Pith. sign in

REVIEW 3 major objections 5 minor 51 references

Promptable Animal Pose Tracking Across Species

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 The paper argues that frozen vision-foundation features can carry anatomical correspondence well enough to track animal pose in video with very little labelled data, and backs it with two routes that share a frozen extractor.

desk verdict A useful promptable animal pose tracker that gets real mileage out of frozen foundation-model features; the main overstatement is the 'training-free' label when the best results need test-time boxes. read the letter →

arxiv 2608.04995 v1 pith:5LSPMOWZ submitted 2026-08-05 cs.CV

Le Li , Daniela Ivanova , Nicolas Pugeault This is my paper
classification cs.CV
keywords animalposetrackingvisionfoundationmodelscross-speciesgeneralisationcorrespondencematchingvideomonitoringDiffusionHyperfeaturespromptablekeypointsunsupervised
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 argues that dense features from pretrained vision foundation models already carry enough anatomical correspondence to track animal keypoints in video, with far less labelled data than conventional pose pipelines. It demonstrates this with two routes on the same frozen features: a supervised route that adds a small keypoint-prompt encoder and a lightweight matcher, reaching 99.1 PCK@0.1 and 94.7 PCK@0.05 on APTv2 and matching the fine-tuned HRNet W48 baseline; and a training-free unsupervised route that reaches 86.6 PCK@0.1 on APTv2 and 83.0 on TigDog with plain nearest-neighbour matching and an optional drift-correction step. The paper also documents a trade-off: supervised matching wins on occlusion and appearance ambiguity, while the unsupervised route holds up better when whole animal families are held out of training. If the claim holds, animal pose tracking shifts from a large-scale supervised learning problem to a prompt-and-match problem, which matters for ecology and conservation where annotation expertise is scarce and species diversity is high.

What carries the argument

Dense per-pixel features from Diffusion Hyperfeatures, a diffusion-based representation that aggregates information across denoising steps and image scales, carry the argument: both routes treat keypoint tracking as feature correspondence between the reference frame and each target frame, never fine-tuning the extractor. The unsupervised route is the simplest version of this, direct nearest-neighbour matching with an optional drift-correction step that restricts the search to bounding-box regions of matched instances so keypoints do not jump between animals. The supervised route adds two small learned mechanisms: a keypoint prompt encoder that turns the reference keypoints into Gaussian heatmaps and then a single structural attention map, used to modulate the reference features so the matcher attends to keypoint-relevant structure; and a coarse cross-frame matcher adapted from a dense feature-matching architecture. The only learnable components in the entire framework are the prompt encoder, a convolutional feature projection, and the matcher, trained for 50 epochs on APTv2.

What would settle it

Take a video of an animal whose body plan lies far from the quadruped mammals covered by APTv2 and TigDog, such as a snake, a bird in flight, or a seahorse, annotate keypoints in one frame, and run both routes with the same frozen Diffusion Hyperfeatures, using a supervised matcher trained only on the paper's benchmark families. If keypoint accuracy collapses relative to the reported numbers and falls at or below the training-free route on familiar species, the cross-species claim is a property of the tested morphology distribution rather than a general property of the features.

Watch

Extended reading notes

Core claim

The paper's central claim is that vision foundation models trained on large datasets can be used effectively to track animal pose with limited labelled data. Concretely, it proposes a promptable pipeline in which keypoints annotated in a single reference frame are propagated to all other frames by matching dense per-pixel features, with Diffusion Hyperfeatures as the primary extractor. The supervised route converts the reference keypoints into Gaussian heatmaps, encodes them into a structural attention map that modulates the reference features, and learns a cross-frame matcher, reaching 99.1 PCK@0.1 and 94.7 PCK@0.05 on APTv2 while the backbone stays frozen, on par with HRNet W48 (98.5 and 95.2), which was pre-trained on human pose data and fine-tuned on the animal dataset. The unsupervised route performs direct nearest-neighbour matching with no task-specific training at all, reaching 86.6 PCK@0.1 on APTv2, and is competitive on several species with dedicated track-any-point models that were trained explicitly on large tracking datasets. The paper further finds a complementary relationship between the routes: supervised learning excels in challenging cases involving occlusion, ambiguity, and large appearance changes, whereas the unsupervised route degrades less when whole animal families are excluded from training.

