Pith. sign in

REVIEW 4 major objections 7 minor 29 references

A Framework for Multi-View Multiple Object Tracking using Single-View Multi-Object Trackers on Fish Data

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A stereo pair of off-the-shelf single-view trackers can yield 3D fish trajectories for ecological studies.

desk verdict A workmanlike stereo fish-tracking pipeline whose central accuracy claim is never tested; the real contribution is the dataset and honest evaluation, not the multi-view enhancement. read the letter →

arxiv 2505.17201 v1 pith:5X453RAN submitted 2025-05-22 cs.CV

classification cs.CV
keywords multi-objecttrackingfishmulti-viewfusionstereomatchingYOLOv8ByteTrack3Dtrajectoryreconstructionunderwatervideoanalysis
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 argues that a multi-view multiple-object-tracking framework can be built by fusing the outputs of ordinary single-view trackers, and that this fusion improves fish tracking in underwater footage while adding a third dimension. The authors run YOLOv8 with ByteTrack on each of two calibrated stereo videos, re-identify fish that the tracker splits into multiple IDs, match the fish across views using epipolar geometry, and triangulate 3D coordinates per fish per frame. The payoff, if the claim holds, is that ecologists can obtain depth-resolved trajectories, speeds, and grouping patterns from stereo cameras using off-the-shelf trackers, without training a dedicated multi-view model. The paper's own results show the pipeline works well on long videos (95–100% stereo-match accuracy) and poorly on short ones (22–36%), where matches had to be corrected by hand.

What carries the argument

The load-bearing mechanism is the post-track fusion pipeline that converts two independent 2D tracking outputs into a 3D trajectory set. Each view is processed by YOLOv8 detection with ByteTrack association; a hand-tuned post-track re-identification pass merges fragmented identities using a 50-pixel proximity threshold, a 100-frame disappearance window, and a 10-frame co-occurrence rule; then stereo correspondence is established through the fundamental matrix and epipolar-line distance thresholds, with the most common per-fish match selected across frames; finally, the calibrated camera parameters feed a triangulation step that produces x, y, z coordinates. The epipolar constraint—that a point in one camera lies on a specific line in the other—is the geometric carrier of the cross-view identity matching.

What would settle it

Run the pipeline on the short stereo videos without manual re-matching and compare the automatically matched fish-ID pairs to a hand-labelled ground-truth correspondence; if the automatic match rate stays at the reported 22–36% rather than near 100%, the framework's central identity-preservation premise fails for short recordings. A complementary check is to compare HOTA/MOTA on the fused multi-view output against the single-view output on the same videos: the claimed enhanced tracking accuracy would be falsified if the fused trajectories do not score higher on association metrics.

Watch

Extended reading notes

Core claim

The central claim is that single-view multi-object trackers can be dual-leveraged—run independently on each stereo view and then fused through geometry—to produce a new 3D output and enhanced tracking accuracy for fish in underwater videos. The framework detects fish with a relative accuracy of about 47% within the ground-truth margin, tracks a subset of fish for full video lengths, and uses the fundamental matrix and epipolar lines to find corresponding fish IDs between views; the most frequent match per fish across frames is chosen, and camera calibration parameters are used to triangulate x, y, z coordinates. In the paper's account, the 3D output enables more realistic speed, depth, and grouping analyses than 2D single-view tracks, and for a select set of videos the multi-view framework overcomes the challenges of small, visually similar fish with rapidly changing appearances. The authors conclude that this adapts off-the-shelf single-view trackers to ecological studies, a gap they identify in prior work.

Load-bearing premise

The whole framework stands on the assumption that the stereo-matching step pairs the same physical fish in the two camera views, so that every triangulated 3D point is the same real fish.

