Pith. sign in

REVIEW 3 major objections 5 minor 45 references

Glance-MCMT: A General MCMT Framework with Glance Initialization and Progressive Association

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Glance-MCMT claims that a short glance phase across overlapping cameras can fix global identities and hold identity growth in check.

desk verdict Workmanlike integration of known MCMT components with a plausible but unproven validation-set result; the glance-phase assumption needs stress-testing. read the letter →

arxiv 2507.10115 v1 pith:EO2TWBYO submitted 2025-07-14 cs.CV

Hamidreza Hashempoor This is my paper
classification cs.CV
keywords multi-cameramulti-targettrackingglobalidentityassignmentglanceinitializationprogressiveassociationBoT-SORTrepresentativefeaturesHOTAmetricchallenge
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

This paper proposes a multi-camera multi-target tracking system that assigns one global identity per physical object across overlapping camera views. Its strategy is to spend a brief glance phase on the first frames from all cameras to establish the complete set of objects, then associate every later local tracklet to one of those global identities using both world-coordinate trajectory similarity and appearance similarity. New global IDs are created only when a tracklet matches no existing identity under strict thresholds, which is meant to prevent identity inflation. On the validation set of a 2025 multi-camera tracking challenge, the pipeline reports HOTA 51.34 with the global-ID-expansion strategy and 43.09 with forced matching. A sympathetic reader would care because the design is modular, class-agnostic, and extendable to online tracking, not because it claims a new theory.

What carries the argument

The load-bearing mechanism is the pair of association gates applied at both stages: trajectory similarity computed from camera-calibration projections of bounding boxes into world coordinates, and appearance similarity computed as cosine distance between DBSCAN-selected representative re-identification embeddings per tracklet. Only tracklets passing both gates are merged into one global ID. The glance phase fixes the initial global count after the first few frames, and the progressive phase reuses the same two gates while allowing one global ID to appear once per view at any given time; leftover tracklets are either force-matched or seed new IDs through GIDE.

What would settle it

Take a validation sequence and deliberately withhold an object from all cameras during the glance window, then let it appear later; if the pipeline assigns that object a brand-new global ID while its true identity already exists from another view, the completeness assumption has been violated and identity fragmentation can be counted directly.

Watch

Extended reading notes

Core claim

The central claim is that global identity assignment across overlapping views can be driven by a two-stage association: first, a glance initialization that merges tracklets across cameras into a fixed set of global IDs using both projected world-space trajectories and re-identification appearance features; second, a progressive association that repeatedly attaches newly available local tracklets to existing global tracklets when their temporally overlapping trajectory segments and appearance features are similar. Identity creation is deliberately conservative: only after all strong matches are exhausted may one remaining long unassigned tracklet become a new global ID, which the paper calls global ID expansion (GIDE). Reported on the validation set, the GIDE variant achieves HOTA 51.34 with DetA 48.19, AssA 54.71, and LocA 66.19, while the forced-matching variant achieves 43.09 with DetA 45.12, AssA 41.16, and LocA 61.58. The comparison is offered as evidence that controlling when new IDs are introduced matters more to identity-consistent tracking than forcing a one-to-one assignment for leftover tracklets.

Load-bearing premise

The load-bearing premise is that during the brief glance window every object in the environment appears in at least one camera view, so the initial global ID set is complete and later ID creation can be kept rare.

Editorial extensions

If this is right

  • If the two-gate matching is sound, the same code path handles both initialization and ongoing association, so adding a camera or a target class requires no new matching logic.
  • Because the glance phase fixes the global count, identity inflation is bounded: no new ID is created unless the longest remaining tracklet cannot be matched, which is exactly the regime where GIDE outperforms forced matching on the validation set.
  • The reported gap between HOTA 51.34 and 43.09 shows that on this benchmark, allowing conservative new-ID creation costs less identity accuracy than blindly assigning leftover tracklets.
  • The stated online extension follows from the design: keep a running history of tracklets, repeat the glance initialization from the first frames, and match new tracklets progressively as they appear.

