Net' throughout the paper.
  2. [Appendix A.5] The text says 'In Table 5 of the main text' when referring to the metric table; the table in question is Table 1 in the main paper.
  3. [Eq. (A2)] The notation for the window parameter is inconsistent: Eq. (A2) uses p and N, while the main text introduces a peak-ratio parameter lambda in Sec. 3.2.1; please align these definitions and define Rs explicitly in relation to the azimuth spectrum.
  4. [Appendix A.6] The sentence 'We use the method in RadSimReal to produce the synthetic dataset B. The synthetic dataset B is a set of the three fitting functions.' appears truncated or repetitive; please clarify exactly what samples dataset B contains.
  5. [References] Several reference entries end with stray trailing numbers (e.g., [1] ends with '2, 6, 1' and [20] with '2, 6, 1'); the bibliography should be cleaned so that only page or section numbers remain where intended.
  6. [Table 6] The runtime values (e.g., '0.30s') are placed in the same row as detection AP values without a clear separation or header, which makes the table hard to read; please split runtime into a separate column or clearly label it.

Circularity Check

1 steps flagged · score 6.0 of 10

Simulation-fidelity validation is circular: the environment tensor E is extracted from the very ground-truth cube it is scored against, so Table 1 reports reconstruction error, not predictive accuracy. Downstream 'sim beats real' claims are also confounded by 2-3x larger training sets.

  1. self definitional [Sec. 4.2 'Simulation Quality and Efficiency'; Eq. (1) in Sec. 3.1; Eq. (3) in Sec. 3.2.2]
    "To accurately assess the similarity between simulation and real data on the same scene, we extract all reflection points and reflection intensities (both scene and noise) from the real radar cube using the CFAR algorithm. We then simulate the radar cube using WARP-Net."

    The reflection environment tensor E used by WARP-Net (Eq. 3) is built from the CFAR-extracted reflection points and intensities of the same ground-truth cube R_gt (Sec. 3.1), and Eq. (1) defines the cube as a weighted superposition of PSFs at exactly those points. The evaluation in Table 1 then compares R_sim against that same R_gt with PPE, PPSE, and FID. Low error therefore measures how well the pipeline reconstructs a sparse encoding of the target cube, not how well it predicts a radar cube from independent scene information. The waveform parameters {sigma, g, Rs, lambda} are also fitted on the same RADDet/Carrada training cubes (Sec. 3.3.1), so the close match is an expected fit diagnostic rather than an external validation of the factorization in Eq. (2).

full rationale

The central simulation-quality claim in Sec. 4.2 reduces to a reconstruction check: E is derived from R_gt via CFAR and R_sim is scored against R_gt, so the measured PPE/PPSE/FID are not an independent prediction. This is a genuine circular validation, but it does not by itself force the downstream results. The downstream comparisons (Tables 2-5) are confounded: Table A3 shows sim-only and combined configurations train on 2-3x more RD slices (e.g., 45,388 vs 22,694 for RADDet; 11,604 vs 5,802 for Carrada), and the simulated sets are generated on the same real training scenarios using waveform parameters measured from those same training sets, so the 'sim beats real' outcome cannot be attributed to simulation quality. These are experimental-design problems rather than derivational equivalences. No load-bearing self-citation chain is present: the PSF idea is attributed to external work (RadSimReal), and the factorization assumption is stated as a modeled prior, not disguised as a prediction. The paper's own limitations paragraph also concedes the waveform-parameter representation is 'an incomplete substitute for real data' for fine-grained tasks such as 3D detection. Overall, one supporting prediction reduces by construction, giving a partial circularity score of 6.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on a stack of assumptions: the PSF superposition model (Eq. 1), the separable functional fit (Eq. 2), the sufficiency of four fitted parameters, the random-point noise model, and the proxy reflection points from LiDAR, camera, or CFAR. None of these is verified against independent radar hardware measurements; parameters are fitted on the same datasets used to train and evaluate the system. The downstream real-test evaluations partially break the circularity, but the training-set-size confound prevents the headline claim from being established.