Editorial extensions

If this is right

  • Ecologists can generate 3D fish trajectories (x, y, z per frame) from calibrated stereo cameras using off-the-shelf single-view trackers and the described fusion steps.
  • The framework produces an enriched behavioral dataset: the paper reports speed over time, acceleration, path length, spatial distribution, density maps, temporal patterns, and depth over time from the triangulated coordinates.
  • On long stereo videos (e.g., 3,117-frame recordings 129 and 406) the stereo-matching step reached 95–100% accuracy, indicating the approach is usable without manual correction on extended footage.
  • Even with a detection recall near 48%, the post-track re-identification step reduced the number of distinct fish IDs by 43.3% on average, consolidating fragmented tracks before fusion.

Reading between the lines

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

  • A direct consequence the authors do not spell out: if off-the-shelf single-view trackers plus epipolar fusion suffice for 3D fish tracking, then the expensive joint multi-view training used in models like DETR3D is unnecessary for this ecological scenario; the bottleneck becomes the per-view detector's recall, not the fusion strategy.
  • The stereo-matching failure on short videos suggests the frequency-based 'most common match' selection needs temporal accumulation: instead of picking the modal match per fish from a 258-frame video, one could accumulate correspondences across the whole recording, or enforce one-to-one matching globally, which might recover the short-video cases that currently require manual re-matching.
  • The claimed 'enhanced tracking accuracy' is asserted but never measured against the single-view baseline with the same MOT metrics; a reader could test it by computing HOTA/MOTA on the fused tracks versus the per-view tracks, which would isolate whether the 3D step improves association or merely adds coordinates.
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 / 7 minor

Summary. The paper proposes a multi-view multiple-object tracking framework for underwater fish videos. The pipeline runs YOLOv8 with ByteTrack independently on each stereo view, applies a post-hoc ID reconciliation heuristic, matches fish identities across views via epipolar geometry, triangulates matched bounding-box centers to 3D coordinates, and produces 3D trajectory visualizations. The authors report single-view MOT metrics, detection-margin statistics, ID-count changes after re-identification, stereo-matching accuracy, and example 3D coordinates. The central claim is that this framework enhances fish tracking accuracy for ecological studies by leveraging stereo inputs and generating a new 3D output.

Significance. If the enhancement claim were supported, the paper would offer ecologists a practical recipe for obtaining 3D fish trajectories from calibrated stereo cameras using off-the-shelf single-view trackers. The authors document a complete pipeline, make code available, and test on a real underwater fish dataset, which are useful engineering contributions. However, the load-bearing assertion that multi-view processing improves tracking accuracy is never tested: the pipeline adds no cross-view feedback to the single-view tracks, and no multi-view versus single-view comparison is reported. The reported metrics are also weak (recall 0.48, HOTA 0.100, DetA 0.023), and the stereo-correspondence step succeeds on only 22-36% of frames for short videos, requiring manual re-matching. As it stands, the contribution reduces to appending triangulated 3D coordinates to independently computed 2D tracks, so the significance of the claimed enhancement is not established.

