Pith. sign in

REVIEW 3 major objections 5 minor 26 references

GenTrack3: Hybrid Stochastic-Deterministic Online Multi-Object Tracking with Cluster-Aware Association

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A cluster-aware association step claims to preserve global tracking assignments while shrinking the cost matrix.

desk verdict The cluster-aware matching idea is new and the code is real, but the exact-global-optimality claim is false in general and the paper itself hedges to 'near-optimal' — a correctable flaw, not a fatal one. read the letter →

arxiv 2608.09581 v1 pith:P5Y7EC3T submitted 2026-08-10 cs.CV cs.RO

Toan Van Nguyen , Rasmus G. K. Christiansen , Dirk Kraft , Leon Bodenhagen This is my paper · ORCID
classification cs.CVcs.RO
keywords multi-objecttrackingdataassociationparticlefilteringswarmoptimizationHungarianalgorithmcluster-awarematchingMOTChallengescalable
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 proposes an online multi-object tracking framework that combines stochastic particle-based state estimation with deterministic data association. Its central new idea is a cluster-aware association step that partitions the global track-detection cost matrix into smaller local matrices, resolves detections that fall into several clusters by a shared-detection reassignment, and still claims to preserve global assignment quality while reducing space complexity from $O(T \cdot D)$ to $\sum_l O(T_l \cdot D_l)$. If this holds, tracking many targets at once becomes less expensive without giving up identity-consistent matches, which matters for crowded scenes and for socially aware robot navigation. The paper reports competitive MOTA values, including 84.86 on MOT17-04 and 78.72 on MOT20-05 for the full GenTrack3 Super variant.

What carries the argument

The load-bearing mechanism is the cluster-aware track-detection association: a per-frame grouping in which an occupancy grid is built from previous target boxes, connected-component labeling forms clusters, a valid-pair mask $M^L$ blanks out spatially distant track-detection pairs, and a shared-detection step compares, for each detection lying in several clusters, the representative cost of its best track in each cluster before fixing the detection to one cluster and running local Hungarian assignments on the remaining submatrices. The paper claims this preserves the global optimum while changing the space complexity of the cost matrix from $O(T \cdot D)$ to $\sum_l O(T_l \cdot D_l)$.

What would settle it

Take the same tracks and detections of a MOT17-04 or MOT20-05 frame, compute the full cost matrix using the paper's Equation (5), run the exact Hungarian algorithm, and run the proposed cluster-aware pipeline on the same input; any frame where the pipeline's total assignment cost exceeds the exact Hungarian optimum falsifies the 'overall matching remains global' claim, and a two-cluster example with one shared detection should already be enough to exhibit such a gap.

Watch

Extended reading notes

Core claim

On its own terms, GenTrack3 establishes a four-stage track-detection association: cluster tracks and detections on a 2-D occupancy grid using connected-component labeling; compute a valid-pair mask and per-cluster cost matrices so implausible pairs are skipped; reassign each shared detection to the cluster whose best track has the lowest representative cost and zero out that detection's costs in all other clusters; then run a local Hungarian assignment in each cluster. The paper's key assertion is that, despite the local computation, 'the overall matching remains global through the pipeline,' so the partitioned procedure reproduces the global assignment while lowering the space complexity of cost computation and assignment from $O(T \cdot D)$ and $O(T \cdot D + T + D)$ to the corresponding per-cluster sums. The framework also modifies the weak-track update from earlier GenTrack versions so that only reliable neighbours, and their median motion, steer occluded tracks.

Load-bearing premise

The argument stands or falls on the claim that assigning each shared detection to the cluster containing its lowest-cost best track, and then excluding that detection from all other clusters, never changes the global Hungarian optimum; the paper asserts this in Section 2.2 without a proof, so if the premise fails the method is an approximation rather than an exact scalable matching.

Editorial extensions

If this is right

  • Matching cost space drops from the full $T \times D$ product to the sum of per-cluster products, and the valid-pair mask skips particle-based cost evaluation for spatially implausible pairs.
  • The same partitioning reduces the practical time of the assignment step, since each local Hungarian runs on a smaller submatrix instead of one global matrix.
  • Because a detection can belong to several clusters before the shared-detection step resolves it, the pipeline supports group tracking without a separate grouping module.
  • When stricter computation is needed, selecting the top $n$ tracks per cluster instead of all valid tracks yields a near-optimal matching while improving scalability.
  • Reported MOTA values on MOT17-04 and MOT20-05 show the full system remains competitive with the GenTrack and GenTrack2 baselines it extends.

