Pith. sign in

REVIEW 4 major objections 4 minor 41 references

CVFusion: Cross-View Fusion of 4D Radar and Camera for 3D Object Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A two-stage cross-view fusion network called CVFusion claims new state-of-the-art 3D detection with 4D radar and camera, matching LiDAR-class accuracy on public benchmarks.

desk verdict Solid two-stage radar-camera fusion paper with first-rate ablations, but the SOTA margins need matched baselines and released code before they should be trusted. read the letter →

arxiv 2507.04587 v1 pith:BGR4LLQH submitted 2025-07-07 cs.CV

Hanzhi Zhong , Zhiyu Xiang , Ruoyu Xu , Jingyun Fu , Peng Xu , Shaohong Wang , Zhihao Yang , Tianyu Pu
show 1 more author
Eryun Liu
This is my paper · ORCID
classification cs.CV
keywords 4Dradarcamera-radarfusion3Dobjectdetectionbird's-eyeviewcross-viewdeformableattentiontwo-stageautonomousdriving
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

CVFusion argues that one-stage BEV fusion, the standard way to combine 4D radar and camera features, leaves accuracy on the table because monocular depth ambiguity misaligns image and radar features. It proposes a two-stage network instead: the first stage fuses radar and camera features in BEV with a radar-guided iterative mechanism to produce high-recall 3D proposals, and the second stage refines each proposal by attending jointly to radar points, image front-view features, and BEV features. On the VoD and TJ4DRadSet benchmarks the method reports 65.41% and 40.00% mAP, surpassing previous radar-camera fusion methods by 9.10 and 3.68 points, and the authors state it is the first radar-camera fusion model whose accuracy is comparable to a classical LiDAR-input network. If that claim is right, autonomous driving perception could rely on inexpensive radar plus camera, which works in bad weather, instead of expensive LiDAR.

What carries the argument

The load-bearing pieces are the RGIter BEV fusion module, the Cross-Modality Deformable Attention (CMDA) block, and the two stage-2 branches, Point-Guided Fusion (PGF) and Grid-Guided Fusion (GGF). RGIter multiplies camera BEV features by radar-derived occupancy weight maps and iterates this gating across scales before concatenating radar and camera features, so that radar's 3D priors correct image depth ambiguity and raise proposal recall. CMDA projects query points, either radar points or regular grid centers, onto another view and samples features there with learned offsets and attention weights; PGF uses radar points as queries into the image, while GGF uses grid centers as queries and fuses image front-view then BEV features sequentially, covering proposals that contain too few radar points. A self-attention layer then merges the point-guided and grid-guided features before the refinement head predicts confidence and box offsets.

What would settle it

Rerun the TJ4DRadSet evaluation and check whether Car AP again falls from 54.08% (stage-1-only) to 51.54% (full model), as in Table 2; a reproducible drop in the dominant class would directly contradict the claim that the second-stage cross-view fusion improves detection across classes.

Watch

Extended reading notes

Core claim

The paper's central claim is that cross-view, two-stage fusion is the right architecture for 4D radar-camera 3D object detection. Rather than fusing radar and camera once in a shared BEV space and predicting boxes in a single pass, CVFusion uses radar-guided iterative BEV fusion to generate high-recall proposals and then refines each proposal with instance-level features gathered from radar points, the image front view, and the BEV map. The authors argue that this stepwise refinement corrects the feature misalignment that comes from the camera's depth ambiguity, and they report new state-of-the-art results on View-of-Delft and TJ4DRadSet, with the full model beating the previous best radar-camera method by 9.10 and 3.68 mAP points and matching or exceeding a classical LiDAR-input PointPillars detector on most metrics.

Load-bearing premise

The load-bearing premise is that radar points, after projection onto the camera image, land on or near the objects they belong to, so the attention mechanism can sample useful image features; radar's known multipath noise and azimuth error could break that alignment.

