Pith. sign in

REVIEW 3 major objections 1 minor 25 references

Horse Eye Blink Detection and Classification for Equine Affective State Assessment

T0 review · 3 major / 1 minor · reviewed 2026-06-28 · grok-4.3

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

Pith's one-line read Three computer vision methods detect and classify horse eye blinks at macro-F1 scores of 0.898 and 0.926 from video.

desk verdict Solid F1 numbers on public horse video data for blink detection, but the affective state claim rests on an unbacked assertion about what the blinks mean. read the letter →

arxiv 2606.05458 v1 pith:Z43MDGDE submitted 2026-06-03 cs.CV

João Alves , Signe M{o}ller-Skuldb{o}l , Pia Haubro Andersen , Rikke Gade This is my paper · ORCID
classification cs.CV
keywords horseblinkdetectionequinewelfarecomputervisionYOLOv12VideoMAEopticalflowfacialactionunitspainassessment
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 tests three approaches to automatically spot half and full eye blinks in horse videos, which serve as subtle signs of pain and stress. These movements are too fine-grained for easy human observation and require frame-by-frame review. The methods reach strong performance numbers on a public dataset, indicating that video analysis could support ongoing equine welfare checks. Results also note remaining difficulties in handling the fine details of these facial actions.

What carries the argument

Evaluation of YOLOv12, optical flow thresholding, and VideoMAE on subtle half and full eye blinks treated as facial action units for affective state monitoring.

What would settle it

Expert re-annotation of the same videos for affective state under controlled conditions that would show whether the automated scores align with actual welfare outcomes.

Watch

Extended reading notes

Core claim

A frame-based YOLOv12 detector, an optical flow magnitude thresholding approach, and a fine-tuned VideoMAE model were applied to horse videos, producing a macro-F1 score of 0.898 for classifying blinks and 0.926 for binary blink detection on a publicly available dataset.

Load-bearing premise

The public dataset reflects real-world horse video variation and the blink categories match established pain and stress indicators without extra validation.

Editorial extensions

If this is right

  • Automated detection reduces the need for manual frame-by-frame inspection in equine welfare monitoring.
  • Binary detection outperforms multi-class classification, indicating simpler tasks may be more immediately practical.
  • The results point to both feasibility and ongoing challenges for fine-grained facial action unit detection in horses.
  • Such methods could extend to continuous monitoring in veterinary or farm environments.

Reading between the lines

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

  • Combining blink detection with other facial action units could build a broader automated system for horse state assessment.
  • Real-time implementation on farm cameras would allow immediate alerts for potential pain or stress.
  • The same video techniques might adapt to detect similar subtle expressions in other large animals.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 / 1 minor

Summary. The manuscript develops and evaluates three methods for automated detection and classification of half and full eye blinks in horse videos (YOLOv12 frame-based detector, optical flow magnitude thresholding, and fine-tuned VideoMAE) on a publicly available dataset. It reports macro-F1 scores of 0.898 for blink classification and 0.926 for binary detection, positioning the work as a step toward automated equine affective state assessment via facial action units.

Significance. If the evaluation details and category validation were provided, the results could offer a practical contribution to fine-grained AU detection in animal videos. The work highlights challenges in micro-expression detection but its significance for welfare applications is limited by the unvalidated link between the specific blink categories and pain/stress indicators.

major comments (3)
  1. [Abstract] Abstract: the claim that half and full blinks are 'recognised indicators of pain and stress' is stated without any citations to equine welfare literature or validation studies, which is load-bearing for the paper's application to affective state assessment.
  2. [Methods] Methods/Results: the abstract reports F1 scores of 0.898 and 0.926 but supplies no dataset size, class distribution, train/test split, cross-validation procedure, or per-class error analysis, preventing assessment of whether the scores are supported by rigorous evaluation.
  3. [Introduction] Introduction/Results: the central application claim requires that the chosen blink categories align with established pain/stress indicators, yet the manuscript provides no independent validation or references for this alignment on the public dataset.
minor comments (1)
  1. [Abstract] Abstract: the three methods are listed but it is unclear which achieves the headline scores or how they compare in the reported results.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for their constructive comments. We address each major comment point by point below, indicating where revisions will be made.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that half and full blinks are 'recognised indicators of pain and stress' is stated without any citations to equine welfare literature or validation studies, which is load-bearing for the paper's application to affective state assessment.

