Pith. sign in

REVIEW 3 major objections 49 references

Continuous video labels and a three-module loop that links detection and association raise open-vocabulary tracking of novel objects to new highs.

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 · grok-4.5

2026-07-13 19:07 UTC pith:FGOI3DBV

load-bearing objection Real data contribution (C-TAO) plus clean incremental modules; SOTA numbers look solid on TAO/BDD, but attribution still hinges on unreleased continuous labels. the 3 major comments →

arxiv 2603.24016 v2 pith:FGOI3DBV submitted 2026-03-25 cs.CV cs.LG

COVTrack++: Learning Open-Vocabulary Multi-Object Tracking from Continuous Videos via a Synergistic Paradigm

Zekun Qian , Wei Feng , Ruize Han , Junhui Hou This is my paper
classification cs.CV cs.LG
keywords open-vocabulary multi-object trackingcontinuous video annotationC-TAOmulti-cue adaptive fusionhierarchical aggregationtemporal confidence propagationdetection-association synergyzero-shot tracking
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.

Open-vocabulary multi-object tracking asks a system to find, name, and keep identities for arbitrary objects, including categories never seen in training. Progress has been blocked by sparse video labels that skip most frames and by trackers that treat detection and association as separate steps. This paper supplies C-TAO, a continuously labeled version of the TAO training set that multiplies annotation density by roughly twenty-six and exposes smooth motion and intermediate states such as occlusion and blur. On that foundation it builds COVTrack++, which fuses appearance, motion, and semantic cues adaptively, strengthens parent objects with reliable child parts in dense boxes, and propagates high-confidence tracks backward and forward to revive flickering detections. The result is a bidirectional loop: better association recovers detections, and recovered detections improve later association. On TAO the method reaches 35.4 percent novel TETA on validation and 30.5 percent on test, with clear gains in association and localization, and it transfers zero-shot to driving video.

Core claim

The paper establishes that continuous frame-by-frame supervision plus a synergistic three-module framework (multi-cue adaptive fusion, hierarchical parent-child aggregation, and temporal confidence propagation) produces a bidirectional reciprocal mechanism between detection and association, yielding state-of-the-art open-vocabulary tracking of novel categories on TAO and strong zero-shot transfer to BDD100K.

What carries the argument

COVTrack++ bidirectional reciprocal mechanism: Multi-Cue Adaptive Fusion (MCF) balances appearance, motion, and semantic cues; Multi-Granularity Hierarchical Aggregation (MGA) lets visible child boxes reinforce occluded parents; Temporal Confidence Propagation (TCP) recovers low-confidence detections from high-confidence temporal neighbors, so association and detection continually improve each other.

Load-bearing premise

The continuous manual labels in C-TAO are accurate, complete, and free of systematic bias relative to the original sparse trajectories, so measured gains can be attributed to continuous supervision and the three modules rather than annotation artifacts.

What would settle it

Retrain the identical COVTrack++ pipeline on the original sparse TAO annotations (or on synthetic image pairs) and show that novel TETA and AssocA on the TAO validation and test sets no longer exceed prior methods by the reported margins; or independently re-annotate a held-out C-TAO subset and demonstrate that label disagreement erases the reported gains.

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

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

3 major / 0 minor

Summary. The paper addresses open-vocabulary multi-object tracking (OVMOT) by tackling a data bottleneck and a framework bottleneck. It introduces C-TAO, a continuously annotated version of the TAO training set (26× denser annotations) that restores frame-by-frame boxes, IDs, and categories so models can learn smooth motion and intermediate states. Building on the authors’ ICCV 2025 COVTrack, it proposes COVTrack++ with three modules that form a bidirectional detection–association loop: Multi-Cue Adaptive Fusion (MCF) balances appearance, motion, and semantic cues via dual-perspective confidence; Multi-Granularity Hierarchical Aggregation (MGA) uses parent–child spatial structure in dense OV detections so visible parts strengthen occluded wholes; and Temporal Confidence Propagation (TCP) recovers flickering low-confidence detections by propagating confidence along bipartite temporal matches. On TAO, the method reports novel TETA of 35.4% (val) and 30.5% (test), with gains of +4.8 AssocA and +5.8 LocA over prior work, plus zero-shot TETA 46.7% on BDD100K. Ablations and qualitative figures support the contribution of C-TAO and each module.

