Pith. sign in

REVIEW 3 major objections 6 minor 38 references

Gait signatures from ordinary wildlife video can identify individual animals without tags or markings.

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 · grok-4.5

2026-07-11 18:07 UTC pith:NM2SUOXO

load-bearing objection Solid multi-species wildlife gait pipeline with clean SAM3 masks and strong reported numbers, but the supervised fine-tuning on the same identities makes the clustering results circular and the open-set claim unproven. the 3 major comments →

arxiv 2607.04518 v1 pith:NM2SUOXO submitted 2026-07-05 cs.CV

A non-invasive video-based method for individual identification of wildlife using gait dynamics

Muhammad Aamir , Matthew Wijers , Sangyun Shin , Andrew Loveridge , Andrew Markham This is my paper
classification cs.CV
keywords gait analysisunique identificationspatiotemporal learningvideo pipelinewildlife biometricsnon-invasive monitoringcosine similarity clustering
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 the way an animal walks is distinctive enough to identify it as an individual, even when the only input is ordinary video filmed in the wild. The authors assemble a fully automatic pipeline: a foundation segmentation model isolates the animal from messy natural backgrounds, a convolutional network extracts body-shape cues, and a video transformer captures the rhythm of the stride. The resulting gait embeddings are compared with cosine similarity, so that clips of the same animal form tight clusters while different animals stay apart. Across five species the method produces high within-animal and low between-animal similarity scores, showing that gait can serve as a non-invasive biometric. A sympathetic reader cares because conservationists still rely on collars, tags or distinctive markings that require capture or close contact; a camera-only alternative would scale monitoring without stressing the animals.

Core claim

Fine-tuned spatial and temporal embeddings extracted from SAM3-segmented lateral walking videos form gait signatures whose cosine similarities are consistently high for the same individual (mean 0.95–0.98) and substantially lower for different individuals (roughly 0.75–0.81), allowing unsupervised clustering of wildlife without physical markings or invasive tagging.

What carries the argument

The fused gait embedding: ResNet18 spatial features concatenated with VideoPrism temporal features, averaged over a segmented video clip and compared by cosine similarity.

Load-bearing premise

The curated set of mostly side-view walking clips is representative enough of real gait variation that the same tight clusters will appear for new animals, new camera angles and new terrains.

What would settle it

Collect a fresh set of walking videos of previously unseen individuals of the same species, filmed from frontal or rear viewpoints or on different substrates, and check whether the new clips still produce high within-animal and low between-animal cosine similarities under the same pipeline.

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

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

3 major / 6 minor

Summary. The paper proposes a fully automated video pipeline for non-invasive individual identification of wildlife from gait. SAM3 produces RGB and binary silhouettes; ResNet18 and VideoPrism are fine-tuned with a classification objective and then used as feature extractors; sequence-level embeddings are compared by cosine similarity and clustered (k-means / hierarchical). On 185 multi-source lateral walking clips of five species (camel, lion, giraffe, zebra, hyena; 4–6 individuals each), the authors report mean intra-animal cosine similarity 0.95–0.98, inter-animal ~0.75–0.81, average silhouette 0.78, and clear block-diagonal structure in similarity matrices and dendrograms. They conclude that gait dynamics alone support scalable, marking-free identification.

Significance. If the empirical claims hold under a proper open-set protocol, the work would be a useful contribution to ecological monitoring: a species-agnostic, non-invasive alternative to collars and tags that leverages foundation models (SAM3, VideoPrism) and requires only lateral walking video. The multi-species design and the explicit fusion of spatial (ResNet18) and temporal (VideoPrism) features are strengths relative to much of the animal-gait literature, which is still dominated by marker-based or single-species studies. The reported numbers (Table 2, silhouette 0.78) would be competitive if they survive held-out-identity evaluation. The manuscript does not ship code, proofs, or parameter-free derivations; its value is empirical and systems-oriented.

