Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Evaluation of an Uncertainty-Aware Late Fusion Algorithm for Multi-Source Bird's Eye View Detections Under Controlled Noise

T0 review · 4 major / 5 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 This paper claims that a Kalman-filter-based late-fusion pipeline, UniKF, beats existing late-fusion baselines for multi-source bird's-eye-view detections under all tested noise levels, with up to 3x lower errors and near-perfect…

desk verdict UniKF is a legitimate incremental late-fusion method with a useful controlled-noise benchmark, but the near-perfect precision/recall and '3x' claims are artifacts of the evaluation protocol. read the letter →

arxiv 2507.03381 v1 pith:Z6PGHT3B submitted 2025-07-04 cs.RO eess.SP

Maryem Fadili (VeDeCom , IRSEEM) , Louis Lecrosnier (IRSEEM) , Steve Pechberti (VeDeCom) , Redouane Khemmar (IRSEEM) This is my paper
classification cs.ROeess.SP
keywords latefusionbird'seyeviewKalmanfilteringuncertainty-awarecollaborativeperceptionmulti-sensorcontrollednoiseevaluationobjectdetection
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

Fusion quality is usually measured together with detection quality, which makes it hard to know whether a fusion algorithm is actually helping. This paper argues that the two can be separated by injecting controlled Gaussian noise into ground-truth bounding boxes, and it uses that protocol to evaluate a late-fusion method it calls UniKF. UniKF merges bird's-eye-view detections from multiple sources with a Kalman filter that carries each box's reported uncertainty and handles synchronous, asynchronous, and delayed measurements by advancing or rolling back the filter in time. Paired with the CSBA association step, it reports consistently lower translation, orientation, and dimension errors than standard NMS, weighted box fusion, and distance-based baselines across three noise levels, with precision and recall near 100%. The stakes are practical: late fusion is the bandwidth-efficient, detector-agnostic option for collaborative driving, so a fusion method that works under controlled uncertainty and a protocol that measures it in isolation are both directly useful.

What carries the argument

The carrying mechanism is a constant-velocity Kalman filter with state $(x, y, v_x, v_y, w, d, \theta)$, updated by five-dimensional BEV measurements $(x, y, w, d, \theta)$ whose covariance is set from each source's reported noise level. To keep time consistent, UniKF distinguishes synchronous measurements, out-of-sequence measurements, and asynchronous measurements: it updates directly in the first case, rolls back to a stored history state before updating in the second, and propagates forward to the measurement time before updating in the third. The CSBA association step scores candidate cross-source pairs by combining center alignment, dimension similarity, and orientation consistency with uncertainty information, and the paper compares UniKF with ground-truth association to show that CSBA contributes most of the association quality.

What would settle it

Feed real detector outputs, such as actual noisy bounding boxes from a 3D object detector on nuScenes or a two-sensor setup, into the same CSBA+UniKF pipeline instead of synthetically perturbed ground truth, and compare mATE, mADE, mAOE, precision, and recall against the reported numbers. If the near-perfect precision and the margin over baselines shrink or disappear, the claim that UniKF maintains its performance under realistic uncertainty is not supported.

Watch

Extended reading notes

Core claim

The central claim is that uncertainty-aware, time-aware Bayesian filtering at the object level is a better late-fusion strategy than heuristic merging, and that this advantage can be shown cleanly once detector noise is controlled. On the nuScenes validation split, with noise injected into ground-truth boxes using distance-dependent standard deviations and multiplicative size noise, CSBA+UniKF achieves up to 3x lower positioning and orientation errors and 2x lower dimension errors than the compared baselines, while precision stays between 99.5% and 100% and recall is 100%. The paper also claims that its own evaluation metrics, which penalize false positives and false negatives instead of scoring only true positives, reveal that some baselines like NMS-STD are not really fusing detections at all and merely retain everything they receive.

Load-bearing premise

The protocol assumes that zero-mean Gaussian noise with hand-set, distance-dependent standard deviations faithfully represents the statistical structure of real detector errors; if real errors are biased, correlated, or otherwise non-Gaussian, the measured fusion gains may not transfer to actual systems.