Reading between the lines

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

  • The exact-global-optimality claim is not proven; the shared-detection rule picks the best single track per cluster and freezes the detection there, which can alter which track wins the detection when the two clusters' internal assignments interact, so the method is at best a strong heuristic unless a proof or counterexample bound is supplied.
  • A direct way to settle the claim is to compare the total assignment cost of the cluster-aware pipeline against the full Hungarian optimum on real MOT frames; any positive gap would quantify the approximation error.
  • The paper's own observation that regional clustering in dense scenes increases the number of shared detections suggests an adaptive grid resolution could balance per-cluster savings against shared-detection overhead.
  • Because the weak-track update uses reliable neighbours, the same cluster structure that cheapens association may constrain which neighbours a weak track can consult, and this interaction is not analysed in the paper.
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 / 5 minor

Summary. The paper presents GenTrack3, an online multi-object tracking framework that combines stochastic particle-based state estimation with deterministic data association. The central novelty is a cluster-aware track-detection matching stage: tracks and detections are grouped via connected components on an occupancy grid, a valid-pair mask restricts cost computation, detections shared across clusters are assigned to a single cluster by a representative-cost rule, and the Hungarian algorithm is run per cluster. The authors claim that this procedure preserves global track-detection matchings while reducing space complexity from O(T·D) to Σ_l O(T_l·D_l). Experiments are reported on one MOT17 sequence (MOT17-04) and one MOT20 sequence (MOT20-05), comparing against ten baselines; GenTrack3 Super reaches 84.86 MOTA on MOT17-04 and 78.72 MOTA on MOT20-05, and an ablation variant GenTrack2 Super+ (the previous pipeline with the new matching component) reaches 87.31 MOTA on MOT17-04. Reference implementations are provided on GitHub.

Significance. If the exact global-preservation claim were correct, the cluster-aware association would be a valuable scalability contribution for dense MOT. The paper does provide a correct complexity inequality for disjoint partitions, and the release of reference implementations with re-evaluable baselines is a real strength. However, the exactness claim is not proven and, as shown below, is false for the described shared-detection rule; the method is at best a near-optimal heuristic. This substantially reduces the significance of the theoretical contribution. The empirical evaluation is also thin: two sequences, no error bars, no runtime or memory measurements, and no sensitivity analysis for the many hand-set parameters. The paper is best positioned as an engineering/empirical contribution, but only after the overclaims are corrected and the evaluation is strengthened.

major comments (3)
  1. [Section 2.2, shared-detection processing] The claim that the proposed pipeline 'preserves global track-detection matchings' and that 'the overall matching remains global through the pipeline' is not established and is in fact false for the described procedure. The representative-cost rule fixes a shared detection in one cluster before running local Hungarians, and this can miss the global optimum. Consider two clusters A and B and one shared detection s with costs: A1-s=0.1, A1-x=0.5, A2-s=2.0, A2-x=1.0; B1-s=0.2, B1-y=1.5, B2-s=2.1, B2-y=0.8. Tracks A1 and B1 are both mutual best matches for s, with representative costs 0.1 and 0.2, so the rule assigns s to cluster A and forbids it in B. The local Hungarians then pay A1-s=0.1, A2-x=1.0, and B2-y=0.8, total 1.9. The true global optimum is A1-x=0.5, B1-s=0.2, B2-y=0.8, total 1.5. The mutual-best condition does not rescue the claim, because A1-s is mutual best in A. The method is therefore a heuristic approximation, not an exact scalable matching; this is also partially conceded later in the same section ('generally ensures global assignments', 'near-optimal solution').
  2. [Section 3, evaluations] The empirical evaluation is not sufficient to support the headline claims of scalability and preserved global assignment. Only one sequence is used per benchmark (MOT17-04 and MOT20-05), with no error bars, no repeated runs, and no measurements of wall-clock time, peak memory, or per-stage cost. The complexity argument in Section 2.2 is about space and time asymptotics, but the paper never reports actual computational savings. The GenTrack2 Super+ comparison is a useful ablation for the matching component, but it is not isolated from other pipeline differences (e.g., the modified weak-track update described in Section 2.1), and it is reported on a single sequence. Add runtime and memory measurements, full-benchmark results, and sensitivity analysis over the introduced parameters (grid cell size, inflation ratio ε, λ_D, and the shared-detection top-n choice).
  3. [Section 3, edge-case discussion and Section 4] The paper's own caveats undermine the central claim. In Section 3, the authors state that in dense scenes the occupancy-grid clustering can collapse into one cluster of 186 targets, so they resort to image-region clustering, which 'results in a higher number of shared detections, increasing the complexity of shared-detection processing and potentially reducing accuracy.' In Section 4, they conclude that the proposed track-to-detection method 'is suited for sparse environments' and that 'the global masking cost matrix provides a better balance between performance in sparse and dense scenarios.' This is inconsistent with the abstract and contribution list, which promise a scalable method that preserves global matchings. The authors should reframe the contribution as an approximate, sparsity-exploiting heuristic, and state clearly where it helps and where it does not.
minor comments (5)
  1. [Section 2.2, final paragraph] Typographical and wording issues: 'overal' should be 'overall'; the phrase 'the overal matching remains global through the pipeline' is vague and should be replaced by a precise statement about what is globally preserved (e.g., the assignment, the cost, or the matching).
  2. [Section 2.2, complexity paragraph] The notation 'w2_j + h2_j' in Eq. (12) should be written as w_j^2 + h_j^2 for clarity, and the symbols T_L, D_L are introduced without defining that L indexes a cluster; define the indexing consistently.
  3. [Section 2.1, Eq. (7)] The recovery trust ζ is defined as sign(ρ_re − f_w_g + Δ_e), which can take negative values, but the text does not explain whether a negative ζ is allowed and how it affects the penalty update; clarify the sign convention.
  4. [Section 2.1, Eq. (8)] The similarity threshold δ_d is used but not defined in the text; define it and state its range.
  5. [Section 3, baselines] The paper says the comparison uses 'identical detections and ground truths' but does not specify the detection source (e.g., public detections or the authors' own detector); state this explicitly for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the complexity reduction is self-contained arithmetic, and the global-optimality concern is a correctness issue, not a circular reduction.