free parameters (4)
  • Per-dataset waveform parameters {sigma, g, Rs, lambda} = RADDet: sigma=2.6, N=8, g=0.6, p=0.1 (N, p map to Rs, lambda); Carrada values not reported
    Fit by Eqs. (A1)-(A3) to real radar cubes in RADDet and Carrada training sets; used as WARP-Net conditioning labels and for downstream simulation generation (Sec. 3.3.1, 4.4).
  • Synthetic dataset B parameter grid = sigma in {2.4,2.5,2.6,2.7,2.8}, N in {6,7,8,9,10}, g in {0.5,0.6,0.7}, p in {0.1,0.2,0.3}
    Hand-chosen ranges in Appendix A.6 cover richer radar attributes; these values define the generative prior for synthetic cubes.
  • Noise reflection point distribution = not specified
    Noise is modeled as randomly distributed reflection points in range, azimuth, and Doppler (Sec. 3.1); count, density, and intensity distribution are not reported but they control the noise floor and affect PPE, PPSE, and FID.
  • CFAR detection parameters = not specified
    Scene and noise reflection points are extracted from real radar cubes with CFAR in Sec. 4.2; thresholds and false-alarm settings are not given and directly determine the environment tensor.
assumptions (6)
  • domain assumption Radar cube is a weighted superposition of an identical, shift-invariant 3D point spread function at each reflection point (Eq. 1).
    Adopted from RadSimReal [3] and reused without independent validation; it is the foundation of every simulated cube.
  • ad hoc to paper The 3D PSF is separable as S_R * S_D * S_A, with Gaussian range, segmented-linear Doppler, and window-spectrum azimuth (Eq. 2, A1-A3).
    This factorization is introduced in this paper and is the basis of the four-parameter attribute embedding; no derivation from radar signal processing is given.
  • ad hoc to paper Four parameters sigma, g, Rs, lambda suffice to characterize radar attributes relevant for simulation.
    The paper states more parameters could be more accurate but these are sufficient (Sec. 3.2.1); no independent hardware validation is provided.
  • ad hoc to paper Noise in the radar cube can be represented by randomly distributed reflection points processed by the same PSF.
    Motivated by visual similarity of 1D slices in Fig. 3; not derived from receiver noise statistics or multipath models.
  • domain assumption Reflection points from LiDAR, camera reconstruction, and CFAR on existing radar are adequate proxies for true radar reflection points.
    Used in Sec. 3.1; robustness is tested only indirectly, and on nuScenes only against simulated test sets.
  • domain assumption Cross-dataset standard 3D reflection waveforms are geometrically similar enough to share one functional family.
    Supported in the paper only by visual comparison in Fig. A1; this assumption underlies attribute transfer between RADDet and Carrada.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Controllable Radar Simulation with Waveform Parameter Embedding." pith.science (2026). https://pith.science/paper/T772L5XR

@misc{pith2026250603134,
  author       = {Pith},
  title        = {Pith review of: Controllable Radar Simulation with Waveform Parameter Embedding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T772L5XR}},
  note         = {Machine review of arXiv:2506.03134}
}
read the original abstract

Autonomous driving simulators still lack high-fidelity radar, even though radar is critical for robust perception in adverse weather. A key obstacle is that raw radar point clouds are extremely sparse and stochastic, making it difficult to model; we argue that simulating the full range-azimuth-Doppler cube is a more principled target. Existing radar cube simulators either rely purely on neural generators, which are opaque and offer little control over sensor attributes, or on detailed electromagnetic pipelines, which are slow, require proprietary hardware specifications, and still struggle to capture real-world complexity. We introduce Ctrl-RS, a controllable radar cube simulation framework that combines the strengths of both worlds. First, we build an environment reflection tensor from diverse sensor sources (including LiDAR, monocular cameras, and existing radar). Second, we abstract radar physics into a compact set of waveform parameters that characterize the 3D point spread function, yielding an intuitive embedding of radar attributes such as range resolution, Doppler broadening, and azimuth beam shape. Third, we train a WARP-Net on a large mixed dataset that fuses real, analytically synthesized, and simulator-generated radar cubes to cover a wide distribution of radar attributes. Ctrl-RS supports viewpoint changes, actor removal, and attribute editing. Experiments on RADDet, Carrada, and nuScenes show that our simulated data can match or surpass real radar in 2D detection and semantic segmentation, and consistently boosts performance in 3D detection when combined with real data. The Project is available at https://github.com/zhuxing0/Ctrl-RS.

