Pith. sign in

REVIEW 4 major objections 3 minor 44 references

A learned link-prediction tracker surpasses handcrafted association rules and end-to-end baselines on multiple multi-object tracking benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 13:52 UTC pith:37N2353S

load-bearing objection Genuinely useful formulation and strong results on non-linear motion benchmarks, but the abstract's 'consistently surpasses SOTA' is contradicted by the paper's own MOT17 numbers. the 4 major comments →

arxiv 2512.22105 v2 pith:37N2353S submitted 2025-12-26 cs.CV

Learning Association via Track-Detection Matching for Multi-Object Tracking

Momir Ad\v{z}emovi\'c This is my paper
classification cs.CV
keywords multi-object trackingdata associationlink predictiontracking-by-detectionbipartite matchingtransformer encodergeometric featuresmetric learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes TDLP, a tracking-by-detection method that treats per-frame data association as link prediction: for each active track and each new detection, a neural network outputs the probability that this detection is the track's continuation, and a Hungarian assignment then decides the final matching. The central claim is that learning this pairwise score function directly from data is more effective than both handcrafted association heuristics and metric-learning (contrastive) objectives, particularly when the available features are low-dimensional geometric cues like bounding boxes. On DanceTrack, SportsMOT, BEE24, and SoccerNet, TDLP reports state-of-the-art HOTA, and the bounding-box-only variant outperforms all heuristic trackers that use appearance cues. The paper also includes a controlled analysis showing link prediction maintains sharp rejection of false detections in threshold tests, while contrastive learning fails, and acknowledges that TDLP underperforms on MOT17 due to limited training data.

Core claim

The discovery is that data association in multi-object tracking can be formulated as bipartite link prediction between tracks and detections, and that this formulation learns a more robust association function than embedding-based metric learning. The model encodes track histories and current detections with transformer encoders, fuses optional cues (pose, appearance), and for each pair predicts a link probability from the concatenation of the two embeddings and their absolute difference, trained with weighted binary cross-entropy. The paper argues that link prediction learns a local similarity function that can selectively weight informative feature differences, whereas contrastive learning

What carries the argument

The load-bearing component is the track–detection link-prediction head: for each pair (i, j), it constructs the vector [z_trk_i, z_det_j, |z_trk_i − z_det_j|] and passes it through a shallow MLP to produce a probability in (0,1). This head is trained with a weighted binary cross-entropy loss over positive and negative track–detection pairs, and at inference its outputs become the cost matrix for linear assignment. Surrounding this head are a temporal transformer encoder (using reversed positional encoding to tolerate truncated trajectories), a geometric feature transform (frame-wise min–max normalization and first-order differences) that converts raw bounding boxes into standardized inputs,

Load-bearing premise

The claim that link prediction is fundamentally more robust than contrastive learning rests on a controlled experiment where negative detections were generated with hand-picked offsets and a limited set of motion patterns; if real benchmarks produce different negative distributions or failure modes, the mechanism may not transfer to the reported benchmark gains.