Reading between the lines

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

  • Beyond the paper's claims: if the glance-completeness assumption fails in a deployment, the method will not detect the failure at initialization; the first sign would be identity splits later, because GIDE can only add new IDs from orphaned tracklets, which may be the same physical object under a second identity.
  • Beyond the paper's claims: the FM-versus-GIDE comparison suggests a design rule for other multi-camera pipelines, namely to make the ID-creation rate an explicit control variable rather than an emergent artifact of clustering; that comparison is a minimal experiment of the variable.
  • Beyond the paper's claims: in scenes where cameras do not overlap, the glance assumption cannot hold, so a straightforward extension would replace the glance phase with re-acquisition after disappearance, a regime the paper does not address.
  • Beyond the paper's claims: because trajectory similarity depends on projecting 2D boxes to world coordinates, the method's ceiling is tied to calibration and box-localization accuracy, and the relatively high LocA compared with AssA hints that spatial alignment is currently the stronger component of the reported HOTA.
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 proposes Glance-MCMT, a modular multi-camera multi-target (MCMT) tracking pipeline. The pipeline begins with a YOLOX detector retrained on the target scenes and BoT-SORT for single-camera tracking, followed by tracklet post-processing via sequential non-maximum suppression and DBSCAN-based representative feature selection. Global identity assignment is performed in two stages: a glance initialization that uses the first few frames from all cameras to construct an initial set of global tracklets via world-coordinate trajectory similarity and appearance similarity, and a progressive association stage that iteratively matches remaining local tracklets to existing global identities. The authors evaluate two strategies for handling unassigned tracklets, forced matching (FM) and global ID expansion (GIDE), reporting HOTA scores of 43.09 and 51.34, respectively, on a single validation split of the AI City Challenge 2025 MTMC Tracking dataset. The central claim is that this framework ensures consistent global identity assignment across overlapping views while controlling identity inflation.

Significance. If the claims were fully supported, the framework would be a useful modular baseline for MCMT tracking, and the public implementation (github link in Section 4.3) is a practical strength. The glance initialization idea is interesting because it decouples initial identity creation from progressive association. However, the evidence presented is far too limited to establish the central claims of identity consistency, controlled ID growth, or generality: there is no comparison to prior MCMT methods, no ablation of the key design choices, no error analysis, and only one synthetic validation split. The contribution is potentially valuable but currently under-validated.

major comments (3)
  1. [Section 3.2.1] The load-bearing assumption in Section 3.2.1 that every object appears in at least one view during the glance window is never tested. The paper does not report the glance window length, does not vary it, and does not measure how many objects enter the scene after the glance phase. Since the design fixes the global ID count after glance and the only recovery mechanism is the GIDE branch, the reported HOTA of 51.34 in Table 1 cannot be interpreted without knowing the frequency of late-appearing objects and whether they are correctly assigned. An experiment that varies the glance window length (e.g., 5, 15, 30, 60 frames) and separately evaluates tracklets of objects present at the start versus later entries is required to support the identity-consistency claim.
  2. [Table 1 and Section 4.4] The experimental section reports only a single validation split with two configurations of the same pipeline. There is no comparison to any baseline, prior MCMT method, or even a simple clustering-based approach, and no error bars or multiple runs. Furthermore, the method's many free parameters (SNMS spatial and temporal thresholds, DBSCAN eps and min_samples, representative feature count, trajectory/feature similarity thresholds, interpolation missing-frame threshold) are not ablated. As a result, the central claim that Glance-MCMT achieves 'consistent global identity assignment' and 'competitive results' is not supported by the evidence presented. At minimum, the authors should compare against a standard hierarchical clustering baseline on the same validation split, report variance across scenes, and ablate the glance window and similarity thresholds.
  3. [Section 4.4 (GIDE)] The Global ID Expansion strategy creates a new global ID from the longest unassigned local tracklet. This rule is not based on any evidence that the tracklet represents a distinct object; it could equally convert a failed match or a fragmented tracklet into a spurious new identity. Because the paper does not report the number of GIDE events or the resulting ID switches/fragmentation, the improvement from FM (43.09) to GIDE (51.34) in Table 1 may be driven by a trade-off that hides identity errors rather than by principled association. The authors should analyze the frequency and correctness of GIDE insertions, for example by checking whether newly created global IDs correspond to ground-truth objects that were not visible during the glance phase.
minor comments (5)
  1. [Section 3.2.2] The text first states that progressive association continues 'until all local tracklets are assigned and no unassociated segments remain' and then, a few sentences later, states that 'a few local tracklets may still remain unassigned.' This contradiction should be resolved with a precise stopping criterion, as it affects the termination condition of the algorithm.
  2. [Section 4.3 and Section 4.4] There are several language and typographical issues, including 'Implementation is avalable' (Section 4.3) and 'taking average in Table 1' (Section 4.4). These should be corrected during revision.
  3. [Section 3.1] The representative feature count is described as 'e.g., 20,' which is vague. The paper should state the exact number used in the experiments and whether it was tuned.
  4. [Section 2.1] Reference [25] is cited as part of the YOLO family, but the actual reference [25] is a paper on occlusion handling in multi-camera people tracking; the citation numbering appears to be incorrect.
  5. [Figure 1] The text in Figure 1 mentions 'G3' and 'G5' without any explanation in the caption or body; please clarify what these labels denote.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the pipeline's claims are empirical, and each stage consumes distinct inputs rather than fitting the reported HOTA.

