Pith. sign in

REVIEW 3 major objections 4 minor 54 references

USVTrack: USV-Based 4D Radar-Camera Tracking Dataset for Autonomous Driving in Inland Waterways

T0 review · 3 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 USVTrack presents the first 4D radar-camera tracking dataset for unmanned surface vehicles on inland waterways, plus a radar-camera matching module that improves tracking accuracy.

desk verdict USVTrack is a genuinely useful first dataset for 4D radar-camera tracking on inland waterways, but the paper needs to show calibration, synchronization, and annotation quality before the RCM numbers carry real weight. read the letter →

arxiv 2506.18737 v1 pith:SMR74IOK submitted 2025-06-23 cs.CV cs.RO

Shanliang Yao , Runwei Guan , Yi Ni , Sen Xu , Yong Yue , Xiaohui Zhu , Ryan Wen Liu This is my paper
classification cs.CVcs.RO
keywords 4Dradarradar-camerafusionobjecttrackingunmannedsurfacevehicleinlandwaterwaysdatasetmulti-objectbenchmark
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

The paper presents USVTrack, described as the first 4D radar-camera tracking dataset collected from an unmanned surface vehicle (USV) on inland waterways, and claims it fills the gap left by road-focused radar-camera datasets. If the claim stands, waterborne autonomous systems gain a shared benchmark with 60 clips, 68,822 camera frames, 85,229 object annotations, 45,091 radar frames, and three tracked classes (ship, boat, vessel) across varied waterways, lighting, weather, and seasons. The paper also proposes RCM, a plug-in matching module for two-stage trackers that replaces IoU-based secondary association with a cost combining camera spatial alignment and radar-derived motion consistency. The reported experiments show RCM raising HOTA and reducing identity switches on four trackers, which the paper reads as evidence that radar-camera matching improves tracking reliability in waterborne environments.

What carries the argument

The load-bearing object is the USVTrack dataset itself, collected from a USV equipped with a 77GHz 4D radar, a monocular camera, GPS, and IMU, with annotations made by projecting radar points onto camera frames and labeling boxes and track IDs within them. The second mechanism is RCM, a cross-modal association module that plugs into the second association stage of two-stage trackers; it clusters radar returns with DBSCAN, then computes a matching cost $C^{\text{rcm}}_{i,j}$ that combines IoU distance $D^{\text{iou}}$ and a normalized radar Mahalanobis distance $\hat{D}^{\text{radar}}$, weighted by $\alpha$ and gated by thresholds $\theta_{\text{iou}}$ and $\theta_{\text{rcm}}$. This cost is what lets low-confidence camera detections be matched using radar motion evidence when pure spatial overlap would fail.

What would settle it

Re-annotate a random subset of frames with a second annotator and compute agreement on bounding boxes and track IDs, and separately perturb the radar-camera projection by small rotations and translations to see how much HOTA and IDSW move; if annotation agreement is low or if small projection perturbations change the reported tracking gains, the central claim of a reliable benchmark does not hold.

Watch

Extended reading notes

Core claim

The paper's central claim is that USVTrack is the first tracking dataset tailored for USV-based autonomous driving in waterborne transportation systems to combine 4D radar, monocular camera, GPS, and IMU data, and that its associated RCM method improves multi-object tracking accuracy and reliability. The dataset supplies 60 annotated clips with three vessel categories, split into train and test halves following the MOT20 methodology, and includes radar point clouds projected onto camera frames with categories and track IDs. RCM is described as a simple radar-camera matching strategy that takes clustered radar detections and low-confidence camera boxes and computes a matching cost from a weighted combination of IoU distance and a normalized Mahalanobis distance on radar dynamic attributes, with thresholds to reject inconsistent matches. The experiments compare four two-stage trackers with and without RCM and report consistent gains, for example Hybrid-SORT rising from 47.208 to 48.302 HOTA while identity switches fall from 207 to 186.

