Pith. sign in

REVIEW 4 major objections 6 minor 27 references

Head Anchor Enhanced Detection and Association for Crowded Pedestrian Tracking

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read FocusTrack matches leading trackers on MOT17 by fusing head keypoints with body-detection features.

desk verdict The iterative Kalman filter in Eq. (19) is internally inconsistent under the paper's own linear measurement model, and the experiments don't back the headline claim—but the feature analysis and the FT25 dataset are worth a second look. read the letter →

arxiv 2508.05514 v1 pith:VLKRCXA4 submitted 2025-08-07 cs.CV

Zewei Wu , César Teixeira , Wei Ke , Zhang Xiong This is my paper · ORCID
classification cs.CV
keywords multi-objecttrackingpedestrianheadkeypointdetectionocclusiontracking-by-detectionKalmanfilterMOT17/MOT20crowdscenes
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes FocusTrack, a multi-object tracker for crowded pedestrian scenes, and argues that combining full-body detector features (classification and regression branches) with head keypoint detection makes identity association more robust under occlusion. The central claim is that head regions stay visible when bodies are occluded, so head keypoints supply occlusion-resistant cues, while full-body features retain the discriminative power needed to tell similar-looking people apart. The tracker also introduces an iterative Kalman filter with 3D priors intended to improve motion prediction during prolonged occlusions. On MOT17 the method reports MOTA 78.3% and IDF1 71.0%, competitive with Bot-SORT without using a dedicated Re-ID network.

What carries the argument

The load-bearing mechanism is the head-aware detector: a YOLOX-style anchor-free detector with an extra head keypoint branch whose classification and regression feature maps, queried at the box center and head position and Gaussian-weighted around the head, provide the appearance cost for the Hungarian association step. The iterative Kalman filter with SE(3) trajectory completion is the supporting motion component.

What would settle it

Run the iterative update in Eq. (19) with the linear H given in Eq. (12) on MOT17 detections and record the residual norm ||z_t - H_t x̂|| across iterations; if it is exactly constant and ||δx^(j+1) - δx^(j)|| never decreases below the threshold in Eq. (20), the claimed iterative Kalman contribution is not realized.

Watch

Extended reading notes

Core claim

The paper's claim is that a tracking-by-detection system can use features taken directly from an anchor-free detector's classification and regression branches, augmented by a head keypoint branch, as a substitute for dedicated Re-ID appearance features, and that this substitution improves tracking stability in crowded scenes. The detector is built on YOLOX and predicts, for each pedestrian box, a head keypoint (x,y) with a visibility score v; the head keypoint is predicted by an independent decoupled branch rather than relative to the box center. During association, the tracker extracts a 15,680-dimensional feature vector from the P3 classification and regression layers at the box center and

Load-bearing premise

The paper's iterative Kalman update requires the innovation to change between iterations, which only happens if the measurement function is nonlinear; the paper defines a linear measurement model, so the update's claimed convergence—and the motion-model improvement it is supposed to justify—does not follow as written.

Editorial extensions

If this is right

  • Trackers can drop dedicated Re-ID networks; detector classification and regression features carry enough identity information to keep associations stable in crowds.
  • Head keypoints extend the observable part of a target, so identity switches should decrease when body occlusion is common and when pedestrians wear similar clothing.
  • The iterative Kalman filter with 3D priors is intended to handle non-linear motion and prolonged occlusion better than the standard linear constant-velocity model.
  • The FT25 dataset, mixing re-annotated MOT20 sequences and CARLA synthetic scenes with head keypoints and pseudo-depth labels, provides a resource for training and benchmarking head-aware trackers.
  • On the paper's reported numbers, a head-enhanced tracker can reach 78.3 MOTA on MOT17 with fewer false positives than ByteTrack (7,723 vs 7,876), indicating the representation also reduces spurious detections.