Significance. If the results hold under independent reimplementation and public release of C-TAO, the work is a solid incremental advance for OVMOT. Continuous video supervision for long-tail tracking is a genuine community need; showing that denser TAO labels help not only the authors’ pipeline but also prior methods would be a lasting contribution. The synergistic framing (association features improve detection recovery, which improves subsequent association) is a clear conceptual step beyond appearance-only or simple multi-cue sum fusion (e.g., SLAck). Strengths include standard TETA breakdowns on base/novel splits, module ablations, zero-shot BDD100K transfer, and explicit disclosure of the conference extension. The free parameters (thresholds, fusion weights, parent–child filters) are typical of tracking systems and do not by themselves invalidate the claim.

major comments (3)
  1. Section III (C-TAO construction) is load-bearing for the central attribution of gains to continuous supervision, yet provides no quantitative quality controls: no inter-annotator agreement, no error rates on intermediate frames, no comparison of continuous labels vs. original sparse TAO trajectories or vs. SLAck-style IoU pseudo-labels on the same videos. Without these, identity drift, category leakage into novel classes, or denser easy boxes could confound both the data-utility claim and the MCF/MGA/TCP ablations. The three-stage pipeline and multi-annotator cross-validation are described only at a high level; a short quality audit (or release of a labeled subset with agreement stats) is needed before the SOTA claim can be fully trusted.
  2. Section IV-B and the surrounding method body are incomplete in the submitted text: multi-cue feature extraction, dual-perspective confidence equations, fusion weights, MGA parent–child criteria (IoU/containment), and TCP bipartite matching / class-conditional propagation are only partially specified after the overview. Reproducibility of the reported +1.1 / +1.6 novel TETA over COVTrack and of the module ablations therefore depends on the promised code release. The authors should restore the full equations and hyperparameter tables (or point to a frozen public repo) so that the bidirectional synergy claim can be verified independently of the unreleased C-TAO.
  3. Attribution of novel AssocA/LocA gains (abstract; Tables implied by the SOTA claim) jointly to C-TAO and to MCF–MGA–TCP requires a cleaner factorial design. Existing ablations (Sections V-C, V-D) appear to mix continuous data with new modules. A controlled comparison—same detector backbone and training protocol, (i) original TAO / synthetic pairs vs. C-TAO, and (ii) COVTrack baseline vs. +MGA vs. +TCP on fixed C-TAO—would isolate whether continuous labels or the new modules drive the bulk of the improvement and would address the risk that denser annotations alone explain the LocA lift.

Circularity Check

1 steps flagged

No derivation circularity: SOTA TETA claims rest on external TAO/BDD100K benchmarks and ablated modules, not on quantities defined to equal the metric or fitted free parameters.

specific steps
  1. self citation load bearing [Extension from Conference Version; Abstract / Sec. I contributions; Table VII (BDD100K)]
    "This paper is an extension of our prior work [13], namely COVTrack, published at ICCV 2025. The major extensions include: ›Novel hierarchical aggregation mechanism: We propose MGA ... ›Novel temporal propagation mechanism: We propose TCP ... ›Improved performance: The extended framework achieves substantially stronger results, with novel TETA improving from 34.3% to 35.4% (+1.1%) on TAO validation and from 28.9% to 30.5% (+1.6%) on TAO test"

    The paper's baseline and part of the association pipeline are the authors' own prior COVTrack. This is disclosed and the new modules are ablated, so it is not a hidden uniqueness import; it is only a minor self-citation burden on the incremental gains, not a reduction of the SOTA claim to a definition.