Load-bearing premise

The benchmark numbers rest on the unmeasured accuracy of the radar-to-camera projection and the manual annotations: if the projection is misaligned or the bounding boxes and track IDs are noisy, every detection and tracking score reported on USVTrack inherits that error.

Editorial extensions

If this is right

  • Researchers in waterborne autonomy get a common testbed with 85,229 camera annotations and 44,271 radar targets, so tracker results become comparable across labs.
  • Fusion-based detectors and trackers can be evaluated on a benchmark that includes rain, fog, snow, dim light, and night, conditions where camera-only tracking is known to fail.
  • The RCM module, being a drop-in replacement for the second association stage, can be appended to any two-stage tracker without retraining, as demonstrated on ByteTrack, OC-SORT, BoT-SORT, and Hybrid-SORT.
  • If the reported gains hold, radar-camera matching becomes a cheap way to reduce identity switches in sparse, occluded, or camera-shaken waterborne scenes.

Reading between the lines

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

  • The dataset's distance distribution, with a peak of roughly 7,658 objects in the 80–100 meter bin, suggests that radar-specific detection ranges could be used to design range-aware tracking confidence, an idea the paper does not pursue.
  • Because RCM relies on radar-derived velocity and direction, its benefit should grow in adverse weather; a testable extension would be to report per-condition HOTA gains to check whether the improvement is concentrated in rain, fog, and snow frames.
  • The unmeasured projection alignment means the published benchmark numbers should be read with caution; an obvious extension is for the authors to release extrinsic calibration error statistics or a re-annotation agreement study.
  • The dataset's categories are limited to ship, boat, and vessel, and kayaks and ferries are present but unannotated; extending the annotations to those classes could make the benchmark directly useful for water rescue and ferry-safety applications.
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 / 4 minor

Summary. The paper introduces USVTrack, a 4D radar-camera tracking dataset collected from an unmanned surface vehicle in inland waterways, containing 60 clips, 68,822 camera frames, 85,229 2D bounding-box annotations, and 45,091 radar frames with 44,271 radar targets across three classes (ship, boat, vessel), together with GPS and IMU data. The authors benchmark nine detectors and eight trackers on the test split and propose RCM, a radar-camera matching cost that replaces the second-stage IoU association in two-stage trackers. In Tables IV and V, RCM improves HOTA by 0.75 to 1.88 and reduces IDSW by 21 to 113 across BoT-SORT, ByteTrack, OC-SORT, and Hybrid-SORT.

Significance. USVTrack addresses a genuine gap: existing waterborne tracking datasets are either camera-only (SMD, SeaDronesSee, BoaTrack) or unannotated (FloW), while road 4D radar-camera datasets are not directly transferable to waterways. The dataset is publicly released, covers diverse seasons, weather, lighting, and waterway types, and reports detailed per-class statistics, which are useful contributions. The benchmark of multiple detectors and trackers on a new domain is also valuable. However, the central claims are conditional: the radar-camera annotation pipeline is not documented with calibration or synchronization evidence, and the RCM gains are reported without statistical support or a validation-based tuning protocol. If these issues are addressed, the dataset and method could be a solid foundation for waterborne multi-object tracking research.