major comments (4)
  1. [Section 6, Section 3.1.4] The central claim that the multi-view framework enhances tracking accuracy is never tested. In the pipeline of Section 3.1.4, YOLOv8+ByteTrack is run independently on each view (step 1), and the stereo matching (step 4) and triangulation (step 6) operate only on the resulting track files; no cross-view information is fed back to correct detections, associations, or missed tracks. Consequently, the 2D tracking metrics (HOTA, MOTA, IDF1) are identical with and without the multi-view stages. The paper reports no multi-view versus single-view comparison and no metric that measures the claimed enhancement. The statement in Section 6 that the framework 'thereby enhancing fish tracking accuracy' is therefore unsupported by the experimental design.
  2. [Section 4.5, Table 6] The stereo-correspondence step, which is essential for the 3D output, is correct in only 22% and 36% of frames for the short videos, and the text states that those matches 'had to be re-matched manually' to be usable. Thus, for the short-video subset the 3D coordinates are not produced by the proposed framework but by manual annotation. The paper does not report how many of the videos and frames used in the 3D evaluation were manually re-matched, so the claimed automatic generation of 3D trajectories is established only for the two long videos.
  3. [Section 3.2] The evaluation relies on an ID mapping function asserted to have '100% accuracy' in mapping tracking IDs to ground-truth IDs, but the paper gives no description of how this mapping is obtained or validated. If this mapping is imperfect, every association-sensitive metric (HOTA, AssA, IDF1, MOTA) and the re-identification evaluation in Section 4.3 are affected. This assertion is load-bearing and must be substantiated.
  4. [Section 4.3, Table 4] The post-track re-identification heuristic (Section 3.1.4 step 2) is not validated. Table 4 shows that after re-identification the number of IDs drops below the ground-truth number for several videos (e.g., video 13_1: GT 13, before re-ID 17, after re-ID 6; video 8_1: GT 9, after re-ID 6), indicating that distinct fish are being merged. The thresholds (50 pixels, 100 frames, 10 frames, 30 frames) are introduced without ablation or sensitivity analysis, so the claimed improvement from re-identification is not established.
minor comments (7)
  1. [Abstract] The phrase 'novel 3D output' overstates standard triangulation from calibrated stereo cameras; suggest rephrasing to '3D output' or similar.
  2. [Section 3.1.4, step 4] The text alternates between 'keypoints' (e.g., corners) and fish bounding-box centers; clarify what features are actually matched and whether fish centers are used or arbitrary corner features.
  3. [Section 4.1, Table 1] The 'Average*' row is not defined (which videos are included, how many), and no standard deviations are reported; add the full protocol and variance.
  4. [Section 4.2, Table 3] The margin criterion (half the average bounding box size and twice the average size) needs a precise definition of which average is used and how the two tiers are combined; currently the reported percentages are difficult to interpret.
  5. [Section 4.3 and Section 3.1.4 step 2] The post-hoc heuristic is called 're-identification,' which may be confused with the learned re-ID branch of MOT models; rename it (e.g., 'ID reconciliation') to avoid ambiguity.
  6. [Section 3.1.3] The model trained by Dr. Nejadasl is described only as having better performance; provide training data, hyperparameters, and validation metrics so the experiments are reproducible.
  7. [Appendix C] Several figure captions in Appendix C appear mis-numbered (multiple figures labeled 'Figure 12'), which should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the multi-view pipeline uses external single-view trackers and standard stereo geometry; the unsupported accuracy-enhancement claim is a validation gap, not a circular reduction.

full rationale

No significant circularity. The derivation chain is: YOLOv8+ByteTrack (external, cited to [15,26]) produces per-view tracks; epipolar matching and triangulation use standard multi-view geometry [6,10,11,18,23]; and the 3D coordinates are outputs of those operations, not inputs to the claimed conclusion. The post-track re-ID thresholds (50 pixels, 100 frames, 10 frames, 30 frames, Section 3.1.4(2)) are hand-chosen heuristics evaluated against external ground-truth ID mappings, so they are not a fitted quantity disguised as a prediction. The self-citations [2,8,12] appear only as related-work context and are not load-bearing. Section 4.5 admits stereo matching failed on short videos and required manual re-matching, and Section 6 claims 'enhancing fish tracking accuracy' without any multi-view versus single-view MOT comparison; these are correctness and validation limitations, not circular reductions. Similarly, calling the standard triangulation output 'novel' is an overclaim, but no quoted equation or construction makes the conclusion equivalent to its inputs.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-tuned pipeline thresholds, unverified assumptions about ground truth and calibration quality, an undisclosed external detection model, and an asserted perfect ID mapping. No new physical entities are introduced. The free parameters directly influence the reported accuracy numbers and were not ablated.