Editorial extensions

If this is right

  • Late-fusion systems can replace fixed IoU or distance thresholds with a filter that needs no per-category tuning, since UniKF associates and fuses boxes using their reported uncertainties.
  • Because the pipeline works on box-level data only, it can combine heterogeneous sensors or agents without sharing raw sensor streams or proprietary model internals.
  • The proposed metrics, which count false positives and negatives, change how fusion methods are ranked: a method that merely keeps all detections no longer looks competitive.
  • Under the tested synthetic noise model, collaborative fusion with UniKF beats any single detector at every noise level, with the largest gains in the highest-noise configuration.

Reading between the lines

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

  • A natural next test is to replace the synthetic Gaussian perturbations with real detector outputs from the same scenes; the paper's own discussion notes that systematic calibration offsets and clock drift are not modeled, so real-world gains could be smaller.
  • Because the Kalman filter remains the best linear unbiased estimator when the update covariance matches the true second moment of the noise, UniKF's advantage should survive moderate distributional shifts but may weaken under heavy-tailed or multimodal errors.
  • The near-perfect precision and recall partly follow from the evaluation design, which assumes correct classification and uses ground-truth IDs for matching; a live system with classification confusion would likely see lower numbers.
  • The same stored state history used for out-of-sequence updates could support tracking evaluation with little extra work, but the paper explicitly defers tracking metrics, so that benefit is not yet demonstrated.
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 / 5 minor

Summary. The paper proposes UniKF, a Kalman-filter-based late-fusion pipeline for BEV detections that explicitly handles synchronous, out-of-sequence, and asynchronous measurements, combined with the CSBA association method from prior work. It also introduces a controlled-noise evaluation protocol in which zero-mean Gaussian perturbations are added to nuScenes ground-truth boxes in order to isolate fusion performance from upstream detection errors. Experiments compare CSBA+UniKF with NMS-based, IoU-based, distance-based, and WLS baselines under three noise levels, reporting lower translation, orientation, and dimension errors together with near-perfect precision and recall.

Significance. If the results are taken at face value, the paper provides a useful modular late-fusion baseline and a clean synthetic benchmark for isolating fusion algorithms from detector errors. The Kalman formulation is standard and reproducible, and the comparison includes several relevant late-fusion baselines. However, the current evaluation does not substantiate the strongest robustness claims: the synthetic protocol contains no missed detections or spurious detections, the reported precision and recall are therefore partly artifacts of the input construction, and the advantage over the closest baseline CSBA+WLS is marginal. With additional stress tests and revised claims, the framework could be a solid contribution; as written, the experimental support for the headline claims is incomplete.