major comments (3)
  1. [Section III-C and Table III] The annotation description is too thin to support the dataset's central claim. The paper does not report the radar-camera extrinsic calibration procedure, reprojection error, or timestamp synchronization policy, yet radar points were annotated by projecting them into image bounding boxes. The frame counts in Table III imply a 712-second mismatch between radar (45,091 frames at 15 Hz is about 3006 seconds) and camera (68,822 frames at 30 Hz is about 2294 seconds), so the frame-level fusion used by RCM needs an explicit synchronization and interpolation rule that is never described. Without this, the projection-based radar labels and all HOTA/IDSW numbers in Table V rest on an unmeasured premise. Please add calibration parameters, reprojection error statistics, synchronization details, and an annotation-quality measure such as double-annotation IoU or inter-annotator agreement.
  2. [Section IV-B3 and Table V] The claim that RCM improves tracking is not statistically supported. The hyperparameters alpha=0.7, lambda=1.2, theta_rcm=0.5 are chosen "based on extensive experience and experimental results" on what appears to be the same test set, with no separate validation split, no sensitivity analysis, and no ablation across alpha and lambda. All metrics are single runs with no variance or significance testing. Given that the reported gains are small (HOTA +0.75 to +1.88, IDSW -21 to -113), one cannot rule out that these improvements are tuning artifacts. Please report a validation-based selection procedure, sensitivity curves, and multiple-seed or run-to-run variance, or a clear statement of why the trackers are deterministic and the tuning protocol is not circular.
  3. [Section IV-A, Eq. (1)] The RCM implementation is not reproducible as described. The method relies on DBSCAN clustering of radar points, but the DBSCAN parameters (eps, minPts) are not given, and the procedure for converting radar clusters into detections compatible with image-space IoU is not specified. Since Eq. (1) computes D_iou between track bounding boxes and detected bounding boxes, and radar clusters do not naturally produce image bounding boxes, the matching pipeline needs a precise definition of how clustered radar detections are projected, sized, and assigned to camera detections. Please provide the full algorithmic steps and parameter values, or make the code available with the dataset release.
minor comments (4)
  1. [Table IV] The CRF-Net row is internally inconsistent: the reported class APs 89.7, 78.6, and 72.8 average to about 80.4, not the listed 86.4. Please correct the table and clarify whether the reported mAP is mAP@50-95 or mAP@50.
  2. [Eqs. (1)-(2)] Eq. (1) thresholds the raw D_radar against theta_rcm, while Eq. (2) defines the normalized D_radar_hat; please clarify which quantity is actually compared with theta_rcm and whether the threshold applies to the normalized or unnormalized distance.
  3. [Contributions list] The third contribution uses the name "RCMatch" while the rest of the paper uses "RCM"; please unify the terminology.
  4. [Section IV-C2] "Trackor++" is a typo for "Tracktor++"; please correct it in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the dataset and RCM benchmark are self-contained and not reduced to their inputs.

full rationale

The paper's derivation chain is empirical rather than formal: it contributes a dataset (USVTrack) and a matching module (RCM) and reports benchmark numbers on its own test split. There is no equation-level reduction: RCM's cost in Eqs. (1)-(2) is a hand-designed fusion of IoU and a normalized radar Mahalanobis distance, and the reported HOTA/IDSW gains are not defined in terms of these costs, so the improvements are not forced by construction. The radar annotations were created by projecting radar points into camera bounding boxes and copying the camera track IDs (Section III-C), which means the dataset's radar labels are camera-derived; this is a limitation for claims of radar-only annotation independence, but RCM is evaluated on the camera-box ground truth and uses raw clustered radar points, not the radar ground-truth labels, so the benchmark result does not reduce to the annotation procedure. The RCM hyperparameters (alpha, lambda, theta_rcm, theta_iou) are set 'based on extensive experience and experimental results' rather than on a separate validation split; this is a methodological weakness that can inflate results, but it is not a circularity because the free parameters are not fitted to the reported metric by a closed-form relation, and the baseline trackers are unchanged in the comparison. The only self-citations ([3], [10], [20]) are to peer-reviewed prior work and are not used to forbid alternatives or to import an unverified uniqueness claim. The unmeasured radar-camera calibration, synchronization, and annotation quality are correctness and verifiability risks, not circularity. Accordingly, no circular step is present.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central numerical claims rest on the quality of the new dataset annotations and the choice of RCM hyperparameters. The annotation and calibration quality are asserted but not measured, and the RCM parameters are set from experimental observation rather than derived.

