Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

MutualForce: Mutual-Aware Enhancement for 4D Radar-LiDAR 3D Object Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Radar and LiDAR mutually enhance each other to reach 71.76% mAP on the VoD validation set.

desk verdict Plausible incremental radar-LiDAR fusion with clean ablations, but the headline mAP is selected on the same validation split used for reporting, so the lead over InterFusion is not yet robust. read the letter →

arxiv 2501.10266 v2 pith:JYHCFEWE submitted 2025-01-17 cs.CV

Xiangyuan Peng , Huawei Sun , Kay Bierzynski , Anton Fischbacher , Lorenzo Servadei , Robert Wille This is my paper · ORCID
classification cs.CV
keywords 4DradarLiDAR3Dobjectdetectionsensorfusionbird's-eyeviewcross-attentioncontrastivelearningautonomousdriving
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

MutualForce is a 3D object detection framework for autonomous driving built on the idea that radar and LiDAR should improve each other rather than merely be merged. It uses two radar signals — radial velocity and radar cross-section — as indicative features that steer geometric feature extraction in both sensor branches, then lets LiDAR's shape information fill in the sparse, incomplete radar bird's-eye-view features. On the View-of-Delft validation set, the paper reports the best mean average precision among the methods it compares: 71.76% over the full area and 86.36% inside the driving corridor, with car AP rising 4.17 and 4.20 points over the InterFusion baseline. The payoff, if the result transfers, is a fusion recipe that keeps LiDAR-grade structure while carrying radar's velocity and material cues and its all-weather robustness.

What carries the argument

The load-bearing objects are the two fusion modules. IRB computes an indicative weight from radar's velocity/RCS pillars via MLP and sigmoid gating, uses it to gate radar spatial features, and uses the same weight as key and value in a dot-product cross-attention whose query comes from LiDAR pillars; this carries radar's dynamic and material cues into both streams. SALC builds per-class shape heatmaps from LiDAR BEV features with a small CNN plus sigmoid, thresholds them at tau = 0.1 to keep foreground grids, derives instance centers following the CenterPoints convention, and applies the multi-class contrastive loss to pull same-class instances together and push different classes apart; the resulting heatmaps are concatenated with radar BEV features. Together they implement the paper's central mechanism: radar tells both encoders where and what kind of object to look for, and LiDAR tells the radar branch what the object looks like.

What would settle it

A concrete check: retrain MutualForce and InterFusion on the VoD training split, then evaluate on a held-out split (or the official test server if it becomes available) with tau fixed without peeking at test labels. If the car AP gain over InterFusion drops materially or reverses, the validation-set result was at least partly a selection artifact; if it persists, the mutual-enhancement claim is supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that mutual, task-directed interaction between 4D radar and LiDAR beats one-way or concatenated fusion. The proposed framework, MutualForce, has two interacting mechanisms: an Indicative Radar-Driven Bidirectional module, where radar's relative and absolute radial velocity plus RCS are turned into a gating weight that sharpens radar pillar features and, through cross-attention, guides LiDAR pillar features toward dynamic and material-relevant regions; and a Shape Awareness LiDAR-Driven Contrastive module, where LiDAR BEV features are converted into class-wise shape heatmaps, supervised by focal loss and a multi-class contrastive 'push-pull' term, and used to enrich the radar BEV features. The reported result is the highest mAP on the VoD validation set among the compared methods, 71.76% overall and 86.36% in the driving corridor, with the largest gains for cars, attributed to their distinctive RCS signatures and symmetric shapes.

Load-bearing premise

The load-bearing premise is that validation-set performance on VoD is an unbiased measure of superiority, because the same validation split is used to tune the BEV threshold tau and then to report the final mAP; if the test server or an independent split gave different rankings, the reported margin could be inflated.

Editorial extensions

If this is right

  • Fusion design shifts from concatenating two feature streams to bidirectional guidance: radar's velocity and RCS act as an attention signal for LiDAR rather than an extra input channel.
  • Cars, the class with the largest gains (4.17 and 4.20 AP over InterFusion), show that symmetric, high-RCS objects are exactly where mutual enhancement pays off.
  • Small road users do not regress: inside the driving corridor, pedestrian AP rises by 1.99 points and cyclist AP by 2.47 points over InterFusion.
  • The full model reports real-time inference at 14.45 FPS with 20.03M parameters, so the added fusion modules do not prevent deployment.
  • Ablations show that each module earns its place: the baseline without IRB and SALC reaches 59.75% mAP, while the complete model reaches 71.76% mAP.

