Pith. sign in

REVIEW 4 major objections 5 minor 41 references

Motion Estimation for Multi-Object Tracking using KalmanNet with Semantic-Independent Encoding

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a learning-aided Kalman filter with a Semantic-Independent Encoder outperforms the classic Kalman filter and prior learned filters for motion estimation in multi-object tracking, reporting roughly 40% higher mean avera

desk verdict A credible incremental improvement with a useful benchmark and code; the MOT-level claims outrun the evidence. read the letter →

arxiv 2509.11323 v1 pith:UMYNMCBW submitted 2025-09-14 cs.CV cs.AI

Jian Song , Wei Mei , Yunfeng Xu , Qiang Fu , Renke Kou , Lina Bu , Yucheng Long This is my paper · ORCID
classification cs.CVcs.AI
keywords multi-objecttrackingmotionestimationKalmanfilterNetlearning-aidedfilteringsemantic-independentencodertracking-by-detection
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 argues that the standard Kalman filter used in multi-object tracking fails when object motion is non-linear or when the filter's noise parameters are mismatched, and that a learning-aided filter can do better. Its proposed SIKNet replaces the Kalman gain with a network that first encodes state-vector elements with the same semantics together, using a 1D convolution along the element dimension, then merges across heterogeneous elements with a fully connected layer. On a semi-simulated benchmark built from public MOT datasets, SIKNet reports higher recall than the classic KF and than KalmanNet and Split-KalmanNet, and improves HOTA when integrated into ByteTrack on DanceTrack and SoccerNet. The paper also provides an open-source framework for reproducible comparison of these motion models.

What carries the argument

The Semantic-Independent Encoder (SIE): a 1D convolution with kernel size 1 applied along the row (semantic) dimension of stacked state vectors, followed by a nonlinear activation, adaptive pooling, and a fully-connected layer. It processes same-semantic elements (e.g., positions across time) independently before mixing different semantics, which reduces the scale mismatch that destabilizes training. The network then tracks the state covariance P and the inverse innovation covariance S^-1 through separate recurrent modules to compute the Kalman gain.

What would settle it

Replace the semi-simulated measurements with actual detector outputs from DanceTrack or MOT17 and run ByteTrack+SIKNet against ByteTrack+KF without oracle detections; if the HOTA gap shrinks to near zero or the KF wins, the central claim fails in the setting it targets.

Watch

Extended reading notes

Core claim

The central claim is that decoupling the embedding of same-semantic state elements from cross-semantic mixing improves both the accuracy and training stability of learned Kalman gains. In their experiments, SIKNet exceeds the traditional KF by about 40% in mean average recall and surpasses existing learning-aided filters by about 6%, with the largest gains at high noise levels and on non-stationary object categories. Integrated with the ByteTrack association algorithm, SIKNet raises HOTA from 49.95 (KF) to 56.19 on DanceTrack and from 72.30 to 76.17 on SoccerNet using oracle detections, while also reducing identity switches.

Load-bearing premise

The evaluation assumes that real detection noise is zero-mean Gaussian added to ground-truth boxes and that integrated tracking can be benchmarked with oracle detections; if actual detectors produce false positives, missed detections, or correlated bias, the reported gains over the Kalman filter may not transfer to real tracking pipelines.

Editorial extensions

If this is right

  • If the claims hold, learned motion models can replace the KF in trackers without per-sequence tuning of noise parameters, which is especially valuable for non-stationary motion like dancers or players.
  • The reported gains grow with measurement noise, so the method is most promising in low-quality detection regimes where the KF's fixed covariance assumptions are most violated.
  • The semi-simulated evaluation protocol lets researchers isolate the motion estimation module from data association, making module-level comparisons cleaner.
  • SIKNet trains with standard backpropagation through time, whereas KalmanNet required truncated BPTT, which could lower training complexity.
  • On the integrated tracker, the improvement over the best prior learned filter is smaller (about 1.5–3 HOTA points) but consistent across nearly all metrics, suggesting the gains are real but not drastic at the tracker level.

