With milk records and treatment histories increasingly supplemented with insights from genomic and sensor technologies, the future of dairy will be an incredibly data rich one. Yet a cow's conformation and structure, the cornerstones of most dairy evaluations, are still largely assessed by eye. The goal of this research was to establish the foundational algorithmic framework needed to extract objective and reproducible measures of anatomical traits from standard quality 2D images acquired in a typical production environment. The training data set consisted of facial images from 108 mature Holstein cows acquired on 3 separate days while restrained at the feedbunk. The standard human facial mesh was adapted to a bovine model, and key anatomical points extracted from each cow image using the MatLab GInput tool. Eye shape was then quantified using 2 schemes: normalized distances between all pairwise combination of points, and a novel geometric approach to biometric extraction that emphasized angles and projection lengths. Compared with the simple normalized distances, the geometric biometrics showed on average a 10% reduction in measurement error associated with human error in point annotation. Geometric measures also demonstrated less correlation among themselves than the normalized distance measures, a desirable trait for the development of stable linear predictive models. To assess the resilience of these metrics to variations in image quality, measurement error across days was regressed against measures designed to reflect changes in image scaling, frame position, and face angle. For the geometric measures, average total correlation between measurement error and measures of image quality was less than 5%, with no correlations exceeding 25%. For normalized distances, average total correlation between measurement error and measures of image quality with 15%, with multiple metric having over 50% of their measurement error attributed to subtle changes in image attributes. These results suggest that a geometric approach to anatomical biometrics could provide a more robust and consistent means of extracting detailed quantitative measures of physical traits from farm quality images.