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

Graphical Abstract Summary: This study evaluated the accuracy of facial feature extraction in dairy calves using pose estimation models, with a focus on the effect of keypoint quantity. Three YOLOv8-based models were trained with 10, 16, and 30 facial keypoints. Model performance was assessed using various training metrics (loss, precision, recall, training time, and epochs) and keypoint prediction accuracy metrics, including object keypoint similarity, average precision, and average recall. Increasing the number of keypoints reduced prediction accuracy and overall model performance. The 10-keypoint model yielded the best results and enabled reliable extraction of relevant phenotypes, such as ear angles and movements, nostril distance, and eye opening.