major comments (4)
  1. [Section III.B, Section III.C, Table II, Section VI.G] The controlled-noise protocol never injects false positives or false negatives: every ground-truth box receives a noisy detection at every frame, and step 3 of Section III.C matches predictions to ground truth by object identifiers. Consequently, recall is 100% for any method that does not discard detections, and precision only measures duplicates created by imperfect association rather than rejection of spurious detector outputs. The abstract and Section VI.A claim that UniKF 'maintains near-perfect precision and recall,' but this is not evidence of robustness to realistic detector errors. Section VI.G itself acknowledges that UniKF cannot recover objects missed by all sensors and will propagate false positives if two agents output associated spurious boxes, yet the evaluation never exercises these failure modes.
  2. [Abstract, Section VI.A, Table II] The reported results do not support the abstract's 'up to 3x lower' claim. In Table II, the largest observed ratios against NMS-STD are 2.75x for mATE under Noise 2 and 2.77x for mADE under Noise 2, and the maximum mATE ratio is 3.44/1.29 = 2.67 under Noise 3. More importantly, CSBA+UniKF and CSBA+WLS are statistically indistinguishable on mATE and mADE in every configuration (e.g., 0.51 +/- 0.12 vs 0.53 +/- 0.15 under Noise 1; 1.29 +/- 0.23 vs 1.30 +/- 0.24 under Noise 3), with only small differences in mAOE. The 'consistently outperforms' claim should be tempered, and paired significance tests should be reported before claiming an advantage over the WLS baseline.
  3. [Section IV.B, Section V, Section VI.A] The time-sensitive update mechanism is a central contribution, but no experiment varies latency or asynchrony. The paper states that nuScenes observed delays are 1-100 ms and that CSBA+WLS with a 100 ms sliding window approximates synchrony, which explains why WLS is nearly as accurate. Without controlled timing scenarios (e.g., synthetic latencies approaching Delta_max, out-of-sequence arrivals, or different sensor rates), the claimed advantage of UniKF's roll-back and forward-propagation logic over WLS is not demonstrated. The evaluation protocol controls noise magnitude but not timing, so the synchronization-handling part of the contribution is untested.
  4. [Section III.B, Section V, Section VI] The validity of the entire evaluation rests on the assumption that zero-mean Gaussian perturbations of ground-truth boxes with hand-set distance-dependent standard deviations replicate the statistical structure of real detector errors. The paper explicitly excludes systematic calibration offsets and localization drift (Section III.B), and it never runs the fusion pipeline on outputs of actual detectors. As a result, the measured gains, especially the near-perfect precision and recall, may not transfer to real multi-sensor or multi-agent systems. At minimum, the conclusions should be framed as valid for the synthetic noise model, and an experiment with a real pretrained BEV detector would be needed to support the broader deployment claims.
minor comments (5)
  1. [Section V] The text states that vision-based detectors achieve an orientation error of 0.2 degrees, which is implausibly low for typical monocular 3D detectors; if this is a typo for 0.2 rad or 2 degrees, it should be corrected because it motivates the Noise 1 orientation sigma.
  2. [Section VI.F] The qualitative-results subsection contains duplicated paragraphs and several typos ('reatain', 'reeceived', 'asssociation', 'In the other hand'); the text should be cleaned and the duplicate paragraph removed.
  3. [Figure 4] The legend mentions 'UniKF with Ground Truth Association (in blue)', but the colors are not easily distinguishable in grayscale; please add distinct markers or a different line style.
  4. [Table IV] The values for CSBA+UniKF differ slightly between the 'SOTA' and 'Proposed' rows (mATE 0.66 vs 0.67, mAOE 4.39 vs 4.38) even though the text says the results are identical; clarify whether these come from different averaging procedures or different trial subsets.
  5. [Section III.C] When multiple predictions correspond to the same ground truth, the protocol retains only the closest prediction as a true positive, but it does not specify how the remaining duplicate predictions enter ATE, AOE, and ADE; since the proposed metrics include false positives, this needs a precise definition.

Circularity Check

1 steps flagged · score 6.0 of 10

Near-perfect precision/recall is structurally guaranteed by the noise-injection protocol, which never introduces false positives or missed objects; the headline robustness numbers on those metrics are a consequence of the input construction.

  1. self definitional [Section III.B (Noise Model) / Section III.C (Evaluation Protocol) / Section VI.F (Qualitative results) and Table II]
    "The noise is sampled from eq. (3) and applied to obtain the perturbed position and orientation in eq. (4). ... We match predicted bounding boxes to ground truth using object identifiers, ensuring a perfect pairing. ... UniKF effectively performs fusion across all bounding boxes, avoiding both false positives and false negatives as long as at least one source provides a detection. This robustness leads to the nearly perfect 100% recall and precision observed in Tab. II."