Figures

Figures reproduced from arXiv: 2506.03134 by the authors.

Figure 1
Figure 1. (a) Ctrl-RS enables controllable and realistic radar simulation by conditioning on customizable radar attributes. It supports [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of the proposed Ctrl-RS. The block diagrams from left to right show the environment simulation process, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of 2D and 1D slices of radar cube from RADDet [ [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualization of the fitting functions and fitting results. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: Comparison of 1D slices of the simulation results for [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Radar simulation visualization based on nuScenes [ [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CRUISE: Cooperative Reconstruction and Editing in V2X Scenarios using Gaussian Splatting

    cs.CV 2025-07 conditional novelty 5.0 of 10

    CRUISE reconstructs real V2X driving scenes as editable Gaussians, then shows that training on its generated data improves 3D detection and tracking on the V2X-Seq benchmark.

Reference graph

Works this paper leans on

63 extracted references · 57 canonical work pages · cited by 1 Pith paper

  1. [1]

    Raysar- 3d sar simulator: Now open source

    Stefan Auer, Richard Bamler, and Peter Reinartz. Raysar- 3d sar simulator: Now open source. In2016 IEEE In- ternational Geoscience and Remote Sensing Symposium (IGARSS), pages 6730–6733. IEEE, 2016. 2, 6, 1

    work page 2016

  2. [2]

    Rctdistill: Cross-modal knowledge distillation framework for radar-camera 3d object detection with temporal fusion

    Geonho Bang, Minjae Seong, Jisong Kim, Geunju Baek, Daye Oh, Junhyung Kim, Junho Koh, and Jun Won Choi. Rctdistill: Cross-modal knowledge distillation framework for radar-camera 3d object detection with temporal fusion. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 25315–25324, 2025. 1

    work page 2025

  3. [3]

    Radsimreal: Bridging the gap between synthetic and real data in radar object detec- tion with simulation

    Oded Bialer and Yuval Haitman. Radsimreal: Bridging the gap between synthetic and real data in radar object detec- tion with simulation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 15407–15416, 2024. 2, 3, 4, 5, 6, 7, 1

    work page 2024

  4. [4]

    Radar fields: Frequency-space neural scene representa- tions for fmcw radar

    David Borts, Erich Liang, Tim Broedermann, Andrea Ra- mazzina, Stefanie Walz, Edoardo Palladin, Jipeng Sun, David Brueggemann, Christos Sakaridis, Luc Van Gool, et al. Radar fields: Frequency-space neural scene representa- tions for fmcw radar. InACM SIGGRAPH 2024 Conference Papers, pages 1–10, 2024. 2

    work page 2024

  5. [5]

    nuscenes: A multi- modal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020. 8, 3, 11

    work page 2020

  6. [6]

    A physically based radar simulator.Journal of Atmospheric and Oceanic Tech- nology, 15(2):593–598, 1998

    Carlo Capsoni and Michele D’Amico. A physically based radar simulator.Journal of Atmospheric and Oceanic Tech- nology, 15(2):593–598, 1998. 2

    work page 1998

  7. [7]

    Doppler-aware lidar-radar fusion for weather-robust 3d detection

    Yujeong Chae, Heejun Park, Hyeonseong Kim, and Kuk-Jin Yoon. Doppler-aware lidar-radar fusion for weather-robust 3d detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 27197–27208, 2025. 2

    work page 2025

  8. [8]

    Racformer: Towards high-quality 3d object detection via query-based radar- camera fusion

    Xiaomeng Chu, Jiajun Deng, Guoliang You, Yifan Duan, Houqiang Li, and Yanyong Zhang. Racformer: Towards high-quality 3d object detection via query-based radar- camera fusion. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 17081–17091, 2025. 1

    work page 2025

Show all 63 references
  1. [9]

    Discrete electro- magnetism with the finite integration technique.Progress In Electromagnetics Research, 32(32):65–87, 2001