Editorial extensions

If this is right

  • Two-stage proposal refinement is a workable design for 4D radar-camera detection and can outperform one-stage radar-camera fusion methods on the same public benchmarks.
  • Radar-guided occupancy gating in stage 1 raises proposal recall from 72.0% to 74.9%, indicating that radar position priors help correct monocular depth ambiguity in image-to-BEV projection.
  • The grid-guided branch improves proposals that contain few or no radar points, so the full model keeps gaining on pedestrians and cyclists where point-only refinement stalls.
  • According to the reported metrics, a radar-camera system reaches accuracy comparable to a classical LiDAR detector, which the authors take as evidence that LiDAR need not be indispensable for 3D perception.
  • Ablations attribute the gains to the individual modules: the second stage adds 5.7 mAP over the first stage, and the stage-1-only model already beats several complete one-stage baselines.

Reading between the lines

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

  • Beyond the paper: LiDAR parity is demonstrated on two urban datasets at moderate ranges; highway speeds, dense traffic, and sustained adverse-weather operation are untested, so the generalization claim remains open.
  • Beyond the paper: because CMDA depends on projecting radar points and grid centers onto the image, accuracy likely depends on calibration quality; an ablation that perturbs the projection matrix would quantify that sensitivity and is not reported.
  • Beyond the paper: the same two-stage point-grid-image-BEV refinement recipe could transfer to other sparse 3D sensors, such as low-cost sparse LiDAR, where proposals often contain very few points and image semantics must fill the gaps.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes CVFusion, a two-stage 4D radar-camera fusion detector for 3D object detection. Stage 1 uses a radar-guided iterative BEV fusion (RGIter) module to generate high-recall 3D proposals; stage 2 refines each proposal with point-guided fusion (PGF) and grid-guided fusion (GGF), which aggregate radar point, image front-view, and BEV features through cross-modality deformable attention. Experiments on the VoD and TJ4DRadSet benchmarks report mAP gains over prior state of the art of +9.10 and +3.68 points, respectively, and the authors claim this is the first radar-camera fusion model to achieve performance comparable to a classical LiDAR-input network. The paper also provides module ablations, backbone ablations, and qualitative visualizations.

Significance. CVFusion's two-stage cross-view design is well motivated, and the ablations consistently attribute positive contributions to the RGIter, PGF, and GGF modules; the backbone-swap experiment in Table 7 is a useful control. If the reported margins survive a matched-protocol evaluation, CVFusion would be a strong new state of the art and the first demonstration that 4D radar-camera fusion can approach LiDAR-level detection, which is significant for adverse-weather autonomy. The main weakness is that the headline numbers rest on single-run results and partly uncorroborated baselines, and the per-class Car regression on TJ4DRadSet needs explanation.