    Every detection fed to the fusion pipeline is generated by adding zero-mean Gaussian perturbations to every ground-truth box (eqs. 3-4), with no object dropout and no spurious boxes. The evaluation protocol then pairs predictions to ground truth by object identifiers, so any fusion output that does not discard an object trivially reaches 100% recall, and precision is reduced only by association-induced duplicates. The paper explicitly states that UniKF does not create or miss data as long as one source supplies a detection, so the reported 99.5-100% precision/recall is a property of the input construction rather than evidence of robustness to detector-level false positives or missed detections.

full rationale

The paper's core error metrics (mATE, mADE, mAOE) compare fusion methods on synthetic noisy detections and are not circular: they quantify how well each method averages noisy inputs, and the comparison against external baselines (NMS, PSA, WBF, DAIR-V2X-Late, InfraDet3D-Late, CSBA+WLS) gives the proposed UniKF independent content. However, the precision/recall component of the headline is structurally guaranteed: because every ground-truth object is always detected by every source (no false positives, no false negatives), and matching uses ground-truth object IDs, a fusion method that preserves objects achieves 100% recall by construction. The paper itself concedes this in Section VI.F. This makes the precision/recall robustness claim circular with respect to the evaluation protocol, though it does not invalidate the error-metric comparisons. Self-citations to the authors' prior work [30] for CSBA and the constant-velocity model are present, but they are not the main source of circularity; the decisive issue is the input-construction guarantee. Overall, partial circularity: one headline claim reduces by construction, while the central fusion-error comparison retains independent content.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central results rest on a self-designed synthetic noise protocol that assumes Gaussian errors and ignores object classification errors. The method itself is standard Kalman filtering with time-alignment heuristics. No new physical entities are introduced.

free parameters (8)
  • Base position noise sigma_x0 = sigma_y0 = 0.2m / 0.5m / 1.0m for Noise 1/2/3
    Hand-set to mimic reported nuScenes detector translation errors (Section V).
  • Base orientation noise sigma_theta0 = 0.2 deg / 5 deg / 10 deg
    Hand-set to mirror vision and LiDAR detector orientation errors (Section V).
  • Size noise sigma_alpha = sigma_beta = 0.2 / 0.5 / 1.0
    Hand-set to match scale error range from state-of-the-art detectors (Section V).
  • Position noise increase rate kx = ky = 0.01
    Chosen by hand in the noise model (Section V).
  • Orientation noise increase rate k_theta = 0.1
    Fixed value in the noise model (Section V).
  • Time tolerance epsilon_s = 10 ms
    Fusion parameter selected by the authors (Section V).
  • Latency threshold Delta_max = 500 ms
    Fusion parameter selected by the authors (Section V).
  • Number of trials N = 5
    Robustness trials, small and hand-chosen (Section V).
assumptions (6)
  • domain assumption All noise sources follow a zero-mean Gaussian distribution.
    The entire evaluation protocol depends on this; acknowledged as a limitation in Section IV.D and VI.G.
  • domain assumption All provided observations are correctly classified.
    Removes categorization errors from the study, so fusion is evaluated only for correctly labeled boxes (Section III.A).
  • domain assumption Constant velocity (CV) motion model is sufficient for object dynamics.
    The Kalman prediction step uses CV; the paper admits it may be insufficient for highly maneuvering objects (Section IV.A).
  • domain assumption nuScenes validation ground-truth annotations are accurate enough to serve as clean references.
    Noise is added to these boxes and the goal is to recover them; any GT error enters the evaluation (Section V).
  • standard math Standard Kalman filter equations are valid for the state-space model.
    Assumes Gaussian process and measurement noise, linear transition and measurement models (Section IV.A).
  • standard math Kalman filter remains the best linear unbiased estimator when covariance matches the true second moment, even for non-Gaussian noise.
    Used to justify robustness beyond Gaussian noise; cited from [45], [46] (Section VI.G).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluation of an Uncertainty-Aware Late Fusion Algorithm for Multi-Source Bird's Eye View Detections Under Controlled Noise." pith.science (2026). https://pith.science/paper/Z6PGHT3B

@misc{pith2026250703381,
  author       = {Pith},
  title        = {Pith review of: Evaluation of an Uncertainty-Aware Late Fusion Algorithm for Multi-Source Bird's Eye View Detections Under Controlled Noise},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z6PGHT3B}},
  note         = {Machine review of arXiv:2507.03381}
}
read the original abstract

Reliable multi-source fusion is crucial for robust perception in autonomous systems. However, evaluating fusion performance independently of detection errors remains challenging. This work introduces a systematic evaluation framework that injects controlled noise into ground-truth bounding boxes to isolate the fusion process. We then propose Unified Kalman Fusion (UniKF), a late-fusion algorithm based on Kalman filtering to merge Bird's Eye View (BEV) detections while handling synchronization issues. Experiments show that UniKF outperforms baseline methods across various noise levels, achieving up to 3x lower object's positioning and orientation errors and 2x lower dimension estimation errors, while maintaining nearperfect precision and recall between 99.5% and 100%.