Reading between the lines

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

  • Pairing head keypoints with a lightweight identity-trained appearance head could strengthen discrimination beyond what the detector's localization features provide.
  • Replacing the pseudo-depth formula z = d + 1/(y+η) with actual depth estimates from a monocular depth network or calibrated stereo might make the SE(3) trajectory completion outperform 2D interpolation, as the paper's ablation suggests 2D currently wins.
  • Extending the association logic toward multi-hypothesis tracking could address the paper's stated limitation about short-term association failing under severe or long occlusions.
  • Fine-tuning the entire detector (including the 2D branches) on the target crowd with the head keypoint supervision, rather than freezing the 2D layers, may reveal larger gains than the reported benchmark numbers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes FocusTrack, a tracking-by-detection framework for crowded pedestrian tracking. The main contributions are: (i) a detector that adds a head keypoint branch to YOLOX, (ii) an association cost built from detector classification/regression/head features rather than a dedicated Re-ID model, (iii) an iterative Kalman filter with 3D priors and SE(3)-based trajectory completion, and (iv) a new dataset FT25 combining re-annotated MOT20 sequences and CARLA-simulated data. The paper reports benchmark results on MOT17, MOT20, and FT25, concluding that FocusTrack achieves performance competitive with existing trackers such as Bot-SORT and ByteTrack.

Significance. If the contributions were sound, the detector-feature analysis and the FT25 dataset would be useful empirical assets: the paper documents that YOLOX classification and regression features can provide discriminative target representations without a separate Re-ID network, and the FT25 dataset with head keypoint annotations and a CARLA-based collection pipeline could support future occlusion-focused tracking research. However, the two central technical claims of the paper are not supported. The iterative Kalman update in Section 3.2 is mathematically inconsistent with the stated linear measurement model, and the trajectory-completion experiment in Section 5.4 shows that plain 2D linear interpolation outperforms the proposed 3D lifting approach. The benchmark results place FocusTrack at or slightly below existing baselines rather than showing the claimed robustness improvement. The paper is therefore not publishable in its current form.

major comments (4)
  1. [Section 3.2, Eqs. (18)-(20)] The iterative Kalman update is not a valid algorithm under the stated model. Equation (12) defines an 8-dimensional linear state vector, and the text states that 'the measurement matrix H maps the state vector to observations,' so h(x) = Hx. With ⊕ as vector addition, the innovation in Eq. (19) becomes z_t - H_t x̂_{t|t-1} - H_t δx_t^(j) + H_t δx_t^(j) = z_t - H_t x̂_{t|t-1}, independent of j. Consequently, the iteration either leaves δ unchanged or adds the same correction indefinitely, so the convergence criterion in Eq. (20) cannot be met unless the innovation is exactly zero. Additionally, the converged δ is never used to update x̂_{t|t} in Eq. (16); the paper only says to update 'with Equation 16 and Equation 17' without defining the relation. The proposed iterative motion-model contribution is therefore mathematically invalid as written.
  2. [Section 5.4, Figure 13] The trajectory-completion experiment directly contradicts a central claim. Figure 13 shows that 2D linear interpolation performs best among the four methods, and the text states that '2D interpolation delivered better results' because the pseudo-depth model in Eq. (22) 'led to inaccurate representations in 3D space.' This undermines the paper's stated contribution that 'dimensional lifting approach enables more accurate trajectory interpolation' and that 3D priors 'better complete motion trajectories' (Section 3.2). The proposed SE(3) Kalman interpolation is not an improvement over a standard baseline.
  3. [Section 5.5, Tables III-V] The benchmark results do not support the central conclusion of superior robustness. On MOT17 total, FocusTrack achieves MOTA/IDF1 of 78.3/71.0, below Bot-SORT's 78.6/74.0. On MOT20 total, FocusTrack's 82.6/68.9 is slightly below SORT's 82.7/69.3. On the introduced FT25 dataset, FocusTrack's 74.0/82.6 is below ByteTrack's 76.0/83.4. The paper describes these results as 'competitive,' but they are at best parity with, or slightly worse than, existing trackers. Given the additional complexity of the head keypoint branch and the iterative filter, the empirical evidence for the proposed method's benefits is weak.
  4. [Algorithm 1 and Section 3.2] The iterative procedure in Algorithm 1 is under-specified and non-reproducible. Line 20 uses 'measurement residuals r > η' without defining r; the while loop says 'update parameters of kalman filters' but does not state which parameters are updated or how Eqs. (18)-(20) are used; and the connection between the converged δx and the state/covariance update is absent. This makes the central motion model impossible to implement from the paper alone.
minor comments (6)
  1. [Section 3.2] Typo: 'varient SORT framework' should be 'variant SORT framework'.
  2. [Section 5.2] Typo in heading: 'Detector Feature Respresentation' should be 'Representation'. Also, the caption of Figure 10 begins 'llustration' instead of 'Illustration'.
  3. [Table I] Typo: 'theshold' should be 'threshold' (appears twice).
  4. [Section 5.5] The text says FocusTrack maintains 'robust identity preservation with only 1248 identity switches' on MOT17, but Table III shows Bot-SORT achieves 990 IDs and SORT 1254 IDs. Calling 1248 'only' is misleading.
  5. [Section 2.2] The subsection heading 'Pedestrian Tracing' should be 'Pedestrian Tracking'.
  6. [Eq. (22)] The pseudo-depth model z = d + 1/(y + η) introduces parameters d and η with no values, ranges, or sensitivity analysis. This is a free parameter in a claimed parameter-light component.

