Pith. sign in

REVIEW 4 major objections 5 minor 41 references

A single fixed configuration of Grounding DINO and SAM 2, with three heuristic adjustments, outperforms trained and zero-shot multi-animal trackers on four wildlife benchmarks without retraining or per-dataset tuning.

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

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

T0 review · deepseek-v4-flash

2026-08-04 00:06 UTC pith:DW4XXUCT

load-bearing objection A competent zero-shot multi-animal tracking adaptation of SAM2MOT with three well-validated heuristics; the main gaps are a missing SAM3 section promised in the abstract, unreported GMOT-40 subset IDs, and unexamined sensitivity to hand-set thresholds that undercut the 'no hyperparameter tuning' claim. the 4 major comments →

arxiv 2511.02591 v2 pith:DW4XXUCT submitted 2025-11-04 cs.CV

Zero-Shot Multi-Animal Tracking in the Wild

Jan Frederik Meier , Timo L\"uddecke This is my paper
classification cs.CV
keywords zero-shot trackingmulti-animal trackingSAM 2Grounding DINOadaptive thresholdingvideo object segmentationwildlife monitoringmulti-object 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.

This paper claims that a zero-shot multi-animal tracker built from a Grounding DINO open-vocabulary detector and the SAM 2 video segmenter, modified with three heuristics, can outperform both trained and existing zero-shot trackers on four diverse animal datasets without any retraining or hyperparameter adaptation between datasets. The three changes are per-sequence adaptive detection thresholds, mask-based track initialization, and density-aware re-prompting. If true, it means a single out-of-the-box configuration can serve as a strong baseline for ecology and conservation monitoring across species and habitats, removing the usual per-dataset engineering burden.

Core claim

The authors show that the main barrier between SAM2MOT and good animal tracking is not the segmentation backbone but two human-centric assumptions: fixed detection thresholds tuned per sequence, and bounding-box overlap tests for track initialization that fail when multiple animals share a box. Replacing the fixed threshold with an Otsu/K-means adaptive threshold computed per sequence, deciding new tracks from SAM 2's masks via normalized mask intersection, and restricting re-prompting to unambiguous best-vs-second-best IoU gaps yields consistent HOTA and AssA gains on ChimpAct, Bird Flock Tracking, AnimalTrack, and GMOT-40-Animal. With one set of hyperparameters, the method beats trained mo

What carries the argument

The load-bearing components are three heuristics sitting on top of Grounding DINO (open-vocabulary detection) and SAM 2 (promptable video segmentation). Adaptive detection thresholds cluster detection scores per sequence with K-Means (approximately Otsu's method) and add a fixed offset, replacing the per-sequence tuned threshold of SAM2MOT. Mask-based track initialization prompts SAM 2 with each detection box and compares the resulting mask to active track masks using normalized mask intersection, so new tracks are spawned only when the mask genuinely differs. Density-aware reconstruction limits re-prompting to detections whose IoU gap between best and second-best track is above a threshold,

Load-bearing premise

The whole zero-shot claim rests on detection scores in every new sequence forming two well-separated clusters — true positives and false positives — so the automatic Otsu/K-Means cut plus a fixed offset gives a reliable threshold; when that bimodality fails, the threshold is miscalibrated and tracks appear or disappear spuriously.

What would settle it

Run the tracker on a clip with no animals but similar background (e.g., an empty enclosure or empty field). All detector outputs are false positives, so the K-Means split will still partition the score distribution into two clusters and promote the higher-scoring cluster to true positives, producing phantom tracks with nonzero HOTA and identity switches. A method that genuinely zero-shots should output zero tracks on such input.

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

If this is right

  • A fixed zero-shot configuration can replace per-dataset fine-tuning for common multi-animal tracking benchmarks, cutting annotation and training cost.
  • The same pipeline extends to non-animal MOT datasets (DanceTrack, SportsMOT) without changing hyperparameters, suggesting broader applicability.
  • Per-sequence adaptive thresholds recover most of the gap to oracle-tuned fixed thresholds, since score distributions drift between sequences within one dataset.
  • The method's gains are largest in association accuracy (AssA), implying the mask-based components improve identity maintenance rather than just detection.