full rationale

Glance-MCMT contains no derivation chain in which an output is equal by construction to an input. The glance phase (Sec. 3.2.1) constructs global IDs from BoT-SORT tracklets using world-coordinate trajectory proximity and ReID feature similarity; progressive association (Sec. 3.2.2) matches remaining local tracklets with the same cues, and the GIDE branch adds new IDs only for unmatched longest tracklets. None of these quantities is defined in terms of the reported HOTA (Sec. 4.4, Table 1), DetA, AssA, or LocA, so the evaluation is an external benchmark measurement rather than a restatement of the method's thresholds or fitted parameters. The main caveats—the completeness assumption that every object appears during the glance window, and the absence of an ablation varying the glance window—are soundness/robustness concerns, not circularity: the components could fail and the HOTA number would simply drop. Self-citations [11,16,17,18,19] appear only as background or as an implementation detail (interpolation 'similar to [16]'); none is invoked as the justification for the central claim, and no uniqueness theorem or prior-work result is imported to force the design. The reported GIDE result (HOTA 51.34) is an empirical outcome on the AI City Challenge 2025 validation set and would be reproducible from held-out data; the method is not a renamed known result wrapped in new coordinates. Score 0 therefore is appropriate.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The pipeline introduces no new physical entities or theoretical constructs. All modules are existing methods, and the paper's novel surface is the ordering and selection of stages. The main unstated costs appear as free thresholds and the two domain assumptions listed above.

free parameters (7)
  • SCMT NMS spatial overlap threshold
    Equation (1) defines the overlap coefficient, but the actual spatial threshold for merging tracklets is not reported.
  • SCMT NMS temporal overlap threshold
    Section 3.1 requires exceeding both spatial and temporal thresholds, but the temporal threshold is not given.
  • DBSCAN eps and min_samples
    Section 3.1 uses DBSCAN clustering on ReID embeddings, but neither eps nor min_samples is specified.
  • Representative feature count = 20
    Section 3.1 says 'a fixed number of samples (e.g., 20)' are chosen from the dominant cluster; the count is a free choice.
  • Trajectory similarity threshold
    Glance initialization and progressive association both require trajectory similarity above a 'predefined threshold', but the threshold value is not reported.
  • Appearance cosine similarity threshold
    Feature matching uses cosine similarity with a threshold that is never quantified in the paper.
  • Interpolation missing-frame threshold
    Section 4.3 interpolates only when consecutive missing frames are below a fixed threshold; the threshold value is not given.
assumptions (4)
  • domain assumption Every object appears in at least one camera view during the glance window.
    Section 3.2.1 explicitly assumes this; it is load-bearing for constructing a complete set of initial global tracklets.
  • domain assumption All local tracklets sharing temporal overlap with a global tracklet correspond to an existing global identity.
    Section 3.2.2 makes this assumption before performing progressive association; if false, matching constraints become invalid.
  • standard math Camera calibration matrices accurately project 2D boxes into a shared world coordinate frame.
    Sections 3.2.1 and 4.2 rely on this projection for trajectory similarity and for the 3D HOTA evaluation.
  • domain assumption BoT-SORT single-camera tracks and OSNet ReID embeddings are reliable upstream inputs.
    No error analysis of single-camera tracking or re-identification is provided, yet the global association quality depends on them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Glance-MCMT: A General MCMT Framework with Glance Initialization and Progressive Association." pith.science (2026). https://pith.science/paper/EO2TWBYO