    Authors: We agree that citations are required to support this claim. We will add relevant references from equine welfare literature in the revised abstract and introduction to substantiate the statement. revision: yes

  2. Referee: [Methods] Methods/Results: the abstract reports F1 scores of 0.898 and 0.926 but supplies no dataset size, class distribution, train/test split, cross-validation procedure, or per-class error analysis, preventing assessment of whether the scores are supported by rigorous evaluation.

    Authors: The full manuscript provides these details in the Methods and Results sections. To address the concern, we will revise the abstract to include key information on dataset size, split, and evaluation procedure while maintaining conciseness. revision: yes

  3. Referee: [Introduction] Introduction/Results: the central application claim requires that the chosen blink categories align with established pain/stress indicators, yet the manuscript provides no independent validation or references for this alignment on the public dataset.

    Authors: We will add references supporting the alignment of half and full blinks with established pain/stress indicators. Independent validation on the public dataset is not included, as the work focuses on detection methods rather than clinical validation. revision: partial

standing simulated objections not resolved
  • Independent validation of the blink categories as pain/stress indicators on the public dataset (beyond adding references), as this is outside the scope of the detection-focused study.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical results on external dataset

full rationale

The paper reports macro-F1 scores from three methods (YOLOv12, optical flow, VideoMAE) evaluated on a publicly available dataset. No equations, derivations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text. The central claims are direct performance measurements rather than reductions to inputs by construction. The assumption about blink categories aligning with affective states is an external validation gap, not a circularity issue per the defined patterns.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

Abstract-only review yields no explicit free parameters or invented entities; the central claim rests on standard computer-vision domain assumptions that the chosen models transfer to equine video data.

assumptions (1)
  • domain assumption Standard assumptions in computer vision for object detection and video analysis hold for equine videos.
    Implicit when applying YOLOv12 and VideoMAE without domain-specific adaptation details.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Horse Eye Blink Detection and Classification for Equine Affective State Assessment." pith.science (2026). https://pith.science/paper/Z43MDGDE

@misc{pith2026260605458,
  author       = {Pith},
  title        = {Pith review of: Horse Eye Blink Detection and Classification for Equine Affective State Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z43MDGDE}},
  note         = {Machine review of arXiv:2606.05458}
}
read the original abstract

Automated detection of equine facial action units (AUs) is a promising yet under-explored avenue for pain and affective state assessment in horses. Half and full-blink movements are recognised indicators of pain and stress, but as micro-expressions, their subtle, fine-grained nature makes them easily missed by the naked eye and only discernible through frame-by-frame video inspection, making reliable automated detection from video a particularly demanding task. We develop and evaluate three methods for automated blink classification from horse videos: a frame-based YOLOv12 detector, an optical flow magnitude thresholding approach, and a fine-tuned VideoMAE model, tested on a publicly available dataset. We achieve a macro-F1 score of 0.898 when doing blink classification and 0.926 on binary blink detection. Our results highlight both the potential and the inherent challenges of fine-grained AU detection for equine welfare monitoring.

Figures

Figures reproduced from arXiv: 2606.05458 by the authors.

Figure 1
Figure 1. Example eye blink movements [6] data: full-blink (left) and half-blink (right). for affective state assessment. We propose and evaluate multiple methodologies on a publicly available dataset [6]. Our contributions are as follows: • A human baseline study demonstrating the perceptual dif￾ficulty of equine full- and half-blink annotation, show￾ing that even naive human observers struggle with half￾blink classification… view at source ↗
Figure 2
Figure 2. Confusion matrices for test set evaluation for each [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Inter-annotator study data class distribution with 22 total [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (8 more)
Figure 5
Figure 5. Figure 5: F1 score comparison naive human annotators and the [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Optical flow based method on example full-length S5 video showing how flow magnitude contributes for blink detection. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Example qualitative results on original S5 video for (a) [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: Sample frames for each of the 12 videos [ [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Qualitative video timelines (S1-S12) for Optical Flow based eyelid closure detection method. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Qualitative video timelines (S1-S12) for Optical Flow based eyelid closure detection method. [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: Qualitative video timelines (S1-S12) for YOLO based eye blink classification method. [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Qualitative video timelines (S1-S12) for VideoMAE based eye blink classification method following a sliding approach with a [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references

  1. [1]