Load-bearing premise

The whole pipeline rests on one premise: the frozen foundation-model features, Diffusion Hyperfeatures in particular, place the same anatomical point of an animal close together in feature space across species, poses, and appearances; if those features lose semantic correspondence, the small modules trained in the supervised route cannot compensate because the backbone is frozen.

Editorial extensions

If this is right

  • Keypoints become a prompt rather than a training label: a researcher can annotate one frame and track the same user-defined keypoints through the rest of the video, including points that are not standard anatomical landmarks.
  • The training-free route (86.6 PCK@0.1 on APTv2, 83.0 on TigDog) works without any task-specific training and without lookahead, so it can run in online monitoring settings where labelled data do not exist yet.
  • On APTv2 the supervised route reaches 99.1 PCK@0.1 and 94.7 PCK@0.05 with a frozen backbone, matching the fine-tuned HRNet W48 baseline (98.5 and 95.2) after only 50 epochs of training small modules.
  • Leave-one-out tests show supervised accuracy drops by roughly 29 to 52 PCK@0.05 points on held-out animal families, often landing below the unsupervised route on the same family, so the two routes offer complementary accuracy-versus-generalisation strengths.
  • Prompting the diffusion feature extractor with the species name adds a consistent accuracy gain (roughly +1 to +2 points on both benchmarks, more for some species), so available ecological metadata improves tracking at no annotation cost.

Reading between the lines

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

  • The cross-species claim is demonstrated on 30 mostly quadruped-mammal species in APTv2 plus tigers and horses in TigDog; extending the same frozen-feature matching to very different body plans (birds, fish, snakes, invertebrates) is a test the paper does not run, and nothing in the design guarantees those features carry the needed semantics.
  • Because each frame is matched to the reference frame rather than to its temporal neighbours, error grows with temporal distance; a testable extension suggested by the paper's own middle-frame experiment is anchor-frame ensembling, where keypoints are matched to several reference frames and the results aggregated.
  • The drift-correction module assumes instance bounding boxes are available, so in real footage its value depends on detector quality; chaining a frozen detector with the unsupervised route on multi-animal videos would settle that dependency.
  • If the premise holds, the hard cases the paper identifies (limb extremities, fast motion, left-right ambiguity) point to temporal context rather than better keypoint heads as the next bottleneck, since pairwise feature matching is already near the level of trained track-any-point models.
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 a promptable animal pose tracking framework with two routes. Given a video and keypoints annotated in a single reference frame, the method extracts dense features with a frozen foundation model (DINOv3, BioCLIP, CleanDIFT, or Diffusion Hyperfeatures) and either performs nearest-neighbour matching directly (unsupervised route) or passes the features through a trained keypoint prompt encoder and correspondence matcher (supervised route). Evaluated on APTv2 and TigDog, the supervised route reports 99.1 PCK@0.1 and 94.7 PCK@0.05 on APTv2, close to HRNet W48, while the unsupervised route reports 86.6 PCK@0.1 and 70.5 PCK@0.05, competitive with trained point trackers such as TAPIR and CoTracker3.