@misc{pith2026250710115,
  author       = {Pith},
  title        = {Pith review of: Glance-MCMT: A General MCMT Framework with Glance Initialization and Progressive Association},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EO2TWBYO}},
  note         = {Machine review of arXiv:2507.10115}
}
read the original abstract

We propose a multi-camera multi-target (MCMT) tracking framework that ensures consistent global identity assignment across views using trajectory and appearance cues. The pipeline starts with BoT-SORT-based single-camera tracking, followed by an initial glance phase to initialize global IDs via trajectory-feature matching. In later frames, new tracklets are matched to existing global identities through a prioritized global matching strategy. New global IDs are only introduced when no sufficiently similar trajectory or feature match is found. 3D positions are estimated using depth maps and calibration for spatial validation.

Figures

Figures reproduced from arXiv: 2507.10115 by the authors.

Figure 1
Figure 1. Overview of the proposed MCMT pipeline. Objects [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Glance initialization. Trajectory similarity is first com [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the global association strategy across multiple overlapping views. Each local tracklet is matched to a global ID [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 30 canonical work pages

  1. [1]

    Bot- sort: Robust associations multi-pedestrian tracking

    Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Bot- sort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 3

    arXiv 2022

  2. [2]

    A survey of ap- proaches and trends in person re-identification

    Apurva Bedagkar-Gala and Shishir K Shah. A survey of ap- proaches and trends in person re-identification. Image and vision computing, 32(4):270–286, 2014. 2

    work page 2014

  3. [3]

    Simple online and realtime tracking

    Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP) , pages 3464–3468. Ieee, 2016. 2

    work page 2016

  4. [4]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. In European confer- ence on computer vision, pages 213–229. Springer, 2020. 2

    2020

  5. [5]

    Mot20: A benchmark for multi object tracking in crowded scenes

    Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taix ´e. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 2

    arXiv 2003

  6. [6]

    Motchallenge: A benchmark for single-camera multiple target tracking

    Patrick Dendorfer, Aljosa Osep, Anton Milan, Konrad Schindler, Daniel Cremers, Ian Reid, Stefan Roth, and Laura Leal-Taix´e. Motchallenge: A benchmark for single-camera multiple target tracking. International Journal of Computer Vision, 129:845–881, 2021. 2

    work page 2021

  7. [7]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 4690–4699, 2019. 2

    2019

  8. [8]

    Strongsort: Make deep- sort great again

    Yunhao Du, Zhicheng Zhao, Yang Song, Yanyun Zhao, Fei Su, Tao Gong, and Hongying Meng. Strongsort: Make deep- sort great again. IEEE Transactions on Multimedia , 25: 8725–8737, 2023. 3

    work page 2023

Show all 45 references
  1. [9]

    Unsuper- vised pre-training for person re-identification

    Dengpan Fu, Dongdong Chen, Jianmin Bao, Hao Yang, Lu Yuan, Lei Zhang, Houqiang Li, and Dong Chen. Unsuper- vised pre-training for person re-identification. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14750–14759, 2021. 2

    2021

  2. [10]

    Yolox: Exceeding yolo series in 2021

    Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 2, 6

    2021 arXiv

  3. [11]

    Deep-lfd: deep robot learning from demonstrations

    Amir Ghalamzan, Kiyanoush Nazari, Hamidreza Hashempour, and Fangxun Zhong. Deep-lfd: deep robot learning from demonstrations. Software Impacts, 9: 100087, 2021. 2

    2021

  4. [12]