major comments (4)
  1. [Sec. 4.2, Tables 1-2] The central SOTA claim is not backed by a fully matched comparison. LXL, RCFusion, and RCBEVDet are cited from their original publications rather than retrained under the same detection range, voxel size, augmentation, training epochs, and post-processing thresholds used for CVFusion (Sec. 4.1.2-4.1.3). Table 7 only swaps the image backbone for a few competitors and does not equalize the complete pipeline. Because the reported +9.10% mAP margin on VoD is the paper's headline, a matched-protocol re-run or a detailed side-by-side configuration table is required before the margin can be attributed to the method rather than to protocol differences. Please also report results over multiple seeds or at least provide error bars.
  2. [Sec. 4.2.2, Table 2] The full model's Car AP on TJ4DRadSet drops from 54.08 (Stage1-only) to 51.54, while Ped, Cyc, and Tru improve; this regression is not discussed anywhere in Sec. 4.3. Since the paper motivates Stage2 as a refinement stage that improves proposals, a class-specific regression on the most important class is a load-bearing anomaly. It may indicate that the point-guided branch (Eqs. 6-7) suffers from radar-to-image projection noise for cars, or that the refinement head is dominated by other classes. Please provide an analysis of this failure case and report Stage1 vs. full per-class and per-distance results on both datasets.
  3. [Sec. 4.2.2] The claim that CVFusion is 'the first time that a radar-camera fusion model can achieve performance comparable to that of a classical LiDAR-input network' is overstated. On TJ4DRadSet, the LiDAR-input PointPillars achieves 43.27 3D mAP while CVFusion achieves 40.00 (Table 2); on VoD CVFusion does beat PointPillars-L on mAP (Table 1), but PointPillars is a relatively simple LiDAR baseline. Please either temper the claim to 'on the VoD benchmark' or compare against a stronger LiDAR detector before making the general statement.
  4. [Sec. 4.1.3 and code availability] The paper states 'Our code will be made publicly available' and gives a repository link, but no code, checkpoints, or evaluation scripts are currently available, and all results appear to be single-run without variance estimates. For a paper whose central contribution is an empirical SOTA claim, this limits reproducibility. Please release at least the evaluation code and training configurations, and report results from multiple runs (or state the random seed and provide error bars).
minor comments (4)
  1. [Fig. 3] The caption contains the stray string '2024. 8.15', which appears to be a leftover artifact; please remove it.
  2. [Eq. (10)] The notation in Eq. (10) is unclear: the transpose operation makes the right-hand side a scalar rather than a 2D coordinate, and the symbols X_size and Y_size are not defined precisely. Please rewrite the projection as an element-wise operation and define all variables.
  3. [Throughout] The dataset name is inconsistently typeset as 'V oD'; please use 'VoD' uniformly.
  4. [Sec. 3.4] The word 'groudtruth' is a typo for 'ground-truth'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CVFusion is an empirical architecture paper whose claims rest on external public benchmarks and component ablations.

full rationale

The paper contains no derivation chain in which a predicted quantity is defined in terms of the same fitted quantity. Its central claims are empirical: CVFusion is evaluated on the public View-of-Delft and TJ4DRadSet benchmarks against previously published methods and against its own retrained baselines, with results reported in Tables 1 and 2. The ablations in Tables 3 through 6 independently remove and add components (RGIter, PGF, GGF, KDE, CMDA, self-attention) and measure the effect on mAP on the same validation set; none of these components is fit to the evaluation metric, and the improvements are not forced by construction. The paper's claim of being "the first time that a radar-camera fusion model can achieve performance comparable to that of a classical LiDAR-input network" is a comparison to an external LiDAR baseline (PointPillars), not a restatement of an input assumption. References to prior work such as Deformable DETR, PV-RCNN, LoGoNet, LXL, and RCBEVDet are used as ordinary building blocks or comparison points, not as load-bearing self-citations; the authors do not rely on any uniqueness theorem or prior result by the same authors to justify their design. Concerns that the reported margins may be sensitive to baseline matching, single-run evaluation, or lack of released code are legitimate reproducibility and robustness issues, but they are not circularity. No step in the paper reduces to its own input by definition, so the appropriate circularity score is 0.

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

Central claim rests on benchmark evaluations and the effectiveness of two new fusion modules; the ledger entries are the hand-chosen design parameters and assumed priors inherited from LiDAR-based perception.