free parameters (5)
  • RCM weight alpha = 0.7
    Chosen from experimental results, Section IV-B-3.
  • RCM radar sensitivity lambda = 1.2
    Set empirically, Section IV-B-3.
  • RCM thresholds theta_rcm and theta_iou = 0.5 and 0.5
    Set to 0.5, Section IV-B-3.
  • Detection confidence thresholds = 0.5 high, 0.1 low, 0.2 track init
    Set in Section IV-B-2 for the tracking experiments.
  • DBSCAN cluster parameters = not reported
    DBSCAN is used to cluster radar points, but eps and minPts are not reported, Section IV-A.
assumptions (3)
  • domain assumption Radar-camera extrinsic calibration is accurate enough to project radar points into image coordinates.
    Used throughout Section III-C and the benchmarks; no calibration error is reported.
  • domain assumption Manual annotations correctly assign categories and tracking IDs across frames.
    Underlies all ground truth; no inter-annotator agreement or quality metric is given.
  • domain assumption DBSCAN clustering of radar points groups points belonging to the same physical object.
    Section IV-A relies on this to form radar detections; parameters are not specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of USVTrack: USV-Based 4D Radar-Camera Tracking Dataset for Autonomous Driving in Inland Waterways." pith.science (2026). https://pith.science/paper/SMR74IOK

@misc{pith2026250618737,
  author       = {Pith},
  title        = {Pith review of: USVTrack: USV-Based 4D Radar-Camera Tracking Dataset for Autonomous Driving in Inland Waterways},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SMR74IOK}},
  note         = {Machine review of arXiv:2506.18737}
}
read the original abstract

Object tracking in inland waterways plays a crucial role in safe and cost-effective applications, including waterborne transportation, sightseeing tours, environmental monitoring and surface rescue. Our Unmanned Surface Vehicle (USV), equipped with a 4D radar, a monocular camera, a GPS, and an IMU, delivers robust tracking capabilities in complex waterborne environments. By leveraging these sensors, our USV collected comprehensive object tracking data, which we present as USVTrack, the first 4D radar-camera tracking dataset tailored for autonomous driving in new generation waterborne transportation systems. Our USVTrack dataset presents rich scenarios, featuring diverse various waterways, varying times of day, and multiple weather and lighting conditions. Moreover, we present a simple but effective radar-camera matching method, termed RCM, which can be plugged into popular two-stage association trackers. Experimental results utilizing RCM demonstrate the effectiveness of the radar-camera matching in improving object tracking accuracy and reliability for autonomous driving in waterborne environments. The USVTrack dataset is public on https://usvtrack.github.io.

Figures

Figures reproduced from arXiv: 2506.18737 by the authors.

Figure 1
Figure 1. Example scenario from our USVTrack dataset from the viewpoint [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Samples in USVTrack dataset. Radar points are projected onto the image plane as colored dots. For each radar point on the image, the color [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Statistics of objects in USVTrack. (a) Wide range of object size. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of object tracking on USVTrack. The first row shows the ground truth, the second displays results from HybridSORT, and the third [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 31 canonical work pages

  1. [1]

    Architecture and key technologies for new generation of waterborne transportation sys- tem,

    X. Yan, C. Li, J. Liu, X. You, S. Wang, and F. Ma, “Architecture and key technologies for new generation of waterborne transportation sys- tem,”Journal of Transportation Systems Engineering and Information Technology, vol. 21, no. 5, p. 22, 2021

    work page 2021

  2. [2]

    The mastr1325 dataset for training deep usv obstacle detection models,

    B. Bovcon, J. Muhovi ˇc, J. Per ˇs, and M. Kristan, “The mastr1325 dataset for training deep usv obstacle detection models,” in2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2019, pp. 3431–3438

    work page 2019

  3. [3]

    Waterscenes: A multi-task 4d radar-camera fusion dataset and benchmarks for autonomous driving on water surfaces,