Significance. If the results hold, the paper offers a practically useful result: a frozen-backbone pose tracker that approaches fully fine-tuned HRNet W48 performance while training only a small prompt encoder and matcher, and an unsupervised route that achieves competitive point tracking without task-specific training. The paper is also thorough in its ablations: it compares four foundation-model backbones, breaks down performance by body region and species, studies temporal degradation, and includes a leave-one-out family generalization experiment. The main weakness is a mismatch between the claimed 'training-free' and 'practical real-world' setting and the actual evaluation protocol, which relies on dataset-provided bounding boxes for the headline unsupervised number and is restricted to the single-instance Easy subset of APTv2.

major comments (3)
  1. [§3.2, Eqs. (4)–(6); §4.2.1, Table 5; Abstract] The headline unsupervised APTv2 result of 86.6 PCK@0.1 is obtained with the drift-correction module enabled, and this module uses dataset-provided bounding boxes at test time (Eqs. (4)–(6)). Table 5 shows the same configuration without the module reaches 85.6 PCK@0.1. Because the abstract and Section 1 present the unsupervised route as 'training-free correspondence matching' and a practical solution, the main comparison should either use a detector-free pipeline (e.g., no boxes, or boxes from an off-the-shelf detector) or clearly separate the oracle-assisted result from the genuinely deployable one.
  2. [Appendix B; §4.2] All APTv2 results are restricted to the Easy subset, where each video contains a single target instance, as stated in Appendix B. The drift-correction module is motivated by inter-instance drift in multi-animal scenes, but multi-animal quantitative evaluation is not provided; the only qualitative multi-animal example (Figure A1) uses bounding boxes estimated from predictions and is not scored. Without Hard-subset results or a dedicated multi-instance evaluation, the claimed advantage of drift correction and the real-world multi-animal use case remain untested.
  3. [§3.1; §4.2, Table 1; §4.3, Table 6] The best backbone, Diffusion Hyperfeatures, is selected based on performance on the same benchmark used for reporting final results, and the paper does not describe a validation-based selection protocol. This introduces potential optimistic bias in the headline numbers. The authors should either report a model-selection procedure on a held-out validation split or explicitly frame the results as best-of-k backbone selection with the corresponding caveat.
minor comments (5)
  1. [§3.2, Eq. (2)] Equation (2) reads 'NN(F r, F r, P r)' but should read 'NN(F r, F t, P r)' to match the surrounding text and the later definition in Eq. (3).
  2. [Table 3, header] The table header contains the fragment 'bone' instead of 'body region' or 'body part'.
  3. [Tables 1, 2, 6] The main tables report a single run without variance or error bars; given that the reported margins around HRNet W48 and the TAP baselines are small, providing standard deviations across runs or seeds would strengthen the comparison.
  4. [Appendix E, Figure A2 caption] The caption refers to 'Figure A1(a)' and 'Figure A1(b)' when it should refer to 'Figure A2(a)' and 'Figure A2(b)'.
  5. [Section 1, second paragraph] 'Levaraging' should be 'Leveraging'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's predictions are evaluated against held-out keypoint ground truth and are not constructed from the targets they predict.

full rationale

The paper's central claims are empirical and are supported by evaluations on external benchmarks (APTv2 and TigDog) using held-out keypoint ground truth. In the supervised route, the trained modules (keypoint prompt encoder, feature projection, correspondence matcher) are optimized on training data and then evaluated on test frames; the frozen backbone, Diffusion Hyperfeatures, is an external pretrained representation, not a model fitted to APTv2 keypoints. In the unsupervised route, keypoint predictions are produced by nearest-neighbour matching between frozen foundation-model features (Eq. 2-3), with no task-specific training. The optional drift-correction module (Eq. 4-6) restricts the search region using dataset-provided bounding boxes, but bounding boxes are coarse instance-level cues, not the keypoint targets being scored, and the paper explicitly states that this correction is optional; the TigDog results are reported without it. The choice of Diffusion Hyperfeatures as the primary backbone is informed by benchmark comparisons (Section 3.1), which is a mild model-selection caveat rather than a circular reduction. Appendix B further notes that evaluation is restricted to the APTv2 Easy subset, a scope limitation, but not a definitional circularity. There are no load-bearing self-citations, no imported uniqueness theorems, and no fitted quantities renamed as predictions. None of the paper's equations reduce the target predictions to their inputs by construction. The main quantitative results therefore stand as independent, externally benchmarked evidence for the paper's claims.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method relies on pretrained foundation-model features (a domain assumption), a hand-set heatmap spread, a learnable modulation scalar, and the choice of backbone based on benchmark performance. No new physical or conceptual entities are introduced; all components are neural network modules and standard algorithmic steps.