Reading between the lines

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

  • Beyond the paper's experiments, the same validation split is used both to choose the BEV threshold tau and to report the final mAP, so the reported margin over InterFusion is likely optimistic; an independent test split or nested cross-validation would give the fair comparison.
  • Although radar's weather robustness motivates the work, MutualForce is only evaluated on the clear-weather View-of-Delft set; a natural next test is fog or rain data, where LiDAR shape cues degrade and the mutual-enhancement balance may shift.
  • The indicative-gating plus contrastive-shape recipe does not depend on LiDAR in principle; a camera could supply the dense shape heatmaps, yielding a radar-camera variant with the same architecture.
  • The threshold tau that separates foreground shape grids is fixed at 0.1 for all classes and scenes; an adaptive per-instance or per-scene threshold could remove a validation-set tuning knob that may not transfer to denser or sparser environments.
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. The paper proposes MutualForce, a 4D radar–LiDAR fusion framework for 3D object detection on the View-of-Delft (VoD) dataset. The method contains two main components: an Indicative Radar-Driven Bidirectional module (IRB), which uses radar velocity and RCS features to guide both radar and LiDAR geometric feature extraction, and a Shape Awareness LiDAR-Driven Contrastive module (SALC), which uses LiDAR-derived BEV shape heatmaps and a contrastive loss to enrich radar BEV features. Experiments on the VoD validation set report a mAP of 71.76% in the whole area and 86.36% in the driving corridor, outperforming prior radar–LiDAR fusion methods including InterFusion, with particularly large gains on the car class. The paper also reports real-time inference at 14.45 FPS.

Significance. If the reported gains are robust, the work makes a useful contribution to multi-modal 3D detection: the idea of using radar's velocity/RCS as indicative features to guide both modalities, and using LiDAR shape information to enrich the sparser radar BEV representation, is well motivated and the module-level ablations in Table II are consistent with each component contributing. The method is simple enough to be reproduced from the description, and the VoD comparison targets a practical fusion scenario. However, the headline claim of state-of-the-art performance rests on a single validation-set evaluation with hyperparameters selected on that same split, and one central equation has a dimension mismatch. These issues need to be addressed before the reported superiority can be accepted.

major comments (3)
  1. [§III-A and Table IV] The reported headline results are selected on the same validation split used for evaluation. Section III-A states that the VoD test server is unavailable, so all results are on the validation set. Table IV then sweeps the BEV threshold τ on that validation set and chooses τ = 0.1 because it gives the best mAP. This is model selection on the evaluation target. The impact is not negligible: changing only τ moves the all-area mAP from 68.11 (τ = 0.2) to 71.76 (τ = 0.1), a difference of 3.65 points, while the claimed margin over InterFusion is only 1.93 points. The paper should provide an unbiased estimate, for example by reporting results with a fixed τ chosen by cross-validation, by using a separate held-out split, or by reporting mean and variance over multiple runs. Without this, the 'highest mAP' claim is an optimistic estimate rather than a demonstrated superiority.
  2. [§II-C, Eq. (3)] Equation (3) as written is not dimensionally well-defined. The matrix S is defined as N×M, where N is the number of classes and M is the maximum number of instance centers among all class sets. S′ is obtained by swapping columns of S, so S′ is also N×M. The numerator uses d(S(h,:), S'(:,h)), where S(h,:) is a row of length M and S'(:,h) is a column of length N. These vectors can be element-wise multiplied only if N = M, which is not generally true and is not stated as an assumption. The same issue affects the denominator terms d(S(h,:), S'(:,w)). The authors should correct the indexing (e.g., by using rows of S′ or by defining S and S′ with matching dimensions) and clarify the distance function so that the contrastive term is reproducible.
  3. [§III-B, Table I] All comparisons are reported as single numbers without error bars or multiple-seed statistics, and several baselines are marked as 'our reproduced results' without describing the reproduction protocol. Since the final claim is a numerical superiority of about 1.93 points over the strongest baseline, the absence of variance information makes it impossible to assess whether the difference is meaningful. This is closely related to the selection issue above, but it is also important even for fixed hyperparameters: the paper should report at least two or three training runs, or a statistical test, for the central comparison.
minor comments (6)
  1. [Table II] The header of Table II is ambiguous: the checkmarks appear under the subcolumns 'R-R', 'R-L', and 'SALC', but the first two are branches of IRB. The caption should explicitly spell out which module configuration each row corresponds to.
  2. [§II-B, Eq. (1)] The word 'concanation' should be 'concatenation'.
  3. [Conclusion] The conclusion contains a stray fragment 'the .' that should be removed.
  4. [Abstract and §II-C] The use of 'V oD' (with a space) is inconsistent; use 'VoD' uniformly.
  5. [§III-A] The evaluation protocol does not specify how the threshold τ in SALC is applied during inference, nor how the 'driving corridor' region is defined. Please provide this information for reproducibility.
  6. [§III-B] The paper reports real-time performance at 14.45 FPS in the conclusion, but the experiments section gives no timing methodology (hardware, batch size, measurement procedure). Either add this detail or remove the claim.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline 71.76% mAP is a validation-selection maximum: τ is tuned on the same split used to claim the 'highest mAP', and the 3.65-point sweep range exceeds the 1.93-point lead over InterFusion.

