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

Area of Science:

• Animal Science
• Artificial Intelligence
• Computer Vision

Background:

• Facial expressions in livestock offer insights into animal health and welfare.
• Artificial intelligence (AI) tools are increasingly used in livestock farming for monitoring.
• Automated analysis of facial features can provide objective welfare indicators.

Purpose of the Study:

• To evaluate the accuracy of pose estimation models for extracting facial features in North American Holstein calves.
• To compare the performance of YOLOv8-based models with varying numbers of facial keypoints.
• To determine the suitability of these models for automated phenotyping and welfare monitoring.

Main Methods:

• Trained three YOLOv8-based pose estimation models using 10, 16, and 30 facial keypoints.
• Evaluated model performance using training metrics (loss, precision, recall, F1 score) and keypoint accuracy (OKSI, AP, AR).
• Assessed the models' ability to extract specific facial phenotypes like nostril distance, eye opening, and ear angles.

Main Results:

• The 10-keypoint model demonstrated superior performance compared to models with 16 and 30 keypoints.
• The 10-keypoint model required less training time and achieved lower loss.
• This model provided the highest prediction accuracy, enabling reliable extraction of facial phenotypes.

Conclusions:

• Pose estimation models, particularly the 10-keypoint YOLOv8 variant, are accurate for analyzing dairy calf facial features.
• Automated phenotyping and welfare monitoring in dairy calves can be achieved using these AI-driven approaches.
• This technology holds significant potential for improving animal welfare assessment in livestock farming.