💻 computer science

PlayClass: Automated Play Behaviour Classification in Poultry

The paper introduces PlayClass, an automated pipeline that combines long-duration tracking with foundation model embeddings to classify positive play behaviors in poultry from top-down video, achieving a 77.0 macro-averaged F1 score despite challenges like occlusion and kinematic similarities between play and non-play actions.

Original authors: Prince Ravi Leow (Section for Health Data Science & AI, University of Copenhagen), Neil Scheidwasser (Section for Health Data Science & AI, University of Copenhagen, Department of Infectious Disease E

Published 2026-05-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Prince Ravi Leow (Section for Health Data Science & AI, University of Copenhagen), Neil Scheidwasser (Section for Health Data Science & AI, University of Copenhagen, Department of Infectious Disease Epidemiology, Imperial College London), Rebecca Oscarsson (AVIAN Behaviour Genomics and Physiology Group, Linköping University), Per Jensen (AVIAN Behaviour Genomics and Physiology Group, Linköping University), Samir Bhatt (Section for Health Data Science & AI, University of Copenhagen, Department of Infectious Disease Epidemiology, Imperial College London), David Alejandro Duchêne (Section for Health Data Science & AI, University of Copenhagen)

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a farmer trying to keep your chickens happy and healthy. For a long time, scientists have been great at spotting when chickens are sick, hurt, or stressed (the "negative" signs). But they've mostly ignored the fun stuff—like when chickens are playing, frolicking, or chasing a worm. It's like having a security camera that only alerts you when someone breaks a window, but never tells you when the kids are having a great time in the backyard.

This paper, called PlayClass, is like building a new kind of "fun detector" for chickens. Here is how they did it, explained simply:

1. The Challenge: The "Chickens in a Crowd" Problem

Chickens look very similar to each other, and they move in big, dense groups. If you try to film them from above, they constantly bump into each other, hide behind one another, or get lost in the crowd.

The Analogy: Imagine trying to follow one specific friend in a crowded concert where everyone is wearing the same shirt. If your friend disappears behind someone for a second, you might lose track of them and start following a different person.
The Fix: The team built a smart tracking system. They used a "spotter" (a tool called YOLO) to find the best moments to reset their tracking, ensuring they didn't lose the identity of a specific chicken even when it got hidden for a while. They also used a powerful AI tool (SAM 3) to draw a digital outline around every chicken in every frame.

2. The Data: A Library of Chicken Play

They recorded 30 videos of 45 young chickens over 15 minutes each.

The Setup: They gave the chickens fake worms and live bugs to play with.
The Labels: A human expert watched the videos and marked down exactly what the chickens were doing:
- Playing: Running fast, jumping, flapping wings, or chasing a worm.
- Not Playing: Just walking, standing still, or pecking at the ground normally.
The Problem: Most of the time, the chickens were not playing. It's like a library where 87% of the books are boring textbooks, and only 13% are exciting adventure stories. The computer had to learn to find the rare "adventure stories" (play) without getting confused by the "textbooks" (normal behavior).

3. The Brain: Teaching the Computer to "See" Play

The researchers tried two different ways to teach the computer what "play" looks like:

Method A: The "Mathematical Detective" (Handcrafted Features)
Instead of showing the computer the whole video, they gave it a list of numbers describing the chicken's movement: How fast is it going? Is it turning sharply? How big is its shape?
- Result: This was surprisingly good! Just by looking at the math of the movement, the computer got about 73% correct. It proved that play has a very specific "dance" to it.
Method B: The "Art Student" (Foundation Models)
They used giant, pre-trained AI brains (like V-JEPA 2.1) that had already learned to understand videos from the internet. They asked these AI brains to look at the chicken videos and guess the behavior.
- Result: The "Art Student" (V-JEPA 2.1) was the best at understanding the visual story. It got better at spotting the subtle differences between "running for fun" and "running to escape."
The Super Combo: When they combined the "Mathematical Detective" (movement numbers) with the "Art Student" (visual understanding), they reached their highest score: 77% accuracy.

4. Where It Still Struggles

Even with this smart system, the computer still gets confused sometimes.

The "Look-Alike" Problem: Sometimes a chicken is running fast because it's playing, and other times it's running fast because it's scared. To the camera, they look exactly the same. The computer can't always tell the difference.
The "Hiding" Problem: When chickens pile on top of each other (occlusion), the computer loses track of who is who, leading to mistakes.

The Bottom Line

This paper shows that we can now automatically spot when chickens are having fun, not just when they are in trouble.

Key Takeaway: You don't need a super-complex AI to do this; simple movement math works almost as well as the most advanced AI. However, combining both gives the best results.
The Goal: This isn't about predicting the future or curing diseases yet. It's about proving that we can build a system to measure positive welfare (happiness/play) in farm animals, which is a huge step forward for understanding how animals feel.