Where Pith is reading between the lines

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

  • The same recipe likely transfers to any category list a user types into the detector, so the practical scope is any open-set tracking task, not just animals; the paper's own DanceTrack/SportsMOT numbers support this reading.
  • Because runtime and memory scale at least linearly with track count, the approach will hit a wall in dense swarms (hundreds of individuals); addressing scalability would widen the wildlife use cases more than further accuracy gains.
  • A testable extension is to replace the static offset δ with a per-sequence calibration based on scene complexity or prompt specificity, which could reduce false-positive tracks in empty or clutter-heavy scenes.
  • The K-Means/Otsu threshold assumes exact bimodality; on scenes where all detections are false positives it will still split the distribution and manufacture two clusters, so a sanity check for empty scenes would harden the zero-shot claim.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a zero-shot multi-animal tracking pipeline that combines Grounding DINO with SAM 2, building on SAM2MOT and introducing three heuristic modifications: adaptive detection thresholds (K-means/Otsu plus a fixed offset), mask-based track initialization, and density-aware reconstruction, together with mask NMS. The method uses the same checkpoints and hyperparameters across all datasets, with only the text prompt changed per dataset. Evaluation on ChimpAct, Bird Flock Tracking, AnimalTrack, and a self-defined GMOT-40-Animal subset shows consistent HOTA/AssA improvements over zero-shot baselines using the same detector, and the ablation indicates each component contributes. The authors claim state-of-the-art zero-shot multi-animal tracking without retraining or per-dataset hyperparameter adaptation.

Significance. If the result holds, the paper provides a strong evidence point that foundation-model-based, zero-shot tracking can outperform specialized trained trackers on animal benchmarks, which is practically relevant for ecology and wildlife monitoring. The main strengths are the consistent gains across four datasets, the controlled comparison against same-detector zero-shot baselines, and a component-wise ablation. The code is released, which supports reproducibility. However, the method relies on several hand-set heuristics and an adaptive threshold whose underlying bimodality assumption is explicitly conceded to be imperfect; the current evidence does not yet establish that the configuration is robust enough to support the 'no hyperparameter adaptation' claim beyond the specific benchmarks tested.

major comments (4)
  1. [Section 2 / Appendix A.3, Eq. (2)] The adaptive detection threshold is a load-bearing component: it contributes the largest single gain in the ChimpAct ablation (Table 2, row 2: HOTA 54.6→56.7). The method assumes a bimodal score distribution, and the paper itself states this 'does not always hold in practice.' Yet no per-sequence validation is provided to show that the Otsu/K-means split plus fixed offset δ=0.1 tracks the oracle threshold, and there is no sensitivity analysis for δ. Table 4 compares against three fixed global thresholds, but not against per-sequence oracle thresholds or non-bimodal sequences. Please add a sensitivity analysis for δ and a per-sequence comparison (e.g., adaptive threshold vs. oracle threshold, or HOTA under varied δ) to support the robustness claim.
  2. [Appendix E / Table 8] Twelve hyperparameters (δ, τ_mask, τ_IoU, occlusion thresholds, N_lost, N_frames, τ_mIoU, τ_Δscore, τ_Δstd, τ_nms) are fixed across datasets, but the paper does not report how these values were chosen or how sensitive the results are to them. If they were selected using the validation splits, the zero-shot claim is weakened: the configuration may be overfit to ChimpAct/BFT validation even if not tuned per test set. At minimum, please report a sensitivity study for the most influential parameters or describe a principled protocol for setting them without validation data.
  3. [Table 2] The ablation shows that Mask NMS reduces BFT HOTA (73.2→72.5) and AssA (76.5→74.6) on the validation split, yet the final model includes Mask NMS for all datasets. If a fixed configuration is a selling point, including a component that consistently hurts one benchmark needs justification. Either show that Mask NMS is necessary for other datasets or discuss why it is retained despite the negative BFT effect; otherwise the design choice appears inconsistent with the ablation evidence.
  4. [Section 3 / Appendix B] GMOT-40-Animal is a self-defined subset of GMOT-40 (16 sequences) without a stated selection criterion beyond containing animals. Since the state-of-the-art claim on this benchmark is based on this nonstandard subset, please specify the selection protocol and report per-sequence results, or compare against published GMOT-40 results in a way that allows the community to assess whether the subset is representative. Without this, the generalization claim on this dataset is difficult to verify.
minor comments (5)
  1. [Abstract / Full text] The abstract states that the recent SAM3 model is evaluated and found to have practical limitations, but no SAM3 results or details appear anywhere in the body or appendices. Either add the SAM3 evaluation or remove this claim from the abstract.
  2. [Naming conventions] The paper uses inconsistent names: 'ChimpAct' vs 'Chimp-Act', 'Grounding Dino' vs 'Grounding DINO', 'DanceTrack' vs 'Danccetrack' in references. Please standardize.
  3. [Equation (1)] The K-means objective in Eq. (1) is written with sets S1,S2 but does not formally define the assignment variable or the constraint that the clusters partition the scores. Consider adding a short description of the standard K-means formulation for clarity.
  4. [Table 7 / Table 8] The implementation details say 'the same model checkpoints and hyperparameters are used across all datasets,' but Table 7 shows dataset-specific text prompts. Clarify that text prompts are treated as input specifications, not hyperparameters, to avoid confusion.
  5. [Table 1] For the 'Trained on dataset' models, several HOTA/AssA entries are '-'. Please state whether these are unavailable, not reported, or not applicable, and consider including the DetA values for the zero-shot comparisons if available.

