Computer vision-based animal phenotyping and analysis in presence of uncertain identification

Animal identification (ID) is key for implementing precision livestock farming technologies. Animal ID algorithms typically generate a probability vector representing the likelihood of each potential individual. Conventionally, the individual with the highest probability is selected as the putative ID. However, this practice may reduce the precision of subsequent downstream analysis by disregarding the inherent uncertainty in the probability distribution. In this study, a mixture model is proposed to incorporate the uncertainty of the ID assignment into downstream analysis, aiming to investigate the impact of ignoring/incorporating the uncertainty of assignment in the subsequent estimations.

We applied our method on two datasets: 1) a publicly available dataset of 3226 images from 30 thoroughbred horses, classified based on body morphometrics using Linear Discriminant Analysis (LDA) with an accuracy of 88 %, where we simulated independent phenotypes with varying group effect sizes and variances, and 2) a dataset comprising 1770 images from 59 Holstein cattle, classified using Support Vector Machines (SVM) with an accuracy of 95 %, where phenotypes were extracted as measurements of body area from each image.

We analyzed the phenotypic data for both datasets to estimate group means and variance components using three approaches: 1) using the correct IDs, 2) using the top probability assignments, and 3) incorporating ID uncertainty through mixture models. Our results indicate that incorporating the mixture model improved the accuracy of variance component estimation and significantly increased the confidence of ID predictions across both datasets.