Abstract
INTRODUCTION: Quantifying natural behavior from video recordings is a key component in ethological studies. Markerless pose estimation methods have provided an important step toward that goal by automatically inferring kinematic body keypoints. Such methodologies warrant efficient organization and interpretation of keypoints sequences into behavioral categories. Existing approaches for behavioral interpretation often overlook the importance of representative samples in learning behavioral classifiers. Consequently, they either require extensive human annotations to train a classifier or rely on a limited set of annotations, resulting in suboptimal performance. METHODS: In this work, we introduce a general toolset which reduces the required human annotations and is applicable to various animal species. In particular, we introduce OpenLabCluster, which clusters temporal keypoint segments into clusters in the latent space, and then employ an Active Learning (AL) approach that refines the clusters and classifies them into behavioral states. The AL approach selects representative examples of segments to be annotated such that the annotation informs clustering and classification of all temporal segments. With these methodologies, OpenLabCluster contributes to faster and more accurate organization of behavioral segments with only a sparse number of them being annotated. RESULTS: We demonstrate OpenLabCluster performance on four different datasets, which include different animal species exhibiting natural behaviors, and show that it boosts clustering and classification compared to existing methods, even when all segments have been annotated. DISCUSSION: OpenLabCluster has been developed as an open-source interactive graphic interface which includes all necessary functions to perform clustering and classification, informs the scientist of the outcomes in each step, and incorporates the choices made by the scientist in further steps.