Circularity Check

0 steps flagged

No circular derivation; the paper is an external benchmark evaluation with heuristic components.

full rationale

The paper's central claim is an empirical result: a fixed configuration of Grounding DINO + SAM 2 plus three heuristic modifications is evaluated on public test sets (ChimpAct, BFT, AnimalTrack, GMOT-40-Animal) via TrackEval. These numbers are measured against external annotations, not derived from the method's own definitions or fitted parameters. The adaptive threshold (Eq. 2) is computed unsupervised from detection-score clusters and is explicitly acknowledged to assume bimodality ('this approach assumes a bimodal score distribution, an assumption that does not always hold in practice'); this is an assumption/limitation affecting robustness, not a circular reduction. The fixed offset delta=0.1 is a hyperparameter constant across datasets, not a prediction constructed from the metrics it is used to explain. The only self-citation (PriMAT, ref [33]) appears in related-work context and is not load-bearing; the method builds on the external SAM2MOT [15] baseline. No equation is equivalent by construction to the reported HOTA/AssA improvements, and no fitted value is renamed as a prediction. Thus no significant circularity.

Axiom & Free-Parameter Ledger

12 free parameters · 6 axioms · 0 invented entities

The central claim rests on the reliability of Grounding DINO detections, SAM2 zero-shot propagation, the bimodality assumption for adaptive thresholds, and on the evaluation setup (public benchmarks, self-defined GMOT subset). All hyperparameters in Appendix E are hand-set design choices rather than learned or derived quantities.