Reading between the lines

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

  • Editorial inference: The reported 40% margin over KF comes from semi-simulated measurements with zero-mean Gaussian noise; real detectors also produce false positives, missed detections, and spatially correlated biases, so the margin could shrink or disappear in end-to-end tracking without oracle detections.
  • Editorial inference: Because the SIE separates semantics by construction, the architecture could transfer to other state-estimation problems where state vectors mix heterogeneous scales, such as 3D multi-object tracking or navigation fusion.
  • Editorial inference: The paper fixes the KF hyperparameters at common defaults; a per-sequence or per-category tuned KF might close some of the gap, so the comparison may overstate the advantage of learning.
  • Editorial inference: The integrated tracker experiments use oracle detections, which are cleaner than realistic detector outputs; a test on the official DanceTrack or MOT17 test sets with a real detector (e.g., YOLOX) would be the natural next validation step.
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. This paper proposes SIKNet, a KalmanNet-style learning-aided filter for the motion estimation module of multi-object tracking. The main contribution is a Semantic-Independent Encoder (SIE) that embeds the input state/measurement vector stack using a 1D convolution along the homogeneous-semantic dimension followed by a fully connected layer, in contrast to the direct concatenation used in KalmanNet. The filter is trained on a semi-simulated benchmark built by adding Gaussian noise to ground-truth box trajectories from MOT17, MOT20, SoccerNet, and DanceTrack, and is evaluated with recall metrics (Re50, Re75, AR) against KF, KalmanNet, and Split-KalmanNet. In a second set of experiments, the learned filters replace the KF inside the BYTE tracker and are evaluated on DanceTrack and SoccerNet validation/test sets with oracle detections. The paper reports consistent gains: roughly 6% mAR over KNet/SKNet and 40% over KF on the synthetic benchmark, and higher HOTA on DanceTrack (56.19 vs 54.60/49.95) and SoccerNet (76.17 vs 75.49/72.30). Code is released through two repositories.

Significance. If the results hold, the contribution is useful: a simple architectural change (SIE) improves training stability and accuracy of learned Kalman gains for box-state motion models, and the unified open-source FilterNet framework enables reproducible comparison. The use of temporal splits and the mismatched-noise experiments are commendable. However, the evaluation does not yet establish the claimed real-MOT superiority: the synthetic measurements are Gaussian with diagonal covariance, and the integrated tracker uses oracle detections. The relative gain over existing learned filters also lacks repeated-seed variance and an ablation isolating SIE from the other changes (input features, loss, training schedule). These are fixable, but they are necessary before the central claim is fully supported.