full rationale

The claimed derivation chain is largely self-contained. The complexity improvement is an arithmetic identity: because final clusters partition tracks and detections, T = sum_l T_l and D = sum_l D_l, so T*D >= sum_l T_l*D_l; this does not depend on any fitted value or on the authors' prior results. The association method is specified in the paper's own equations (5), (6), and (12), and the shared-detection rule is described algorithmically rather than inherited. The global-optimality assertion in Section 2.2 is the weakest point, but it is a correctness gap, not circularity: the described single-best representative-cost rule can miss the global Hungarian optimum, and the paper itself hedges by saying the shared-detection processing 'generally ensures global assignments' and that 'a near-optimal solution can be obtained.' That is an unsupported equivalence, not a reduction of a prediction to its input. The heavy self-citations to [16,17] for particle sampling, PSO fitness, and weak-track updates are not load-bearing for the novel matching claim: the relevant equations are restated in the text, and the final system is evaluated on external MOT17/MOT20 benchmarks against multiple external trackers, so the empirical results are independently falsifiable. No specific step was found in which an output is defined as, or fitted to, its input by construction.

Assumptions & free parameters 13 free parameters · 6 assumptions · 0 invented entities

The central algorithm rests on roughly a dozen hand-set weights and thresholds, most with unreported numerical values, and many are inherited from the authors' own GenTrack/GenTrack2 papers. No new physical entities are introduced. The mathematical complexity reduction is self-contained given the clustering, but the global-optimality preservation is an unproven ad hoc assumption.