full rationale

This is an empirical computer-vision systems paper, not a first-principles derivation. The claimed chain is: (1) construct denser continuous labels C-TAO from TAO training videos; (2) train MCF/MGA/TCP modules that fuse cues, aggregate parent-child detections, and propagate confidence; (3) evaluate novel TETA / AssocA / LocA on held-out TAO validation/test and zero-shot BDD100K. None of these steps is circular by construction. TETA-style metrics are standard external benchmarks; the modules are not defined so that their outputs equal the reported scores; and no free parameter is fitted to a subset of the evaluation metric and then re-reported as a prediction. The only self-reference is the disclosed extension of the authors' own COVTrack (ICCV 2025), with incremental modules (MGA, TCP) and ablations; that is ordinary incremental work, not a load-bearing uniqueness theorem or ansatz smuggled in as external fact. C-TAO quality and unreproducible gains are real correctness/reproducibility risks, but they do not make the reported numbers tautological. Score 1 for minor self-extension only.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The central empirical claim rests on standard tracking-by-detection assumptions, a frozen open-vocabulary detector, continuous manual labels on TAO train, and several design choices in fusion/aggregation/propagation. No new physical entities; free parameters are the usual deep-learning and association thresholds rather than constants fitted to invent a law.

free parameters (4)
  • Detection confidence thresholds and TCP recovery threshold
    Low-confidence candidates are revived only when temporal matching and propagated confidence exceed design thresholds; these cutoffs directly affect LocA/AssocA and are not uniquely determined by theory.
  • MCF dual-perspective confidence fusion weights / temperature
    Intra- and inter-frame confidence estimators and their combination into adaptive cue weights are learned or tuned; reliability balancing is not parameter-free.
  • MGA parent-child spatial relation criteria (IoU/containment filters)
    Which boxes count as children of a parent and which children are filtered as unreliable are rule- or score-based choices that shape hierarchical aggregation.
  • Association matching cost / bipartite matching hyperparameters
    Standard MOT association costs and graph matching settings control ID assignment and TCP edges; values are chosen for the benchmark.
axioms (5)
  • domain assumption Tracking-by-detection: per-frame detections can be associated into identities over time.
    Stated throughout Sections I–IV as the operating paradigm for OVMOT.
  • domain assumption A frozen open-vocabulary detector (as in OVTrack/SLAck) provides usable appearance and semantic features for novel classes.
    Section IV-B: backbone kept frozen for fair comparison; all cues build on its outputs.
  • ad hoc to paper Dense open-vocabulary detections form useful parent-child spatial hierarchies (parts help wholes under occlusion).
    Core premise of MGA (Section IV-C); not universally true for all category pairs or detectors.
  • domain assumption C-TAO continuous manual labels faithfully extend original TAO trajectories without systematic identity or category error.
    Section III construction and quality protocol; all continuous-supervision claims depend on it.
  • domain assumption Temporal cycle consistency / bipartite matching across adjacent frames is a reliable signal of feature stability and detection recovery.
    Used in MCF inter-frame confidence and TCP (Sections IV-B, IV-D).
invented entities (3)
  • C-TAO continuous training dataset no independent evidence
    purpose: Provide frame-by-frame boxes, IDs, and categories so models can learn motion dynamics and intermediate states for OVMOT.
    New resource constructed by the authors; independent utility is partially supported by the paper's claim that other methods also improve when trained on it, but public release is still promised.
  • Multi-Granularity Hierarchical Aggregation (MGA) module no independent evidence
    purpose: Aggregate reliable child-part features into parent association features under occlusion.
    Architectural construct introduced in this extension; evidence is ablation/qualitative only within the paper.
  • Temporal Confidence Propagation (TCP) module no independent evidence
    purpose: Recover flickering low-confidence detections via graph matching to high-confidence tracks.
    Architectural construct of this extension; validated only by internal ablations and visualizations.

pith-pipeline@v1.1.0-grok45 · 18047 in / 3437 out tokens · 31292 ms · 2026-07-13T19:07:30.956432+00:00 · methodology

0 comments
read the original abstract

Multi-Object Tracking (MOT) has traditionally focused on a few specific categories, restricting its applicability to real-world scenarios involving diverse objects. Open-Vocabulary Multi-Object Tracking (OVMOT) addresses this by enabling tracking of arbitrary categories, including novel objects unseen during training. However, current progress is constrained by two challenges: the lack of continuously annotated video data for training, and the lack of a customized OVMOT framework to synergistically handle detection and association. We address the data bottleneck by constructing C-TAO, the first continuously annotated training set for OVMOT, which increases annotation density by 26x over the original TAO and captures smooth motion dynamics and intermediate object states. For the framework bottleneck, we propose COVTrack++, a synergistic framework that achieves a bidirectional reciprocal mechanism between detection and association through three modules: (1) Multi-Cue Adaptive Fusion (MCF) dynamically balances appearance, motion, and semantic cues for association feature learning; (2) Multi-Granularity Hierarchical Aggregation (MGA) exploits hierarchical spatial relationships in dense detections, where visible child nodes (e.g., object parts) assist occluded parent objects (e.g., whole body) for association feature enhancement; (3) Temporal Confidence Propagation (TCP) recovers flickering detections through high-confidence tracked objects boosting low-confidence candidates across frames, stabilizing trajectories. Extensive experiments on TAO demonstrate state-of-the-art performance, with novel TETA reaching 35.4% and 30.5% on validation and test sets, improving novel AssocA by 4.8% and novel LocA by 5.8% over previous methods, and show strong zero-shot generalization on BDD100K.