free parameters (12)
  • Static detection threshold offset δ = 0.1
    Added to the Otsu/K-means weighted centroid threshold to bias toward precision; hand-set and kept constant across all experiments (Eq. 2, Appendix E Table 8).
  • Mask initialization threshold τ_mask = 0.4
    A new track is initialized only if the normalized mask intersection with all existing tracks is below this value (Eq. 3, Table 8).
  • Re-prompt IoU margin τ_IoU = 0.3
    Minimum difference between best and second-best detection-track IoU for re-prompting (Eq. 5, Table 8).
  • τ_reliable occlusion score = 8
    Tracks with occlusion score above this are considered reliable and not re-prompted (Table 8).
  • τ_pending occlusion score = 6
    Tracks with occlusion score above this are considered pending, enabling re-prompting (Table 8).
  • τ_lost occlusion score = 2
    Tracks with occlusion score below this are considered lost and can be terminated (Table 8).
  • N_lost consecutive frames = 25
    Number of consecutive frames below τ_lost before a track is terminated (Table 8).
  • N_frames for occlusion std = 10
    Window size for computing occlusion-score standard deviation in cross-object interaction (Table 8).
  • τ_mIoU = 0.8
    Mask overlap threshold for two tracks to be considered in an occlusion relationship (Table 8).
  • τ_Δscore = 2
    Minimum occlusion score difference for deciding which object is occluded (Table 8).
  • τ_Δstd = 0.2
    Minimum occlusion score standard-deviation difference for deciding which object is occluded (Table 8).
  • τ_nms = 0.95
    Mask IoU threshold for non-maximum suppression of track masks (Table 8).
axioms (6)
  • domain assumption Grounding DINO zero-shot detector can detect target animal species when prompted with their common names.
    Foundation of the pipeline; if the detector cannot find animals, the tracker cannot initialize or re-prompt.
  • domain assumption SAM 2 generalizes zero-shot to video of unseen animal species and produces temporally consistent masks.
    Core assumption of the propagation stage; SAM2MOT is used as-is for mask propagation.
  • ad hoc to paper Detection confidence scores from Grounding DINO are approximately bimodal per sequence (true vs false positives).
    Required for adaptive thresholding via Otsu/K-means; the authors concede the assumption does not always hold (Section 2).
  • domain assumption MOT metrics HOTA/DetA/AssA on public benchmarks measure tracking quality relevant to animal behavior.
    Evaluation framework; the paper uses these metrics without justification beyond standard MOT practice.
  • ad hoc to paper The self-defined GMOT-40-Animal subset of 16 sequences is representative of generic animal tracking.
    Subset is only described as 'animal sequences'; sequence IDs are not listed, so representativeness cannot be checked.
  • domain assumption No test-set information was used to choose hyperparameters.
    Central to the zero-shot claim; not verifiable from the paper alone.

pith-pipeline@v1.3.0-alltime-deepseek · 10878 in / 11925 out tokens · 108415 ms · 2026-08-04T00:06:00.205307+00:00 · methodology

0 comments
read the original abstract

Multi-animal tracking is crucial for understanding animal ecology and behavior, yet remains challenging due to variations in habitat, motion patterns, and species appearance. Traditional approaches typically require extensive fine-tuning and heuristic design for each new scenario. In this work, we explore vision foundation models for zero-shot multi-animal tracking. Building on SAM2MOT, we combine Grounding DINO with the Segment Anything Model2 (SAM 2) and introduce three targeted modifications to adapt the framework to animal appearance and behavior without any retraining or hyperparameter tuning between datasets. We also evaluate the recent SAM3 model, but identify practical limitations that restrict its applicability to multi-animal tracking in the wild. Our method achieves state-of-the-art results across Chimp-Act, Bird Flock Tracking, AnimalTrack, and a subset of GMOT-40, demonstrating robust generalization across diverse species and environments. The code is available at https://github.com/ecker-lab/SAM2-Animal-Tracking.

Figures

Figures reproduced from arXiv: 2511.02591 by Jan Frederik Meier, Timo L\"uddecke.

Figure 1
Figure 1. Figure 1: Overview of our model architecture. Blue [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Detection score distributions and adaptive thresholding. (a) In-domain detector scores, showing varied effective [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Detection score distribution from different sequences of the ChimpAct test split. The distribution differs significantly [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Runtime and VRAM requirements for different numbers of tracks. The inference speed and memory consumption [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

41 extracted references · 1 canonical work pages

  1. [1]

    Gmot-40: A benchmark for generic multiple object tracking

    Hexin Bai, Wensheng Cheng, Peng Chu, Juehuan Liu, Kai Zhang, and Haibin Ling. Gmot-40: A benchmark for generic multiple object tracking. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6719–6728, 2021

    2021

  2. [2]

    Simple online and realtime tracking

    Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In2016 IEEE international conference on image processing (ICIP), pages 3464–3468. Ieee, 2016

    2016

  3. [3]

    G. Bradski. The OpenCV Library.Dr. Dobb’s Journal of Software Tools, 2000

    2000

  4. [4]

    Hicks, Sorrel Jones, Kevin Lee, Vera Leinert, Juan Lapuente, Maureen S

    Otto Brookes, Majid Mirmehdi, Colleen Stephens, Samuel Angedakin, Katherine Corogenes, Dervla Dowd, Paula Dieguez, Thurston C. Hicks, Sorrel Jones, Kevin Lee, Vera Leinert, Juan Lapuente, Maureen S. McCarthy, Amelia Meier, Mizuki Murai, Emmanuelle Normand, Virginie Vergnes, Erin G. Wessling, Roman M. Wittig, Kevin Langergraber, Nuria Maldonado, Xinyu Yang...