free parameters (3)
  • Gaussian heatmap spread sigma = not reported (hand-set)
    Controls the spread of the prompt heatmap in Eq. 9, affecting how much nearby context is emphasized around each keypoint.
  • Feature modulation scalar alpha = learned
    Learnable scalar in Eq. 11 that controls the strength of the keypoint attention modulation.
  • Backbone choice = Diffusion Hyperfeatures
    Selected as the primary feature extractor based on benchmark performance among DINOv3, BioCLIP, CleanDIFT, and DHf.
assumptions (4)
  • domain assumption Pretrained foundation-model features (DHf) preserve semantic keypoint correspondence across animal species and poses.
    Section 3.1 adopts Diffusion Hyperfeatures as the feature extractor; the entire pipeline relies on these features transferring without backbone fine-tuning.
  • domain assumption Evaluation on the APTv2 Easy subset (single target instance) is representative of the real-world multi-animal settings discussed in the introduction.
    Appendix B limits experiments to the Easy subset; claims about practical real-world use extend beyond this evaluation.
  • domain assumption Bounding boxes are available or can be estimated at test time for drift correction.
    Section 3.2 uses dataset-provided boxes in Eq. 4 to Eq. 6; the paper notes they can be estimated (Figure A1) but this adds test-time dependency.
  • standard math Standard machine learning background: nearest-neighbour matching, Gaussian heatmaps, supervised training objective.
    Used throughout Sections 3.2 and 3.3 without proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Promptable Animal Pose Tracking Across Species." pith.science (2026). https://pith.science/paper/5LSPMOWZ

@misc{pith2026260804995,
  author       = {Pith},
  title        = {Pith review of: Promptable Animal Pose Tracking Across Species},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5LSPMOWZ}},
  note         = {Machine review of arXiv:2608.04995}
}
read the original abstract

Animal pose estimation and tracking is important for wildlife monitoring and conservation research, and with limited expert time for labelling automated approaches are imperative. While human pose estimation and tracking has seen rapid progress thanks to large annotated datasets, animal pose remain challenging, due to large morphological and behavioural differences between species and limited annotated data. Existing approaches either optimise generic keypoint localisation from annotated datasets (such as APTv2) with poor generalisation, or track custom keypoints using visual tracking, at the cost of performance. In this paper, we demonstrate that vision foundation models trained on large datasets can be used effectively to track animal pose with limited labelled data. We propose two models, one unsupervised and the other supervised, to track user-selected keypoints in videos. The supervised approach delivers superior tracking accuracy by employing a keypoint prompt encoder to explicitly inject structural priors from a reference frame into feature matching. In parallel, the unsupervised route provides strong cross-species robustness by leveraging diverse foundation-model features for training-free correspondence matching. Extensive evaluation on challenging animal video benchmarks APTv2 and TigDog demonstrates that our framework achieves strong performance while maintaining an effective balance between accuracy and generalisation, offering a practical solution for real-world animal behaviour analysis and conservation applications.

Figures

Figures reproduced from arXiv: 2608.04995 by the authors.

