Machine Learning Approaches to Predicting Calving Difficulty from Skeletal Key Points

Abstract

We have used the Convolutional Pose Machines (CPM) model, designed to estimate joint positions in the human body, to estimate the lower limb and pelvis key points in cattle. Here we examined whether the differences in distance and angle between key points could be used to predict calving difficulty. Skeletal key points can be acquired by both manual annotation and inferred by CPM. Annotated points are considered to be correct. Inferred points are estimated from images and contain errors. We compared the accuracy of judgement of the degree of calving difficulty by machine learning based on both annotated and estimated key points, using the distance and angle between key points as explanatory variables. Side view images had 20 explanatory variables and back view images had 8. Of 5 machine learning models (support vector machine, random forest, linear discriminant analysis, logistic regression, and decision tree) tested, the support vector machine trained and tested with annotated key points had the highest accuracy: 0.79 for the side view images and 0.77 for the back view images. However, when it used the estimated key points, the accuracy was lower, at 0.52 and 0.53, respectively. Our work shows that with correct skeletal key points, calving difficulty can be accurately predicted.