Résumé
Free dyadic dance is an ecological context in which people spontaneously synchronize and imitate each other. Yet, automatically measuring these mutual behavioral adjustments remains challenging. Here, we compare a data-driven approach with human perception for detecting motion Interpersonal coordination (IC) in dyadic dance. Using a motion capture dataset where visual and auditory coupling between dancers was experimentally manipulated, we formulated binary classification tasks contrasting coupled and uncoupled partners. We benchmarked machine learning models for multivariate time series classification. Then we compared their predictions with perceptual judgments from an online experiment (N = 87). Machine learning models are evaluated on two types of input data: raw velocities and features based on the synchrony of pairwise sensors across the two dancers. Results show that models’ performance improved significantly when using synchrony features over raw velocities. Humans reached scores similar to models using raw velocities(≃ 60% accuracy), with substantial inter-agreements (≃ 70%). Correlational analysis with variables derived from the mutual quantity of motions (QoM) of dancers revealed a human bias toward judging partners with similar QoM as coupled.