free parameters (13)
  • epsilon_X, epsilon_V (state and velocity perturbation scales, Eq. 1)
    Control how far particles explore each frame; values not reported, presumably tuned or inherited from GenTrack/GenTrack2.
  • lambda_X, lambda_V (motion model weights, Eq. 1)
    Constrained by lambda_X + lambda_V = 1; the split is not reported.
  • sigma_h, sigma_p, sigma_i (PSO fitness weights, Eq. 2)
    Sum to 1; individual values are not reported.
  • lambda_s, lambda_m (feature and motion fitness weights, Eq. 3)
    Sum to 1; values not reported.
  • xi_p, xi_V (social fitness weights, Eq. 4)
    Sum to 1; values not reported.
  • lambda_p, lambda_d, lambda_h (matching cost weights, Eq. 5)
    Sum to 1; values not reported.
  • sigma_u, sigma_d (IoU and distance weights, Eq. 6)
    Sum to 1; values not reported.
  • lambda_D (valid-mask range factor, Eq. 12)
    Controls the spatial range around each detection; value not reported.
  • grid cell size gx, gy = 5 pixels
    Hand-chosen grid resolution for the occupancy map; no sensitivity analysis is given.
  • epsilon inflation ratio for occupied grid cells
    Inflation factor in the grid occupancy update; value not reported.
  • max age (partial_max) and recovery trust (rho_re, eta) in Eq. 7
    Control weak-track lifetime and recovery; values not reported.
  • velocity suppression threshold tau_V in Eq. 9
    Suppresses weak-track updates below a speed threshold; value not reported.
  • velocity regression window H and frame step F
    Used for trend-based velocity estimation; values not reported.
assumptions (6)
  • standard math The Hungarian algorithm returns a minimum-cost one-to-one assignment.
    Used as the final assignment step within each cluster in Section 2.2.
  • domain assumption MCMC proposals and PSO optimization approximate each target's optimal state without full posterior sampling.
    Section 2.1 relies on this to replace full Bayesian filtering; no convergence guarantees are provided.
  • domain assumption The random motion model in Eq. (1) with bounded perturbations covers the relevant target dynamics.
    All particle generation uses this model; strongly non-linear or abrupt motions may violate the bounds.
  • ad hoc to paper Occupancy-grid clustering with shared-detection reassignment preserves the global Hungarian optimum.
    Asserted without proof in Section 2.2; this is the paper's weakest load-bearing premise.
  • domain assumption The valid-mask inequality in Eq. (12) rejects only impossible track-detection pairs.
    A single distance threshold based on lambda_D determines which pairs are considered; no missed-match analysis is given.
  • domain assumption Two selected sequences, MOT17-04 and MOT20-05, are representative of the MOT17 and MOT20 benchmarks.
    Generalization from one sparse and one crowded sequence is asserted without full-benchmark validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GenTrack3: Hybrid Stochastic-Deterministic Online Multi-Object Tracking with Cluster-Aware Association." pith.science (2026). https://pith.science/paper/P5Y7EC3T

@misc{pith2026260809581,
  author       = {Pith},
  title        = {Pith review of: GenTrack3: Hybrid Stochastic-Deterministic Online Multi-Object Tracking with Cluster-Aware Association},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P5Y7EC3T}},
  note         = {Machine review of arXiv:2608.09581}
}
read the original abstract

Multi-object tracking (MOT) involves maintaining consistent target identities as objects dynamically enter and leave a scene. Deterministic approaches, such as tracking-by-detection with data association, produce reproducible results and are computationally efficient, but they rely heavily on motion models and are sensitive to noisy detections that can lead to association errors. In contrast, stochastic methods explicitly model uncertainty and can better handle complex non-linear dynamics, albeit at the cost of increased computational complexity and variability arising from random sampling. This paper presents an online MOT framework that integrates deterministic and stochastic principles to achieve robust tracking under uncertainty. Furthermore, a novel track-to-detection matching approach is introduced to enhance scalability with increasing target numbers while supporting group tracking. The tracking inference mechanism employs a tracklet that includes identifiers, states, velocities, track penalties and track ages of targets, supporting a systematic tracking pipeline. Each target is associated with a stochastic particle set to compute the matching cost to detections. Reference implementations of the proposed approach and baseline trackers can be found on GitHub: https://github.com/SDU-VelKoTek/GenTrack3.

Figures

Figures reproduced from arXiv: 2608.09581 by the authors.

Figure 1
Figure 1. Overview of the tracking pipeline for temporal and scalable numbers of targets. Section 3 evalu￾ates the proposed tracker on MOT17 and MOT20. Finally, the conclusions are presented in Section 4. 2 Methodology To effectively address non-linear dynamics and non-Gaussian noise, multi-object tracking can be formulated as a particle filtering problem. Here, the system state is represented as {Kt , Xt} = {Ki,t , Xi,t} k i… view at source ↗
Figure 2
Figure 2. Pipeline for track-detection matching The first stage clusters tracks and detections to partition the global cost matrix into smaller, independent submatrices for efficient matching, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Track and detection grouping sampled from the MOT17-04 sequence [ [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Example of a valid-pair mask within an individual [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Example scenarios where a detection is assigned [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: A shared detection - shown as a blue bounding box [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Crowded scene from MOT20-05 sequence [24] show￾ing 188 targets in the current frame. Three clusters are identified: clusters 1 and 2 each contain a single target, while the last cluster contains the remaining 186 targets [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Image-region clustering: the image is partitioned into [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 24 canonical work pages

  1. [1]

    Multiple object tracking by hierarchical association of spatio-temporal data,

    Csaba Beleznai and David Schreiber, "Multiple object tracking by hierarchical association of spatio-temporal data," in 2010 IEEE International Conference on Image Processing, 2010, pp. 41-44

    work page 2010

  2. [2]