Figure 1
Figure 1. A concise overview of our proposed promptable animal pose tracking pipeline. The supervised route [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall pipeline of the unsupervised route. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overall framework of the supervised route. The keypoint prompt encoder, feature projection, and [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of tracking results across increasing tem [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Comparison between supervised and unsupervised routes. Left: failure cases of the unsupervised method [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 45 canonical work pages

  1. [1]

    Consistent multi-animal pose estimation in cattle using dynamic kalman filter based tracking

    Maarten Perneel, Ines Adriaens, Ben Aernouts, and Jan V erwaeren. Consistent multi-animal pose estimation in cattle using dynamic kalman filter based tracking. Smart agricultural technology, 11:101014, 2025

    work page 2025

  2. [2]

    Real-time cattle pose estimation based on improved rtmpose

    Xiaowu Li, Kun Sun, Hongbo Fan, and Zihan He. Real-time cattle pose estimation based on improved rtmpose. Agriculture, 13:1938, 2023

    work page 1938

  3. [3]

    Multi-animal 3d social pose estimation, identification and behaviour embedding with a few-shot learning framework

    Y aning Han, Ke Chen, Y unke Wang, Wenhao Liu, Zhouwei Wang, Xiaojing Wang, Chuanliang Han, Jiahui Liao, Kang Huang, Shengyuan Cai, et al. Multi-animal 3d social pose estimation, identification and behaviour embedding with a few-shot learning framework. Nature machine intelligence, 6:48–61, 2024

    work page 2024

  4. [4]

    Whole-Herd Elephant Pose Estimation from Drone Data for Collective Behavior Analysis

    Brody McNutt, Libby Zhang, Angus Carey-Douglas, Fritz V ollrath, Frank Pope, and Leandra Brickson. Whole- herd elephant pose estimation from drone data for collective behavior analysis. arXiv preprint arXiv:2411.00196, 2024

    work page Pith review arXiv 2024

  5. [5]

    Omni-supervised joint detection and pose estimation for wild animals

    Teng Zhang, Liangchen Liu, Kun Zhao, Arnold Wiliem, Graham Hemson, and Brian Lovell. Omni-supervised joint detection and pose estimation for wild animals. Pattern recognition letters, 132:84–90, 2020

    work page 2020

  6. [6]

    Towards multi-modal animal pose estimation: A survey and in-depth analysis

    Qianyi Deng, Oishi Deb, Amir Patel, Christian Rupprecht, Philip Torr, Niki Trigoni, and Andrew Markham. Towards multi-modal animal pose estimation: A survey and in-depth analysis. arXiv preprint arXiv:2410.09312, 2024. 11

    arXiv 2024

  7. [7]

    Multi-animal pose estimation and tracking with deeplabcut

    Jessy Lauer, Mu Zhou, Shaokai Y e, William Menegas, Tanmay Nath, Mohammed Mostafizur Rahman, V alentina Di Santo, Daniel Soberanes, Guoping Feng, V enkatesh N Murthy, et al. Multi-animal pose estimation and tracking with deeplabcut. BioRxiv, pages 2021–04, 2021

    work page 2021

  8. [8]

    Multi-animal pose estimation, identification and tracking with deeplabcut

    Jessy Lauer, Mu Zhou, Shaokai Y e, William Menegas, Steffen Schneider, Tanmay Nath, Mohammed Mostafizur Rahman, V alentina Di Santo, Daniel Soberanes, Guoping Feng, et al. Multi-animal pose estimation, identification and tracking with deeplabcut. Nature methods, 19:496–504, 2022

    work page 2022

Show all 51 references
  1. [9]

    Sleap: Multi-animal pose tracking

    Talmo D Pereira, Nathaniel Tabris, Junyu Li, Shruthi Ravindranath, Eleni S Papadoyannis, Z Y an Wang, David M Turner, Grace McKenzie-Smith, Sarah D Kocher, Annegret L Falkner, et al. Sleap: Multi-animal pose tracking. BioRxiv, pages 2020–08, 2020

    2020

  2. [10]

    Fast animal pose estimation using deep neural networks

    Talmo D Pereira, Diego E Aldarondo, Lindsay Willmore, Mikhail Kislin, Samuel S-H Wang, Mala Murthy, and Joshua W Shaevitz. Fast animal pose estimation using deep neural networks. Nature methods, 16:117–125, 2019

    2019

  3. [11]