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

49 extracted references · 5 linked inside Pith

  1. [1]

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

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

    Pith/arXiv arXiv 2003

  2. [2]

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

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

    2022

  3. [3]

    Vision meets robotics: The kitti dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The kitti dataset,”The international journal of robotics research, vol. 32, no. 11, pp. 1231–1237, 2013

    2013

  4. [4]

    Ov- track: Open-vocabulary multiple object tracking,

    S. Li, T. Fischer, L. Ke, H. Ding, M. Danelljan, and F. Yu, “Ov- track: Open-vocabulary multiple object tracking,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 5567–5577

    2023

  5. [5]

    Matching anything by segmenting anything,

    S. Li, L. Ke, M. Danelljan, L. Piccinelli, M. Segu, L. Van Gool, and F. Yu, “Matching anything by segmenting anything,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 18 963–18 973

    2024

  6. [6]

    Slack: Semantic, location, and appearance aware open- vocabulary tracking,

    S. Li, L. Ke, Y .-H. Yang, L. Piccinelli, M. Seg `u, M. Danelljan, and L. Van Gool, “Slack: Semantic, location, and appearance aware open- vocabulary tracking,” inProceedings of the European conference on computer vision (ECCV), 2024

    2024

  7. [7]

    Open-vocabulary multi-object tracking with domain generalized and temporally adaptive features,

    R. Li, D. Zhang, Y . Wang, Y . Jiang, Z. Zheng, S.-W. Jeon, and H. Wang, “Open-vocabulary multi-object tracking with domain generalized and temporally adaptive features,”IEEE Transactions on Multimedia, 2025

    2025

  8. [8]

    Ovtr: End-to-end open-vocabulary multiple object tracking with transformer,

    J. Li, E. Yu, S. Chen, and W. Tao, “Ovtr: End-to-end open-vocabulary multiple object tracking with transformer,” inThe Thirteenth Interna- tional Conference on Learning Representations (ICLR), 2025, arXiv preprint arXiv:2503.10616

    Pith/arXiv arXiv 2025

  9. [9]

    Attention to trajec- tory: Trajectory-aware open-vocabulary tracking,

    Y . Li, Y . Jiao, D. Meng, H. Fan, and L. Zhang, “Attention to trajec- tory: Trajectory-aware open-vocabulary tracking,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025

    2025

  10. [10]

    Lvis: A dataset for large vocabulary instance segmentation,

    A. Gupta, P. Dollar, and R. Girshick, “Lvis: A dataset for large vocabulary instance segmentation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 5356– 5364

    2019

  11. [11]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment anything,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4015–4026

    2023

  12. [12]

    Tao: A large-scale benchmark for tracking any object,

    A. Dave, T. Khurana, P. Tokmakov, C. Schmid, and D. Ramanan, “Tao: A large-scale benchmark for tracking any object,” inComputer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16. Springer, 2020, pp. 436–454

    2020

  13. [13]

    Covtrack: Continuous open-vocabulary tracking via adaptive multi-cue fusion,

    Z. Qian, R. Han, Z. Wang, J. Hou, and W. Feng, “Covtrack: Continuous open-vocabulary tracking via adaptive multi-cue fusion,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 10 054–10 063

    2025

  14. [14]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning,

    F. Yu, H. Chen, X. Wang, W. Xian, Y . Chen, F. Liu, V . Madhavan, and T. Darrell, “Bdd100k: A diverse driving dataset for heterogeneous multitask learning,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 2636–2645

