Ayushi Priyadashani

My Project

Abstract

​

Dance choreography involves intricate coordination between music and complex movement, making automation highly challenging. Recent advancements in AI-driven choreography, particularly with models like the Full Attention Cross-modal Transformer (FACT) and pose map-based systems, present exciting opportunities. FACT models generate high-quality dance sequences conditioned on music, showcasing the power of cross-modal architectures. While challenges like determinism and missing physical constraints persist, these AI systems demonstrate the immense potential to revolutionize the technical and creative aspects of dance.

​

Introduction

​

Dance, a universal form of human expression, has become increasingly popular in the digital age, with viral phenomena like "Baby Shark" and "Gangnam Style" showcasing its global appeal. However, generating new and unique choreography is challenging due to the vast corpus of existing movements and the intersection of art and technology.

Artificial Intelligence (AI) has emerged as a promising tool in choreography, enabling the creation of synchronized dance sequences and opening new creative possibilities. Systems like the Full Attention Cross-modal Transformer (FACT) and pose-based models aim to overcome these challenges by leveraging massive datasets like AIST++, which contains over 56 hours of 3D dance data.

The integration of AI in choreography serves as a collaborative tool, pushing the boundaries of creative exploration in dance. By generating sequences that inspire choreographers to explore new artistic possibilities, AI enhances both technical efficiency and creative artistry.

​

Methodology

The FACT model utilizes cross-modal learning to generate dance sequences conditioned on music by fusing audio and motion embeddings. Key features include:

Transformer Network: Focuses on learning music-motion correlations.
Scalable Data Structures: Processes large datasets like AIST++ for realistic motion generation.
Evaluation Metrics: Includes Frechet Inception Distance (FID) and Beat Alignment Score (BeatAlign) to assess synchronization and quality.
Pose Map Systems
A complementary approach involves pose maps, which encode movements as spatial trajectories. Using kinematic principles, these maps facilitate exploration of physical constraints and new motion patterns. This system extends beyond traditional autoencoders to generate novel trajectories while maintaining realism.

​

Conclusion

AI-driven choreography shows significant promise in both technical and creative domains. The FACT model excels in generating synchronized, music-driven sequences, while pose maps expand the boundaries of improvisational choreography. Challenges such as determinism and physical realism remain, but the synergy of AI systems like FACT and pose maps underscores the potential for collaborative artistic processes. Future directions include improving physical constraints, incorporating dancer feedback, and fostering real-time creative collaboration.

​

Acknowledgments

Special thanks to Mohit Nadkarni and the BrainBridge Connections team for their guidance and mentorship. This work would not have been possible without their support.

References

Yang, W., & Woojin, M. (2015). “Kinematic Imagination: Exploring the Possibilities of Combining AI and Dance.” Proceedings of the Twenty-fourth International Joint Conference on Artificial Intelligence (IJCAI).
Tiepong R., Shan Y.W., David A. R., & Ross K. (2021). “The Influence of Artificial Intelligence in Dance Choreography.” Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV).
“Exploring AI’s Role as a Facilitator in Dance Creativity.” (2021). IPHS 200: Programming Humanity.