major comments (3)
  1. §3.3 (and abstract): Both ResNet18 and VideoPrism are fine-tuned with a supervised classification loss L_cls whose labels are the individual identities later used for cosine-similarity clustering. The embeddings are therefore optimized to separate those same identities; the subsequent “unsupervised” clustering largely recovers a partition already injected into the feature space. This makes the intra/inter gap in Table 2 and Figures 3–5 circular with respect to the central claim that gait signatures alone enable identification without reliance on labeled identity. A load-bearing fix is required: leave-one-individual-out (or leave-one-clip-out with no identity overlap) fine-tuning, or a pure frozen-backbone / self-supervised protocol, plus open-set verification metrics (e.g., ROC/EER on held-out animals).
  2. §3.1 and §4: The dataset comprises only 4–6 individuals per species and 5–20 snippets each, predominantly lateral walks, with no held-out individuals, no multi-view (frontal/rear) test set, and no cross-environment or cross-speed protocol. The weakest assumption of the paper—that the observed block-diagonal structure generalizes to new animals, viewpoints, and terrains—is never stress-tested. Without at least one held-out-identity experiment and a clear statement of closed-set vs open-set performance, the claim of “scalable ecological monitoring” is not supported by the current evidence.
  3. §4.3 / Table 2: Ablations compare ResNet18-only, VideoPrism-only, and fusion, but there is no comparison to standard gait baselines used in the literature the authors cite (GEI, GaitSet, silhouette CNNs, pose/skeleton GCNs, or simple optical-flow features). Without such baselines on the same clips, it is impossible to judge whether the reported 0.95+ intra-similarity is due to gait dynamics, residual appearance after SAM3, or the supervised fine-tuning itself. Adding at least one classical and one modern gait re-ID baseline under the same evaluation protocol is necessary for the comparative claim.
minor comments (6)
  1. §3.2: SAM3 is cited via arXiv:2511.16719; ensure the model name and citation match the publicly available Segment Anything lineage (SAM / SAM2) or clarify if SAM3 is a distinct release, to avoid reproducibility confusion.
  2. Figure 5: Confusion matrices are described as “derived from cosine similarity” but the mapping from continuous similarity to discrete confusion cells is not specified (threshold? nearest-neighbor assignment?). State the decision rule explicitly.
  3. §3.5: Identification decision uses both argmax_g sim and a threshold τ, but τ is never reported or cross-validated. Report the operating threshold and sensitivity to it.
  4. Notation: temporal window τ for VideoPrism and similarity threshold τ share the same symbol; rename one to avoid collision.
  5. Table 1 and literature review: several human gait methods are listed; a short paragraph situating the present multi-species wildlife setting against recent animal re-ID / camera-trap work (beyond DeepLabCut) would strengthen positioning.
  6. Minor language: “Identifable” → “identifiable” (§5); “specie locomotion” → “species’ locomotion” (§2.2); inconsistent hyphenation of “spatiotemporal” / “spatio-temporal”.

Circularity Check

1 steps flagged

Supervised classification fine-tuning on the same individual identities later clustered by cosine similarity makes the reported intra/inter gap and silhouette scores recover the training signal by construction rather than demonstrate open-set gait biometrics.

specific steps
  1. fitted input called prediction [§3.3 Deep Feature Extraction (loss L_cls) + §4 Results (Table 2, Figs. 3–5)]
    "Both models are fine-tuned using a classification objective and subsequently used as feature extractors to generate discriminative gait representations. ... L_cls = −∑_i log exp(W h̄_i + b)_y_i / ∑_c exp(W h̄_i + b)_c ... Cosine similarity is then used to compare gait signatures, enabling similarity-based clustering of individuals ... Intra-animal cosine similarity values ranged between 0.91 and 0.99 ... average silhouette score of 0.78"

    The classification labels y_i are the individual identities. Fine-tuning therefore optimizes the very embeddings h̄ later compared by cosine similarity and clustered. With no held-out identities, the high intra-similarity and silhouette scores are statistically forced by the supervised objective rather than an independent test of gait dynamics; the “unsupervised” clustering recovers the partition already injected during fine-tuning.

full rationale

The paper's central empirical claim (high intra-animal cosine similarity 0.95–0.98, inter ~0.75–0.81, silhouette 0.78, and clean k-means/hierarchical clusters across five species) rests on embeddings extracted after fine-tuning ResNet18 and VideoPrism with a supervised classification loss whose class labels are precisely the individual animals. No held-out-identity, leave-one-individual-out, or open-set protocol is reported; the same labeled clips (or overlapping individuals) therefore appear in both the fine-tuning stage and the subsequent “unsupervised” similarity/clustering stage. The block-diagonal structure and high silhouette therefore largely recover the supervision already injected into the feature space rather than independently discovering gait signatures. This is a classic fitted-input-called-prediction circularity of moderate severity: the pipeline still contains non-circular engineering (SAM3 segmentation, dual spatial-temporal fusion, multi-species data), but the load-bearing quantitative support for the identification claim reduces by construction to the classification objective. Score 6 reflects partial circularity confined to the evaluation of the embeddings; the rest of the pipeline is self-contained.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard deep-learning practice plus several domain and experimental assumptions that are not independently verified outside this paper. No new physical entities are invented; free parameters are the usual neural-network and clustering knobs plus the similarity threshold τ.