free parameters (5)
  • reid_distance_threshold = 50 pixels
    Post-track re-ID merges a new fish into an old ID if within 50 px (Eq. 9); threshold chosen without ablation, Section 3.1.4(2).
  • reid_frame_window = 100 frames
    A 'new' fish is considered a reappearance if it appears within a 100-frame window of the 'old' fish's disappearance, Section 3.1.4(2).
  • reid_overlap_frames = 10 frames
    New ID must overlap the old ID for 10 or fewer frames, and only at the end, to qualify as the same fish, Section 3.1.4(2).
  • false_positive_min_frames = 30 frames
    Entities appearing in fewer than 30 frames are removed as false positives, Section 3.1.4(2).
  • detection_margin = 0.5 average GT box size (and 2x for near miss)
    The 'relative accuracy of 47%' is defined as detection centers within half the average ground-truth bounding box size; this margin choice affects all headline accuracy numbers, Section 3.2.
assumptions (4)
  • standard math Epipolar geometry, fundamental matrix estimation, and triangulation from calibrated stereo pairs are valid.
    Sections 3.1.4(4)-(6) apply standard two-view geometry [6,10,11,18,23].
  • domain assumption The ground-truth MOT files and stereo calibration .mat parameters are correct.
    Section 3.1.1 uses GT track files labeled by students/supervisors and camera intrinsics/extrinsics as ground truth without error analysis.
  • ad hoc to paper The pre-trained fish detection model supplied by Dr. Nejadasl, replacing locally trained models, is trained on comparable data and is representative.
    Section 3.1.3 states a model trained by a co-author was chosen after local training performed poorly; no training specification or data split is given.
  • ad hoc to paper The ID mapping function used for evaluation achieves 100% accuracy in mapping tracking IDs to ground truth IDs.
    Section 3.2 asserts this mapping without describing the algorithm or validating it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Framework for Multi-View Multiple Object Tracking using Single-View Multi-Object Trackers on Fish Data." pith.science (2026). https://pith.science/paper/5X453RAN

@misc{pith2026250517201,
  author       = {Pith},
  title        = {Pith review of: A Framework for Multi-View Multiple Object Tracking using Single-View Multi-Object Trackers on Fish Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5X453RAN}},
  note         = {Machine review of arXiv:2505.17201}
}
read the original abstract

Multi-object tracking (MOT) in computer vision has made significant advancements, yet tracking small fish in underwater environments presents unique challenges due to complex 3D motions and data noise. Traditional single-view MOT models often fall short in these settings. This thesis addresses these challenges by adapting state-of-the-art single-view MOT models, FairMOT and YOLOv8, for underwater fish detecting and tracking in ecological studies. The core contribution of this research is the development of a multi-view framework that utilizes stereo video inputs to enhance tracking accuracy and fish behavior pattern recognition. By integrating and evaluating these models on underwater fish video datasets, the study aims to demonstrate significant improvements in precision and reliability compared to single-view approaches. The proposed framework detects fish entities with a relative accuracy of 47% and employs stereo-matching techniques to produce a novel 3D output, providing a more comprehensive understanding of fish movements and interactions

Figures

Figures reproduced from arXiv: 2505.17201 by the authors.

