87. Comparisons of Ultralytics YOLO Models for Individual Cow Identification and Estrus Detection in Dairy Cattle

Abstract Computer vision has the potential to improve reproductive management in dairy cattle by providing a cost-effective alternative to visual observation and wearable biosensors for estrus detection. Despite widespread use of You Only Look Once (YOLO) architectures for object detection, limited research has compared the performance of Ultralytics-compatible YOLO models for livestock applications. This study evaluated six models (YOLOv5su, YOLOv8s, YOLOv9s, YOLOv10s, YOLOv11s, and YOLOv12s) for estrus behavior detection and individual cow identification. Thirty lactating Holstein-Friesian cows housed in a single pen within a freestall barn were observed across three 10-day trial periods (n = 10 cows/trial). On day 2 of each trial, cows were synchronized using a single intramuscular injection of Lutalyse® (Dinoprost tromethamine, 5 mg/mL; Zoetis Inc.). Continuous 1080p video footage was collected via four corner-mounted cameras, resulting in a total of 2,880 hours of observation. A 95-hour subset containing observable estrus behaviors (classified as chin resting and mounting behaviors) were extracted and segmented into 135,850 frames. The resulting frames were manually annotated to generate four datasets: estrus detection (3,747 frames) and three individual identification datasets corresponding to each trial (1,229, 976, and 937 frames, respectively). Each dataset was subsequently partitioned into training and validation subsets at an 80:20 ratio. All six model architectures were trained on each dataset, resulting in 24 trained models. Training was conducted for 100 epochs with a batch size of 16 at 640 × 640 p resolution. Models were trained on a Windows 11 workstation equipped with an Intel® Core™ i7-13850HX CPU, 32 GB of RAM, and an NVIDIA RTX™ 2000 Ada GPU with 8 GB of VRAM. Model performance was evaluated in R (version 4.4.1), with statistical significance set at p < 0.05. Evaluation metrics included total training and validation loss, image processing time, and plateau epoch (epoch at which mAP50 ceased to improve) as well as standard performance indicators: precision, recall, mAP50, mAP50-95, and F1-score. Results indicate that model architecture significantly influenced training and validation loss (p < 0.01). YOLOv10s exhibited higher loss values than all other models (p < 0.01) suggesting less stable convergence and weaker localization performance. However, no significant differences were observed for precision, recall, mAP50, mAP50-95, F1-score, or plateau epoch. Image processing time did not differ across models for preprocessing, inference time, or total processing time. However, post processing time varied (p < 0.01), with YOLOv10 performing faster than YOLOv5su and YOLOv8s. These findings indicate that the tested Ultralytics-compatible YOLO models achieved comparable performance in estrus detection and individual animal identification tasks. While YOLOv10s demonstrated higher training loss, overall detection accuracy remained consistent, supporting the applicability of these models for automated estrus detection and animal identification in dairy cattle.