free parameters (5)
  • classification fine-tuning weights (W, b) and network parameters θ_R, θ_V
    Optimized on the same individual labels later recovered by clustering; values not reported.
  • temporal window length τ for VideoPrism
    Controls how much motion context is encoded; chosen by authors, value not given.
  • similarity threshold τ for identification decision
    Used in the match-if-max-sim ≥ τ rule; not specified numerically.
  • number of clusters k and hierarchical linkage parameters
    Chosen to match known number of individuals; affects silhouette and dendrogram appearance.
  • λ_temp (temporal consistency weight in SAM3 refinement)
    Balances data and temporal terms in the mask objective; value not reported.
axioms (5)
  • domain assumption Gait dynamics of an individual are sufficiently stable across recording sessions and sufficiently distinct from other individuals of the same species to serve as a biometric under natural conditions.
    Stated throughout Introduction and §2.3; never independently measured outside the present embeddings.
  • domain assumption SAM3 produces temporally consistent, identity-preserving masks without species-specific training.
    Invoked in §3.2; performance is illustrated qualitatively but not quantified with IoU or mask-error metrics.
  • domain assumption Cosine similarity of mean-pooled spatiotemporal embeddings is an adequate metric for individual identity.
    Used as the sole comparison method in §3.4–3.5.
  • ad hoc to paper Lateral walking sequences contain the kinematic information necessary for reliable identification; frontal/rear views are deferred to future work.
    Explicitly restricted in Conclusion; all quantitative claims are conditioned on this viewpoint.
  • standard math Standard cross-entropy classification loss on individual labels produces embeddings that generalize to new clips of the same animals.
    Ordinary deep-metric-learning practice; used in §3.3.

pith-pipeline@v1.1.0-grok45 · 14931 in / 3423 out tokens · 56360 ms · 2026-07-11T18:07:56.574205+00:00 · methodology

0 comments
read the original abstract

Gait is a distinctive behavioral characteristic that enables non-invasive individual identification without requiring physical interaction with an animal. While gait-based analysis has been extensively studied in humans, its application to wildlife remains limited due to environmental variability and the lack of scalable identification methods. This paper presents a fully automated, video-based pipeline for wildlife gait analysis and individual identification using deep spatiotemporal representation learning. The proposed pipeline uses the Segment Anything Model 3 (SAM3) to generate high-quality RGB and binary silhouette masks, robustly isolating animals from complex natural backgrounds. Segmented video sequences are processed using a convolutional neural network (ResNet18) for spatial feature extraction and a transformer-based video model (VideoPrism) for temporal motion modeling. Both models are fine-tuned using a classification objective and subsequently used as feature extractors to generate discriminative gait representations. Cosine similarity is then used to compare gait signatures, enabling similarity-based clustering of individuals without reliance on physical markings or invasive tagging. Experiments conducted on multi-source wildlife video data across multiple species demonstrate strong intra-individual consistency and clear inter-individual separation. Quantitative results using cosine similarity distributions and silhouette scores confirm the effectiveness of the proposed method. These findings demonstrate that gait dynamics provide a viable, non-invasive approach for individual identification in wildlife and highlight the potential of video-based deep learning pipelines for scalable ecological monitoring.

Figures

Figures reproduced from arXiv: 2607.04518 by Andrew Loveridge, Andrew Markham, Matthew Wijers, Muhammad Aamir, Sangyun Shin.