  1. fitted input called prediction [Abstract; Section III-A (Dataset and Metrics); Section III-B (Main Results); Section III-C (Analysis of the BEV threshold), Tables I and IV]
    "Due to the unavailability of the test server, our evaluation is conducted on the validation set. [...] Results with different τ are in Table IV. Our model achieves the best mAP when τ equals 0.1. [...] our method outperforms other approaches with the mAP of 71.76% across the entire area [...] achieving the highest mAP of 71.76%."

    τ is a SALC hyperparameter ('A threshold τ is set to filter out the background grids'), and Section III-C chooses it on the validation split because the model 'achieves the best mAP when τ equals 0.1': Table IV sweeps τ={0.05, 0.1, 0.2} to mAP {68.42, 71.76, 68.11}. Since 'the test server' is 'unavailable' (Section III-A), that same validation set is then used to report the result, so the headline 71.76% is by construction the maximum of the sweep over the evaluation split, not an independent estimate. The sweep range (3.65 points) nearly doubles the claimed 1.93-point lead over InterFusion (71.76 vs 69.83, Table I); choosing τ=0.2 would erase most of the reported advantage. The fitted threshold is thus renamed as a prediction of superiority.

full rationale

The method itself is not circular: IRB (Eqs. 1–2) uses raw radar velocity/RCS pillars to gate radar features and to provide cross-attention keys/values for LiDAR features; SALC builds LiDAR BEV shape heatmaps supervised by Focal Loss and the MCCont loss (Eqs. 3–4) on ground-truth object centers; the final loss (Eq. 5) combines RPN and shape losses. No quantity is defined in terms of its own target, no uniqueness theorem is invoked, and the one overlapping-author citation ([33], MUFASA) is only a comparison baseline, not load-bearing. The circular step is in the evaluation protocol. Section III-A acknowledges the test server is unavailable, so evaluation is on the validation set; Section III-C then selects τ because it 'achieves the best mAP' in the Table IV sweep; and the Abstract/Section III-B report that same selected maximum as 'the highest mAP of 71.76%' and as a 1.93-point improvement over InterFusion. Because the threshold is selected on the same split that is used to declare superiority, the headline number is a maximum of the hyperparameter grid by construction, and the 3.65-point sweep range is nearly twice the claimed 1.93-point advantage. The reported lead is therefore partly produced by the selection procedure rather than by an unbiased evaluation of the architecture — a partial circularity (score 6), not a self-definitional or self-citation collapse. The architectural content remains independently meaningful, which is why the score is not higher.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method's performance depends on domain assumptions about radar and LiDAR signal properties and about the VoD validation set. No new physical entities or constants are introduced.

free parameters (2)
  • BEV shape heatmap threshold τ = 0.1
    Chosen as the value giving the best mAP on the validation set in Table IV.
  • Shape-aware loss weight α = 1.0
    Set manually to 1.0 without a sensitivity analysis.
assumptions (4)
  • domain assumption Radar relative and absolute radial velocity and RCS are indicative of object class and motion state.
    Section II-B and Fig. 2 use these features to guide feature learning; the paper provides histograms as support.
  • domain assumption LiDAR shape information can enrich radar BEV features, especially for symmetric objects like cars.
    Section II-C proposes SALC based on this assumption, supported only by the final detection results.
  • domain assumption The VoD validation split is representative and the annotations are correct, despite being the only evaluation set.
    Section III-A states the test server is unavailable and evaluation is on the validation set.
  • standard math Standard attention, focal loss, and contrastive loss formulations from cited works are valid building blocks.
    Equations (1) to (5) rely on these standard components from [25], [26], [29].