    2020

  15. [15]

    People-tracking-by-detection and people-detection-by-tracking,

    M. Andriluka, S. Roth, and B. Schiele, “People-tracking-by-detection and people-detection-by-tracking,” in2008 IEEE Conference on com- puter vision and pattern recognition. IEEE, 2008, pp. 1–8

    2008

  16. [16]

    Tracking without bells and whistles,

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

    2019

  17. [17]

    Qdtrack: Quasi-dense similarity learning for appearance-only multiple object tracking,

    T. Fischer, T. E. Huang, J. Pang, L. Qiu, H. Chen, T. Darrell, and F. Yu, “Qdtrack: Quasi-dense similarity learning for appearance-only multiple object tracking,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    2023

  18. [18]

    Learning by track- ing: Siamese cnn for robust target association,

    L. Leal-Taix ´e, C. Canton-Ferrer, and K. Schindler, “Learning by track- ing: Siamese cnn for robust target association,” inProceedings of the IEEE conference on computer vision and pattern recognition workshops, 2016, pp. 33–40

    2016

  19. [19]

    Quasi- dense similarity learning for multiple object tracking,

    J. Pang, L. Qiu, X. Li, H. Chen, Q. Li, T. Darrell, and F. Yu, “Quasi- dense similarity learning for multiple object tracking,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 164–173

    2021

  20. [20]

    Tracking the untrackable: Learning to track multiple cues with long-term dependencies,

    A. Sadeghian, A. Alahi, and S. Savarese, “Tracking the untrackable: Learning to track multiple cues with long-term dependencies,” inPro- ceedings of the IEEE international conference on computer vision, 2017, pp. 300–311

    2017

  21. [21]

    Simple online and realtime tracking with a deep association metric,

    N. Wojke, A. Bewley, and D. Paulus, “Simple online and realtime tracking with a deep association metric,” in2017 IEEE international conference on image processing (ICIP). IEEE, 2017, pp. 3645–3649. 17

    2017

  22. [22]

    Memot: Multi-object tracking with memory,

    J. Cai, M. Xu, W. Li, Y . Xiong, W. Xia, Z. Tu, and S. Soatto, “Memot: Multi-object tracking with memory,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 8090–8100

    2022

  23. [23]

    Tracking objects as points,

    X. Zhou, V . Koltun, and P. Kr ¨ahenb¨uhl, “Tracking objects as points,” in European conference on computer vision. Springer, 2020, pp. 474–490

    2020

  24. [24]

    Probabilistic tracklet scoring and inpainting for multiple object track- ing,

    F. Saleh, S. Aliakbarian, H. Rezatofighi, M. Salzmann, and S. Gould, “Probabilistic tracklet scoring and inpainting for multiple object track- ing,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 14 329–14 339

    2021

  25. [25]

    Simple baselines for human pose estimation and tracking,

    B. Xiao, H. Wu, and Y . Wei, “Simple baselines for human pose estimation and tracking,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 466–481

    2018

  26. [26]

    Motiontrack: Learning robust short-term and long-term motions for multi-object tracking,

    Z. Qin, S. Zhou, L. Wang, J. Duan, G. Hua, and W. Tang, “Motiontrack: Learning robust short-term and long-term motions for multi-object tracking,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 17 939–17 948

    2023

  27. [27]

    Strong- sort: Make deepsort great again,

    Y . Du, Z. Zhao, Y . Song, Y . Zhao, F. Su, T. Gong, and H. Meng, “Strong- sort: Make deepsort great again,”IEEE Transactions on Multimedia, 2023

    2023

  28. [28]

    Delving into motion-aware matching for monocular 3d object tracking,

    K.-C. Huang, M.-H. Yang, and Y .-H. Tsai, “Delving into motion-aware matching for monocular 3d object tracking,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 6909–6918

    2023

  29. [29]

    Track to reconstruct and reconstruct to track,

    J. Luiten, T. Fischer, and B. Leibe, “Track to reconstruct and reconstruct to track,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 1803– 1810, 2020

    2020

  30. [30]

    Camo-mot: Combined appearance-motion optimization for 3d multi-object tracking with camera-lidar fusion,