Figure 1
Figure 1. Figure 1: The proposed video based pipeline for unique identification using gaits analysis [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Masks generation using SAM3 from random frames of Lion, Giraffe, and Hyena videos. First column represent [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: K-mean clustering using Multidimensional Scaling (MDS) to project the cosine similarity into 2D space [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Hierarchical clustering dendrograms illustrating gait-based similarity between individual animals for each [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Similarity in the form of confusion matrix for the all five species [PITH_FULL_IMAGE:figures/full_fig_p011_5.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

38 extracted references · 3 linked inside Pith

  1. [1]

    K. Y. Pouran, Biomechanics, behaviour dynamics and archaeology: Integrative attempts to study animal domestication and husbandry, Ph.D. thesis, Universitat Autònoma de Barcelona, 2023

    2023

  2. [2]

    M. W. Whittle, Gait analysis: an introduction, Butterworth-Heinemann, 2014

    2014

  3. [3]

    Muybridge, Animals in motion dover publications, Mineóla NY (1957)

    E. Muybridge, Animals in motion dover publications, Mineóla NY (1957)

    1957

  4. [4]

    Kirtley, Clinical gait analysis: theory and practice, Elsevier Health Sciences, 2006

    C. Kirtley, Clinical gait analysis: theory and practice, Elsevier Health Sciences, 2006

    2006

  5. [5]

    T. E. Dennis, S. F. Shah, Assessing acute effects of trapping, handling, and tagging on the behavior of wildlife using gps telemetry: a case study of the common brushtail possum, Journal of Applied Animal Welfare Science 15 (2012) 189–207

    2012

  6. [6]

    M. S. Norouzzadeh, A. Nguyen, M. Kosmala, A. Swanson, M. S. Palmer, C. Packer, J. Clune, Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning, Proceedings of the National Academy of Sciences 115 (2018) E5716–E5725

    2018

  7. [7]

    Carion, L

    N. Carion, L. Gustafson, Y.-T. Hu, S. Debnath, R. Hu, D. Suris, C. Ryali, K. V. Alwala, H. Khedr, A.Huang, etal., Sam3: Segmentanythingwithconcepts, arXivpreprintarXiv:2511.16719(2025)

    Pith/arXiv arXiv 2025

  8. [8]

    N. Ravi, V. Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rol- land, L. Gustafson, et al., Sam 2: Segment anything in images and videos, arXiv preprint arXiv:2408.00714 (2024)

    Pith/arXiv arXiv 2024

  9. [9]

    Z. Chen, Y. Jiang, X. Zhang, R. Zheng, R. Qiu, Y. Sun, C. Zhao, H. Shang, Resnet18dnn: prediction approach of drug-induced liver injury by deep neural network with resnet18, Briefings in bioinformatics 23 (2022) bbab503

    2022

  10. [10]

    L. Zhao, N. B. Gundavarapu, L. Yuan, H. Zhou, S. Yan, J. J. Sun, L. Friedman, R. Qian, T. Weyand, Y. Zhao, et al., Videoprism: A foundational visual encoder for video understanding, arXiv preprint arXiv:2402.13217 (2024)

    Pith/arXiv arXiv 2024

  11. [11]

    M. L. George, T. Govindarajan, K. Angamuthu Rajasekaran, S. R. Bandi, A robust similarity based deep siamese convolutional neural network for gait recognition across views, Computational Intelligence 36 (2020) 1290–1319

    2020

  12. [12]

    M. S. Nixon, T. Tan, R. Chellappa, Human identification based on gait, volume 4, Springer Science & Business Media, 2010. 12

    2010

  13. [13]

    G. A. Borelli, P. Maquet, Borelli’s on the Movement of Animals-on the Force of Percussion, Springer, 2015

    2015

  14. [14]

    Braun, Eadweard Muybridge, Reaktion Books, 2012

    M. Braun, Eadweard Muybridge, Reaktion Books, 2012

    2012

  15. [15]

    Ahmadi, F

    A. Ahmadi, F. Destelle, L. Unzueta, D. S. Monaghan, M. T. Linaza, K. Moran, N. E. O’Connor, 3d human gait reconstruction and monitoring using body-worn inertial sensors and kinematic modeling, IEEE Sensors Journal 16 (2016) 8823–8831

    2016

  16. [16]

    J. R. Hutchinson, The evolutionary biomechanics of locomotor function in giant land animals, Journal of Experimental Biology 224 (2021) jeb217463

    2021

  17. [17]

    W. Liu, C. Zhang, H. Ma, S. Li, Learning efficient spatial-temporal gait features with deep learning for human identification, Neuroinformatics 16 (2018) 457–471

    2018

  18. [18]

    J. Han, B. Bhanu, Individual recognition using gait energy image, IEEE transactions on pattern analysis and machine intelligence 28 (2005) 316–322

    2005

  19. [19]

    S. V. Purish, M. V. Lobachev, Gait recognition methods in the task of biometric human identifi- cation, Herald of Advanced Information Technology 6 (2023) 13–25