free parameters (4)
  • Stage2 RoI grid number U = 6
    The proposal is divided into 6x6x6 grids (Sec. 3.3). This is a hand-chosen resolution controlling feature granularity; no ablation over U is reported.
  • Voxel size for radar point cloud = 0.05m x 0.05m x 0.1m
    Set in Sec. 4.1.2. Very fine voxels inherited from LiDAR backbones; no radar-specific ablation is shown, and it directly affects point feature density and training cost.
  • Post-processing thresholds = confidence 0.1, NMS 0.01
    Given in Sec. 4.1.3; chosen by hand and not ablated, these thresholds affect final precision/recall and hence reported mAP.
  • KDE bandwidth = not reported
    KDE is introduced in PGF (Sec. 3.3.1) following [7], but the bandwidth parameter is not specified, making the exact implementation ambiguous.
assumptions (4)
  • domain assumption Depth-based view transformation from monocular image to BEV is a sufficient approximation for aligning image features with radar BEV features.
    Used in Sec. 3.2.1, 'inspired by [24,32]'. If the depth distribution is poorly estimated, the camera BEV feature is misaligned and RGIter weighting cannot fully correct it.
  • domain assumption Radar points inside a proposal box, after projection to the image, provide queries that land on the correct object locations.
    Core assumption of PGF (Eq. 6-7). Radar noise and azimuth errors, acknowledged in Sec. 1, can break this assumption.
  • domain assumption The public annotations, splits, and evaluation protocols of VoD and TJ4DRadSet are correct and comparable across methods.
    All experimental claims rely on these external benchmarks (Sec. 4.1.1).
  • standard math SECOND and PV-RCNN training recipes (losses, augmentations, proposal assignment) transfer to 4D radar point clouds.
    The backbone and two-stage losses are taken from LiDAR literature [26,35] without modification (Sec. 3.2.1, 3.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of CVFusion: Cross-View Fusion of 4D Radar and Camera for 3D Object Detection." pith.science (2026). https://pith.science/paper/BGR4LLQH

@misc{pith2026250704587,
  author       = {Pith},
  title        = {Pith review of: CVFusion: Cross-View Fusion of 4D Radar and Camera for 3D Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BGR4LLQH}},
  note         = {Machine review of arXiv:2507.04587}
}
read the original abstract

4D radar has received significant attention in autonomous driving thanks to its robustness under adverse weathers. Due to the sparse points and noisy measurements of the 4D radar, most of the research finish the 3D object detection task by integrating images from camera and perform modality fusion in BEV space. However, the potential of the radar and the fusion mechanism is still largely unexplored, hindering the performance improvement. In this study, we propose a cross-view two-stage fusion network called CVFusion. In the first stage, we design a radar guided iterative (RGIter) BEV fusion module to generate high-recall 3D proposal boxes. In the second stage, we aggregate features from multiple heterogeneous views including points, image, and BEV for each proposal. These comprehensive instance level features greatly help refine the proposals and generate high-quality predictions. Extensive experiments on public datasets show that our method outperforms the previous state-of-the-art methods by a large margin, with 9.10% and 3.68% mAP improvements on View-of-Delft (VoD) and TJ4DRadSet, respectively. Our code will be made publicly available.

Figures

Figures reproduced from arXiv: 2507.04587 by the authors.

Figure 1
Figure 1. Comparison between the current method and our two [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of CVFusion. The first stage is responsible for preliminary feature fusion and proposal generation, while the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. The architecture of the Cross Modality Deformable At [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Qualitative comparison of the PGF-only and the PGF+GGF (full stage2) detector. The scene image embedded with radar points [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 32 canonical work pages

  1. [1]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. In European confer- ence on computer vision, pages 213–229. Springer, 2020. 3

    2020

  2. [2]

    Futr3d: A unified sensor fusion framework for 3d detection

    Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. Inproceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 172–181,

    work page

  3. [3]

    V oxel r-cnn: Towards high performance voxel-based 3d object detection

    Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. V oxel r-cnn: Towards high performance voxel-based 3d object detection. In Pro- ceedings of the AAAI conference on artificial intelligence , pages 1201–1209, 2021. 2

    work page 2021

  4. [4]

    DPFT: Dual Perspective Fusion Transformer for Camera-Radar-based Object Detection

    Felix Fent, Andras Palffy, and Holger Caesar. Dpft: Dual perspective fusion transformer for camera-radar-based object detection. arXiv preprint arXiv:2404.03015, 2024. 2

    work page Pith review arXiv 2024

  5. [5]

    Simple-bev: What really mat- ters for multi-sensor bev perception? In 2023 IEEE Inter- national Conference on Robotics and Automation (ICRA) , pages 2759–2765

    Adam W Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. Simple-bev: What really mat- ters for multi-sensor bev perception? In 2023 IEEE Inter- national Conference on Robotics and Automation (ICRA) , pages 2759–2765. IEEE, 2023. 2

    work page 2023

  6. [6]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 8

    work page 2016

  7. [7]