how reviews work

0 comments
Cite this review

Pith. "Pith review of MutualForce: Mutual-Aware Enhancement for 4D Radar-LiDAR 3D Object Detection." pith.science (2026). https://pith.science/paper/JYHCFEWE

@misc{pith2026250110266,
  author       = {Pith},
  title        = {Pith review of: MutualForce: Mutual-Aware Enhancement for 4D Radar-LiDAR 3D Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JYHCFEWE}},
  note         = {Machine review of arXiv:2501.10266}
}
read the original abstract

Radar and LiDAR have been widely used in autonomous driving as LiDAR provides rich structure information, and radar demonstrates high robustness under adverse weather. Recent studies highlight the effectiveness of fusing radar and LiDAR point clouds. However, challenges remain due to the modality misalignment and information loss during feature extractions. To address these issues, we propose a 4D radar-LiDAR framework to mutually enhance their representations. Initially, the indicative features from radar are utilized to guide both radar and LiDAR geometric feature learning. Subsequently, to mitigate their sparsity gap, the shape information from LiDAR is used to enrich radar BEV features. Extensive experiments on the View-of-Delft (VoD) dataset demonstrate our approach's superiority over existing methods, achieving the highest mAP of 71.76% across the entire area and 86.36\% within the driving corridor. Especially for cars, we improve the AP by 4.17% and 4.20% due to the strong indicative features and symmetric shapes.

Figures

Figures reproduced from arXiv: 2501.10266 by the authors.

Figure 1
Figure 1. The overall structure of MutualForce. • Extensive experiments on the VoD dataset [14] demon￾strate the effectiveness of our proposed method. II. PROPOSED METHOD A. Overall Structure In this section, we introduce a fusion framework, Mu￾tualForce, which leverages the advantages of both sensors to mutually enhance their features. Its overall structure is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. The Structure of IRB module. the raw point clouds, radar indicative pillars p c r and spatial pillars p s r have the dimension of (N,P,D) and (N,P,C1) while LiDAR pillars p s l are in dimension (M,P,C2). M and N represent the number of pillars, P denotes the number of points in each pillar. C1 and C2 are the feature channels. D = 3 indicates the three chosen radar indicative features, relative and absolute radial ve… view at source ↗
Figure 2
Figure 2. The distribution of radar indicative features for different [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The structure of SALC module. In the SALC module, LiDAR BEV features B ∈ R H×W×C from PointPillars backbones [25] are fed into a shape￾awareness network composed of three 2D CNNs and one sig￾moid layer. To obtain the global shape distribution, the shape￾awareness netwo…
Figure 5
Figure 5. Figure 5: Visualization of the detection on three methods. Ground [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Continue with ORCID 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. ELMAR: Enhancing LiDAR Detection with 4D Radar Motion Awareness and Cross-modal Uncertainty

    cs.CV 2025-06

Reference graph

Works this paper leans on

37 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [1]

    Scene adaptive sparse transformer for event-based object detection,

    Y . Peng, H. Li, Y . Zhang, X. Sun, and F. Wu, “Scene adaptive sparse transformer for event-based object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 16 794–16 804

    2024

  2. [2]

    CARPE-ID: Continuously Adaptable Re-identification for Personalized Robot Assistance

    F. Rollo, A. Zunino, N. Tsagarakis, E. M. Hoffman, and A. Ajoudani, “Carpe-id: Continuously adaptable re-identification for personalized robot assistance,” arXiv preprint arXiv:2310.19413 , 2023

    work page Pith review arXiv 2023

  3. [3]

    Time to shine: Fine-tuning object detection models with synthetic adverse weather images,

    T. Rothmeier, W. Huber, and A. C. Knoll, “Time to shine: Fine-tuning object detection models with synthetic adverse weather images,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 4447–4456

    work page 2024

  4. [4]

    Revisiting token pruning for object detection and instance segmentation,

    Y . Liu, M. Gehrig, N. Messikommer, M. Cannici, and D. Scara- muzza, “Revisiting token pruning for object detection and instance segmentation,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2024, pp. 2658–2668

    2024

  5. [5]

    Bounding box-guided pseudo point clouds early- fusion and density optimize for 3d object detection,

    S. Zhao and S. Yan, “Bounding box-guided pseudo point clouds early- fusion and density optimize for 3d object detection,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 4205–4209

    work page 2024

  6. [6]