major comments (4)
  1. [Section 5.2, Implementation details and Table 3] The integrated experiments use oracle detections, as stated at the end of Section 5.2 ('DanceTrack and SoccerNet use oracle detections'). The improvements in HOTA/AssA/MOTA/IDs are therefore measured under ideal detection. Real MOT pipelines face false positives, missed detections, and spatially correlated detector noise; learned filters may not transfer under those conditions. The conclusion itself concedes that SIKNet has not been integrated into a complete MOT framework. Please either evaluate with the datasets' public detection outputs, or add a detection-noise model with false positives/misses to the semi-simulated data, and qualify the abstract/conclusion accordingly.
  2. [Section 3.2, Eq. (21) and Section 5.1] The semi-simulated measurements are generated by adding zero-mean Gaussian noise with diagonal covariance R_t (Eqs. 14-15 and 21) to ground-truth boxes. This is exactly the measurement model assumed by the KF baseline. The benchmark therefore measures robustness to process-model mismatch only, not to realistic detector error structure such as outliers, spatially correlated biases, or false positives/missed detections. The claim of 'superior robustness' for MOT is not supported by this setup. Please include experiments with non-Gaussian noise, outliers, or detector-residual statistics, or restrict the robustness claim to the synthetic Gaussian scenario.
  3. [Section 5.1.1 and Table 1] All tables report single runs. The central ~6% mAR margin over KNet/SKNet in Table 1 could be within run-to-run variance; the HOTA differences in Table 3 (e.g., 56.19 vs 54.60) are also single numbers. Moreover, KNet and SKNet are trained with the hyperparameters from their original papers, while SIKNet is trained with a tuned cosine-annealing schedule (50 epochs, batch size 32). Please report mean±std over at least 3-5 seeds and use a common hyperparameter/search budget for all learned filters, or show that the gain is robust to these choices.
  4. [Section 4 and Section 5.1] SIKNet differs from SKNet in several respects: the SIE encoder, the additional original-feature groups Z_in^3 and Z_in^4 (Eqs. 24-25), the Smooth L1 loss (Eqs. 27-28), and the training schedule. The paper attributes the gains to SIE, but no ablation isolates this component. Add experiments with at least SKNet+SIE and SIKNet without SIE, using the same input features and loss, to verify the architectural contribution of SIE.
minor comments (5)
  1. [Table 1/2] The table formatting in the arXiv text is ambiguous: values such as '0.63890.71710.6067' appear to have missing or concatenated columns. Please ensure each cell is clearly separated and the column structure matches the header.
  2. [Section 2.1.1 vs Eq. (21)] The system model defines e_t and v_t as zero-mean Gaussian with unknown covariance, while Eq. (21) generates v_t from a known R_t for the semi-simulated data. Clarify this notational tension.
  3. [Eq. (17) and state mode] IoU is computed on bounding-box coordinates, but the state x_t also contains velocities. Clarify the conversion from state estimate to bbox for the IoU metric.
  4. [Fig. 3(b) and Section 4.1] The description of the 1D convolution is confusing: the text says '1D convolution with a 1×N kernel' and 'kernel size 1' at the same time. Specify the convolution axis and kernel shape explicitly so the reader can reproduce the SIE.
  5. [Abstract and Conclusion] The conclusion's caveat that SIKNet has not been fully integrated into complete MOT frameworks should be reflected in the abstract's claim of 'outperforms ... for MOT'. Please qualify the scope of the claim.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: SIKNet is evaluated on held-out temporal splits and real-data tracking benchmarks; the only self-citation is a non-load-bearing training implementation detail.

full rationale

The paper's derivation and evaluation chain are not circular. The central claim — SIKNet outperforms KF and existing learning-aided filters — is supported by held-out test evaluation. In Sec. 5.1, the semi-simulated dataset is generated by Eq. (21), y_t^b = x_t^b + v_t with v_t ~ N(0,R_t), where R_t is defined in Eqs. (14)-(15). The KF baseline is deliberately given the same R_t (α_p equal to the test set's α_p), so the learned filters are not exploiting a deliberately mismatched KF. Training and test sets are disjoint temporal halves of each sequence: 'Each semi-simulated dataset was divided into training and test sets by splitting each sequence into two equal temporal segments,' and all Table 1/2 results are on those test sets. The mismatched-noise experiments in Sec. 5.1.4 also evaluate on noise levels different from those used in training. For the integrated tracking experiments, Sec. 5.2 uses the real DanceTrack and SoccerNet benchmarks with the BYTE association algorithm, reporting HOTA/MOTA/IDF1/IDs; these are externally specified metrics and datasets, not numbers reconstructed from fitted training data. The only author self-citation is [35], used as an implementation detail for training the KNet baseline: 'the alternative truncated back-propagation through time algorithm proposed in [35] is employed to ensure successful convergence.' This is not the source of SIKNet's gains and is not a load-bearing premise of the paper's novelty. The paper's own limitation statement — 'SIKNet has not yet been fully integrated into complete MOT frameworks for end-to-end training with detection components' — narrows the scope of the claim but is not a circular step. No equation is defined in terms of the target conclusion, no fitted parameter is relabeled as a prediction, and no uniqueness theorem is imported from the authors' prior work. The score of 2 reflects only the minor, non-load-bearing self-citation; there is no substantive circularity.

