Using Machine Learning to Collect Data About Dogs’ Head Orientation in Experimental Settings

Markerless pose estimation, a key advancement in the field of ethology, has revolutionized the study of animal behavior by overcoming the inherent limitations of traditional methods. This thesis examines the methodological transition from manual observation and commercial tracking software to the utilization of deep learning frameworks, with a specific focus on the capabilities of DeepLabCut (DLC). While manual methods are prone to subjectivity and are resource-intensive, and commercial systems like EthoVision are constrained by their lack of customizability and adaptability, DLC presents a robust, open-source alternative. By leveraging convolutional neural networks and the principles of transfer learning, DLC facilitates high-precision tracking of user-defined body parts, requiring only a minimal number of training samples. To demonstrate its efficacy, here details a case study focused on training a DLC model to detect dogs head orientation from a top-down perspective. Through three distinct trials, we systematically increased the dataset's size and diversity, including multiple dog breeds and coat colors. The results confirm that dataset volume and variability are critical for achieving a robust and generalizable model. The findings highlight DLC's potential to provide repeatable, high-resolution data on subtle behaviors, which are difficult to quantify with other methods. This work underscores the importance of integrating advanced computational tools into ethological research and suggests a future of hybrid approaches where automated systems augment, rather than replace, human observation.