    2023

  20. [20]

    P. C. Cattin, Biometric authentication system using human gait, Ph.D. thesis, ETH Zurich, 2002

    2002

  21. [21]

    Parashar, A

    A. Parashar, A. Parashar, R. S. Shekhawat, A robust covariate-invariant gait recognition based on pose features, IET Biometrics 11 (2022) 601–613

    2022

  22. [22]

    H. S. Kühl, T. Burghardt, Animal biometrics: quantifying and detecting phenotypic appearance, Trends in ecology & evolution 28 (2013) 432–441

    2013

  23. [23]

    J. E. Boyd, J. J. Little, Biometric gait recognition, in: Advanced Studies in Biometrics: Summer School on Biometrics, Alghero, Italy, June 2-6, 2003. Revised Selected Lectures and Papers, Springer, 2005, pp. 19–42

    2003

  24. [24]

    Mathis, P

    A. Mathis, P. Mamidanna, K. M. Cury, T. Abe, V. N. Murthy, M. W. Mathis, M. Bethge, Deeplabcut: markerless pose estimation of user-defined body parts with deep learning, Nature neuroscience 21 (2018) 1281–1289

    2018

  25. [25]

    G. H. Martınez, Openpose: Whole-body pose estimation, Ph. D. dissertation (2019)

    2019

  26. [26]

    A. Raza, F. Hanif, H. A. Mohammed, Analyzing the enhancement of cnn-yolo and transformer based architectures for real-time animal detection in complex ecological environments, Scientific Reports 15 (2025) 39142

    2025

  27. [27]

    Q. Zeng, P. Liu, N. Yu, J. Wu, W. Huo, J. Han, Video-based quantification of gait impairments in parkinson’s disease using skeleton-silhouette fusion convolution network, IEEE Transactions on Neural Systems and Rehabilitation Engineering 31 (2023) 2912–2922

    2023

  28. [28]

    Arnab, M

    A. Arnab, M. Dehghani, G. Heigold, C. Sun, M. Lučić, C. Schmid, Vivit: A video vision trans- former, in: Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 6836–6846

    2021

  29. [29]

    H. Chao, Y. He, J. Zhang, J. Feng, Gaitset: Regarding gait as a set for cross-view gait recognition, in: Proceedings of the AAAI conference on artificial intelligence, volume 33, 2019, pp. 8126–8133

    2019

  30. [30]

    Y. Cui, Y. Kang, Multi-modal gait recognition via effective spatial-temporal feature fusion, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 17949–17957

    2023

  31. [31]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  32. [32]

    Makihara, M

    Y. Makihara, M. S. Nixon, Y. Yagi, Gait recognition: Databases, representations, and applica- tions, in: Computer Vision: A Reference Guide, Springer, 2021, pp. 1–13. 13

    2021

  33. [33]

    Sheth, M

    A. Sheth, M. Sharath, S. C. Reddy, K. Sindhu, Gait recognition using convolutional neural network., International Journal of Online & Biomedical Engineering 19 (2023)

    2023

  34. [34]

    Zhang, C

    X. Zhang, C. Xu, X. Tian, D. Tao, Graph edge convolutional neural networks for skeleton-based action recognition, IEEE transactions on neural networks and learning systems 31 (2019) 3047– 3060

    2019

  35. [35]

    Z. Cao, G. Hidalgo, T. Simon, S.-E. Wei, Y. Sheikh, Openpose: Realtime multi-person 2d pose estimation using part affinity fields, IEEE transactions on pattern analysis and machine intelligence 43 (2019) 172–186

    2019

  36. [36]

    Martinez, R

    J. Martinez, R. Hossain, J. Romero, J. J. Little, A simple yet effective baseline for 3d human pose estimation, in: Proceedings of the IEEE international conference on computer vision, 2017, pp. 2640–2649

    2017

  37. [37]

    Shrestha, J

    A. Shrestha, J. Le Kernec, F. Fioranelli, J. F. Marshall, L. Voute, Gait analysis of horses for lameness detection with radar sensors, in: International conference on radar systems (radar 2017), IET, 2017, p. 62

    2017

  38. [38]

    H. Gill, J. Charles, R. Grant, J. Gardiner, K. Bates, C. Brassey, Machine learning-based gait tracking in dogs: A markerless approach for controlled settings, Available at SSRN 5262411 (????). 14