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

The central claim rests on a benchmark where measurement noise is synthetic Gaussian with covariance proportional to box dimensions, and on a training objective that directly supervises the estimated box against GT. The free parameters alpha_p and alpha_v define the noise levels and the KF baseline's assumed statistics. The temporal split assumption and the CV model prior are structural premises that could affect the reported margins if violated.

free parameters (2)
  • alpha_p (measurement noise factor) = 0.05, 0.1, 0.2, 0.4
    Controls the synthetic measurement noise level in Eq (21) and is also the KF's assumed noise factor. Matched experiments train and test on the same alpha_p, which is favorable to learned filters.
  • alpha_v (KF process noise velocity factor) = 0.00625
    Adopted common value for KF's process noise covariance Q in Eq (9); not tuned per dataset or category.
assumptions (4)
  • domain assumption Detection measurements are generated by adding zero-mean Gaussian noise with covariance R_t as in Eqs (14)-(15) to ground-truth boxes.
    Underlies the semi-simulated dataset (Section 3.2, Eq 21); real detector noise may be non-Gaussian, correlated, or include false positives and misses, which would change the comparison.
  • domain assumption Linear constant-velocity model with known F and H (Eqs 6-7, 12-13) describes the motion prior used by KF and inherited by the learned filters.
    The filters predict using the same transition and measurement matrices; only the Kalman gain is learned. If the CV model is a poor prior for a category, SIKNet's gains could vanish.
  • domain assumption Temporal splitting of each sequence into two equal halves yields a valid train/test split.
    Adjacent frames are temporally correlated; splitting by trajectory segment may let the model exploit trajectory continuity, and the two halves may have different motion statistics.
  • standard math Smooth L1 loss between estimated and GT boxes is a valid training objective for motion estimation.
    Standard supervised regression loss; the paper assumes minimizing it improves IoU-based recall and downstream tracking metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Motion Estimation for Multi-Object Tracking using KalmanNet with Semantic-Independent Encoding." pith.science (2026). https://pith.science/paper/UMYNMCBW

@misc{pith2026250911323,
  author       = {Pith},
  title        = {Pith review of: Motion Estimation for Multi-Object Tracking using KalmanNet with Semantic-Independent Encoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UMYNMCBW}},
  note         = {Machine review of arXiv:2509.11323}
}
read the original abstract

Motion estimation is a crucial component in multi-object tracking (MOT). It predicts the trajectory of objects by analyzing the changes in their positions in consecutive frames of images, reducing tracking failures and identity switches. The Kalman filter (KF) based on the linear constant-velocity model is one of the most commonly used methods in MOT. However, it may yield unsatisfactory results when KF's parameters are mismatched and objects move in non-stationary. In this work, we utilize the learning-aided filter to handle the motion estimation of MOT. In particular, we propose a novel method named Semantic-Independent KalmanNet (SIKNet), which encodes the state vector (the input feature) using a Semantic-Independent Encoder (SIE) by two steps. First, the SIE uses a 1D convolution with a kernel size of 1, which convolves along the dimension of homogeneous-semantic elements across different state vectors to encode independent semantic information. Then it employs a fully-connected layer and a nonlinear activation layer to encode nonlinear and cross-dependency information between heterogeneous-semantic elements. To independently evaluate the performance of the motion estimation module in MOT, we constructed a large-scale semi-simulated dataset from several open-source MOT datasets. Experimental results demonstrate that the proposed SIKNet outperforms the traditional KF and achieves superior robustness and accuracy than existing learning-aided filters. The code is available at (https://github.com/SongJgit/filternet and https://github.com/SongJgit/TBDTracker).

Figures

Figures reproduced from arXiv: 2509.11323 by the authors.