    S. Yao, R. Guan, Z. Wu, Y . Ni, Z. Huang, R. W. Liu, Y . Yue, W. Ding, E. G. Lim, H. Seoet al., “Waterscenes: A multi-task 4d radar-camera fusion dataset and benchmarks for autonomous driving on water surfaces,”IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 11, pp. 16 584–16 598, 2024

    work page 2024

  4. [4]

    Are we ready for unmanned surface vehicles in inland waterways? the usvinland multisensor dataset and benchmark,

    Y . Cheng, M. Jiang, J. Zhu, and Y . Liu, “Are we ready for unmanned surface vehicles in inland waterways? the usvinland multisensor dataset and benchmark,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 3964–3970, 2021

    work page 2021

  5. [5]

    3rd workshop on maritime computer vision (macvi) 2025: Challenge results,

    B. Kiefer, L. Zust, M. Kristan, J. Pers, M. Tersek, U. Mudenagudi, C. Desai, A. Wiliem, M. Kreis, N. Akalwadiet al., “3rd workshop on maritime computer vision (macvi) 2025: Challenge results,” in Proceedings of the Winter Conference on Applications of Computer Vision, 2025, pp. 1542–1569

    work page 2025

  6. [6]

    A benchmark for deep learning based object detection in maritime environments,

    S. Moosbauer, D. Konig, J. Jakel, and M. Teutsch, “A benchmark for deep learning based object detection in maritime environments,” in2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). IEEE Computer Society, 2019, pp. 916–925

    work page 2019

  7. [7]

    Flow: A dataset and benchmark for floating waste detection in inland waters,

    Y . Cheng, J. Zhu, M. Jiang, J. Fu, C. Pang, P. Wang, K. Sankaran, O. Onabola, Y . Liu, D. Liuet al., “Flow: A dataset and benchmark for floating waste detection in inland waters,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 953–10 962

    work page 2021

  8. [8]

    Seadronessee: A maritime benchmark for detecting humans in open water,

    L. A. Varga, B. Kiefer, M. Messmer, and A. Zell, “Seadronessee: A maritime benchmark for detecting humans in open water,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2022, pp. 2260–2270

    2022

Show all 54 references
  1. [9]

    2nd workshop on maritime computer vision (macvi) 2024: Challenge results,

    B. Kiefer, L. ˇZust, M. Kristan, J. Per ˇs, M. Ter ˇsek, A. Wiliem, M. Messmer, C.-Y . Yang, H.-W. Huang, Z. Jianget al., “2nd workshop on maritime computer vision (macvi) 2024: Challenge results,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vi...

    2024

  2. [10]

    Exploring radar data representations in autonomous driving: A comprehensive review,

    S. Yao, R. Guan, Z. Peng, C. Xu, Y . Shi, Y . Yue, E. G. Lim, H. Seo, K. L. Man, X. Zhuet al., “Exploring radar data representations in autonomous driving: A comprehensive review,”IEEE Transactions on Intelligent Transportation Systems, vol. 26, no. 6, pp. 7401–7425, 2025

    2025

  3. [11]

    Seeing through fog without seeing fog: Deep multi- modal sensor fusion in unseen adverse weather,

    M. Bijelic, T. Gruber, F. Mannan, F. Kraus, W. Ritter, K. Dietmayer, and F. Heide, “Seeing through fog without seeing fog: Deep multi- modal sensor fusion in unseen adverse weather,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp....

    2020

  4. [12]

    Comparative analysis of radar and lidar technologies for automotive applications,

    I. Bilik, “Comparative analysis of radar and lidar technologies for automotive applications,”IEEE Intelligent Transportation Systems Magazine, vol. 15, no. 1, pp. 244–269, 2022

    2022

  5. [13]

    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 Vehicles, vol. 9, no. 1, pp. 799–812, 2023

    2023

  6. [14]

