MAGIC: A Dataset Capturing Mid-Air Gesture Performance for Interaction and Feature Analysis

Mid-air gestures are increasingly used as a natural modality in interactive systems, ranging from smart home controls and automotive interfaces to augmented reality and public displays. Despite their growing adoption across diverse domains, we still lack a comprehensive understanding of what makes gestures effective, expressive, and robust across participants and application domains. We conducted a data collection study in which we captured a novel dataset of 18 widely used mid-air gestures, each performed by 42 participants. A total of 4,265 gesture samples were recorded using synchronized electromagnetic sensors and depth cameras. We analyzed gestures using kinematic and geometric measures, and quantified consistency in terms of hand shape and palm trajectory using Multidimensional Dynamic Time Warping. We further trained LSTM-based classifiers as recognition baselines and computed a deep-feature consistency measure from their learned embedding space. Results revealed that consistency and recognizability vary substantially across gesture types, and consistency scores were positively associated with recognition performance. These findings provide empirical guidance for the design and selection of mid-air gesture vocabularies in real-world interaction systems.