What would settle it

Run TDLP and a contrastive baseline on a benchmark with frequent detector misses and dense overlapping tracks (e.g., a crowded pedestrian scene with heavy occlusion), and measure the ID-switch rate when the true detection is absent. If the contrastive baseline yields fewer ID switches, or if TDLP fails a threshold test with negatives drawn from real benchmark failure modes (e.g., offsets of a few pixels in multiple dimensions), the central robustness claim collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Tracking-by-detection pipelines can retain the efficiency and modularity of SORT-style systems while matching or exceeding end-to-end trackers on benchmarks with non-linear motion.
  • The bounding-box-only variant outperforming appearance-based heuristic trackers suggests that appearance cues may be unnecessary on such benchmarks, enabling cheaper tracking pipelines.
  • Link prediction provides a natural way to fuse heterogeneous cues (bounding boxes, pose, appearance) by letting the model learn how to weight each cue per pair rather than relying on hand-tuned distances.
  • The demonstrated robustness to false negatives implies TDLP may preserve identities better when detectors miss objects, a common failure in occlusion-heavy scenes.
  • The paper acknowledges that TDLP underperforms on MOT17 and that its computational cost scales quadratically with the number of tracks and detections, so crowded scenes remain a practical limitation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to apply the same link-prediction head to other matching problems with heterogeneous, low-dimensional cues, such as sensor-fusion association or object re-identification across camera views, where contrastive embeddings may be equally fragile.
  • The threshold-test analysis suggests a practical probe for any learned tracker: simulate the absence of the true detection and check whether negatives are reliably rejected; this could be part of a benchmarking protocol for association methods.
  • The choice of negative offsets in the controlled experiment (s=0.1) may underrepresent real-world failure modes like dense, overlapping boxes; a wider sweep of negative distributions could sharpen or qualify the claimed advantage over contrastive learning.
  • Since TDLP only demonstrates the link-prediction advantage with low-dimensional geometric features, a natural next experiment is to examine whether the advantage narrows when high-dimensional appearance embeddings dominate, which would clarify where each objective is preferable.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. The paper proposes Track-Detection Link Prediction (TDLP), a tracking-by-detection MOT method that replaces handcrafted association heuristics with a learned per-frame link predictor between track and detection embeddings. The pipeline encodes track histories and current detections with transformer encoders, fuses optional pose/appearance cues, and uses a shallow MLP to output link probabilities that are fed to Hungarian matching. Training uses weighted BCE on ground-truth track–detection pairs. The method is evaluated on DanceTrack, SportsMOT, BEE24, SoccerNet, and MOT17. TDLP reports state-of-the-art HOTA on the first four benchmarks, while the full model underperforms on MOT17; a bbox-only variant (TDLP-bbox) is also analyzed. A component ablation against a contrastive baseline (CTDP) and a synthetic experiment are used to argue that link prediction is more effective than contrastive learning for geometric features.

Significance. If the reported results hold, TDLP is a useful contribution: it shows that learned link prediction can replace handcrafted association rules in a tracking-by-detection pipeline and achieve strong results on motion-heavy benchmarks (DanceTrack, SportsMOT, BEE24, SoccerNet) while remaining modular and more efficient than end-to-end trackers. The paper also provides an instructive ablation against a contrastive objective under a fixed architecture, and the code is released. However, the headline claim of 'consistently surpassing' state-of-the-art is contradicted by the paper's own MOT17 numbers, and the lack of error bars makes it difficult to assess whether the reported gains, some as small as +0.8 HOTA, are significant. The central methodological idea is sound, but the claims need to be qualified and the evidence needs to be reported more rigorously.

major comments (4)
  1. [Abstract; Section 1; Tables 2 and 6] The abstract states that TDLP 'consistently surpasses state-of-the-art performance across both tracking-by-detection and end-to-end methods,' and Section 1 claims the bbox-only variant 'outperforms all heuristic-based trackers, including those that use object appearance cues.' Both claims are contradicted by the paper's own tables. On MOT17 (Table 6), TDLP achieves HOTA 60.6, below ByteTrack (62.8) and CAMELTrack (62.4), and the text openly concedes underperformance. On SportsMOT (Table 2), TDLP-bbox reaches HOTA 74.8, but DeepMoveSORT (Train+Val) reaches 78.7 and Deep-EIoU (Train+Val) 77.2, both heuristic-based trackers using appearance cues. The claims must be restricted to the benchmarks and settings where they actually hold, e.g., 'surpasses SOTA on non-linear-motion benchmarks' and 'outperforms prior motion-only bbox trackers.'
  2. [Section 4, Tables 1–3, 6, 7] No error bars, confidence intervals, or multiple-seed results are reported for any benchmark. Gains over the previous SOTA are modest in several cases (e.g., +0.8 HOTA on DanceTrack, +1.5 on SportsMOT, +1.6 on BEE24), and without run-to-run variance it is impossible to know whether these differences are significant or whether the 'consistently surpasses' wording is justified. I request mean±std over at least three seeds, or a clear statement if the benchmark protocol permits only a single submission and why the reported differences should be considered reliable.
  3. [Section 3.2, Eq. (4)] Frame-wise min–max normalization is not well-defined when x_max = x_min, which occurs whenever a feature dimension is constant across all detections in a frame (e.g., a frame with a single detection, or identical widths/heights). The paper neither specifies an epsilon nor a fallback for this case. Since Eq. (4) is applied to all geometric features before training and inference, the implementation must handle this degenerate condition; please state the exact procedure (e.g., if denominator is zero, keep the value unchanged or set to 0).
  4. [Section 4.2 / Appendix B] The synthetic experiment that supports the mechanistic claim 'link prediction is more robust in threshold tests' uses hand-chosen negative detection offsets (s=0.1) and a small set of motion patterns. No sensitivity analysis is provided, so the conclusion that TDLP 'consistently suppresses negatives' may be an artifact of this specific negative distribution. Since this analysis is used to explain the benchmark gains, please add a sensitivity sweep over s and negative-generation strategy, or explicitly soften the claim to 'under the tested synthetic configurations.'
minor comments (3)
  1. [Table 1] Typo: 'MeMORT' should be 'MeMOTR' to match the reference and the rest of the text.
  2. [Section 4.2, Figure 4] The caption refers to 'n0–n3' but does not define these shorthand labels; please define them either in the caption or by referencing Appendix B more explicitly.
  3. [Appendix A] The implementation details do not mention the random seed, the number of training runs, or the hardware used for the main experiments. Please add this information for reproducibility.