    V2x-radar: A multi-modal dataset with 4d radar for cooperative perception,

    L. Yang, X. Zhang, J. Li, C. Wang, Z. Song, T. Zhao, Z. Song, L. Wang, M. Zhou, Y . Shenet al., “V2x-radar: A multi-modal dataset with 4d radar for cooperative perception,”arXiv preprint arXiv:2411.10962, 2024

    2024

  7. [15]

    Radarocc: Robust 3d occupancy prediction with 4d imaging radar,

    F. Ding, X. Wen, Y . Zhu, Y . Li, and C. X. Lu, “Radarocc: Robust 3d occupancy prediction with 4d imaging radar,”Advances in Neural Information Processing Systems, vol. 37, pp. 101 589–101 617, 2025

    2025

  8. [16]

    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,” Advances in Neural Information Processing Systems, vol. 35, pp. 3819–3829, 2022

    2022

  9. [17]

    Tj4dradset: A 4d radar dataset for au- tonomous driving,

    L. Zheng, Z. Ma, X. Zhu, B. Tan, S. Li, K. Long, W. Sun, S. Chen, L. Zhang, M. Wanet al., “Tj4dradset: A 4d radar dataset for au- tonomous driving,” in2022 IEEE 25th international conference on intelligent transportation systems (ITSC). IEEE, 2022, pp. 493–498

    2022

  10. [18]

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

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

    2023 arXiv

  11. [19]

    Ntu4dradlm: 4d radar-centric multi-modal dataset for localization and mapping,

    J. Zhang, H. Zhuge, Y . Liu, G. Peng, Z. Wu, H. Zhang, Q. Lyu, H. Li, C. Zhao, D. Kircaliet al., “Ntu4dradlm: 4d radar-centric multi-modal dataset for localization and mapping,” in2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2023,...

    2023

  12. [20]

    Radar-camera fusion for object detection and semantic segmentation in autonomous driving: A comprehensive review,

    S. Yao, R. Guan, X. Huang, Z. Li, X. Sha, Y . Yue, E. G. Lim, H. Seo, K. L. Man, X. Zhuet al., “Radar-camera fusion for object detection and semantic segmentation in autonomous driving: A comprehensive review,”IEEE Transactions on Intelligent Vehicles, vol. 9, no. 1, pp. 2094–...

    2023

  13. [21]

    Sgdet3d: Semantics and geometry fusion for 3d object detection using 4d radar and camera,

    X. Bai, Z. Yu, L. Zheng, X. Zhang, Z. Zhou, X. Zhang, F. Wang, J. Bai, and H.-L. Shen, “Sgdet3d: Semantics and geometry fusion for 3d object detection using 4d radar and camera,”IEEE Robotics and Automation Letters, 2024

    2024

  14. [22]

    Lxlv2: Enhanced lidar excluded lean 3d object detection with fusion of 4d radar and camera,

    W. Xiong, Z. Zou, Q. Zhao, F. He, and B. Zhu, “Lxlv2: Enhanced lidar excluded lean 3d object detection with fusion of 4d radar and camera,”IEEE Robotics and Automation Letters, 2025

    2025

  15. [23]

    Faster r-cnn: Towards real- time object detection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real- time object detection with region proposal networks,”Advances in neural information processing systems, vol. 28, 2015

    2015

  16. [24]

    Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,

    C.-Y . Wang, A. Bochkovskiy, and H.-Y . M. Liao, “Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 7464–7475

    2023

  17. [25]

    YOLO by Ultralytics,

    G. Jocher, A. Chaurasia, and J. Qiu, “YOLO by Ultralytics,” Jan

  18. [26]

    Yolov12: Attention-centric real- time object detectors,

    Y . Tian, Q. Ye, and D. Doermann, “Yolov12: Attention-centric real- time object detectors,”arXiv preprint arXiv:2502.12524, 2025

    2025 arXiv

  19. [27]

    The hungarian method for the assignment problem,

    H. W. Kuhn, “The hungarian method for the assignment problem,” Naval research logistics quarterly, vol. 2, no. 1-2, pp. 83–97, 1955