Circularity Check

1 steps flagged · score 6.0 of 10

Eq. (19)'s 'iterative' Kalman update collapses to the standard one-shot Kalman update under the paper's own linear H; the motion-model pillar is a renamed Kalman filter by construction.

  1. self definitional [Section 3.2, Iterative Motion Prediction (Eqs. 12-20)]
    "the measurement matrix H maps the state vector to observations. ... δx_t^(j+1) = δx_t^(j) + K_t^(j)( z_t − h( x̂_{t|t−1} ⊕ δ x_t^(j) ) + H_t δ x_t^(j) ) ... After convergence at iterations, the final state x̂_t|t and covariance P_t|t updates with Equation 16 and Equation 17."

    Under the paper's own linear measurement model—'the measurement matrix H maps the state vector to observations'—one has h(x̂⊕δ)=H(x̂⊕δ)=Hx̂+Hδ (with ⊕ the vector addition used throughout the Kalman equations). Substituting into the bracketed residual of Eq. (19) yields z_t − H_t x̂_{t|t−1} − H_t δ + H_t δ = z_t − H_t x̂_{t|t−1}, which is independent of j. Thus Eq. (19) iterates δx^(j+1)=δx^(j)+K_t^(j)(z_t−H_t x̂_{t|t−1}), adding the same correction each time. The convergence test Eq. (20) can never be satisfied unless that correction is zero, and the text says the final state is updated by Eqs. (16)-(17), i.e., the ordinary one-shot Kalman update. So the proposed 'iterative Kalman filtering' is the standard Kalman filter by construction; the claimed motion-model innovation is definitionall

full rationale

The paper's central motion-model contribution, the iterative Kalman update, is not an independent advance: with the linear measurement matrix H stated in Section 3.2, the residual in Eq. (19) is constant across iterations, so the iteration either repeats the same correction indefinitely or, when the final update is made through Eqs. (16)-(17), reduces exactly to the standard one-shot Kalman filter. This is a reduction by construction, not a difference in fitted constants. The appearance side—head keypoint regression combined with detector classification and regression features—is evaluated against external benchmarks with controlled detections and does not reduce to its inputs; the hand-set weights and noise covariances are tuned constants, not fitted parameters later relabeled as predictions. The self-citation [20] shares two co-authors with the present paper but is used only as related-work motivation, not as load-bearing support. The paper's own trajectory-completion experiments (Section 5.4) show 2D linear interpolation outperforming the proposed 3D interpolation, further weakening but not circularizing that component. Overall, the circularity is partial and localized to the motion-model pillar, so the score is 6 rather than higher.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities or forces. It introduces a new dataset (FT25) but that is a data artifact, not an invented entity. The free parameters are mostly hand-set weights and ad hoc depth constants, and the principal modeling assumptions are standard surveillance-scene priors.

