PlayClass: Automated Play Behaviour Classification in Poultry
Prince Ravi Leow, Neil Scheidwasser, Rebecca Oscarsson, Per Jensen, Samir Bhatt, David Alejandro Duchêne, [6pt] Section for Health Data Science & AI, University of Copenhagen · 2026
Read PlayClass: Automated Play Behaviour Classification in Poultry by Prince Ravi Leow, Neil Scheidwasser, Rebecca Oscarsson, Per Jensen, Samir Bhatt, David Alejandro Duchêne, [6pt] Section for Health Data Science & AI, University of Copenhagen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-27304.txt.
Introduction
Prince Ravi Leow and colleagues present PlayClass, a pipeline that classifies play behavior in poultry from overhead pen video. It tracks birds with long-duration tracking and classifies actions from frozen image and video embeddings. The best backbone reaches 77 percent macro-averaged F1, though similar motion profiles and birds blocking each other keep the dataset hard.
Catalog record
| CREATOR | Prince Ravi Leow, Neil Scheidwasser, Rebecca Oscarsson, Per Jensen, Samir Bhatt, David Alejandro Duchêne, [6pt] Section for Health Data Science & AI, University of Copenhagen |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 23121 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.27304
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
PlayClass: Automated Play Behaviour Classification in Poultry Prince Ravi Leow, Neil Scheidwasser, Rebecca Oscarsson, Per Jensen, Samir Bhatt, David Alejandro Duchêne, [6pt] Section for Health Data Science & AI, University of Copenhagen Abstract Automated monitoring of animal welfare has largely targeted negative indicators, leaving positive welfare behaviours such as play underexplored. To address this gap, we present PlayClass , a pipeline for play-behaviour classification in poultry from top-down pen video. The pipeline leverages long-duration tracking with SAM 3 via YOLO-guided chunk boundaries to minimise identity errors in point-based prompting, and frozen embeddings from image and video foundation models for play action classification. Although handcrafted motion features from tracked masks alone achieved competitive accuracy, V-JEPA 2.1 consistently outperformed all other backbones across model scales, reaching 77.0 macro-averaged F 1 when combined with handcrafted features. Despite this result, the dataset remains challenging due to play sub-types sharing similar kinematic profiles with non-play and inter-bird occlusion. Overall, our work provides encouraging evidence towards automated frameworks for play behaviour classification in poultry. † † footnotetext: ∗ Equal contribution. † Supervising author. 1 Introduction Precise measurement of animal behaviour is a fundamental problem in many areas in biology, including understanding neural mechanisms 2 and monitoring welfare in agricultural 10 and conservation 36 settings. Recent advances in computer vision have played a central role in this endeavour. Transfer learning approaches have greatly advanced the automation of animal pose estimation and tracking 27 . Vision foundation models 8 ; 37 ; 42 ; 3 pretrained on large, diverse datasets extend these capabilities with general-purpose representations that can be adapted effectively to specialised domains with limited labelled data. In agriculture, poultry constitutes an important case study. From a public health perspective, poultry farming is a cornerstone of food security and a major target of zoonosis surveillance 30 . From a technical perspective, poultry settings constitute a challenging vision domain, as they typically feature visually similar individuals in dense groups, which may lead to occlusions and identity confusion. Within poultry behaviour analysis, health and welfare assessment has been a dominant focus 12 . A substantial body of work has leveraged optical flow methods to measure flock-level patterns of activity, impaired gait, or infection in commercial settings 11 ; 9 ; 15 . Conversely, methods leveraging deep learning-based tracking are emerging for individual-level recognition of health-related markers (e.g., posture, feeding, sneezing) in experimental settings 7 ; 21 . However, these efforts have largely targeted negative health and welfare indicators such as pain or disease 41 , leaving positive behaviours such as play 6 ; 17 ; 38 relatively underexplored. Play is a ubiquitous behaviour among animals, in spite of being energetically costly and physically risky 16 . Despite its biological and welfare significance, automatic play assessment is hindered by data scarcity and by definitional ambiguity, as play often shares motor patterns with other behaviours and can transition rapidly between states 6 . In this work, we present PlayClass ‡ † † footnotetext: ‡ https://github.com/sbhattlab/PlayClassCV4Animals (Fig. 1 ), a pipeline for individual-level play-behaviour classification in poultry from top-down video, and evaluate it on a dataset of 30 recordings covering 45 tracked individuals in a severely imbalanced regime. Our contributions are: 1. A tracking strategy that extends SAM 3 8 to 15-minute recordings via YOLO-guided chunk-boundary selection and cross-chunk identity matching, maintaining individual identity through frequent occlusions. 