    I-muppet: Interactive multi-pigeon pose estimation and tracking

    Urs Waldmann, Hemal Naik, Nagy Máté, Fumihiro Kano, Iain D Couzin, Oliver Deussen, and Bastian Goldlücke. I-muppet: Interactive multi-pigeon pose estimation and tracking. In Dagm German conference on pattern recog- nition, pages 513–528, 2022

    2022

  4. [12]

    3d-muppet: 3d multi-pigeon pose estimation and tracking

    Urs Waldmann, Alex Hoi Hang Chan, Hemal Naik, Máté Nagy, Iain D Couzin, Oliver Deussen, Bastian Gold- luecke, and Fumihiro Kano. 3d-muppet: 3d multi-pigeon pose estimation and tracking. International journal of computer vision, 132:4235–4252, 2024

    2024

  5. [13]

    Pigpose: A realtime frame- work for farm animal pose estimation and tracking

    Milan Kresovic, Thong Nguyen, Mohib Ullah, Hina Afridi, and Faouzi Alaya Cheikh. Pigpose: A realtime frame- work for farm animal pose estimation and tracking. In IFIP International conference on artificial intelligence applications and innovations, pages 204–215, 2022

    2022

  6. [14]

    Adoption of mmpose, a general purpose pose estimation library, for animal tracking

    Jessica D Choi and Vivek Kumar. Adoption of mmpose, a general purpose pose estimation library, for animal tracking. BioRxiv, pages 2026–03, 2026

    2026

  7. [15]

    Learning from synthetic animals

    Jiteng Mu, Weichao Qiu, Gregory D Hager, and Alan L Y uille. Learning from synthetic animals. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 12386–12395, 2020

    2020

  8. [16]

    From synthetic to real: Unsupervised domain adaptation for animal pose estimation

    Chen Li and Gim Hee Lee. From synthetic to real: Unsupervised domain adaptation for animal pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 1482–1491, 2021

    2021

  9. [17]

    Sydog-video: A synthetic dog video dataset for temporal pose estimation

    Moira Shooter, Charles Malleson, and Adrian Hilton. Sydog-video: A synthetic dog video dataset for temporal pose estimation. International journal of computer vision , 132:1986–2002, 2024

    1986

  10. [18]

    Scarcenet: Animal pose estimation with scarce annotations

    Chen Li and Gim Hee Lee. Scarcenet: Animal pose estimation with scarce annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 17174–17183, 2023

    2023

  11. [19]

    Superanimal pretrained pose estimation models for behavioral analysis.Nature communications, 15:5165, 2024

    Shaokai Y e, Anastasiia Filippova, Jessy Lauer, Steffen Schneider, Maxime Vidal, Tian Qiu, Alexander Mathis, and Mackenzie Weygandt Mathis. Superanimal pretrained pose estimation models for behavioral analysis.Nature communications, 15:5165, 2024

    2024

  12. [20]

    Oriane Siméoni, Huy V V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, V asil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, et al. Dinov3. arXiv preprint arXiv:2508.10104, 2025

    2025 arXiv

  13. [21]

    Diffusion hyperfeatures: Searching through time and space for semantic correspondence

    Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Advances in neural information processing systems, 36:47500–47510, 2023

    2023

  14. [22]

    Cleandift: Diffusion features without noise

    Nick Stracke, Stefan Andreas Baumann, Kolja Bauer, Frank Fundel, and Björn Ommer. Cleandift: Diffusion features without noise. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 117–127, 2025

    2025

  15. [23]

    Dino-tracker: Taming dino for self-supervised point tracking in a single video

    Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. In European conference on computer vision , pages 367–385. Springer, 2024

    2024

  16. [24]

    Zebrapose: Zebra detection and pose estimation using only synthetic data

    Elia Bonetto and Aamir Ahmad. Zebrapose: Zebra detection and pose estimation using only synthetic data. In Proceedings of the IEEE winter conference on applications of computer vision , pages 6611–6620, 2026

    2026

  17. [25]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 770–778, 2016

    2016

  18. [26]

    Stacked hourglass networks for human pose estimation

    Alejandro Newell, Kaiyu Y ang, and Jia Deng. Stacked hourglass networks for human pose estimation. In European conference on computer vision , pages 483–499, 2016

    2016

  19. [27]

    Openpose: Realtime multi-person 2d pose estimation using part affinity fields

    Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Y aser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence , 43:172–186, 2019. 12

    2019

  20. [28]

    Deep high-resolution representation learning for visual recognition

    Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Y ang Zhao, Dong Liu, Y adong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence , 43:3349–...