    Fast r-cnn

    Ross Girshick. Fast r-cnn. In Proceedings of the IEEE inter- national conference on computer vision , pages 1440–1448,

  5. [13]

    Rich feature hierarchies for accurate object detection and semantic segmentation

    Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 580–587, 2014. 2

    2014

  6. [14]

    Viewpoint invariant pedestrian recognition with an ensemble of localized features

    Douglas Gray and Hai Tao. Viewpoint invariant pedestrian recognition with an ensemble of localized features. In Com- puter Vision–ECCV 2008: 10th European Conference on Computer Vision, Marseille, France, October 12-18, 2008, Proceedings, Part I 10, pages 262–275. Springer, 2008. 2

    2008

  7. [15]

    Dimensional- ity reduction by learning an invariant mapping

    Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensional- ity reduction by learning an invariant mapping. In2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06), pages 1735–1742. IEEE, 2006. 2

    2006

  8. [16]

    Deep learning based data-assisted channel estimation and detection

    Hamidreza Hashempoor and Wan Choi. Deep learning based data-assisted channel estimation and detection. IEEE Trans- actions on Machine Learning in Communications and Net- working, 2025. 6

    2025

  9. [17]

    Featuresort: essential features for effective tracking

    Hamidreza Hashempoor, Rosemary Koikara, and Yu Dong Hwang. Featuresort: essential features for effective tracking. arXiv preprint arXiv:2407.04249, 2024. 2

    2024

  10. [18]

    Gated in- ference network: Inference and learning state-space mod- els

    Hamidreza Hashempoorikderi and Wan Choi. Gated in- ference network: Inference and learning state-space mod- els. Advances in Neural Information Processing Systems , 37:39036–39073, 2024. 2

    2024

  11. [19]

    A data-set of piercing needle through de- formable objects for deep learning from demonstrations

    Hamidreza Hashempour, Kiyanoush Nazari, Fangxun Zhong, et al. A data-set of piercing needle through de- formable objects for deep learning from demonstrations. arXiv preprint arXiv:2012.02458, 2020. 2

    2012 arXiv

  12. [20]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Gir- shick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017. 2

    2017

  13. [21]

    Deep metric learning using triplet network

    Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-based pattern recognition: third inter- national workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84–92. Springer,

    2015

  14. [22]

    Enhancing multi-camera people tracking with anchor-guided clustering and spatio-temporal consistency id re-assignment

    Hsiang-Wei Huang, Cheng-Yen Yang, Zhongyu Jiang, Pyong-Kun Kim, Kyoungoh Lee, Kwangju Kim, Samartha Ramkumar, Chaitanya Mullapudi, In-Su Jang, Chung-I Huang, et al. Enhancing multi-camera people tracking with anchor-guided clustering and spatio-temporal consistency id re-assig...

    2023

  15. [23]

    Ap- pearance modeling for tracking in multiple non-overlapping cameras

    Omar Javed, Khurram Shafique, and Mubarak Shah. Ap- pearance modeling for tracking in multiple non-overlapping cameras. In 2005 IEEE computer society conference on com- puter vision and pattern recognition (CVPR’05) , pages 26–

    2005

  16. [24]

    Leveraging future trajectory prediction for multi- camera people tracking

    Yuntae Jeon, Dai Quoc Tran, Minsoo Park, and Seunghee Park. Leveraging future trajectory prediction for multi- camera people tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 5399–5408, 2023. 2

    2023

  17. [25]

    Addressing the occlusion problem in multi-camera people tracking with human pose estimation

    Jeongho Kim, Wooksu Shin, Hancheol Park, and Jongwon Baek. Addressing the occlusion problem in multi-camera people tracking with human pose estimation. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5463–5469, 2023. 2 7

    2023

  18. [26]

    Cluster self-refinement for enhanced online multi- camera people tracking

    Jeongho Kim, Wooksu Shin, Hancheol Park, and Donghyuk Choi. Cluster self-refinement for enhanced online multi- camera people tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7190–7197, 2024. 3

    2024

  19. [27]