    Emanuela Dalla Costa, Michela Minero, Dirk Lebelt, Diana Stucke, Elisabetta Canali, and Matthew C. Leach. Develop- ment of the Horse Grimace Scale (HGS) as a Pain Assess- ment Tool in Horses Undergoing Routine Castration.PLoS ONE, 9(3):e92281, March 2014. 1

    2014

  2. [2]

    Burrows, Bridget M

    Jen Wathan, Anne M. Burrows, Bridget M. Waller, and Karen McComb. EquiFACS: The Equine Facial Action Cod- ing System.PLOS ONE, 10(8):e0131738, August 2015. 1

    2015

  3. [3]

    Ekman, W

    P. Ekman, W. V . Friesen, and J. C. Hager. Facial action cod- ing system (FACS).Research Nexus, 2002. 1

    2002

  4. [4]

    Going Deeper than Tracking: A Sur- vey of Computer-Vision Based Recognition of Animal Pain and Emotions.International Journal of Computer Vision, 131(2):572–590, February 2023

    Sofia Broom ´e, Marcelo Feighelstein, Anna Zamansky, Gabriel Carreira Lencioni, Pia Haubro Andersen, Francisca Pessanha, Marwa Mahmoud, Hedvig Kjellstr ¨om, and Al- bert Ali Salah. Going Deeper than Tracking: A Sur- vey of Computer-Vision Based Recognition of Animal Pain and Emotions.International Journal of Computer Vision, 131(2):572–590, February 2023. 1

    2023

  5. [5]

    Towards Machine Recognition of Facial Expressions of Pain in Horses.Ani- mals, 11(6):1643, June 2021

    Pia Haubro Andersen, Sofia Broom ´e, Maheen Rashid, Jo- han Lundblad, Katrina Ask, Zhenghong Li, Elin Hernlund, Marie Rhodin, and Hedvig Kjellstr ¨om. Towards Machine Recognition of Facial Expressions of Pain in Horses.Ani- mals, 11(6):1643, June 2021. 1, 2

    2021

  6. [6]

    Equine Facial Action Coding System for determination of pain-related facial responses in videos of horses.PLOS ONE, 15(11):e0231608, November

    Maheen Rashid, Alina Silventoinen, Karina Bech Gleerup, and Pia Haubro Andersen. Equine Facial Action Coding System for determination of pain-related facial responses in videos of horses.PLOS ONE, 15(11):e0231608, November

  7. [7]

    Swedish University of Agricultural Sciences, 2024

    Johan Lundblad.Exploring Facial Expressions in Horses : Biological and Methodological Approaches, volume 2024:38 ofActa Universitatis Agriculturae Sueciae. Swedish University of Agricultural Sciences, 2024. 1

    2024

  8. [8]

    An equine pain face.Veterinary Anaes- thesia and Analgesia, 42(1):103–114, January 2015

    Karina B Gleerup, Bj ¨orn Forkman, Casper Lindegaard, and Pia H Andersen. An equine pain face.Veterinary Anaes- thesia and Analgesia, 42(1):103–114, January 2015. 1, 2, 4

    2015

Show all 25 references
  1. [9]

    Changes in the equine facial repertoire during different orthopedic pain intensities

    Katrina Ask, Marie Rhodin, Maheen Rashid-Engstr ¨om, Elin Hernlund, and Pia Haubro Andersen. Changes in the equine facial repertoire during different orthopedic pain intensities. Scientific Reports, 14(1):129, January 2024. 1

    2024

  2. [10]

    Automated Detection of Equine Facial Action Units, 2021

    Zhenghong Li, Sofia Broom ´e, Pia Haubro Andersen, and Hedvig Kjellstr ¨om. Automated Detection of Equine Facial Action Units, 2021. 2

    2021

  3. [11]

    Read my ears! horse ear movement detection for equine affective state assessment

    Jo ˜ao Moreira Alves, Pia Haubro Andersen, and Rikke Gade. Read my ears! horse ear movement detection for equine affective state assessment. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). 2

    2025

  4. [12]

    Auto- matic horse blink detection using computer vision and deep nets

    Stefani Dimitrova, Emily Orchard, Bishnu Paudel, Sebastian McBride, Andrew Hemmings, and Otar Akanyeti. Auto- matic horse blink detection using computer vision and deep nets. In George Panoutsos, Mahdi Mahfouf, and Lyud- mila S Mihaylova, editors,Advances in Computational In- ...