    Svqnet: Sparse voxel-adjacent query network for 4d spatio-temporal lidar se- mantic segmentation,

    X. Chen, S. Xu, X. Zou, T. Cao, D.-Y . Yeung, and L. Fang, “Svqnet: Sparse voxel-adjacent query network for 4d spatio-temporal lidar se- mantic segmentation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 8569–8578

    2023

  7. [7]

    Rethinking normals: Direction guided point cloud recognition,

    K. Liu, Y . Zhu, Z. Wang, K. Wang, and G. Zhou, “Rethinking normals: Direction guided point cloud recognition,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 8110–8114

    work page 2024

  8. [8]

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

    S. Shi, C. Guo, L. Jiang, Z. Wang, J. Shi, X. Wang, and H. Li, “Pv-rcnn: Point-voxel feature set abstraction for 3d object detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 10 529–10 538

    2020

Show all 37 references
  1. [9]

    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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 162–15 172

    2024

  2. [10]

    Fog simulation on real lidar point clouds for 3d object detection in adverse weather,

    M. Hahner, C. Sakaridis, D. Dai, and L. Van Gool, “Fog simulation on real lidar point clouds for 3d object detection in adverse weather,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 15 283–15 292

    2021

  3. [11]

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

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

    2022

  4. [12]

    Centerpoint transformer for bev object detection with automotive radar,

    L. Saini, Y . Su, H. Tercan, and T. Meisen, “Centerpoint transformer for bev object detection with automotive radar,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2024, pp. 4451–4460

    2024

  5. [13]

    Dual radar: A multi-modal dataset with dual 4d radar for autononous driving,

    X. Zhang, L. Wang, J. Chen, C. Fang, L. Yang, Z. Song, G. Yang, Y . Wang, X. Zhang, and J. Li, “Dual radar: A multi-modal dataset with dual 4d radar for autononous driving,” arXiv preprint arXiv:2310.07602, 2023

    2023 arXiv

  6. [14]

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

    A. Palffy, E. Pool, S. Baratam, J. F. Kooij, and D. M. Gavrila, “Multi- class road user detection with 3+ 1d radar in the view-of-delft dataset,” IEEE Robotics and Automation Letters , vol. 7, no. 2, pp. 4961–4968, 2022

    2022

  7. [15]

    Tj4dradset: A 4d radar dataset for autonomous driving,

    L. Zheng, Z. Ma, X. Zhu, B. Tan, S. Li, K. Long, W. Sun, S. Chen, L. Zhang, M. Wan et al., “Tj4dradset: A 4d radar dataset for autonomous driving,” in 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC) . IEEE, 2022, pp. 493–498

    2022

  8. [16]

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

    L. Wang, X. Zhang, B. Xv, J. Zhang, R. Fu, X. Wang, L. Zhu, H. Ren, P. Lu, J. Li et al. , “Interfusion: Interaction-based 4d radar and lidar fusion for 3d object detection,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2022, pp. 12...

    2022

  9. [17]

    See further than cfar: a data-driven radar detector trained by lidar,

    I. Roldan, A. Palffy, J. F. Kooij, D. M. Gavrila, F. Fioranelli, and A. Yarovoy, “See further than cfar: a data-driven radar detector trained by lidar,” in 2024 IEEE Radar Conference (RadarConf24) . IEEE, 2024, pp. 1–6

    2024

  10. [18]

    Traffic object detection for autonomous driving fusing lidar and pseudo 4d-radar under bird’s- eye-view,

    Z. Meng, Y . Song, Y . Zhang, Y . Nan, and Z. Bai, “Traffic object detection for autonomous driving fusing lidar and pseudo 4d-radar under bird’s- eye-view,” IEEE Transactions on Intelligent Transportation Systems , 2024

    2024

  11. [19]

    Lirafusion: Deep adaptive lidar- radar fusion for 3d object detection,

    J. Song, L. Zhao, and K. A. Skinner, “Lirafusion: Deep adaptive lidar- radar fusion for 3d object detection,” arXiv preprint arXiv:2402.11735 , 2024

    2024 arXiv

  12. [20]

    Bi-lrfusion: Bi-directional lidar-radar fusion for 3d dynamic object detection,

    Y . Wang, J. Deng, Y . Li, J. Hu, C. Liu, Y . Zhang, J. Ji, W. Ouyang, and Y . Zhang, “Bi-lrfusion: Bi-directional lidar-radar fusion for 3d dynamic object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 13 394–13 403

    2023

  13. [21]