    2020

  21. [29]

    Ap-10k: A benchmark for animal pose estimation in the wild

    Hang Y u, Y ufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. Ap-10k: A benchmark for animal pose estimation in the wild. arXiv preprint arXiv:2108.12617, 2021

    2021 arXiv

  22. [30]

    Cross-domain adapta- tion for animal pose estimation

    Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Y u-Wing Tai. Cross-domain adapta- tion for animal pose estimation. In Proceedings of the IEEE international conference on computer vision , pages 9498–9507, 2019

    2019

  23. [31]

    Bioclip: A vision foundation model for the tree of life

    Samuel Stevens, Jiaman Wu, Matthew J Thompson, Elizabeth G Campolongo, Chan Hee Song, David Edward Carlyn, Li Dong, Wasila M Dahdul, Charles Stewart, Tanya Berger-Wolf, et al. Bioclip: A vision foundation model for the tree of life. In Proceedings of the IEEE conference on com...

    2024

  24. [32]

    Diffpose: Spatiotemporal diffusion model for video-based human pose estimation

    Runyang Feng, Yixing Gao, Tze Ho Elden Tse, Xueqing Ma, and Hyung Jin Chang. Diffpose: Spatiotemporal diffusion model for video-based human pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 14861–14872, 2023

    2023

  25. [33]

    Diffusion model with temporal constraint for 3d human pose estimation: Z

    Zhangmeng Chen, Ju Dai, Junjun Pan, and Feng Zhou. Diffusion model with temporal constraint for 3d human pose estimation: Z. chen et al. The visual computer, 41:5961–5977, 2025

    2025

  26. [34]

    Diffusion-based action recognition gener- alizes to untrained domains

    Rogério Guimarães, Frank Xiao, Pietro Perona, and Markus Marks. Diffusion-based action recognition gener- alizes to untrained domains. In Proceedings of the IEEE winter conference on applications of computer vision , pages 5919–5933, 2026

    2026

  27. [35]

    Dposer-x: Diffusion model as robust 3d whole-body human pose prior

    Junzhe Lu, Jing Lin, Hongkun Dou, Ailing Zeng, Y ue Deng, Xian Liu, Zhongang Cai, Lei Y ang, Y ulun Zhang, Haoqian Wang, et al. Dposer-x: Diffusion model as robust 3d whole-body human pose prior. In Proceedings of the IEEE international conference on computer vision , pages 99...

    2025

  28. [36]

    T-leap: Occlusion-robust pose estimation of walking cows using temporal information

    Helena Russello, Rik V an Der Tol, and Gert Kootstra. T-leap: Occlusion-robust pose estimation of walking cows using temporal information. Computers and electronics in agriculture, 192:106559, 2022

    2022

  29. [37]

    Deeplabcut: markerless pose estimation of user-defined body parts with deep learning

    Alexander Mathis, Pranav Mamidanna, Kevin M Cury, Taiga Abe, V enkatesh N Murthy, Mackenzie Weygandt Mathis, and Matthias Bethge. Deeplabcut: markerless pose estimation of user-defined body parts with deep learning. Nature neuroscience, 21:1281–1289, 2018

    2018

  30. [38]

    Simple online and realtime tracking

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

    2016

  31. [39]

    Animal pose labeling using general-purpose point trackers

    Zhuoyang Pan, Boxiao Pan, Guandao Y ang, Adam W Harley, and Leonidas Guibas. Animal pose labeling using general-purpose point trackers. arXiv preprint arXiv:2506.03868, 2025