    Point density-aware voxels for lidar 3d object detection

    Jordan SK Hu, Tianshu Kuai, and Steven L Waslander. Point density-aware voxels for lidar 3d object detection. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8469–8478, 2022. 2, 5

    work page 2022

  8. [8]

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

    Youngseok Kim, Jun Won Choi, and Dongsuk Kum. Grif net: Gated region of interest fusion network for robust 3d object detection from radar point cloud and monocular im- age. In 2020 IEEE/RSJ International Conference on Intelli- gent Robots and Systems (IROS), pages 10857–10864. IEEE,

    work page 2020

Show all 41 references
  1. [9]

    Craft: Camera-radar 3d object detection with spatio-contextual fusion transformer

    Youngseok Kim, Sanmin Kim, Jun Won Choi, and Dong- suk Kum. Craft: Camera-radar 3d object detection with spatio-contextual fusion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence , pages 1160– 1168, 2023. 3

    2023

  2. [10]

    Crn: Camera radar net for accurate, robust, efficient 3d perception

    Youngseok Kim, Juyeb Shin, Sanmin Kim, In-Jae Lee, Jun Won Choi, and Dongsuk Kum. Crn: Camera radar net for accurate, robust, efficient 3d perception. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 17615–17626, 2023. 2

    2023

  3. [11]

    Rtnh+: Enhanced 4d radar object detection network using combined cfar-based two-level preprocessing and vertical encoding

    Seung-Hyun Kong, Dong-Hee Paek, and Sangjae Cho. Rtnh+: Enhanced 4d radar object detection network using combined cfar-based two-level preprocessing and vertical encoding. arXiv preprint arXiv:2310.17659, 2023. 2

    2023 arXiv

  4. [12]

    Pointpillars: Fast encoders for object detection from point clouds

    Alex H Lang, Sourabh V ora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697–12705, 2019. 2, 6

    2019

  5. [13]

    Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion

    Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern R...

    2023

  6. [14]

    Rcbevdet: Radar-camera fusion in bird’s eye view for 3d object detection

    Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yong- tao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird’s eye view for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti...

    2024

  7. [15]

    Smurf: Spatial multi- representation fusion for 3d object detection with 4d imaging radar

    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, 2023. 2, 6

    2023

  8. [16]

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

    Yang Liu, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Echoes Beyond Points: Unleashing the Power of Raw Radar Data in Multi-modality Fusion. In NeurIPS, 2023. 2

    2023

  9. [17]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3, 6, 8

    2021

  10. [18]

    Bevfusion: Multi- task multi-sensor fusion with unified bird’s-eye view repre- sentation

    Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi- task multi-sensor fusion with unified bird’s-eye view repre- sentation. In 2023 IEEE international conference on robotics and automation (ICRA) , pages 2774–2781. ...

    2023

  11. [19]

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

    Ramin Nabati and Hairong Qi. Centerfusion: Center-based radar and camera fusion for 3d object detection. In Proceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527–1536, 2021. 2

    2021

  12. [20]

    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. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Bench- marks Track, 2022. 2

    2022

  13. [21]

    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

    Andras Palffy, Ewoud Pool, Srimannarayana Baratam, Ju- lian FP 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. 5

    2022

  14. [22]

    Transcar: Transformer-based camera-and-radar fusion for 3d object detection

    Su Pang, Daniel Morris, and Hayder Radha. Transcar: Transformer-based camera-and-radar fusion for 3d object detection. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10902–10909. IEEE, 2023. 3

    2023

  15. [23]

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

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660,

  16. [24]

    Categorical depth distribution network for monocular 3d object detection

    Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555–8564, 2021. 3

    2021

  17. [25]

    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,

  18. [26]

    Pv-rcnn: Point- voxel feature set abstraction for 3d object detection

    Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point- voxel feature set abstraction for 3d object detection. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 2, 5, 6

    2020

  19. [27]

    Smiformer: Learning spatial fea- ture representation for 3d object detection from 4d imaging radar via multi-view interactive transformers