    1955

  20. [28]

    An introduction to the kalman filter,

    G. Welch, G. Bishopet al., “An introduction to the kalman filter,” 1995

    1995

  21. [29]

    Simple online and realtime tracking with a deep association metric,

    N. Wojke, A. Bewley, and D. Paulus, “Simple online and realtime tracking with a deep association metric,” in2017 IEEE international conference on image processing (ICIP). IEEE, 2017, pp. 3645–3649

    2017

  22. [30]

    Bytetrack: Multi-object tracking by associating every detection box,

    Y . Zhang, P. Sun, Y . Jiang, D. Yu, F. Weng, Z. Yuan, P. Luo, W. Liu, and X. Wang, “Bytetrack: Multi-object tracking by associating every detection box,” inEuropean conference on computer vision. Springer, 2022, pp. 1–21

    2022

  23. [31]

    Bot-sort: Robust asso- ciations multi-pedestrian tracking,

    N. Aharon, R. Orfaig, and B.-Z. Bobrovsky, “Bot-sort: Robust asso- ciations multi-pedestrian tracking,”arXiv preprint arXiv:2206.14651, 2022

    2022 arXiv

  24. [32]

    Observation- centric sort: Rethinking sort for robust multi-object tracking,

    J. Cao, J. Pang, X. Weng, R. Khirodkar, and K. Kitani, “Observation- centric sort: Rethinking sort for robust multi-object tracking,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 9686–9696

    2023

  25. [33]

    Fairmot: On the fairness of detection and re-identification in multiple object tracking,

    Y . Zhang, C. Wang, X. Wang, W. Zeng, and W. Liu, “Fairmot: On the fairness of detection and re-identification in multiple object tracking,” International journal of computer vision, vol. 129, pp. 3069–3087, 2021

    2021

  26. [34]

    Tracking without bells and whistles,

    P. Bergmann, T. Meinhardt, and L. Leal-Taixe, “Tracking without bells and whistles,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 941–951

    2019

  27. [35]

    Track- former: Multi-object tracking with transformers,

    T. Meinhardt, A. Kirillov, L. Leal-Taixe, and C. Feichtenhofer, “Track- former: Multi-object tracking with transformers,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8844–8854

    2022

  28. [36]

    Motr: End-to-end multiple-object tracking with transformer,

    F. Zeng, B. Dong, Y . Zhang, T. Wang, X. Zhang, and Y . Wei, “Motr: End-to-end multiple-object tracking with transformer,” inEuropean conference on computer vision. Springer Nature Switzerland Cham, 2022, pp. 659–675

    2022

  29. [37]

    Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors,

    Y . Zhang, T. Wang, and X. Zhang, “Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 22 056–22 065

    2023

  30. [38]

    Motrv3: Release-fetch supervision for end-to-end multi-object tracking,

    E. Yu, T. Wang, Z. Li, Y . Zhang, X. Zhang, and W. Tao, “Motrv3: Release-fetch supervision for end-to-end multi-object tracking,”arXiv preprint arXiv:2305.14298, 2023

    2023 arXiv

  31. [39]

    Multiple object tracking as id prediction,

    R. Gao, Y . Zhang, and L. Wang, “Multiple object tracking as id prediction,”arXiv preprint arXiv:2403.16848, 2024

    2024 arXiv

  32. [40]

    Mot20: A bench- mark for multi object tracking in crowded scenes,

    P. Dendorfer, H. Rezatofighi, A. Milan, J. Shi, D. Cremers, I. Reid, S. Roth, K. Schindler, and L. Leal-Taix ´e, “Mot20: A bench- mark for multi object tracking in crowded scenes,”arXiv preprint arXiv:2003.09003, 2020

    2003 arXiv

  33. [41]