Figure 1
Figure 1. IoU on adjacent frames. Compared to MOT17 and MOT20, SoccerNet and DanceTrack have lower IoU values, which indicates that the objects in these datasets have more complex motion patterns. 3. Semi-simulated datasets The large-scale semi-simulated datasets are built upon the commonly used open-source MOT datasets. 3.1. Base datasets MOT17 [21] and MOT20 [22]. These are large datasets used to evaluate the performance of… view at source ↗
Figure 2
Figure 2. Selected semi-simulated sequences on DanceTrack video. Green solid lines denote ground-truth bounding-box, and the red dashed lines represent the noisy bounding-box 𝒚𝑡 derived from Eq. (21), and the IoU metric quantifies their spatial overlap. (a) Vertical Stacking in KNet (b) Semantic-Independent Encoder in SIKNet Stacking FC … M×N M× 𝐶𝐶 1D Conv(k=1, s=1) FC Activation AvgPooling 1D Conv … 1z 2 z in Z emb Z [PITH_… view at source ↗
Figure 3
Figure 3. Feature embedding in learning-aided filter. (a) Vertical stacking in KNet. (b) Semantic-Independent Encoder in SIKNet. 𝒁in is composed of 𝑁 𝑀-dimension state (column) vectors 𝒛𝑗 ∈ ℝ𝑀×1 (𝑗 = 1, 2, … , 𝑁). Among them, elements in the same row vector of 𝒁in = [𝒛1 , 𝒛2 , …, 𝒛𝑁 ], e.g., 𝒛1 = 𝒙̂ 𝑡−1|𝑡−1 and 𝒛2 = 𝒙̂ 𝑡|𝑡−1, have the same or similar semantic information. By performing a 1D convolution with a convolution kern… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Block diagram of the proposed Semantic-Independent KalmanNet 0.05 0.1 0.2 0.4 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 BoxMode(XYAH) BoxMode(XYWH) Observation KF KNet SKNet Ours αp Mean Average Recall [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: Line graph of mean average recall corresponding to [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Average recall on different categories. (a) 𝛼𝑝 = 0.2. (b) 𝛼𝑝 = 0.4. 0.05 0.1 0.2 0.4 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 BoxMode(XYAH) BoxMode(XYWH) KF KNet SKNet Ours αp Mean Average Recall (a) Trained on 𝛼𝑝 = 0.05 0.05 0.1 0.2 0.4 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 BoxM…
Figure 7
Figure 7. Figure 7: Mean average recall for scenarios where training and testing noise parameters are mismatched. (a) Trained on 𝛼𝑝 = 0.05. (b) Trained on 𝛼𝑝 = 0.4. : Preprint submitted to Elsevier Page 15 of 12 [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Visual results of selected sequences obtained by different motion estimation methods on the DanceTrack in XYAH mode. Training and test sets of the sequences are generated using 𝛼𝑝 = 0.2. The frame interval of the sequences is 5 and the images were cropped to focus on t…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 2 canonical work pages

  1. [1]

    X. Weng, J. Wang, D. Held, K. Kitani, 3d multi-object tracking: A baseline and new evaluation metrics, in: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS, IEEE, 2020, pp. 10359–10366.doi:10.1109/IROS45743.2020.9341164

    arXiv 2020

  2. [2]

    14227–14233.doi:10.1109/ICRA48506.2021.9561754

    H.-K.Chiu,J.Li,R.Ambrus,J.Bohg,Probabilistic3dmulti-modal,multi-objecttrackingforautonomousdriving,in:2021IEEEInternational Conference on Robotics and Automation, ICRA, IEEE, 2021, pp. 14227–14233.doi:10.1109/ICRA48506.2021.9561754

    arXiv 2021

  3. [3]

    Cioppa, S

    A. Cioppa, S. Giancola, A. Deliege, L. Kang, X. Zhou, Z. Cheng, B. Ghanem, M. Van Droogenbroeck, Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos, in: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPRW, IEEE, 2022, pp. 3490–3501.doi:10.1109/CVPRW56347.2022.00393

    arXiv 2022

  4. [4]

    Girshick, Fast r-cnn, in: 2015 IEEE International Conference on Computer Vision, ICCV, IEEE, 2015, pp

    R. Girshick, Fast r-cnn, in: 2015 IEEE International Conference on Computer Vision, ICCV, IEEE, 2015, pp. 1440–1448.doi:10.1109/ ICCV.2015.169

    2015

  5. [5]

    S.Ren,K.He,R.Girshick,J.Sun,Fasterr-cnn:Towardsreal-timeobjectdetectionwithregionproposalnetworks,IEEETrans.PatternAnal. Mach. Intell. 39 (6) (2017) 1137–1149.doi:10.1109/TPAMI.2016.2577031

    arXiv 2017

  6. [6]