    Improving Multi- target Tracking via Social Grouping,

    Zhen Qin and Christian R. Shelton, "Improving Multi- target Tracking via Social Grouping," in 2012 IEEE Con- ference on Computer Vision and Pattern Recognition, 2012, pp. 1972-1978

    work page 2012

  3. [3]

    Simple online and realtime track- ing,

    Bewley Alex and Ge Zongyuan and Ott Lionel and Ramos Fabio and Upcroft Ben, "Simple online and realtime track- ing," in 2016 IEEE International Conference on Image Processing (ICIP), 2016, pp. 3464-3468

    work page 2016

  4. [4]

    Sim- ple online and realtime tracking with a deep association metric,

    Nicolai Wojke and Alex Bewley and Dietrich Paulus, "Sim- ple online and realtime tracking with a deep association metric," in 2017 IEEE International Conference on Image Processing (ICIP), 2017, pp. 3645-3649

    work page 2017

  5. [5]

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

    Yifu Zhang and Peize Sun and Yi Jiang and Dongdong Yu and Fucheng Weng and Zehuan Yuan and Ping Luo and Wenyu Liu and Xinggang Wang, "ByteTrack: Multi-object tracking by associating every detection box," in Computer Vision – ECCV 2022. Lecture Notes in Computer Science, vol. 13682, 2022, Springer Cham

    work page 2022

  6. [6]

    BoT-SORT: Robust associations multi-pedestrian track- ing,

    Nir Aharon and Roy Orfaig and Ben-Zion Bobrovsky, "BoT-SORT: Robust associations multi-pedestrian track- ing," arXiv, 2022. Available at: https://doi.org/10. 48550/arXiv.2206.14651

    arXiv doi:10.48550/arxiv.2206.14651 2022

  7. [7]

    Observation-Centric SORT: Rethinking SORT for robust multi-object tracking,

    Jinkun Cao and Jiangmiao Pang and Xinshuo Weng and Rawal Khirodkar and Kris Kitani, "Observation-Centric SORT: Rethinking SORT for robust multi-object tracking," in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 9686-9696

    work page 2023

  8. [8]

    SMILEtrack: Similarity learning for occlusion-aware mul- tiple object tracking,

    Yu-Hsiang Wang and Jun-Wei Hsieh and Ping-Yang Chen and Ming-Ching Chang and Hung-Hin So and Xin Li, "SMILEtrack: Similarity learning for occlusion-aware mul- tiple object tracking," in Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence, vol. 38, no. 6, pp. 5740-5748, 2024

    work page 2024

Show all 26 references
  1. [9]

    ConfTrack: Kalman filter-based multi- person tracking by utilizing confidence score of detection box,

    Hyeonchul Jang and Seokjun Kang and Takgen Kim and HyeongKi Kim, "ConfTrack: Kalman filter-based multi- person tracking by utilizing confidence score of detection box," in 2024 IEEE/CVF Winter Conference on Applica- tions of Computer Vision (W ACV), 2024, pp. 6583-6592

    2024

  2. [10]

    Boost- Track: Boosting the similarity measure and detection con- fidence for improved multiple object tracking,

    Vukasin D. Stanojevic and Branimir T. Todorovic, "Boost- Track: Boosting the similarity measure and detection con- fidence for improved multiple object tracking," Machine Vision and Applications, vol. 35, no. 53, 2024

    2024

  3. [11]

    BraMBLe: A Bayesian multiple-blob tracker,

    M. Isard and J. MacCormick, "BraMBLe: A Bayesian multiple-blob tracker," in Proceedings Eighth IEEE Inter- national Conference on Computer Vision (ICCV 2001), 2001, pp. 34-41

    2001

  4. [12]

    MCMC-based particle filter for tracking a variable number of interacting targets,

    Zia Khan and T. Balch and F. Dellaert, "MCMC-based particle filter for tracking a variable number of interacting targets," IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, vol. 27, no .11, pp. 1805-1819, 2005

    2005

  5. [13]

    Robust tracking-by-detection using a detector confidence particle filter,

    Michael D. Breitenstein and Fabian Reichlin and Bastian Leibe and Esther Koller-Meier and Luc Van Gool, "Robust tracking-by-detection using a detector confidence particle filter," in 2009 IEEE 12th International Conference on Computer Vision (ICCV), 2009, pp. 1515-1522

    2009

  6. [14]