Circularity Check

0 steps flagged

No significant circularity: TDLP's benchmark results are held-out evaluations of a learned association objective, and self-citations are controlled by including the cited method as a baseline.

full rationale

The paper's central derivation is self-contained: TDLP trains link probabilities S_ij with a BCE loss (Eq. 3) on ground-truth identities from training clips, and the reported HOTA/IDF1 values come from held-out test sets (Tables 1–3, 7), so the predictions are not the fitted inputs by construction. The CTDP baseline is an internal ablation sharing the TDLP architecture; it is used to isolate the effect of the link-prediction objective, not to generate the SOTA claim. The main reuse from the author's prior work is the temporal/motion encoder: 'we adopt the motion encoder from TransFilter [2]' (Section 3.2). This is a self-citation, but it is not circular: [2] is a published, externally evaluated tracker, and DeepMoveSORT [2] appears as a baseline in the same tables, so the comparison controls for the shared component. The synthetic rank/threshold experiment (Section 4.2/Appendix B) uses hand-chosen negative offsets (s=0.1), but it serves as an illustrative mechanistic analysis and is not the basis for the benchmark claims. The paper's own admission that 'our model underperforms compared to heuristic-based methods' on MOT17 (Table 6) is a correctness/scoping problem with the abstract's 'consistently surpasses' wording, not a circularity problem. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported, and no load-bearing step reduces to the paper's own inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 3 axioms · 0 invented entities

The method rests on standard deep learning components (transformers, MLPs, BCE loss) and a standard matching algorithm; the main free parameters are per-dataset thresholds and the clip length. The controlled CTDP-vs-TDLP experiment introduces hand-chosen synthetic negatives. No new physical or conceptual entities are postulated.

free parameters (6)
  • BCE positive weight w+ = 10
    Controls class imbalance in the track–detection link prediction loss (Eq. 3); chosen by hand.
  • link gate threshold theta_link = 0.015, 0.01, 0.65, 0.05 per dataset
    Used at inference to prune candidate links before linear assignment; tuned per benchmark (Appendix A) and central to the threshold-test analysis.
  • detection threshold theta_det = 0.4, 0.1, 0.6, 0.5 per dataset
    Filters detections before scoring; standard SORT hyperparameter tuned per dataset.
  • clip length L = 50 or 150 frames per dataset
    Length of track history window; chosen to support long occlusions and affects training.
  • tracker lifecycle thresholds (Tinit, theta_new, Tlost) = e.g., Tinit=3,1,0,1; theta_new=0.9,0.4,0.6,0.55; Tlost=50,150,50,50
    SORT-like lifecycle hyperparameters for track initialization and termination; tuned per dataset.
  • negative-generation offsets in CTDP-vs-TDLP experiment = s=0.1 and four offset vectors (Appendix B)
    Hand-chosen synthetic negatives for the controlled rank/threshold tests; the threshold-test conclusions depend on these choices.
axioms (3)
  • standard math Hungarian algorithm solves the bipartite assignment optimally
    Used at inference (Sec. 3.1) to assign tracks to detections.
  • domain assumption Detections from an external detector are fixed and not learned
    The method operates in tracking-by-detection; all results use public YOLOX or provided detectors.
  • ad hoc to paper Frame-wise min-max normalization is well-defined and invertible
    Eq. (4) divides by xmax − xmin; with a single detection the denominator is zero. No epsilon or fallback is stated.

