Introduction
Monitoring animal welfare in commercial poultry houses is a critical but challenging task. Precise localization of broilers is necessary to identify issues like "wet litter" or to track activity levels. However, these environments are characterized by extreme crowding and frequent occlusions from feeders and drinkers. While multi-view camera systems offer a solution by providing multiple perspectives, they introduce a significant data bottleneck: manually annotating ground-plane coordinates for thousands of moving birds across synchronized feeds is prohibitively labor-intensive. This paper presents a label-efficient framework for training high-performance multi-view detectors using pseudo-labels, eliminating the need for manual ground-truth annotations.
Methodology: From Geometric to Learned Fusion
The core of the proposed approach is a transition from simple geometric projection to a sophisticated end-to-end learned model.
Pseudo-Label Generation (Geometric-Fusion)
The process begins with "Geometric-Fusion." A standard YOLOv8 detector identifies broilers in individual 2D camera views. These detections are then projected onto a 3D ground plane using camera calibration data. A clustering algorithm aggregates these points to generate initial, albeit noisy, ground-plane "pseudo-labels."
End-to-End Training (Learned-Fusion)
These pseudo-labels serve as the training signal for a "Learned-Fusion" model based on the MVDet architecture. Unlike the geometric approach, which processes views independently before fusion, the learned model takes raw feature maps from all cameras simultaneously. This allows the network to learn how to aggregate multi-view evidence, effectively filtering out false positives and correcting for minor calibration errors.
Age-Specific Adaptation
Broilers undergo rapid physical growth over their six-week lifespan. To account for this, the researchers introduced age-specific Gaussian kernels. During label generation, the size of the "occupancy blob" for each bird is adjusted based on its growth stage (Starter, Grower, or Finisher). Furthermore, the Non-Maximum Suppression (NMS) distance threshold used in post-processing adapts linearly with the birds' age to prevent redundant detections.
Experimental Setup and Dataset
The framework was evaluated using an extensive real-world dataset comprising approximately 800 hours of video from four synchronized cameras in a commercial broiler house. The data spans a full growth cycle and is categorized into three stages:
- Starter: <10 days
- Grower: 10-23 days
- Finisher: >23 days
For evaluation purposes only, a subset of the data was manually annotated (4,739 birds for validation, 1,967 for testing).
Key Results and Insights
The primary performance metric, Multi-Object Detection Accuracy (MODA), revealed that the learned-fusion model consistently outperformed its "teacher."
- Superior Accuracy: The learned-fusion model achieved a peak MODA of 82.0%, surpassing the 80.9% achieved by the geometric-fusion baseline.
- Robustness to Noise: The improvement was most pronounced when using smaller, noisier initial detectors (e.g., YOLOv8S), where the learned model acted as a powerful filter to clean up the detection signal.
- Growth Stage Performance: The system performed best during the Finisher stage (83.9% MODA), where larger birds are more distinct. Performance in the Starter stage was lower due to the difficulty of resolving dense clusters of small chicks at the available image resolution.
Conclusion
This research demonstrates that high-quality multi-view object detection can be achieved without expensive manual ground-plane labels. By leveraging noisy pseudo-labels and incorporating domain-specific knowledge like age-based growth patterns, the "Learned-Fusion" approach provides a scalable solution for lifelong monitoring in precision livestock farming. Future work will explore continuous model adaptation and more advanced post-processing techniques to further enhance detection in extremely crowded conditions.
Summary updated