    Ciaparrone, F

    G. Ciaparrone, F. Luque Sánchez, S. Tabik, L. Troiano, R. Tagliaferri, F. Herrera, Deep learning in video multi-object tracking: A survey, Neurocomputing 381 (2020) 61–88.doi:10.1016/j.neucom.2019.11.023

    work page doi:10.1016/j.neucom.2019.11.023 2020

  7. [7]

    2107.08430

    Z.Ge,S.Liu,F.Wang,Z.Li,J.Sun,Yolox:Exceedingyoloseriesin2021,2021,arXivpreprintarXiv:2107.08430.doi:10.48550/arXiv. 2107.08430

    arXiv doi:10.48550/arxiv.2107.08430 2021

  8. [8]

    Bewley, Z

    A. Bewley, Z. Ge, L. Ott, F. Ramos, B. Upcroft, Simple online and realtime tracking, in: 2016 IEEE International Conference on Image Processing, ICIP, IEEE, 2016, pp. 3464–3468.doi:10.1109/ICIP.2016.7533003

    arXiv 2016

Show all 41 references
  1. [9]

    Wojke, A

    N. Wojke, A. Bewley, D. Paulus, Simple online and realtime tracking with a deep association metric, in: 2017 IEEE International Conference on Image Processing, ICIP, IEEE, 2017, pp. 3645–3649.doi:10.1109/ICIP.2017.8296962

    2017

  2. [10]

    Zhang, C

    Y. Zhang, C. Wang, X. Wang, W. Zeng, W. Liu, Fairmot: On the fairness of detection and re-identification in multiple object tracking, Int. J. Comput. Vis. 129 (11) (2021) 3069–3087.doi:10.1007/s11263-021-01513-4

    2021 doi

  3. [11]

    Y. Du, J. Wan, Y. Zhao, B. Zhang, Z. Tong, J. Dong, Giaotracker: A comprehensive framework for mcmot with global information and optimizing strategies in visdrone 2021, in: 2021 IEEE/CVF International Conference on Computer Vision Workshops, ICCVW, IEEE, 2021, pp. 2809–2819.do...

    2021

  4. [12]

    Zhang, P

    Y. Zhang, P. Sun, Y. Jiang, D. Yu, F. Weng, Z. Yuan, P. Luo, W. Liu, X. Wang, Bytetrack: Multi-object tracking by associating every detection box, in: European Conference on Computer Vision, ECCV, Springer Nature Switzerland, 2022, pp. 1–21.doi:10.1007/ 978-3-031-20047-2_1

    2022

  5. [13]

    F. Yang, S. Odashima, S. Masui, S. Jiang, Hard to track objects with irregular motions and similar appearances? make it easier by buffering the matching space, in: 2023 IEEE/CVF Winter Conference on Applications of Computer Vision, WACV, IEEE, 2023, pp. 4788–4797. doi:10.1109/...

    2023

  6. [14]

    99 (2024) 104064.doi:10.1016/j.jvcir.2024.104064

    K.Qi,W.Xu,W.Chen,X.Tao,P.Chen,Multipleobjecttrackingwithsegmentationandinteractivemultiplemodel,J.VisualCommun.Image Represent. 99 (2024) 104064.doi:10.1016/j.jvcir.2024.104064

    2024

  7. [15]

    X. Wang, Z. Sun, A. Chehri, G. Jeon, Y. Song, Deep learning and multi-modal fusion for real-time multi-object tracking: Algorithms, challenges, datasets, and comparative study, Inf. Fusion 105 (2024) 102247.doi:10.1016/j.inffus.2024.102247

    2024

  8. [16]

    H.W.Kuhn,Thehungarianmethodfortheassignmentproblem,Nav.Res.Logist.Q.2(1-2)(1955)83–97.doi:10.1002/nav.3800020109

    1955 doi

  9. [17]

    P. Zhu, L. Wen, D. Du, X. Bian, H. Fan, Q. Hu, H. Ling, Detection and tracking meet drones challenge, IEEE Trans. Pattern Anal. Mach. Intell. 44 (11) (2022) 7380–7399.doi:10.1109/TPAMI.2021.3119563

