Improving the Annotation Efficiency for Animal Activity Recognition using Active Learning

Animal activity recognition (AAR) is essential for the conservation of endangered species and the well-being of livestock. Small resource-constraint devices using inertial measurement units can be attached to animals to automatically classify the performed activities such as running or eating based on their bodily motions. The IMU time-series data is enriched with ground-truth annotations by experts to train artificial intelligence models as classifiers. Annotating AAR data is tedious and time-consuming. Experts that can provide high-quality annotations are scarce and their time is costly. It is difficult for annotators to determine what data is essential to annotate, thus they tend to annotate large amounts of data to improve the performance of the trained classifiers. In this paper, we show that active learning (AL) increases annotation efficiency by selecting only the most informative data for annotation and algorithm training. We use real-world IMU data from four horses performing various activities. We compare five AL-based algorithms (three uncertainty sampling and two disagreement-based sampling (DBS)) with random sampling. Our results show that DBS increases the annotation efficiency, especially when the AAR problem is harder by increasing the number of classified activities from 6 to 8. Furthermore, we show that random sampling can be an effective method to improve annotation efficiency when the AAR task is not too difficult.