A study on effective data collection methods to increace gesture varaibility

Abstract

Among various modalities, Body guesture, which is more intuitive and highly usable, is the main interaction tool between human and computer. Introducing these core interaction tools into the system requires techniques that can more accurately recognize gestures that fit the user's intentions. To obtain stable generalization performance of the model in a more practical environment, researchers have proposed various approaches, including building large-scale datasets, applying augmentation techniques, or utilizing various deep learning models. Among these methods, in this paper, i aim to verify the practical effectiveness of the styling word method, which can collect high-quality data by inducing the gesture variability in controlled environments. To confirm the practical effectiveness of these data collection methods, in this work, I intend to build various collection environments to collect data and verify the effectiveness of the styling word method. Furthermore, I quantitatively analyze the behavioral characteristics of gestures for each collection environment with a gesture feature analysis from Skeleton-based. In practice, an entity shall identify how data in accordance with the collection environment affects the performance of the recognizer.