Figures

Figures reproduced from arXiv: 2507.03381 by the authors.

Figure 1
Figure 1. Overview of the proposed modular late fusion pipeline. The system is agnostic to both sensor modality and detection algorithm, supporting arbitrary [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. UniKF time consistency handling for different sensor types. Sensor A provides synchronous data updates within a predefined tolerance window, allowing direct state updates. Sensor B delivers delayed measurements; thus, the filter rolls back to the nearest historical state, incorporates the delayed measurement, and propagates forward to the current timestamp. Sensor C provides asynchronous data at irregular intervals,… view at source ↗
Figure 3
Figure 3. Per-class comparison of mATE and mAOE for our method CSBA+UniKF, DAIR-V2X-Late [34], and NMS-STD [4] under the combined Noise 1 and Noise 3 configuration [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Impact of different association algorithms combined with [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results of fusion algorithms of heterogeneous detectors [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 35 canonical work pages

  1. [30]

    Weighted least-squares multi-detection fusion and kalman filter-based tracking for collaborative perception systems

    M. Fadili, L. Lecrosnier, S. Pechberti, and R. Khemmar, “Weighted least-squares multi-detection fusion and kalman filter-based tracking for collaborative perception systems.” [Online]. Available: https: //hal.science/hal-04910986

    work page

  2. [1]

    Collaborative perception in autonomous driving: Methods, datasets, and challenges,

    Y . Han, H. Zhang, H. Li, Y . Jin, C. Lang, and Y . Li, “Collaborative perception in autonomous driving: Methods, datasets, and challenges,” IEEE Intelligent Transportation Systems Magazine, vol. 15, no. 6, 2023

    work page 2023

  3. [2]

    Survey on cooperative perception in an automotive context,

    A. Caillot, S. Ouerghi, P. Vasseur, R. Boutteau, and Y . Dupuis, “Survey on cooperative perception in an automotive context,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 9, pp. 14 204–14 223, 2022

    work page 2022

  4. [3]

    Multi-modal sensor fusion for auto driving perception: A survey

    K. Huang, B. Shi, X. Li, X. Li, S. Huang, and Y . Li, “Multi-modal sensor fusion for auto driving perception: A survey.” [Online]. Available: http://arxiv.org/abs/2202.02703

    arXiv

  5. [4]

    Efficient non-maximum suppression,

    A. Neubeck and L. Van Gool, “Efficient non-maximum suppression,” in 18th International Conference on Pattern Recognition (ICPR’06) . IEEE, 2006, pp. 850–855. [Online]. Available: http://ieeexplore.ieee. org/document/1699659/

    work page arXiv 2006

  6. [5]

    CooPre: Cooperative pretraining for v2x cooperative perception

    S. Z. Zhao, H. Xiang, C. Xu, X. Xia, B. Zhou, and J. Ma, “CooPre: Cooperative pretraining for v2x cooperative perception.” [Online]. Available: http://arxiv.org/abs/2408.11241

    arXiv

  7. [6]

    COˆ3: Cooperative unsupervised 3d representation learning for autonomous driving,

    R. Chen, Y . Mu, R. Xu, W. Shao, C. Jiang, H. Xu, Z. Li, and P. Luo, “COˆ3: Cooperative unsupervised 3d representation learning for autonomous driving,” publisher: arXiv Version Number: 2. [Online]. Available: https://arxiv.org/abs/2206.04028

    work page arXiv

  8. [7]