    2024

  5. [5]

    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. InEuropean conference on computer vision, pages 213–229. Springer, 2020

    2020

  6. [6]

    Ehrlich, Anthony D

    Gerardo Ceballos, Paul R. Ehrlich, Anthony D. Barnosky, Andrés García, Robert M. Pringle, and Todd M. Palmer. Accelerated modern human–induced species losses: Entering the sixth mass extinction.Science Advances, June 2015. Publisher: American Association for the Advancement of Science

    2015

  7. [7]

    Ehrlich, and Peter H

    Gerardo Ceballos, Paul R. Ehrlich, and Peter H. Raven. Vertebrates on the brink as indicators of biological annihilation and the sixth mass extinction.Proceedings of the National Academy of Sciences, 117(24): 13596–13602, June 2020. doi: 10.1073/pnas.1922686117

    work page doi:10.1073/pnas.1922686117 2020

  8. [8]

    Unifying short and long-term tracking with graph hierarchies

    Orcun Cetintas, Guillem Brasó, and Laura Leal-Taixé. Unifying short and long-term tracking with graph hierarchies. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22877–22887, 2023

    2023

  9. [9]

    MMDetection: Open MMLab Detection Toolbox and Benchmark.arXiv preprint arXiv:1906.07155, 2019

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open MMLab Detection Toolbox and B...

    Pith/arXiv arXiv 1906

  10. [10]

    Sportsmot: A large multi-object tracking dataset in multiple sports scenes

    Yutao Cui, Chenkai Zeng, Xiaoyu Zhao, Yichun Yang, Gangshan Wu, and Limin Wang. Sportsmot: A large multi-object tracking dataset in multiple sports scenes. InProceedings of the IEEE/CVF international conference on computer vision, pages 9921–9931, 2023

    2023

  11. [11]

    Sam2long: Enhancing sam 2 for long video segmentation with a training-free memory tree.arXiv preprint arXiv:2410.16268, 2024

    Shuangrui Ding, Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Yuwei Guo, Dahua Lin, and Jiaqi Wang. Sam2long: Enhancing sam 2 for long video segmentation with a training-free memory tree.arXiv preprint arXiv:2410.16268, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Situating Ecology as a Big-Data Science: Current Advances, Challenges, and Solutions.BioScience, 68(8):563–576, August 2018

    Scott S Farley, Andria Dawson, Simon J Goring, and John W Williams. Situating Ecology as a Big-Data Science: Current Advances, Challenges, and Solutions.BioScience, 68(8):563–576, August 2018. ISSN 0006-3568

    2018

  13. [13]

    Multiple object tracking as id prediction

    Ruopeng Gao, Ji Qi, and Limin Wang. Multiple object tracking as id prediction. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 27883–27893, 2025

    2025

  14. [14]

    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

    Pith/arXiv arXiv 2021

  15. [15]

    SAM2MOT: A Novel Paradigm of Multi-Object Tracking by Segmentation, May 2025

    Junjie Jiang, Zelin Wang, Manqi Zhao, Yin Li, and DongSheng Jiang. SAM2MOT: A Novel Paradigm of Multi-Object Tracking by Segmentation, May 2025. arXiv:2504.04519 [cs]

    arXiv 2025

  16. [16]

    TrackEval, 2020

    Arne Hoffhues Jonathon Luiten. TrackEval, 2020

    2020

  17. [17]