    Detection driven adaptive multi-cue integration for mul- tiple human tracking,

    Ming Yang and Gengjun Lv and Wei Xu and Yihong Gong, "Detection driven adaptive multi-cue integration for mul- tiple human tracking," in 2009 IEEE 12th International Conference on Computer Vision (ICCV), 2009, pp. 1554- 1561

    2009

  7. [15]

    A boosted particle filter: multitarget detection and tracking,

    Kenji Okuma and Ali Taleghani and Nando de Freitas and James J. Little and David G. Lowe, "A boosted particle filter: multitarget detection and tracking," in Proceedings of European Conference on Computer Vision (ECCV), 2024, pp.28-39

    2024

  8. [16]

    GenTrack: A New Genera- tion of Multi-Object Tracking,

    Toan Van Nguyen and Rasmus G. K. Christiansen and Dirk Kraft and Leon Bodenhagen, "GenTrack: A New Genera- tion of Multi-Object Tracking," arXiv, 2025. Available at: https://doi.org/10.48550/arXiv.2510.24399

    2025 doi arXiv

  9. [17]

    Toan Van Nguyen and Rasmus G. K. Christiansen and Dirk Kraft and Leon Bodenhagen, "GenTrack2: An Improved Hybrid Approach for Multi-Object Track- ing, arXiv, 2025. Available at: https://doi.org/10. 48550/arXiv.2510.24410

    2025 doi arXiv

  10. [18]

    Particle swarm optimiza- tion,

    J. Kennedy and R. Eberhart, "Particle swarm optimiza- tion," in Proceedings of IEEE International Conference on Neural Networks, 1995, pp. 1942-1948

    1995

  11. [19]

    The particle swarm - explosion, stability, and convergence in a multidimensional complex space,

    M. Clerc and J. Kennedy, "The particle swarm - explosion, stability, and convergence in a multidimensional complex space," IEEE Transactions on Evolutionary Computation, vol. 6, no. 1, pp. 58-73, 2002

    2002

  12. [20]

    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, 1995

    1995

  13. [21]

    Improved Hungarian algorithm for assignment problems of serial- parallel systems,

    Tingpeng Li and Yue Li and Yanling Qian, "Improved Hungarian algorithm for assignment problems of serial- parallel systems," Journal of Systems Engineering and Electronics, vol. 27, no. 4, pp. 858-870, 2016. Preprint– GenTrack3: HybridStochastic–DeterministicOnlineMulti-Object...

    2016

  14. [22]

    The Dynamic Hungarian Algorithm for the Assignment Problem with Changing Costs,

    Ayorkor Mills-Tettey and Anthony Stent and M. Bernar- dine Dias, "The Dynamic Hungarian Algorithm for the Assignment Problem with Changing Costs," Robotics In- stitute, Carnegie Mellon University, 2007

    2007

  15. [23]

    MOTChallenge: A benchmark for single-camera multiple target tracking,

    Patrick Dendorfer and Aljosa Osep and Anton Milan and Konrad Schindler and Daniel Cremers and Ian Reid and Stefan Roth and Laura Leal-Taixe, "MOTChallenge: A benchmark for single-camera multiple target tracking," International Journal of Computer Vision, vol. 129, pp. 845-881, 2021

    2021

  16. [24]

    MOT20: A benchmark for multi object tracking in crowded scenes,

    Patrick Dendorfer and Hamid Rezatofighi and Anton Mi- lan and Javen Shi and Daniel Cremers and Ian Reid and Stefan Roth and Konrad Schindler and Laura Leal- Taixe, "MOT20: A benchmark for multi object tracking in crowded scenes," arXiv, 2020. Available at: https: //doi.org/10....

    2020 doi arXiv

  17. [25]

    Performance evaluation of object detec- tion and tracking in video,

    Vasant Manohar and Padmanabhan Soundararajan and Har- ish Raju and Dmitry Goldgof and Rangachar Kasturi and John Garofolo, "Performance evaluation of object detec- tion and tracking in video, "Springer, Berlin, Heidelberg, 2006, in Narayanan, P.J., Nayar, S.K., Shum, HY . (eds...

    2006

  18. [26]

    HOTA: A higher order metric for evaluating multi-object tracking,

    Jonathon Luiten and Aljosa Osep and Patrick Dendorfer and Philip Torr and Andreas Geiger and Laura Leal-Taixe and Bastian Leibe, "HOTA: A higher order metric for evaluating multi-object tracking," International Journal of Computer Vision, vol. 129, pp. 548-578, 2021

    2021

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.