    Collaborative multi-object tracking with conformal uncertainty propagation,

    S. Su, S. Han, Y . Li, Z. Zhang, C. Feng, C. Ding, and F. Miao, “Collaborative multi-object tracking with conformal uncertainty propagation,” vol. 9, no. 4, pp. 3323–3330, conference Name: IEEE Robotics and Automation Letters. [Online]. Available: https: //ieeexplore.ieee.org/document/10430224

    work page arXiv

Show all 47 references
  1. [8]

    When2com: Multi-agent perception via communication graph grouping

    Y .-C. Liu, J. Tian, N. Glaser, and Z. Kira, “When2com: Multi-agent perception via communication graph grouping.” [Online]. Available: http://arxiv.org/abs/2006.00176

    2006 arXiv

  2. [9]

    V2x-DGW: Domain generalization for multi-agent perception under adverse weather conditions

    B. Li, J. Li, X. Liu, R. Xu, Z. Tu, J. Guo, X. Li, and H. Yu, “V2x-DGW: Domain generalization for multi-agent perception under adverse weather conditions.” [Online]. Available: http://arxiv.org/abs/2403.11371

    arXiv

  3. [10]

    CoBEVFusion: Cooperative perception with LiDAR-camera bird’s-eye view fusion

    D. Qiao and F. Zulkernine, “CoBEVFusion: Cooperative perception with LiDAR-camera bird’s-eye view fusion.” [Online]. Available: http://arxiv.org/abs/2310.06008

    arXiv

  4. [11]

    TUMTraf v2x cooperative perception dataset

    W. Zimmer, G. A. Wardana, S. Sritharan, X. Zhou, R. Song, and A. C. Knoll, “TUMTraf v2x cooperative perception dataset.” [Online]. Available: http://arxiv.org/abs/2403.01316

    arXiv

  5. [12]

    Data optimization strategies for collaborative perception,

    B. Abdali, Q. Picard, and M. Fadili, “Data optimization strategies for collaborative perception,” Electronic Imaging, vol. 37, pp. 1–5, 2025

    2025

  6. [13]

    Soft-NMS — improving object detection with one line of code,

    N. Bodla, B. Singh, R. Chellappa, and L. S. Davis, “Soft-NMS — improving object detection with one line of code,” in 2017 IEEE International Conference on Computer Vision (ICCV). IEEE, pp. 5562–

    2017

  7. [14]

    Model- agnostic multi-agent perception framework,

    R. Xu, W. Chen, H. Xiang, X. Xia, L. Liu, and J. Ma, “Model- agnostic multi-agent perception framework,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) , 2023, pp. 1471–1478

    2023

  8. [15]

    A competitive method to VIPriors object detection challenge,

    F. Shen, X. He, M. Wei, and Y . Xie, “A competitive method to VIPriors object detection challenge,” version Number: 1. [Online]. Available: https://arxiv.org/abs/2104.09059

    arXiv

  9. [16]

    InfraDet3d: Multi-modal 3d object detection based on roadside infrastructure camera and LiDAR sensors

    W. Zimmer, J. Birkner, M. Brucker, H. T. Nguyen, S. Petrovski, B. Wang, and A. C. Knoll, “InfraDet3d: Multi-modal 3d object detection based on roadside infrastructure camera and LiDAR sensors.” [Online]. Available: http://arxiv.org/abs/2305.00314

    arXiv

  10. [17]

    DAIR-v2x: A large-scale dataset for vehicle-infrastructure cooperative 3d object detection

    H. Yu, Y . Luo, M. Shu, Y . Huo, Z. Yang, Y . Shi, Z. Guo, H. Li, X. Hu, J. Yuan, and Z. Nie, “DAIR-v2x: A large-scale dataset for vehicle-infrastructure cooperative 3d object detection.” [Online]. Available: http://arxiv.org/abs/2204.05575

    arXiv

  11. [18]

    Decentralized perception system with multiple viewpoints,

    Q. Picard, M. Morice, M. Fadili, and S. Pechberti, “Decentralized perception system with multiple viewpoints,” Oct. 2024, working paper or preprint. [Online]. Available: https://hal.science/hal-04744167

    2024

  12. [19]

    Multi-Modal 3D Object Detection in Autonomous Driving: A Survey and Taxonomy,

    L. Wang, X. Zhang, Z. Song, J. Bi, G. Zhang, H. Wei, L. Tang, L. Yang, J. Li, C. Jia, and L. Zhao, “Multi-Modal 3D Object Detection in Autonomous Driving: A Survey and Taxonomy,” IEEE Transactions on Intelligent Vehicles , vol. 8, no. 7, pp. 3781–3798, Jul. 2023. [Online]. Ava...