    2022

  10. [18]

    K. Yi, K. Luo, X. Luo, J. Huang, H. Wu, R. Hu, W. Hao, Ucmctrack: Multi-object tracking with uniform camera motion compensation, in: Proceedings of the AAAI Conference on Artificial Intelligence, AAAI, AAAI, 2024, pp. 6702–6710.doi:10.1609/aaai.v38i7.28493

    2024 doi

  11. [19]

    W. Lv, Y. Huang, N. Zhang, R.-S. Lin, M. Han, D. Zeng, Diffmot: A real-time diffusion-based multiple object tracker with non-linear prediction, in: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, IEEE, 2024, pp. 19321–19330.doi: 10.1109/CVPR52733.2024.01828

    2024

  12. [20]

    Y. Wang, Y. Qing, K. Huang, C. Dang, Z. Wu, Preformer mot: A transformer-based approach for multi-object tracking with global trajectory prediction, Fundam. Res. (2025).doi:10.1016/j.fmre.2024.06.015

    2025 doi

  13. [21]

    doi:10.48550/arXiv.1603.00831

    A.Milan,L.Leal-Taixe,I.Reid,S.Roth,K.Schindler,Mot16:Abenchmarkformulti-objecttracking,2016,arXivpreprintarXiv.1603.00831. doi:10.48550/arXiv.1603.00831

    2016 doi arXiv

  14. [22]

    P.Dendorfer,H.Rezatofighi,A.Milan,J.Shi,D.Cremers,I.Reid,S.Roth,K.Schindler,L.Leal-Taixé,Mot20:Abenchmarkformultiobject tracking in crowded scenes, 2020, arXiv preprint arXiv:2003.09003.doi:10.48550/arXiv.2003.09003

    2020 doi arXiv

  15. [23]

    Cioppa, A

    A. Cioppa, A. Deliège, F. Magera, S. Giancola, O. Barnich, B. Ghanem, M. Van Droogenbroeck, Camera calibration and player localization insoccernet-v2andinvestigationoftheirrepresentationsforactionspotting,in:2021IEEE/CVFConferenceonComputerVisionandPattern Recognition Workshop...

    2021

  16. [24]

    P. Sun, J. Cao, Y. Jiang, Z. Yuan, S. Bai, K. Kitani, P. Luo, Dancetrack: Multi-object tracking in uniform appearance and diverse motion, in: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, IEEE, 2022, pp. 20961–20970.doi:10.1109/CVPR52688. 2022.02032

    2022

  17. [25]

    A. Patil,S. Malla, H.Gang, Y.-T. Chen,The h3d datasetfor full-surround 3dmulti-object detection andtracking in crowdedurban scenes, in: 2019 International Conference on Robotics and Automation, ICRA, IEEE, 2019, pp. 9552–9557.doi:10.1109/ICRA.2019.8793925

    2019

  18. [26]

    P. Liu, Z. Duan, An imm-enabled adaptive 3d multi-object tracker for autonomous driving, in: 2021 IEEE 24th International Conference on Information Fusion, FUSION, IEEE, 2021, pp. 1–8.doi:10.23919/FUSION49465.2021.9626913

    2021

  19. [27]

    N.Shlezinger,Y.C.Eldar,S.P.Boyd,Model-baseddeeplearning:Ontheintersectionofdeeplearningandoptimization,IEEEaccess:pract. innov. open solut. 10 (2022) 115384–115398.doi:10.1109/ACCESS.2022.3218802

    2022

  20. [28]

    2023.3247480

    N.Shlezinger,J.Whang,Y.C.Eldar,A.G.Dimakis,Model-baseddeeplearning,Proc.IEEE111(5)(2023)465–499.doi:10.1109/JPROC. 2023.3247480

    2023

  21. [29]

    Revach, N

    G. Revach, N. Shlezinger, R. J. G. Van Sloun, Y. C. Eldar, Kalmannet: Data-driven kalman filtering, in: 2021 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP, IEEE, 2021, pp. 3905–3909.doi:10.1109/ICASSP39728.2021.9413750