    Kerby, Jacob M

    Benjamin Koger, Adwait Deshpande, Jeffrey T. Kerby, Jacob M. Graving, Blair R. Costelloe, and Iain D. Couzin. Quantifying the movement, behaviour and environmental context of group-living animals using drones and computer vision.Journal of Animal Ecology, 92(7):1357–1371, 2023. ISSN 1365-2656. _eprint: https://besjournals.onlinelibrary.wiley.com/doi/pdf/1...

    arXiv 2023

  18. [18]

    Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection, July 2024

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection, July 2024. arXiv:2303.05499 [cs]

    Pith/arXiv arXiv 2024

  19. [19]

    Deep learning in multiple animal tracking: A survey.Computers and Electronics in Agriculture, 224:109161, September 2024

    Yeqiang Liu, Weiran Li, Xue Liu, Zhenbo Li, and Jun Yue. Deep learning in multiple animal tracking: A survey.Computers and Electronics in Agriculture, 224:109161, September 2024. ISSN 0168-1699

    2024

  20. [20]

    Hota: A higher order metric for evaluating multi-object tracking.International journal of computer vision, 129(2):548–578, 2021

    Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixé, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking.International journal of computer vision, 129(2):548–578, 2021. Publisher: Springer

    2021

  21. [21]

    Multiple object tracking: A literature review.Artificial Intelligence, 293:103448, April 2021

    Wenhan Luo, Junliang Xing, Anton Milan, Xiaoqin Zhang, Wei Liu, and Tae-Kyun Kim. Multiple object tracking: A literature review.Artificial Intelligence, 293:103448, April 2021. ISSN 0004-3702

    2021

  22. [22]

    Chimpact: A longitudinal dataset for understanding chimpanzee behaviors.Advances in Neural Information Processing Systems, 36:27501–27531, 2023

    Xiaoxuan Ma, Stephan Kaufhold, Jiajun Su, Wentao Zhu, Jack Terwilliger, Andres Meza, Yixin Zhu, Federico Rossano, and Yizhou Wang. Chimpact: A longitudinal dataset for understanding chimpanzee behaviors.Advances in Neural Information Processing Systems, 36:27501–27531, 2023

    2023

  23. [23]

    AlphaChimp: Tracking and Behavior Recognition of Chimpanzees

    Xiaoxuan Ma, Yutang Lin, Yuan Xu, Stephan P Kaufhold, Jack Terwilliger, Andres Meza, Yixin Zhu, Federico Rossano, and Yizhou Wang. AlphaChimp: Tracking and Behavior Recognition of Chimpanzees. arXiv preprint arXiv:2410.17136, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Trackformer: Multi- object tracking with transformers

    Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi- object tracking with transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8844–8854, 2022

    2022

  25. [25]

    NOOUGAT: Towards Unified Online and Offline Multi-Object Tracking.arXiv preprint arXiv:2509.02111, 2025

    Benjamin Missaoui, Orcun Cetintas, Guillem Brasó, Tim Meinhardt, and Laura Leal-Taixé. NOOUGAT: Towards Unified Online and Offline Multi-Object Tracking.arXiv preprint arXiv:2509.02111, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    A Threshold Selection Method from Gray-Level Histograms.IEEE Transactions on Systems, Man, and Cybernetics, 9(1):62–66, January 1979

    Nobuyuki Otsu. A Threshold Selection Method from Gray-Level Histograms.IEEE Transactions on Systems, Man, and Cybernetics, 9(1):62–66, January 1979. ISSN 2168-2909. 6

    1979

  27. [27]

    Deep MAnTra: deep learning- based multi-animal tracking for Japanese macaques.Artif

    Riza Rae Pineda, Takatomi Kubo, Masaki Shimada, and Kazushi Ikeda. Deep MAnTra: deep learning- based multi-animal tracking for Japanese macaques.Artif. Life Robot., 28(1):127–138, December 2022. ISSN 1433-5298

    2022

  28. [28]

    SAM 2: Segment Anything in Images and Videos, October 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. SAM 2: Segment Anything in Images and Videos, October 2024. arXiv:2408.00714 [cs]

    Pith/arXiv arXiv 2024

  29. [29]

    CAMELTrack: Context-Aware Multi-cue ExpLoitation for Online Multi-Object Tracking.arXiv preprint arXiv:2505.01257, 2025