    L. Wang, X. Zhang, W. Qin, X. Li, J. Gao, L. Yang, Z. Li, J. Li, L. Zhu, H. Wanget al., “Camo-mot: Combined appearance-motion optimization for 3d multi-object tracking with camera-lidar fusion,” IEEE Transactions on Intelligent Transportation Systems, 2023

    2023

  31. [31]

    Pedestrian tracking with monocular camera using unconstrained 3d motion model,

    J. Krej ˇc´ı, O. Kost, O. Straka, and J. Dun ´ık, “Pedestrian tracking with monocular camera using unconstrained 3d motion model,”arXiv preprint arXiv:2403.11978, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    Track, then decide: Category-agnostic vision-based multi-object tracking,

    A. O ˇsep, W. Mehner, P. V oigtlaender, and B. Leibe, “Track, then decide: Category-agnostic vision-based multi-object tracking,” in2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2018, pp. 3494–3501

    2018

  33. [33]

    Simple online and realtime tracking,

    A. Bewley, Z. Ge, L. Ott, F. Ramos, and B. Upcroft, “Simple online and realtime tracking,” in2016 IEEE international conference on image processing (ICIP). IEEE, 2016, pp. 3464–3468

    2016

  34. [34]

    Walker: self-supervised multiple object tracking by walking on tem- poral appearance graphs,

    M. Segu, L. Piccinelli, S. Li, L. Van Gool, F. Yu, and B. Schiele, “Walker: self-supervised multiple object tracking by walking on tem- poral appearance graphs,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 1–18

    2024

  35. [35]

    Towards real-time multi-object tracking,

    Z. Wang, L. Zheng, Y . Liu, Y . Li, and S. Wang, “Towards real-time multi-object tracking,” inEuropean conference on computer vision. Springer, 2020, pp. 107–122

    2020

  36. [36]

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

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

    2022

  37. [37]

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

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

    2021

  38. [38]

    Track- former: Multi-object tracking with transformers,

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

    2022

  39. [39]

    Transtrack: Multiple object tracking with transformer,

    P. Sun, J. Cao, Y . Jiang, R. Zhang, E. Xie, Z. Yuan, C. Wang, and P. Luo, “Transtrack: Multiple object tracking with transformer,”arXiv preprint arXiv:2012.15460, 2020

    Pith/arXiv arXiv 2012

  40. [40]

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

    F. Zeng, B. Dong, Y . Zhang, T. Wang, X. Zhang, and Y . Wei, “Motr: End-to-end multiple-object tracking with transformer,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 659–675

    2022

  41. [41]

    1st place solution to eccv-tao-2020: Detect and represent any object for tracking,

    F. Du, B. Xu, J. Tang, Y . Zhang, F. Wang, and H. Li, “1st place solution to eccv-tao-2020: Detect and represent any object for tracking,”arXiv preprint arXiv:2101.08040, 2021

    Pith/arXiv arXiv 2020

  42. [42]

    Global tracking transformers,

    X. Zhou, T. Yin, V . Koltun, and P. Kr ¨ahenb¨uhl, “Global tracking transformers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 8771–8780

    2022

  43. [43]

    Tracking every thing in the wild,

    S. Li, M. Danelljan, H. Ding, T. E. Huang, and F. Yu, “Tracking every thing in the wild,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 498–515

    2022

  44. [44]

    Taking mobile multi-object tracking to the next level: People, unknown objects, and carried items,

    D. Mitzel and B. Leibe, “Taking mobile multi-object tracking to the next level: People, unknown objects, and carried items,” inComputer Vision– ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12. Springer, 2012, pp. 566–579

    2012

  45. [45]

    Opening up open world tracking,

    Y . Liu, I. E. Zulfikar, J. Luiten, A. Dave, D. Ramanan, B. Leibe, A. Oˇsep, and L. Leal-Taix´e, “Opening up open world tracking,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 19 045–19 055

    2022

  46. [46]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PMLR, 2021, pp. 8748–8763

    2021

  47. [47]

    Open-vocabulary object detec- tion via vision and language knowledge distillation,

    X. Gu, T.-Y . Lin, W. Kuo, and Y . Cui, “Open-vocabulary object detec- tion via vision and language knowledge distillation,” inInternational Conference on Learning Representations, 2022

    2022

  48. [48]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  49. [49]

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

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

    2023