Validating Foundation Models for Automated Cattle Detection

Automated monitoring of cattle behavior through computer vision requires robust detection as a foundational step for tracking, re-identification, and behavior analysis. However, training accurate detection models typically demands extensive manually annotated datasets, creating a significant bottleneck for scaling cattle monitoring systems. In this work, we investigate whether automated annotation using foundation models can reliably replace manual labeling for cattle detection tasks. We introduce EMA (Extensive Mitrovac Annotations), a dataset of barn images with manually annotated cows with oriented bounding boxes including head orientation, posture labels (standing/lying), and visibility status (whole/partially visible). We systematically compare manualy annotated oriented bounding boxes with those generated by the Segment Anything Model 3 (SAM 3), demonstrating high agreement between automated and ground truth annotations. Furthermore, we train YOLO11-OBB detectors on both manual and SAM-generated annotations, showing that models trained on automated annotations achieve comparable performance to those trained on manual labels when evaluated on our ground truth test set. Our analysis reveals that only a small fraction of SAM-annotated data is sufficient to achieve high detection accuracy, proving the feasibility of automated annotation at scale. These findings suggest that foundation models show promise for generating training data in cattle detection pipelines under controlled conditions, potentially reducing annotation costs and supporting scalable deployment of monitoring systems. The EMA dataset and trained models are publicly available to support further research in precision livestock farming.