    2021

  22. [30]

    Signal Process

    G.Revach,N.Shlezinger,X.Ni,A.L.Escoriza,R.J.G.VanSloun,Y.C.Eldar,Kalmannet:Neuralnetworkaidedkalmanfilteringforpartially known dynamics, IEEE Trans. Signal Process. 70 (2022) 1532–1547.doi:10.1109/TSP.2022.3158588

    2022

  23. [31]

    G.Choi,J.Park,N.Shlezinger,Y.C.Eldar,N.Lee,Split-kalmannet:Arobustmodel-baseddeeplearningapproachforstateestimation,IEEE Trans. Veh. Technol. 72 (9) (2023) 12326–12331.doi:10.1109/TVT.2023.3270353

    2023

  24. [32]

    Y. Du, Z. Zhao, Y. Song, Y. Zhao, F. Su, T. Gong, H. Meng, Strongsort: Make deepsort great again, IEEE Trans. Multimedia 25 (2023) 8725–8737.doi:10.1109/TMM.2023.3240881

    2023

  25. [33]

    J. Cao, J. Pang, X. Weng, R. Khirodkar, K. Kitani, Observation-centric sort: Rethinking sort for robust multi-object tracking, in: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, IEEE, 2023, pp. 9686–9696.doi:10.1109/CVPR52729. 2023.00934

    2023

  26. [34]

    Aharon, R

    N. Aharon, R. Orfaig, B.-Z. Bobrovsky, Bot-sort: Robust associations multi-pedestrian tracking, Jul. 2022, arXiv preprint arXiv:2206.14651. arXiv:2206.14651,doi:10.48550/arXiv.2206.14651

    2022 doi arXiv

  27. [35]

    J. Song, W. Mei, Y. Xu, Q. Fu, L. Bu, Practical implementation of kalmannet for accurate data fusion in integrated navigation, IEEE Signal Process. Lett. 31 (2024) 1890–1894.doi:10.1109/LSP.2024.3431443

    2024

  28. [36]

    Donghao, W

    L. Donghao, W. Xue, Moderntcn: A modern pure convolution structure for general time series analysis, in: The Twelfth International Conference on Learning Representations, ICLR, 2024

    2024

  29. [37]

    K. Kang, H. Li, J. Yan, X. Zeng, B. Yang, T. Xiao, C. Zhang, Z. Wang, R. Wang, X. Wang, W. Ouyang, T-cnn: Tubelets with convolutional neural networks for object detection from videos, IEEE Trans. Circuits Syst. Video Technol. 28 (10) (2018) 2896–2907.doi:10.1109/ TCSVT.2017.2736553

    2018

  30. [38]

    S. Bai, J. Z. Kolter, V. Koltun, An empirical evaluation of generic convolutional and recurrent networks for sequence modeling, 2018, arXiv. doi:10.48550/arXiv.1803.01271

    2018 doi arXiv

  31. [39]

    Hosang, R

    J. Hosang, R. Benenson, P. Dollár, B. Schiele, What makes for effective detection proposals?, IEEE Trans. Pattern Anal. Mach. Intell. 38 (4) (2016) 814–830.doi:10.1109/TPAMI.2015.2465908

    2016

  32. [40]

    J.Luiten,A.O ˘sep,P.Dendorfer,P.Torr,A.Geiger,L.Leal-Taixé,B.Leibe,Hota:Ahigherordermetricforevaluatingmulti-objecttracking, Int. J. Comput. Vision 129 (2) (2021) 548–578.doi:10.1007/s11263-020-01375-2

    2021 doi

  33. [41]

    Bernardin, R

    K. Bernardin, R. Stiefelhagen, Evaluating multiple object tracking performance: The clear mot metrics, EURASIP J. Image Video Process. 2008 (2008) 1–10.doi:10.1155/2008/246309. :Preprint submitted to Elsevier Page 12 of 12 (a)The sequence are generated using𝛼𝑝=0.05 (b)The sequ...

    2008 doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.