pith-pipeline@v1.3.0-alltime-deepseek · 13697 in / 10588 out tokens · 107808 ms · 2026-08-03T13:52:00.524940+00:00 · methodology

0 comments
read the original abstract

Multi-object tracking aims to maintain object identities over time by associating detections across video frames. Two dominant paradigms exist in literature: tracking-by-detection methods, which are computationally efficient but rely on handcrafted association heuristics, and end-to-end approaches, which learn association from data at the cost of higher computational complexity. We propose Track-Detection Link Prediction (TDLP), a tracking-by-detection method that performs per-frame association via link prediction between tracks and detections, i.e., by predicting the correct continuation of each track at every frame. TDLP is architecturally designed primarily for geometric features such as bounding boxes, while optionally incorporating additional cues, including pose and appearance. Unlike heuristic-based methods, TDLP learns association directly from data without handcrafted rules, while remaining modular and computationally efficient compared to end-to-end trackers. Extensive experiments on multiple benchmarks demonstrate that TDLP consistently surpasses state-of-the-art performance across both tracking-by-detection and end-to-end methods. Finally, we provide a detailed analysis comparing link prediction with metric learning-based association and show that link prediction is more effective, particularly when handling heterogeneous features such as detection bounding boxes. Our code is available at \href{https://github.com/Robotmurlock/TDLP}{https://github.com/Robotmurlock/TDLP}.

Figures

Figures reproduced from arXiv: 2512.22105 by Momir Ad\v{z}emovi\'c.

Figure 1
Figure 1. Figure 1: High-level overview of the proposed association pipeline. Encoded track histories and current detections are linked via link prediction and associated using bipartite matching. In this work, we propose Track–Detection Link Prediction (TDLP), a tracking￾by-detection method that lies between heuristic-based association and fully end-to-end learning. TDLP formulates data association as a per-frame link predic… view at source ↗
Figure 2
Figure 2. Figure 2: Architecture overview of the proposed TDLP method. Tracks and detec￾tions first undergo a geometric feature transformation, while optional additional cues (e.g., pose keypoints and appearance) are extracted using dedicated models. The static encoder processes per-frame detections from both current detections and track histories, while a motion encoder extracts track-specific temporal features (e.g., veloci… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of CTDP and TDLP across different bounding-box features (X, Y, W, H, C) measured by HOTA on SportsMOT validation set. TDLP consistently outperforms CTDP across all feature subsets, with the gap widening as more features are introduced. This highlights a fundamental difference between the objectives: contrastive learning enforces a global embedding structure in which all features must align withi… view at source ↗
Figure 4
Figure 4. Figure 4: Pass/fail matrices for CTDP and TDLP. Dots indicate correct outcomes; shaded cells (X) mark failures leading to ID switches. See Appendix B. demonstrating robustness in scenarios where linear assignment most often fails. This difference arises from the contrastive objective of CTDP, which maps spatially nearby boxes to similar embeddings, whereas TDLP maintains sharper discriminative boundaries. Full exper… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 5 canonical work pages

  1. [1]

    Adˇ zemovi´ c, M.: Deep learning-based multi-object tracking: A comprehensive survey from foundations to state-of-the-art (2025),https://arxiv.org/abs/2506.13457

    Pith/arXiv arXiv 2025

  2. [2]

    arXiv preprint arXiv:2407.00738 (2024), https: //arxiv.org/abs/2407.00738

    Adˇ zemovi´ c, M., Tadi´ c, P., Petrovi´ c, A., Nikoli´ c, M.: Engineering an efficient object tracker for non-linear motion. arXiv preprint arXiv:2407.00738 (2024), https: //arxiv.org/abs/2407.00738

    arXiv 2024

  3. [3]

    Machine Vision and Applications 36(1), 20 (2024)