    Dancetrack: Multi-object tracking in uniform appearance and diverse motion,

    P. Sun, J. Cao, Y . Jiang, Z. Yuan, S. Bai, K. Kitani, and P. Luo, “Dancetrack: Multi-object tracking in uniform appearance and diverse motion,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 20 993–21 002

    2022

  34. [42]

    Detection and tracking meet drones challenge,

    P. Zhu, L. Wen, D. Du, X. Bian, H. Fan, Q. Hu, and H. Ling, “Detection and tracking meet drones challenge,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 11, pp. 7380–7399, 2021

    2021

  35. [43]

    A density-based algorithm for discovering clusters in large spatial databases with noise,

    M. Ester, H.-P. Kriegel, J. Sander, X. Xuet al., “A density-based algorithm for discovering clusters in large spatial databases with noise,” inkdd, vol. 96, no. 34, 1996, pp. 226–231

    1996

  36. [44]

    Objects as points,

    X. Zhou, D. Wang, and P. Kr ¨ahenb¨uhl, “Objects as points,”arXiv preprint arXiv:1904.07850, 2019

    1904 arXiv

  37. [45]

    Deformable detr: Deformable transformers for end-to-end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable detr: Deformable transformers for end-to-end object detection,”arXiv preprint arXiv:2010.04159, 2020

    2010 arXiv

  38. [46]

    Yolox: Exceeding yolo series in 2021,

    Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “Yolox: Exceeding yolo series in 2021,”arXiv preprint arXiv:2107.08430, 2021

    2021 arXiv

  39. [47]

    Yolov9: Learning what you want to learn using programmable gradient information,

    C.-Y . Wang, I.-H. Yeh, and H.-Y . Mark Liao, “Yolov9: Learning what you want to learn using programmable gradient information,” inEu- ropean conference on computer vision. Springer Nature Switzerland Cham, 2024, pp. 1–21

    2024

  40. [48]

    Yolov10: Real-time end-to-end object detection,

    A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Hanet al., “Yolov10: Real-time end-to-end object detection,”Advances in Neural Informa- tion Processing Systems, vol. 37, pp. 107 984–108 011, 2025

    2025

  41. [49]

    Spatial attention fusion for obstacle detection using mmwave radar and vision sensor,

    S. Chang, Y . Zhang, F. Zhang, X. Zhao, S. Huang, Z. Feng, and Z. Wei, “Spatial attention fusion for obstacle detection using mmwave radar and vision sensor,”Sensors, vol. 20, no. 4, p. 956, 2020

    2020

  42. [50]

    A deep learning-based radar and camera sensor fusion architecture for object detection,

    F. Nobis, M. Geisslinger, M. Weber, J. Betz, and M. Lienkamp, “A deep learning-based radar and camera sensor fusion architecture for object detection,” in2019 Sensor Data Fusion: Trends, Solutions, Applications (SDF). IEEE, 2019, pp. 1–7

    2019

  43. [51]

    Strongsort: Make deepsort great again,

    Y . Du, Z. Zhao, Y . Song, Y . Zhao, F. Su, T. Gong, and H. Meng, “Strongsort: Make deepsort great again,”IEEE Transactions on Mul- timedia, vol. 25, pp. 8725–8737, 2023

    2023

  44. [52]

    Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification,

    G. Maggiolino, A. Ahmad, J. Cao, and K. Kitani, “Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification,” in2023 IEEE International conference on image processing (ICIP). IEEE, 2023, pp. 3025–3029

    2023

  45. [53]

    Hybrid-sort: Weak cues matter for online multi-object tracking,

    M. Yang, G. Han, B. Yan, W. Zhang, J. Qi, H. Lu, and D. Wang, “Hybrid-sort: Weak cues matter for online multi-object tracking,” in Proceedings of the AAAI conference on artificial intelligence, vol. 38, no. 7, 2024, pp. 6504–6512

    2024

  46. [2023]

    Available: https://github.com/ultralytics/ultralytics

    [Online]. Available: https://github.com/ultralytics/ultralytics

Pith tools

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