    Markus Clemens and Thomas Weiland. Discrete electro- magnetism with the finite integration technique.Progress In Electromagnetics Research, 32(32):65–87, 2001. 2

    2001

  2. [10]

    Gen- erating synthetic short-range fmcw range-doppler maps us- ing generative adversarial networks and deep convolutional autoencoders

    Marcio L Lima De Oliveira and Marco JG Bekooij. Gen- erating synthetic short-range fmcw range-doppler maps us- ing generative adversarial networks and deep convolutional autoencoders. In2020 IEEE Radar Conference (Radar- Conf20), pages 1–6. IEEE, 2020. 2

    2020

  3. [11]

    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 Workshops, pages 102– 103, 2020. 2

    2020

  4. [12]

    Carla: An open urban driv- ing simulator

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Anto- nio Lopez, and Vladlen Koltun. Carla: An open urban driv- ing simulator. InConference on robot learning, pages 1–16. PMLR, 2017. 1

    2017

  5. [13]

    Automotive radar signal processing: Research di- rections and practical challenges.IEEE Journal of Selected Topics in Signal Processing, 15(4):865–878, 2021

    Florian Engels, Philipp Heidenreich, Markus Winterman- tel, Lukas St ¨acker, Muhammed Al Kadi, and Abdelhak M Zoubir. Automotive radar signal processing: Research di- rections and practical challenges.IEEE Journal of Selected Topics in Signal Processing, 15(4):865–878, 2021. 1, 2

    2021

  6. [14]

    A review of cfar detection techniques in radar systems.Microwave Journal, 29:115, 1986

    Alfonso Farina and Francisco A Studer. A review of cfar detection techniques in radar systems.Microwave Journal, 29:115, 1986. 3

    1986

  7. [15]

    Generation of realistic synthetic raw radar data for automated driving applications using genera- tive adversarial networks.arXiv preprint arXiv:2308.02632,

    Eduardo C Fidelis, Fabio Reway, Herick Ribeiro, Pietro L Campos, Werner Huber, Christian Icking, Lester A Faria, and Torsten Sch ¨on. Generation of realistic synthetic raw radar data for automated driving applications using genera- tive adversarial networks.arXiv preprint arXi...

    arXiv

  8. [16]

    Ray-tracing simulator for radar signals propagation in radar networks

    Demetrio Gubelli, Oleg A Krasnov, and Olexander Yarovyi. Ray-tracing simulator for radar signals propagation in radar networks. In2013 European Radar Conference, pages 73–

  9. [17]

    A history of the unity game engine

    John K Haas. A history of the unity game engine. 2014. 1

    2014

  10. [18]

    Doppdrive: Doppler-driven temporal aggregation for improved radar object detection

    Yuval Haitman and Oded Bialer. Doppdrive: Doppler-driven temporal aggregation for improved radar object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 26085–26094, 2025. 2

    2025

  11. [19]

    Channel measure- ment and ray-tracing simulation for 77 ghz automotive radar

    Danping He, Ke Guan, Bo Ai, Zhangdui Zhong, Junhyeong Kim, Heesang Chung, and Andrej Hrovat. Channel measure- ment and ray-tracing simulation for 77 ghz automotive radar. IEEE Transactions on Intelligent Transportation Systems, 24 (7):7746–7756, 2022. 2, 4

    2022

  12. [20]

    A ray launching approach for modeling an fmcw radar system

    Nils Hirsenkorn, Paul Subkowski, Timo Hanke, Alexander Schaermann, Andreas Rauch, Ralph Rasshofer, and Erwin Biebl. A ray launching approach for modeling an fmcw radar system. In2017 18th International Radar Symposium (IRS), pages 1–10. IEEE, 2017. 2, 6, 1

  13. [21]

    Dart: Im- plicit doppler tomography for radar novel view synthesis

    Tianshu Huang, John Miller, Akarsh Prabhakara, Tao Jin, Tarana Laroia, Zico Kolter, and Anthony Rowe. Dart: Im- plicit doppler tomography for radar novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 24118–24129, 2024. 2

    2024

  14. [22]

    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, pages 24655–24665, 2025. 2

    2025

  15. [23]

    V2x-r: Cooperative lidar-4d radar fusion with denoising diffusion for 3d object detection

    Xun Huang, Jinlong Wang, Qiming Xia, Siheng Chen, Bisheng Yang, Xin Li, Cheng Wang, and Chenglu Wen. V2x-r: Cooperative lidar-4d radar fusion with denoising diffusion for 3d object detection. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 27390–...