Technical Summary: PlayClass – Automated Play Behaviour Classification in Poultry

Problem Statement

Automated animal welfare monitoring has historically focused on negative indicators (e.g., disease, pain), leaving positive welfare behaviors like play significantly underexplored. While computer vision has advanced animal pose estimation and tracking, applying these to poultry presents unique challenges: dense groups of visually similar individuals lead to frequent occlusions and identity confusion. Furthermore, play behavior is inherently difficult to classify automatically due to data scarcity, definitional ambiguity (play often shares motor patterns with non-play behaviors), and rapid state transitions. Existing work has not addressed automated play-behavior classification in poultry at the individual level.

Methodology

The authors present PlayClass, a pipeline designed to classify individual-level play behaviors from top-down poultry pen videos. The system integrates three core components:

1. Tracking Strategy (SAM 3 with Adaptive Chunking)

To handle 15-minute recordings where end-to-end tracking is limited by GPU memory and error accumulation:

Chunking: Videos are processed in 60-second chunks.
Initialization: Tracking is initialized with a text prompt ("bird") or an adaptive grounding strategy (running SAM 3 over the first 5 seconds to find a high-confidence, well-separated frame).
Identity Maintenance: To minimize identity switches across chunks, the system uses a YOLO-guided chunk boundary selection. A YOLO detector and BoT-SORT tracker identify frames with maximum inter-bird separation to serve as chunk boundaries, facilitating robust identity transfer.
Output: The pipeline generates over 1.8 million tracked masks with bounding boxes across 30 videos (45 individuals).

2. Feature Extraction

The system extracts two complementary feature streams per observation window:

Handcrafted Motion Features: 19 per-frame features derived from tracked masks, capturing spatial shape (area, solidity), temporal dynamics (velocity, acceleration), and social context (distances to others). These are summarized into a 171-dimensional vector using statistical moments and percentiles.
Foundation Model Embeddings: The pipeline evaluates frozen embeddings from image and video foundation models:
- Image Models: DINOv3 (extracting CLS-token embeddings from bounding box crops).
- Video Models: V-JEPA 2, V-JEPA 2.1, and VideoPrism (processing non-overlapping clips and spatially mean-pooling).

3. Classification Architecture

The task is a 3-class classification problem: Locomotor Play, Object Play, and Other (non-play).

Class Imbalance: The dataset is severely imbalanced (86.7% Other, 9.3% Object, 4.0% Locomotor).
Models: Two architectures were tested: a Multi-Layer Perceptron (MLP) on mean-pooled features and a 1D-CNN to preserve temporal signals.
Validation: A Leave-One-Cage-Out (LOCO) cross-validation strategy was employed to prevent environmental leakage, ensuring the model generalizes across different physical cages.

Key Results

The study was evaluated on a dataset of 30 recordings covering 45 Red Junglefowl × White Leghorn chicks, annotated into 14,515 windows.

Baseline Performance: Handcrafted motion features alone, processed via a simple MLP, achieved a 73.4 macro-averaged F1 score, establishing a strong baseline without learned visual representations.
Foundation Model Performance: Among frozen backbones, V-JEPA 2.1 consistently outperformed all other models (DINOv3, VideoPrism, V-JEPA 2) across both ViT-B and ViT-L scales.
Best Performance: The hybrid approach combining handcrafted features with V-JEPA 2.1 (ViT-L) embeddings yielded the highest performance: 77.0 macro-averaged F1.
Scaling: Increasing model size (e.g., DINOv3 to ViT-H) yielded minimal gains, suggesting representations saturate early for this specific domain.
Error Analysis:
- The model struggled most with Object Play (66.1% recall), often misclassifying it as "Other."
- Errors were driven by kinematic ambiguity (e.g., "worm pecking" shares motion profiles with non-play) and occlusion.
- Frolicking (a locomotor sub-type) formed a distinct cluster with high recall (88%), likely due to its unique multi-directional movement signature.
- Crowding significantly impacted object play accuracy (Spearman $\rho = 0.17$ ).

Significance and Claims

The paper claims to provide the first automated framework for individual-level play-behavior classification in poultry. Its primary contributions are:

A Robust Tracking Pipeline: Extending SAM 3 to long-duration recordings via YOLO-guided chunking and adaptive grounding, maintaining identity through frequent occlusions.
Benchmarking Foundation Models: Establishing V-JEPA 2.1 as the superior backbone for animal behavior classification, outperforming image-based and other video models.
Insight into Play Recognition: Demonstrating that while frozen video representations provide complementary signal, handcrafted kinematic and shape features alone constitute a strong, interpretable baseline (73.4 F1). This suggests that much of the task-relevant signal in poultry play is captured by motion and proximity statistics.

The authors conclude that while the pipeline shows encouraging evidence for automated play detection, challenges remain regarding occlusion and the definitional ambiguity of play sub-types. They note that current reliance on manual post-processing for identity correction limits fully autonomous deployment, suggesting future work should focus on domain-adaptive pretraining and improved cross-chunk identity matching.