    Adˇ zemovi´ c, M., Tadi´ c, P., Petrovi´ c, A., Nikoli´ c, M.: Beyond kalman filters: deep learning-based filters for improved object tracking. Machine Vision and Applications 36(1), 20 (2024). https://doi.org/10.1007/s00138-024-01644-x , https://doi. org/10.1007/s00138-024-01644-x

    work page doi:10.1007/s00138-024-01644-x 2024

  4. [4]

    In: arXiv preprint (2022), arxiv: 2206.14651

    Aharon, N., Orfaig, R., Bobrovsky, B.Z.: Bot-sort: Robust associations multi- pedestrian tracking. In: arXiv preprint (2022), arxiv: 2206.14651

    Pith/arXiv arXiv 2022

  5. [5]

    In: arXiv preprint (2016), arxiv: 1607.06450

    Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. In: arXiv preprint (2016), arxiv: 1607.06450

    Pith/arXiv arXiv 2016

  6. [6]

    In: 2016 IEEE International Conference on Image Processing (ICIP)

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

    arXiv 2016

  7. [7]

    Bozek, K., Hebert, L., Mikheyev, A., Stephens, G.: Towards dense object tracking in a 2d honeybee hive. pp. 4185–4193 (06 2018). https://doi.org/10.1109/CVPR. 2018.00440

    arXiv 2018

  8. [8]

    In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)

    Bras´ o, G., Leal-Taix´ e, L.: Learning a neural solver for multiple object tracking. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)

    2020

  9. [9]

    arXiv preprint arXiv:1903.11027 (2019)

    Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. arXiv preprint arXiv:1903.11027 (2019)

    Pith/arXiv arXiv 1903

  10. [12]

    IEEE Transactions on Image Processing34, 743–758 (2025)

    Cao, X., Zheng, Y., Yao, Y., Qin, H., Cao, X., Guo, S.: Topic: A parallel association paradigm for multi-object tracking under complex motions and diverse scenes. IEEE Transactions on Image Processing34, 743–758 (2025). https://doi.org/10.1109/ TIP.2025.3526066

    arXiv 2025

  11. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Cetintas, O., Bras´ o, G., Leal-Taix´ e, L.: Unifying short and long-term tracking with graph hierarchies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22877–22887 (June 2023)

    2023

  12. [14]

    arXiv preprint arXiv:2203.05328 (2022)

    Chen, B., Li, P., Bai, L., Qiao, L., Shen, Q., Li, B., Gan, W., Wu, W., Ouyang, W.: Backbone is all your need: A simplified architecture for visual object tracking. arXiv preprint arXiv:2203.05328 (2022)

    Pith/arXiv arXiv 2022

  13. [15]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Cioppa, A., Giancola, S., Deliege, A., Kang, L., Zhou, X., Cheng, Z., Ghanem, B., Van Droogenbroeck, M.: Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3491–3502 (2022) 16 M. Adˇ zemovi´ c

    2022

  14. [17]

    International Journal of Computer Vision129(4), 845– 881 (2021)

    Dendorfer, P., Oˇ sep, A., Milan, A., Schindler, K., Cremers, D., Reid, I., Roth, S., Leal-Taix´ e, L.: Motchallenge: A benchmark for single-camera mul- tiple target tracking. International Journal of Computer Vision129(4), 845– 881 (2021). https://doi.org/10.1007/s11263-020-01393-0 , https://doi.org/ 10.1007/s11263-020-01393-0

    doi:10.1007/s11263-020-01393-0 2021

  15. [18]

    Neural Networks107, 3– 11 (2018)

    Elfwing, S., Uchibe, E., Doya, K.: Sigmoid-weighted linear units for neural net- work function approximation in reinforcement learning. Neural Networks107, 3– 11 (2018). https://doi.org/https://doi.org/10.1016/j.neunet.2017.12.012, special issue on deep reinforcement learning

    doi:10.1016/j.neunet.2017.12.012 2018

  16. [19]

    In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV)

    Gao, R., Wang, L.: Memotr: Long-term memory-augmented transformer for multi- object tracking. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9867–9876 (2023). https://doi.org/10.1109/ICCV51070.2023.00908

    arXiv 2023

  17. [20]

    Gao, R., Zhang, Y., Wang, L.: Multiple object tracking as id prediction (2024), https://arxiv.org/abs/2403.16848

    Pith/arXiv arXiv 2024

  18. [21]

    Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021 (2021), https://arxiv.org/abs/2107.08430

    Pith/arXiv arXiv 2021

  19. [22]

    Applied Intelligence55(1), 33 (2024)

    Han, X., Oishi, N., Tian, Y., Ucurum, E., Young, R., Chatwin, C., Birch, P.: Ettrack: enhanced temporal motion predictor for multi-object tracking. Applied Intelligence55(1), 33 (2024). https://doi.org/10.1007/s10489-024-05866-4 , https://doi.org/10.1007/s10489-024-05866-4

    work page doi:10.1007/s10489-024-05866-4 2024

  20. [23]

    In: 2024 International Conference on Machine Intelligence and Smart Innovation (ICMISI)

    Hossam, A., Ramadan, A., Magdy, M., Abdelwahab, R., Ashraf, S., Mohamed, Z.: Revolutionizing retail analytics: Advancing inventory and customer insight with ai. In: 2024 International Conference on Machine Intelligence and Smart Innovation (ICMISI). pp. 64–69 (2024). https://doi.org/10.1109/ICMISI61517. 2024.10580424

    arXiv 2024

  21. [24]

    In: 2024 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (W ACVW)

    Huang, H.W., Yang, C.Y., Sun, J., Kim, P.K., Kim, K.J., Lee, K., Huang, C.I., Hwang, J.N.: Iterative Scale-Up ExpansionIoU and Deep Features Association for Multi-Object Tracking in Sports . In: 2024 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (W ACVW). pp. 163–172 (2024). https://doi.org/10.1109/WACVW60836.2024.00024

    arXiv 2024

  22. [25]

    https://doi

    Liu, Z., Wang, X., Wang, C., Liu, W., Bai, X.: Sparsetrack: Multi-object tracking by performing scene decomposition based on pseudo-depth (2025). https://doi. org/10.1109/TCSVT.2024.3524670

    arXiv 2025

  23. [26]

    arXiv preprint arXiv:1608.03983 (2017)

    Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2017)

    Pith/arXiv arXiv 2017

  24. [27]

    In: 7th Interna- tional Conference on Learning Representations (ICLR) (2019)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: 7th Interna- tional Conference on Learning Representations (ICLR) (2019)

    2019

  25. [28]

    International Journal of Computer Vision129(2), 548–578 (2021)

    Luiten, J., Oˇ sep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taix´ e, L., Leibe, B.: Hota: A higher order metric for evaluating multi-object tracking. International Journal of Computer Vision129(2), 548–578 (2021). https://doi.org/10.1007/ s11263-020-01375-2

    2021

  26. [29]

    In: 2023 IEEE International Conference on Image Processing (ICIP)

    Maggiolino, G., Ahmad, A., Cao, J., Kitani, K.: Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 3025–3029 (2023). https://doi.org/10.1109/ ICIP49359.2023.10222576 Title Suppressed Due to Excessive Length 17

    arXiv 2023

  27. [31]

    Pattern Recognition160, 111169 (2025)

    Miah, M., Bilodeau, G.A., Saunier, N.: Learning data association for multi- object tracking using only coordinates. Pattern Recognition160, 111169 (2025). https://doi.org/https://doi.org/10.1016/j.patcog.2024.111169, https:// www.sciencedirect.com/science/article/pii/S0031320324009208

    arXiv 2025

  28. [32]

    arXiv preprint arXiv:1807.03748 (2018)

    Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    Pith/arXiv arXiv 2018

  29. [33]

    In: 3rd International Conference on Learning Representations (ICLR) (2015)

    P, D., Kingma, Ba, J.: Adam: A method for stochastic optimization. In: 3rd International Conference on Learning Representations (ICLR) (2015)

    2015

  30. [34]

    In: HP Laboratories (2012)

    Ramshaw, L., E, R., Tarjan: On minimum-cost assignments in unbalanced bipartite graphs. In: HP Laboratories (2012)

    2012

  31. [35]

    In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Seidenschwarz, J., Bras´ o, G., Serrano, V.C., Elezi, I., Leal-Taix´ e, L.: Simple cues lead to a strong multi-object tracker. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13813–13823 (2023). https://doi. org/10.1109/CVPR52729.2023.01327

    arXiv 2023

  32. [36]