    Yolov6: A single-stage object detec- tion framework for industrial applications

    Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, Weiqiang Nie, et al. Yolov6: A single-stage object detec- tion framework for industrial applications. arXiv preprint arXiv:2209.02976, 2022. 2

    2022 arXiv

  20. [28]

    Simpletrack: Rethinking and improving the jde approach for multi-object tracking

    Jiaxin Li, Yan Ding, Hua-Liang Wei, Yutong Zhang, and Wenxiang Lin. Simpletrack: Rethinking and improving the jde approach for multi-object tracking. Sensors, 22(15): 5863, 2022. 3

    2022

  21. [29]

    Hierarchical clustering and refinement for generalized multi-camera person tracking

    Zongyi Li, Runsheng Wang, He Li, Bohao Wei, Yuxuan Shi, Hefei Ling, Jiazhong Chen, Boyuan Liu, Zhongyang Li, and Hanqing Zheng. Hierarchical clustering and refinement for generalized multi-camera person tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and...

    2023

  22. [30]

    Hota: A higher order metric for evaluating multi-object tracking

    Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taix´e, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129:548– 578, 2021. 6

    2021

  23. [31]

    Multi-camera people tracking with mixture of realistic and synthetic knowledge

    Quang Qui-Vinh Nguyen, Huy Dinh-Anh Le, Truc Thi- Thanh Chau, Duc Trung Luu, Nhat Minh Chung, and Synh Viet-Uyen Ha. Multi-camera people tracking with mixture of realistic and synthetic knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni...

    2023

  24. [32]

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

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information process- ing systems, 28, 2015. 2

    2015

  25. [33]

    Learning discriminative appearance-based models using partial least squares

    William Robson Schwartz and Larry S Davis. Learning discriminative appearance-based models using partial least squares. In 2009 XXII Brazilian symposium on computer graphics and image processing, pages 322–329. IEEE, 2009. 2

    2009

  26. [34]

    Reidtrack: Reid- only multi-target multi-camera tracking

    Andreas Specker and J ¨urgen Beyerer. Reidtrack: Reid- only multi-target multi-camera tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5442–5452, 2023. 2

    2023

  27. [35]

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

    Chien-Yao Wang, Alexey Bochkovskiy, and Hong- Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7464–7475, 2023. 2

    2023

  28. [36]

    Integrating appearance and spatial- temporal information for multi-camera people tracking

    Wenjie Yang, Zhenyu Xie, Yaoming Wang, Yang Zhang, Xiao Ma, and Bing Hao. Integrating appearance and spatial- temporal information for multi-camera people tracking. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 5260–5269, 2023. 2

    2023

  29. [37]

    Deep learning for person re- identification: A survey and outlook

    Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re- identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872–2893,

  30. [38]

    Overlap suppression clustering for offline multi-camera people tracking

    Ryuto Yoshida, Junichi Okubo, Junichiro Fujii, Masazumi Amakata, and Takayoshi Yamashita. Overlap suppression clustering for offline multi-camera people tracking. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7153–7162, 2024. 3

    2024

  31. [39]

    Utm: A unified multiple object tracking model with identity-aware feature enhancement

    Sisi You, Hantao Yao, Bing-Kun Bao, and Changsheng Xu. Utm: A unified multiple object tracking model with identity-aware feature enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21876–21886, 2023. 2

    2023

  32. [40]

    Dino: Detr with improved denoising anchor boxes for end-to-end object detection

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2

    2022 arXiv

  33. [41]

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

    Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International journal of computer vision, 129:3069–3087, 2021. 2

    2021

  34. [42]

    Bytetrack: Multi-object tracking by associating every detection box

    Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In European conference on computer vision, pages 1–21. Springer, 2022. 6

    2022

  35. [43]

    Scalable person re-identification: A benchmark

    Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jing- dong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international con- ference on computer vision, pages 1116–1124, 2015. 2

    2015

  36. [44]

    Omni-scale feature learning for person re- identification

    Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Omni-scale feature learning for person re- identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3702–3712, 2019. 2, 6

    2019

  37. [45]

    Deformable detr: Deformable trans- formers for end-to-end object detection

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable trans- formers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2 8

    2010 arXiv

Pith tools

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