    2025

  16. [24]

    John Wiley & Sons, 2015

    Jian-Ming Jin.The finite element method in electromagnet- ics. John Wiley & Sons, 2015. 2

    2015

  17. [25]

    4d radar ground truth augmentation with lidar-to-4d radar data synthesis.arXiv preprint arXiv:2503.03637, 2025

    Woo-Jin Jung, Dong-Hee Paek, and Seung-Hyun Kong. 4d radar ground truth augmentation with lidar-to-4d radar data synthesis.arXiv preprint arXiv:2503.03637, 2025. 2

    2025 arXiv

  18. [26]

    Radarsplat: Radar gaussian splatting for high-fidelity data synthesis and 3d reconstruc- tion of autonomous driving scenes

    Pou-Chun Kung, Skanda Harisha, Ram Vasudevan, Aline Eid, and Katherine A Skinner. Radarsplat: Radar gaussian splatting for high-fidelity data synthesis and 3d reconstruc- tion of autonomous driving scenes. InProceedings of the IEEE/CVF International Conference on Computer Visi...

    2025

  19. [27]

    Non-line-of- sight 3d reconstruction with radar

    Haowen Lai, Zitong Lan, and Mingmin Zhao. Non-line-of- sight 3d reconstruction with radar. InThe Thirty-ninth An- nual Conference on Neural Information Processing Systems. 2

  20. [28]

    Sar-nerf: Neural radiance fields for synthetic aperture radar multi-view representation.IEEE Transactions on Geoscience and Remote Sensing, 2024

    Zhengxin Lei, Feng Xu, Jiangtao Wei, Feng Cai, Feng Wang, and Ya-Qiu Jin. Sar-nerf: Neural radiance fields for synthetic aperture radar multi-view representation.IEEE Transactions on Geoscience and Remote Sensing, 2024. 2

    2024

  21. [29]

    Uniscene: Unified occupancy-centric driving scene generation.arXiv preprint arXiv:2412.05435, 2024

    Bohan Li, Jiazhe Guo, Hongsi Liu, Yingshuang Zou, Yikang Ding, Xiwu Chen, Hu Zhu, Feiyang Tan, Chi Zhang, Tiancai Wang, et al. Uniscene: Unified occupancy-centric driving scene generation.arXiv preprint arXiv:2412.05435, 2024. 1

    2024 arXiv

  22. [30]

    John Wiley & Sons, 2008

    Jian Li and Petre Stoica.MIMO radar signal processing. John Wiley & Sons, 2008. 1, 2

    2008

  23. [31]

    Spatial alignment and temporal match- ing adapter for video-radar remote physiological measure- ment

    Qian Liang, Ruixu Geng, Jinbo Chen, Haoyu Wang, Yan Chen, and Yang Hu. Spatial alignment and temporal match- ing adapter for video-radar remote physiological measure- ment. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 8623–8633, 2025. 3

    2025

  24. [32]

    Riccardo: Radar hit prediction and convolution for camera-radar 3d object detection

    Yunfei Long, Abhinav Kumar, Xiaoming Liu, and Daniel Morris. Riccardo: Radar hit prediction and convolution for camera-radar 3d object detection. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 22276–22285, 2025. 3

    2025

  25. [33]

    Comparing methods of measurement.Clini- cal and Experimental Pharmacology and Physiology, 24(2): 198–203, 1997

    John Ludbrook. Comparing methods of measurement.Clini- cal and Experimental Pharmacology and Physiology, 24(2): 198–203, 1997. 5

    1997

  26. [34]

    Rc-autocalib: An end-to-end radar-camera automatic calibration network

    Van-Tin Luu, Yon-Lin Cai, Vu-Hoang Tran, Wei-Chen Chiu, Yi-Ting Chen, and Ching-Chun Huang. Rc-autocalib: An end-to-end radar-camera automatic calibration network. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6700–6709, 2025. 3