Figure 1
Figure 1. Model Output VS Ground Truth Identification for a [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 5
Figure 5. Tracking Before and After Re-Identification Video [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 2
Figure 2. Fish margins of fish entity 7 in Video 23_1: The left [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figures from the paper (16 more)
Figure 3
Figure 3. Figure 3: Missed Identification of fish entity 4 in Video 23_1: [PITH_FULL_IMAGE:figures/full_fig_p008_3.png]
Figure 4
Figure 4. Figure 4: False Positive Identification of fish entity in Video [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: 3D Coordinates Visualization of Video 129 For [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Stereo Matching DataFrame Video 33 8 Appendix B: Mot Data Cleaning Example [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Mot Data Cleaning Before and After 9 Appendix C: Visualization Output 9.1 Fish 0 Trajectory In [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Fish 0 Trajectory Video 129 9.2 Fish Speed In [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Fish 0 Speed Video 129 9.3 Fish Acceleration In [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Fish 0 Acceleration Video 129 9.4 Fish Path Lengths In [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Fish Path lengths Video 129 9.5 Fish Spatial Distribution In [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Fish Spatial Distributions Video 129 9.6 Fish Density Map In [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: Fish Density Map Video 129 9.7 Fish Temporal Patterns In [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]
Figure 15
Figure 15. Figure 15: Average Fish Temporal Patterns Video 129 [PITH_FULL_IMAGE:figures/full_fig_p014_15.png]
Figure 16
Figure 16. Figure 16: Fish Depths Video 129 10 Appendix D [PITH_FULL_IMAGE:figures/full_fig_p015_16.png]
Figure 17
Figure 17. Figure 17: 3D Coordinate Construction Using Epipolar Geometry with the gray region being the epipolar plane, the orange line [PITH_FULL_IMAGE:figures/full_fig_p015_17.png]
Figure 18
Figure 18. Figure 18: Training Pipeline Visualization 11.2 Fish Tracking Pipeline [PITH_FULL_IMAGE:figures/full_fig_p016_18.png]
Figure 19
Figure 19. Figure 19: Training Pipeline Visualization 12 Appendix F: Full Evaluation Metrics Table [PITH_FULL_IMAGE:figures/full_fig_p016_19.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 19 canonical work pages

  1. [1]

    Derya Akkaynak and Tali Treibitz. 2019. Sea-thru: A method for removing water from underwater images. (2019), 1682–1691

  2. [2]

    Karim Anwar and Seyed Sahand Mohammadi Ziabari. 2025. Attention to the Branches: A Comparative Analysis of FairMOT with Transformers on Fish Dataset. InMulti-disciplinary Trends in Artificial Intelligence, Chattrakul Sombat- theera, Paul Weng, and Jun Pang (Eds.). Springer Nature Singapore, Singapore, 64–76. https://doi.org/10.1007/978-981-96-0692-4_6

  3. [3]

    Keni Bernardin and Rainer Stiefelhagen. 2008. Evaluating multiple object track- ing performance: the clear mot metrics.EURASIP Journal on Image and Video Processing2008 (2008), 1–10

  4. [4]

    Akash D. 2023. yolov8_in_depth. https://github.com/akashAD98/yolov8_in_ depth Accessed: 2024-06-10

  5. [5]

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

  6. [6]

    Victor H Diaz-Ramirez, Martin Gonzalez-Ruiz, Vitaly Kober, and Rigoberto Juarez- Salazar. 2022. Stereo image matching using adaptive morphological correlation. Sensors22, 23 (2022), 9050

  7. [7]

    Meng Joo Er, Jie Chen, Yani Zhang, and Wenxiao Gao. 2023. Research Challenges, Recent Advances, and Popular Datasets in Deep Learning-Based Underwater Marine Object Detection: A Review.Sensors23, 4 (Feb. 2023), 1990. https: //doi.org/10.3390/s23041990

  8. [8]

    Faber, Seyed Sahand Mohammadi Ziabari, and Fatemeh Karimi Nejadasl

    Niels G. Faber, Seyed Sahand Mohammadi Ziabari, and Fatemeh Karimi Nejadasl

Show all 29 references
  1. [9]

    Rui Gao and Lu Wang. 2023. MeMOTR: Long-Term Memory-Augmented Trans- former for Multi-Object Tracking.arXiv(2023). arXiv:2307.15700 [cs.CV]

  2. [10]

    2003.Multiple view geometry in computer vision

    Richard Hartley and Andrew Zisserman. 2003.Multiple view geometry in computer vision. Cambridge university press

  3. [11]

    Heiko Hirschmuller. 2007. Stereo processing by semiglobal matching and mutual information.IEEE Transactions on pattern analysis and machine intelligence30, 2 (2007), 328–341

  4. [12]

    Katona, S.S

    Z. Katona, S.S. Mohammadi Ziabari, and F. Karimi Nejadasl. 2025. MARINE: A Computer Vision Model for Detecting Rare Predator-Prey Interactions in Animal Videos. InBig Data and Artificial Intelligence. BDA 2024 (Lecture Notes in Computer Science, Vol. 15526), A. Dasgupta, R.U. ...