    2024

  5. [13]

    Eye-LRCN: A long-term recurrent con- volutional network for eye blink completeness detection

    Gonzalo De La Cruz, Madalena Lira, Oscar Luaces, and Beatriz Remeseiro. Eye-LRCN: A long-term recurrent con- volutional network for eye blink completeness detection. IEEE Transactions on Neural Networks and Learning Sys- tems, 35(4):5130–5140, 2024. 2

    2024

  6. [14]

    Blin- kLinMulT: Transformer-based eye blink detection.Journal of Imaging, 9(10):196, 2023

    ´Ad´am Fodor, Kristian Fenech, and Andr ´as L ˝orincz. Blin- kLinMulT: Transformer-based eye blink detection.Journal of Imaging, 9(10):196, 2023. 2

    2023

  7. [15]

    Robust eye blink detection using dual embedding video vi- sion transformer

    Jeongmin Hong, Joseph Shin, Juhee Choi, and Minsam Ko. Robust eye blink detection using dual embedding video vi- sion transformer. In2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6362–6372. IEEE, 2024. 2

    2024

  8. [16]

    Eyes closeness detection from still images with multi-scale histograms of principal oriented gradients.Pattern Recogni- tion, 47(9):2825–2838, 2014

    Fengyi Song, Xiaoyang Tan, Xue Liu, and Songcan Chen. Eyes closeness detection from still images with multi-scale histograms of principal oriented gradients.Pattern Recogni- tion, 47(9):2825–2838, 2014. 2

    2014

  9. [17]

    Eyeblink-based anti-spoofing in face recognition from a generic webcamera

    Gang Pan, Lin Sun, Zhaohui Wu, and Shihong Lao. Eyeblink-based anti-spoofing in face recognition from a generic webcamera. In2007 IEEE 11th International Con- ference on Computer Vision, pages 1–8. IEEE, 2007. 2

    2007

  10. [18]

    Pupil localization using geodesic dis- tance

    Radovan Fusek. Pupil localization using geodesic dis- tance. In George Bebis, Richard Boyle, Bahram Parvin, Darko Koracin, Matt Turek, Srikumar Ramalingam, Kai Xu, Stephen Lin, Bilal Alsallakh, Jing Yang, Eduardo Cuervo, and Jonathan Ventura, editors,Advances in Visual Comput-...

    2018

  11. [19]

    RT- BENE: A dataset and baselines for real-time blink estimation in natural environments

    Kevin Cortacero, Tobias Fischer, and Yiannis Demiris. RT- BENE: A dataset and baselines for real-time blink estimation in natural environments. In2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 1159–1168. IEEE, 2019. 2

    2019

  12. [20]

    Eye blink complete- ness detection.Computer Vision and Image Understanding, 176-177:78–85, 2018

    Andrej Fogelton and Wanda Benesova. Eye blink complete- ness detection.Computer Vision and Image Understanding, 176-177:78–85, 2018. 2

    2018

  13. [21]

    Two-Frame Motion Estimation Based on Polynomial Expansion

    Gunnar Farneb ¨ack. Two-Frame Motion Estimation Based on Polynomial Expansion. In Gerhard Goos, Juris Hartmanis, Jan Van Leeuwen, Josef Bigun, and Tomas Gustavsson, edi- tors,Image Analysis, volume 2749, pages 363–370. Springer Berlin Heidelberg, Berlin, Heidelberg, 2003. 2

    2003

  14. [22]

    Horse Eye Blink Detection and Classification for Equine Affective State Assessment

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders are Data-Efficient Learn- ers for Self-Supervised Video Pre-Training.NeurIPS, 2022. 3 Horse Eye Blink Detection and Classification for Equine Affective State Assessment Supplementary Material This ...

    2022

  15. [23]

    Dataset subjects sample data To provide further insight into the dataset used, we provide sample frames of each of the 12 test videos from [6] in this supplementary material (see Figure 8)

  16. [24]

    Human baseline study Additional results of the human baseline study are presented in Table 3

  17. [25]

    Figure 8

    Supplementary qualitative results In this section we show the qualitative results obtained from using each method on the original full length dataset videos (see Figures 10, 11 and 12). Figure 8. Sample frames for each of the 12 videos [8] in the test dataset in row-major orde...

Pith tools

Reviewed June 28, 2026 · model on record in the stance chip above.