    2025

  27. [35]

    Rtmdet: An empirical study of designing real-time object detectors, 2022

    Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors, 2022. 6, 7, 5

    2022

  28. [36]

    Rapid and precise millimeter-wave radar simulation for adas virtual assess- ment

    Takashi Machida and Takashi Owaki. Rapid and precise millimeter-wave radar simulation for adas virtual assess- ment. In2019 IEEE Intelligent Transportation Systems Con- ference (ITSC), pages 431–436. IEEE, 2019. 2

    2019

  29. [37]

    Chapman and Hall/CRC,

    Bassem R Mahafza and Atef Elsherbeni.MATLAB simu- lations for radar systems design. Chapman and Hall/CRC,

  30. [38]

    Graph convolutional networks for 3d object detection on radar data

    Michael Meyer, Georg Kuschk, and Sven Tomforde. Graph convolutional networks for 3d object detection on radar data. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3060–3069, 2021. 2

    2021

  31. [39]

    Multi-view radar semantic segmen- tation

    Arthur Ouaknine, Alasdair Newson, Patrick P ´erez, Florence Tupin, and Julien Rebut. Multi-view radar semantic segmen- tation. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 15671–15680, 2021. 7, 6

    2021

  32. [40]

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

    Arthur Ouaknine, Alasdair Newson, Julien Rebut, Florence Tupin, and Patrick P ´erez. Carrada dataset: Camera and automotive radar with range-angle-doppler annotations. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 5068–5075. IEEE, 2021. 5, 6, 7, 3, 4

    2020

  33. [41]

    Automotive radars: A review of signal processing techniques.IEEE Signal Processing Magazine, 34(2):22–35,

    Sujeet Milind Patole, Murat Torlak, Dan Wang, and Mur- taza Ali. Automotive radars: A review of signal processing techniques.IEEE Signal Processing Magazine, 34(2):22–35,

  34. [42]

    Synthesis of through-wall micro-doppler signatures of human motions using generative adversarial networks.arXiv preprint arXiv:2404.08739, 2024

    Kainat Yasmeen Shobha Sundar Ram. Synthesis of through-wall micro-doppler signatures of human motions using generative adversarial networks.arXiv preprint arXiv:2404.08739, 2024. 2

    2024 arXiv

  35. [43]

    Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159,

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159,

    arXiv

  36. [44]

    A realistic radar ray tracing simulator for large mimo-arrays in automo- tive environments.IEEE Journal of Microwaves, 1(4):962– 974, 2021

    Christian Sch ¨ußler, Marcel Hoffmann, Johanna Br¨aunig, In- grid Ullmann, Randolf Ebelt, and Martin V ossiek. A realistic radar ray tracing simulator for large mimo-arrays in automo- tive environments.IEEE Journal of Microwaves, 1(4):962– 974, 2021. 2, 3, 6, 1

    2021

  37. [45]

    New York: Wiley, 1999

    Mehrdad Soumekh.Synthetic aperture radar signal process- ing. New York: Wiley, 1999. 1, 2

    1999

  38. [46]

    Neurad: Neural rendering for autonomous driving

    Adam Tonderski, Carl Lindstr ¨om, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14895–14904, 2024. 1

    2024

  39. [47]

    Vggt: Vi- sual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Vi- sual geometry grounded transformer. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 5294–5306, 2025. 3

    2025

  40. [48]

    L2r gan: Lidar-to-radar translation

    Leichen Wang, Bastian Goldluecke, and Carsten Anklam. L2r gan: Lidar-to-radar translation. InProceedings of the Asian Conference on Computer Vision, 2020. 2

    2020

  41. [49]

    S3e: Self-supervised state estimation for radar-inertial sys- tem

    Shengpeng Wang, Yulong Xie, Qing Liao, and Wei Wang. S3e: Self-supervised state estimation for radar-inertial sys- tem. InProceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 26686–26695, 2025. 2

    2025

  42. [50]

    Yizhou Wang, Zhongyu Jiang, Yudong Li, Jenq-Neng Hwang, Guanbin Xing, and Hui 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, 15(4):954–967, 2021. 4, 2

    2021

  43. [51]

    Tacodepth: Towards efficient radar-camera depth estima- tion with one-stage fusion

    Yiran Wang, Jiaqi Li, Chaoyi Hong, Ruibo Li, Liusheng Sun, Xiao Song, Zhe Wang, Zhiguo Cao, and Guosheng Lin. Tacodepth: Towards efficient radar-camera depth estima- tion with one-stage fusion. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 1052...