    Vladimir Somers, Baptiste Standaert, Victor Joos, Alexandre Alahi, and Christophe De Vleeschouwer. CAMELTrack: Context-Aware Multi-cue ExpLoitation for Online Multi-Object Tracking.arXiv preprint arXiv:2505.01257, 2025

    Pith/arXiv arXiv 2025

  30. [30]

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

    Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi- object tracking in uniform appearance and diverse motion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20993–21002, 2022

    2022

  31. [31]

    Costelloe, Silvia Zuffi, Benjamin Risse, Alexander Mathis, Mackenzie W

    Devis Tuia, Benjamin Kellenberger, Sara Beery, Blair R. Costelloe, Silvia Zuffi, Benjamin Risse, Alexander Mathis, Mackenzie W. Mathis, Frank van Langevelde, Tilo Burghardt, Roland Kays, Holger Klinck, Martin Wikelski, Iain D. Couzin, Grant van Horn, Margaret C. Crofoot, Charles V . Stewart, and Tanya Berger-Wolf. Perspectives in machine learning for wild...

  32. [32]

    A distractor-aware memory for visual object tracking with sam2

    Jovana Videnovic, Alan Lukezic, and Matej Kristan. A distractor-aware memory for visual object tracking with sam2. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 24255– 24264, 2025

    2025

  33. [33]

    Weis, Timo Lüddecke, Elif Karakoç, Zurna Ahmed, Sofia M

    Richard V ogg, Matthias Nuske, Marissa A. Weis, Timo Lüddecke, Elif Karakoç, Zurna Ahmed, Sofia M. Pereira, Suchinda Malaivijitnond, Suthirote Meesawat, Derek Murphy, Julia Fischer, Florentin Wörgötter, Peter M. Kappeler, Alexander Gail, Julia Ostner, Oliver Schülke, Claudia Fichtel, and Alexander S. Ecker. PriMAT: A robust multi-animal tracking model for...

    2025

  34. [34]

    Simple online and realtime tracking with a deep association metric

    Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In2017 IEEE international conference on image processing (ICIP), pages 3645–3649. IEEE, 2017

    2017

  35. [35]

    Samurai: Adapting segment anything model for zero-shot visual tracking with motion-aware memory.arXiv preprint arXiv:2411.11922, 2024

    Cheng-Yen Yang, Hsiang-Wei Huang, Wenhao Chai, Zhongyu Jiang, and Jenq-Neng Hwang. Samurai: Adapting segment anything model for zero-shot visual tracking with motion-aware memory.arXiv preprint arXiv:2411.11922, 2024

    Pith/arXiv arXiv 2024

  36. [36]

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

    Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. InEuropean conference on computer vision, pages 659–675. Springer, 2022

    2022

  37. [37]

    AnimalTrack: A Benchmark for Multi-Animal Tracking in the Wild.International Journal of Computer Vision, 131(2):496–513, February 2023

    Libo Zhang, Junyuan Gao, Zhen Xiao, and Heng Fan. AnimalTrack: A Benchmark for Multi-Animal Tracking in the Wild.International Journal of Computer Vision, 131(2):496–513, February 2023. ISSN 1573-1405

    2023

  38. [38]

    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. InEuropean conference on computer vision, pages 1–21. Springer, 2022

    2022

  39. [39]

    Progressive end-to-end object detection in crowded scenes

    Anlin Zheng, Yuang Zhang, Xiangyu Zhang, Xiaojuan Qi, and Jian Sun. Progressive end-to-end object detection in crowded scenes. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 857–866, 2022. 7

    2022

  40. [40]

    torch.cuda.memory.memory_allocated

    Guangze Zheng, Shijie Lin, Haobo Zuo, Changhong Fu, and Jia Pan. Nettrack: Tracking highly dy- namic objects with a net. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19145–19155, 2024. 8 Appendix A Model architecture A.1 SAM 2 SAM 2 [28] is a unified model for both image and video segmentation, treating images...

    2024

  41. [2022]

    Publisher: Nature Publishing Group

    ISSN 2041-1723. Publisher: Nature Publishing Group

    2041