free parameters (8)
  • alpha
    Weighting factor in the assignment cost (Eq. 2); value not given, chosen by hand.
  • beta = 1e5
    Positional constraint weight in Eq. (2), set to 1e5 to exclude anchors outside the center region.
  • w_app, w_mot
    Balancing coefficients in the association cost matrix (Eq. 11); values not specified, set empirically.
  • d
    Minimum depth term in the pseudo-depth formula (Eq. 22); value not given, ad hoc constant.
  • eta
    Denominator offset in the pseudo-depth formula (Eq. 22) to avoid division by zero; value not given.
  • R, Q
    Measurement and process noise covariances in the Kalman filter, described as determined empirically.
  • eps = 0.01
    Convergence threshold in Eq. (20), typically 0.01.
  • w (patience window), g (gating threshold)
    Algorithm hyperparameters in Algorithm 1; values not given.
assumptions (4)
  • domain assumption The head is anatomically positioned at the highest point and is less prone to occlusion than the full body.
    Introduction and Section 2.3 rely on this to justify head keypoint detection as the occlusion-robust cue.
  • domain assumption Targets move on a common ground plane with the camera mounted at an acute angle, a typical surveillance setup.
    Invoked in Section 3.2 to justify the pseudo-depth formula z = d + 1/(y + eta) for 2D-to-3D lifting.
  • domain assumption A constant-velocity motion model with an 8D state is adequate for pedestrian motion.
    Used in the Kalman filter state definition (Eq. 12) and transition matrix F in Section 3.2.
  • domain assumption Detector classification and regression features contain discriminative target representations suitable for appearance matching.
    The whole feature-selection strategy rests on this hypothesis argued from similarity-matrix visualizations in Section 2.3 and 5.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Head Anchor Enhanced Detection and Association for Crowded Pedestrian Tracking." pith.science (2026). https://pith.science/paper/VLKRCXA4

@misc{pith2026250805514,
  author       = {Pith},
  title        = {Pith review of: Head Anchor Enhanced Detection and Association for Crowded Pedestrian Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VLKRCXA4}},
  note         = {Machine review of arXiv:2508.05514}
}
read the original abstract

Visual pedestrian tracking represents a promising research field, with extensive applications in intelligent surveillance, behavior analysis, and human-computer interaction. However, real-world applications face significant occlusion challenges. When multiple pedestrians interact or overlap, the loss of target features severely compromises the tracker's ability to maintain stable trajectories. Traditional tracking methods, which typically rely on full-body bounding box features extracted from {Re-ID} models and linear constant-velocity motion assumptions, often struggle in severe occlusion scenarios. To address these limitations, this work proposes an enhanced tracking framework that leverages richer feature representations and a more robust motion model. Specifically, the proposed method incorporates detection features from both the regression and classification branches of an object detector, embedding spatial and positional information directly into the feature representations. To further mitigate occlusion challenges, a head keypoint detection model is introduced, as the head is less prone to occlusion compared to the full body. In terms of motion modeling, we propose an iterative Kalman filtering approach designed to align with modern detector assumptions, integrating 3D priors to better complete motion trajectories in complex scenes. By combining these advancements in appearance and motion modeling, the proposed method offers a more robust solution for multi-object tracking in crowded environments where occlusions are prevalent.

Figures

Figures reproduced from arXiv: 2508.05514 by the authors.