    Somers, V., Standaert, B., Joos, V., Alahi, A., Vleeschouwer, C.D.: Cameltrack: Context-aware multi-cue exploitation for online multi-object tracking (2025), https: //arxiv.org/abs/2505.01257

    Pith/arXiv arXiv 2025

  33. [37]

    Journal of Machine Learning Research15(56), 1929–1958 (2014)

    Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research15(56), 1929–1958 (2014)

    1929

  34. [38]

    Machine Vision and Applications35(3), 53 (April 2024)

    Stanojevic, V.D., Todorovic, B.T.: Boosttrack: boosting the similarity mea- sure and detection confidence for improved multiple object tracking. Machine Vision and Applications35(3), 53 (April 2024). https://doi.org/10.1007/ s00138-024-01531-5,https://doi.org/10.1007/s00138-024-01531-5

    work page doi:10.1007/s00138-024-01531-5 2024

  35. [39]

    In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

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

    arXiv 2022

  36. [40]

    In: 5th Workshop on Long-term Human Motion Prediction (2021), arxiv: 2106.01153

    Urbann, O., Bredtmann, O., Otten, M., Richter, J.P., Bauer, T., Zibriczky, D.: Online and real-time tracking in a surveillance scenario. In: 5th Workshop on Long-term Human Motion Prediction (2021), arxiv: 2106.01153

    Pith/arXiv arXiv 2021

  37. [41]

    In: 2017 IEEE International Conference on Image Processing (ICIP)

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

    arXiv 2017

  38. [42]

    Neural Networks179, 106539 (2024).https://doi.org/https://doi.org/10.1016/j.neunet.2024.106539

    Xiao, C., Cao, Q., Zhong, Y., Lan, L., Zhang, X., Luo, Z., Tao, D.: Motiontrack: Learning motion predictor for multiple object tracking. Neural Networks179, 106539 (2024).https://doi.org/https://doi.org/10.1016/j.neunet.2024.106539

    arXiv 2024

  39. [43]

    IEEE Robotics and Automation Letters9(1), 651–658 (2024)

    Xu, Z., Zhan, X., Xiu, Y., Suzuki, C., Shimada, K.: Onboard dynamic-object detection and tracking for autonomous robot navigation with rgb-d camera. IEEE Robotics and Automation Letters9(1), 651–658 (2024). https://doi.org/10. 1109/LRA.2023.3334683

    arXiv 2024

  40. [44]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Yang, M., Han, G., Yan, B., Zhang, W., Qi, J., Lu, H., Wang, D.: Hybrid-sort: Weak cues matter for online multi-object tracking. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6504–6512 (2024). https://doi. org/10.1609/aaai.v38i7.28471 18 M. Adˇ zemovi´ c

    work page doi:10.1609/aaai.v38i7.28471 2024

  41. [45]

    Proceedings of the AAAI Conference on Artificial Intelligence38(7), 6702–6710 (Mar 2024)

    Yi, K., Luo, K., Luo, X., Huang, J., Wu, H., Hu, R., Hao, W.: Ucmctrack: Multi- object tracking with uniform camera motion compensation. Proceedings of the AAAI Conference on Artificial Intelligence38(7), 6702–6710 (Mar 2024). https: //doi.org/10.1609/aaai.v38i7.28493

    work page doi:10.1609/aaai.v38i7.28493 2024

  42. [46]

    In: European Conference on Computer Vision (ECCV) (2022)

    Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. In: European Conference on Computer Vision (ECCV) (2022)

    2022

  43. [47]

    In: Avidan, S., Brostow, G., Ciss´ e, M., Farinella, G.M., Hassner, T

    Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: Bytetrack: Multi-object tracking by associating every detection box. In: Avidan, S., Brostow, G., Ciss´ e, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision – ECCV 2022. pp. 1–21. Springer Nature Switzerland, Cham (2022). https: //doi.org/10.1007/978-3-031-20047-2_1

    doi:10.1007/978-3-031-20047-2_1 2022

  44. [48]

    International Journal of Computer Vision129(11), 3069–3087 (Sep 2021)

    Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision129(11), 3069–3087 (Sep 2021). https://doi.org/10.1007/ s11263-021-01513-4

    2021