Facial phenotyping of Holstein calves using pose estimation models with varying keypoints

Artificial intelligence is increasingly applied in livestock farming and offers valuable tools for monitoring animal health and welfare. Among these, facial expressions have emerged as informative indicators of well-being in animals. In this context, the main objective of this study was to evaluate the accuracy of pose estimation models for extracting facial features in North American Holstein calves. Three YOLOv8-based models were trained using 10, 16, and 30 facial keypoints, and their performance was evaluated using training metrics (loss, precision, recall, F1 score, training time, and number of epochs) and keypoint accuracy measures (object keypoint similarity, average precision, average recall). The 10-keypoint model outperformed the others, requiring less training time, achieving lower loss, and yielding the highest prediction accuracy. This model enabled reliable extraction of facial phenotypes, including nostril distance, eye opening, and ear angles and movements. These findings demonstrate the potential of pose estimation models for automated phenotyping and welfare monitoring in dairy calves.