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

    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 system,” arXiv preprint arXiv:2211.06108 , 2022

    2022 arXiv

  14. [22]

    Robust multimodal vehicle detection in foggy weather using complementary lidar and radar sig- nals,

    K. Qian, S. Zhu, X. Zhang, and L. E. Li, “Robust multimodal vehicle detection in foggy weather using complementary lidar and radar sig- nals,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2021, pp. 444–453

    2021

  15. [23]

    The rcs scaled measurement and inversion method for metal targets with rough surfaces based on the distribution of surface elements in the thz band,

    S. Pang, Y . Zeng, H. Wang, Q. Yang, and B. Deng, “The rcs scaled measurement and inversion method for metal targets with rough surfaces based on the distribution of surface elements in the thz band,” IEEE Transactions on Antennas and Propagation , 2024

    2024

  16. [24]

    Transloc4d: Transformer-based 4d radar place recognition,

    G. Peng, H. Li, Y . Zhao, J. Zhang, Z. Wu, P. Zheng, and D. Wang, “Transloc4d: Transformer-based 4d radar place recognition,” in Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17 595–17 605

    2024

  17. [25]

    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,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 12 697–12 705

    2019

  18. [26]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2980–2988

    2017

  19. [27]

    Bsh-det3d: improving 3d object detection with bev shape heatmap,

    Y . Shen, Y . Zhang, Y . Wu, Z. Wang, L. Yang, S. Coleman, and D. Kerr, “Bsh-det3d: improving 3d object detection with bev shape heatmap,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 5730–5737

    2023

  20. [28]

    Center-based 3d object detection and tracking,

    T. Yin, X. Zhou, and P. Krahenbuhl, “Center-based 3d object detection and tracking,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 11 784–11 793

    2021

  21. [29]

    Coin: Contrastive instance feature mining for outdoor 3d object detection with very limited annotations,

    Q. Xia, J. Deng, C. Wen, H. Wu, S. Shi, X. Li, and C. Wang, “Coin: Contrastive instance feature mining for outdoor 3d object detection with very limited annotations,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 6254–6263

    2023

  22. [30]

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

    O. Team et al. , “Openpcdet: An open-source toolbox for 3d object detection from point clouds,” OD Team, 2020

    2020

  23. [31]

    Mvfan: Multi-view feature assisted network for 4d radar object detection,

    Q. Yan and Y . Wang, “Mvfan: Multi-view feature assisted network for 4d radar object detection,” in International Conference on Neural Information Processing. Springer, 2023, pp. 493–511

    2023

  24. [32]

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

    J. Liu, Q. Zhao, W. Xiong, T. Huang, Q.-L. Han, and B. Zhu, “Smurf: Spatial multi-representation fusion for 3d object detection with 4d imaging radar,” IEEE Transactions on Intelligent V ehicles, vol. 9, no. 1, pp. 799–812, 2024

    2024

  25. [33]

    Mufasa: Multi-view fusion and adaptation network with spatial aware- ness for radar object detection,

    X. Peng, M. Tang, H. Sun, K. Bierzynski, L. Servadei, and R. Wille, “Mufasa: Multi-view fusion and adaptation network with spatial aware- ness for radar object detection,” 2024. [Online]. Available: https://arxiv. org/abs/2408.00565

    2024 arXiv

  26. [34]

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

    Z. Liu, H. Tang, A. Amini, X. Yang, H. Mao, D. L. Rus, and S. Han, “Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) , 2023, pp. 2774–2781

    2023

  27. [35]

    Rcfusion: Fusing 4d radar and camera with bird’s-eye view features for 3d object detection,

    L. Zheng, S. Li, B. Tan, L. Yang, S. Chen, L. Huang, J. Bai, X. Zhu, and Z. Ma, “Rcfusion: Fusing 4d radar and camera with bird’s-eye view features for 3d object detection,” IEEE Transactions on Instrumentation and Measurement, 2023

    2023

  28. [36]

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

    Z. Lin, Z. Liu, Z. Xia, X. Wang, Y . Wang, S. Qi, Y . Dong, N. Dong, L. Zhang, and C. 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 Recognition , 2024, pp. 14 928–14 937

    2024

  29. [37]

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

    W. Xiong, J. Liu, T. Huang, Q.-L. Han, Y . Xia, and B. Zhu, “Lxl: Lidar excluded lean 3d object detection with 4d imaging radar and camera fusion,” IEEE Transactions on Intelligent V ehicles , 2023

    2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.