Figure 1
Figure 1. Illustration of occlusion events in real-world ap￾plications. The green bounding boxes indicate the full-body detection of pedestrians. The red boxes highlight the head regions of targets. The blue lines demonstrate the spatial relationships between head and body regions. To address these challenges, we propose a new track￾ing framework that combines object detection with head keypoint regression to enhance feature … view at source ↗
Figure 2
Figure 2. 𝑏 captures the influence of spatial relationships on t = 1 t = 4 t = 1 t = 4 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of Heatmaps from different features using LayerCAM. First Author et al.: Preprint submitted to Elsevier Page 3 of 13 [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Illustration of the detector model. The backbone network employs CSPDarknet53 architecture, while the neck network utilizes PAFPNet structure. As illustrated in the figure, the detection heads are decoupled into separate branches for classification, regression, and hea…
Figure 5
Figure 5. Figure 5: Illustration of the false due to strong motion prior. Trajectory Completion via Dimensional Lifting For non-real-time scenarios, we propose a trajectory completion mechanism based on dimensional lifting to handle occlu￾sions and missing detections. The approach lifts 2…
Figure 6
Figure 6. Figure 6: Illustration of the annotated examples from FT25-01 and FT25-02 sequences. First Author et al.: Preprint submitted to Elsevier Page 7 of 13 [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Illustration of the annotated examples from FT25-03 and FT25-04 sequences. To comprehensively evaluate our proposed method, we constructed a simple simulated dataset FT25-03, FT25-04 using the CARLA simulation platform, illustrated in Fig￾ure 7. The training and valida…
Figure 9
Figure 9. Figure 9: Illustration of examples from our simulated scenario. The figure illustrates diverse viewpoints and scene samples from our dataset, along with corresponding annotation in￾formation provided for comprehensive evaluation. The data collection pipeline also supports the ge…
Figure 8
Figure 8. Figure 8: Illustration of the systematic pipeline of our dataset collection. This process consists of initialization, pedestrian generation, data acquisition, and subsequent processing stages. As demonstrated in [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 10
Figure 10. Figure 10: llustration of different detector features for Re-ID task. target features from each frame. The similarity between these features is computed using cosine distance and vi￾sualized through similarity matrices. The figure displays similarity matrices between targets acr…
Figure 11
Figure 11. Figure 11: Illustration of different scale features for Re-ID task. To better investigate the impact of feature scale on sim￾ilarity discrimination, we conducted additional comparative experiments by sampling features from the FT03 dataset. The comparison results are illustrated…
Figure 12
Figure 12. Figure 12: Illustration of different P3 features for Re-ID task. The weighted head features enhance the original box-based features through Gaussian weighting centered on the head keypoint position. This weighting mechanism emphasizes the spatial information around the head regi…
Figure 13
Figure 13. Figure 13: Performances of different post-process with in￾terpolation. The figure illustrates four distinct interpolation approaches: 2D linear interpolation, 3D linear interpolation, 3D SE(3) linear interpolation, and 3D SE(3) Kalman filter-based interpolation. We compared the …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 19 canonical work pages

  1. [1]

    Ciaparrone, F

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

    work page 2020

  2. [2]

    W. Luo, J. Xing, A. Milan, X. Zhang, W. Liu, T.-K. Kim, Multiple objecttracking:Aliteraturereview,ArtificialIntelligence293(2021) 103448

    work page 2021

  3. [3]

    S. Ren, K. He, R. Girshick, J. Sun, Faster r-cnn: Towards real-time object detection with region proposal networks, Advances in neural information processing systems 28 (2015)

    2015

  4. [4]

    Bochkovskiy, C.-Y

    A. Bochkovskiy, C.-Y. Wang, H.-Y. M. Liao, Yolov4: Optimal speed and accuracy of object detection, arXiv preprint arXiv:2004.10934 (2020)

    arXiv 2004

  5. [5]

    Z.Ge,S.Liu,F.Wang,Z.Li,J.Sun,Yolox:Exceedingyoloseriesin 2021, arXiv preprint arXiv:2107.08430 (2021)

    arXiv 2021

  6. [6]

    Bewley, Z

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

    work page 2016

  7. [7]

    3645–3649

    N.Wojke,A.Bewley,D.Paulus, Simpleonlineandrealtimetracking with a deep association metric, in: 2017 IEEE international confer- ence on image processing (ICIP), IEEE, 2017, pp. 3645–3649

    work page 2017

  8. [8]

    Zhang, P

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

    work page 2022

Show all 27 references
  1. [9]

    Zhang, C

    Y. Zhang, C. Wang, X. Wang, W. Zeng, W. Liu, Fairmot: On the fairness of detection and re-identification in multiple object tracking, International journal of computer vision 129 (2021) 3069–3087

    2021

  2. [10]

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

    2022 arXiv

  3. [11]

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

    2023

  4. [12]

    Sundararaman, C

    R. Sundararaman, C. De Almeida Braga, E. Marchand, J. Pettre, Tracking pedestrian heads in dense crowd, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 3865–3875

    2021

  5. [13]