    2023

  13. [20]

    Multi-Sensor Fusion Technology for 3D Object Detection in Autonomous Driving: A Review,

    X. Wang, K. Li, and A. Chehri, “Multi-Sensor Fusion Technology for 3D Object Detection in Autonomous Driving: A Review,” IEEE Transactions on Intelligent Transportation Systems , vol. 25, no. 2, pp. 1148–1165, Feb. 2024, conference Name: IEEE Transactions on Intelligent Transp...

    2024

  14. [21]

    Multi-view 3d object detection network for autonomous driving,

    X. Chen, H. Ma, J. Wan, B. Li, and T. Xia, “Multi-view 3d object detection network for autonomous driving,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , July 2017

    2017

  15. [22]

    Joint 3d proposal generation and object detection from view aggregation,

    J. Ku, M. Mozifian, J. Lee, A. Harakeh, and S. L. Waslander, “Joint 3d proposal generation and object detection from view aggregation,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018, pp. 1–8

    2018

  16. [23]

    Clocs: Camera-lidar object candi- dates fusion for 3d object detection,

    S. Pang, D. Morris, and H. Radha, “Clocs: Camera-lidar object candi- dates fusion for 3d object detection,” 2020

    2020

  17. [24]

    Fast-clocs: Fast camera-lidar object candidates fusion for 3d object detection,

    ——, “Fast-clocs: Fast camera-lidar object candidates fusion for 3d object detection,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , January 2022, pp. 187–196

    2022

  18. [25]

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

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

    2023

  19. [26]

    A review of data fusion techniques,

    F. Castanedo, “A review of data fusion techniques,” vol. 2013, p. 704504

    2013

  20. [27]

    Uncertainty quantification of collaborative detection for self-driving,

    S. Su, Y . Li, S. He, S. Han, C. Feng, C. Ding, and F. Miao, “Uncertainty quantification of collaborative detection for self-driving,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 5588–5594

    2023

  21. [28]

    Uncertainty prediction for monocular 3d object detection,

    J. Mun and H. Choi, “Uncertainty prediction for monocular 3d object detection,” vol. 23, no. 12, p. 5395, number: 12 Publisher: Multidisciplinary Digital Publishing Institute. [Online]. Available: https://www.mdpi.com/1424-8220/23/12/5395

  22. [29]

    Probabilis- tic 3d multi-object cooperative tracking for autonomous driving via differentiable multi-sensor kalman filter,

    H.-K. Chiu, C.-Y . Wang, M.-H. Chen, and S. F. Smith, “Probabilis- tic 3d multi-object cooperative tracking for autonomous driving via differentiable multi-sensor kalman filter,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) , 2024

    2024

  23. [31]

    Object detection and identi- fication using vision and radar data fusion system for ground-based navigation,

    H. Jha, V . Lodhi, and D. Chakravarty, “Object detection and identi- fication using vision and radar data fusion system for ground-based navigation,” in 2019 6th International Conference on Signal Processing and Integrated Networks (SPIN) , 2019, pp. 590–593

    2019

  24. [32]

    Vision meets robotics: The KITTI dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The KITTI dataset,” vol. 32, no. 11, pp. 1231–1237. [Online]. Available: http://journals.sagepub.com/doi/10.1177/0278364913491297

    doi

  25. [33]

    nuScenes: A multimodal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuScenes: A multimodal dataset for autonomous driving,” 2020. [Online]. Available: http://arxiv.org/abs/1903.11027