    2025

  44. [52]

    There and back again: Learning to simulate radar data for real- world applications

    Rob Weston, Oiwi Parker Jones, and Ingmar Posner. There and back again: Learning to simulate radar data for real- world applications. In2021 IEEE international conference on robotics and automation (ICRA), pages 12809–12816. IEEE, 2021. 2

    2021

  45. [53]

    Deep stochastic radar models

    Tim A Wheeler, Martin Holder, Hermann Winner, and Mykel J Kochenderfer. Deep stochastic radar models. In 2017 IEEE Intelligent Vehicles Symposium (IV), pages 47–

    2017

  46. [54]

    Tars: Traffic-aware radar scene flow estimation

    Jialong Wu, Marco Braun, Dominic Spata, and Matthias Rottmann. Tars: Traffic-aware radar scene flow estimation. arXiv preprint arXiv:2503.10210, 2025. 2

    2025 arXiv

  47. [55]

    Road boundary detec- tion using 4d mmwave radar for autonomous driving.arXiv preprint arXiv:2503.01930, 2025

    Yuyan Wu and Hae Young Noh. Road boundary detec- tion using 4d mmwave radar for autonomous driving.arXiv preprint arXiv:2503.01930, 2025. 1

    2025 arXiv

  48. [56]

    Street gaussians: Modeling dynamic urban scenes with gaussian splatting

    Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians: Modeling dynamic urban scenes with gaussian splatting. InEuropean Conference on Computer Vision, pages 156–173. Springer, 2024. 1

    2024

  49. [57]

    V2x-radar: A multi-modal dataset with 4d radar for cooper- ative perception.arXiv preprint arXiv:2411.10962, 2024

    Lei Yang, Xinyu Zhang, Jun Li, Chen Wang, Jiaqi Ma, Zhiy- ing Song, Tong Zhao, Ziying Song, Li Wang, Mo Zhou, et al. V2x-radar: A multi-modal dataset with 4d radar for cooper- ative perception.arXiv preprint arXiv:2411.10962, 2024. 2

    2024

  50. [58]

    Numerical solution of initial boundary value prob- lems involving maxwell’s equations in isotropic media.IEEE Transactions on antennas and propagation, 14(3):302–307,

    Kane Yee. Numerical solution of initial boundary value prob- lems involving maxwell’s equations in isotropic media.IEEE Transactions on antennas and propagation, 14(3):302–307,

  51. [59]

    Ray tracing for radio propagation modeling: Principles and applications.IEEE access, 3:1089–1100, 2015

    Zhengqing Yun and Magdy F Iskander. Ray tracing for radio propagation modeling: Principles and applications.IEEE access, 3:1089–1100, 2015. 2, 6, 1

    2015

  52. [60]

    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. IEEE, 2021. 4, 5, 6, 7, 8, 1, 2

    2021

  53. [61]

    Peakconv: Learn- ing peak receptive field for radar semantic segmentation

    Liwen Zhang, Xinyan Zhang, Youcheng Zhang, Yufei Guo, Yuanpei Chen, Xuhui Huang, and Zhe Ma. Peakconv: Learn- ing peak receptive field for radar semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 17577–17586, 2023. 4, 6

    2023

  54. [62]

    Cvfusion: Cross-view fusion of 4d radar and camera for 3d object detection

    Hanzhi Zhong, Zhiyu Xiang, Ruoyu Xu, Jingyun Fu, Peng Xu, Shaohong Wang, Zhihao Yang, Tianyu Pu, and Eryun Liu. Cvfusion: Cross-view fusion of 4d radar and camera for 3d object detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 28188– 2...

    2025

  55. [63]

    Drivinggaussian: Composite gaussian splatting for surrounding dynamic au- tonomous driving scenes

    Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic au- tonomous driving scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.