    2025 arXiv

  32. [40]

    Co- tracker3: Simpler and better point tracking by pseudo-labelling real videos

    Nikita Karaev, Y uri Makarov, Jianyuan Wang, Natalia Neverova, Andrea V edaldi, and Christian Rupprecht. Co- tracker3: Simpler and better point tracking by pseudo-labelling real videos. In Proceedings of the IEEE interna- tional conference on computer vision , pages 6013–6022, 2025

    2025

  33. [41]

    Tapir: Tracking any point with per-frame initialization and temporal refinement

    Carl Doersch, Yi Y ang, Mel V ecerik, Dilara Gokay, Ankush Gupta, Y usuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE international conference on computer vision , pages 1...

    2023

  34. [42]

    Alltracker: Efficient dense point tracking at high resolution

    Adam W Harley, Y ang Y ou, Xinglong Sun, Y ang Zheng, Nikhil Raghuraman, Y unqi Gu, Sheldon Liang, Wen- Hsuan Chu, Achal Dave, Suya Y ou, et al. Alltracker: Efficient dense point tracking at high resolution. In Proceedings of the IEEE international conference on computer vision...

    2025

  35. [43]

    Apt-36k: A large-scale bench- mark for animal pose estimation and tracking

    Y uxiang Y ang, Junjie Y ang, Y ufei Xu, Jing Zhang, Long Lan, and Dacheng Tao. Apt-36k: A large-scale bench- mark for animal pose estimation and tracking. Advances in neural information processing systems , 35:17301– 17313, 2022

    2022

  36. [44]

    Roma v2: Harder better faster denser feature matching

    Johan Edstedt, David Nordström, Y ushan Zhang, Georg Bökman, Jonathan Astermark, Viktor Larsson, Anders Heyden, Fredrik Kahl, Mårten Wadenbäck, and Michael Felsberg. Roma v2: Harder better faster denser feature matching. arXiv preprint arXiv:2511.15706, 2025

    2025 arXiv

  37. [45]

    Aptv2: Benchmarking animal pose estimation and tracking with a large-scale dataset and beyond

    Y uxiang Y ang, Yingqi Deng, Y ufei Xu, and Jing Zhang. Aptv2: Benchmarking animal pose estimation and tracking with a large-scale dataset and beyond. arXiv preprint arXiv:2312.15612, 2023

    2023 arXiv

  38. [46]

    Behavior discovery and alignment of articulated object classes from unstructured video

    Luca Del Pero, Susanna Ricco, Rahul Sukthankar, and Vittorio Ferrari. Behavior discovery and alignment of articulated object classes from unstructured video. International journal of computer vision , 121:303–325, 2017. 13

    2017

  39. [47]

    Tap-vid: A benchmark for tracking any point in a video

    Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Y usuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Y ang. Tap-vid: A benchmark for tracking any point in a video. Advances in neural information processing systems, 35:13610–13626, 2022

    2022

  40. [48]

    Dynamicstereo: Consistent dynamic depth from stereo videos

    Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea V edaldi, and Christian Rupprecht. Dynamicstereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE conference on com- puter vision and pattern recognition , pages 13229–13239, 2023

    2023

  41. [49]

    Pointodyssey: A large- scale synthetic dataset for long-term point tracking

    Y ang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large- scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE international conference on computer vision, pages 19855–19865, 2023

    2023

  42. [50]

    Bootstap: Bootstrapped training for tracking-any-point

    Carl Doersch, Pauline Luc, Yi Y ang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, Joao Carreira, et al. Bootstap: Bootstrapped training for tracking-any-point. In Proceed- ings of the asian conference on computer vision , pages 3257...

    2024

  43. [51]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision , pages 740–755, 2014. 14 Promptable Animal Pose Tracking Across Spe...

    2014

Pith tools

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