    2020 arXiv

  26. [34]

    Dair-v2x: A large-scale real-world v2i collaborative perception dataset,

    C. Yu, Y . Li, and R. Zhang, “Dair-v2x: A large-scale real-world v2i collaborative perception dataset,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  27. [35]

    Uncertainty estimation for deep neural object detectors in safety-critical applications,

    M. T. Le, F. Diehl, T. Brunner, and A. Knoll, “Uncertainty estimation for deep neural object detectors in safety-critical applications,” in 2018 21st International Conference on Intelligent Transportation Systems (ITSC) , 2018, pp. 3873–3878

    2018

  28. [36]

    The truncated normal distribution

    J. Burkardt, “The truncated normal distribution.”

  29. [37]

    Derivation of a constant velocity motion model for visual tracking

    N. L. Baisa, “Derivation of a constant velocity motion model for visual tracking.” [Online]. Available: http://arxiv.org/abs/2005.00844

    2005 arXiv

  30. [38]

    Opv2v: Opencda-based v2v collaborative perception dataset,

    R. Xu, T. Wang, and Y . Chen, “Opv2v: Opencda-based v2v collaborative perception dataset,” in Proceedings of the IEEE International Confer- ence on Robotics and Automation (ICRA) , 2022

    2022

  31. [39]

    V2xset: A large-scale open simulation dataset for v2x perception,

    Z. Hu, F. Liu, and Y . Shen, “V2xset: A large-scale open simulation dataset for v2x perception,” in Proceedings of the European Conference on Computer Vision (ECCV) , 2022

    2022

  32. [40]

    V2x-sim: Multi-agent collaborative perception dataset,

    Q. Li, T. Wang, and Y . Chen, “V2x-sim: Multi-agent collaborative perception dataset,” IEEE Robotics and Automation Letters , 2021

    2021

  33. [41]

    A survey of uncertainty in deep neural networks,

    J. Gawlikowski, C. R. N. Tassi, M. Ali, J. Lee, M. Humt, J. Feng, A. Kruspe, R. Triebel, P. Jung, R. Roscher, M. Shahzad, W. Yang, R. Bamler, and X. X. Zhu, “A survey of uncertainty in deep neural networks,” Artificial Intelligence Review , vol. 56, no. S1, pp. 1513–1589, Oct....

    2023 doi

  34. [42]

    AB3dmot: A baseline for 3d multi-object tracking and new evaluation metrics

    X. Weng, J. Wang, D. Held, and K. Kitani, “AB3dmot: A baseline for 3d multi-object tracking and new evaluation metrics.” [Online]. Available: http://arxiv.org/abs/2008.08063

    2008 arXiv

  35. [43]

    V2v4real: A real-world multimodal collaborative perception dataset for vehicle-to-vehicle interaction,

    Y . Yang, X. Ma, and R. Wang, “V2v4real: A real-world multimodal collaborative perception dataset for vehicle-to-vehicle interaction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  36. [44]

    Generalized intersection over union: A metric and a loss for bounding box regression

    H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese, “Generalized intersection over union: A metric and a loss for bounding box regression.” [Online]. Available: http://arxiv.org/abs/1902.09630

    1902 arXiv

  37. [45]

    The kantorovich inequality for error analysis of the kalman filter with unknown noise distributions,

    J. C. Spall, “The kantorovich inequality for error analysis of the kalman filter with unknown noise distributions,” Automatica, vol. 31, no. 10, pp. 1513–1517, 1995. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/0005109895000699

    1995

  38. [46]

    Use of the kalman filter for inference in state-space models with unknown noise distributions,

    J. Maryak, J. Spall, and B. Heydon, “Use of the kalman filter for inference in state-space models with unknown noise distributions,” IEEE Transactions on Automatic Control , vol. 49, no. 1, pp. 87–90, 2004

    2004

  39. [5570]

    Available: http://ieeexplore.ieee.org/document/8237855/

    [Online]. Available: http://ieeexplore.ieee.org/document/8237855/

Pith tools

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