    Pedersen, J

    M. Pedersen, J. B. Haurum, S. H. Bengtson, T. B. Moeslund, 3d-zef: A 3d zebrafish tracking benchmark dataset, in: Proceedings of the IEEE/CVFConferenceonComputerVisionandPatternRecognition, 2020, pp. 2426–2436

    2020

  6. [14]

    Redmon, S

    J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once:Unified,real-timeobjectdetection,in:ProceedingsoftheIEEE conference on computer vision and pattern recognition, 2016, pp. 779–788

    2016

  7. [15]

    K. Duan, S. Bai, L. Xie, H. Qi, Q. Huang, Q. Tian, Centernet: Key- point triplets for object detection, in: Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6569–6578

    2019

  8. [16]

    Idrees, K

    H. Idrees, K. Soomro, M. Shah, Detecting humans in dense crowds using locally-consistent scale prior and global occlusion reason- ing, IEEE transactions on pattern analysis and machine intelligence 37 (10) (2015) 1986–1998

    2015

  9. [17]

    S. Shao, Z. Zhao, B. Li, T. Xiao, G. Yu, X. Zhang, J. Sun, Crowd- human: A benchmark for detecting human in a crowd, arXiv preprint arXiv:1805.00123 (2018). First Author et al.:Preprint submitted to Elsevier Page 12 of 13 Short Title of the Article

    2018 arXiv

  10. [18]

    R. E. Kalman, A new approach to linear filtering and prediction problems, Transactions of the ASME–Journal of Basic Engineering 82 (Series D) (1960) 35–45

    1960

  11. [19]

    S. Wang, H. Sheng, D. Yang, Y. Zhang, Y. Wu, S. Wang, Extend- able multiple nodes recurrent tracking framework with rtu++, IEEE Transactions on Image Processing 31 (2022) 5257–5271

    2022

  12. [20]

    Y.Zhang,H.Sheng,Y.Wu,S.Wang,W.Lyu,W.Ke,Z.Xiong,Long- term tracking with deep tracklet association, IEEE Transactions on Image Processing 29 (2020) 6694–6706

    2020

  13. [21]

    2528–2533

    D.Peng,Z.Sun,Z.Chen,Z.Cai,L.Xie,L.Jin,Detectingheadsusing featurerefinenetandcascadedmulti-scalearchitecture,in:201824th InternationalConferenceonPatternRecognition(ICPR),IEEE,2018, pp. 2528–2533

    2018

  14. [22]

    Milan, L

    A. Milan, L. Leal-Taixé, I. Reid, S. Roth, K. Schindler, Mot16: A benchmark for multi-object tracking, arXiv preprint arXiv:1603.00831 (2016)

    2016 arXiv

  15. [23]

    Z.Zheng,P.Wang,D.Ren,W.Liu,R.Ye,Q.Hu,W.Zuo,Enhancing geometricfactorsinmodellearningandinferenceforobjectdetection and instance segmentation, IEEE Transactions on cybernetics 52 (8) (2021) 8574–8586

    2021

  16. [24]

    Dendorfer, H

    P. Dendorfer, H. Rezatofighi, A. Milan, J. Shi, D. Cremers, I. Reid, S. Roth, K. Schindler, L. Leal-Taixé, Mot20: A benchmark for multi object tracking in crowded scenes, arXiv preprint arXiv:2003.09003 (2020)

    2003 arXiv

  17. [25]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al., Pytorch: An imperative style, high-performance deep learning library, Advances in neural information processing systems 32 (2019)

    2019

  18. [26]

    Bernardin, R

    K. Bernardin, R. Stiefelhagen, Evaluating multiple object tracking performance: the clear mot metrics, EURASIP Journal on Image and Video Processing 2008 (2008) 1–10

    2008

  19. [27]

    M. Yang, G. Han, B. Yan, W. Zhang, J. Qi, H. Lu, D. Wang, Hybrid-sort: Weak cues matter for online multi-object tracking, in: ProceedingsoftheAAAIconferenceonartificialintelligence,Vol.38, 2024, pp. 6504–6512. First Author et al.:Preprint submitted to Elsevier Page 13 of 13

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.