Scalable Group Choreography via Variational Phase Manifold Learning

TL;DR

This study proposes a phase-based variational generative model for group dance generation on learning a generative manifold that achieves high-fidelity group dance motion and enables the generation with an unlimited number of dancers while consuming only a minimal and constant amount of memory.

Abstract

Generating group dance motion from the music is a challenging task with several industrial applications. Although several methods have been proposed to tackle this problem, most of them prioritize optimizing the fidelity in dancing movement, constrained by predetermined dancer counts in datasets. This limitation impedes adaptability to real-world applications. Our study addresses the scalability problem in group choreography while preserving naturalness and synchronization. In particular, we propose a phase-based variational generative model for group dance generation on learning a generative manifold. Our method achieves high-fidelity group dance motion and enables the generation with an unlimited number of dancers while consuming only a minimal and constant amount of memory. The intensive experiments on two public datasets show that our proposed method outperforms recent state-of-the-art approaches by a large margin and is scalable to a great number of dancers beyond the training data.

Figures (5)

• Figure 2: Overview of our Phase-conditioned Dance VAE (PDVAE) for scalable group dance generation. It consists of an Encoder, a Prior, and a Decoder network. During training, we encode the corresponding motion and music inputs into a latent phase manifold, which is variational and parameterized by the frequency domain parameters of periodic functions. The latent phases can be sampled from the manifold and then decoded back to the original data space to obtain new motions. The consistency loss $\mathcal{L}_{\text{csc}}$ is further imposed to constrain the manifold to be consistently unified for dancers that belong to the same group. At inference stage, only the Prior and the Decoder are used to synthesize group dances efficiently.
• Figure 3: Visualization of a dancing sample between different methods. GDanceR displays monotonous, repetitive, or sinking dance motions. GCD exhibits more divergence in dance motions, yet dancers may intersect since their optimization does not address this issue explicitly. Blue boxes mark these issues. In contrast, our manifold-based solution ensures the divergence of dancing motions, while the phase motion path demonstrates its effectiveness in addressing floating and crossing issues in group dances.
• Figure 4: Memory usage vs. number of dancers in different dance generators.
• Figure 5: Visualization of scalable dancers.
• Figure 6: Realism between different methods when number of dancers is varied.