    Weigang Shi, Ziming Zhu, Kezhi Zhang, Huanlei Chen, Zhuoping Yu, and Yu Zhu. Smiformer: Learning spatial fea- ture representation for 3d object detection from 4d imaging radar via multi-view interactive transformers. Sensors, 23 (23):9429, 2023. 2, 6

    2023

  20. [28]

    3-d object detection for multiframe 4-d automotive millimeter-wave radar point cloud

    Bin Tan, Zhixiong Ma, Xichan Zhu, Sen Li, Lianqing Zheng, Sihan Chen, Libo Huang, and Jie Bai. 3-d object detection for multiframe 4-d automotive millimeter-wave radar point cloud. IEEE Sensors Journal, 23(11):11125–11138, 2022. 2

    2022

  21. [29]

    Openpcdet: An open- source toolbox for 3d object detection from point clouds

    OpenPCDet Development Team. Openpcdet: An open- source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet ,

  22. [30]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5

    2017

  23. [31]

    Mvfusion: Multi-view 3d object detection with semantic-aligned radar and camera fusion

    Zizhang Wu, Guilian Chen, Yuanzhu Gan, Lei Wang, and Jian Pu. Mvfusion: Multi-view 3d object detection with semantic-aligned radar and camera fusion. In 2023 IEEE In- ternational Conference on Robotics and Automation (ICRA), pages 2766–2773. IEEE, 2023. 2

    2023

  24. [32]

    Lxl: Lidar excluded lean 3d object detection with 4d imaging radar and camera fusion

    Weiyi Xiong, Jianan Liu, Tao Huang, Qing-Long Han, Yux- uan Xia, and Bing Zhu. Lxl: Lidar excluded lean 3d object detection with 4d imaging radar and camera fusion. IEEE Transactions on Intelligent Vehicles, 2023. 1, 2, 3, 6

    2023

  25. [33]

    Rpfa-net: A 4d radar pillar feature attention network for 3d object detection

    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 detection. In 2021 IEEE International Intelligent Transportation Sys- tems Conference (ITSC), pages 3061–3066. IE...

    2021

  26. [34]

    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. In International Conference on Neural Information Processing , pages 493–

  27. [35]

    Second: Sparsely embed- ded convolutional detection

    Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embed- ded convolutional detection. Sensors, 18(10):3337, 2018. 2, 3, 6

    2018

  28. [36]

    Exploring radar data representations in au- tonomous driving: A comprehensive review, 2024

    Shanliang Yao, Runwei Guan, Zitian Peng, Chenhang Xu, Yilu Shi, Weiping Ding, Eng Gee Lim, Yong Yue, Hyungjoon Seo, Ka Lok Man, Jieming Ma, Xiaohui Zhu, and Yutao Yue. Exploring radar data representations in au- tonomous driving: A comprehensive review, 2024. 2

    2024

  29. [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. In 2022 IEEE 25th International Con- ference on Intelligent Transportation Systems ...

    2022

  30. [38]

    Rcfu- sion: Fusing 4d radar and camera with bird’s-eye view fea- tures for 3d object detection

    Lianqing Zheng, Sen Li, Bin Tan, Long Yang, Sihan Chen, Libo Huang, Jie Bai, Xichan Zhu, and Zhixiong Ma. Rcfu- sion: Fusing 4d radar and camera with bird’s-eye view fea- tures for 3d object detection. IEEE Transactions on Instru- mentation and Measurement, 2023. 1, 2, 3, 6, 8

    2023

  31. [39]

    Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection

    Taohua Zhou, Junjie Chen, Yining Shi, Kun Jiang, Meng- meng Yang, and Diange Yang. Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection. IEEE Transactions on Intelligent Vehicles, 8(2):1523–1535, 2023. 2

    2023

  32. [40]

    Deformable detr: Deformable trans- formers for end-to-end object detection

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable trans- formers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4

    2010 arXiv

  33. [511]

    Springer, 2023. 2, 6

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.