  5. [13]

    Haitong Lou, Xuehu Duan, Junmei Guo, Haiying Liu, Jason Gu, Lingyun Bi, and Haonan Chen. 2023. DC-YOLOv8: small-size object detection algorithm based on camera sensor.Electronics12, 10 (2023), 2323

  6. [14]

    = 2020, month = oct, pages = 548–578

    Jonathon Luiten, Aljoša Ošep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixé, and Bastian Leibe. = 2020, month = oct, pages = 548–578. HOTA: A Higher Order Metric for Evaluating Multi-object Tracking.International Journal of Computer Vision129, 2 (= 2020, mon...

  7. [15]

    Muhammad Rizwan Munawar, Glenn Jocher, and Ayush Chaurasia. 2023. Home — docs.ultralytics.com. https://docs.ultralytics.com/. [Accessed 24-05-2024]

  8. [16]

    David MW Powers. 2020. Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation.arXiv preprint arXiv:2010.16061 (2020)

  9. [17]

    Zou, Rita Cucchiara, and Carlo Tomasi

    Ergys Ristani, Francesco Solera, Roger S. Zou, Rita Cucchiara, and Carlo Tomasi

  10. [18]

    Olgierd Stankiewicz, Gauthier Lafruit, and Marek Domański. 2018. Multiview video: Acquisition, processing, compression, and virtual view rendering. (2018), 3–74

  11. [19]

    Ultralytics. 2023. Are class and box losses calculated the same in YoloV8 and YoloV5? https://github.com/ultralytics/ultralytics/issues/2789 Accessed: 2024- 06-10

  12. [20]

    Ultralytics. 2023. yolo8 loss function for classification task. https://github.com/ ultralytics/ultralytics/issues/4684 Accessed: 2024-06-10

  13. [21]

    Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. 2022. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. InConference on Robot Learning. PMLR, 180–191

  14. [22]

    Xin Xiao and Xinlong Feng. 2023. Multi-object pedestrian tracking using im- proved YOLOv8 and OC-SORT.Sensors23, 20 (2023), 8439

  15. [23]

    Atsushi Yamashita, Toru Kaneko, Shinya Matsushita, Kenjiro T Miura, and Sue- kichi Isogai. 2003. Camera Calibration and 3-D Measurement with an Active Stereo Vision System for Handling Moving Objects.Journal of Robotics and Mechatronics15, 3 (2003), 304–313

  16. [24]

    En Yu et al. 2023. MOTRv3: Release-Fetch Supervision for End-to-End Multi- Object Tracking.arXiv preprint arXiv:2305.14298(2023)

  17. [25]

    Fangao Zeng et al. 2022. Motr: End-to-end multiple-object tracking with trans- former. InEuropean Conference on Computer Vision. Springer Nature Switzerland

  18. [26]

    Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. 2022. ByteTrack: Multi-Object Tracking by Associating Every Detection Box. arXiv:2110.06864 [cs.CV]

  19. [27]

    Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. 2021. FairMOT: On the Fairness of Detection and Re-identification in Multiple Object Tracking.International Journal of Computer Vision129, 11 (Sept. 2021), 3069–3087. https://doi.org/10.1007/s11263-021-01513-4...

  20. [2016]

    arXiv:1609.01775 [cs.CV]

    Performance Measures and a Data Set for Multi-Target, Multi-Camera Tracking. arXiv:1609.01775 [cs.CV]

  21. [2024]

    Leveraging Foundation Models via Knowledge Distillation in Multi-Object Tracking: Distilling DINOv2 Features to FairMOT. arXiv:2407.18288 [cs.CV] https://arxiv.org/abs/2407.18288 A Framework for Multi-View Multiple Object Tracking using Single-View Multi-Object Trackers on Fis...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.