Opera Performance Pose Transfer Using Transformer Models from an Embodiment Perspective

This study addresses the issues of incomplete motion-capture data, as well as insufficient detail and temporal continuity in existing opera performance datasets, and aims to enable high-fidelity motion transfer between different performers. Within the framework of embodiment theory, the body is conceptualized as a dynamic carrier of action, while serving as a central component of performance cognition that encodes performers' cognition, emotions, and sense of rhythm. From the perspective of practical adaptability, the body extends beyond “physical training” into “performance communication,” and its movement patterns directly shape audience perception and artistic expression. Accordingly, this study integrates body posture, rhythmic perception, mechanical constraints, and culturally grounded aesthetic significance into the modeling of motion features, thereby ensuring that the generated transfer results better conform to the physical logic of opera performance. The proposed method enhances the model's understanding of motion by embedding performers' bodily constraints and expressive intentions into the feature representation of actions. In addition, a multi-scale temporal self-attention module is designed to simultaneously capture short-term motion details and long-term structural dependencies, ensuring both fine-grained realism and global coherence during the transfer process. To enable style transfer across different performers, a conditional pose generation strategy is further introduced, improving the diversity and personalization of generated results. Experiments were conducted on the public AIST Dance Video Database, comparing a conventional Long Short-Term Memory sequence-to-sequence (LSTM Seq2Seq) model, a Graph Convolutional Network (GCN), and the proposed improved Transformer model. The results show that, in terms of Mean Per Joint Position Error (MPJPE), the Transformer achieved 10.6 cm, representing reductions of 42.1% and 30.3% compared with LSTM Seq2Seq (18.3 cm) and GCN (15.2 cm), respectively. For Fréchet Inception Distance (FID), which evaluates the naturalness of generated motion, the Transformer obtained a score of 28.7, outperforming LSTM Seq2Seq (42.5) and GCN (36.8). Regarding Acceleration Error (ACC), the Transformer achieved 0.034, whereas LSTM Seq2Seq and GCN reached 0.057 and 0.046, respectively. These results demonstrate that the proposed method outperforms baseline models in motion accuracy, naturalness, and smoothness. Overall, the findings confirm that embedding performers' embodied sensations and expressive intentions into a Transformer-based architecture enables more accurate and natural motion transfer. This approach provides a practical pathway for future applications, including digitally assisted opera education and the development of virtual opera performers.