2. A benchmark of image (DINOv3 37 ) and video (V-JEPA 2/2.1 3 ; 29 , VideoPrism 42 ) foundation models against a competitive baseline of handcrafted motion descriptors, where V-JEPA 2.1 emerged as the strongest backbone. 3. An error analysis showing that misclassification is driven by dominant kinematic profiles within behaviour groups as well as occlusion scenarios. Figure 1 : Overview of the PlayClass pipeline for play action classification in poultry, combining handcrafted motion features and frozen foundation-model embeddings extracted from 1.8 million SAM 3-tracked masks across 30 top-down video recordings. 2 Related Work Open-source pose estimation tools such as DeepLabCut 26 ; 23 have become a common starting point for behavioural analysis. Although pretrained pose models can generalise across dozens of species without manual labels 40 , many scenarios still require de novo keypoint labelling. As a label-light alternative, DINOv2 31 image embeddings from tracking masks have supported action classification in sheep 39 , while domain-adaptive pretraining of the V-JEPA 4 video foundation model has shown strong results for primate behaviour classification 28 . V-JEPA 2.1 29 extends the latter with a dense predictive objective that grounds representations in fine-grained spatiotemporal structure, but has not yet been evaluated for animal behaviour. For tracking, SAM 3 8 enables unprecedented promptable video segmentation, but remains difficult to apply to long clips ( > 1 min) due to memory constraints and error accumulation 14 . Poultry-specific video data remains scarce, due in part to privacy and ownership constraints in agricultural settings. Recent efforts are beginning to address this gap, with datasets covering behaviour in general scenarios 21 ; 7 and under infection 13 ; 35 . Play behaviour video has been collected 5 ; 32 but without automated classification. To our knowledge, no prior work addresses automated play-behaviour classification in poultry. 3 Methods 3.1 Data From a controlled play study 32 on 45 Red Junglefowl \times White Leghorn chicks (housed in five cages of nine), we used the 30 top-down video recordings ( 704{\times}576 , 25 fps, 15 min each) collected on days 28 and 29 of rearing. On each recording day, cages were split into three focal groups of three, and each group was filmed separately in a dedicated arena provisioned with a fake worm and live mealworms to elicit object play. The groups of three differed across days. A trained observer scored play events per bird using 1/0 sampling in 5 s windows according to an ethogram designed across two natural granularity levels: three play-behaviour categories (locomotor play, object play, social play) and 14 fine-grained sub-behaviours, detailed in Table 1 . We classify at the category level because the fine-grained scheme has severely uneven support (0 to 12,585 windows), with several sub-types too sparse to evaluate reliably under LOCO. Social play was excluded due to insufficient samples (201 windows, 1.4%). The resulting dataset comprises 14,515 windows across 45 individuals, with severe class imbalance: 86.7% other (non-play), 9.3% object, 4.0% locomotor. Table 1 : Ethogram of scored poultry play behaviours. N : number of labelled windows. Number of no-play windows: 12,585. Behaviour Definition N Locomotor play 585 Frolicking Spontaneous and rapid running and/or jumping while wings either flapping or raised, often including rapid direction changes. 531 Wing flapping Rapid vertical movement with both wings while stationary or walking up to 2 steps. Excludes wing flaps performed by a bird to balance itself or correct its feathers. 25 Running Spontaneous forward movement at \geq 2\times normal walking pace, often including rapid direction changes. Can start either while walking or being stationary. Wing flapping does not occur simultaneously. 22 Spinning Circling movement around the bird’s own axis with \geq 2\times the normal walking pace. Wing flapping does not occur simultaneously. 0 Spinning while wing flapping As above, but wing flapping occurs simultaneously. 7 Object play 1345 Worm pecking Bird pecks at fake worm or mealworm on the