Few-Shot Learning Approach for Avatar Action Identification in the Metaverse

Avatars are vital in the metaverse as they represent users and their interactions within the digital environment. This research delves into understanding and identifying avatar actions in the metaverse to enhance virtual interactions for increased safety and contextual awareness. While avatar actions share similarities with human movements, they also exhibit distinct patterns and levels of freedom. As a result, the study focuses on transferring knowledge from human actions to comprehend avatars' behavior in the metaverse. To improve the interpretation of avatar movements, a model pre-trained on human actions is trained using a small amount of avatar action data. The study compares this “few-shot” model with a “zero-shot” model (which doesn't use any avatar data) and a model exclusively trained on avatar data. The analysis employs the Temporal Convolutional Networks (TCN) model for handling avatar skeleton data and an MLP (Multi-Layer Perceptron) model for classifying avatar actions. Ultimately, the study finds that the few-shot model surpasses the zero-shot model in terms of effectiveness and adaptability to avatars. Moreover, in comparison to the model trained solely on avatar data, the few-shot model demonstrates high generalizability and emerges as a robust performer, showcasing its ability to effectively classify actions across various types of avatars.